Update bcachefs sources to 717b356d1d bcachefs: Convert journal validation to bkey_invalid_flags

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

52 files changed, 5775 insertions, 5467 deletions

diff --git a/.bcachefs_revision b/.bcachefs_revision
index b2d87414..d78b2e28 100644
--- a/.bcachefs_revision
+++ b/.bcachefs_revision
@@ -1 +1 @@
-5b8c4a1366df20bc043404cb882230ce86296590
+717b356d1dfdf178ac46e217c81bb710b7e77032
diff --git a/cmd_device.c b/cmd_device.c
index c59d3709..1914629b 100644
--- a/cmd_device.c
+++ b/cmd_device.c
@@ -16,6 +16,7 @@
 #include "libbcachefs/bcachefs_ioctl.h"
 #include "libbcachefs/errcode.h"
 #include "libbcachefs/journal.h"
+#include "libbcachefs/sb-members.h"
 #include "libbcachefs/super-io.h"
 #include "cmds.h"
 #include "libbcachefs.h"
diff --git a/cmd_dump.c b/cmd_dump.c
index ab94707a..e1f0a5aa 100644
--- a/cmd_dump.c
+++ b/cmd_dump.c
@@ -13,6 +13,7 @@
 #include "libbcachefs/btree_iter.h"
 #include "libbcachefs/error.h"
 #include "libbcachefs/extents.h"
+#include "libbcachefs/sb-members.h"
 #include "libbcachefs/super.h"
 
 static void dump_usage(void)
diff --git a/cmd_kill_btree_node.c b/cmd_kill_btree_node.c
index a8915a1f..e9b8265d 100644
--- a/cmd_kill_btree_node.c
+++ b/cmd_kill_btree_node.c
@@ -11,6 +11,7 @@
 #include "libbcachefs/btree_iter.h"
 #include "libbcachefs/errcode.h"
 #include "libbcachefs/error.h"
+#include "libbcachefs/sb-members.h"
 #include "libbcachefs/super.h"
 
 static void kill_btree_node_usage(void)
diff --git a/libbcachefs/alloc_foreground.h b/libbcachefs/alloc_foreground.h
index fee195f7..7aaeec44 100644
--- a/libbcachefs/alloc_foreground.h
+++ b/libbcachefs/alloc_foreground.h
@@ -5,7 +5,7 @@
 #include "bcachefs.h"
 #include "alloc_types.h"
 #include "extents.h"
-#include "super.h"
+#include "sb-members.h"
 
 #include <linux/hash.h>
 
diff --git a/libbcachefs/bcachefs.h b/libbcachefs/bcachefs.h
index e1f1e8e8..30b3d7b9 100644
--- a/libbcachefs/bcachefs.h
+++ b/libbcachefs/bcachefs.h
@@ -294,8 +294,8 @@ do {									\
 
 #define bch_err_fn(_c, _ret)						\
 	 bch_err(_c, "%s(): error %s", __func__, bch2_err_str(_ret))
-#define bch_err_msg(_c, _ret, _msg)					\
-	 bch_err(_c, "%s(): error " _msg " %s", __func__, bch2_err_str(_ret))
+#define bch_err_msg(_c, _ret, _msg, ...)				\
+	 bch_err(_c, "%s(): error " _msg " %s", __func__, ##__VA_ARGS__, bch2_err_str(_ret))
 
 #define bch_verbose(c, fmt, ...)					\
 do {									\
@@ -995,6 +995,7 @@ struct bch_fs {
 	enum bch_recovery_pass	curr_recovery_pass;
 	/* bitmap of explicitly enabled recovery passes: */
 	u64			recovery_passes_explicit;
+	u64			recovery_passes_complete;
 
 	/* DEBUG JUNK */
 	struct dentry		*fs_debug_dir;
@@ -1139,22 +1140,6 @@ static inline bool bch2_dev_exists2(const struct bch_fs *c, unsigned dev)
 	return dev < c->sb.nr_devices && c->devs[dev];
 }
 
-/*
- * For when we need to rewind recovery passes and run a pass we skipped:
- */
-static inline int bch2_run_explicit_recovery_pass(struct bch_fs *c,
-						  enum bch_recovery_pass pass)
-{
-	c->recovery_passes_explicit |= BIT_ULL(pass);
-
-	if (c->curr_recovery_pass >= pass) {
-		c->curr_recovery_pass = pass;
-		return -BCH_ERR_restart_recovery;
-	} else {
-		return 0;
-	}
-}
-
 #define BKEY_PADDED_ONSTACK(key, pad)				\
 	struct { struct bkey_i key; __u64 key ## _pad[pad]; }
 
diff --git a/libbcachefs/bkey.c b/libbcachefs/bkey.c
index ee7ba700..d6960e25 100644
--- a/libbcachefs/bkey.c
+++ b/libbcachefs/bkey.c
@@ -7,14 +7,6 @@
 #include "bset.h"
 #include "util.h"
 
-#undef EBUG_ON
-
-#ifdef DEBUG_BKEYS
-#define EBUG_ON(cond)		BUG_ON(cond)
-#else
-#define EBUG_ON(cond)
-#endif
-
 const struct bkey_format bch2_bkey_format_current = BKEY_FORMAT_CURRENT;
 
 void bch2_bkey_packed_to_binary_text(struct printbuf *out,
@@ -185,6 +177,28 @@ static u64 get_inc_field(struct unpack_state *state, unsigned field)
 }
 
 __always_inline
+static void __set_inc_field(struct pack_state *state, unsigned field, u64 v)
+{
+	unsigned bits = state->format->bits_per_field[field];
+
+	if (bits) {
+		if (bits > state->bits) {
+			bits -= state->bits;
+			/* avoid shift by 64 if bits is 64 - bits is never 0 here: */
+			state->w |= (v >> 1) >> (bits - 1);
+
+			*state->p = state->w;
+			state->p = next_word(state->p);
+			state->w = 0;
+			state->bits = 64;
+		}
+
+		state->bits -= bits;
+		state->w |= v << state->bits;
+	}
+}
+
+__always_inline
 static bool set_inc_field(struct pack_state *state, unsigned field, u64 v)
 {
 	unsigned bits = state->format->bits_per_field[field];
@@ -198,20 +212,7 @@ static bool set_inc_field(struct pack_state *state, unsigned field, u64 v)
 	if (fls64(v) > bits)
 		return false;
 
-	if (bits > state->bits) {
-		bits -= state->bits;
-		/* avoid shift by 64 if bits is 0 - bits is never 64 here: */
-		state->w |= (v >> 1) >> (bits - 1);
-
-		*state->p = state->w;
-		state->p = next_word(state->p);
-		state->w = 0;
-		state->bits = 64;
-	}
-
-	state->bits -= bits;
-	state->w |= v << state->bits;
-
+	__set_inc_field(state, field, v);
 	return true;
 }
 
@@ -380,19 +381,7 @@ static bool set_inc_field_lossy(struct pack_state *state, unsigned field, u64 v)
 		ret = false;
 	}
 
-	if (bits > state->bits) {
-		bits -= state->bits;
-		state->w |= (v >> 1) >> (bits - 1);
-
-		*state->p = state->w;
-		state->p = next_word(state->p);
-		state->w = 0;
-		state->bits = 64;
-	}
-
-	state->bits -= bits;
-	state->w |= v << state->bits;
-
+	__set_inc_field(state, field, v);
 	return ret;
 }
 
@@ -435,6 +424,24 @@ static bool bkey_packed_successor(struct bkey_packed *out,
 
 	return false;
 }
+
+static bool bkey_format_has_too_big_fields(const struct bkey_format *f)
+{
+	for (unsigned i = 0; i < f->nr_fields; i++) {
+		unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
+		u64 unpacked_max = ~((~0ULL << 1) << (unpacked_bits - 1));
+		u64 packed_max = f->bits_per_field[i]
+			? ~((~0ULL << 1) << (f->bits_per_field[i] - 1))
+			: 0;
+		u64 field_offset = le64_to_cpu(f->field_offset[i]);
+
+		if (packed_max + field_offset < packed_max ||
+		    packed_max + field_offset > unpacked_max)
+			return true;
+	}
+
+	return false;
+}
 #endif
 
 /*
@@ -515,7 +522,8 @@ enum bkey_pack_pos_ret bch2_bkey_pack_pos_lossy(struct bkey_packed *out,
 
 		BUG_ON(bkey_cmp_left_packed(b, out, &orig) >= 0);
 		BUG_ON(bkey_packed_successor(&successor, b, *out) &&
-		       bkey_cmp_left_packed(b, &successor, &orig) < 0);
+		       bkey_cmp_left_packed(b, &successor, &orig) < 0 &&
+		       !bkey_format_has_too_big_fields(f));
 	}
 #endif
 
@@ -604,40 +612,74 @@ struct bkey_format bch2_bkey_format_done(struct bkey_format_state *s)
 		}
 	}
 
-	EBUG_ON(bch2_bkey_format_validate(&ret));
+#ifdef CONFIG_BCACHEFS_DEBUG
+	{
+		struct printbuf buf = PRINTBUF;
+
+		BUG_ON(bch2_bkey_format_invalid(NULL, &ret, 0, &buf));
+		printbuf_exit(&buf);
+	}
+#endif
 	return ret;
 }
 
-const char *bch2_bkey_format_validate(struct bkey_format *f)
+int bch2_bkey_format_invalid(struct bch_fs *c,
+			     struct bkey_format *f,
+			     enum bkey_invalid_flags flags,
+			     struct printbuf *err)
 {
 	unsigned i, bits = KEY_PACKED_BITS_START;
 
-	if (f->nr_fields != BKEY_NR_FIELDS)
-		return "incorrect number of fields";
+	if (f->nr_fields != BKEY_NR_FIELDS) {
+		prt_printf(err, "incorrect number of fields: got %u, should be %u",
+			   f->nr_fields, BKEY_NR_FIELDS);
+		return -BCH_ERR_invalid;
+	}
 
 	/*
 	 * Verify that the packed format can't represent fields larger than the
 	 * unpacked format:
 	 */
 	for (i = 0; i < f->nr_fields; i++) {
-		unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
-		u64 unpacked_max = ~((~0ULL << 1) << (unpacked_bits - 1));
-		u64 packed_max = f->bits_per_field[i]
-			? ~((~0ULL << 1) << (f->bits_per_field[i] - 1))
-			: 0;
-		u64 field_offset = le64_to_cpu(f->field_offset[i]);
-
-		if (packed_max + field_offset < packed_max ||
-		    packed_max + field_offset > unpacked_max)
-			return "field too large";
+		if (!c || c->sb.version_min >= bcachefs_metadata_version_snapshot) {
+			unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
+			u64 unpacked_max = ~((~0ULL << 1) << (unpacked_bits - 1));
+			u64 packed_max = f->bits_per_field[i]
+				? ~((~0ULL << 1) << (f->bits_per_field[i] - 1))
+				: 0;
+			u64 field_offset = le64_to_cpu(f->field_offset[i]);
+
+			if (packed_max + field_offset < packed_max ||
+			    packed_max + field_offset > unpacked_max) {
+				prt_printf(err, "field %u too large: %llu + %llu > %llu",
+					   i, packed_max, field_offset, unpacked_max);
+				return -BCH_ERR_invalid;
+			}
+		}
 
 		bits += f->bits_per_field[i];
 	}
 
-	if (f->key_u64s != DIV_ROUND_UP(bits, 64))
-		return "incorrect key_u64s";
+	if (f->key_u64s != DIV_ROUND_UP(bits, 64)) {
+		prt_printf(err, "incorrect key_u64s: got %u, should be %u",
+			   f->key_u64s, DIV_ROUND_UP(bits, 64));
+		return -BCH_ERR_invalid;
+	}
+
+	return 0;
+}
 
-	return NULL;
+void bch2_bkey_format_to_text(struct printbuf *out, const struct bkey_format *f)
+{
+	prt_printf(out, "u64s %u fields ", f->key_u64s);
+
+	for (unsigned i = 0; i < ARRAY_SIZE(f->bits_per_field); i++) {
+		if (i)
+			prt_str(out, ", ");
+		prt_printf(out, "%u:%llu",
+			   f->bits_per_field[i],
+			   le64_to_cpu(f->field_offset[i]));
+	}
 }
 
 /*
diff --git a/libbcachefs/bkey.h b/libbcachefs/bkey.h
index e81fb3e0..51969a46 100644
--- a/libbcachefs/bkey.h
+++ b/libbcachefs/bkey.h
@@ -9,6 +9,12 @@
 #include "util.h"
 #include "vstructs.h"
 
+enum bkey_invalid_flags {
+	BKEY_INVALID_WRITE		= (1U << 0),
+	BKEY_INVALID_COMMIT		= (1U << 1),
+	BKEY_INVALID_JOURNAL		= (1U << 2),
+};
+
 #if 0
 
 /*
@@ -769,6 +775,8 @@ static inline void bch2_bkey_format_add_key(struct bkey_format_state *s, const s
 
 void bch2_bkey_format_add_pos(struct bkey_format_state *, struct bpos);
 struct bkey_format bch2_bkey_format_done(struct bkey_format_state *);
-const char *bch2_bkey_format_validate(struct bkey_format *);
+int bch2_bkey_format_invalid(struct bch_fs *, struct bkey_format *,
+			     enum bkey_invalid_flags, struct printbuf *);
+void bch2_bkey_format_to_text(struct printbuf *, const struct bkey_format *);
 
 #endif /* _BCACHEFS_BKEY_H */
diff --git a/libbcachefs/bkey_methods.h b/libbcachefs/bkey_methods.h
index d7b63769..668f595e 100644
--- a/libbcachefs/bkey_methods.h
+++ b/libbcachefs/bkey_methods.h
@@ -13,12 +13,6 @@ enum btree_node_type;
 extern const char * const bch2_bkey_types[];
 extern const struct bkey_ops bch2_bkey_null_ops;
 
-enum bkey_invalid_flags {
-	BKEY_INVALID_WRITE		= (1U << 0),
-	BKEY_INVALID_COMMIT		= (1U << 1),
-	BKEY_INVALID_JOURNAL		= (1U << 2),
-};
-
 /*
  * key_invalid: checks validity of @k, returns 0 if good or -EINVAL if bad. If
  * invalid, entire key will be deleted.
diff --git a/libbcachefs/btree_cache.c b/libbcachefs/btree_cache.c
index 13c88d95..a8283fdc 100644
--- a/libbcachefs/btree_cache.c
+++ b/libbcachefs/btree_cache.c
@@ -1214,7 +1214,6 @@ wait_on_io:
 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
 			     const struct btree *b)
 {
-	const struct bkey_format *f = &b->format;
 	struct bset_stats stats;
 
 	memset(&stats, 0, sizeof(stats));
@@ -1228,9 +1227,13 @@ void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
 	prt_printf(out, ":\n"
 	       "    ptrs: ");
 	bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
+	prt_newline(out);
 
-	prt_printf(out, "\n"
-	       "    format: u64s %u fields %u %u %u %u %u\n"
+	prt_printf(out,
+	       "    format: ");
+	bch2_bkey_format_to_text(out, &b->format);
+
+	prt_printf(out,
 	       "    unpack fn len: %u\n"
 	       "    bytes used %zu/%zu (%zu%% full)\n"
 	       "    sib u64s: %u, %u (merge threshold %u)\n"
@@ -1238,12 +1241,6 @@ void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
 	       "    nr unpacked keys %u\n"
 	       "    floats %zu\n"
 	       "    failed unpacked %zu\n",
-	       f->key_u64s,
-	       f->bits_per_field[0],
-	       f->bits_per_field[1],
-	       f->bits_per_field[2],
-	       f->bits_per_field[3],
-	       f->bits_per_field[4],
 	       b->unpack_fn_len,
 	       b->nr.live_u64s * sizeof(u64),
 	       btree_bytes(c) - sizeof(struct btree_node),
diff --git a/libbcachefs/btree_gc.c b/libbcachefs/btree_gc.c
index 49e9822d..a5f685ff 100644
--- a/libbcachefs/btree_gc.c
+++ b/libbcachefs/btree_gc.c
@@ -9,6 +9,7 @@
 #include "alloc_foreground.h"
 #include "bkey_methods.h"
 #include "bkey_buf.h"
+#include "btree_journal_iter.h"
 #include "btree_key_cache.h"
 #include "btree_locking.h"
 #include "btree_update_interior.h"
@@ -43,7 +44,7 @@
 static bool should_restart_for_topology_repair(struct bch_fs *c)
 {
 	return c->opts.fix_errors != FSCK_FIX_no &&
-		!(c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology));
+		!(c->recovery_passes_complete & BIT_ULL(BCH_RECOVERY_PASS_check_topology));
 }
 
 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
diff --git a/libbcachefs/btree_gc.h b/libbcachefs/btree_gc.h
index b45e382f..607575f8 100644
--- a/libbcachefs/btree_gc.h
+++ b/libbcachefs/btree_gc.h
@@ -2,6 +2,7 @@
 #ifndef _BCACHEFS_BTREE_GC_H
 #define _BCACHEFS_BTREE_GC_H
 
+#include "bkey.h"
 #include "btree_types.h"
 
 int bch2_check_topology(struct bch_fs *);
diff --git a/libbcachefs/btree_io.c b/libbcachefs/btree_io.c
index c049876e..3b654841 100644
--- a/libbcachefs/btree_io.c
+++ b/libbcachefs/btree_io.c
@@ -17,6 +17,7 @@
 #include "io.h"
 #include "journal_reclaim.h"
 #include "journal_seq_blacklist.h"
+#include "recovery.h"
 #include "super-io.h"
 #include "trace.h"
 
@@ -543,31 +544,7 @@ static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 	prt_str(out, ": ");
 }
 
-enum btree_err_type {
-	/*
-	 * We can repair this locally, and we're after the checksum check so
-	 * there's no need to try another replica:
-	 */
-	BTREE_ERR_FIXABLE,
-	/*
-	 * We can repair this if we have to, but we should try reading another
-	 * replica if we can:
-	 */
-	BTREE_ERR_WANT_RETRY,
-	/*
-	 * Read another replica if we have one, otherwise consider the whole
-	 * node bad:
-	 */
-	BTREE_ERR_MUST_RETRY,
-	BTREE_ERR_BAD_NODE,
-	BTREE_ERR_INCOMPATIBLE,
-};
-
-enum btree_validate_ret {
-	BTREE_RETRY_READ = 64,
-};
-
-static int __btree_err(enum btree_err_type type,
+static int __btree_err(int ret,
 		       struct bch_fs *c,
 		       struct bch_dev *ca,
 		       struct btree *b,
@@ -578,7 +555,6 @@ static int __btree_err(enum btree_err_type type,
 {
 	struct printbuf out = PRINTBUF;
 	va_list args;
-	int ret = -BCH_ERR_fsck_fix;
 
 	btree_err_msg(&out, c, ca, b, i, b->written, write);
 
@@ -594,27 +570,26 @@ static int __btree_err(enum btree_err_type type,
 		goto out;
 	}
 
-	if (!have_retry && type == BTREE_ERR_WANT_RETRY)
-		type = BTREE_ERR_FIXABLE;
-	if (!have_retry && type == BTREE_ERR_MUST_RETRY)
-		type = BTREE_ERR_BAD_NODE;
+	if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry)
+		ret = -BCH_ERR_btree_node_read_err_fixable;
+	if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry)
+		ret = -BCH_ERR_btree_node_read_err_bad_node;
 
-	switch (type) {
-	case BTREE_ERR_FIXABLE:
+	switch (ret) {
+	case -BCH_ERR_btree_node_read_err_fixable:
 		mustfix_fsck_err(c, "%s", out.buf);
 		ret = -BCH_ERR_fsck_fix;
 		break;
-	case BTREE_ERR_WANT_RETRY:
-	case BTREE_ERR_MUST_RETRY:
+	case -BCH_ERR_btree_node_read_err_want_retry:
+	case -BCH_ERR_btree_node_read_err_must_retry:
 		bch2_print_string_as_lines(KERN_ERR, out.buf);
-		ret = BTREE_RETRY_READ;
 		break;
-	case BTREE_ERR_BAD_NODE:
+	case -BCH_ERR_btree_node_read_err_bad_node:
 		bch2_print_string_as_lines(KERN_ERR, out.buf);
 		bch2_topology_error(c);
 		ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology) ?: -EIO;
 		break;
-	case BTREE_ERR_INCOMPATIBLE:
+	case -BCH_ERR_btree_node_read_err_incompatible:
 		bch2_print_string_as_lines(KERN_ERR, out.buf);
 		ret = -BCH_ERR_fsck_errors_not_fixed;
 		break;
@@ -631,8 +606,11 @@ fsck_err:
 ({									\
 	int _ret = __btree_err(type, c, ca, b, i, write, have_retry, msg, ##__VA_ARGS__);\
 									\
-	if (_ret != -BCH_ERR_fsck_fix)					\
+	if (_ret != -BCH_ERR_fsck_fix) {				\
+		ret = _ret;						\
 		goto fsck_err;						\
+	}								\
+									\
 	*saw_error = true;						\
 })
 
@@ -696,19 +674,18 @@ static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 			 int write, bool have_retry, bool *saw_error)
 {
 	unsigned version = le16_to_cpu(i->version);
-	const char *err;
 	struct printbuf buf1 = PRINTBUF;
 	struct printbuf buf2 = PRINTBUF;
 	int ret = 0;
 
 	btree_err_on(!bch2_version_compatible(version),
-		     BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
+		     -BCH_ERR_btree_node_read_err_incompatible, c, ca, b, i,
 		     "unsupported bset version %u.%u",
 		     BCH_VERSION_MAJOR(version),
 		     BCH_VERSION_MINOR(version));
 
 	if (btree_err_on(version < c->sb.version_min,
-			 BTREE_ERR_FIXABLE, c, NULL, b, i,
+			 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 			 "bset version %u older than superblock version_min %u",
 			 version, c->sb.version_min)) {
 		mutex_lock(&c->sb_lock);
@@ -719,7 +696,7 @@ static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 
 	if (btree_err_on(BCH_VERSION_MAJOR(version) >
 			 BCH_VERSION_MAJOR(c->sb.version),
-			 BTREE_ERR_FIXABLE, c, NULL, b, i,
+			 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 			 "bset version %u newer than superblock version %u",
 			 version, c->sb.version)) {
 		mutex_lock(&c->sb_lock);
@@ -729,11 +706,11 @@ static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 	}
 
 	btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
-		     BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
+		     -BCH_ERR_btree_node_read_err_incompatible, c, ca, b, i,
 		     "BSET_SEPARATE_WHITEOUTS no longer supported");
 
 	if (btree_err_on(offset + sectors > btree_sectors(c),
-			 BTREE_ERR_FIXABLE, c, ca, b, i,
+			 -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
 			 "bset past end of btree node")) {
 		i->u64s = 0;
 		ret = 0;
@@ -741,12 +718,12 @@ static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 	}
 
 	btree_err_on(offset && !i->u64s,
-		     BTREE_ERR_FIXABLE, c, ca, b, i,
+		     -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
 		     "empty bset");
 
 	btree_err_on(BSET_OFFSET(i) &&
 		     BSET_OFFSET(i) != offset,
-		     BTREE_ERR_WANT_RETRY, c, ca, b, i,
+		     -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 		     "bset at wrong sector offset");
 
 	if (!offset) {
@@ -760,16 +737,16 @@ static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 
 			/* XXX endianness */
 			btree_err_on(bp->seq != bn->keys.seq,
-				     BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
+				     -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 				     "incorrect sequence number (wrong btree node)");
 		}
 
 		btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
-			     BTREE_ERR_MUST_RETRY, c, ca, b, i,
+			     -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
 			     "incorrect btree id");
 
 		btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
-			     BTREE_ERR_MUST_RETRY, c, ca, b, i,
+			     -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
 			     "incorrect level");
 
 		if (!write)
@@ -786,7 +763,7 @@ static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 			}
 
 			btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
-				     BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
+				     -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 				     "incorrect min_key: got %s should be %s",
 				     (printbuf_reset(&buf1),
 				      bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
@@ -795,7 +772,7 @@ static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 		}
 
 		btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
-			     BTREE_ERR_MUST_RETRY, c, ca, b, i,
+			     -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
 			     "incorrect max key %s",
 			     (printbuf_reset(&buf1),
 			      bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
@@ -804,10 +781,12 @@ static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 			compat_btree_node(b->c.level, b->c.btree_id, version,
 					  BSET_BIG_ENDIAN(i), write, bn);
 
-		err = bch2_bkey_format_validate(&bn->format);
-		btree_err_on(err,
-			     BTREE_ERR_BAD_NODE, c, ca, b, i,
-			     "invalid bkey format: %s", err);
+		btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1),
+			     -BCH_ERR_btree_node_read_err_bad_node, c, ca, b, i,
+			     "invalid bkey format: %s\n  %s", buf1.buf,
+			     (printbuf_reset(&buf2),
+			      bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf));
+		printbuf_reset(&buf1);
 
 		compat_bformat(b->c.level, b->c.btree_id, version,
 			       BSET_BIG_ENDIAN(i), write,
@@ -847,14 +826,14 @@ static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 		struct bkey tmp;
 
 		if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
-				 BTREE_ERR_FIXABLE, c, NULL, b, i,
+				 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 				 "key extends past end of bset")) {
 			i->u64s = cpu_to_le16((u64 *) k - i->_data);
 			break;
 		}
 
 		if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
-				 BTREE_ERR_FIXABLE, c, NULL, b, i,
+				 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 				 "invalid bkey format %u", k->format)) {
 			i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 			memmove_u64s_down(k, bkey_p_next(k),
@@ -878,7 +857,7 @@ static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 			prt_printf(&buf, "\n  ");
 			bch2_bkey_val_to_text(&buf, c, u.s_c);
 
-			btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
+			btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf);
 
 			i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 			memmove_u64s_down(k, bkey_p_next(k),
@@ -902,7 +881,7 @@ static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 
 			bch2_dump_bset(c, b, i, 0);
 
-			if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf)) {
+			if (btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf)) {
 				i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 				memmove_u64s_down(k, bkey_p_next(k),
 						  (u64 *) vstruct_end(i) - (u64 *) k);
@@ -945,16 +924,16 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 	iter->size = (btree_blocks(c) + 1) * 2;
 
 	if (bch2_meta_read_fault("btree"))
-		btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
+		btree_err(-BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 			  "dynamic fault");
 
 	btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
-		     BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
+		     -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 		     "bad magic: want %llx, got %llx",
 		     bset_magic(c), le64_to_cpu(b->data->magic));
 
 	btree_err_on(!b->data->keys.seq,
-		     BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
+		     -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 		     "bad btree header: seq 0");
 
 	if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
@@ -962,7 +941,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 			&bkey_i_to_btree_ptr_v2(&b->key)->v;
 
 		btree_err_on(b->data->keys.seq != bp->seq,
-			     BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
+			     -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 			     "got wrong btree node (seq %llx want %llx)",
 			     b->data->keys.seq, bp->seq);
 	}
@@ -977,7 +956,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 			i = &b->data->keys;
 
 			btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
-				     BTREE_ERR_WANT_RETRY, c, ca, b, i,
+				     -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 				     "unknown checksum type %llu",
 				     BSET_CSUM_TYPE(i));
 
@@ -985,7 +964,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 			csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 
 			btree_err_on(bch2_crc_cmp(csum, b->data->csum),
-				     BTREE_ERR_WANT_RETRY, c, ca, b, i,
+				     -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 				     "invalid checksum");
 
 			ret = bset_encrypt(c, i, b->written << 9);
@@ -995,7 +974,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 
 			btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
 				     !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
-				     BTREE_ERR_INCOMPATIBLE, c, NULL, b, NULL,
+				     -BCH_ERR_btree_node_read_err_incompatible, c, NULL, b, NULL,
 				     "btree node does not have NEW_EXTENT_OVERWRITE set");
 
 			sectors = vstruct_sectors(b->data, c->block_bits);
@@ -1007,7 +986,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 				break;
 
 			btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
-				     BTREE_ERR_WANT_RETRY, c, ca, b, i,
+				     -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 				     "unknown checksum type %llu",
 				     BSET_CSUM_TYPE(i));
 
@@ -1015,7 +994,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 			csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 
 			btree_err_on(bch2_crc_cmp(csum, bne->csum),
-				     BTREE_ERR_WANT_RETRY, c, ca, b, i,
+				     -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 				     "invalid checksum");
 
 			ret = bset_encrypt(c, i, b->written << 9);
@@ -1048,12 +1027,12 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 					true);
 
 		btree_err_on(blacklisted && first,
-			     BTREE_ERR_FIXABLE, c, ca, b, i,
+			     -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
 			     "first btree node bset has blacklisted journal seq (%llu)",
 			     le64_to_cpu(i->journal_seq));
 
 		btree_err_on(blacklisted && ptr_written,
-			     BTREE_ERR_FIXABLE, c, ca, b, i,
+			     -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
 			     "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
 			     le64_to_cpu(i->journal_seq),
 			     b->written, b->written + sectors, ptr_written);
@@ -1072,7 +1051,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 
 	if (ptr_written) {
 		btree_err_on(b->written < ptr_written,
-			     BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
+			     -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, NULL,
 			     "btree node data missing: expected %u sectors, found %u",
 			     ptr_written, b->written);
 	} else {
@@ -1083,7 +1062,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 				     !bch2_journal_seq_is_blacklisted(c,
 								      le64_to_cpu(bne->keys.journal_seq),
 								      true),
-				     BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
+				     -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, NULL,
 				     "found bset signature after last bset");
 
 		/*
@@ -1137,7 +1116,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 			prt_printf(&buf, "\n  ");
 			bch2_bkey_val_to_text(&buf, c, u.s_c);
 
-			btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
+			btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf);
 
 			btree_keys_account_key_drop(&b->nr, 0, k);
 
@@ -1177,7 +1156,8 @@ out:
 	printbuf_exit(&buf);
 	return retry_read;
 fsck_err:
-	if (ret == BTREE_RETRY_READ)
+	if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
+	    ret == -BCH_ERR_btree_node_read_err_must_retry)
 		retry_read = 1;
 	else
 		set_btree_node_read_error(b);
@@ -1363,14 +1343,14 @@ static void btree_node_read_all_replicas_done(struct closure *cl)
 		}
 
 		written2 = btree_node_sectors_written(c, ra->buf[i]);
-		if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
+		if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
 				 "btree node sectors written mismatch: %u != %u",
 				 written, written2) ||
 		    btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
-				 BTREE_ERR_FIXABLE, c, NULL, b, NULL,
+				 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
 				 "found bset signature after last bset") ||
 		    btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
-				 BTREE_ERR_FIXABLE, c, NULL, b, NULL,
+				 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
 				 "btree node replicas content mismatch"))
 			dump_bset_maps = true;
 
diff --git a/libbcachefs/btree_iter.c b/libbcachefs/btree_iter.c
index dfb77b23..06dbb617 100644
--- a/libbcachefs/btree_iter.c
+++ b/libbcachefs/btree_iter.c
@@ -5,6 +5,7 @@
 #include "bkey_buf.h"
 #include "btree_cache.h"
 #include "btree_iter.h"
+#include "btree_journal_iter.h"
 #include "btree_key_cache.h"
 #include "btree_locking.h"
 #include "btree_update.h"
@@ -12,7 +13,6 @@
 #include "error.h"
 #include "extents.h"
 #include "journal.h"
-#include "recovery.h"
 #include "replicas.h"
 #include "subvolume.h"
 #include "trace.h"
diff --git a/libbcachefs/btree_iter.h b/libbcachefs/btree_iter.h
index c472aa8c..8876f2b8 100644
--- a/libbcachefs/btree_iter.h
+++ b/libbcachefs/btree_iter.h
@@ -221,6 +221,22 @@ struct btree_path *bch2_path_get(struct btree_trans *, enum btree_id, struct bpo
 				 unsigned, unsigned, unsigned, unsigned long);
 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *, struct bkey *);
 
+/*
+ * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
+ * different snapshot:
+ */
+static inline struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
+{
+	struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
+
+	if (k.k && bpos_eq(path->pos, k.k->p))
+		return k;
+
+	bkey_init(u);
+	u->p = path->pos;
+	return (struct bkey_s_c) { u, NULL };
+}
+
 struct bkey_i *bch2_btree_journal_peek_slot(struct btree_trans *,
 					struct btree_iter *, struct bpos);
 
diff --git a/libbcachefs/btree_journal_iter.c b/libbcachefs/btree_journal_iter.c
new file mode 100644
index 00000000..58a981bc
--- /dev/null
+++ b/libbcachefs/btree_journal_iter.c
@@ -0,0 +1,531 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "bset.h"
+#include "btree_journal_iter.h"
+#include "journal_io.h"
+
+#include <linux/sort.h>
+
+/*
+ * For managing keys we read from the journal: until journal replay works normal
+ * btree lookups need to be able to find and return keys from the journal where
+ * they overwrite what's in the btree, so we have a special iterator and
+ * operations for the regular btree iter code to use:
+ */
+
+static int __journal_key_cmp(enum btree_id	l_btree_id,
+			     unsigned		l_level,
+			     struct bpos	l_pos,
+			     const struct journal_key *r)
+{
+	return (cmp_int(l_btree_id,	r->btree_id) ?:
+		cmp_int(l_level,	r->level) ?:
+		bpos_cmp(l_pos,	r->k->k.p));
+}
+
+static int journal_key_cmp(const struct journal_key *l, const struct journal_key *r)
+{
+	return __journal_key_cmp(l->btree_id, l->level, l->k->k.p, r);
+}
+
+static inline size_t idx_to_pos(struct journal_keys *keys, size_t idx)
+{
+	size_t gap_size = keys->size - keys->nr;
+
+	if (idx >= keys->gap)
+		idx += gap_size;
+	return idx;
+}
+
+static inline struct journal_key *idx_to_key(struct journal_keys *keys, size_t idx)
+{
+	return keys->d + idx_to_pos(keys, idx);
+}
+
+static size_t __bch2_journal_key_search(struct journal_keys *keys,
+					enum btree_id id, unsigned level,
+					struct bpos pos)
+{
+	size_t l = 0, r = keys->nr, m;
+
+	while (l < r) {
+		m = l + ((r - l) >> 1);
+		if (__journal_key_cmp(id, level, pos, idx_to_key(keys, m)) > 0)
+			l = m + 1;
+		else
+			r = m;
+	}
+
+	BUG_ON(l < keys->nr &&
+	       __journal_key_cmp(id, level, pos, idx_to_key(keys, l)) > 0);
+
+	BUG_ON(l &&
+	       __journal_key_cmp(id, level, pos, idx_to_key(keys, l - 1)) <= 0);
+
+	return l;
+}
+
+static size_t bch2_journal_key_search(struct journal_keys *keys,
+				      enum btree_id id, unsigned level,
+				      struct bpos pos)
+{
+	return idx_to_pos(keys, __bch2_journal_key_search(keys, id, level, pos));
+}
+
+struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *c, enum btree_id btree_id,
+					   unsigned level, struct bpos pos,
+					   struct bpos end_pos, size_t *idx)
+{
+	struct journal_keys *keys = &c->journal_keys;
+	unsigned iters = 0;
+	struct journal_key *k;
+search:
+	if (!*idx)
+		*idx = __bch2_journal_key_search(keys, btree_id, level, pos);
+
+	while ((k = *idx < keys->nr ? idx_to_key(keys, *idx) : NULL)) {
+		if (__journal_key_cmp(btree_id, level, end_pos, k) < 0)
+			return NULL;
+
+		if (__journal_key_cmp(btree_id, level, pos, k) <= 0 &&
+		    !k->overwritten)
+			return k->k;
+
+		(*idx)++;
+		iters++;
+		if (iters == 10) {
+			*idx = 0;
+			goto search;
+		}
+	}
+
+	return NULL;
+}
+
+struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *c, enum btree_id btree_id,
+					   unsigned level, struct bpos pos)
+{
+	size_t idx = 0;
+
+	return bch2_journal_keys_peek_upto(c, btree_id, level, pos, pos, &idx);
+}
+
+static void journal_iters_fix(struct bch_fs *c)
+{
+	struct journal_keys *keys = &c->journal_keys;
+	/* The key we just inserted is immediately before the gap: */
+	size_t gap_end = keys->gap + (keys->size - keys->nr);
+	struct btree_and_journal_iter *iter;
+
+	/*
+	 * If an iterator points one after the key we just inserted, decrement
+	 * the iterator so it points at the key we just inserted - if the
+	 * decrement was unnecessary, bch2_btree_and_journal_iter_peek() will
+	 * handle that:
+	 */
+	list_for_each_entry(iter, &c->journal_iters, journal.list)
+		if (iter->journal.idx == gap_end)
+			iter->journal.idx = keys->gap - 1;
+}
+
+static void journal_iters_move_gap(struct bch_fs *c, size_t old_gap, size_t new_gap)
+{
+	struct journal_keys *keys = &c->journal_keys;
+	struct journal_iter *iter;
+	size_t gap_size = keys->size - keys->nr;
+
+	list_for_each_entry(iter, &c->journal_iters, list) {
+		if (iter->idx > old_gap)
+			iter->idx -= gap_size;
+		if (iter->idx >= new_gap)
+			iter->idx += gap_size;
+	}
+}
+
+int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
+				 unsigned level, struct bkey_i *k)
+{
+	struct journal_key n = {
+		.btree_id	= id,
+		.level		= level,
+		.k		= k,
+		.allocated	= true,
+		/*
+		 * Ensure these keys are done last by journal replay, to unblock
+		 * journal reclaim:
+		 */
+		.journal_seq	= U32_MAX,
+	};
+	struct journal_keys *keys = &c->journal_keys;
+	size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
+
+	BUG_ON(test_bit(BCH_FS_RW, &c->flags));
+
+	if (idx < keys->size &&
+	    journal_key_cmp(&n, &keys->d[idx]) == 0) {
+		if (keys->d[idx].allocated)
+			kfree(keys->d[idx].k);
+		keys->d[idx] = n;
+		return 0;
+	}
+
+	if (idx > keys->gap)
+		idx -= keys->size - keys->nr;
+
+	if (keys->nr == keys->size) {
+		struct journal_keys new_keys = {
+			.nr			= keys->nr,
+			.size			= max_t(size_t, keys->size, 8) * 2,
+		};
+
+		new_keys.d = kvmalloc_array(new_keys.size, sizeof(new_keys.d[0]), GFP_KERNEL);
+		if (!new_keys.d) {
+			bch_err(c, "%s: error allocating new key array (size %zu)",
+				__func__, new_keys.size);
+			return -BCH_ERR_ENOMEM_journal_key_insert;
+		}
+
+		/* Since @keys was full, there was no gap: */
+		memcpy(new_keys.d, keys->d, sizeof(keys->d[0]) * keys->nr);
+		kvfree(keys->d);
+		*keys = new_keys;
+
+		/* And now the gap is at the end: */
+		keys->gap = keys->nr;
+	}
+
+	journal_iters_move_gap(c, keys->gap, idx);
+
+	move_gap(keys->d, keys->nr, keys->size, keys->gap, idx);
+	keys->gap = idx;
+
+	keys->nr++;
+	keys->d[keys->gap++] = n;
+
+	journal_iters_fix(c);
+
+	return 0;
+}
+
+/*
+ * Can only be used from the recovery thread while we're still RO - can't be
+ * used once we've got RW, as journal_keys is at that point used by multiple
+ * threads:
+ */
+int bch2_journal_key_insert(struct bch_fs *c, enum btree_id id,
+			    unsigned level, struct bkey_i *k)
+{
+	struct bkey_i *n;
+	int ret;
+
+	n = kmalloc(bkey_bytes(&k->k), GFP_KERNEL);
+	if (!n)
+		return -BCH_ERR_ENOMEM_journal_key_insert;
+
+	bkey_copy(n, k);
+	ret = bch2_journal_key_insert_take(c, id, level, n);
+	if (ret)
+		kfree(n);
+	return ret;
+}
+
+int bch2_journal_key_delete(struct bch_fs *c, enum btree_id id,
+			    unsigned level, struct bpos pos)
+{
+	struct bkey_i whiteout;
+
+	bkey_init(&whiteout.k);
+	whiteout.k.p = pos;
+
+	return bch2_journal_key_insert(c, id, level, &whiteout);
+}
+
+void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
+				  unsigned level, struct bpos pos)
+{
+	struct journal_keys *keys = &c->journal_keys;
+	size_t idx = bch2_journal_key_search(keys, btree, level, pos);
+
+	if (idx < keys->size &&
+	    keys->d[idx].btree_id	== btree &&
+	    keys->d[idx].level		== level &&
+	    bpos_eq(keys->d[idx].k->k.p, pos))
+		keys->d[idx].overwritten = true;
+}
+
+static void bch2_journal_iter_advance(struct journal_iter *iter)
+{
+	if (iter->idx < iter->keys->size) {
+		iter->idx++;
+		if (iter->idx == iter->keys->gap)
+			iter->idx += iter->keys->size - iter->keys->nr;
+	}
+}
+
+static struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
+{
+	struct journal_key *k = iter->keys->d + iter->idx;
+
+	while (k < iter->keys->d + iter->keys->size &&
+	       k->btree_id	== iter->btree_id &&
+	       k->level		== iter->level) {
+		if (!k->overwritten)
+			return bkey_i_to_s_c(k->k);
+
+		bch2_journal_iter_advance(iter);
+		k = iter->keys->d + iter->idx;
+	}
+
+	return bkey_s_c_null;
+}
+
+static void bch2_journal_iter_exit(struct journal_iter *iter)
+{
+	list_del(&iter->list);
+}
+
+static void bch2_journal_iter_init(struct bch_fs *c,
+				   struct journal_iter *iter,
+				   enum btree_id id, unsigned level,
+				   struct bpos pos)
+{
+	iter->btree_id	= id;
+	iter->level	= level;
+	iter->keys	= &c->journal_keys;
+	iter->idx	= bch2_journal_key_search(&c->journal_keys, id, level, pos);
+}
+
+static struct bkey_s_c bch2_journal_iter_peek_btree(struct btree_and_journal_iter *iter)
+{
+	return bch2_btree_node_iter_peek_unpack(&iter->node_iter,
+						iter->b, &iter->unpacked);
+}
+
+static void bch2_journal_iter_advance_btree(struct btree_and_journal_iter *iter)
+{
+	bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
+}
+
+void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
+{
+	if (bpos_eq(iter->pos, SPOS_MAX))
+		iter->at_end = true;
+	else
+		iter->pos = bpos_successor(iter->pos);
+}
+
+struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *iter)
+{
+	struct bkey_s_c btree_k, journal_k, ret;
+again:
+	if (iter->at_end)
+		return bkey_s_c_null;
+
+	while ((btree_k = bch2_journal_iter_peek_btree(iter)).k &&
+	       bpos_lt(btree_k.k->p, iter->pos))
+		bch2_journal_iter_advance_btree(iter);
+
+	while ((journal_k = bch2_journal_iter_peek(&iter->journal)).k &&
+	       bpos_lt(journal_k.k->p, iter->pos))
+		bch2_journal_iter_advance(&iter->journal);
+
+	ret = journal_k.k &&
+		(!btree_k.k || bpos_le(journal_k.k->p, btree_k.k->p))
+		? journal_k
+		: btree_k;
+
+	if (ret.k && iter->b && bpos_gt(ret.k->p, iter->b->data->max_key))
+		ret = bkey_s_c_null;
+
+	if (ret.k) {
+		iter->pos = ret.k->p;
+		if (bkey_deleted(ret.k)) {
+			bch2_btree_and_journal_iter_advance(iter);
+			goto again;
+		}
+	} else {
+		iter->pos = SPOS_MAX;
+		iter->at_end = true;
+	}
+
+	return ret;
+}
+
+void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *iter)
+{
+	bch2_journal_iter_exit(&iter->journal);
+}
+
+void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
+						  struct bch_fs *c,
+						  struct btree *b,
+						  struct btree_node_iter node_iter,
+						  struct bpos pos)
+{
+	memset(iter, 0, sizeof(*iter));
+
+	iter->b = b;
+	iter->node_iter = node_iter;
+	bch2_journal_iter_init(c, &iter->journal, b->c.btree_id, b->c.level, pos);
+	INIT_LIST_HEAD(&iter->journal.list);
+	iter->pos = b->data->min_key;
+	iter->at_end = false;
+}
+
+/*
+ * this version is used by btree_gc before filesystem has gone RW and
+ * multithreaded, so uses the journal_iters list:
+ */
+void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
+						struct bch_fs *c,
+						struct btree *b)
+{
+	struct btree_node_iter node_iter;
+
+	bch2_btree_node_iter_init_from_start(&node_iter, b);
+	__bch2_btree_and_journal_iter_init_node_iter(iter, c, b, node_iter, b->data->min_key);
+	list_add(&iter->journal.list, &c->journal_iters);
+}
+
+/* sort and dedup all keys in the journal: */
+
+void bch2_journal_entries_free(struct bch_fs *c)
+{
+	struct journal_replay **i;
+	struct genradix_iter iter;
+
+	genradix_for_each(&c->journal_entries, iter, i)
+		if (*i)
+			kvpfree(*i, offsetof(struct journal_replay, j) +
+				vstruct_bytes(&(*i)->j));
+	genradix_free(&c->journal_entries);
+}
+
+/*
+ * When keys compare equal, oldest compares first:
+ */
+static int journal_sort_key_cmp(const void *_l, const void *_r)
+{
+	const struct journal_key *l = _l;
+	const struct journal_key *r = _r;
+
+	return  journal_key_cmp(l, r) ?:
+		cmp_int(l->journal_seq, r->journal_seq) ?:
+		cmp_int(l->journal_offset, r->journal_offset);
+}
+
+void bch2_journal_keys_free(struct journal_keys *keys)
+{
+	struct journal_key *i;
+
+	move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
+	keys->gap = keys->nr;
+
+	for (i = keys->d; i < keys->d + keys->nr; i++)
+		if (i->allocated)
+			kfree(i->k);
+
+	kvfree(keys->d);
+	keys->d = NULL;
+	keys->nr = keys->gap = keys->size = 0;
+}
+
+static void __journal_keys_sort(struct journal_keys *keys)
+{
+	struct journal_key *src, *dst;
+
+	sort(keys->d, keys->nr, sizeof(keys->d[0]), journal_sort_key_cmp, NULL);
+
+	src = dst = keys->d;
+	while (src < keys->d + keys->nr) {
+		while (src + 1 < keys->d + keys->nr &&
+		       src[0].btree_id	== src[1].btree_id &&
+		       src[0].level	== src[1].level &&
+		       bpos_eq(src[0].k->k.p, src[1].k->k.p))
+			src++;
+
+		*dst++ = *src++;
+	}
+
+	keys->nr = dst - keys->d;
+}
+
+int bch2_journal_keys_sort(struct bch_fs *c)
+{
+	struct genradix_iter iter;
+	struct journal_replay *i, **_i;
+	struct jset_entry *entry;
+	struct bkey_i *k;
+	struct journal_keys *keys = &c->journal_keys;
+	size_t nr_keys = 0, nr_read = 0;
+
+	genradix_for_each(&c->journal_entries, iter, _i) {
+		i = *_i;
+
+		if (!i || i->ignore)
+			continue;
+
+		for_each_jset_key(k, entry, &i->j)
+			nr_keys++;
+	}
+
+	if (!nr_keys)
+		return 0;
+
+	keys->size = roundup_pow_of_two(nr_keys);
+
+	keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
+	if (!keys->d) {
+		bch_err(c, "Failed to allocate buffer for sorted journal keys (%zu keys); trying slowpath",
+			nr_keys);
+
+		do {
+			keys->size >>= 1;
+			keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
+		} while (!keys->d && keys->size > nr_keys / 8);
+
+		if (!keys->d) {
+			bch_err(c, "Failed to allocate %zu size buffer for sorted journal keys; exiting",
+				keys->size);
+			return -BCH_ERR_ENOMEM_journal_keys_sort;
+		}
+	}
+
+	genradix_for_each(&c->journal_entries, iter, _i) {
+		i = *_i;
+
+		if (!i || i->ignore)
+			continue;
+
+		cond_resched();
+
+		for_each_jset_key(k, entry, &i->j) {
+			if (keys->nr == keys->size) {
+				__journal_keys_sort(keys);
+
+				if (keys->nr > keys->size * 7 / 8) {
+					bch_err(c, "Too many journal keys for slowpath; have %zu compacted, buf size %zu, processed %zu/%zu",
+						keys->nr, keys->size, nr_read, nr_keys);
+					return -BCH_ERR_ENOMEM_journal_keys_sort;
+				}
+			}
+
+			keys->d[keys->nr++] = (struct journal_key) {
+				.btree_id	= entry->btree_id,
+				.level		= entry->level,
+				.k		= k,
+				.journal_seq	= le64_to_cpu(i->j.seq),
+				.journal_offset	= k->_data - i->j._data,
+			};
+
+			nr_read++;
+		}
+	}
+
+	__journal_keys_sort(keys);
+	keys->gap = keys->nr;
+
+	bch_verbose(c, "Journal keys: %zu read, %zu after sorting and compacting", nr_keys, keys->nr);
+	return 0;
+}
diff --git a/libbcachefs/btree_journal_iter.h b/libbcachefs/btree_journal_iter.h
new file mode 100644
index 00000000..5d64e7e2
--- /dev/null
+++ b/libbcachefs/btree_journal_iter.h
@@ -0,0 +1,57 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_BTREE_JOURNAL_ITER_H
+#define _BCACHEFS_BTREE_JOURNAL_ITER_H
+
+struct journal_iter {
+	struct list_head	list;
+	enum btree_id		btree_id;
+	unsigned		level;
+	size_t			idx;
+	struct journal_keys	*keys;
+};
+
+/*
+ * Iterate over keys in the btree, with keys from the journal overlaid on top:
+ */
+
+struct btree_and_journal_iter {
+	struct btree		*b;
+	struct btree_node_iter	node_iter;
+	struct bkey		unpacked;
+
+	struct journal_iter	journal;
+	struct bpos		pos;
+	bool			at_end;
+};
+
+struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *, enum btree_id,
+				unsigned, struct bpos, struct bpos, size_t *);
+struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *, enum btree_id,
+					   unsigned, struct bpos);
+
+int bch2_journal_key_insert_take(struct bch_fs *, enum btree_id,
+				 unsigned, struct bkey_i *);
+int bch2_journal_key_insert(struct bch_fs *, enum btree_id,
+			    unsigned, struct bkey_i *);
+int bch2_journal_key_delete(struct bch_fs *, enum btree_id,
+			    unsigned, struct bpos);
+void bch2_journal_key_overwritten(struct bch_fs *, enum btree_id,
+				  unsigned, struct bpos);
+
+void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *);
+struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *);
+
+void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *);
+void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *,
+				struct bch_fs *, struct btree *,
+				struct btree_node_iter, struct bpos);
+void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *,
+						struct bch_fs *,
+						struct btree *);
+
+void bch2_journal_keys_free(struct journal_keys *);
+void bch2_journal_entries_free(struct bch_fs *);
+
+int bch2_journal_keys_sort(struct bch_fs *);
+
+#endif /* _BCACHEFS_BTREE_JOURNAL_ITER_H */
diff --git a/libbcachefs/btree_update_leaf.c b/libbcachefs/btree_trans_commit.c
index 369e37a4..78a09aa0 100644
--- a/libbcachefs/btree_update_leaf.c
+++ b/libbcachefs/btree_trans_commit.c
@@ -1,45 +1,22 @@
 // SPDX-License-Identifier: GPL-2.0
 
 #include "bcachefs.h"
-#include "btree_update.h"
-#include "btree_update_interior.h"
 #include "btree_gc.h"
 #include "btree_io.h"
 #include "btree_iter.h"
+#include "btree_journal_iter.h"
 #include "btree_key_cache.h"
-#include "btree_locking.h"
+#include "btree_update_interior.h"
 #include "btree_write_buffer.h"
 #include "buckets.h"
-#include "debug.h"
 #include "errcode.h"
 #include "error.h"
-#include "extent_update.h"
 #include "journal.h"
 #include "journal_reclaim.h"
-#include "keylist.h"
-#include "recovery.h"
-#include "subvolume.h"
 #include "replicas.h"
-#include "trace.h"
+#include "subvolume.h"
 
 #include <linux/prefetch.h>
-#include <linux/sort.h>
-
-/*
- * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
- * different snapshot:
- */
-static struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
-{
-	struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
-
-	if (k.k && bpos_eq(path->pos, k.k->p))
-		return k;
-
-	bkey_init(u);
-	u->p = path->pos;
-	return (struct bkey_s_c) { u, NULL };
-}
 
 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
 {
@@ -64,20 +41,6 @@ static void verify_update_old_key(struct btree_trans *trans, struct btree_insert
 #endif
 }
 
-static int __must_check
-bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
-			  struct bkey_i *, enum btree_update_flags,
-			  unsigned long ip);
-
-static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
-					 const struct btree_insert_entry *r)
-{
-	return   cmp_int(l->btree_id,	r->btree_id) ?:
-		 cmp_int(l->cached,	r->cached) ?:
-		 -cmp_int(l->level,	r->level) ?:
-		 bpos_cmp(l->k->k.p,	r->k->k.p);
-}
-
 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
 {
 	return i->path->l + i->level;
@@ -1191,917 +1154,3 @@ err:
 
 	goto retry;
 }
-
-static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
-					  enum btree_id id,
-					  struct bpos pos)
-{
-	struct bch_fs *c = trans->c;
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	int ret;
-
-	bch2_trans_iter_init(trans, &iter, id, pos,
-			     BTREE_ITER_NOT_EXTENTS|
-			     BTREE_ITER_ALL_SNAPSHOTS);
-	while (1) {
-		k = bch2_btree_iter_prev(&iter);
-		ret = bkey_err(k);
-		if (ret)
-			break;
-
-		if (!k.k)
-			break;
-
-		if (!bkey_eq(pos, k.k->p))
-			break;
-
-		if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
-			ret = 1;
-			break;
-		}
-	}
-	bch2_trans_iter_exit(trans, &iter);
-
-	return ret;
-}
-
-static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
-					  enum btree_id id,
-					  struct bpos pos)
-{
-	if (!btree_type_has_snapshots(id) ||
-	    bch2_snapshot_is_leaf(trans->c, pos.snapshot))
-		return 0;
-
-	return __check_pos_snapshot_overwritten(trans, id, pos);
-}
-
-static noinline int extent_front_merge(struct btree_trans *trans,
-				       struct btree_iter *iter,
-				       struct bkey_s_c k,
-				       struct bkey_i **insert,
-				       enum btree_update_flags flags)
-{
-	struct bch_fs *c = trans->c;
-	struct bkey_i *update;
-	int ret;
-
-	update = bch2_bkey_make_mut_noupdate(trans, k);
-	ret = PTR_ERR_OR_ZERO(update);
-	if (ret)
-		return ret;
-
-	if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
-		return 0;
-
-	ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
-		check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
-	if (ret < 0)
-		return ret;
-	if (ret)
-		return 0;
-
-	ret = bch2_btree_delete_at(trans, iter, flags);
-	if (ret)
-		return ret;
-
-	*insert = update;
-	return 0;
-}
-
-static noinline int extent_back_merge(struct btree_trans *trans,
-				      struct btree_iter *iter,
-				      struct bkey_i *insert,
-				      struct bkey_s_c k)
-{
-	struct bch_fs *c = trans->c;
-	int ret;
-
-	ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
-		check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
-	if (ret < 0)
-		return ret;
-	if (ret)
-		return 0;
-
-	bch2_bkey_merge(c, bkey_i_to_s(insert), k);
-	return 0;
-}
-
-/*
- * When deleting, check if we need to emit a whiteout (because we're overwriting
- * something in an ancestor snapshot)
- */
-static int need_whiteout_for_snapshot(struct btree_trans *trans,
-				      enum btree_id btree_id, struct bpos pos)
-{
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	u32 snapshot = pos.snapshot;
-	int ret;
-
-	if (!bch2_snapshot_parent(trans->c, pos.snapshot))
-		return 0;
-
-	pos.snapshot++;
-
-	for_each_btree_key_norestart(trans, iter, btree_id, pos,
-			   BTREE_ITER_ALL_SNAPSHOTS|
-			   BTREE_ITER_NOPRESERVE, k, ret) {
-		if (!bkey_eq(k.k->p, pos))
-			break;
-
-		if (bch2_snapshot_is_ancestor(trans->c, snapshot,
-					      k.k->p.snapshot)) {
-			ret = !bkey_whiteout(k.k);
-			break;
-		}
-	}
-	bch2_trans_iter_exit(trans, &iter);
-
-	return ret;
-}
-
-int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
-				   enum btree_id id,
-				   struct bpos old_pos,
-				   struct bpos new_pos)
-{
-	struct bch_fs *c = trans->c;
-	struct btree_iter old_iter, new_iter = { NULL };
-	struct bkey_s_c old_k, new_k;
-	snapshot_id_list s;
-	struct bkey_i *update;
-	int ret;
-
-	if (!bch2_snapshot_has_children(c, old_pos.snapshot))
-		return 0;
-
-	darray_init(&s);
-
-	bch2_trans_iter_init(trans, &old_iter, id, old_pos,
-			     BTREE_ITER_NOT_EXTENTS|
-			     BTREE_ITER_ALL_SNAPSHOTS);
-	while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
-	       !(ret = bkey_err(old_k)) &&
-	       bkey_eq(old_pos, old_k.k->p)) {
-		struct bpos whiteout_pos =
-			SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
-
-		if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
-		    snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
-			continue;
-
-		new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
-					   BTREE_ITER_NOT_EXTENTS|
-					   BTREE_ITER_INTENT);
-		ret = bkey_err(new_k);
-		if (ret)
-			break;
-
-		if (new_k.k->type == KEY_TYPE_deleted) {
-			update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
-			ret = PTR_ERR_OR_ZERO(update);
-			if (ret)
-				break;
-
-			bkey_init(&update->k);
-			update->k.p		= whiteout_pos;
-			update->k.type		= KEY_TYPE_whiteout;
-
-			ret = bch2_trans_update(trans, &new_iter, update,
-						BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
-		}
-		bch2_trans_iter_exit(trans, &new_iter);
-
-		ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
-		if (ret)
-			break;
-	}
-	bch2_trans_iter_exit(trans, &new_iter);
-	bch2_trans_iter_exit(trans, &old_iter);
-	darray_exit(&s);
-
-	return ret;
-}
-
-int bch2_trans_update_extent_overwrite(struct btree_trans *trans,
-				       struct btree_iter *iter,
-				       enum btree_update_flags flags,
-				       struct bkey_s_c old,
-				       struct bkey_s_c new)
-{
-	enum btree_id btree_id = iter->btree_id;
-	struct bkey_i *update;
-	struct bpos new_start = bkey_start_pos(new.k);
-	bool front_split = bkey_lt(bkey_start_pos(old.k), new_start);
-	bool back_split  = bkey_gt(old.k->p, new.k->p);
-	int ret = 0, compressed_sectors;
-
-	/*
-	 * If we're going to be splitting a compressed extent, note it
-	 * so that __bch2_trans_commit() can increase our disk
-	 * reservation:
-	 */
-	if (((front_split && back_split) ||
-	     ((front_split || back_split) && old.k->p.snapshot != new.k->p.snapshot)) &&
-	    (compressed_sectors = bch2_bkey_sectors_compressed(old)))
-		trans->extra_journal_res += compressed_sectors;
-
-	if (front_split) {
-		update = bch2_bkey_make_mut_noupdate(trans, old);
-		if ((ret = PTR_ERR_OR_ZERO(update)))
-			return ret;
-
-		bch2_cut_back(new_start, update);
-
-		ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
-					old.k->p, update->k.p) ?:
-			bch2_btree_insert_nonextent(trans, btree_id, update,
-					BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
-		if (ret)
-			return ret;
-	}
-
-	/* If we're overwriting in a different snapshot - middle split: */
-	if (old.k->p.snapshot != new.k->p.snapshot &&
-	    (front_split || back_split)) {
-		update = bch2_bkey_make_mut_noupdate(trans, old);
-		if ((ret = PTR_ERR_OR_ZERO(update)))
-			return ret;
-
-		bch2_cut_front(new_start, update);
-		bch2_cut_back(new.k->p, update);
-
-		ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
-					old.k->p, update->k.p) ?:
-			bch2_btree_insert_nonextent(trans, btree_id, update,
-					  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
-		if (ret)
-			return ret;
-	}
-
-	if (bkey_le(old.k->p, new.k->p)) {
-		update = bch2_trans_kmalloc(trans, sizeof(*update));
-		if ((ret = PTR_ERR_OR_ZERO(update)))
-			return ret;
-
-		bkey_init(&update->k);
-		update->k.p = old.k->p;
-		update->k.p.snapshot = new.k->p.snapshot;
-
-		if (new.k->p.snapshot != old.k->p.snapshot) {
-			update->k.type = KEY_TYPE_whiteout;
-		} else if (btree_type_has_snapshots(btree_id)) {
-			ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
-			if (ret < 0)
-				return ret;
-			if (ret)
-				update->k.type = KEY_TYPE_whiteout;
-		}
-
-		ret = bch2_btree_insert_nonextent(trans, btree_id, update,
-					  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
-		if (ret)
-			return ret;
-	}
-
-	if (back_split) {
-		update = bch2_bkey_make_mut_noupdate(trans, old);
-		if ((ret = PTR_ERR_OR_ZERO(update)))
-			return ret;
-
-		bch2_cut_front(new.k->p, update);
-
-		ret = bch2_trans_update_by_path(trans, iter->path, update,
-					  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
-					  flags, _RET_IP_);
-		if (ret)
-			return ret;
-	}
-
-	return 0;
-}
-
-static int bch2_trans_update_extent(struct btree_trans *trans,
-				    struct btree_iter *orig_iter,
-				    struct bkey_i *insert,
-				    enum btree_update_flags flags)
-{
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	enum btree_id btree_id = orig_iter->btree_id;
-	int ret = 0;
-
-	bch2_trans_iter_init(trans, &iter, btree_id, bkey_start_pos(&insert->k),
-			     BTREE_ITER_INTENT|
-			     BTREE_ITER_WITH_UPDATES|
-			     BTREE_ITER_NOT_EXTENTS);
-	k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
-	if ((ret = bkey_err(k)))
-		goto err;
-	if (!k.k)
-		goto out;
-
-	if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
-		if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
-			ret = extent_front_merge(trans, &iter, k, &insert, flags);
-			if (ret)
-				goto err;
-		}
-
-		goto next;
-	}
-
-	while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
-		bool done = bkey_lt(insert->k.p, k.k->p);
-
-		ret = bch2_trans_update_extent_overwrite(trans, &iter, flags, k, bkey_i_to_s_c(insert));
-		if (ret)
-			goto err;
-
-		if (done)
-			goto out;
-next:
-		bch2_btree_iter_advance(&iter);
-		k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
-		if ((ret = bkey_err(k)))
-			goto err;
-		if (!k.k)
-			goto out;
-	}
-
-	if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
-		ret = extent_back_merge(trans, &iter, insert, k);
-		if (ret)
-			goto err;
-	}
-out:
-	if (!bkey_deleted(&insert->k))
-		ret = bch2_btree_insert_nonextent(trans, btree_id, insert, flags);
-err:
-	bch2_trans_iter_exit(trans, &iter);
-
-	return ret;
-}
-
-static noinline int flush_new_cached_update(struct btree_trans *trans,
-					    struct btree_path *path,
-					    struct btree_insert_entry *i,
-					    enum btree_update_flags flags,
-					    unsigned long ip)
-{
-	struct btree_path *btree_path;
-	struct bkey k;
-	int ret;
-
-	btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
-				   BTREE_ITER_INTENT, _THIS_IP_);
-	ret = bch2_btree_path_traverse(trans, btree_path, 0);
-	if (ret)
-		goto out;
-
-	/*
-	 * The old key in the insert entry might actually refer to an existing
-	 * key in the btree that has been deleted from cache and not yet
-	 * flushed. Check for this and skip the flush so we don't run triggers
-	 * against a stale key.
-	 */
-	bch2_btree_path_peek_slot_exact(btree_path, &k);
-	if (!bkey_deleted(&k))
-		goto out;
-
-	i->key_cache_already_flushed = true;
-	i->flags |= BTREE_TRIGGER_NORUN;
-
-	btree_path_set_should_be_locked(btree_path);
-	ret = bch2_trans_update_by_path(trans, btree_path, i->k, flags, ip);
-out:
-	bch2_path_put(trans, btree_path, true);
-	return ret;
-}
-
-static int __must_check
-bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
-			  struct bkey_i *k, enum btree_update_flags flags,
-			  unsigned long ip)
-{
-	struct bch_fs *c = trans->c;
-	struct btree_insert_entry *i, n;
-	u64 seq = 0;
-	int cmp;
-
-	EBUG_ON(!path->should_be_locked);
-	EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
-	EBUG_ON(!bpos_eq(k->k.p, path->pos));
-
-	/*
-	 * The transaction journal res hasn't been allocated at this point.
-	 * That occurs at commit time. Reuse the seq field to pass in the seq
-	 * of a prejournaled key.
-	 */
-	if (flags & BTREE_UPDATE_PREJOURNAL)
-		seq = trans->journal_res.seq;
-
-	n = (struct btree_insert_entry) {
-		.flags		= flags,
-		.bkey_type	= __btree_node_type(path->level, path->btree_id),
-		.btree_id	= path->btree_id,
-		.level		= path->level,
-		.cached		= path->cached,
-		.path		= path,
-		.k		= k,
-		.seq		= seq,
-		.ip_allocated	= ip,
-	};
-
-#ifdef CONFIG_BCACHEFS_DEBUG
-	trans_for_each_update(trans, i)
-		BUG_ON(i != trans->updates &&
-		       btree_insert_entry_cmp(i - 1, i) >= 0);
-#endif
-
-	/*
-	 * Pending updates are kept sorted: first, find position of new update,
-	 * then delete/trim any updates the new update overwrites:
-	 */
-	trans_for_each_update(trans, i) {
-		cmp = btree_insert_entry_cmp(&n, i);
-		if (cmp <= 0)
-			break;
-	}
-
-	if (!cmp && i < trans->updates + trans->nr_updates) {
-		EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
-
-		bch2_path_put(trans, i->path, true);
-		i->flags	= n.flags;
-		i->cached	= n.cached;
-		i->k		= n.k;
-		i->path		= n.path;
-		i->seq		= n.seq;
-		i->ip_allocated	= n.ip_allocated;
-	} else {
-		array_insert_item(trans->updates, trans->nr_updates,
-				  i - trans->updates, n);
-
-		i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
-		i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
-
-		if (unlikely(trans->journal_replay_not_finished)) {
-			struct bkey_i *j_k =
-				bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
-
-			if (j_k) {
-				i->old_k = j_k->k;
-				i->old_v = &j_k->v;
-			}
-		}
-	}
-
-	__btree_path_get(i->path, true);
-
-	/*
-	 * If a key is present in the key cache, it must also exist in the
-	 * btree - this is necessary for cache coherency. When iterating over
-	 * a btree that's cached in the key cache, the btree iter code checks
-	 * the key cache - but the key has to exist in the btree for that to
-	 * work:
-	 */
-	if (path->cached && bkey_deleted(&i->old_k))
-		return flush_new_cached_update(trans, path, i, flags, ip);
-
-	return 0;
-}
-
-int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
-				   struct bkey_i *k, enum btree_update_flags flags)
-{
-	struct btree_path *path = iter->update_path ?: iter->path;
-	struct bkey_cached *ck;
-	int ret;
-
-	if (iter->flags & BTREE_ITER_IS_EXTENTS)
-		return bch2_trans_update_extent(trans, iter, k, flags);
-
-	if (bkey_deleted(&k->k) &&
-	    !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
-	    (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
-		ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
-		if (unlikely(ret < 0))
-			return ret;
-
-		if (ret)
-			k->k.type = KEY_TYPE_whiteout;
-	}
-
-	/*
-	 * Ensure that updates to cached btrees go to the key cache:
-	 */
-	if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
-	    !path->cached &&
-	    !path->level &&
-	    btree_id_cached(trans->c, path->btree_id)) {
-		if (!iter->key_cache_path ||
-		    !iter->key_cache_path->should_be_locked ||
-		    !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
-			if (!iter->key_cache_path)
-				iter->key_cache_path =
-					bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
-						      BTREE_ITER_INTENT|
-						      BTREE_ITER_CACHED, _THIS_IP_);
-
-			iter->key_cache_path =
-				bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
-							iter->flags & BTREE_ITER_INTENT,
-							_THIS_IP_);
-
-			ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
-						       BTREE_ITER_CACHED);
-			if (unlikely(ret))
-				return ret;
-
-			ck = (void *) iter->key_cache_path->l[0].b;
-
-			if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
-				trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
-				return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
-			}
-
-			btree_path_set_should_be_locked(iter->key_cache_path);
-		}
-
-		path = iter->key_cache_path;
-	}
-
-	return bch2_trans_update_by_path(trans, path, k, flags, _RET_IP_);
-}
-
-/*
- * Add a transaction update for a key that has already been journaled.
- */
-int __must_check bch2_trans_update_seq(struct btree_trans *trans, u64 seq,
-				       struct btree_iter *iter, struct bkey_i *k,
-				       enum btree_update_flags flags)
-{
-	trans->journal_res.seq = seq;
-	return bch2_trans_update(trans, iter, k, flags|BTREE_UPDATE_NOJOURNAL|
-						 BTREE_UPDATE_PREJOURNAL);
-}
-
-int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
-					    enum btree_id btree,
-					    struct bkey_i *k)
-{
-	struct btree_write_buffered_key *i;
-	int ret;
-
-	EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
-	EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
-
-	trans_for_each_wb_update(trans, i) {
-		if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
-			bkey_copy(&i->k, k);
-			return 0;
-		}
-	}
-
-	if (!trans->wb_updates ||
-	    trans->nr_wb_updates == trans->wb_updates_size) {
-		struct btree_write_buffered_key *u;
-
-		if (trans->nr_wb_updates == trans->wb_updates_size) {
-			struct btree_transaction_stats *s = btree_trans_stats(trans);
-
-			BUG_ON(trans->wb_updates_size > U8_MAX / 2);
-			trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
-			if (s)
-				s->wb_updates_size = trans->wb_updates_size;
-		}
-
-		u = bch2_trans_kmalloc_nomemzero(trans,
-					trans->wb_updates_size *
-					sizeof(struct btree_write_buffered_key));
-		ret = PTR_ERR_OR_ZERO(u);
-		if (ret)
-			return ret;
-
-		if (trans->nr_wb_updates)
-			memcpy(u, trans->wb_updates, trans->nr_wb_updates *
-			       sizeof(struct btree_write_buffered_key));
-		trans->wb_updates = u;
-	}
-
-	trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
-		.btree	= btree,
-	};
-
-	bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
-	trans->nr_wb_updates++;
-
-	return 0;
-}
-
-int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
-			     enum btree_id btree, struct bpos end)
-{
-	struct bkey_s_c k;
-	int ret = 0;
-
-	bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
-	k = bch2_btree_iter_prev(iter);
-	ret = bkey_err(k);
-	if (ret)
-		goto err;
-
-	bch2_btree_iter_advance(iter);
-	k = bch2_btree_iter_peek_slot(iter);
-	ret = bkey_err(k);
-	if (ret)
-		goto err;
-
-	BUG_ON(k.k->type != KEY_TYPE_deleted);
-
-	if (bkey_gt(k.k->p, end)) {
-		ret = -BCH_ERR_ENOSPC_btree_slot;
-		goto err;
-	}
-
-	return 0;
-err:
-	bch2_trans_iter_exit(trans, iter);
-	return ret;
-}
-
-void bch2_trans_commit_hook(struct btree_trans *trans,
-			    struct btree_trans_commit_hook *h)
-{
-	h->next = trans->hooks;
-	trans->hooks = h;
-}
-
-int bch2_btree_insert_nonextent(struct btree_trans *trans,
-				enum btree_id btree, struct bkey_i *k,
-				enum btree_update_flags flags)
-{
-	struct btree_iter iter;
-	int ret;
-
-	bch2_trans_iter_init(trans, &iter, btree, k->k.p,
-			     BTREE_ITER_NOT_EXTENTS|
-			     BTREE_ITER_INTENT);
-	ret   = bch2_btree_iter_traverse(&iter) ?:
-		bch2_trans_update(trans, &iter, k, flags);
-	bch2_trans_iter_exit(trans, &iter);
-	return ret;
-}
-
-int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
-			struct bkey_i *k, enum btree_update_flags flags)
-{
-	struct btree_iter iter;
-	int ret;
-
-	bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
-			     BTREE_ITER_CACHED|
-			     BTREE_ITER_INTENT);
-	ret   = bch2_btree_iter_traverse(&iter) ?:
-		bch2_trans_update(trans, &iter, k, flags);
-	bch2_trans_iter_exit(trans, &iter);
-	return ret;
-}
-
-/**
- * bch2_btree_insert - insert keys into the extent btree
- * @c:			pointer to struct bch_fs
- * @id:			btree to insert into
- * @insert_keys:	list of keys to insert
- * @hook:		insert callback
- */
-int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
-		      struct bkey_i *k,
-		      struct disk_reservation *disk_res,
-		      u64 *journal_seq, int flags)
-{
-	return bch2_trans_do(c, disk_res, journal_seq, flags,
-			     __bch2_btree_insert(&trans, id, k, 0));
-}
-
-int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
-				unsigned len, unsigned update_flags)
-{
-	struct bkey_i *k;
-
-	k = bch2_trans_kmalloc(trans, sizeof(*k));
-	if (IS_ERR(k))
-		return PTR_ERR(k);
-
-	bkey_init(&k->k);
-	k->k.p = iter->pos;
-	bch2_key_resize(&k->k, len);
-	return bch2_trans_update(trans, iter, k, update_flags);
-}
-
-int bch2_btree_delete_at(struct btree_trans *trans,
-			 struct btree_iter *iter, unsigned update_flags)
-{
-	return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
-}
-
-int bch2_btree_delete_at_buffered(struct btree_trans *trans,
-				  enum btree_id btree, struct bpos pos)
-{
-	struct bkey_i *k;
-
-	k = bch2_trans_kmalloc(trans, sizeof(*k));
-	if (IS_ERR(k))
-		return PTR_ERR(k);
-
-	bkey_init(&k->k);
-	k->k.p = pos;
-	return bch2_trans_update_buffered(trans, btree, k);
-}
-
-int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
-				  struct bpos start, struct bpos end,
-				  unsigned update_flags,
-				  u64 *journal_seq)
-{
-	u32 restart_count = trans->restart_count;
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	int ret = 0;
-
-	bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
-	while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
-		struct disk_reservation disk_res =
-			bch2_disk_reservation_init(trans->c, 0);
-		struct bkey_i delete;
-
-		ret = bkey_err(k);
-		if (ret)
-			goto err;
-
-		bkey_init(&delete.k);
-
-		/*
-		 * This could probably be more efficient for extents:
-		 */
-
-		/*
-		 * For extents, iter.pos won't necessarily be the same as
-		 * bkey_start_pos(k.k) (for non extents they always will be the
-		 * same). It's important that we delete starting from iter.pos
-		 * because the range we want to delete could start in the middle
-		 * of k.
-		 *
-		 * (bch2_btree_iter_peek() does guarantee that iter.pos >=
-		 * bkey_start_pos(k.k)).
-		 */
-		delete.k.p = iter.pos;
-
-		if (iter.flags & BTREE_ITER_IS_EXTENTS)
-			bch2_key_resize(&delete.k,
-					bpos_min(end, k.k->p).offset -
-					iter.pos.offset);
-
-		ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
-			bch2_trans_commit(trans, &disk_res, journal_seq,
-					  BTREE_INSERT_NOFAIL);
-		bch2_disk_reservation_put(trans->c, &disk_res);
-err:
-		/*
-		 * the bch2_trans_begin() call is in a weird place because we
-		 * need to call it after every transaction commit, to avoid path
-		 * overflow, but don't want to call it if the delete operation
-		 * is a no-op and we have no work to do:
-		 */
-		bch2_trans_begin(trans);
-
-		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
-			ret = 0;
-		if (ret)
-			break;
-	}
-	bch2_trans_iter_exit(trans, &iter);
-
-	if (!ret && trans_was_restarted(trans, restart_count))
-		ret = -BCH_ERR_transaction_restart_nested;
-	return ret;
-}
-
-/*
- * bch_btree_delete_range - delete everything within a given range
- *
- * Range is a half open interval - [start, end)
- */
-int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
-			    struct bpos start, struct bpos end,
-			    unsigned update_flags,
-			    u64 *journal_seq)
-{
-	int ret = bch2_trans_run(c,
-			bch2_btree_delete_range_trans(&trans, id, start, end,
-						      update_flags, journal_seq));
-	if (ret == -BCH_ERR_transaction_restart_nested)
-		ret = 0;
-	return ret;
-}
-
-int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
-		       struct bpos pos, bool set)
-{
-	struct bkey_i *k;
-	int ret = 0;
-
-	k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
-	ret = PTR_ERR_OR_ZERO(k);
-	if (unlikely(ret))
-		return ret;
-
-	bkey_init(&k->k);
-	k->k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
-	k->k.p = pos;
-
-	return bch2_trans_update_buffered(trans, btree, k);
-}
-
-static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
-{
-	struct printbuf buf = PRINTBUF;
-	struct jset_entry_log *l;
-	unsigned u64s;
-	int ret;
-
-	prt_vprintf(&buf, fmt, args);
-	ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
-	if (ret)
-		goto err;
-
-	u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
-
-	ret = darray_make_room(entries, jset_u64s(u64s));
-	if (ret)
-		goto err;
-
-	l = (void *) &darray_top(*entries);
-	l->entry.u64s		= cpu_to_le16(u64s);
-	l->entry.btree_id	= 0;
-	l->entry.level		= 1;
-	l->entry.type		= BCH_JSET_ENTRY_log;
-	l->entry.pad[0]		= 0;
-	l->entry.pad[1]		= 0;
-	l->entry.pad[2]		= 0;
-	memcpy(l->d, buf.buf, buf.pos);
-	while (buf.pos & 7)
-		l->d[buf.pos++] = '\0';
-
-	entries->nr += jset_u64s(u64s);
-err:
-	printbuf_exit(&buf);
-	return ret;
-}
-
-static int
-__bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
-		  va_list args)
-{
-	int ret;
-
-	if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
-		ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
-	} else {
-		ret = bch2_trans_do(c, NULL, NULL,
-			BTREE_INSERT_LAZY_RW|commit_flags,
-			__bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
-	}
-
-	return ret;
-}
-
-int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
-{
-	va_list args;
-	int ret;
-
-	va_start(args, fmt);
-	ret = __bch2_fs_log_msg(c, 0, fmt, args);
-	va_end(args);
-	return ret;
-}
-
-/*
- * Use for logging messages during recovery to enable reserved space and avoid
- * blocking.
- */
-int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
-{
-	va_list args;
-	int ret;
-
-	va_start(args, fmt);
-	ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);
-	va_end(args);
-	return ret;
-}
diff --git a/libbcachefs/btree_update.c b/libbcachefs/btree_update.c
new file mode 100644
index 00000000..612fba60
--- /dev/null
+++ b/libbcachefs/btree_update.c
@@ -0,0 +1,943 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "btree_update.h"
+#include "btree_iter.h"
+#include "btree_journal_iter.h"
+#include "btree_locking.h"
+#include "buckets.h"
+#include "debug.h"
+#include "errcode.h"
+#include "error.h"
+#include "extents.h"
+#include "keylist.h"
+#include "subvolume.h"
+#include "trace.h"
+
+static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
+					 const struct btree_insert_entry *r)
+{
+	return   cmp_int(l->btree_id,	r->btree_id) ?:
+		 cmp_int(l->cached,	r->cached) ?:
+		 -cmp_int(l->level,	r->level) ?:
+		 bpos_cmp(l->k->k.p,	r->k->k.p);
+}
+
+static int __must_check
+bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
+			  struct bkey_i *, enum btree_update_flags,
+			  unsigned long ip);
+
+static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
+					  enum btree_id id,
+					  struct bpos pos)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+
+	bch2_trans_iter_init(trans, &iter, id, pos,
+			     BTREE_ITER_NOT_EXTENTS|
+			     BTREE_ITER_ALL_SNAPSHOTS);
+	while (1) {
+		k = bch2_btree_iter_prev(&iter);
+		ret = bkey_err(k);
+		if (ret)
+			break;
+
+		if (!k.k)
+			break;
+
+		if (!bkey_eq(pos, k.k->p))
+			break;
+
+		if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
+			ret = 1;
+			break;
+		}
+	}
+	bch2_trans_iter_exit(trans, &iter);
+
+	return ret;
+}
+
+static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
+					  enum btree_id id,
+					  struct bpos pos)
+{
+	if (!btree_type_has_snapshots(id) ||
+	    bch2_snapshot_is_leaf(trans->c, pos.snapshot))
+		return 0;
+
+	return __check_pos_snapshot_overwritten(trans, id, pos);
+}
+
+static noinline int extent_front_merge(struct btree_trans *trans,
+				       struct btree_iter *iter,
+				       struct bkey_s_c k,
+				       struct bkey_i **insert,
+				       enum btree_update_flags flags)
+{
+	struct bch_fs *c = trans->c;
+	struct bkey_i *update;
+	int ret;
+
+	update = bch2_bkey_make_mut_noupdate(trans, k);
+	ret = PTR_ERR_OR_ZERO(update);
+	if (ret)
+		return ret;
+
+	if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
+		return 0;
+
+	ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
+		check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
+	if (ret < 0)
+		return ret;
+	if (ret)
+		return 0;
+
+	ret = bch2_btree_delete_at(trans, iter, flags);
+	if (ret)
+		return ret;
+
+	*insert = update;
+	return 0;
+}
+
+static noinline int extent_back_merge(struct btree_trans *trans,
+				      struct btree_iter *iter,
+				      struct bkey_i *insert,
+				      struct bkey_s_c k)
+{
+	struct bch_fs *c = trans->c;
+	int ret;
+
+	ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
+		check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
+	if (ret < 0)
+		return ret;
+	if (ret)
+		return 0;
+
+	bch2_bkey_merge(c, bkey_i_to_s(insert), k);
+	return 0;
+}
+
+/*
+ * When deleting, check if we need to emit a whiteout (because we're overwriting
+ * something in an ancestor snapshot)
+ */
+static int need_whiteout_for_snapshot(struct btree_trans *trans,
+				      enum btree_id btree_id, struct bpos pos)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	u32 snapshot = pos.snapshot;
+	int ret;
+
+	if (!bch2_snapshot_parent(trans->c, pos.snapshot))
+		return 0;
+
+	pos.snapshot++;
+
+	for_each_btree_key_norestart(trans, iter, btree_id, pos,
+			   BTREE_ITER_ALL_SNAPSHOTS|
+			   BTREE_ITER_NOPRESERVE, k, ret) {
+		if (!bkey_eq(k.k->p, pos))
+			break;
+
+		if (bch2_snapshot_is_ancestor(trans->c, snapshot,
+					      k.k->p.snapshot)) {
+			ret = !bkey_whiteout(k.k);
+			break;
+		}
+	}
+	bch2_trans_iter_exit(trans, &iter);
+
+	return ret;
+}
+
+int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
+				   enum btree_id id,
+				   struct bpos old_pos,
+				   struct bpos new_pos)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter old_iter, new_iter = { NULL };
+	struct bkey_s_c old_k, new_k;
+	snapshot_id_list s;
+	struct bkey_i *update;
+	int ret;
+
+	if (!bch2_snapshot_has_children(c, old_pos.snapshot))
+		return 0;
+
+	darray_init(&s);
+
+	bch2_trans_iter_init(trans, &old_iter, id, old_pos,
+			     BTREE_ITER_NOT_EXTENTS|
+			     BTREE_ITER_ALL_SNAPSHOTS);
+	while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
+	       !(ret = bkey_err(old_k)) &&
+	       bkey_eq(old_pos, old_k.k->p)) {
+		struct bpos whiteout_pos =
+			SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
+
+		if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
+		    snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
+			continue;
+
+		new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
+					   BTREE_ITER_NOT_EXTENTS|
+					   BTREE_ITER_INTENT);
+		ret = bkey_err(new_k);
+		if (ret)
+			break;
+
+		if (new_k.k->type == KEY_TYPE_deleted) {
+			update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
+			ret = PTR_ERR_OR_ZERO(update);
+			if (ret)
+				break;
+
+			bkey_init(&update->k);
+			update->k.p		= whiteout_pos;
+			update->k.type		= KEY_TYPE_whiteout;
+
+			ret = bch2_trans_update(trans, &new_iter, update,
+						BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+		}
+		bch2_trans_iter_exit(trans, &new_iter);
+
+		ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
+		if (ret)
+			break;
+	}
+	bch2_trans_iter_exit(trans, &new_iter);
+	bch2_trans_iter_exit(trans, &old_iter);
+	darray_exit(&s);
+
+	return ret;
+}
+
+int bch2_trans_update_extent_overwrite(struct btree_trans *trans,
+				       struct btree_iter *iter,
+				       enum btree_update_flags flags,
+				       struct bkey_s_c old,
+				       struct bkey_s_c new)
+{
+	enum btree_id btree_id = iter->btree_id;
+	struct bkey_i *update;
+	struct bpos new_start = bkey_start_pos(new.k);
+	bool front_split = bkey_lt(bkey_start_pos(old.k), new_start);
+	bool back_split  = bkey_gt(old.k->p, new.k->p);
+	int ret = 0, compressed_sectors;
+
+	/*
+	 * If we're going to be splitting a compressed extent, note it
+	 * so that __bch2_trans_commit() can increase our disk
+	 * reservation:
+	 */
+	if (((front_split && back_split) ||
+	     ((front_split || back_split) && old.k->p.snapshot != new.k->p.snapshot)) &&
+	    (compressed_sectors = bch2_bkey_sectors_compressed(old)))
+		trans->extra_journal_res += compressed_sectors;
+
+	if (front_split) {
+		update = bch2_bkey_make_mut_noupdate(trans, old);
+		if ((ret = PTR_ERR_OR_ZERO(update)))
+			return ret;
+
+		bch2_cut_back(new_start, update);
+
+		ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
+					old.k->p, update->k.p) ?:
+			bch2_btree_insert_nonextent(trans, btree_id, update,
+					BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
+		if (ret)
+			return ret;
+	}
+
+	/* If we're overwriting in a different snapshot - middle split: */
+	if (old.k->p.snapshot != new.k->p.snapshot &&
+	    (front_split || back_split)) {
+		update = bch2_bkey_make_mut_noupdate(trans, old);
+		if ((ret = PTR_ERR_OR_ZERO(update)))
+			return ret;
+
+		bch2_cut_front(new_start, update);
+		bch2_cut_back(new.k->p, update);
+
+		ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
+					old.k->p, update->k.p) ?:
+			bch2_btree_insert_nonextent(trans, btree_id, update,
+					  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
+		if (ret)
+			return ret;
+	}
+
+	if (bkey_le(old.k->p, new.k->p)) {
+		update = bch2_trans_kmalloc(trans, sizeof(*update));
+		if ((ret = PTR_ERR_OR_ZERO(update)))
+			return ret;
+
+		bkey_init(&update->k);
+		update->k.p = old.k->p;
+		update->k.p.snapshot = new.k->p.snapshot;
+
+		if (new.k->p.snapshot != old.k->p.snapshot) {
+			update->k.type = KEY_TYPE_whiteout;
+		} else if (btree_type_has_snapshots(btree_id)) {
+			ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
+			if (ret < 0)
+				return ret;
+			if (ret)
+				update->k.type = KEY_TYPE_whiteout;
+		}
+
+		ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+					  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
+		if (ret)
+			return ret;
+	}
+
+	if (back_split) {
+		update = bch2_bkey_make_mut_noupdate(trans, old);
+		if ((ret = PTR_ERR_OR_ZERO(update)))
+			return ret;
+
+		bch2_cut_front(new.k->p, update);
+
+		ret = bch2_trans_update_by_path(trans, iter->path, update,
+					  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+					  flags, _RET_IP_);
+		if (ret)
+			return ret;
+	}
+
+	return 0;
+}
+
+static int bch2_trans_update_extent(struct btree_trans *trans,
+				    struct btree_iter *orig_iter,
+				    struct bkey_i *insert,
+				    enum btree_update_flags flags)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	enum btree_id btree_id = orig_iter->btree_id;
+	int ret = 0;
+
+	bch2_trans_iter_init(trans, &iter, btree_id, bkey_start_pos(&insert->k),
+			     BTREE_ITER_INTENT|
+			     BTREE_ITER_WITH_UPDATES|
+			     BTREE_ITER_NOT_EXTENTS);
+	k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
+	if ((ret = bkey_err(k)))
+		goto err;
+	if (!k.k)
+		goto out;
+
+	if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
+		if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
+			ret = extent_front_merge(trans, &iter, k, &insert, flags);
+			if (ret)
+				goto err;
+		}
+
+		goto next;
+	}
+
+	while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
+		bool done = bkey_lt(insert->k.p, k.k->p);
+
+		ret = bch2_trans_update_extent_overwrite(trans, &iter, flags, k, bkey_i_to_s_c(insert));
+		if (ret)
+			goto err;
+
+		if (done)
+			goto out;
+next:
+		bch2_btree_iter_advance(&iter);
+		k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
+		if ((ret = bkey_err(k)))
+			goto err;
+		if (!k.k)
+			goto out;
+	}
+
+	if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
+		ret = extent_back_merge(trans, &iter, insert, k);
+		if (ret)
+			goto err;
+	}
+out:
+	if (!bkey_deleted(&insert->k))
+		ret = bch2_btree_insert_nonextent(trans, btree_id, insert, flags);
+err:
+	bch2_trans_iter_exit(trans, &iter);
+
+	return ret;
+}
+
+static noinline int flush_new_cached_update(struct btree_trans *trans,
+					    struct btree_path *path,
+					    struct btree_insert_entry *i,
+					    enum btree_update_flags flags,
+					    unsigned long ip)
+{
+	struct btree_path *btree_path;
+	struct bkey k;
+	int ret;
+
+	btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
+				   BTREE_ITER_INTENT, _THIS_IP_);
+	ret = bch2_btree_path_traverse(trans, btree_path, 0);
+	if (ret)
+		goto out;
+
+	/*
+	 * The old key in the insert entry might actually refer to an existing
+	 * key in the btree that has been deleted from cache and not yet
+	 * flushed. Check for this and skip the flush so we don't run triggers
+	 * against a stale key.
+	 */
+	bch2_btree_path_peek_slot_exact(btree_path, &k);
+	if (!bkey_deleted(&k))
+		goto out;
+
+	i->key_cache_already_flushed = true;
+	i->flags |= BTREE_TRIGGER_NORUN;
+
+	btree_path_set_should_be_locked(btree_path);
+	ret = bch2_trans_update_by_path(trans, btree_path, i->k, flags, ip);
+out:
+	bch2_path_put(trans, btree_path, true);
+	return ret;
+}
+
+static int __must_check
+bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
+			  struct bkey_i *k, enum btree_update_flags flags,
+			  unsigned long ip)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_insert_entry *i, n;
+	u64 seq = 0;
+	int cmp;
+
+	EBUG_ON(!path->should_be_locked);
+	EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
+	EBUG_ON(!bpos_eq(k->k.p, path->pos));
+
+	/*
+	 * The transaction journal res hasn't been allocated at this point.
+	 * That occurs at commit time. Reuse the seq field to pass in the seq
+	 * of a prejournaled key.
+	 */
+	if (flags & BTREE_UPDATE_PREJOURNAL)
+		seq = trans->journal_res.seq;
+
+	n = (struct btree_insert_entry) {
+		.flags		= flags,
+		.bkey_type	= __btree_node_type(path->level, path->btree_id),
+		.btree_id	= path->btree_id,
+		.level		= path->level,
+		.cached		= path->cached,
+		.path		= path,
+		.k		= k,
+		.seq		= seq,
+		.ip_allocated	= ip,
+	};
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+	trans_for_each_update(trans, i)
+		BUG_ON(i != trans->updates &&
+		       btree_insert_entry_cmp(i - 1, i) >= 0);
+#endif
+
+	/*
+	 * Pending updates are kept sorted: first, find position of new update,
+	 * then delete/trim any updates the new update overwrites:
+	 */
+	trans_for_each_update(trans, i) {
+		cmp = btree_insert_entry_cmp(&n, i);
+		if (cmp <= 0)
+			break;
+	}
+
+	if (!cmp && i < trans->updates + trans->nr_updates) {
+		EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
+
+		bch2_path_put(trans, i->path, true);
+		i->flags	= n.flags;
+		i->cached	= n.cached;
+		i->k		= n.k;
+		i->path		= n.path;
+		i->seq		= n.seq;
+		i->ip_allocated	= n.ip_allocated;
+	} else {
+		array_insert_item(trans->updates, trans->nr_updates,
+				  i - trans->updates, n);
+
+		i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
+		i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
+
+		if (unlikely(trans->journal_replay_not_finished)) {
+			struct bkey_i *j_k =
+				bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
+
+			if (j_k) {
+				i->old_k = j_k->k;
+				i->old_v = &j_k->v;
+			}
+		}
+	}
+
+	__btree_path_get(i->path, true);
+
+	/*
+	 * If a key is present in the key cache, it must also exist in the
+	 * btree - this is necessary for cache coherency. When iterating over
+	 * a btree that's cached in the key cache, the btree iter code checks
+	 * the key cache - but the key has to exist in the btree for that to
+	 * work:
+	 */
+	if (path->cached && bkey_deleted(&i->old_k))
+		return flush_new_cached_update(trans, path, i, flags, ip);
+
+	return 0;
+}
+
+int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
+				   struct bkey_i *k, enum btree_update_flags flags)
+{
+	struct btree_path *path = iter->update_path ?: iter->path;
+	struct bkey_cached *ck;
+	int ret;
+
+	if (iter->flags & BTREE_ITER_IS_EXTENTS)
+		return bch2_trans_update_extent(trans, iter, k, flags);
+
+	if (bkey_deleted(&k->k) &&
+	    !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
+	    (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
+		ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
+		if (unlikely(ret < 0))
+			return ret;
+
+		if (ret)
+			k->k.type = KEY_TYPE_whiteout;
+	}
+
+	/*
+	 * Ensure that updates to cached btrees go to the key cache:
+	 */
+	if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
+	    !path->cached &&
+	    !path->level &&
+	    btree_id_cached(trans->c, path->btree_id)) {
+		if (!iter->key_cache_path ||
+		    !iter->key_cache_path->should_be_locked ||
+		    !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
+			if (!iter->key_cache_path)
+				iter->key_cache_path =
+					bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
+						      BTREE_ITER_INTENT|
+						      BTREE_ITER_CACHED, _THIS_IP_);
+
+			iter->key_cache_path =
+				bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
+							iter->flags & BTREE_ITER_INTENT,
+							_THIS_IP_);
+
+			ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
+						       BTREE_ITER_CACHED);
+			if (unlikely(ret))
+				return ret;
+
+			ck = (void *) iter->key_cache_path->l[0].b;
+
+			if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+				trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
+				return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
+			}
+
+			btree_path_set_should_be_locked(iter->key_cache_path);
+		}
+
+		path = iter->key_cache_path;
+	}
+
+	return bch2_trans_update_by_path(trans, path, k, flags, _RET_IP_);
+}
+
+/*
+ * Add a transaction update for a key that has already been journaled.
+ */
+int __must_check bch2_trans_update_seq(struct btree_trans *trans, u64 seq,
+				       struct btree_iter *iter, struct bkey_i *k,
+				       enum btree_update_flags flags)
+{
+	trans->journal_res.seq = seq;
+	return bch2_trans_update(trans, iter, k, flags|BTREE_UPDATE_NOJOURNAL|
+						 BTREE_UPDATE_PREJOURNAL);
+}
+
+int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
+					    enum btree_id btree,
+					    struct bkey_i *k)
+{
+	struct btree_write_buffered_key *i;
+	int ret;
+
+	EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
+	EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
+
+	trans_for_each_wb_update(trans, i) {
+		if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
+			bkey_copy(&i->k, k);
+			return 0;
+		}
+	}
+
+	if (!trans->wb_updates ||
+	    trans->nr_wb_updates == trans->wb_updates_size) {
+		struct btree_write_buffered_key *u;
+
+		if (trans->nr_wb_updates == trans->wb_updates_size) {
+			struct btree_transaction_stats *s = btree_trans_stats(trans);
+
+			BUG_ON(trans->wb_updates_size > U8_MAX / 2);
+			trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
+			if (s)
+				s->wb_updates_size = trans->wb_updates_size;
+		}
+
+		u = bch2_trans_kmalloc_nomemzero(trans,
+					trans->wb_updates_size *
+					sizeof(struct btree_write_buffered_key));
+		ret = PTR_ERR_OR_ZERO(u);
+		if (ret)
+			return ret;
+
+		if (trans->nr_wb_updates)
+			memcpy(u, trans->wb_updates, trans->nr_wb_updates *
+			       sizeof(struct btree_write_buffered_key));
+		trans->wb_updates = u;
+	}
+
+	trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
+		.btree	= btree,
+	};
+
+	bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
+	trans->nr_wb_updates++;
+
+	return 0;
+}
+
+int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
+			     enum btree_id btree, struct bpos end)
+{
+	struct bkey_s_c k;
+	int ret = 0;
+
+	bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
+	k = bch2_btree_iter_prev(iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	bch2_btree_iter_advance(iter);
+	k = bch2_btree_iter_peek_slot(iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	BUG_ON(k.k->type != KEY_TYPE_deleted);
+
+	if (bkey_gt(k.k->p, end)) {
+		ret = -BCH_ERR_ENOSPC_btree_slot;
+		goto err;
+	}
+
+	return 0;
+err:
+	bch2_trans_iter_exit(trans, iter);
+	return ret;
+}
+
+void bch2_trans_commit_hook(struct btree_trans *trans,
+			    struct btree_trans_commit_hook *h)
+{
+	h->next = trans->hooks;
+	trans->hooks = h;
+}
+
+int bch2_btree_insert_nonextent(struct btree_trans *trans,
+				enum btree_id btree, struct bkey_i *k,
+				enum btree_update_flags flags)
+{
+	struct btree_iter iter;
+	int ret;
+
+	bch2_trans_iter_init(trans, &iter, btree, k->k.p,
+			     BTREE_ITER_NOT_EXTENTS|
+			     BTREE_ITER_INTENT);
+	ret   = bch2_btree_iter_traverse(&iter) ?:
+		bch2_trans_update(trans, &iter, k, flags);
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
+			struct bkey_i *k, enum btree_update_flags flags)
+{
+	struct btree_iter iter;
+	int ret;
+
+	bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
+			     BTREE_ITER_CACHED|
+			     BTREE_ITER_INTENT);
+	ret   = bch2_btree_iter_traverse(&iter) ?:
+		bch2_trans_update(trans, &iter, k, flags);
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+/**
+ * bch2_btree_insert - insert keys into the extent btree
+ * @c:			pointer to struct bch_fs
+ * @id:			btree to insert into
+ * @insert_keys:	list of keys to insert
+ * @hook:		insert callback
+ */
+int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
+		      struct bkey_i *k,
+		      struct disk_reservation *disk_res,
+		      u64 *journal_seq, int flags)
+{
+	return bch2_trans_do(c, disk_res, journal_seq, flags,
+			     __bch2_btree_insert(&trans, id, k, 0));
+}
+
+int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
+				unsigned len, unsigned update_flags)
+{
+	struct bkey_i *k;
+
+	k = bch2_trans_kmalloc(trans, sizeof(*k));
+	if (IS_ERR(k))
+		return PTR_ERR(k);
+
+	bkey_init(&k->k);
+	k->k.p = iter->pos;
+	bch2_key_resize(&k->k, len);
+	return bch2_trans_update(trans, iter, k, update_flags);
+}
+
+int bch2_btree_delete_at(struct btree_trans *trans,
+			 struct btree_iter *iter, unsigned update_flags)
+{
+	return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
+}
+
+int bch2_btree_delete_at_buffered(struct btree_trans *trans,
+				  enum btree_id btree, struct bpos pos)
+{
+	struct bkey_i *k;
+
+	k = bch2_trans_kmalloc(trans, sizeof(*k));
+	if (IS_ERR(k))
+		return PTR_ERR(k);
+
+	bkey_init(&k->k);
+	k->k.p = pos;
+	return bch2_trans_update_buffered(trans, btree, k);
+}
+
+int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
+				  struct bpos start, struct bpos end,
+				  unsigned update_flags,
+				  u64 *journal_seq)
+{
+	u32 restart_count = trans->restart_count;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
+	while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
+		struct disk_reservation disk_res =
+			bch2_disk_reservation_init(trans->c, 0);
+		struct bkey_i delete;
+
+		ret = bkey_err(k);
+		if (ret)
+			goto err;
+
+		bkey_init(&delete.k);
+
+		/*
+		 * This could probably be more efficient for extents:
+		 */
+
+		/*
+		 * For extents, iter.pos won't necessarily be the same as
+		 * bkey_start_pos(k.k) (for non extents they always will be the
+		 * same). It's important that we delete starting from iter.pos
+		 * because the range we want to delete could start in the middle
+		 * of k.
+		 *
+		 * (bch2_btree_iter_peek() does guarantee that iter.pos >=
+		 * bkey_start_pos(k.k)).
+		 */
+		delete.k.p = iter.pos;
+
+		if (iter.flags & BTREE_ITER_IS_EXTENTS)
+			bch2_key_resize(&delete.k,
+					bpos_min(end, k.k->p).offset -
+					iter.pos.offset);
+
+		ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
+			bch2_trans_commit(trans, &disk_res, journal_seq,
+					  BTREE_INSERT_NOFAIL);
+		bch2_disk_reservation_put(trans->c, &disk_res);
+err:
+		/*
+		 * the bch2_trans_begin() call is in a weird place because we
+		 * need to call it after every transaction commit, to avoid path
+		 * overflow, but don't want to call it if the delete operation
+		 * is a no-op and we have no work to do:
+		 */
+		bch2_trans_begin(trans);
+
+		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+			ret = 0;
+		if (ret)
+			break;
+	}
+	bch2_trans_iter_exit(trans, &iter);
+
+	if (!ret && trans_was_restarted(trans, restart_count))
+		ret = -BCH_ERR_transaction_restart_nested;
+	return ret;
+}
+
+/*
+ * bch_btree_delete_range - delete everything within a given range
+ *
+ * Range is a half open interval - [start, end)
+ */
+int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
+			    struct bpos start, struct bpos end,
+			    unsigned update_flags,
+			    u64 *journal_seq)
+{
+	int ret = bch2_trans_run(c,
+			bch2_btree_delete_range_trans(&trans, id, start, end,
+						      update_flags, journal_seq));
+	if (ret == -BCH_ERR_transaction_restart_nested)
+		ret = 0;
+	return ret;
+}
+
+int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
+		       struct bpos pos, bool set)
+{
+	struct bkey_i *k;
+	int ret = 0;
+
+	k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
+	ret = PTR_ERR_OR_ZERO(k);
+	if (unlikely(ret))
+		return ret;
+
+	bkey_init(&k->k);
+	k->k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
+	k->k.p = pos;
+
+	return bch2_trans_update_buffered(trans, btree, k);
+}
+
+static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
+{
+	struct printbuf buf = PRINTBUF;
+	struct jset_entry_log *l;
+	unsigned u64s;
+	int ret;
+
+	prt_vprintf(&buf, fmt, args);
+	ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
+	if (ret)
+		goto err;
+
+	u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
+
+	ret = darray_make_room(entries, jset_u64s(u64s));
+	if (ret)
+		goto err;
+
+	l = (void *) &darray_top(*entries);
+	l->entry.u64s		= cpu_to_le16(u64s);
+	l->entry.btree_id	= 0;
+	l->entry.level		= 1;
+	l->entry.type		= BCH_JSET_ENTRY_log;
+	l->entry.pad[0]		= 0;
+	l->entry.pad[1]		= 0;
+	l->entry.pad[2]		= 0;
+	memcpy(l->d, buf.buf, buf.pos);
+	while (buf.pos & 7)
+		l->d[buf.pos++] = '\0';
+
+	entries->nr += jset_u64s(u64s);
+err:
+	printbuf_exit(&buf);
+	return ret;
+}
+
+static int
+__bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
+		  va_list args)
+{
+	int ret;
+
+	if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
+		ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
+	} else {
+		ret = bch2_trans_do(c, NULL, NULL,
+			BTREE_INSERT_LAZY_RW|commit_flags,
+			__bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
+	}
+
+	return ret;
+}
+
+int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
+{
+	va_list args;
+	int ret;
+
+	va_start(args, fmt);
+	ret = __bch2_fs_log_msg(c, 0, fmt, args);
+	va_end(args);
+	return ret;
+}
+
+/*
+ * Use for logging messages during recovery to enable reserved space and avoid
+ * blocking.
+ */
+int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
+{
+	va_list args;
+	int ret;
+
+	va_start(args, fmt);
+	ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);
+	va_end(args);
+	return ret;
+}
diff --git a/libbcachefs/btree_update_interior.c b/libbcachefs/btree_update_interior.c
index f42ef46c..986dd541 100644
--- a/libbcachefs/btree_update_interior.c
+++ b/libbcachefs/btree_update_interior.c
@@ -5,6 +5,7 @@
 #include "bkey_methods.h"
 #include "btree_cache.h"
 #include "btree_gc.h"
+#include "btree_journal_iter.h"
 #include "btree_update.h"
 #include "btree_update_interior.h"
 #include "btree_io.h"
@@ -17,7 +18,6 @@
 #include "journal.h"
 #include "journal_reclaim.h"
 #include "keylist.h"
-#include "recovery.h"
 #include "replicas.h"
 #include "super-io.h"
 #include "trace.h"
diff --git a/libbcachefs/buckets.h b/libbcachefs/buckets.h
index a418f664..f192809f 100644
--- a/libbcachefs/buckets.h
+++ b/libbcachefs/buckets.h
@@ -10,7 +10,31 @@
 
 #include "buckets_types.h"
 #include "extents.h"
-#include "super.h"
+#include "sb-members.h"
+
+static inline size_t sector_to_bucket(const struct bch_dev *ca, sector_t s)
+{
+	return div_u64(s, ca->mi.bucket_size);
+}
+
+static inline sector_t bucket_to_sector(const struct bch_dev *ca, size_t b)
+{
+	return ((sector_t) b) * ca->mi.bucket_size;
+}
+
+static inline sector_t bucket_remainder(const struct bch_dev *ca, sector_t s)
+{
+	u32 remainder;
+
+	div_u64_rem(s, ca->mi.bucket_size, &remainder);
+	return remainder;
+}
+
+static inline size_t sector_to_bucket_and_offset(const struct bch_dev *ca, sector_t s,
+						 u32 *offset)
+{
+	return div_u64_rem(s, ca->mi.bucket_size, offset);
+}
 
 #define for_each_bucket(_b, _buckets)				\
 	for (_b = (_buckets)->b + (_buckets)->first_bucket;	\
@@ -292,6 +316,27 @@ int bch2_trans_mark_metadata_bucket(struct btree_trans *, struct bch_dev *,
 				    size_t, enum bch_data_type, unsigned);
 int bch2_trans_mark_dev_sb(struct bch_fs *, struct bch_dev *);
 
+static inline bool is_superblock_bucket(struct bch_dev *ca, u64 b)
+{
+	struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
+	u64 b_offset	= bucket_to_sector(ca, b);
+	u64 b_end	= bucket_to_sector(ca, b + 1);
+	unsigned i;
+
+	if (!b)
+		return true;
+
+	for (i = 0; i < layout->nr_superblocks; i++) {
+		u64 offset = le64_to_cpu(layout->sb_offset[i]);
+		u64 end = offset + (1 << layout->sb_max_size_bits);
+
+		if (!(offset >= b_end || end <= b_offset))
+			return true;
+	}
+
+	return false;
+}
+
 /* disk reservations: */
 
 static inline void bch2_disk_reservation_put(struct bch_fs *c,
diff --git a/libbcachefs/checksum.c b/libbcachefs/checksum.c
index a08997a5..a0ef85eb 100644
--- a/libbcachefs/checksum.c
+++ b/libbcachefs/checksum.c
@@ -360,7 +360,7 @@ struct bch_csum bch2_checksum_merge(unsigned type, struct bch_csum a,
 
 	state.type = type;
 	bch2_checksum_init(&state);
-	state.seed = a.lo;
+	state.seed = (u64 __force) a.lo;
 
 	BUG_ON(!bch2_checksum_mergeable(type));
 
@@ -371,7 +371,7 @@ struct bch_csum bch2_checksum_merge(unsigned type, struct bch_csum a,
 				page_address(ZERO_PAGE(0)), b);
 		b_len -= b;
 	}
-	a.lo = bch2_checksum_final(&state);
+	a.lo = (__le64 __force) bch2_checksum_final(&state);
 	a.lo ^= b.lo;
 	a.hi ^= b.hi;
 	return a;
@@ -426,7 +426,7 @@ int bch2_rechecksum_bio(struct bch_fs *c, struct bio *bio,
 		merged = bch2_checksum_bio(c, crc_old.csum_type,
 				extent_nonce(version, crc_old), bio);
 
-	if (bch2_crc_cmp(merged, crc_old.csum)) {
+	if (bch2_crc_cmp(merged, crc_old.csum) && !c->opts.no_data_io) {
 		bch_err(c, "checksum error in bch2_rechecksum_bio() (memory corruption or bug?)\n"
 			"expected %0llx:%0llx got %0llx:%0llx (old type %s new type %s)",
 			crc_old.csum.hi,
@@ -458,6 +458,48 @@ int bch2_rechecksum_bio(struct bch_fs *c, struct bio *bio,
 	return 0;
 }
 
+/* BCH_SB_FIELD_crypt: */
+
+static int bch2_sb_crypt_validate(struct bch_sb *sb,
+				  struct bch_sb_field *f,
+				  struct printbuf *err)
+{
+	struct bch_sb_field_crypt *crypt = field_to_type(f, crypt);
+
+	if (vstruct_bytes(&crypt->field) < sizeof(*crypt)) {
+		prt_printf(err, "wrong size (got %zu should be %zu)",
+		       vstruct_bytes(&crypt->field), sizeof(*crypt));
+		return -BCH_ERR_invalid_sb_crypt;
+	}
+
+	if (BCH_CRYPT_KDF_TYPE(crypt)) {
+		prt_printf(err, "bad kdf type %llu", BCH_CRYPT_KDF_TYPE(crypt));
+		return -BCH_ERR_invalid_sb_crypt;
+	}
+
+	return 0;
+}
+
+static void bch2_sb_crypt_to_text(struct printbuf *out, struct bch_sb *sb,
+				  struct bch_sb_field *f)
+{
+	struct bch_sb_field_crypt *crypt = field_to_type(f, crypt);
+
+	prt_printf(out, "KFD:               %llu", BCH_CRYPT_KDF_TYPE(crypt));
+	prt_newline(out);
+	prt_printf(out, "scrypt n:          %llu", BCH_KDF_SCRYPT_N(crypt));
+	prt_newline(out);
+	prt_printf(out, "scrypt r:          %llu", BCH_KDF_SCRYPT_R(crypt));
+	prt_newline(out);
+	prt_printf(out, "scrypt p:          %llu", BCH_KDF_SCRYPT_P(crypt));
+	prt_newline(out);
+}
+
+const struct bch_sb_field_ops bch_sb_field_ops_crypt = {
+	.validate	= bch2_sb_crypt_validate,
+	.to_text	= bch2_sb_crypt_to_text,
+};
+
 #ifdef __KERNEL__
 static int __bch2_request_key(char *key_description, struct bch_key *key)
 {
@@ -597,7 +639,7 @@ int bch2_disable_encryption(struct bch_fs *c)
 	if (ret)
 		goto out;
 
-	crypt->key.magic	= BCH_KEY_MAGIC;
+	crypt->key.magic	= cpu_to_le64(BCH_KEY_MAGIC);
 	crypt->key.key		= key;
 
 	SET_BCH_SB_ENCRYPTION_TYPE(c->disk_sb.sb, 0);
@@ -625,7 +667,7 @@ int bch2_enable_encryption(struct bch_fs *c, bool keyed)
 	if (ret)
 		goto err;
 
-	key.magic = BCH_KEY_MAGIC;
+	key.magic = cpu_to_le64(BCH_KEY_MAGIC);
 	get_random_bytes(&key.key, sizeof(key.key));
 
 	if (keyed) {
diff --git a/libbcachefs/checksum.h b/libbcachefs/checksum.h
index 1ad1d5f0..c7b1a8fc 100644
--- a/libbcachefs/checksum.h
+++ b/libbcachefs/checksum.h
@@ -72,6 +72,8 @@ static inline int bch2_encrypt_bio(struct bch_fs *c, unsigned type,
 		: 0;
 }
 
+extern const struct bch_sb_field_ops bch_sb_field_ops_crypt;
+
 int bch2_decrypt_sb_key(struct bch_fs *, struct bch_sb_field_crypt *,
 			struct bch_key *);
 
diff --git a/libbcachefs/disk_groups.c b/libbcachefs/disk_groups.c
index de14ca3a..f36472c4 100644
--- a/libbcachefs/disk_groups.c
+++ b/libbcachefs/disk_groups.c
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 #include "bcachefs.h"
 #include "disk_groups.h"
+#include "sb-members.h"
 #include "super-io.h"
 
 #include <linux/sort.h>
diff --git a/libbcachefs/errcode.h b/libbcachefs/errcode.h
index 735eb241..f7fa8744 100644
--- a/libbcachefs/errcode.h
+++ b/libbcachefs/errcode.h
@@ -213,6 +213,12 @@
 	x(BCH_ERR_invalid_sb,		invalid_sb_quota)			\
 	x(BCH_ERR_invalid,		invalid_bkey)				\
 	x(BCH_ERR_operation_blocked,    nocow_lock_blocked)			\
+	x(EIO,				btree_node_read_err)			\
+	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_fixable)		\
+	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_want_retry)		\
+	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_must_retry)		\
+	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_bad_node)		\
+	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_incompatible)
 
 enum bch_errcode {
 	BCH_ERR_START		= 2048,
diff --git a/libbcachefs/extents.c b/libbcachefs/extents.c
index c13e0afc..7a3f42f3 100644
--- a/libbcachefs/extents.c
+++ b/libbcachefs/extents.c
@@ -517,13 +517,13 @@ static void bch2_extent_crc_pack(union bch_extent_crc *dst,
 	switch (type) {
 	case BCH_EXTENT_ENTRY_crc32:
 		set_common_fields(dst->crc32, src);
-		memcpy(&dst->crc32.csum, &src.csum.lo, sizeof(dst->crc32.csum));
+		dst->crc32.csum	 	= (u32 __force) *((__le32 *) &src.csum.lo);
 		break;
 	case BCH_EXTENT_ENTRY_crc64:
 		set_common_fields(dst->crc64, src);
 		dst->crc64.nonce	= src.nonce;
-		dst->crc64.csum_lo	= src.csum.lo;
-		dst->crc64.csum_hi	= *((__le16 *) &src.csum.hi);
+		dst->crc64.csum_lo	= (u64 __force) src.csum.lo;
+		dst->crc64.csum_hi	= (u64 __force) *((__le16 *) &src.csum.hi);
 		break;
 	case BCH_EXTENT_ENTRY_crc128:
 		set_common_fields(dst->crc128, src);
diff --git a/libbcachefs/extents.h b/libbcachefs/extents.h
index 6e9d23a0..7ee8d031 100644
--- a/libbcachefs/extents.h
+++ b/libbcachefs/extents.h
@@ -155,7 +155,7 @@ bch2_extent_crc_unpack(const struct bkey *k, const union bch_extent_crc *crc)
 			common_fields(crc->crc32),
 		};
 
-		memcpy(&ret.csum.lo, &crc->crc32.csum, sizeof(crc->crc32.csum));
+		*((__le32 *) &ret.csum.lo) = (__le32 __force) crc->crc32.csum;
 		return ret;
 	}
 	case BCH_EXTENT_ENTRY_crc64: {
@@ -165,8 +165,8 @@ bch2_extent_crc_unpack(const struct bkey *k, const union bch_extent_crc *crc)
 			.csum.lo		= (__force __le64) crc->crc64.csum_lo,
 		};
 
-		u16 hi = crc->crc64.csum_hi;
-		memcpy(&ret.csum.hi, &hi, sizeof(hi));
+		*((__le16 *) &ret.csum.hi) = (__le16 __force) crc->crc64.csum_hi;
+
 		return ret;
 	}
 	case BCH_EXTENT_ENTRY_crc128: {
diff --git a/libbcachefs/fs-io-buffered.c b/libbcachefs/fs-io-buffered.c
new file mode 100644
index 00000000..f6c8d216
--- /dev/null
+++ b/libbcachefs/fs-io-buffered.c
@@ -0,0 +1,1102 @@
+// SPDX-License-Identifier: GPL-2.0
+#ifndef NO_BCACHEFS_FS
+
+#include "bcachefs.h"
+#include "alloc_foreground.h"
+#include "bkey_buf.h"
+#include "fs-io.h"
+#include "fs-io-buffered.h"
+#include "fs-io-direct.h"
+#include "fs-io-pagecache.h"
+#include "io.h"
+
+#include <linux/backing-dev.h>
+#include <linux/pagemap.h>
+#include <linux/writeback.h>
+
+static inline bool bio_full(struct bio *bio, unsigned len)
+{
+	if (bio->bi_vcnt >= bio->bi_max_vecs)
+		return true;
+	if (bio->bi_iter.bi_size > UINT_MAX - len)
+		return true;
+	return false;
+}
+
+/* readpage(s): */
+
+static void bch2_readpages_end_io(struct bio *bio)
+{
+	struct folio_iter fi;
+
+	bio_for_each_folio_all(fi, bio) {
+		if (!bio->bi_status) {
+			folio_mark_uptodate(fi.folio);
+		} else {
+			folio_clear_uptodate(fi.folio);
+			folio_set_error(fi.folio);
+		}
+		folio_unlock(fi.folio);
+	}
+
+	bio_put(bio);
+}
+
+struct readpages_iter {
+	struct address_space	*mapping;
+	unsigned		idx;
+	folios			folios;
+};
+
+static int readpages_iter_init(struct readpages_iter *iter,
+			       struct readahead_control *ractl)
+{
+	struct folio **fi;
+	int ret;
+
+	memset(iter, 0, sizeof(*iter));
+
+	iter->mapping = ractl->mapping;
+
+	ret = bch2_filemap_get_contig_folios_d(iter->mapping,
+				ractl->_index << PAGE_SHIFT,
+				(ractl->_index + ractl->_nr_pages) << PAGE_SHIFT,
+				0, mapping_gfp_mask(iter->mapping),
+				&iter->folios);
+	if (ret)
+		return ret;
+
+	darray_for_each(iter->folios, fi) {
+		ractl->_nr_pages -= 1U << folio_order(*fi);
+		__bch2_folio_create(*fi, __GFP_NOFAIL|GFP_KERNEL);
+		folio_put(*fi);
+		folio_put(*fi);
+	}
+
+	return 0;
+}
+
+static inline struct folio *readpage_iter_peek(struct readpages_iter *iter)
+{
+	if (iter->idx >= iter->folios.nr)
+		return NULL;
+	return iter->folios.data[iter->idx];
+}
+
+static inline void readpage_iter_advance(struct readpages_iter *iter)
+{
+	iter->idx++;
+}
+
+static bool extent_partial_reads_expensive(struct bkey_s_c k)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	struct bch_extent_crc_unpacked crc;
+	const union bch_extent_entry *i;
+
+	bkey_for_each_crc(k.k, ptrs, crc, i)
+		if (crc.csum_type || crc.compression_type)
+			return true;
+	return false;
+}
+
+static int readpage_bio_extend(struct btree_trans *trans,
+			       struct readpages_iter *iter,
+			       struct bio *bio,
+			       unsigned sectors_this_extent,
+			       bool get_more)
+{
+	/* Don't hold btree locks while allocating memory: */
+	bch2_trans_unlock(trans);
+
+	while (bio_sectors(bio) < sectors_this_extent &&
+	       bio->bi_vcnt < bio->bi_max_vecs) {
+		struct folio *folio = readpage_iter_peek(iter);
+		int ret;
+
+		if (folio) {
+			readpage_iter_advance(iter);
+		} else {
+			pgoff_t folio_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
+
+			if (!get_more)
+				break;
+
+			folio = xa_load(&iter->mapping->i_pages, folio_offset);
+			if (folio && !xa_is_value(folio))
+				break;
+
+			folio = filemap_alloc_folio(readahead_gfp_mask(iter->mapping), 0);
+			if (!folio)
+				break;
+
+			if (!__bch2_folio_create(folio, GFP_KERNEL)) {
+				folio_put(folio);
+				break;
+			}
+
+			ret = filemap_add_folio(iter->mapping, folio, folio_offset, GFP_KERNEL);
+			if (ret) {
+				__bch2_folio_release(folio);
+				folio_put(folio);
+				break;
+			}
+
+			folio_put(folio);
+		}
+
+		BUG_ON(folio_sector(folio) != bio_end_sector(bio));
+
+		BUG_ON(!bio_add_folio(bio, folio, folio_size(folio), 0));
+	}
+
+	return bch2_trans_relock(trans);
+}
+
+static void bchfs_read(struct btree_trans *trans,
+		       struct bch_read_bio *rbio,
+		       subvol_inum inum,
+		       struct readpages_iter *readpages_iter)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_buf sk;
+	int flags = BCH_READ_RETRY_IF_STALE|
+		BCH_READ_MAY_PROMOTE;
+	u32 snapshot;
+	int ret = 0;
+
+	rbio->c = c;
+	rbio->start_time = local_clock();
+	rbio->subvol = inum.subvol;
+
+	bch2_bkey_buf_init(&sk);
+retry:
+	bch2_trans_begin(trans);
+	iter = (struct btree_iter) { NULL };
+
+	ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+	if (ret)
+		goto err;
+
+	bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
+			     SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
+			     BTREE_ITER_SLOTS);
+	while (1) {
+		struct bkey_s_c k;
+		unsigned bytes, sectors, offset_into_extent;
+		enum btree_id data_btree = BTREE_ID_extents;
+
+		/*
+		 * read_extent -> io_time_reset may cause a transaction restart
+		 * without returning an error, we need to check for that here:
+		 */
+		ret = bch2_trans_relock(trans);
+		if (ret)
+			break;
+
+		bch2_btree_iter_set_pos(&iter,
+				POS(inum.inum, rbio->bio.bi_iter.bi_sector));
+
+		k = bch2_btree_iter_peek_slot(&iter);
+		ret = bkey_err(k);
+		if (ret)
+			break;
+
+		offset_into_extent = iter.pos.offset -
+			bkey_start_offset(k.k);
+		sectors = k.k->size - offset_into_extent;
+
+		bch2_bkey_buf_reassemble(&sk, c, k);
+
+		ret = bch2_read_indirect_extent(trans, &data_btree,
+					&offset_into_extent, &sk);
+		if (ret)
+			break;
+
+		k = bkey_i_to_s_c(sk.k);
+
+		sectors = min(sectors, k.k->size - offset_into_extent);
+
+		if (readpages_iter) {
+			ret = readpage_bio_extend(trans, readpages_iter, &rbio->bio, sectors,
+						  extent_partial_reads_expensive(k));
+			if (ret)
+				break;
+		}
+
+		bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
+		swap(rbio->bio.bi_iter.bi_size, bytes);
+
+		if (rbio->bio.bi_iter.bi_size == bytes)
+			flags |= BCH_READ_LAST_FRAGMENT;
+
+		bch2_bio_page_state_set(&rbio->bio, k);
+
+		bch2_read_extent(trans, rbio, iter.pos,
+				 data_btree, k, offset_into_extent, flags);
+
+		if (flags & BCH_READ_LAST_FRAGMENT)
+			break;
+
+		swap(rbio->bio.bi_iter.bi_size, bytes);
+		bio_advance(&rbio->bio, bytes);
+
+		ret = btree_trans_too_many_iters(trans);
+		if (ret)
+			break;
+	}
+err:
+	bch2_trans_iter_exit(trans, &iter);
+
+	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+		goto retry;
+
+	if (ret) {
+		bch_err_inum_offset_ratelimited(c,
+				iter.pos.inode,
+				iter.pos.offset << 9,
+				"read error %i from btree lookup", ret);
+		rbio->bio.bi_status = BLK_STS_IOERR;
+		bio_endio(&rbio->bio);
+	}
+
+	bch2_bkey_buf_exit(&sk, c);
+}
+
+void bch2_readahead(struct readahead_control *ractl)
+{
+	struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	struct bch_io_opts opts;
+	struct btree_trans trans;
+	struct folio *folio;
+	struct readpages_iter readpages_iter;
+	int ret;
+
+	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
+
+	ret = readpages_iter_init(&readpages_iter, ractl);
+	BUG_ON(ret);
+
+	bch2_trans_init(&trans, c, 0, 0);
+
+	bch2_pagecache_add_get(inode);
+
+	while ((folio = readpage_iter_peek(&readpages_iter))) {
+		unsigned n = min_t(unsigned,
+				   readpages_iter.folios.nr -
+				   readpages_iter.idx,
+				   BIO_MAX_VECS);
+		struct bch_read_bio *rbio =
+			rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
+						   GFP_KERNEL, &c->bio_read),
+				  opts);
+
+		readpage_iter_advance(&readpages_iter);
+
+		rbio->bio.bi_iter.bi_sector = folio_sector(folio);
+		rbio->bio.bi_end_io = bch2_readpages_end_io;
+		BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
+
+		bchfs_read(&trans, rbio, inode_inum(inode),
+			   &readpages_iter);
+		bch2_trans_unlock(&trans);
+	}
+
+	bch2_pagecache_add_put(inode);
+
+	bch2_trans_exit(&trans);
+	darray_exit(&readpages_iter.folios);
+}
+
+static void __bchfs_readfolio(struct bch_fs *c, struct bch_read_bio *rbio,
+			     subvol_inum inum, struct folio *folio)
+{
+	struct btree_trans trans;
+
+	bch2_folio_create(folio, __GFP_NOFAIL);
+
+	rbio->bio.bi_opf = REQ_OP_READ|REQ_SYNC;
+	rbio->bio.bi_iter.bi_sector = folio_sector(folio);
+	BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
+
+	bch2_trans_init(&trans, c, 0, 0);
+	bchfs_read(&trans, rbio, inum, NULL);
+	bch2_trans_exit(&trans);
+}
+
+static void bch2_read_single_folio_end_io(struct bio *bio)
+{
+	complete(bio->bi_private);
+}
+
+int bch2_read_single_folio(struct folio *folio, struct address_space *mapping)
+{
+	struct bch_inode_info *inode = to_bch_ei(mapping->host);
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	struct bch_read_bio *rbio;
+	struct bch_io_opts opts;
+	int ret;
+	DECLARE_COMPLETION_ONSTACK(done);
+
+	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
+
+	rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_KERNEL, &c->bio_read),
+			 opts);
+	rbio->bio.bi_private = &done;
+	rbio->bio.bi_end_io = bch2_read_single_folio_end_io;
+
+	__bchfs_readfolio(c, rbio, inode_inum(inode), folio);
+	wait_for_completion(&done);
+
+	ret = blk_status_to_errno(rbio->bio.bi_status);
+	bio_put(&rbio->bio);
+
+	if (ret < 0)
+		return ret;
+
+	folio_mark_uptodate(folio);
+	return 0;
+}
+
+int bch2_read_folio(struct file *file, struct folio *folio)
+{
+	int ret;
+
+	ret = bch2_read_single_folio(folio, folio->mapping);
+	folio_unlock(folio);
+	return bch2_err_class(ret);
+}
+
+/* writepages: */
+
+struct bch_writepage_io {
+	struct bch_inode_info		*inode;
+
+	/* must be last: */
+	struct bch_write_op		op;
+};
+
+struct bch_writepage_state {
+	struct bch_writepage_io	*io;
+	struct bch_io_opts	opts;
+	struct bch_folio_sector	*tmp;
+	unsigned		tmp_sectors;
+};
+
+static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
+								  struct bch_inode_info *inode)
+{
+	struct bch_writepage_state ret = { 0 };
+
+	bch2_inode_opts_get(&ret.opts, c, &inode->ei_inode);
+	return ret;
+}
+
+static void bch2_writepage_io_done(struct bch_write_op *op)
+{
+	struct bch_writepage_io *io =
+		container_of(op, struct bch_writepage_io, op);
+	struct bch_fs *c = io->op.c;
+	struct bio *bio = &io->op.wbio.bio;
+	struct folio_iter fi;
+	unsigned i;
+
+	if (io->op.error) {
+		set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
+
+		bio_for_each_folio_all(fi, bio) {
+			struct bch_folio *s;
+
+			folio_set_error(fi.folio);
+			mapping_set_error(fi.folio->mapping, -EIO);
+
+			s = __bch2_folio(fi.folio);
+			spin_lock(&s->lock);
+			for (i = 0; i < folio_sectors(fi.folio); i++)
+				s->s[i].nr_replicas = 0;
+			spin_unlock(&s->lock);
+		}
+	}
+
+	if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
+		bio_for_each_folio_all(fi, bio) {
+			struct bch_folio *s;
+
+			s = __bch2_folio(fi.folio);
+			spin_lock(&s->lock);
+			for (i = 0; i < folio_sectors(fi.folio); i++)
+				s->s[i].nr_replicas = 0;
+			spin_unlock(&s->lock);
+		}
+	}
+
+	/*
+	 * racing with fallocate can cause us to add fewer sectors than
+	 * expected - but we shouldn't add more sectors than expected:
+	 */
+	WARN_ON_ONCE(io->op.i_sectors_delta > 0);
+
+	/*
+	 * (error (due to going RO) halfway through a page can screw that up
+	 * slightly)
+	 * XXX wtf?
+	   BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
+	 */
+
+	/*
+	 * PageWriteback is effectively our ref on the inode - fixup i_blocks
+	 * before calling end_page_writeback:
+	 */
+	bch2_i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
+
+	bio_for_each_folio_all(fi, bio) {
+		struct bch_folio *s = __bch2_folio(fi.folio);
+
+		if (atomic_dec_and_test(&s->write_count))
+			folio_end_writeback(fi.folio);
+	}
+
+	bio_put(&io->op.wbio.bio);
+}
+
+static void bch2_writepage_do_io(struct bch_writepage_state *w)
+{
+	struct bch_writepage_io *io = w->io;
+
+	w->io = NULL;
+	closure_call(&io->op.cl, bch2_write, NULL, NULL);
+}
+
+/*
+ * Get a bch_writepage_io and add @page to it - appending to an existing one if
+ * possible, else allocating a new one:
+ */
+static void bch2_writepage_io_alloc(struct bch_fs *c,
+				    struct writeback_control *wbc,
+				    struct bch_writepage_state *w,
+				    struct bch_inode_info *inode,
+				    u64 sector,
+				    unsigned nr_replicas)
+{
+	struct bch_write_op *op;
+
+	w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
+					      REQ_OP_WRITE,
+					      GFP_KERNEL,
+					      &c->writepage_bioset),
+			     struct bch_writepage_io, op.wbio.bio);
+
+	w->io->inode		= inode;
+	op			= &w->io->op;
+	bch2_write_op_init(op, c, w->opts);
+	op->target		= w->opts.foreground_target;
+	op->nr_replicas		= nr_replicas;
+	op->res.nr_replicas	= nr_replicas;
+	op->write_point		= writepoint_hashed(inode->ei_last_dirtied);
+	op->subvol		= inode->ei_subvol;
+	op->pos			= POS(inode->v.i_ino, sector);
+	op->end_io		= bch2_writepage_io_done;
+	op->devs_need_flush	= &inode->ei_devs_need_flush;
+	op->wbio.bio.bi_iter.bi_sector = sector;
+	op->wbio.bio.bi_opf	= wbc_to_write_flags(wbc);
+}
+
+static int __bch2_writepage(struct folio *folio,
+			    struct writeback_control *wbc,
+			    void *data)
+{
+	struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	struct bch_writepage_state *w = data;
+	struct bch_folio *s;
+	unsigned i, offset, f_sectors, nr_replicas_this_write = U32_MAX;
+	loff_t i_size = i_size_read(&inode->v);
+	int ret;
+
+	EBUG_ON(!folio_test_uptodate(folio));
+
+	/* Is the folio fully inside i_size? */
+	if (folio_end_pos(folio) <= i_size)
+		goto do_io;
+
+	/* Is the folio fully outside i_size? (truncate in progress) */
+	if (folio_pos(folio) >= i_size) {
+		folio_unlock(folio);
+		return 0;
+	}
+
+	/*
+	 * The folio straddles i_size.  It must be zeroed out on each and every
+	 * writepage invocation because it may be mmapped.  "A file is mapped
+	 * in multiples of the folio size.  For a file that is not a multiple of
+	 * the  folio size, the remaining memory is zeroed when mapped, and
+	 * writes to that region are not written out to the file."
+	 */
+	folio_zero_segment(folio,
+			   i_size - folio_pos(folio),
+			   folio_size(folio));
+do_io:
+	f_sectors = folio_sectors(folio);
+	s = bch2_folio(folio);
+
+	if (f_sectors > w->tmp_sectors) {
+		kfree(w->tmp);
+		w->tmp = kzalloc(sizeof(struct bch_folio_sector) *
+				 f_sectors, __GFP_NOFAIL);
+		w->tmp_sectors = f_sectors;
+	}
+
+	/*
+	 * Things get really hairy with errors during writeback:
+	 */
+	ret = bch2_get_folio_disk_reservation(c, inode, folio, false);
+	BUG_ON(ret);
+
+	/* Before unlocking the page, get copy of reservations: */
+	spin_lock(&s->lock);
+	memcpy(w->tmp, s->s, sizeof(struct bch_folio_sector) * f_sectors);
+
+	for (i = 0; i < f_sectors; i++) {
+		if (s->s[i].state < SECTOR_dirty)
+			continue;
+
+		nr_replicas_this_write =
+			min_t(unsigned, nr_replicas_this_write,
+			      s->s[i].nr_replicas +
+			      s->s[i].replicas_reserved);
+	}
+
+	for (i = 0; i < f_sectors; i++) {
+		if (s->s[i].state < SECTOR_dirty)
+			continue;
+
+		s->s[i].nr_replicas = w->opts.compression
+			? 0 : nr_replicas_this_write;
+
+		s->s[i].replicas_reserved = 0;
+		bch2_folio_sector_set(folio, s, i, SECTOR_allocated);
+	}
+	spin_unlock(&s->lock);
+
+	BUG_ON(atomic_read(&s->write_count));
+	atomic_set(&s->write_count, 1);
+
+	BUG_ON(folio_test_writeback(folio));
+	folio_start_writeback(folio);
+
+	folio_unlock(folio);
+
+	offset = 0;
+	while (1) {
+		unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
+		u64 sector;
+
+		while (offset < f_sectors &&
+		       w->tmp[offset].state < SECTOR_dirty)
+			offset++;
+
+		if (offset == f_sectors)
+			break;
+
+		while (offset + sectors < f_sectors &&
+		       w->tmp[offset + sectors].state >= SECTOR_dirty) {
+			reserved_sectors += w->tmp[offset + sectors].replicas_reserved;
+			dirty_sectors += w->tmp[offset + sectors].state == SECTOR_dirty;
+			sectors++;
+		}
+		BUG_ON(!sectors);
+
+		sector = folio_sector(folio) + offset;
+
+		if (w->io &&
+		    (w->io->op.res.nr_replicas != nr_replicas_this_write ||
+		     bio_full(&w->io->op.wbio.bio, sectors << 9) ||
+		     w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
+		     (BIO_MAX_VECS * PAGE_SIZE) ||
+		     bio_end_sector(&w->io->op.wbio.bio) != sector))
+			bch2_writepage_do_io(w);
+
+		if (!w->io)
+			bch2_writepage_io_alloc(c, wbc, w, inode, sector,
+						nr_replicas_this_write);
+
+		atomic_inc(&s->write_count);
+
+		BUG_ON(inode != w->io->inode);
+		BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio,
+				     sectors << 9, offset << 9));
+
+		/* Check for writing past i_size: */
+		WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
+			  round_up(i_size, block_bytes(c)) &&
+			  !test_bit(BCH_FS_EMERGENCY_RO, &c->flags),
+			  "writing past i_size: %llu > %llu (unrounded %llu)\n",
+			  bio_end_sector(&w->io->op.wbio.bio) << 9,
+			  round_up(i_size, block_bytes(c)),
+			  i_size);
+
+		w->io->op.res.sectors += reserved_sectors;
+		w->io->op.i_sectors_delta -= dirty_sectors;
+		w->io->op.new_i_size = i_size;
+
+		offset += sectors;
+	}
+
+	if (atomic_dec_and_test(&s->write_count))
+		folio_end_writeback(folio);
+
+	return 0;
+}
+
+int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
+{
+	struct bch_fs *c = mapping->host->i_sb->s_fs_info;
+	struct bch_writepage_state w =
+		bch_writepage_state_init(c, to_bch_ei(mapping->host));
+	struct blk_plug plug;
+	int ret;
+
+	blk_start_plug(&plug);
+	ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
+	if (w.io)
+		bch2_writepage_do_io(&w);
+	blk_finish_plug(&plug);
+	kfree(w.tmp);
+	return bch2_err_class(ret);
+}
+
+/* buffered writes: */
+
+int bch2_write_begin(struct file *file, struct address_space *mapping,
+		     loff_t pos, unsigned len,
+		     struct page **pagep, void **fsdata)
+{
+	struct bch_inode_info *inode = to_bch_ei(mapping->host);
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	struct bch2_folio_reservation *res;
+	struct folio *folio;
+	unsigned offset;
+	int ret = -ENOMEM;
+
+	res = kmalloc(sizeof(*res), GFP_KERNEL);
+	if (!res)
+		return -ENOMEM;
+
+	bch2_folio_reservation_init(c, inode, res);
+	*fsdata = res;
+
+	bch2_pagecache_add_get(inode);
+
+	folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
+				FGP_LOCK|FGP_WRITE|FGP_CREAT|FGP_STABLE,
+				mapping_gfp_mask(mapping));
+	if (IS_ERR_OR_NULL(folio))
+		goto err_unlock;
+
+	if (folio_test_uptodate(folio))
+		goto out;
+
+	offset = pos - folio_pos(folio);
+	len = min_t(size_t, len, folio_end_pos(folio) - pos);
+
+	/* If we're writing entire folio, don't need to read it in first: */
+	if (!offset && len == folio_size(folio))
+		goto out;
+
+	if (!offset && pos + len >= inode->v.i_size) {
+		folio_zero_segment(folio, len, folio_size(folio));
+		flush_dcache_folio(folio);
+		goto out;
+	}
+
+	if (folio_pos(folio) >= inode->v.i_size) {
+		folio_zero_segments(folio, 0, offset, offset + len, folio_size(folio));
+		flush_dcache_folio(folio);
+		goto out;
+	}
+readpage:
+	ret = bch2_read_single_folio(folio, mapping);
+	if (ret)
+		goto err;
+out:
+	ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
+	if (ret)
+		goto err;
+
+	ret = bch2_folio_reservation_get(c, inode, folio, res, offset, len);
+	if (ret) {
+		if (!folio_test_uptodate(folio)) {
+			/*
+			 * If the folio hasn't been read in, we won't know if we
+			 * actually need a reservation - we don't actually need
+			 * to read here, we just need to check if the folio is
+			 * fully backed by uncompressed data:
+			 */
+			goto readpage;
+		}
+
+		goto err;
+	}
+
+	*pagep = &folio->page;
+	return 0;
+err:
+	folio_unlock(folio);
+	folio_put(folio);
+	*pagep = NULL;
+err_unlock:
+	bch2_pagecache_add_put(inode);
+	kfree(res);
+	*fsdata = NULL;
+	return bch2_err_class(ret);
+}
+
+int bch2_write_end(struct file *file, struct address_space *mapping,
+		   loff_t pos, unsigned len, unsigned copied,
+		   struct page *page, void *fsdata)
+{
+	struct bch_inode_info *inode = to_bch_ei(mapping->host);
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	struct bch2_folio_reservation *res = fsdata;
+	struct folio *folio = page_folio(page);
+	unsigned offset = pos - folio_pos(folio);
+
+	lockdep_assert_held(&inode->v.i_rwsem);
+	BUG_ON(offset + copied > folio_size(folio));
+
+	if (unlikely(copied < len && !folio_test_uptodate(folio))) {
+		/*
+		 * The folio needs to be read in, but that would destroy
+		 * our partial write - simplest thing is to just force
+		 * userspace to redo the write:
+		 */
+		folio_zero_range(folio, 0, folio_size(folio));
+		flush_dcache_folio(folio);
+		copied = 0;
+	}
+
+	spin_lock(&inode->v.i_lock);
+	if (pos + copied > inode->v.i_size)
+		i_size_write(&inode->v, pos + copied);
+	spin_unlock(&inode->v.i_lock);
+
+	if (copied) {
+		if (!folio_test_uptodate(folio))
+			folio_mark_uptodate(folio);
+
+		bch2_set_folio_dirty(c, inode, folio, res, offset, copied);
+
+		inode->ei_last_dirtied = (unsigned long) current;
+	}
+
+	folio_unlock(folio);
+	folio_put(folio);
+	bch2_pagecache_add_put(inode);
+
+	bch2_folio_reservation_put(c, inode, res);
+	kfree(res);
+
+	return copied;
+}
+
+static noinline void folios_trunc(folios *folios, struct folio **fi)
+{
+	while (folios->data + folios->nr > fi) {
+		struct folio *f = darray_pop(folios);
+
+		folio_unlock(f);
+		folio_put(f);
+	}
+}
+
+static int __bch2_buffered_write(struct bch_inode_info *inode,
+				 struct address_space *mapping,
+				 struct iov_iter *iter,
+				 loff_t pos, unsigned len)
+{
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	struct bch2_folio_reservation res;
+	folios folios;
+	struct folio **fi, *f;
+	unsigned copied = 0, f_offset;
+	u64 end = pos + len, f_pos;
+	loff_t last_folio_pos = inode->v.i_size;
+	int ret = 0;
+
+	BUG_ON(!len);
+
+	bch2_folio_reservation_init(c, inode, &res);
+	darray_init(&folios);
+
+	ret = bch2_filemap_get_contig_folios_d(mapping, pos, end,
+				   FGP_LOCK|FGP_WRITE|FGP_STABLE|FGP_CREAT,
+				   mapping_gfp_mask(mapping),
+				   &folios);
+	if (ret)
+		goto out;
+
+	BUG_ON(!folios.nr);
+
+	f = darray_first(folios);
+	if (pos != folio_pos(f) && !folio_test_uptodate(f)) {
+		ret = bch2_read_single_folio(f, mapping);
+		if (ret)
+			goto out;
+	}
+
+	f = darray_last(folios);
+	end = min(end, folio_end_pos(f));
+	last_folio_pos = folio_pos(f);
+	if (end != folio_end_pos(f) && !folio_test_uptodate(f)) {
+		if (end >= inode->v.i_size) {
+			folio_zero_range(f, 0, folio_size(f));
+		} else {
+			ret = bch2_read_single_folio(f, mapping);
+			if (ret)
+				goto out;
+		}
+	}
+
+	ret = bch2_folio_set(c, inode_inum(inode), folios.data, folios.nr);
+	if (ret)
+		goto out;
+
+	f_pos = pos;
+	f_offset = pos - folio_pos(darray_first(folios));
+	darray_for_each(folios, fi) {
+		struct folio *f = *fi;
+		u64 f_len = min(end, folio_end_pos(f)) - f_pos;
+
+		/*
+		 * XXX: per POSIX and fstests generic/275, on -ENOSPC we're
+		 * supposed to write as much as we have disk space for.
+		 *
+		 * On failure here we should still write out a partial page if
+		 * we aren't completely out of disk space - we don't do that
+		 * yet:
+		 */
+		ret = bch2_folio_reservation_get(c, inode, f, &res, f_offset, f_len);
+		if (unlikely(ret)) {
+			folios_trunc(&folios, fi);
+			if (!folios.nr)
+				goto out;
+
+			end = min(end, folio_end_pos(darray_last(folios)));
+			break;
+		}
+
+		f_pos = folio_end_pos(f);
+		f_offset = 0;
+	}
+
+	if (mapping_writably_mapped(mapping))
+		darray_for_each(folios, fi)
+			flush_dcache_folio(*fi);
+
+	f_pos = pos;
+	f_offset = pos - folio_pos(darray_first(folios));
+	darray_for_each(folios, fi) {
+		struct folio *f = *fi;
+		u64 f_len = min(end, folio_end_pos(f)) - f_pos;
+		unsigned f_copied = copy_page_from_iter_atomic(&f->page, f_offset, f_len, iter);
+
+		if (!f_copied) {
+			folios_trunc(&folios, fi);
+			break;
+		}
+
+		if (!folio_test_uptodate(f) &&
+		    f_copied != folio_size(f) &&
+		    pos + copied + f_copied < inode->v.i_size) {
+			folio_zero_range(f, 0, folio_size(f));
+			folios_trunc(&folios, fi);
+			break;
+		}
+
+		flush_dcache_folio(f);
+		copied += f_copied;
+
+		if (f_copied != f_len) {
+			folios_trunc(&folios, fi + 1);
+			break;
+		}
+
+		f_pos = folio_end_pos(f);
+		f_offset = 0;
+	}
+
+	if (!copied)
+		goto out;
+
+	end = pos + copied;
+
+	spin_lock(&inode->v.i_lock);
+	if (end > inode->v.i_size)
+		i_size_write(&inode->v, end);
+	spin_unlock(&inode->v.i_lock);
+
+	f_pos = pos;
+	f_offset = pos - folio_pos(darray_first(folios));
+	darray_for_each(folios, fi) {
+		struct folio *f = *fi;
+		u64 f_len = min(end, folio_end_pos(f)) - f_pos;
+
+		if (!folio_test_uptodate(f))
+			folio_mark_uptodate(f);
+
+		bch2_set_folio_dirty(c, inode, f, &res, f_offset, f_len);
+
+		f_pos = folio_end_pos(f);
+		f_offset = 0;
+	}
+
+	inode->ei_last_dirtied = (unsigned long) current;
+out:
+	darray_for_each(folios, fi) {
+		folio_unlock(*fi);
+		folio_put(*fi);
+	}
+
+	/*
+	 * If the last folio added to the mapping starts beyond current EOF, we
+	 * performed a short write but left around at least one post-EOF folio.
+	 * Clean up the mapping before we return.
+	 */
+	if (last_folio_pos >= inode->v.i_size)
+		truncate_pagecache(&inode->v, inode->v.i_size);
+
+	darray_exit(&folios);
+	bch2_folio_reservation_put(c, inode, &res);
+
+	return copied ?: ret;
+}
+
+static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
+{
+	struct file *file = iocb->ki_filp;
+	struct address_space *mapping = file->f_mapping;
+	struct bch_inode_info *inode = file_bch_inode(file);
+	loff_t pos = iocb->ki_pos;
+	ssize_t written = 0;
+	int ret = 0;
+
+	bch2_pagecache_add_get(inode);
+
+	do {
+		unsigned offset = pos & (PAGE_SIZE - 1);
+		unsigned bytes = iov_iter_count(iter);
+again:
+		/*
+		 * Bring in the user page that we will copy from _first_.
+		 * Otherwise there's a nasty deadlock on copying from the
+		 * same page as we're writing to, without it being marked
+		 * up-to-date.
+		 *
+		 * Not only is this an optimisation, but it is also required
+		 * to check that the address is actually valid, when atomic
+		 * usercopies are used, below.
+		 */
+		if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
+			bytes = min_t(unsigned long, iov_iter_count(iter),
+				      PAGE_SIZE - offset);
+
+			if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
+				ret = -EFAULT;
+				break;
+			}
+		}
+
+		if (unlikely(fatal_signal_pending(current))) {
+			ret = -EINTR;
+			break;
+		}
+
+		ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
+		if (unlikely(ret < 0))
+			break;
+
+		cond_resched();
+
+		if (unlikely(ret == 0)) {
+			/*
+			 * If we were unable to copy any data at all, we must
+			 * fall back to a single segment length write.
+			 *
+			 * If we didn't fallback here, we could livelock
+			 * because not all segments in the iov can be copied at
+			 * once without a pagefault.
+			 */
+			bytes = min_t(unsigned long, PAGE_SIZE - offset,
+				      iov_iter_single_seg_count(iter));
+			goto again;
+		}
+		pos += ret;
+		written += ret;
+		ret = 0;
+
+		balance_dirty_pages_ratelimited(mapping);
+	} while (iov_iter_count(iter));
+
+	bch2_pagecache_add_put(inode);
+
+	return written ? written : ret;
+}
+
+ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
+{
+	struct file *file = iocb->ki_filp;
+	struct bch_inode_info *inode = file_bch_inode(file);
+	ssize_t ret;
+
+	if (iocb->ki_flags & IOCB_DIRECT) {
+		ret = bch2_direct_write(iocb, from);
+		goto out;
+	}
+
+	/* We can write back this queue in page reclaim */
+	current->backing_dev_info = inode_to_bdi(&inode->v);
+	inode_lock(&inode->v);
+
+	ret = generic_write_checks(iocb, from);
+	if (ret <= 0)
+		goto unlock;
+
+	ret = file_remove_privs(file);
+	if (ret)
+		goto unlock;
+
+	ret = file_update_time(file);
+	if (ret)
+		goto unlock;
+
+	ret = bch2_buffered_write(iocb, from);
+	if (likely(ret > 0))
+		iocb->ki_pos += ret;
+unlock:
+	inode_unlock(&inode->v);
+	current->backing_dev_info = NULL;
+
+	if (ret > 0)
+		ret = generic_write_sync(iocb, ret);
+out:
+	return bch2_err_class(ret);
+}
+
+void bch2_fs_fs_io_buffered_exit(struct bch_fs *c)
+{
+	bioset_exit(&c->writepage_bioset);
+}
+
+int bch2_fs_fs_io_buffered_init(struct bch_fs *c)
+{
+	if (bioset_init(&c->writepage_bioset,
+			4, offsetof(struct bch_writepage_io, op.wbio.bio),
+			BIOSET_NEED_BVECS))
+		return -BCH_ERR_ENOMEM_writepage_bioset_init;
+
+	return 0;
+}
+
+#endif /* NO_BCACHEFS_FS */
diff --git a/libbcachefs/fs-io-buffered.h b/libbcachefs/fs-io-buffered.h
new file mode 100644
index 00000000..74eba219
--- /dev/null
+++ b/libbcachefs/fs-io-buffered.h
@@ -0,0 +1,28 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_FS_IO_BUFFERED_H
+#define _BCACHEFS_FS_IO_BUFFERED_H
+
+#ifndef NO_BCACHEFS_FS
+
+int bch2_read_single_folio(struct folio *, struct address_space *);
+int bch2_read_folio(struct file *, struct folio *);
+
+int bch2_writepages(struct address_space *, struct writeback_control *);
+void bch2_readahead(struct readahead_control *);
+
+int bch2_write_begin(struct file *, struct address_space *, loff_t,
+		     unsigned, struct page **, void **);
+int bch2_write_end(struct file *, struct address_space *, loff_t,
+		   unsigned, unsigned, struct page *, void *);
+
+ssize_t bch2_read_iter(struct kiocb *, struct iov_iter *);
+ssize_t bch2_write_iter(struct kiocb *, struct iov_iter *);
+
+void bch2_fs_fs_io_buffered_exit(struct bch_fs *);
+int bch2_fs_fs_io_buffered_init(struct bch_fs *);
+#else
+static inline void bch2_fs_fs_io_buffered_exit(struct bch_fs *c) {}
+static inline int bch2_fs_fs_io_buffered_init(struct bch_fs *c) { return 0; }
+#endif
+
+#endif /* _BCACHEFS_FS_IO_BUFFERED_H */
diff --git a/libbcachefs/fs-io-direct.c b/libbcachefs/fs-io-direct.c
new file mode 100644
index 00000000..a5cadc5d
--- /dev/null
+++ b/libbcachefs/fs-io-direct.c
@@ -0,0 +1,678 @@
+// SPDX-License-Identifier: GPL-2.0
+#ifndef NO_BCACHEFS_FS
+
+#include "bcachefs.h"
+#include "alloc_foreground.h"
+#include "fs.h"
+#include "fs-io.h"
+#include "fs-io-direct.h"
+#include "fs-io-pagecache.h"
+#include "io.h"
+
+#include <linux/kthread.h>
+#include <linux/pagemap.h>
+#include <linux/task_io_accounting_ops.h>
+
+/* O_DIRECT reads */
+
+struct dio_read {
+	struct closure			cl;
+	struct kiocb			*req;
+	long				ret;
+	bool				should_dirty;
+	struct bch_read_bio		rbio;
+};
+
+static void bio_check_or_release(struct bio *bio, bool check_dirty)
+{
+	if (check_dirty) {
+		bio_check_pages_dirty(bio);
+	} else {
+		bio_release_pages(bio, false);
+		bio_put(bio);
+	}
+}
+
+static void bch2_dio_read_complete(struct closure *cl)
+{
+	struct dio_read *dio = container_of(cl, struct dio_read, cl);
+
+	dio->req->ki_complete(dio->req, dio->ret);
+	bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
+}
+
+static void bch2_direct_IO_read_endio(struct bio *bio)
+{
+	struct dio_read *dio = bio->bi_private;
+
+	if (bio->bi_status)
+		dio->ret = blk_status_to_errno(bio->bi_status);
+
+	closure_put(&dio->cl);
+}
+
+static void bch2_direct_IO_read_split_endio(struct bio *bio)
+{
+	struct dio_read *dio = bio->bi_private;
+	bool should_dirty = dio->should_dirty;
+
+	bch2_direct_IO_read_endio(bio);
+	bio_check_or_release(bio, should_dirty);
+}
+
+static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
+{
+	struct file *file = req->ki_filp;
+	struct bch_inode_info *inode = file_bch_inode(file);
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	struct bch_io_opts opts;
+	struct dio_read *dio;
+	struct bio *bio;
+	loff_t offset = req->ki_pos;
+	bool sync = is_sync_kiocb(req);
+	size_t shorten;
+	ssize_t ret;
+
+	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
+
+	if ((offset|iter->count) & (block_bytes(c) - 1))
+		return -EINVAL;
+
+	ret = min_t(loff_t, iter->count,
+		    max_t(loff_t, 0, i_size_read(&inode->v) - offset));
+
+	if (!ret)
+		return ret;
+
+	shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
+	iter->count -= shorten;
+
+	bio = bio_alloc_bioset(NULL,
+			       bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
+			       REQ_OP_READ,
+			       GFP_KERNEL,
+			       &c->dio_read_bioset);
+
+	bio->bi_end_io = bch2_direct_IO_read_endio;
+
+	dio = container_of(bio, struct dio_read, rbio.bio);
+	closure_init(&dio->cl, NULL);
+
+	/*
+	 * this is a _really_ horrible hack just to avoid an atomic sub at the
+	 * end:
+	 */
+	if (!sync) {
+		set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
+		atomic_set(&dio->cl.remaining,
+			   CLOSURE_REMAINING_INITIALIZER -
+			   CLOSURE_RUNNING +
+			   CLOSURE_DESTRUCTOR);
+	} else {
+		atomic_set(&dio->cl.remaining,
+			   CLOSURE_REMAINING_INITIALIZER + 1);
+	}
+
+	dio->req	= req;
+	dio->ret	= ret;
+	/*
+	 * This is one of the sketchier things I've encountered: we have to skip
+	 * the dirtying of requests that are internal from the kernel (i.e. from
+	 * loopback), because we'll deadlock on page_lock.
+	 */
+	dio->should_dirty = iter_is_iovec(iter);
+
+	goto start;
+	while (iter->count) {
+		bio = bio_alloc_bioset(NULL,
+				       bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
+				       REQ_OP_READ,
+				       GFP_KERNEL,
+				       &c->bio_read);
+		bio->bi_end_io		= bch2_direct_IO_read_split_endio;
+start:
+		bio->bi_opf		= REQ_OP_READ|REQ_SYNC;
+		bio->bi_iter.bi_sector	= offset >> 9;
+		bio->bi_private		= dio;
+
+		ret = bio_iov_iter_get_pages(bio, iter);
+		if (ret < 0) {
+			/* XXX: fault inject this path */
+			bio->bi_status = BLK_STS_RESOURCE;
+			bio_endio(bio);
+			break;
+		}
+
+		offset += bio->bi_iter.bi_size;
+
+		if (dio->should_dirty)
+			bio_set_pages_dirty(bio);
+
+		if (iter->count)
+			closure_get(&dio->cl);
+
+		bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
+	}
+
+	iter->count += shorten;
+
+	if (sync) {
+		closure_sync(&dio->cl);
+		closure_debug_destroy(&dio->cl);
+		ret = dio->ret;
+		bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
+		return ret;
+	} else {
+		return -EIOCBQUEUED;
+	}
+}
+
+ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
+{
+	struct file *file = iocb->ki_filp;
+	struct bch_inode_info *inode = file_bch_inode(file);
+	struct address_space *mapping = file->f_mapping;
+	size_t count = iov_iter_count(iter);
+	ssize_t ret;
+
+	if (!count)
+		return 0; /* skip atime */
+
+	if (iocb->ki_flags & IOCB_DIRECT) {
+		struct blk_plug plug;
+
+		if (unlikely(mapping->nrpages)) {
+			ret = filemap_write_and_wait_range(mapping,
+						iocb->ki_pos,
+						iocb->ki_pos + count - 1);
+			if (ret < 0)
+				goto out;
+		}
+
+		file_accessed(file);
+
+		blk_start_plug(&plug);
+		ret = bch2_direct_IO_read(iocb, iter);
+		blk_finish_plug(&plug);
+
+		if (ret >= 0)
+			iocb->ki_pos += ret;
+	} else {
+		bch2_pagecache_add_get(inode);
+		ret = generic_file_read_iter(iocb, iter);
+		bch2_pagecache_add_put(inode);
+	}
+out:
+	return bch2_err_class(ret);
+}
+
+/* O_DIRECT writes */
+
+struct dio_write {
+	struct kiocb			*req;
+	struct address_space		*mapping;
+	struct bch_inode_info		*inode;
+	struct mm_struct		*mm;
+	unsigned			loop:1,
+					extending:1,
+					sync:1,
+					flush:1,
+					free_iov:1;
+	struct quota_res		quota_res;
+	u64				written;
+
+	struct iov_iter			iter;
+	struct iovec			inline_vecs[2];
+
+	/* must be last: */
+	struct bch_write_op		op;
+};
+
+static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
+				       u64 offset, u64 size,
+				       unsigned nr_replicas, bool compressed)
+{
+	struct btree_trans trans;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	u64 end = offset + size;
+	u32 snapshot;
+	bool ret = true;
+	int err;
+
+	bch2_trans_init(&trans, c, 0, 0);
+retry:
+	bch2_trans_begin(&trans);
+
+	err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
+	if (err)
+		goto err;
+
+	for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
+			   SPOS(inum.inum, offset, snapshot),
+			   BTREE_ITER_SLOTS, k, err) {
+		if (bkey_ge(bkey_start_pos(k.k), POS(inum.inum, end)))
+			break;
+
+		if (k.k->p.snapshot != snapshot ||
+		    nr_replicas > bch2_bkey_replicas(c, k) ||
+		    (!compressed && bch2_bkey_sectors_compressed(k))) {
+			ret = false;
+			break;
+		}
+	}
+
+	offset = iter.pos.offset;
+	bch2_trans_iter_exit(&trans, &iter);
+err:
+	if (bch2_err_matches(err, BCH_ERR_transaction_restart))
+		goto retry;
+	bch2_trans_exit(&trans);
+
+	return err ? false : ret;
+}
+
+static noinline bool bch2_dio_write_check_allocated(struct dio_write *dio)
+{
+	struct bch_fs *c = dio->op.c;
+	struct bch_inode_info *inode = dio->inode;
+	struct bio *bio = &dio->op.wbio.bio;
+
+	return bch2_check_range_allocated(c, inode_inum(inode),
+				dio->op.pos.offset, bio_sectors(bio),
+				dio->op.opts.data_replicas,
+				dio->op.opts.compression != 0);
+}
+
+static void bch2_dio_write_loop_async(struct bch_write_op *);
+static __always_inline long bch2_dio_write_done(struct dio_write *dio);
+
+/*
+ * We're going to return -EIOCBQUEUED, but we haven't finished consuming the
+ * iov_iter yet, so we need to stash a copy of the iovec: it might be on the
+ * caller's stack, we're not guaranteed that it will live for the duration of
+ * the IO:
+ */
+static noinline int bch2_dio_write_copy_iov(struct dio_write *dio)
+{
+	struct iovec *iov = dio->inline_vecs;
+
+	/*
+	 * iov_iter has a single embedded iovec - nothing to do:
+	 */
+	if (iter_is_ubuf(&dio->iter))
+		return 0;
+
+	/*
+	 * We don't currently handle non-iovec iov_iters here - return an error,
+	 * and we'll fall back to doing the IO synchronously:
+	 */
+	if (!iter_is_iovec(&dio->iter))
+		return -1;
+
+	if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
+		iov = kmalloc_array(dio->iter.nr_segs, sizeof(*iov),
+				    GFP_KERNEL);
+		if (unlikely(!iov))
+			return -ENOMEM;
+
+		dio->free_iov = true;
+	}
+
+	memcpy(iov, dio->iter.__iov, dio->iter.nr_segs * sizeof(*iov));
+	dio->iter.__iov = iov;
+	return 0;
+}
+
+static void bch2_dio_write_flush_done(struct closure *cl)
+{
+	struct dio_write *dio = container_of(cl, struct dio_write, op.cl);
+	struct bch_fs *c = dio->op.c;
+
+	closure_debug_destroy(cl);
+
+	dio->op.error = bch2_journal_error(&c->journal);
+
+	bch2_dio_write_done(dio);
+}
+
+static noinline void bch2_dio_write_flush(struct dio_write *dio)
+{
+	struct bch_fs *c = dio->op.c;
+	struct bch_inode_unpacked inode;
+	int ret;
+
+	dio->flush = 0;
+
+	closure_init(&dio->op.cl, NULL);
+
+	if (!dio->op.error) {
+		ret = bch2_inode_find_by_inum(c, inode_inum(dio->inode), &inode);
+		if (ret) {
+			dio->op.error = ret;
+		} else {
+			bch2_journal_flush_seq_async(&c->journal, inode.bi_journal_seq, &dio->op.cl);
+			bch2_inode_flush_nocow_writes_async(c, dio->inode, &dio->op.cl);
+		}
+	}
+
+	if (dio->sync) {
+		closure_sync(&dio->op.cl);
+		closure_debug_destroy(&dio->op.cl);
+	} else {
+		continue_at(&dio->op.cl, bch2_dio_write_flush_done, NULL);
+	}
+}
+
+static __always_inline long bch2_dio_write_done(struct dio_write *dio)
+{
+	struct kiocb *req = dio->req;
+	struct bch_inode_info *inode = dio->inode;
+	bool sync = dio->sync;
+	long ret;
+
+	if (unlikely(dio->flush)) {
+		bch2_dio_write_flush(dio);
+		if (!sync)
+			return -EIOCBQUEUED;
+	}
+
+	bch2_pagecache_block_put(inode);
+
+	if (dio->free_iov)
+		kfree(dio->iter.__iov);
+
+	ret = dio->op.error ?: ((long) dio->written << 9);
+	bio_put(&dio->op.wbio.bio);
+
+	/* inode->i_dio_count is our ref on inode and thus bch_fs */
+	inode_dio_end(&inode->v);
+
+	if (ret < 0)
+		ret = bch2_err_class(ret);
+
+	if (!sync) {
+		req->ki_complete(req, ret);
+		ret = -EIOCBQUEUED;
+	}
+	return ret;
+}
+
+static __always_inline void bch2_dio_write_end(struct dio_write *dio)
+{
+	struct bch_fs *c = dio->op.c;
+	struct kiocb *req = dio->req;
+	struct bch_inode_info *inode = dio->inode;
+	struct bio *bio = &dio->op.wbio.bio;
+
+	req->ki_pos	+= (u64) dio->op.written << 9;
+	dio->written	+= dio->op.written;
+
+	if (dio->extending) {
+		spin_lock(&inode->v.i_lock);
+		if (req->ki_pos > inode->v.i_size)
+			i_size_write(&inode->v, req->ki_pos);
+		spin_unlock(&inode->v.i_lock);
+	}
+
+	if (dio->op.i_sectors_delta || dio->quota_res.sectors) {
+		mutex_lock(&inode->ei_quota_lock);
+		__bch2_i_sectors_acct(c, inode, &dio->quota_res, dio->op.i_sectors_delta);
+		__bch2_quota_reservation_put(c, inode, &dio->quota_res);
+		mutex_unlock(&inode->ei_quota_lock);
+	}
+
+	bio_release_pages(bio, false);
+
+	if (unlikely(dio->op.error))
+		set_bit(EI_INODE_ERROR, &inode->ei_flags);
+}
+
+static __always_inline long bch2_dio_write_loop(struct dio_write *dio)
+{
+	struct bch_fs *c = dio->op.c;
+	struct kiocb *req = dio->req;
+	struct address_space *mapping = dio->mapping;
+	struct bch_inode_info *inode = dio->inode;
+	struct bch_io_opts opts;
+	struct bio *bio = &dio->op.wbio.bio;
+	unsigned unaligned, iter_count;
+	bool sync = dio->sync, dropped_locks;
+	long ret;
+
+	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
+
+	while (1) {
+		iter_count = dio->iter.count;
+
+		EBUG_ON(current->faults_disabled_mapping);
+		current->faults_disabled_mapping = mapping;
+
+		ret = bio_iov_iter_get_pages(bio, &dio->iter);
+
+		dropped_locks = fdm_dropped_locks();
+
+		current->faults_disabled_mapping = NULL;
+
+		/*
+		 * If the fault handler returned an error but also signalled
+		 * that it dropped & retook ei_pagecache_lock, we just need to
+		 * re-shoot down the page cache and retry:
+		 */
+		if (dropped_locks && ret)
+			ret = 0;
+
+		if (unlikely(ret < 0))
+			goto err;
+
+		if (unlikely(dropped_locks)) {
+			ret = bch2_write_invalidate_inode_pages_range(mapping,
+					req->ki_pos,
+					req->ki_pos + iter_count - 1);
+			if (unlikely(ret))
+				goto err;
+
+			if (!bio->bi_iter.bi_size)
+				continue;
+		}
+
+		unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
+		bio->bi_iter.bi_size -= unaligned;
+		iov_iter_revert(&dio->iter, unaligned);
+
+		if (!bio->bi_iter.bi_size) {
+			/*
+			 * bio_iov_iter_get_pages was only able to get <
+			 * blocksize worth of pages:
+			 */
+			ret = -EFAULT;
+			goto err;
+		}
+
+		bch2_write_op_init(&dio->op, c, opts);
+		dio->op.end_io		= sync
+			? NULL
+			: bch2_dio_write_loop_async;
+		dio->op.target		= dio->op.opts.foreground_target;
+		dio->op.write_point	= writepoint_hashed((unsigned long) current);
+		dio->op.nr_replicas	= dio->op.opts.data_replicas;
+		dio->op.subvol		= inode->ei_subvol;
+		dio->op.pos		= POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
+		dio->op.devs_need_flush	= &inode->ei_devs_need_flush;
+
+		if (sync)
+			dio->op.flags |= BCH_WRITE_SYNC;
+		dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
+
+		ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
+						 bio_sectors(bio), true);
+		if (unlikely(ret))
+			goto err;
+
+		ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
+						dio->op.opts.data_replicas, 0);
+		if (unlikely(ret) &&
+		    !bch2_dio_write_check_allocated(dio))
+			goto err;
+
+		task_io_account_write(bio->bi_iter.bi_size);
+
+		if (unlikely(dio->iter.count) &&
+		    !dio->sync &&
+		    !dio->loop &&
+		    bch2_dio_write_copy_iov(dio))
+			dio->sync = sync = true;
+
+		dio->loop = true;
+		closure_call(&dio->op.cl, bch2_write, NULL, NULL);
+
+		if (!sync)
+			return -EIOCBQUEUED;
+
+		bch2_dio_write_end(dio);
+
+		if (likely(!dio->iter.count) || dio->op.error)
+			break;
+
+		bio_reset(bio, NULL, REQ_OP_WRITE);
+	}
+out:
+	return bch2_dio_write_done(dio);
+err:
+	dio->op.error = ret;
+
+	bio_release_pages(bio, false);
+
+	bch2_quota_reservation_put(c, inode, &dio->quota_res);
+	goto out;
+}
+
+static noinline __cold void bch2_dio_write_continue(struct dio_write *dio)
+{
+	struct mm_struct *mm = dio->mm;
+
+	bio_reset(&dio->op.wbio.bio, NULL, REQ_OP_WRITE);
+
+	if (mm)
+		kthread_use_mm(mm);
+	bch2_dio_write_loop(dio);
+	if (mm)
+		kthread_unuse_mm(mm);
+}
+
+static void bch2_dio_write_loop_async(struct bch_write_op *op)
+{
+	struct dio_write *dio = container_of(op, struct dio_write, op);
+
+	bch2_dio_write_end(dio);
+
+	if (likely(!dio->iter.count) || dio->op.error)
+		bch2_dio_write_done(dio);
+	else
+		bch2_dio_write_continue(dio);
+}
+
+ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
+{
+	struct file *file = req->ki_filp;
+	struct address_space *mapping = file->f_mapping;
+	struct bch_inode_info *inode = file_bch_inode(file);
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	struct dio_write *dio;
+	struct bio *bio;
+	bool locked = true, extending;
+	ssize_t ret;
+
+	prefetch(&c->opts);
+	prefetch((void *) &c->opts + 64);
+	prefetch(&inode->ei_inode);
+	prefetch((void *) &inode->ei_inode + 64);
+
+	inode_lock(&inode->v);
+
+	ret = generic_write_checks(req, iter);
+	if (unlikely(ret <= 0))
+		goto err;
+
+	ret = file_remove_privs(file);
+	if (unlikely(ret))
+		goto err;
+
+	ret = file_update_time(file);
+	if (unlikely(ret))
+		goto err;
+
+	if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
+		goto err;
+
+	inode_dio_begin(&inode->v);
+	bch2_pagecache_block_get(inode);
+
+	extending = req->ki_pos + iter->count > inode->v.i_size;
+	if (!extending) {
+		inode_unlock(&inode->v);
+		locked = false;
+	}
+
+	bio = bio_alloc_bioset(NULL,
+			       bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
+			       REQ_OP_WRITE,
+			       GFP_KERNEL,
+			       &c->dio_write_bioset);
+	dio = container_of(bio, struct dio_write, op.wbio.bio);
+	dio->req		= req;
+	dio->mapping		= mapping;
+	dio->inode		= inode;
+	dio->mm			= current->mm;
+	dio->loop		= false;
+	dio->extending		= extending;
+	dio->sync		= is_sync_kiocb(req) || extending;
+	dio->flush		= iocb_is_dsync(req) && !c->opts.journal_flush_disabled;
+	dio->free_iov		= false;
+	dio->quota_res.sectors	= 0;
+	dio->written		= 0;
+	dio->iter		= *iter;
+	dio->op.c		= c;
+
+	if (unlikely(mapping->nrpages)) {
+		ret = bch2_write_invalidate_inode_pages_range(mapping,
+						req->ki_pos,
+						req->ki_pos + iter->count - 1);
+		if (unlikely(ret))
+			goto err_put_bio;
+	}
+
+	ret = bch2_dio_write_loop(dio);
+err:
+	if (locked)
+		inode_unlock(&inode->v);
+	return ret;
+err_put_bio:
+	bch2_pagecache_block_put(inode);
+	bio_put(bio);
+	inode_dio_end(&inode->v);
+	goto err;
+}
+
+void bch2_fs_fs_io_direct_exit(struct bch_fs *c)
+{
+	bioset_exit(&c->dio_write_bioset);
+	bioset_exit(&c->dio_read_bioset);
+}
+
+int bch2_fs_fs_io_direct_init(struct bch_fs *c)
+{
+	if (bioset_init(&c->dio_read_bioset,
+			4, offsetof(struct dio_read, rbio.bio),
+			BIOSET_NEED_BVECS))
+		return -BCH_ERR_ENOMEM_dio_read_bioset_init;
+
+	if (bioset_init(&c->dio_write_bioset,
+			4, offsetof(struct dio_write, op.wbio.bio),
+			BIOSET_NEED_BVECS))
+		return -BCH_ERR_ENOMEM_dio_write_bioset_init;
+
+	return 0;
+}
+
+#endif /* NO_BCACHEFS_FS */
diff --git a/libbcachefs/fs-io-direct.h b/libbcachefs/fs-io-direct.h
new file mode 100644
index 00000000..8e950cca
--- /dev/null
+++ b/libbcachefs/fs-io-direct.h
@@ -0,0 +1,15 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_FS_IO_DIRECT_H
+#define _BCACHEFS_FS_IO_DIRECT_H
+
+#ifndef NO_BCACHEFS_FS
+ssize_t bch2_direct_write(struct kiocb *, struct iov_iter *);
+
+void bch2_fs_fs_io_direct_exit(struct bch_fs *);
+int bch2_fs_fs_io_direct_init(struct bch_fs *);
+#else
+static inline void bch2_fs_fs_io_direct_exit(struct bch_fs *c) {}
+static inline int bch2_fs_fs_io_direct_init(struct bch_fs *c) { return 0; }
+#endif
+
+#endif /* _BCACHEFS_FS_IO_DIRECT_H */
diff --git a/libbcachefs/fs-io-pagecache.c b/libbcachefs/fs-io-pagecache.c
new file mode 100644
index 00000000..07c4bfe3
--- /dev/null
+++ b/libbcachefs/fs-io-pagecache.c
@@ -0,0 +1,777 @@
+// SPDX-License-Identifier: GPL-2.0
+#ifndef NO_BCACHEFS_FS
+
+#include "bcachefs.h"
+#include "btree_iter.h"
+#include "extents.h"
+#include "fs-io.h"
+#include "fs-io-pagecache.h"
+#include "subvolume.h"
+
+#include <linux/pagevec.h>
+#include <linux/writeback.h>
+
+int bch2_filemap_get_contig_folios_d(struct address_space *mapping,
+				     loff_t start, u64 end,
+				     int fgp_flags, gfp_t gfp,
+				     folios *folios)
+{
+	struct folio *f;
+	u64 pos = start;
+	int ret = 0;
+
+	while (pos < end) {
+		if ((u64) pos >= (u64) start + (1ULL << 20))
+			fgp_flags &= ~FGP_CREAT;
+
+		ret = darray_make_room_gfp(folios, 1, gfp & GFP_KERNEL);
+		if (ret)
+			break;
+
+		f = __filemap_get_folio(mapping, pos >> PAGE_SHIFT, fgp_flags, gfp);
+		if (IS_ERR_OR_NULL(f))
+			break;
+
+		BUG_ON(folios->nr && folio_pos(f) != pos);
+
+		pos = folio_end_pos(f);
+		darray_push(folios, f);
+	}
+
+	if (!folios->nr && !ret && (fgp_flags & FGP_CREAT))
+		ret = -ENOMEM;
+
+	return folios->nr ? 0 : ret;
+}
+
+/* pagecache_block must be held */
+int bch2_write_invalidate_inode_pages_range(struct address_space *mapping,
+					    loff_t start, loff_t end)
+{
+	int ret;
+
+	/*
+	 * XXX: the way this is currently implemented, we can spin if a process
+	 * is continually redirtying a specific page
+	 */
+	do {
+		if (!mapping->nrpages)
+			return 0;
+
+		ret = filemap_write_and_wait_range(mapping, start, end);
+		if (ret)
+			break;
+
+		if (!mapping->nrpages)
+			return 0;
+
+		ret = invalidate_inode_pages2_range(mapping,
+				start >> PAGE_SHIFT,
+				end >> PAGE_SHIFT);
+	} while (ret == -EBUSY);
+
+	return ret;
+}
+
+static const char * const bch2_folio_sector_states[] = {
+#define x(n)	#n,
+	BCH_FOLIO_SECTOR_STATE()
+#undef x
+	NULL
+};
+
+static inline enum bch_folio_sector_state
+folio_sector_dirty(enum bch_folio_sector_state state)
+{
+	switch (state) {
+	case SECTOR_unallocated:
+		return SECTOR_dirty;
+	case SECTOR_reserved:
+		return SECTOR_dirty_reserved;
+	default:
+		return state;
+	}
+}
+
+static inline enum bch_folio_sector_state
+folio_sector_undirty(enum bch_folio_sector_state state)
+{
+	switch (state) {
+	case SECTOR_dirty:
+		return SECTOR_unallocated;
+	case SECTOR_dirty_reserved:
+		return SECTOR_reserved;
+	default:
+		return state;
+	}
+}
+
+static inline enum bch_folio_sector_state
+folio_sector_reserve(enum bch_folio_sector_state state)
+{
+	switch (state) {
+	case SECTOR_unallocated:
+		return SECTOR_reserved;
+	case SECTOR_dirty:
+		return SECTOR_dirty_reserved;
+	default:
+		return state;
+	}
+}
+
+/* for newly allocated folios: */
+struct bch_folio *__bch2_folio_create(struct folio *folio, gfp_t gfp)
+{
+	struct bch_folio *s;
+
+	s = kzalloc(sizeof(*s) +
+		    sizeof(struct bch_folio_sector) *
+		    folio_sectors(folio), gfp);
+	if (!s)
+		return NULL;
+
+	spin_lock_init(&s->lock);
+	folio_attach_private(folio, s);
+	return s;
+}
+
+struct bch_folio *bch2_folio_create(struct folio *folio, gfp_t gfp)
+{
+	return bch2_folio(folio) ?: __bch2_folio_create(folio, gfp);
+}
+
+static unsigned bkey_to_sector_state(struct bkey_s_c k)
+{
+	if (bkey_extent_is_reservation(k))
+		return SECTOR_reserved;
+	if (bkey_extent_is_allocation(k.k))
+		return SECTOR_allocated;
+	return SECTOR_unallocated;
+}
+
+static void __bch2_folio_set(struct folio *folio,
+			     unsigned pg_offset, unsigned pg_len,
+			     unsigned nr_ptrs, unsigned state)
+{
+	struct bch_folio *s = bch2_folio(folio);
+	unsigned i, sectors = folio_sectors(folio);
+
+	BUG_ON(pg_offset >= sectors);
+	BUG_ON(pg_offset + pg_len > sectors);
+
+	spin_lock(&s->lock);
+
+	for (i = pg_offset; i < pg_offset + pg_len; i++) {
+		s->s[i].nr_replicas	= nr_ptrs;
+		bch2_folio_sector_set(folio, s, i, state);
+	}
+
+	if (i == sectors)
+		s->uptodate = true;
+
+	spin_unlock(&s->lock);
+}
+
+/*
+ * Initialize bch_folio state (allocated/unallocated, nr_replicas) from the
+ * extents btree:
+ */
+int bch2_folio_set(struct bch_fs *c, subvol_inum inum,
+		   struct folio **folios, unsigned nr_folios)
+{
+	struct btree_trans trans;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	struct bch_folio *s;
+	u64 offset = folio_sector(folios[0]);
+	unsigned folio_idx;
+	u32 snapshot;
+	bool need_set = false;
+	int ret;
+
+	for (folio_idx = 0; folio_idx < nr_folios; folio_idx++) {
+		s = bch2_folio_create(folios[folio_idx], GFP_KERNEL);
+		if (!s)
+			return -ENOMEM;
+
+		need_set |= !s->uptodate;
+	}
+
+	if (!need_set)
+		return 0;
+
+	folio_idx = 0;
+	bch2_trans_init(&trans, c, 0, 0);
+retry:
+	bch2_trans_begin(&trans);
+
+	ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
+	if (ret)
+		goto err;
+
+	for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
+			   SPOS(inum.inum, offset, snapshot),
+			   BTREE_ITER_SLOTS, k, ret) {
+		unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k);
+		unsigned state = bkey_to_sector_state(k);
+
+		while (folio_idx < nr_folios) {
+			struct folio *folio = folios[folio_idx];
+			u64 folio_start	= folio_sector(folio);
+			u64 folio_end	= folio_end_sector(folio);
+			unsigned folio_offset = max(bkey_start_offset(k.k), folio_start) - folio_start;
+			unsigned folio_len = min(k.k->p.offset, folio_end) - folio_offset - folio_start;
+
+			BUG_ON(k.k->p.offset < folio_start);
+			BUG_ON(bkey_start_offset(k.k) > folio_end);
+
+			if (!bch2_folio(folio)->uptodate)
+				__bch2_folio_set(folio, folio_offset, folio_len, nr_ptrs, state);
+
+			if (k.k->p.offset < folio_end)
+				break;
+			folio_idx++;
+		}
+
+		if (folio_idx == nr_folios)
+			break;
+	}
+
+	offset = iter.pos.offset;
+	bch2_trans_iter_exit(&trans, &iter);
+err:
+	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+		goto retry;
+	bch2_trans_exit(&trans);
+
+	return ret;
+}
+
+void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k)
+{
+	struct bvec_iter iter;
+	struct folio_vec fv;
+	unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
+		? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
+	unsigned state = bkey_to_sector_state(k);
+
+	bio_for_each_folio(fv, bio, iter)
+		__bch2_folio_set(fv.fv_folio,
+				 fv.fv_offset >> 9,
+				 fv.fv_len >> 9,
+				 nr_ptrs, state);
+}
+
+void bch2_mark_pagecache_unallocated(struct bch_inode_info *inode,
+				     u64 start, u64 end)
+{
+	pgoff_t index = start >> PAGE_SECTORS_SHIFT;
+	pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
+	struct folio_batch fbatch;
+	unsigned i, j;
+
+	if (end <= start)
+		return;
+
+	folio_batch_init(&fbatch);
+
+	while (filemap_get_folios(inode->v.i_mapping,
+				  &index, end_index, &fbatch)) {
+		for (i = 0; i < folio_batch_count(&fbatch); i++) {
+			struct folio *folio = fbatch.folios[i];
+			u64 folio_start = folio_sector(folio);
+			u64 folio_end = folio_end_sector(folio);
+			unsigned folio_offset = max(start, folio_start) - folio_start;
+			unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
+			struct bch_folio *s;
+
+			BUG_ON(end <= folio_start);
+
+			folio_lock(folio);
+			s = bch2_folio(folio);
+
+			if (s) {
+				spin_lock(&s->lock);
+				for (j = folio_offset; j < folio_offset + folio_len; j++)
+					s->s[j].nr_replicas = 0;
+				spin_unlock(&s->lock);
+			}
+
+			folio_unlock(folio);
+		}
+		folio_batch_release(&fbatch);
+		cond_resched();
+	}
+}
+
+void bch2_mark_pagecache_reserved(struct bch_inode_info *inode,
+				  u64 start, u64 end)
+{
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	pgoff_t index = start >> PAGE_SECTORS_SHIFT;
+	pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
+	struct folio_batch fbatch;
+	s64 i_sectors_delta = 0;
+	unsigned i, j;
+
+	if (end <= start)
+		return;
+
+	folio_batch_init(&fbatch);
+
+	while (filemap_get_folios(inode->v.i_mapping,
+				  &index, end_index, &fbatch)) {
+		for (i = 0; i < folio_batch_count(&fbatch); i++) {
+			struct folio *folio = fbatch.folios[i];
+			u64 folio_start = folio_sector(folio);
+			u64 folio_end = folio_end_sector(folio);
+			unsigned folio_offset = max(start, folio_start) - folio_start;
+			unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
+			struct bch_folio *s;
+
+			BUG_ON(end <= folio_start);
+
+			folio_lock(folio);
+			s = bch2_folio(folio);
+
+			if (s) {
+				spin_lock(&s->lock);
+				for (j = folio_offset; j < folio_offset + folio_len; j++) {
+					i_sectors_delta -= s->s[j].state == SECTOR_dirty;
+					bch2_folio_sector_set(folio, s, j, folio_sector_reserve(s->s[j].state));
+				}
+				spin_unlock(&s->lock);
+			}
+
+			folio_unlock(folio);
+		}
+		folio_batch_release(&fbatch);
+		cond_resched();
+	}
+
+	bch2_i_sectors_acct(c, inode, NULL, i_sectors_delta);
+}
+
+static inline unsigned sectors_to_reserve(struct bch_folio_sector *s,
+					  unsigned nr_replicas)
+{
+	return max(0, (int) nr_replicas -
+		   s->nr_replicas -
+		   s->replicas_reserved);
+}
+
+int bch2_get_folio_disk_reservation(struct bch_fs *c,
+				struct bch_inode_info *inode,
+				struct folio *folio, bool check_enospc)
+{
+	struct bch_folio *s = bch2_folio_create(folio, 0);
+	unsigned nr_replicas = inode_nr_replicas(c, inode);
+	struct disk_reservation disk_res = { 0 };
+	unsigned i, sectors = folio_sectors(folio), disk_res_sectors = 0;
+	int ret;
+
+	if (!s)
+		return -ENOMEM;
+
+	for (i = 0; i < sectors; i++)
+		disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
+
+	if (!disk_res_sectors)
+		return 0;
+
+	ret = bch2_disk_reservation_get(c, &disk_res,
+					disk_res_sectors, 1,
+					!check_enospc
+					? BCH_DISK_RESERVATION_NOFAIL
+					: 0);
+	if (unlikely(ret))
+		return ret;
+
+	for (i = 0; i < sectors; i++)
+		s->s[i].replicas_reserved +=
+			sectors_to_reserve(&s->s[i], nr_replicas);
+
+	return 0;
+}
+
+void bch2_folio_reservation_put(struct bch_fs *c,
+			struct bch_inode_info *inode,
+			struct bch2_folio_reservation *res)
+{
+	bch2_disk_reservation_put(c, &res->disk);
+	bch2_quota_reservation_put(c, inode, &res->quota);
+}
+
+int bch2_folio_reservation_get(struct bch_fs *c,
+			struct bch_inode_info *inode,
+			struct folio *folio,
+			struct bch2_folio_reservation *res,
+			unsigned offset, unsigned len)
+{
+	struct bch_folio *s = bch2_folio_create(folio, 0);
+	unsigned i, disk_sectors = 0, quota_sectors = 0;
+	int ret;
+
+	if (!s)
+		return -ENOMEM;
+
+	BUG_ON(!s->uptodate);
+
+	for (i = round_down(offset, block_bytes(c)) >> 9;
+	     i < round_up(offset + len, block_bytes(c)) >> 9;
+	     i++) {
+		disk_sectors += sectors_to_reserve(&s->s[i],
+						res->disk.nr_replicas);
+		quota_sectors += s->s[i].state == SECTOR_unallocated;
+	}
+
+	if (disk_sectors) {
+		ret = bch2_disk_reservation_add(c, &res->disk, disk_sectors, 0);
+		if (unlikely(ret))
+			return ret;
+	}
+
+	if (quota_sectors) {
+		ret = bch2_quota_reservation_add(c, inode, &res->quota,
+						 quota_sectors, true);
+		if (unlikely(ret)) {
+			struct disk_reservation tmp = {
+				.sectors = disk_sectors
+			};
+
+			bch2_disk_reservation_put(c, &tmp);
+			res->disk.sectors -= disk_sectors;
+			return ret;
+		}
+	}
+
+	return 0;
+}
+
+static void bch2_clear_folio_bits(struct folio *folio)
+{
+	struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	struct bch_folio *s = bch2_folio(folio);
+	struct disk_reservation disk_res = { 0 };
+	int i, sectors = folio_sectors(folio), dirty_sectors = 0;
+
+	if (!s)
+		return;
+
+	EBUG_ON(!folio_test_locked(folio));
+	EBUG_ON(folio_test_writeback(folio));
+
+	for (i = 0; i < sectors; i++) {
+		disk_res.sectors += s->s[i].replicas_reserved;
+		s->s[i].replicas_reserved = 0;
+
+		dirty_sectors -= s->s[i].state == SECTOR_dirty;
+		bch2_folio_sector_set(folio, s, i, folio_sector_undirty(s->s[i].state));
+	}
+
+	bch2_disk_reservation_put(c, &disk_res);
+
+	bch2_i_sectors_acct(c, inode, NULL, dirty_sectors);
+
+	bch2_folio_release(folio);
+}
+
+void bch2_set_folio_dirty(struct bch_fs *c,
+			  struct bch_inode_info *inode,
+			  struct folio *folio,
+			  struct bch2_folio_reservation *res,
+			  unsigned offset, unsigned len)
+{
+	struct bch_folio *s = bch2_folio(folio);
+	unsigned i, dirty_sectors = 0;
+
+	WARN_ON((u64) folio_pos(folio) + offset + len >
+		round_up((u64) i_size_read(&inode->v), block_bytes(c)));
+
+	BUG_ON(!s->uptodate);
+
+	spin_lock(&s->lock);
+
+	for (i = round_down(offset, block_bytes(c)) >> 9;
+	     i < round_up(offset + len, block_bytes(c)) >> 9;
+	     i++) {
+		unsigned sectors = sectors_to_reserve(&s->s[i],
+						res->disk.nr_replicas);
+
+		/*
+		 * This can happen if we race with the error path in
+		 * bch2_writepage_io_done():
+		 */
+		sectors = min_t(unsigned, sectors, res->disk.sectors);
+
+		s->s[i].replicas_reserved += sectors;
+		res->disk.sectors -= sectors;
+
+		dirty_sectors += s->s[i].state == SECTOR_unallocated;
+
+		bch2_folio_sector_set(folio, s, i, folio_sector_dirty(s->s[i].state));
+	}
+
+	spin_unlock(&s->lock);
+
+	bch2_i_sectors_acct(c, inode, &res->quota, dirty_sectors);
+
+	if (!folio_test_dirty(folio))
+		filemap_dirty_folio(inode->v.i_mapping, folio);
+}
+
+vm_fault_t bch2_page_fault(struct vm_fault *vmf)
+{
+	struct file *file = vmf->vma->vm_file;
+	struct address_space *mapping = file->f_mapping;
+	struct address_space *fdm = faults_disabled_mapping();
+	struct bch_inode_info *inode = file_bch_inode(file);
+	vm_fault_t ret;
+
+	if (fdm == mapping)
+		return VM_FAULT_SIGBUS;
+
+	/* Lock ordering: */
+	if (fdm > mapping) {
+		struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
+
+		if (bch2_pagecache_add_tryget(inode))
+			goto got_lock;
+
+		bch2_pagecache_block_put(fdm_host);
+
+		bch2_pagecache_add_get(inode);
+		bch2_pagecache_add_put(inode);
+
+		bch2_pagecache_block_get(fdm_host);
+
+		/* Signal that lock has been dropped: */
+		set_fdm_dropped_locks();
+		return VM_FAULT_SIGBUS;
+	}
+
+	bch2_pagecache_add_get(inode);
+got_lock:
+	ret = filemap_fault(vmf);
+	bch2_pagecache_add_put(inode);
+
+	return ret;
+}
+
+vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
+{
+	struct folio *folio = page_folio(vmf->page);
+	struct file *file = vmf->vma->vm_file;
+	struct bch_inode_info *inode = file_bch_inode(file);
+	struct address_space *mapping = file->f_mapping;
+	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	struct bch2_folio_reservation res;
+	unsigned len;
+	loff_t isize;
+	vm_fault_t ret;
+
+	bch2_folio_reservation_init(c, inode, &res);
+
+	sb_start_pagefault(inode->v.i_sb);
+	file_update_time(file);
+
+	/*
+	 * Not strictly necessary, but helps avoid dio writes livelocking in
+	 * bch2_write_invalidate_inode_pages_range() - can drop this if/when we get
+	 * a bch2_write_invalidate_inode_pages_range() that works without dropping
+	 * page lock before invalidating page
+	 */
+	bch2_pagecache_add_get(inode);
+
+	folio_lock(folio);
+	isize = i_size_read(&inode->v);
+
+	if (folio->mapping != mapping || folio_pos(folio) >= isize) {
+		folio_unlock(folio);
+		ret = VM_FAULT_NOPAGE;
+		goto out;
+	}
+
+	len = min_t(loff_t, folio_size(folio), isize - folio_pos(folio));
+
+	if (bch2_folio_set(c, inode_inum(inode), &folio, 1) ?:
+	    bch2_folio_reservation_get(c, inode, folio, &res, 0, len)) {
+		folio_unlock(folio);
+		ret = VM_FAULT_SIGBUS;
+		goto out;
+	}
+
+	bch2_set_folio_dirty(c, inode, folio, &res, 0, len);
+	bch2_folio_reservation_put(c, inode, &res);
+
+	folio_wait_stable(folio);
+	ret = VM_FAULT_LOCKED;
+out:
+	bch2_pagecache_add_put(inode);
+	sb_end_pagefault(inode->v.i_sb);
+
+	return ret;
+}
+
+void bch2_invalidate_folio(struct folio *folio, size_t offset, size_t length)
+{
+	if (offset || length < folio_size(folio))
+		return;
+
+	bch2_clear_folio_bits(folio);
+}
+
+bool bch2_release_folio(struct folio *folio, gfp_t gfp_mask)
+{
+	if (folio_test_dirty(folio) || folio_test_writeback(folio))
+		return false;
+
+	bch2_clear_folio_bits(folio);
+	return true;
+}
+
+/* fseek: */
+
+static int folio_data_offset(struct folio *folio, loff_t pos,
+			     unsigned min_replicas)
+{
+	struct bch_folio *s = bch2_folio(folio);
+	unsigned i, sectors = folio_sectors(folio);
+
+	if (s)
+		for (i = folio_pos_to_s(folio, pos); i < sectors; i++)
+			if (s->s[i].state >= SECTOR_dirty &&
+			    s->s[i].nr_replicas + s->s[i].replicas_reserved >= min_replicas)
+				return i << SECTOR_SHIFT;
+
+	return -1;
+}
+
+loff_t bch2_seek_pagecache_data(struct inode *vinode,
+				loff_t start_offset,
+				loff_t end_offset,
+				unsigned min_replicas,
+				bool nonblock)
+{
+	struct folio_batch fbatch;
+	pgoff_t start_index	= start_offset >> PAGE_SHIFT;
+	pgoff_t end_index	= end_offset >> PAGE_SHIFT;
+	pgoff_t index		= start_index;
+	unsigned i;
+	loff_t ret;
+	int offset;
+
+	folio_batch_init(&fbatch);
+
+	while (filemap_get_folios(vinode->i_mapping,
+				  &index, end_index, &fbatch)) {
+		for (i = 0; i < folio_batch_count(&fbatch); i++) {
+			struct folio *folio = fbatch.folios[i];
+
+			if (!nonblock) {
+				folio_lock(folio);
+			} else if (!folio_trylock(folio)) {
+				folio_batch_release(&fbatch);
+				return -EAGAIN;
+			}
+
+			offset = folio_data_offset(folio,
+					max(folio_pos(folio), start_offset),
+					min_replicas);
+			if (offset >= 0) {
+				ret = clamp(folio_pos(folio) + offset,
+					    start_offset, end_offset);
+				folio_unlock(folio);
+				folio_batch_release(&fbatch);
+				return ret;
+			}
+			folio_unlock(folio);
+		}
+		folio_batch_release(&fbatch);
+		cond_resched();
+	}
+
+	return end_offset;
+}
+
+static int folio_hole_offset(struct address_space *mapping, loff_t *offset,
+			      unsigned min_replicas, bool nonblock)
+{
+	struct folio *folio;
+	struct bch_folio *s;
+	unsigned i, sectors;
+	bool ret = true;
+
+	folio = __filemap_get_folio(mapping, *offset >> PAGE_SHIFT,
+				    FGP_LOCK|(nonblock ? FGP_NOWAIT : 0), 0);
+	if (folio == ERR_PTR(-EAGAIN))
+		return -EAGAIN;
+	if (IS_ERR_OR_NULL(folio))
+		return true;
+
+	s = bch2_folio(folio);
+	if (!s)
+		goto unlock;
+
+	sectors = folio_sectors(folio);
+	for (i = folio_pos_to_s(folio, *offset); i < sectors; i++)
+		if (s->s[i].state < SECTOR_dirty ||
+		    s->s[i].nr_replicas + s->s[i].replicas_reserved < min_replicas) {
+			*offset = max(*offset,
+				      folio_pos(folio) + (i << SECTOR_SHIFT));
+			goto unlock;
+		}
+
+	*offset = folio_end_pos(folio);
+	ret = false;
+unlock:
+	folio_unlock(folio);
+	folio_put(folio);
+	return ret;
+}
+
+loff_t bch2_seek_pagecache_hole(struct inode *vinode,
+				loff_t start_offset,
+				loff_t end_offset,
+				unsigned min_replicas,
+				bool nonblock)
+{
+	struct address_space *mapping = vinode->i_mapping;
+	loff_t offset = start_offset;
+
+	while (offset < end_offset &&
+	       !folio_hole_offset(mapping, &offset, min_replicas, nonblock))
+		;
+
+	return min(offset, end_offset);
+}
+
+int bch2_clamp_data_hole(struct inode *inode,
+			 u64 *hole_start,
+			 u64 *hole_end,
+			 unsigned min_replicas,
+			 bool nonblock)
+{
+	loff_t ret;
+
+	ret = bch2_seek_pagecache_hole(inode,
+		*hole_start << 9, *hole_end << 9, min_replicas, nonblock) >> 9;
+	if (ret < 0)
+		return ret;
+
+	*hole_start = ret;
+
+	if (*hole_start == *hole_end)
+		return 0;
+
+	ret = bch2_seek_pagecache_data(inode,
+		*hole_start << 9, *hole_end << 9, min_replicas, nonblock) >> 9;
+	if (ret < 0)
+		return ret;
+
+	*hole_end = ret;
+	return 0;
+}
+
+#endif /* NO_BCACHEFS_FS */
diff --git a/libbcachefs/fs-io-pagecache.h b/libbcachefs/fs-io-pagecache.h
new file mode 100644
index 00000000..f1c747cb
--- /dev/null
+++ b/libbcachefs/fs-io-pagecache.h
@@ -0,0 +1,175 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_FS_IO_PAGECACHE_H
+#define _BCACHEFS_FS_IO_PAGECACHE_H
+
+#include <linux/pagemap.h>
+
+typedef DARRAY(struct folio *) folios;
+
+int bch2_filemap_get_contig_folios_d(struct address_space *, loff_t,
+				     u64, int, gfp_t, folios *);
+int bch2_write_invalidate_inode_pages_range(struct address_space *, loff_t, loff_t);
+
+/*
+ * Use u64 for the end pos and sector helpers because if the folio covers the
+ * max supported range of the mapping, the start offset of the next folio
+ * overflows loff_t. This breaks much of the range based processing in the
+ * buffered write path.
+ */
+static inline u64 folio_end_pos(struct folio *folio)
+{
+	return folio_pos(folio) + folio_size(folio);
+}
+
+static inline size_t folio_sectors(struct folio *folio)
+{
+	return PAGE_SECTORS << folio_order(folio);
+}
+
+static inline loff_t folio_sector(struct folio *folio)
+{
+	return folio_pos(folio) >> 9;
+}
+
+static inline u64 folio_end_sector(struct folio *folio)
+{
+	return folio_end_pos(folio) >> 9;
+}
+
+#define BCH_FOLIO_SECTOR_STATE()	\
+	x(unallocated)			\
+	x(reserved)			\
+	x(dirty)			\
+	x(dirty_reserved)		\
+	x(allocated)
+
+enum bch_folio_sector_state {
+#define x(n)	SECTOR_##n,
+	BCH_FOLIO_SECTOR_STATE()
+#undef x
+};
+
+struct bch_folio_sector {
+	/* Uncompressed, fully allocated replicas (or on disk reservation): */
+	unsigned		nr_replicas:4;
+
+	/* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */
+	unsigned		replicas_reserved:4;
+
+	/* i_sectors: */
+	enum bch_folio_sector_state state:8;
+};
+
+struct bch_folio {
+	spinlock_t		lock;
+	atomic_t		write_count;
+	/*
+	 * Is the sector state up to date with the btree?
+	 * (Not the data itself)
+	 */
+	bool			uptodate;
+	struct bch_folio_sector	s[];
+};
+
+/* Helper for when we need to add debug instrumentation: */
+static inline void bch2_folio_sector_set(struct folio *folio,
+			     struct bch_folio *s,
+			     unsigned i, unsigned n)
+{
+	s->s[i].state = n;
+}
+
+/* file offset (to folio offset) to bch_folio_sector index */
+static inline int folio_pos_to_s(struct folio *folio, loff_t pos)
+{
+	u64 f_offset = pos - folio_pos(folio);
+	BUG_ON(pos < folio_pos(folio) || pos >= folio_end_pos(folio));
+	return f_offset >> SECTOR_SHIFT;
+}
+
+/* for newly allocated folios: */
+static inline void __bch2_folio_release(struct folio *folio)
+{
+	kfree(folio_detach_private(folio));
+}
+
+static inline void bch2_folio_release(struct folio *folio)
+{
+	EBUG_ON(!folio_test_locked(folio));
+	__bch2_folio_release(folio);
+}
+
+static inline struct bch_folio *__bch2_folio(struct folio *folio)
+{
+	return folio_has_private(folio)
+		? (struct bch_folio *) folio_get_private(folio)
+		: NULL;
+}
+
+static inline struct bch_folio *bch2_folio(struct folio *folio)
+{
+	EBUG_ON(!folio_test_locked(folio));
+
+	return __bch2_folio(folio);
+}
+
+struct bch_folio *__bch2_folio_create(struct folio *, gfp_t);
+struct bch_folio *bch2_folio_create(struct folio *, gfp_t);
+
+struct bch2_folio_reservation {
+	struct disk_reservation	disk;
+	struct quota_res	quota;
+};
+
+static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
+{
+	/* XXX: this should not be open coded */
+	return inode->ei_inode.bi_data_replicas
+		? inode->ei_inode.bi_data_replicas - 1
+		: c->opts.data_replicas;
+}
+
+static inline void bch2_folio_reservation_init(struct bch_fs *c,
+			struct bch_inode_info *inode,
+			struct bch2_folio_reservation *res)
+{
+	memset(res, 0, sizeof(*res));
+
+	res->disk.nr_replicas = inode_nr_replicas(c, inode);
+}
+
+int bch2_folio_set(struct bch_fs *, subvol_inum, struct folio **, unsigned);
+void bch2_bio_page_state_set(struct bio *, struct bkey_s_c);
+
+void bch2_mark_pagecache_unallocated(struct bch_inode_info *, u64, u64);
+void bch2_mark_pagecache_reserved(struct bch_inode_info *, u64, u64);
+
+int bch2_get_folio_disk_reservation(struct bch_fs *,
+				struct bch_inode_info *,
+				struct folio *, bool);
+
+void bch2_folio_reservation_put(struct bch_fs *,
+			struct bch_inode_info *,
+			struct bch2_folio_reservation *);
+int bch2_folio_reservation_get(struct bch_fs *,
+			struct bch_inode_info *,
+			struct folio *,
+			struct bch2_folio_reservation *,
+			unsigned, unsigned);
+
+void bch2_set_folio_dirty(struct bch_fs *,
+			  struct bch_inode_info *,
+			  struct folio *,
+			  struct bch2_folio_reservation *,
+			  unsigned, unsigned);
+
+vm_fault_t bch2_page_fault(struct vm_fault *);
+vm_fault_t bch2_page_mkwrite(struct vm_fault *);
+void bch2_invalidate_folio(struct folio *, size_t, size_t);
+bool bch2_release_folio(struct folio *, gfp_t);
+
+loff_t bch2_seek_pagecache_data(struct inode *, loff_t, loff_t, unsigned, bool);
+loff_t bch2_seek_pagecache_hole(struct inode *, loff_t, loff_t, unsigned, bool);
+int bch2_clamp_data_hole(struct inode *, u64 *, u64 *, unsigned, bool);
+
+#endif /* _BCACHEFS_FS_IO_PAGECACHE_H */
diff --git a/libbcachefs/fs-io.c b/libbcachefs/fs-io.c
index d433f4d5..e846d1eb 100644
--- a/libbcachefs/fs-io.c
+++ b/libbcachefs/fs-io.c
@@ -3,7 +3,7 @@
 
 #include "bcachefs.h"
 #include "alloc_foreground.h"
-#include "bkey_buf.h"
+//#include "bkey_buf.h"
 #include "btree_update.h"
 #include "buckets.h"
 #include "clock.h"
@@ -12,6 +12,9 @@
 #include "extent_update.h"
 #include "fs.h"
 #include "fs-io.h"
+#include "fs-io-buffered.h"
+//#include "fs-io-direct.h"
+#include "fs-io-pagecache.h"
 #include "fsck.h"
 #include "inode.h"
 #include "journal.h"
@@ -31,116 +34,9 @@
 #include <linux/sched/signal.h>
 #include <linux/task_io_accounting_ops.h>
 #include <linux/uio.h>
-#include <linux/writeback.h>
 
 #include <trace/events/writeback.h>
 
-static int bch2_clamp_data_hole(struct inode *, u64 *, u64 *, unsigned, bool);
-
-struct folio_vec {
-	struct folio	*fv_folio;
-	size_t		fv_offset;
-	size_t		fv_len;
-};
-
-static inline struct folio_vec biovec_to_foliovec(struct bio_vec bv)
-{
-
-	struct folio *folio	= page_folio(bv.bv_page);
-	size_t offset		= (folio_page_idx(folio, bv.bv_page) << PAGE_SHIFT) +
-		bv.bv_offset;
-	size_t len = min_t(size_t, folio_size(folio) - offset, bv.bv_len);
-
-	return (struct folio_vec) {
-		.fv_folio	= folio,
-		.fv_offset	= offset,
-		.fv_len		= len,
-	};
-}
-
-static inline struct folio_vec bio_iter_iovec_folio(struct bio *bio,
-						    struct bvec_iter iter)
-{
-	return biovec_to_foliovec(bio_iter_iovec(bio, iter));
-}
-
-#define __bio_for_each_folio(bvl, bio, iter, start)			\
-	for (iter = (start);						\
-	     (iter).bi_size &&						\
-		((bvl = bio_iter_iovec_folio((bio), (iter))), 1);	\
-	     bio_advance_iter_single((bio), &(iter), (bvl).fv_len))
-
-/**
- * bio_for_each_folio - iterate over folios within a bio
- *
- * Like other non-_all versions, this iterates over what bio->bi_iter currently
- * points to. This version is for drivers, where the bio may have previously
- * been split or cloned.
- */
-#define bio_for_each_folio(bvl, bio, iter)				\
-	__bio_for_each_folio(bvl, bio, iter, (bio)->bi_iter)
-
-/*
- * Use u64 for the end pos and sector helpers because if the folio covers the
- * max supported range of the mapping, the start offset of the next folio
- * overflows loff_t. This breaks much of the range based processing in the
- * buffered write path.
- */
-static inline u64 folio_end_pos(struct folio *folio)
-{
-	return folio_pos(folio) + folio_size(folio);
-}
-
-static inline size_t folio_sectors(struct folio *folio)
-{
-	return PAGE_SECTORS << folio_order(folio);
-}
-
-static inline loff_t folio_sector(struct folio *folio)
-{
-	return folio_pos(folio) >> 9;
-}
-
-static inline u64 folio_end_sector(struct folio *folio)
-{
-	return folio_end_pos(folio) >> 9;
-}
-
-typedef DARRAY(struct folio *) folios;
-
-static int filemap_get_contig_folios_d(struct address_space *mapping,
-				       loff_t start, u64 end,
-				       int fgp_flags, gfp_t gfp,
-				       folios *folios)
-{
-	struct folio *f;
-	u64 pos = start;
-	int ret = 0;
-
-	while (pos < end) {
-		if ((u64) pos >= (u64) start + (1ULL << 20))
-			fgp_flags &= ~FGP_CREAT;
-
-		ret = darray_make_room_gfp(folios, 1, gfp & GFP_KERNEL);
-		if (ret)
-			break;
-
-		f = __filemap_get_folio(mapping, pos >> PAGE_SHIFT, fgp_flags, gfp);
-		if (IS_ERR_OR_NULL(f))
-			break;
-
-		BUG_ON(folios->nr && folio_pos(f) != pos);
-
-		pos = folio_end_pos(f);
-		darray_push(folios, f);
-	}
-
-	if (!folios->nr && !ret && (fgp_flags & FGP_CREAT))
-		ret = -ENOMEM;
-
-	return folios->nr ? 0 : ret;
-}
-
 struct nocow_flush {
 	struct closure	*cl;
 	struct bch_dev	*ca;
@@ -157,9 +53,9 @@ static void nocow_flush_endio(struct bio *_bio)
 	bio_put(&bio->bio);
 }
 
-static void bch2_inode_flush_nocow_writes_async(struct bch_fs *c,
-						struct bch_inode_info *inode,
-						struct closure *cl)
+void bch2_inode_flush_nocow_writes_async(struct bch_fs *c,
+					 struct bch_inode_info *inode,
+					 struct closure *cl)
 {
 	struct nocow_flush *bio;
 	struct bch_dev *ca;
@@ -207,170 +103,6 @@ static int bch2_inode_flush_nocow_writes(struct bch_fs *c,
 	return 0;
 }
 
-static inline bool bio_full(struct bio *bio, unsigned len)
-{
-	if (bio->bi_vcnt >= bio->bi_max_vecs)
-		return true;
-	if (bio->bi_iter.bi_size > UINT_MAX - len)
-		return true;
-	return false;
-}
-
-static inline struct address_space *faults_disabled_mapping(void)
-{
-	return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
-}
-
-static inline void set_fdm_dropped_locks(void)
-{
-	current->faults_disabled_mapping =
-		(void *) (((unsigned long) current->faults_disabled_mapping)|1);
-}
-
-static inline bool fdm_dropped_locks(void)
-{
-	return ((unsigned long) current->faults_disabled_mapping) & 1;
-}
-
-struct quota_res {
-	u64				sectors;
-};
-
-struct bch_writepage_io {
-	struct bch_inode_info		*inode;
-
-	/* must be last: */
-	struct bch_write_op		op;
-};
-
-struct dio_write {
-	struct kiocb			*req;
-	struct address_space		*mapping;
-	struct bch_inode_info		*inode;
-	struct mm_struct		*mm;
-	unsigned			loop:1,
-					extending:1,
-					sync:1,
-					flush:1,
-					free_iov:1;
-	struct quota_res		quota_res;
-	u64				written;
-
-	struct iov_iter			iter;
-	struct iovec			inline_vecs[2];
-
-	/* must be last: */
-	struct bch_write_op		op;
-};
-
-struct dio_read {
-	struct closure			cl;
-	struct kiocb			*req;
-	long				ret;
-	bool				should_dirty;
-	struct bch_read_bio		rbio;
-};
-
-/* pagecache_block must be held */
-static noinline int write_invalidate_inode_pages_range(struct address_space *mapping,
-					      loff_t start, loff_t end)
-{
-	int ret;
-
-	/*
-	 * XXX: the way this is currently implemented, we can spin if a process
-	 * is continually redirtying a specific page
-	 */
-	do {
-		if (!mapping->nrpages)
-			return 0;
-
-		ret = filemap_write_and_wait_range(mapping, start, end);
-		if (ret)
-			break;
-
-		if (!mapping->nrpages)
-			return 0;
-
-		ret = invalidate_inode_pages2_range(mapping,
-				start >> PAGE_SHIFT,
-				end >> PAGE_SHIFT);
-	} while (ret == -EBUSY);
-
-	return ret;
-}
-
-/* quotas */
-
-#ifdef CONFIG_BCACHEFS_QUOTA
-
-static void __bch2_quota_reservation_put(struct bch_fs *c,
-					 struct bch_inode_info *inode,
-					 struct quota_res *res)
-{
-	BUG_ON(res->sectors > inode->ei_quota_reserved);
-
-	bch2_quota_acct(c, inode->ei_qid, Q_SPC,
-			-((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
-	inode->ei_quota_reserved -= res->sectors;
-	res->sectors = 0;
-}
-
-static void bch2_quota_reservation_put(struct bch_fs *c,
-				       struct bch_inode_info *inode,
-				       struct quota_res *res)
-{
-	if (res->sectors) {
-		mutex_lock(&inode->ei_quota_lock);
-		__bch2_quota_reservation_put(c, inode, res);
-		mutex_unlock(&inode->ei_quota_lock);
-	}
-}
-
-static int bch2_quota_reservation_add(struct bch_fs *c,
-				      struct bch_inode_info *inode,
-				      struct quota_res *res,
-				      u64 sectors,
-				      bool check_enospc)
-{
-	int ret;
-
-	if (test_bit(EI_INODE_SNAPSHOT, &inode->ei_flags))
-		return 0;
-
-	mutex_lock(&inode->ei_quota_lock);
-	ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
-			      check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
-	if (likely(!ret)) {
-		inode->ei_quota_reserved += sectors;
-		res->sectors += sectors;
-	}
-	mutex_unlock(&inode->ei_quota_lock);
-
-	return ret;
-}
-
-#else
-
-static void __bch2_quota_reservation_put(struct bch_fs *c,
-					 struct bch_inode_info *inode,
-					 struct quota_res *res) {}
-
-static void bch2_quota_reservation_put(struct bch_fs *c,
-				       struct bch_inode_info *inode,
-				       struct quota_res *res) {}
-
-static int bch2_quota_reservation_add(struct bch_fs *c,
-				      struct bch_inode_info *inode,
-				      struct quota_res *res,
-				      unsigned sectors,
-				      bool check_enospc)
-{
-	return 0;
-}
-
-#endif
-
 /* i_size updates: */
 
 struct inode_new_size {
@@ -409,7 +141,7 @@ int __must_check bch2_write_inode_size(struct bch_fs *c,
 	return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 }
 
-static void __i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
+void __bch2_i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 			   struct quota_res *quota_res, s64 sectors)
 {
 	bch2_fs_inconsistent_on((s64) inode->v.i_blocks + sectors < 0, c,
@@ -433,2344 +165,6 @@ static void __i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 #endif
 }
 
-static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
-			   struct quota_res *quota_res, s64 sectors)
-{
-	if (sectors) {
-		mutex_lock(&inode->ei_quota_lock);
-		__i_sectors_acct(c, inode, quota_res, sectors);
-		mutex_unlock(&inode->ei_quota_lock);
-	}
-}
-
-/* page state: */
-
-/* stored in page->private: */
-
-#define BCH_FOLIO_SECTOR_STATE()	\
-	x(unallocated)			\
-	x(reserved)			\
-	x(dirty)			\
-	x(dirty_reserved)		\
-	x(allocated)
-
-enum bch_folio_sector_state {
-#define x(n)	SECTOR_##n,
-	BCH_FOLIO_SECTOR_STATE()
-#undef x
-};
-
-static const char * const bch2_folio_sector_states[] = {
-#define x(n)	#n,
-	BCH_FOLIO_SECTOR_STATE()
-#undef x
-	NULL
-};
-
-static inline enum bch_folio_sector_state
-folio_sector_dirty(enum bch_folio_sector_state state)
-{
-	switch (state) {
-	case SECTOR_unallocated:
-		return SECTOR_dirty;
-	case SECTOR_reserved:
-		return SECTOR_dirty_reserved;
-	default:
-		return state;
-	}
-}
-
-static inline enum bch_folio_sector_state
-folio_sector_undirty(enum bch_folio_sector_state state)
-{
-	switch (state) {
-	case SECTOR_dirty:
-		return SECTOR_unallocated;
-	case SECTOR_dirty_reserved:
-		return SECTOR_reserved;
-	default:
-		return state;
-	}
-}
-
-static inline enum bch_folio_sector_state
-folio_sector_reserve(enum bch_folio_sector_state state)
-{
-	switch (state) {
-	case SECTOR_unallocated:
-		return SECTOR_reserved;
-	case SECTOR_dirty:
-		return SECTOR_dirty_reserved;
-	default:
-		return state;
-	}
-}
-
-struct bch_folio_sector {
-	/* Uncompressed, fully allocated replicas (or on disk reservation): */
-	unsigned		nr_replicas:4;
-
-	/* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */
-	unsigned		replicas_reserved:4;
-
-	/* i_sectors: */
-	enum bch_folio_sector_state state:8;
-};
-
-struct bch_folio {
-	spinlock_t		lock;
-	atomic_t		write_count;
-	/*
-	 * Is the sector state up to date with the btree?
-	 * (Not the data itself)
-	 */
-	bool			uptodate;
-	struct bch_folio_sector	s[];
-};
-
-static inline void folio_sector_set(struct folio *folio,
-			     struct bch_folio *s,
-			     unsigned i, unsigned n)
-{
-	s->s[i].state = n;
-}
-
-/* file offset (to folio offset) to bch_folio_sector index */
-static inline int folio_pos_to_s(struct folio *folio, loff_t pos)
-{
-	u64 f_offset = pos - folio_pos(folio);
-	BUG_ON(pos < folio_pos(folio) || pos >= folio_end_pos(folio));
-	return f_offset >> SECTOR_SHIFT;
-}
-
-static inline struct bch_folio *__bch2_folio(struct folio *folio)
-{
-	return folio_has_private(folio)
-		? (struct bch_folio *) folio_get_private(folio)
-		: NULL;
-}
-
-static inline struct bch_folio *bch2_folio(struct folio *folio)
-{
-	EBUG_ON(!folio_test_locked(folio));
-
-	return __bch2_folio(folio);
-}
-
-/* for newly allocated folios: */
-static void __bch2_folio_release(struct folio *folio)
-{
-	kfree(folio_detach_private(folio));
-}
-
-static void bch2_folio_release(struct folio *folio)
-{
-	EBUG_ON(!folio_test_locked(folio));
-	__bch2_folio_release(folio);
-}
-
-/* for newly allocated folios: */
-static struct bch_folio *__bch2_folio_create(struct folio *folio, gfp_t gfp)
-{
-	struct bch_folio *s;
-
-	s = kzalloc(sizeof(*s) +
-		    sizeof(struct bch_folio_sector) *
-		    folio_sectors(folio), gfp);
-	if (!s)
-		return NULL;
-
-	spin_lock_init(&s->lock);
-	folio_attach_private(folio, s);
-	return s;
-}
-
-static struct bch_folio *bch2_folio_create(struct folio *folio, gfp_t gfp)
-{
-	return bch2_folio(folio) ?: __bch2_folio_create(folio, gfp);
-}
-
-static unsigned bkey_to_sector_state(struct bkey_s_c k)
-{
-	if (bkey_extent_is_reservation(k))
-		return SECTOR_reserved;
-	if (bkey_extent_is_allocation(k.k))
-		return SECTOR_allocated;
-	return SECTOR_unallocated;
-}
-
-static void __bch2_folio_set(struct folio *folio,
-			     unsigned pg_offset, unsigned pg_len,
-			     unsigned nr_ptrs, unsigned state)
-{
-	struct bch_folio *s = bch2_folio(folio);
-	unsigned i, sectors = folio_sectors(folio);
-
-	BUG_ON(pg_offset >= sectors);
-	BUG_ON(pg_offset + pg_len > sectors);
-
-	spin_lock(&s->lock);
-
-	for (i = pg_offset; i < pg_offset + pg_len; i++) {
-		s->s[i].nr_replicas	= nr_ptrs;
-		folio_sector_set(folio, s, i, state);
-	}
-
-	if (i == sectors)
-		s->uptodate = true;
-
-	spin_unlock(&s->lock);
-}
-
-/*
- * Initialize bch_folio state (allocated/unallocated, nr_replicas) from the
- * extents btree:
- */
-static int bch2_folio_set(struct bch_fs *c, subvol_inum inum,
-			  struct folio **folios, unsigned nr_folios)
-{
-	struct btree_trans trans;
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	struct bch_folio *s;
-	u64 offset = folio_sector(folios[0]);
-	unsigned folio_idx;
-	u32 snapshot;
-	bool need_set = false;
-	int ret;
-
-	for (folio_idx = 0; folio_idx < nr_folios; folio_idx++) {
-		s = bch2_folio_create(folios[folio_idx], GFP_KERNEL);
-		if (!s)
-			return -ENOMEM;
-
-		need_set |= !s->uptodate;
-	}
-
-	if (!need_set)
-		return 0;
-
-	folio_idx = 0;
-	bch2_trans_init(&trans, c, 0, 0);
-retry:
-	bch2_trans_begin(&trans);
-
-	ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
-	if (ret)
-		goto err;
-
-	for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
-			   SPOS(inum.inum, offset, snapshot),
-			   BTREE_ITER_SLOTS, k, ret) {
-		unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k);
-		unsigned state = bkey_to_sector_state(k);
-
-		while (folio_idx < nr_folios) {
-			struct folio *folio = folios[folio_idx];
-			u64 folio_start	= folio_sector(folio);
-			u64 folio_end	= folio_end_sector(folio);
-			unsigned folio_offset = max(bkey_start_offset(k.k), folio_start) - folio_start;
-			unsigned folio_len = min(k.k->p.offset, folio_end) - folio_offset - folio_start;
-
-			BUG_ON(k.k->p.offset < folio_start);
-			BUG_ON(bkey_start_offset(k.k) > folio_end);
-
-			if (!bch2_folio(folio)->uptodate)
-				__bch2_folio_set(folio, folio_offset, folio_len, nr_ptrs, state);
-
-			if (k.k->p.offset < folio_end)
-				break;
-			folio_idx++;
-		}
-
-		if (folio_idx == nr_folios)
-			break;
-	}
-
-	offset = iter.pos.offset;
-	bch2_trans_iter_exit(&trans, &iter);
-err:
-	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
-		goto retry;
-	bch2_trans_exit(&trans);
-
-	return ret;
-}
-
-static void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k)
-{
-	struct bvec_iter iter;
-	struct folio_vec fv;
-	unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
-		? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
-	unsigned state = bkey_to_sector_state(k);
-
-	bio_for_each_folio(fv, bio, iter)
-		__bch2_folio_set(fv.fv_folio,
-				 fv.fv_offset >> 9,
-				 fv.fv_len >> 9,
-				 nr_ptrs, state);
-}
-
-static void mark_pagecache_unallocated(struct bch_inode_info *inode,
-				       u64 start, u64 end)
-{
-	pgoff_t index = start >> PAGE_SECTORS_SHIFT;
-	pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
-	struct folio_batch fbatch;
-	unsigned i, j;
-
-	if (end <= start)
-		return;
-
-	folio_batch_init(&fbatch);
-
-	while (filemap_get_folios(inode->v.i_mapping,
-				  &index, end_index, &fbatch)) {
-		for (i = 0; i < folio_batch_count(&fbatch); i++) {
-			struct folio *folio = fbatch.folios[i];
-			u64 folio_start = folio_sector(folio);
-			u64 folio_end = folio_end_sector(folio);
-			unsigned folio_offset = max(start, folio_start) - folio_start;
-			unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
-			struct bch_folio *s;
-
-			BUG_ON(end <= folio_start);
-
-			folio_lock(folio);
-			s = bch2_folio(folio);
-
-			if (s) {
-				spin_lock(&s->lock);
-				for (j = folio_offset; j < folio_offset + folio_len; j++)
-					s->s[j].nr_replicas = 0;
-				spin_unlock(&s->lock);
-			}
-
-			folio_unlock(folio);
-		}
-		folio_batch_release(&fbatch);
-		cond_resched();
-	}
-}
-
-static void mark_pagecache_reserved(struct bch_inode_info *inode,
-				    u64 start, u64 end)
-{
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	pgoff_t index = start >> PAGE_SECTORS_SHIFT;
-	pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
-	struct folio_batch fbatch;
-	s64 i_sectors_delta = 0;
-	unsigned i, j;
-
-	if (end <= start)
-		return;
-
-	folio_batch_init(&fbatch);
-
-	while (filemap_get_folios(inode->v.i_mapping,
-				  &index, end_index, &fbatch)) {
-		for (i = 0; i < folio_batch_count(&fbatch); i++) {
-			struct folio *folio = fbatch.folios[i];
-			u64 folio_start = folio_sector(folio);
-			u64 folio_end = folio_end_sector(folio);
-			unsigned folio_offset = max(start, folio_start) - folio_start;
-			unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
-			struct bch_folio *s;
-
-			BUG_ON(end <= folio_start);
-
-			folio_lock(folio);
-			s = bch2_folio(folio);
-
-			if (s) {
-				spin_lock(&s->lock);
-				for (j = folio_offset; j < folio_offset + folio_len; j++) {
-					i_sectors_delta -= s->s[j].state == SECTOR_dirty;
-					folio_sector_set(folio, s, j, folio_sector_reserve(s->s[j].state));
-				}
-				spin_unlock(&s->lock);
-			}
-
-			folio_unlock(folio);
-		}
-		folio_batch_release(&fbatch);
-		cond_resched();
-	}
-
-	i_sectors_acct(c, inode, NULL, i_sectors_delta);
-}
-
-static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
-{
-	/* XXX: this should not be open coded */
-	return inode->ei_inode.bi_data_replicas
-		? inode->ei_inode.bi_data_replicas - 1
-		: c->opts.data_replicas;
-}
-
-static inline unsigned sectors_to_reserve(struct bch_folio_sector *s,
-					  unsigned nr_replicas)
-{
-	return max(0, (int) nr_replicas -
-		   s->nr_replicas -
-		   s->replicas_reserved);
-}
-
-static int bch2_get_folio_disk_reservation(struct bch_fs *c,
-				struct bch_inode_info *inode,
-				struct folio *folio, bool check_enospc)
-{
-	struct bch_folio *s = bch2_folio_create(folio, 0);
-	unsigned nr_replicas = inode_nr_replicas(c, inode);
-	struct disk_reservation disk_res = { 0 };
-	unsigned i, sectors = folio_sectors(folio), disk_res_sectors = 0;
-	int ret;
-
-	if (!s)
-		return -ENOMEM;
-
-	for (i = 0; i < sectors; i++)
-		disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
-
-	if (!disk_res_sectors)
-		return 0;
-
-	ret = bch2_disk_reservation_get(c, &disk_res,
-					disk_res_sectors, 1,
-					!check_enospc
-					? BCH_DISK_RESERVATION_NOFAIL
-					: 0);
-	if (unlikely(ret))
-		return ret;
-
-	for (i = 0; i < sectors; i++)
-		s->s[i].replicas_reserved +=
-			sectors_to_reserve(&s->s[i], nr_replicas);
-
-	return 0;
-}
-
-struct bch2_folio_reservation {
-	struct disk_reservation	disk;
-	struct quota_res	quota;
-};
-
-static void bch2_folio_reservation_init(struct bch_fs *c,
-			struct bch_inode_info *inode,
-			struct bch2_folio_reservation *res)
-{
-	memset(res, 0, sizeof(*res));
-
-	res->disk.nr_replicas = inode_nr_replicas(c, inode);
-}
-
-static void bch2_folio_reservation_put(struct bch_fs *c,
-			struct bch_inode_info *inode,
-			struct bch2_folio_reservation *res)
-{
-	bch2_disk_reservation_put(c, &res->disk);
-	bch2_quota_reservation_put(c, inode, &res->quota);
-}
-
-static int bch2_folio_reservation_get(struct bch_fs *c,
-			struct bch_inode_info *inode,
-			struct folio *folio,
-			struct bch2_folio_reservation *res,
-			unsigned offset, unsigned len)
-{
-	struct bch_folio *s = bch2_folio_create(folio, 0);
-	unsigned i, disk_sectors = 0, quota_sectors = 0;
-	int ret;
-
-	if (!s)
-		return -ENOMEM;
-
-	BUG_ON(!s->uptodate);
-
-	for (i = round_down(offset, block_bytes(c)) >> 9;
-	     i < round_up(offset + len, block_bytes(c)) >> 9;
-	     i++) {
-		disk_sectors += sectors_to_reserve(&s->s[i],
-						res->disk.nr_replicas);
-		quota_sectors += s->s[i].state == SECTOR_unallocated;
-	}
-
-	if (disk_sectors) {
-		ret = bch2_disk_reservation_add(c, &res->disk, disk_sectors, 0);
-		if (unlikely(ret))
-			return ret;
-	}
-
-	if (quota_sectors) {
-		ret = bch2_quota_reservation_add(c, inode, &res->quota,
-						 quota_sectors, true);
-		if (unlikely(ret)) {
-			struct disk_reservation tmp = {
-				.sectors = disk_sectors
-			};
-
-			bch2_disk_reservation_put(c, &tmp);
-			res->disk.sectors -= disk_sectors;
-			return ret;
-		}
-	}
-
-	return 0;
-}
-
-static void bch2_clear_folio_bits(struct folio *folio)
-{
-	struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	struct bch_folio *s = bch2_folio(folio);
-	struct disk_reservation disk_res = { 0 };
-	int i, sectors = folio_sectors(folio), dirty_sectors = 0;
-
-	if (!s)
-		return;
-
-	EBUG_ON(!folio_test_locked(folio));
-	EBUG_ON(folio_test_writeback(folio));
-
-	for (i = 0; i < sectors; i++) {
-		disk_res.sectors += s->s[i].replicas_reserved;
-		s->s[i].replicas_reserved = 0;
-
-		dirty_sectors -= s->s[i].state == SECTOR_dirty;
-		folio_sector_set(folio, s, i, folio_sector_undirty(s->s[i].state));
-	}
-
-	bch2_disk_reservation_put(c, &disk_res);
-
-	i_sectors_acct(c, inode, NULL, dirty_sectors);
-
-	bch2_folio_release(folio);
-}
-
-static void bch2_set_folio_dirty(struct bch_fs *c,
-			struct bch_inode_info *inode,
-			struct folio *folio,
-			struct bch2_folio_reservation *res,
-			unsigned offset, unsigned len)
-{
-	struct bch_folio *s = bch2_folio(folio);
-	unsigned i, dirty_sectors = 0;
-
-	WARN_ON((u64) folio_pos(folio) + offset + len >
-		round_up((u64) i_size_read(&inode->v), block_bytes(c)));
-
-	BUG_ON(!s->uptodate);
-
-	spin_lock(&s->lock);
-
-	for (i = round_down(offset, block_bytes(c)) >> 9;
-	     i < round_up(offset + len, block_bytes(c)) >> 9;
-	     i++) {
-		unsigned sectors = sectors_to_reserve(&s->s[i],
-						res->disk.nr_replicas);
-
-		/*
-		 * This can happen if we race with the error path in
-		 * bch2_writepage_io_done():
-		 */
-		sectors = min_t(unsigned, sectors, res->disk.sectors);
-
-		s->s[i].replicas_reserved += sectors;
-		res->disk.sectors -= sectors;
-
-		dirty_sectors += s->s[i].state == SECTOR_unallocated;
-
-		folio_sector_set(folio, s, i, folio_sector_dirty(s->s[i].state));
-	}
-
-	spin_unlock(&s->lock);
-
-	i_sectors_acct(c, inode, &res->quota, dirty_sectors);
-
-	if (!folio_test_dirty(folio))
-		filemap_dirty_folio(inode->v.i_mapping, folio);
-}
-
-vm_fault_t bch2_page_fault(struct vm_fault *vmf)
-{
-	struct file *file = vmf->vma->vm_file;
-	struct address_space *mapping = file->f_mapping;
-	struct address_space *fdm = faults_disabled_mapping();
-	struct bch_inode_info *inode = file_bch_inode(file);
-	vm_fault_t ret;
-
-	if (fdm == mapping)
-		return VM_FAULT_SIGBUS;
-
-	/* Lock ordering: */
-	if (fdm > mapping) {
-		struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
-
-		if (bch2_pagecache_add_tryget(inode))
-			goto got_lock;
-
-		bch2_pagecache_block_put(fdm_host);
-
-		bch2_pagecache_add_get(inode);
-		bch2_pagecache_add_put(inode);
-
-		bch2_pagecache_block_get(fdm_host);
-
-		/* Signal that lock has been dropped: */
-		set_fdm_dropped_locks();
-		return VM_FAULT_SIGBUS;
-	}
-
-	bch2_pagecache_add_get(inode);
-got_lock:
-	ret = filemap_fault(vmf);
-	bch2_pagecache_add_put(inode);
-
-	return ret;
-}
-
-vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
-{
-	struct folio *folio = page_folio(vmf->page);
-	struct file *file = vmf->vma->vm_file;
-	struct bch_inode_info *inode = file_bch_inode(file);
-	struct address_space *mapping = file->f_mapping;
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	struct bch2_folio_reservation res;
-	unsigned len;
-	loff_t isize;
-	vm_fault_t ret;
-
-	bch2_folio_reservation_init(c, inode, &res);
-
-	sb_start_pagefault(inode->v.i_sb);
-	file_update_time(file);
-
-	/*
-	 * Not strictly necessary, but helps avoid dio writes livelocking in
-	 * write_invalidate_inode_pages_range() - can drop this if/when we get
-	 * a write_invalidate_inode_pages_range() that works without dropping
-	 * page lock before invalidating page
-	 */
-	bch2_pagecache_add_get(inode);
-
-	folio_lock(folio);
-	isize = i_size_read(&inode->v);
-
-	if (folio->mapping != mapping || folio_pos(folio) >= isize) {
-		folio_unlock(folio);
-		ret = VM_FAULT_NOPAGE;
-		goto out;
-	}
-
-	len = min_t(loff_t, folio_size(folio), isize - folio_pos(folio));
-
-	if (bch2_folio_set(c, inode_inum(inode), &folio, 1) ?:
-	    bch2_folio_reservation_get(c, inode, folio, &res, 0, len)) {
-		folio_unlock(folio);
-		ret = VM_FAULT_SIGBUS;
-		goto out;
-	}
-
-	bch2_set_folio_dirty(c, inode, folio, &res, 0, len);
-	bch2_folio_reservation_put(c, inode, &res);
-
-	folio_wait_stable(folio);
-	ret = VM_FAULT_LOCKED;
-out:
-	bch2_pagecache_add_put(inode);
-	sb_end_pagefault(inode->v.i_sb);
-
-	return ret;
-}
-
-void bch2_invalidate_folio(struct folio *folio, size_t offset, size_t length)
-{
-	if (offset || length < folio_size(folio))
-		return;
-
-	bch2_clear_folio_bits(folio);
-}
-
-bool bch2_release_folio(struct folio *folio, gfp_t gfp_mask)
-{
-	if (folio_test_dirty(folio) || folio_test_writeback(folio))
-		return false;
-
-	bch2_clear_folio_bits(folio);
-	return true;
-}
-
-/* readpage(s): */
-
-static void bch2_readpages_end_io(struct bio *bio)
-{
-	struct folio_iter fi;
-
-	bio_for_each_folio_all(fi, bio) {
-		if (!bio->bi_status) {
-			folio_mark_uptodate(fi.folio);
-		} else {
-			folio_clear_uptodate(fi.folio);
-			folio_set_error(fi.folio);
-		}
-		folio_unlock(fi.folio);
-	}
-
-	bio_put(bio);
-}
-
-struct readpages_iter {
-	struct address_space	*mapping;
-	unsigned		idx;
-	folios			folios;
-};
-
-static int readpages_iter_init(struct readpages_iter *iter,
-			       struct readahead_control *ractl)
-{
-	struct folio **fi;
-	int ret;
-
-	memset(iter, 0, sizeof(*iter));
-
-	iter->mapping = ractl->mapping;
-
-	ret = filemap_get_contig_folios_d(iter->mapping,
-				ractl->_index << PAGE_SHIFT,
-				(ractl->_index + ractl->_nr_pages) << PAGE_SHIFT,
-				0, mapping_gfp_mask(iter->mapping),
-				&iter->folios);
-	if (ret)
-		return ret;
-
-	darray_for_each(iter->folios, fi) {
-		ractl->_nr_pages -= 1U << folio_order(*fi);
-		__bch2_folio_create(*fi, __GFP_NOFAIL|GFP_KERNEL);
-		folio_put(*fi);
-		folio_put(*fi);
-	}
-
-	return 0;
-}
-
-static inline struct folio *readpage_iter_peek(struct readpages_iter *iter)
-{
-	if (iter->idx >= iter->folios.nr)
-		return NULL;
-	return iter->folios.data[iter->idx];
-}
-
-static inline void readpage_iter_advance(struct readpages_iter *iter)
-{
-	iter->idx++;
-}
-
-static bool extent_partial_reads_expensive(struct bkey_s_c k)
-{
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	struct bch_extent_crc_unpacked crc;
-	const union bch_extent_entry *i;
-
-	bkey_for_each_crc(k.k, ptrs, crc, i)
-		if (crc.csum_type || crc.compression_type)
-			return true;
-	return false;
-}
-
-static int readpage_bio_extend(struct btree_trans *trans,
-			       struct readpages_iter *iter,
-			       struct bio *bio,
-			       unsigned sectors_this_extent,
-			       bool get_more)
-{
-	/* Don't hold btree locks while allocating memory: */
-	bch2_trans_unlock(trans);
-
-	while (bio_sectors(bio) < sectors_this_extent &&
-	       bio->bi_vcnt < bio->bi_max_vecs) {
-		struct folio *folio = readpage_iter_peek(iter);
-		int ret;
-
-		if (folio) {
-			readpage_iter_advance(iter);
-		} else {
-			pgoff_t folio_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
-
-			if (!get_more)
-				break;
-
-			folio = xa_load(&iter->mapping->i_pages, folio_offset);
-			if (folio && !xa_is_value(folio))
-				break;
-
-			folio = filemap_alloc_folio(readahead_gfp_mask(iter->mapping), 0);
-			if (!folio)
-				break;
-
-			if (!__bch2_folio_create(folio, GFP_KERNEL)) {
-				folio_put(folio);
-				break;
-			}
-
-			ret = filemap_add_folio(iter->mapping, folio, folio_offset, GFP_KERNEL);
-			if (ret) {
-				__bch2_folio_release(folio);
-				folio_put(folio);
-				break;
-			}
-
-			folio_put(folio);
-		}
-
-		BUG_ON(folio_sector(folio) != bio_end_sector(bio));
-
-		BUG_ON(!bio_add_folio(bio, folio, folio_size(folio), 0));
-	}
-
-	return bch2_trans_relock(trans);
-}
-
-static void bchfs_read(struct btree_trans *trans,
-		       struct bch_read_bio *rbio,
-		       subvol_inum inum,
-		       struct readpages_iter *readpages_iter)
-{
-	struct bch_fs *c = trans->c;
-	struct btree_iter iter;
-	struct bkey_buf sk;
-	int flags = BCH_READ_RETRY_IF_STALE|
-		BCH_READ_MAY_PROMOTE;
-	u32 snapshot;
-	int ret = 0;
-
-	rbio->c = c;
-	rbio->start_time = local_clock();
-	rbio->subvol = inum.subvol;
-
-	bch2_bkey_buf_init(&sk);
-retry:
-	bch2_trans_begin(trans);
-	iter = (struct btree_iter) { NULL };
-
-	ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
-	if (ret)
-		goto err;
-
-	bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
-			     SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
-			     BTREE_ITER_SLOTS);
-	while (1) {
-		struct bkey_s_c k;
-		unsigned bytes, sectors, offset_into_extent;
-		enum btree_id data_btree = BTREE_ID_extents;
-
-		/*
-		 * read_extent -> io_time_reset may cause a transaction restart
-		 * without returning an error, we need to check for that here:
-		 */
-		ret = bch2_trans_relock(trans);
-		if (ret)
-			break;
-
-		bch2_btree_iter_set_pos(&iter,
-				POS(inum.inum, rbio->bio.bi_iter.bi_sector));
-
-		k = bch2_btree_iter_peek_slot(&iter);
-		ret = bkey_err(k);
-		if (ret)
-			break;
-
-		offset_into_extent = iter.pos.offset -
-			bkey_start_offset(k.k);
-		sectors = k.k->size - offset_into_extent;
-
-		bch2_bkey_buf_reassemble(&sk, c, k);
-
-		ret = bch2_read_indirect_extent(trans, &data_btree,
-					&offset_into_extent, &sk);
-		if (ret)
-			break;
-
-		k = bkey_i_to_s_c(sk.k);
-
-		sectors = min(sectors, k.k->size - offset_into_extent);
-
-		if (readpages_iter) {
-			ret = readpage_bio_extend(trans, readpages_iter, &rbio->bio, sectors,
-						  extent_partial_reads_expensive(k));
-			if (ret)
-				break;
-		}
-
-		bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
-		swap(rbio->bio.bi_iter.bi_size, bytes);
-
-		if (rbio->bio.bi_iter.bi_size == bytes)
-			flags |= BCH_READ_LAST_FRAGMENT;
-
-		bch2_bio_page_state_set(&rbio->bio, k);
-
-		bch2_read_extent(trans, rbio, iter.pos,
-				 data_btree, k, offset_into_extent, flags);
-
-		if (flags & BCH_READ_LAST_FRAGMENT)
-			break;
-
-		swap(rbio->bio.bi_iter.bi_size, bytes);
-		bio_advance(&rbio->bio, bytes);
-
-		ret = btree_trans_too_many_iters(trans);
-		if (ret)
-			break;
-	}
-err:
-	bch2_trans_iter_exit(trans, &iter);
-
-	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
-		goto retry;
-
-	if (ret) {
-		bch_err_inum_offset_ratelimited(c,
-				iter.pos.inode,
-				iter.pos.offset << 9,
-				"read error %i from btree lookup", ret);
-		rbio->bio.bi_status = BLK_STS_IOERR;
-		bio_endio(&rbio->bio);
-	}
-
-	bch2_bkey_buf_exit(&sk, c);
-}
-
-void bch2_readahead(struct readahead_control *ractl)
-{
-	struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	struct bch_io_opts opts;
-	struct btree_trans trans;
-	struct folio *folio;
-	struct readpages_iter readpages_iter;
-	int ret;
-
-	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
-
-	ret = readpages_iter_init(&readpages_iter, ractl);
-	BUG_ON(ret);
-
-	bch2_trans_init(&trans, c, 0, 0);
-
-	bch2_pagecache_add_get(inode);
-
-	while ((folio = readpage_iter_peek(&readpages_iter))) {
-		unsigned n = min_t(unsigned,
-				   readpages_iter.folios.nr -
-				   readpages_iter.idx,
-				   BIO_MAX_VECS);
-		struct bch_read_bio *rbio =
-			rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
-						   GFP_KERNEL, &c->bio_read),
-				  opts);
-
-		readpage_iter_advance(&readpages_iter);
-
-		rbio->bio.bi_iter.bi_sector = folio_sector(folio);
-		rbio->bio.bi_end_io = bch2_readpages_end_io;
-		BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
-
-		bchfs_read(&trans, rbio, inode_inum(inode),
-			   &readpages_iter);
-		bch2_trans_unlock(&trans);
-	}
-
-	bch2_pagecache_add_put(inode);
-
-	bch2_trans_exit(&trans);
-	darray_exit(&readpages_iter.folios);
-}
-
-static void __bchfs_readfolio(struct bch_fs *c, struct bch_read_bio *rbio,
-			     subvol_inum inum, struct folio *folio)
-{
-	struct btree_trans trans;
-
-	bch2_folio_create(folio, __GFP_NOFAIL);
-
-	rbio->bio.bi_opf = REQ_OP_READ|REQ_SYNC;
-	rbio->bio.bi_iter.bi_sector = folio_sector(folio);
-	BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
-
-	bch2_trans_init(&trans, c, 0, 0);
-	bchfs_read(&trans, rbio, inum, NULL);
-	bch2_trans_exit(&trans);
-}
-
-static void bch2_read_single_folio_end_io(struct bio *bio)
-{
-	complete(bio->bi_private);
-}
-
-static int bch2_read_single_folio(struct folio *folio,
-				  struct address_space *mapping)
-{
-	struct bch_inode_info *inode = to_bch_ei(mapping->host);
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	struct bch_read_bio *rbio;
-	struct bch_io_opts opts;
-	int ret;
-	DECLARE_COMPLETION_ONSTACK(done);
-
-	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
-
-	rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_KERNEL, &c->bio_read),
-			 opts);
-	rbio->bio.bi_private = &done;
-	rbio->bio.bi_end_io = bch2_read_single_folio_end_io;
-
-	__bchfs_readfolio(c, rbio, inode_inum(inode), folio);
-	wait_for_completion(&done);
-
-	ret = blk_status_to_errno(rbio->bio.bi_status);
-	bio_put(&rbio->bio);
-
-	if (ret < 0)
-		return ret;
-
-	folio_mark_uptodate(folio);
-	return 0;
-}
-
-int bch2_read_folio(struct file *file, struct folio *folio)
-{
-	int ret;
-
-	ret = bch2_read_single_folio(folio, folio->mapping);
-	folio_unlock(folio);
-	return bch2_err_class(ret);
-}
-
-/* writepages: */
-
-struct bch_writepage_state {
-	struct bch_writepage_io	*io;
-	struct bch_io_opts	opts;
-	struct bch_folio_sector	*tmp;
-	unsigned		tmp_sectors;
-};
-
-static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
-								  struct bch_inode_info *inode)
-{
-	struct bch_writepage_state ret = { 0 };
-
-	bch2_inode_opts_get(&ret.opts, c, &inode->ei_inode);
-	return ret;
-}
-
-static void bch2_writepage_io_done(struct bch_write_op *op)
-{
-	struct bch_writepage_io *io =
-		container_of(op, struct bch_writepage_io, op);
-	struct bch_fs *c = io->op.c;
-	struct bio *bio = &io->op.wbio.bio;
-	struct folio_iter fi;
-	unsigned i;
-
-	if (io->op.error) {
-		set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
-
-		bio_for_each_folio_all(fi, bio) {
-			struct bch_folio *s;
-
-			folio_set_error(fi.folio);
-			mapping_set_error(fi.folio->mapping, -EIO);
-
-			s = __bch2_folio(fi.folio);
-			spin_lock(&s->lock);
-			for (i = 0; i < folio_sectors(fi.folio); i++)
-				s->s[i].nr_replicas = 0;
-			spin_unlock(&s->lock);
-		}
-	}
-
-	if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
-		bio_for_each_folio_all(fi, bio) {
-			struct bch_folio *s;
-
-			s = __bch2_folio(fi.folio);
-			spin_lock(&s->lock);
-			for (i = 0; i < folio_sectors(fi.folio); i++)
-				s->s[i].nr_replicas = 0;
-			spin_unlock(&s->lock);
-		}
-	}
-
-	/*
-	 * racing with fallocate can cause us to add fewer sectors than
-	 * expected - but we shouldn't add more sectors than expected:
-	 */
-	WARN_ON_ONCE(io->op.i_sectors_delta > 0);
-
-	/*
-	 * (error (due to going RO) halfway through a page can screw that up
-	 * slightly)
-	 * XXX wtf?
-	   BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
-	 */
-
-	/*
-	 * PageWriteback is effectively our ref on the inode - fixup i_blocks
-	 * before calling end_page_writeback:
-	 */
-	i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
-
-	bio_for_each_folio_all(fi, bio) {
-		struct bch_folio *s = __bch2_folio(fi.folio);
-
-		if (atomic_dec_and_test(&s->write_count))
-			folio_end_writeback(fi.folio);
-	}
-
-	bio_put(&io->op.wbio.bio);
-}
-
-static void bch2_writepage_do_io(struct bch_writepage_state *w)
-{
-	struct bch_writepage_io *io = w->io;
-
-	w->io = NULL;
-	closure_call(&io->op.cl, bch2_write, NULL, NULL);
-}
-
-/*
- * Get a bch_writepage_io and add @page to it - appending to an existing one if
- * possible, else allocating a new one:
- */
-static void bch2_writepage_io_alloc(struct bch_fs *c,
-				    struct writeback_control *wbc,
-				    struct bch_writepage_state *w,
-				    struct bch_inode_info *inode,
-				    u64 sector,
-				    unsigned nr_replicas)
-{
-	struct bch_write_op *op;
-
-	w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
-					      REQ_OP_WRITE,
-					      GFP_KERNEL,
-					      &c->writepage_bioset),
-			     struct bch_writepage_io, op.wbio.bio);
-
-	w->io->inode		= inode;
-	op			= &w->io->op;
-	bch2_write_op_init(op, c, w->opts);
-	op->target		= w->opts.foreground_target;
-	op->nr_replicas		= nr_replicas;
-	op->res.nr_replicas	= nr_replicas;
-	op->write_point		= writepoint_hashed(inode->ei_last_dirtied);
-	op->subvol		= inode->ei_subvol;
-	op->pos			= POS(inode->v.i_ino, sector);
-	op->end_io		= bch2_writepage_io_done;
-	op->devs_need_flush	= &inode->ei_devs_need_flush;
-	op->wbio.bio.bi_iter.bi_sector = sector;
-	op->wbio.bio.bi_opf	= wbc_to_write_flags(wbc);
-}
-
-static int __bch2_writepage(struct folio *folio,
-			    struct writeback_control *wbc,
-			    void *data)
-{
-	struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	struct bch_writepage_state *w = data;
-	struct bch_folio *s;
-	unsigned i, offset, f_sectors, nr_replicas_this_write = U32_MAX;
-	loff_t i_size = i_size_read(&inode->v);
-	int ret;
-
-	EBUG_ON(!folio_test_uptodate(folio));
-
-	/* Is the folio fully inside i_size? */
-	if (folio_end_pos(folio) <= i_size)
-		goto do_io;
-
-	/* Is the folio fully outside i_size? (truncate in progress) */
-	if (folio_pos(folio) >= i_size) {
-		folio_unlock(folio);
-		return 0;
-	}
-
-	/*
-	 * The folio straddles i_size.  It must be zeroed out on each and every
-	 * writepage invocation because it may be mmapped.  "A file is mapped
-	 * in multiples of the folio size.  For a file that is not a multiple of
-	 * the  folio size, the remaining memory is zeroed when mapped, and
-	 * writes to that region are not written out to the file."
-	 */
-	folio_zero_segment(folio,
-			   i_size - folio_pos(folio),
-			   folio_size(folio));
-do_io:
-	f_sectors = folio_sectors(folio);
-	s = bch2_folio(folio);
-
-	if (f_sectors > w->tmp_sectors) {
-		kfree(w->tmp);
-		w->tmp = kzalloc(sizeof(struct bch_folio_sector) *
-				 f_sectors, __GFP_NOFAIL);
-		w->tmp_sectors = f_sectors;
-	}
-
-	/*
-	 * Things get really hairy with errors during writeback:
-	 */
-	ret = bch2_get_folio_disk_reservation(c, inode, folio, false);
-	BUG_ON(ret);
-
-	/* Before unlocking the page, get copy of reservations: */
-	spin_lock(&s->lock);
-	memcpy(w->tmp, s->s, sizeof(struct bch_folio_sector) * f_sectors);
-
-	for (i = 0; i < f_sectors; i++) {
-		if (s->s[i].state < SECTOR_dirty)
-			continue;
-
-		nr_replicas_this_write =
-			min_t(unsigned, nr_replicas_this_write,
-			      s->s[i].nr_replicas +
-			      s->s[i].replicas_reserved);
-	}
-
-	for (i = 0; i < f_sectors; i++) {
-		if (s->s[i].state < SECTOR_dirty)
-			continue;
-
-		s->s[i].nr_replicas = w->opts.compression
-			? 0 : nr_replicas_this_write;
-
-		s->s[i].replicas_reserved = 0;
-		folio_sector_set(folio, s, i, SECTOR_allocated);
-	}
-	spin_unlock(&s->lock);
-
-	BUG_ON(atomic_read(&s->write_count));
-	atomic_set(&s->write_count, 1);
-
-	BUG_ON(folio_test_writeback(folio));
-	folio_start_writeback(folio);
-
-	folio_unlock(folio);
-
-	offset = 0;
-	while (1) {
-		unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
-		u64 sector;
-
-		while (offset < f_sectors &&
-		       w->tmp[offset].state < SECTOR_dirty)
-			offset++;
-
-		if (offset == f_sectors)
-			break;
-
-		while (offset + sectors < f_sectors &&
-		       w->tmp[offset + sectors].state >= SECTOR_dirty) {
-			reserved_sectors += w->tmp[offset + sectors].replicas_reserved;
-			dirty_sectors += w->tmp[offset + sectors].state == SECTOR_dirty;
-			sectors++;
-		}
-		BUG_ON(!sectors);
-
-		sector = folio_sector(folio) + offset;
-
-		if (w->io &&
-		    (w->io->op.res.nr_replicas != nr_replicas_this_write ||
-		     bio_full(&w->io->op.wbio.bio, sectors << 9) ||
-		     w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
-		     (BIO_MAX_VECS * PAGE_SIZE) ||
-		     bio_end_sector(&w->io->op.wbio.bio) != sector))
-			bch2_writepage_do_io(w);
-
-		if (!w->io)
-			bch2_writepage_io_alloc(c, wbc, w, inode, sector,
-						nr_replicas_this_write);
-
-		atomic_inc(&s->write_count);
-
-		BUG_ON(inode != w->io->inode);
-		BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio,
-				     sectors << 9, offset << 9));
-
-		/* Check for writing past i_size: */
-		WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
-			  round_up(i_size, block_bytes(c)) &&
-			  !test_bit(BCH_FS_EMERGENCY_RO, &c->flags),
-			  "writing past i_size: %llu > %llu (unrounded %llu)\n",
-			  bio_end_sector(&w->io->op.wbio.bio) << 9,
-			  round_up(i_size, block_bytes(c)),
-			  i_size);
-
-		w->io->op.res.sectors += reserved_sectors;
-		w->io->op.i_sectors_delta -= dirty_sectors;
-		w->io->op.new_i_size = i_size;
-
-		offset += sectors;
-	}
-
-	if (atomic_dec_and_test(&s->write_count))
-		folio_end_writeback(folio);
-
-	return 0;
-}
-
-int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
-{
-	struct bch_fs *c = mapping->host->i_sb->s_fs_info;
-	struct bch_writepage_state w =
-		bch_writepage_state_init(c, to_bch_ei(mapping->host));
-	struct blk_plug plug;
-	int ret;
-
-	blk_start_plug(&plug);
-	ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
-	if (w.io)
-		bch2_writepage_do_io(&w);
-	blk_finish_plug(&plug);
-	kfree(w.tmp);
-	return bch2_err_class(ret);
-}
-
-/* buffered writes: */
-
-int bch2_write_begin(struct file *file, struct address_space *mapping,
-		     loff_t pos, unsigned len,
-		     struct page **pagep, void **fsdata)
-{
-	struct bch_inode_info *inode = to_bch_ei(mapping->host);
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	struct bch2_folio_reservation *res;
-	struct folio *folio;
-	unsigned offset;
-	int ret = -ENOMEM;
-
-	res = kmalloc(sizeof(*res), GFP_KERNEL);
-	if (!res)
-		return -ENOMEM;
-
-	bch2_folio_reservation_init(c, inode, res);
-	*fsdata = res;
-
-	bch2_pagecache_add_get(inode);
-
-	folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
-				FGP_LOCK|FGP_WRITE|FGP_CREAT|FGP_STABLE,
-				mapping_gfp_mask(mapping));
-	if (IS_ERR_OR_NULL(folio))
-		goto err_unlock;
-
-	if (folio_test_uptodate(folio))
-		goto out;
-
-	offset = pos - folio_pos(folio);
-	len = min_t(size_t, len, folio_end_pos(folio) - pos);
-
-	/* If we're writing entire folio, don't need to read it in first: */
-	if (!offset && len == folio_size(folio))
-		goto out;
-
-	if (!offset && pos + len >= inode->v.i_size) {
-		folio_zero_segment(folio, len, folio_size(folio));
-		flush_dcache_folio(folio);
-		goto out;
-	}
-
-	if (folio_pos(folio) >= inode->v.i_size) {
-		folio_zero_segments(folio, 0, offset, offset + len, folio_size(folio));
-		flush_dcache_folio(folio);
-		goto out;
-	}
-readpage:
-	ret = bch2_read_single_folio(folio, mapping);
-	if (ret)
-		goto err;
-out:
-	ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
-	if (ret)
-		goto err;
-
-	ret = bch2_folio_reservation_get(c, inode, folio, res, offset, len);
-	if (ret) {
-		if (!folio_test_uptodate(folio)) {
-			/*
-			 * If the folio hasn't been read in, we won't know if we
-			 * actually need a reservation - we don't actually need
-			 * to read here, we just need to check if the folio is
-			 * fully backed by uncompressed data:
-			 */
-			goto readpage;
-		}
-
-		goto err;
-	}
-
-	*pagep = &folio->page;
-	return 0;
-err:
-	folio_unlock(folio);
-	folio_put(folio);
-	*pagep = NULL;
-err_unlock:
-	bch2_pagecache_add_put(inode);
-	kfree(res);
-	*fsdata = NULL;
-	return bch2_err_class(ret);
-}
-
-int bch2_write_end(struct file *file, struct address_space *mapping,
-		   loff_t pos, unsigned len, unsigned copied,
-		   struct page *page, void *fsdata)
-{
-	struct bch_inode_info *inode = to_bch_ei(mapping->host);
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	struct bch2_folio_reservation *res = fsdata;
-	struct folio *folio = page_folio(page);
-	unsigned offset = pos - folio_pos(folio);
-
-	lockdep_assert_held(&inode->v.i_rwsem);
-	BUG_ON(offset + copied > folio_size(folio));
-
-	if (unlikely(copied < len && !folio_test_uptodate(folio))) {
-		/*
-		 * The folio needs to be read in, but that would destroy
-		 * our partial write - simplest thing is to just force
-		 * userspace to redo the write:
-		 */
-		folio_zero_range(folio, 0, folio_size(folio));
-		flush_dcache_folio(folio);
-		copied = 0;
-	}
-
-	spin_lock(&inode->v.i_lock);
-	if (pos + copied > inode->v.i_size)
-		i_size_write(&inode->v, pos + copied);
-	spin_unlock(&inode->v.i_lock);
-
-	if (copied) {
-		if (!folio_test_uptodate(folio))
-			folio_mark_uptodate(folio);
-
-		bch2_set_folio_dirty(c, inode, folio, res, offset, copied);
-
-		inode->ei_last_dirtied = (unsigned long) current;
-	}
-
-	folio_unlock(folio);
-	folio_put(folio);
-	bch2_pagecache_add_put(inode);
-
-	bch2_folio_reservation_put(c, inode, res);
-	kfree(res);
-
-	return copied;
-}
-
-static noinline void folios_trunc(folios *folios, struct folio **fi)
-{
-	while (folios->data + folios->nr > fi) {
-		struct folio *f = darray_pop(folios);
-
-		folio_unlock(f);
-		folio_put(f);
-	}
-}
-
-static int __bch2_buffered_write(struct bch_inode_info *inode,
-				 struct address_space *mapping,
-				 struct iov_iter *iter,
-				 loff_t pos, unsigned len)
-{
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	struct bch2_folio_reservation res;
-	folios folios;
-	struct folio **fi, *f;
-	unsigned copied = 0, f_offset;
-	u64 end = pos + len, f_pos;
-	loff_t last_folio_pos = inode->v.i_size;
-	int ret = 0;
-
-	BUG_ON(!len);
-
-	bch2_folio_reservation_init(c, inode, &res);
-	darray_init(&folios);
-
-	ret = filemap_get_contig_folios_d(mapping, pos, end,
-				   FGP_LOCK|FGP_WRITE|FGP_STABLE|FGP_CREAT,
-				   mapping_gfp_mask(mapping),
-				   &folios);
-	if (ret)
-		goto out;
-
-	BUG_ON(!folios.nr);
-
-	f = darray_first(folios);
-	if (pos != folio_pos(f) && !folio_test_uptodate(f)) {
-		ret = bch2_read_single_folio(f, mapping);
-		if (ret)
-			goto out;
-	}
-
-	f = darray_last(folios);
-	end = min(end, folio_end_pos(f));
-	last_folio_pos = folio_pos(f);
-	if (end != folio_end_pos(f) && !folio_test_uptodate(f)) {
-		if (end >= inode->v.i_size) {
-			folio_zero_range(f, 0, folio_size(f));
-		} else {
-			ret = bch2_read_single_folio(f, mapping);
-			if (ret)
-				goto out;
-		}
-	}
-
-	ret = bch2_folio_set(c, inode_inum(inode), folios.data, folios.nr);
-	if (ret)
-		goto out;
-
-	f_pos = pos;
-	f_offset = pos - folio_pos(darray_first(folios));
-	darray_for_each(folios, fi) {
-		struct folio *f = *fi;
-		u64 f_len = min(end, folio_end_pos(f)) - f_pos;
-
-		/*
-		 * XXX: per POSIX and fstests generic/275, on -ENOSPC we're
-		 * supposed to write as much as we have disk space for.
-		 *
-		 * On failure here we should still write out a partial page if
-		 * we aren't completely out of disk space - we don't do that
-		 * yet:
-		 */
-		ret = bch2_folio_reservation_get(c, inode, f, &res, f_offset, f_len);
-		if (unlikely(ret)) {
-			folios_trunc(&folios, fi);
-			if (!folios.nr)
-				goto out;
-
-			end = min(end, folio_end_pos(darray_last(folios)));
-			break;
-		}
-
-		f_pos = folio_end_pos(f);
-		f_offset = 0;
-	}
-
-	if (mapping_writably_mapped(mapping))
-		darray_for_each(folios, fi)
-			flush_dcache_folio(*fi);
-
-	f_pos = pos;
-	f_offset = pos - folio_pos(darray_first(folios));
-	darray_for_each(folios, fi) {
-		struct folio *f = *fi;
-		u64 f_len = min(end, folio_end_pos(f)) - f_pos;
-		unsigned f_copied = copy_page_from_iter_atomic(&f->page, f_offset, f_len, iter);
-
-		if (!f_copied) {
-			folios_trunc(&folios, fi);
-			break;
-		}
-
-		if (!folio_test_uptodate(f) &&
-		    f_copied != folio_size(f) &&
-		    pos + copied + f_copied < inode->v.i_size) {
-			folio_zero_range(f, 0, folio_size(f));
-			folios_trunc(&folios, fi);
-			break;
-		}
-
-		flush_dcache_folio(f);
-		copied += f_copied;
-
-		if (f_copied != f_len) {
-			folios_trunc(&folios, fi + 1);
-			break;
-		}
-
-		f_pos = folio_end_pos(f);
-		f_offset = 0;
-	}
-
-	if (!copied)
-		goto out;
-
-	end = pos + copied;
-
-	spin_lock(&inode->v.i_lock);
-	if (end > inode->v.i_size)
-		i_size_write(&inode->v, end);
-	spin_unlock(&inode->v.i_lock);
-
-	f_pos = pos;
-	f_offset = pos - folio_pos(darray_first(folios));
-	darray_for_each(folios, fi) {
-		struct folio *f = *fi;
-		u64 f_len = min(end, folio_end_pos(f)) - f_pos;
-
-		if (!folio_test_uptodate(f))
-			folio_mark_uptodate(f);
-
-		bch2_set_folio_dirty(c, inode, f, &res, f_offset, f_len);
-
-		f_pos = folio_end_pos(f);
-		f_offset = 0;
-	}
-
-	inode->ei_last_dirtied = (unsigned long) current;
-out:
-	darray_for_each(folios, fi) {
-		folio_unlock(*fi);
-		folio_put(*fi);
-	}
-
-	/*
-	 * If the last folio added to the mapping starts beyond current EOF, we
-	 * performed a short write but left around at least one post-EOF folio.
-	 * Clean up the mapping before we return.
-	 */
-	if (last_folio_pos >= inode->v.i_size)
-		truncate_pagecache(&inode->v, inode->v.i_size);
-
-	darray_exit(&folios);
-	bch2_folio_reservation_put(c, inode, &res);
-
-	return copied ?: ret;
-}
-
-static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
-{
-	struct file *file = iocb->ki_filp;
-	struct address_space *mapping = file->f_mapping;
-	struct bch_inode_info *inode = file_bch_inode(file);
-	loff_t pos = iocb->ki_pos;
-	ssize_t written = 0;
-	int ret = 0;
-
-	bch2_pagecache_add_get(inode);
-
-	do {
-		unsigned offset = pos & (PAGE_SIZE - 1);
-		unsigned bytes = iov_iter_count(iter);
-again:
-		/*
-		 * Bring in the user page that we will copy from _first_.
-		 * Otherwise there's a nasty deadlock on copying from the
-		 * same page as we're writing to, without it being marked
-		 * up-to-date.
-		 *
-		 * Not only is this an optimisation, but it is also required
-		 * to check that the address is actually valid, when atomic
-		 * usercopies are used, below.
-		 */
-		if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
-			bytes = min_t(unsigned long, iov_iter_count(iter),
-				      PAGE_SIZE - offset);
-
-			if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
-				ret = -EFAULT;
-				break;
-			}
-		}
-
-		if (unlikely(fatal_signal_pending(current))) {
-			ret = -EINTR;
-			break;
-		}
-
-		ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
-		if (unlikely(ret < 0))
-			break;
-
-		cond_resched();
-
-		if (unlikely(ret == 0)) {
-			/*
-			 * If we were unable to copy any data at all, we must
-			 * fall back to a single segment length write.
-			 *
-			 * If we didn't fallback here, we could livelock
-			 * because not all segments in the iov can be copied at
-			 * once without a pagefault.
-			 */
-			bytes = min_t(unsigned long, PAGE_SIZE - offset,
-				      iov_iter_single_seg_count(iter));
-			goto again;
-		}
-		pos += ret;
-		written += ret;
-		ret = 0;
-
-		balance_dirty_pages_ratelimited(mapping);
-	} while (iov_iter_count(iter));
-
-	bch2_pagecache_add_put(inode);
-
-	return written ? written : ret;
-}
-
-/* O_DIRECT reads */
-
-static void bio_check_or_release(struct bio *bio, bool check_dirty)
-{
-	if (check_dirty) {
-		bio_check_pages_dirty(bio);
-	} else {
-		bio_release_pages(bio, false);
-		bio_put(bio);
-	}
-}
-
-static void bch2_dio_read_complete(struct closure *cl)
-{
-	struct dio_read *dio = container_of(cl, struct dio_read, cl);
-
-	dio->req->ki_complete(dio->req, dio->ret);
-	bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
-}
-
-static void bch2_direct_IO_read_endio(struct bio *bio)
-{
-	struct dio_read *dio = bio->bi_private;
-
-	if (bio->bi_status)
-		dio->ret = blk_status_to_errno(bio->bi_status);
-
-	closure_put(&dio->cl);
-}
-
-static void bch2_direct_IO_read_split_endio(struct bio *bio)
-{
-	struct dio_read *dio = bio->bi_private;
-	bool should_dirty = dio->should_dirty;
-
-	bch2_direct_IO_read_endio(bio);
-	bio_check_or_release(bio, should_dirty);
-}
-
-static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
-{
-	struct file *file = req->ki_filp;
-	struct bch_inode_info *inode = file_bch_inode(file);
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	struct bch_io_opts opts;
-	struct dio_read *dio;
-	struct bio *bio;
-	loff_t offset = req->ki_pos;
-	bool sync = is_sync_kiocb(req);
-	size_t shorten;
-	ssize_t ret;
-
-	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
-
-	if ((offset|iter->count) & (block_bytes(c) - 1))
-		return -EINVAL;
-
-	ret = min_t(loff_t, iter->count,
-		    max_t(loff_t, 0, i_size_read(&inode->v) - offset));
-
-	if (!ret)
-		return ret;
-
-	shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
-	iter->count -= shorten;
-
-	bio = bio_alloc_bioset(NULL,
-			       bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
-			       REQ_OP_READ,
-			       GFP_KERNEL,
-			       &c->dio_read_bioset);
-
-	bio->bi_end_io = bch2_direct_IO_read_endio;
-
-	dio = container_of(bio, struct dio_read, rbio.bio);
-	closure_init(&dio->cl, NULL);
-
-	/*
-	 * this is a _really_ horrible hack just to avoid an atomic sub at the
-	 * end:
-	 */
-	if (!sync) {
-		set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
-		atomic_set(&dio->cl.remaining,
-			   CLOSURE_REMAINING_INITIALIZER -
-			   CLOSURE_RUNNING +
-			   CLOSURE_DESTRUCTOR);
-	} else {
-		atomic_set(&dio->cl.remaining,
-			   CLOSURE_REMAINING_INITIALIZER + 1);
-	}
-
-	dio->req	= req;
-	dio->ret	= ret;
-	/*
-	 * This is one of the sketchier things I've encountered: we have to skip
-	 * the dirtying of requests that are internal from the kernel (i.e. from
-	 * loopback), because we'll deadlock on page_lock.
-	 */
-	dio->should_dirty = iter_is_iovec(iter);
-
-	goto start;
-	while (iter->count) {
-		bio = bio_alloc_bioset(NULL,
-				       bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
-				       REQ_OP_READ,
-				       GFP_KERNEL,
-				       &c->bio_read);
-		bio->bi_end_io		= bch2_direct_IO_read_split_endio;
-start:
-		bio->bi_opf		= REQ_OP_READ|REQ_SYNC;
-		bio->bi_iter.bi_sector	= offset >> 9;
-		bio->bi_private		= dio;
-
-		ret = bio_iov_iter_get_pages(bio, iter);
-		if (ret < 0) {
-			/* XXX: fault inject this path */
-			bio->bi_status = BLK_STS_RESOURCE;
-			bio_endio(bio);
-			break;
-		}
-
-		offset += bio->bi_iter.bi_size;
-
-		if (dio->should_dirty)
-			bio_set_pages_dirty(bio);
-
-		if (iter->count)
-			closure_get(&dio->cl);
-
-		bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
-	}
-
-	iter->count += shorten;
-
-	if (sync) {
-		closure_sync(&dio->cl);
-		closure_debug_destroy(&dio->cl);
-		ret = dio->ret;
-		bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
-		return ret;
-	} else {
-		return -EIOCBQUEUED;
-	}
-}
-
-ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
-{
-	struct file *file = iocb->ki_filp;
-	struct bch_inode_info *inode = file_bch_inode(file);
-	struct address_space *mapping = file->f_mapping;
-	size_t count = iov_iter_count(iter);
-	ssize_t ret;
-
-	if (!count)
-		return 0; /* skip atime */
-
-	if (iocb->ki_flags & IOCB_DIRECT) {
-		struct blk_plug plug;
-
-		if (unlikely(mapping->nrpages)) {
-			ret = filemap_write_and_wait_range(mapping,
-						iocb->ki_pos,
-						iocb->ki_pos + count - 1);
-			if (ret < 0)
-				goto out;
-		}
-
-		file_accessed(file);
-
-		blk_start_plug(&plug);
-		ret = bch2_direct_IO_read(iocb, iter);
-		blk_finish_plug(&plug);
-
-		if (ret >= 0)
-			iocb->ki_pos += ret;
-	} else {
-		bch2_pagecache_add_get(inode);
-		ret = generic_file_read_iter(iocb, iter);
-		bch2_pagecache_add_put(inode);
-	}
-out:
-	return bch2_err_class(ret);
-}
-
-/* O_DIRECT writes */
-
-static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
-				       u64 offset, u64 size,
-				       unsigned nr_replicas, bool compressed)
-{
-	struct btree_trans trans;
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	u64 end = offset + size;
-	u32 snapshot;
-	bool ret = true;
-	int err;
-
-	bch2_trans_init(&trans, c, 0, 0);
-retry:
-	bch2_trans_begin(&trans);
-
-	err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
-	if (err)
-		goto err;
-
-	for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
-			   SPOS(inum.inum, offset, snapshot),
-			   BTREE_ITER_SLOTS, k, err) {
-		if (bkey_ge(bkey_start_pos(k.k), POS(inum.inum, end)))
-			break;
-
-		if (k.k->p.snapshot != snapshot ||
-		    nr_replicas > bch2_bkey_replicas(c, k) ||
-		    (!compressed && bch2_bkey_sectors_compressed(k))) {
-			ret = false;
-			break;
-		}
-	}
-
-	offset = iter.pos.offset;
-	bch2_trans_iter_exit(&trans, &iter);
-err:
-	if (bch2_err_matches(err, BCH_ERR_transaction_restart))
-		goto retry;
-	bch2_trans_exit(&trans);
-
-	return err ? false : ret;
-}
-
-static noinline bool bch2_dio_write_check_allocated(struct dio_write *dio)
-{
-	struct bch_fs *c = dio->op.c;
-	struct bch_inode_info *inode = dio->inode;
-	struct bio *bio = &dio->op.wbio.bio;
-
-	return bch2_check_range_allocated(c, inode_inum(inode),
-				dio->op.pos.offset, bio_sectors(bio),
-				dio->op.opts.data_replicas,
-				dio->op.opts.compression != 0);
-}
-
-static void bch2_dio_write_loop_async(struct bch_write_op *);
-static __always_inline long bch2_dio_write_done(struct dio_write *dio);
-
-/*
- * We're going to return -EIOCBQUEUED, but we haven't finished consuming the
- * iov_iter yet, so we need to stash a copy of the iovec: it might be on the
- * caller's stack, we're not guaranteed that it will live for the duration of
- * the IO:
- */
-static noinline int bch2_dio_write_copy_iov(struct dio_write *dio)
-{
-	struct iovec *iov = dio->inline_vecs;
-
-	/*
-	 * iov_iter has a single embedded iovec - nothing to do:
-	 */
-	if (iter_is_ubuf(&dio->iter))
-		return 0;
-
-	/*
-	 * We don't currently handle non-iovec iov_iters here - return an error,
-	 * and we'll fall back to doing the IO synchronously:
-	 */
-	if (!iter_is_iovec(&dio->iter))
-		return -1;
-
-	if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
-		iov = kmalloc_array(dio->iter.nr_segs, sizeof(*iov),
-				    GFP_KERNEL);
-		if (unlikely(!iov))
-			return -ENOMEM;
-
-		dio->free_iov = true;
-	}
-
-	memcpy(iov, dio->iter.__iov, dio->iter.nr_segs * sizeof(*iov));
-	dio->iter.__iov = iov;
-	return 0;
-}
-
-static void bch2_dio_write_flush_done(struct closure *cl)
-{
-	struct dio_write *dio = container_of(cl, struct dio_write, op.cl);
-	struct bch_fs *c = dio->op.c;
-
-	closure_debug_destroy(cl);
-
-	dio->op.error = bch2_journal_error(&c->journal);
-
-	bch2_dio_write_done(dio);
-}
-
-static noinline void bch2_dio_write_flush(struct dio_write *dio)
-{
-	struct bch_fs *c = dio->op.c;
-	struct bch_inode_unpacked inode;
-	int ret;
-
-	dio->flush = 0;
-
-	closure_init(&dio->op.cl, NULL);
-
-	if (!dio->op.error) {
-		ret = bch2_inode_find_by_inum(c, inode_inum(dio->inode), &inode);
-		if (ret) {
-			dio->op.error = ret;
-		} else {
-			bch2_journal_flush_seq_async(&c->journal, inode.bi_journal_seq, &dio->op.cl);
-			bch2_inode_flush_nocow_writes_async(c, dio->inode, &dio->op.cl);
-		}
-	}
-
-	if (dio->sync) {
-		closure_sync(&dio->op.cl);
-		closure_debug_destroy(&dio->op.cl);
-	} else {
-		continue_at(&dio->op.cl, bch2_dio_write_flush_done, NULL);
-	}
-}
-
-static __always_inline long bch2_dio_write_done(struct dio_write *dio)
-{
-	struct kiocb *req = dio->req;
-	struct bch_inode_info *inode = dio->inode;
-	bool sync = dio->sync;
-	long ret;
-
-	if (unlikely(dio->flush)) {
-		bch2_dio_write_flush(dio);
-		if (!sync)
-			return -EIOCBQUEUED;
-	}
-
-	bch2_pagecache_block_put(inode);
-
-	if (dio->free_iov)
-		kfree(dio->iter.__iov);
-
-	ret = dio->op.error ?: ((long) dio->written << 9);
-	bio_put(&dio->op.wbio.bio);
-
-	/* inode->i_dio_count is our ref on inode and thus bch_fs */
-	inode_dio_end(&inode->v);
-
-	if (ret < 0)
-		ret = bch2_err_class(ret);
-
-	if (!sync) {
-		req->ki_complete(req, ret);
-		ret = -EIOCBQUEUED;
-	}
-	return ret;
-}
-
-static __always_inline void bch2_dio_write_end(struct dio_write *dio)
-{
-	struct bch_fs *c = dio->op.c;
-	struct kiocb *req = dio->req;
-	struct bch_inode_info *inode = dio->inode;
-	struct bio *bio = &dio->op.wbio.bio;
-
-	req->ki_pos	+= (u64) dio->op.written << 9;
-	dio->written	+= dio->op.written;
-
-	if (dio->extending) {
-		spin_lock(&inode->v.i_lock);
-		if (req->ki_pos > inode->v.i_size)
-			i_size_write(&inode->v, req->ki_pos);
-		spin_unlock(&inode->v.i_lock);
-	}
-
-	if (dio->op.i_sectors_delta || dio->quota_res.sectors) {
-		mutex_lock(&inode->ei_quota_lock);
-		__i_sectors_acct(c, inode, &dio->quota_res, dio->op.i_sectors_delta);
-		__bch2_quota_reservation_put(c, inode, &dio->quota_res);
-		mutex_unlock(&inode->ei_quota_lock);
-	}
-
-	bio_release_pages(bio, false);
-
-	if (unlikely(dio->op.error))
-		set_bit(EI_INODE_ERROR, &inode->ei_flags);
-}
-
-static __always_inline long bch2_dio_write_loop(struct dio_write *dio)
-{
-	struct bch_fs *c = dio->op.c;
-	struct kiocb *req = dio->req;
-	struct address_space *mapping = dio->mapping;
-	struct bch_inode_info *inode = dio->inode;
-	struct bch_io_opts opts;
-	struct bio *bio = &dio->op.wbio.bio;
-	unsigned unaligned, iter_count;
-	bool sync = dio->sync, dropped_locks;
-	long ret;
-
-	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
-
-	while (1) {
-		iter_count = dio->iter.count;
-
-		EBUG_ON(current->faults_disabled_mapping);
-		current->faults_disabled_mapping = mapping;
-
-		ret = bio_iov_iter_get_pages(bio, &dio->iter);
-
-		dropped_locks = fdm_dropped_locks();
-
-		current->faults_disabled_mapping = NULL;
-
-		/*
-		 * If the fault handler returned an error but also signalled
-		 * that it dropped & retook ei_pagecache_lock, we just need to
-		 * re-shoot down the page cache and retry:
-		 */
-		if (dropped_locks && ret)
-			ret = 0;
-
-		if (unlikely(ret < 0))
-			goto err;
-
-		if (unlikely(dropped_locks)) {
-			ret = write_invalidate_inode_pages_range(mapping,
-					req->ki_pos,
-					req->ki_pos + iter_count - 1);
-			if (unlikely(ret))
-				goto err;
-
-			if (!bio->bi_iter.bi_size)
-				continue;
-		}
-
-		unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
-		bio->bi_iter.bi_size -= unaligned;
-		iov_iter_revert(&dio->iter, unaligned);
-
-		if (!bio->bi_iter.bi_size) {
-			/*
-			 * bio_iov_iter_get_pages was only able to get <
-			 * blocksize worth of pages:
-			 */
-			ret = -EFAULT;
-			goto err;
-		}
-
-		bch2_write_op_init(&dio->op, c, opts);
-		dio->op.end_io		= sync
-			? NULL
-			: bch2_dio_write_loop_async;
-		dio->op.target		= dio->op.opts.foreground_target;
-		dio->op.write_point	= writepoint_hashed((unsigned long) current);
-		dio->op.nr_replicas	= dio->op.opts.data_replicas;
-		dio->op.subvol		= inode->ei_subvol;
-		dio->op.pos		= POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
-		dio->op.devs_need_flush	= &inode->ei_devs_need_flush;
-
-		if (sync)
-			dio->op.flags |= BCH_WRITE_SYNC;
-		dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
-
-		ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
-						 bio_sectors(bio), true);
-		if (unlikely(ret))
-			goto err;
-
-		ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
-						dio->op.opts.data_replicas, 0);
-		if (unlikely(ret) &&
-		    !bch2_dio_write_check_allocated(dio))
-			goto err;
-
-		task_io_account_write(bio->bi_iter.bi_size);
-
-		if (unlikely(dio->iter.count) &&
-		    !dio->sync &&
-		    !dio->loop &&
-		    bch2_dio_write_copy_iov(dio))
-			dio->sync = sync = true;
-
-		dio->loop = true;
-		closure_call(&dio->op.cl, bch2_write, NULL, NULL);
-
-		if (!sync)
-			return -EIOCBQUEUED;
-
-		bch2_dio_write_end(dio);
-
-		if (likely(!dio->iter.count) || dio->op.error)
-			break;
-
-		bio_reset(bio, NULL, REQ_OP_WRITE);
-	}
-out:
-	return bch2_dio_write_done(dio);
-err:
-	dio->op.error = ret;
-
-	bio_release_pages(bio, false);
-
-	bch2_quota_reservation_put(c, inode, &dio->quota_res);
-	goto out;
-}
-
-static noinline __cold void bch2_dio_write_continue(struct dio_write *dio)
-{
-	struct mm_struct *mm = dio->mm;
-
-	bio_reset(&dio->op.wbio.bio, NULL, REQ_OP_WRITE);
-
-	if (mm)
-		kthread_use_mm(mm);
-	bch2_dio_write_loop(dio);
-	if (mm)
-		kthread_unuse_mm(mm);
-}
-
-static void bch2_dio_write_loop_async(struct bch_write_op *op)
-{
-	struct dio_write *dio = container_of(op, struct dio_write, op);
-
-	bch2_dio_write_end(dio);
-
-	if (likely(!dio->iter.count) || dio->op.error)
-		bch2_dio_write_done(dio);
-	else
-		bch2_dio_write_continue(dio);
-}
-
-static noinline
-ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
-{
-	struct file *file = req->ki_filp;
-	struct address_space *mapping = file->f_mapping;
-	struct bch_inode_info *inode = file_bch_inode(file);
-	struct bch_fs *c = inode->v.i_sb->s_fs_info;
-	struct dio_write *dio;
-	struct bio *bio;
-	bool locked = true, extending;
-	ssize_t ret;
-
-	prefetch(&c->opts);
-	prefetch((void *) &c->opts + 64);
-	prefetch(&inode->ei_inode);
-	prefetch((void *) &inode->ei_inode + 64);
-
-	inode_lock(&inode->v);
-
-	ret = generic_write_checks(req, iter);
-	if (unlikely(ret <= 0))
-		goto err;
-
-	ret = file_remove_privs(file);
-	if (unlikely(ret))
-		goto err;
-
-	ret = file_update_time(file);
-	if (unlikely(ret))
-		goto err;
-
-	if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
-		goto err;
-
-	inode_dio_begin(&inode->v);
-	bch2_pagecache_block_get(inode);
-
-	extending = req->ki_pos + iter->count > inode->v.i_size;
-	if (!extending) {
-		inode_unlock(&inode->v);
-		locked = false;
-	}
-
-	bio = bio_alloc_bioset(NULL,
-			       bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
-			       REQ_OP_WRITE,
-			       GFP_KERNEL,
-			       &c->dio_write_bioset);
-	dio = container_of(bio, struct dio_write, op.wbio.bio);
-	dio->req		= req;
-	dio->mapping		= mapping;
-	dio->inode		= inode;
-	dio->mm			= current->mm;
-	dio->loop		= false;
-	dio->extending		= extending;
-	dio->sync		= is_sync_kiocb(req) || extending;
-	dio->flush		= iocb_is_dsync(req) && !c->opts.journal_flush_disabled;
-	dio->free_iov		= false;
-	dio->quota_res.sectors	= 0;
-	dio->written		= 0;
-	dio->iter		= *iter;
-	dio->op.c		= c;
-
-	if (unlikely(mapping->nrpages)) {
-		ret = write_invalidate_inode_pages_range(mapping,
-						req->ki_pos,
-						req->ki_pos + iter->count - 1);
-		if (unlikely(ret))
-			goto err_put_bio;
-	}
-
-	ret = bch2_dio_write_loop(dio);
-err:
-	if (locked)
-		inode_unlock(&inode->v);
-	return ret;
-err_put_bio:
-	bch2_pagecache_block_put(inode);
-	bio_put(bio);
-	inode_dio_end(&inode->v);
-	goto err;
-}
-
-ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
-{
-	struct file *file = iocb->ki_filp;
-	struct bch_inode_info *inode = file_bch_inode(file);
-	ssize_t ret;
-
-	if (iocb->ki_flags & IOCB_DIRECT) {
-		ret = bch2_direct_write(iocb, from);
-		goto out;
-	}
-
-	/* We can write back this queue in page reclaim */
-	current->backing_dev_info = inode_to_bdi(&inode->v);
-	inode_lock(&inode->v);
-
-	ret = generic_write_checks(iocb, from);
-	if (ret <= 0)
-		goto unlock;
-
-	ret = file_remove_privs(file);
-	if (ret)
-		goto unlock;
-
-	ret = file_update_time(file);
-	if (ret)
-		goto unlock;
-
-	ret = bch2_buffered_write(iocb, from);
-	if (likely(ret > 0))
-		iocb->ki_pos += ret;
-unlock:
-	inode_unlock(&inode->v);
-	current->backing_dev_info = NULL;
-
-	if (ret > 0)
-		ret = generic_write_sync(iocb, ret);
-out:
-	return bch2_err_class(ret);
-}
 
 /* fsync: */
 
@@ -2911,10 +305,10 @@ static int __bch2_truncate_folio(struct bch_inode_info *inode,
 		s->s[i].nr_replicas	= 0;
 
 		i_sectors_delta -= s->s[i].state == SECTOR_dirty;
-		folio_sector_set(folio, s, i, SECTOR_unallocated);
+		bch2_folio_sector_set(folio, s, i, SECTOR_unallocated);
 	}
 
-	i_sectors_acct(c, inode, NULL, i_sectors_delta);
+	bch2_i_sectors_acct(c, inode, NULL, i_sectors_delta);
 
 	/*
 	 * Caller needs to know whether this folio will be written out by
@@ -3105,7 +499,7 @@ int bch2_truncate(struct mnt_idmap *idmap,
 	ret = bch2_fpunch(c, inode_inum(inode),
 			round_up(iattr->ia_size, block_bytes(c)) >> 9,
 			U64_MAX, &i_sectors_delta);
-	i_sectors_acct(c, inode, NULL, i_sectors_delta);
+	bch2_i_sectors_acct(c, inode, NULL, i_sectors_delta);
 
 	bch2_fs_inconsistent_on(!inode->v.i_size && inode->v.i_blocks &&
 				!bch2_journal_error(&c->journal), c,
@@ -3159,7 +553,7 @@ static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len
 		ret = bch2_fpunch(c, inode_inum(inode),
 				  block_start >> 9, block_end >> 9,
 				  &i_sectors_delta);
-		i_sectors_acct(c, inode, NULL, i_sectors_delta);
+		bch2_i_sectors_acct(c, inode, NULL, i_sectors_delta);
 	}
 
 	mutex_lock(&inode->ei_update_lock);
@@ -3210,7 +604,7 @@ static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
 
 	new_size = inode->v.i_size + shift;
 
-	ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
+	ret = bch2_write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
 	if (ret)
 		return ret;
 
@@ -3226,7 +620,7 @@ static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
 		ret = bch2_fpunch(c, inode_inum(inode),
 				  offset >> 9, (offset + len) >> 9,
 				  &i_sectors_delta);
-		i_sectors_acct(c, inode, NULL, i_sectors_delta);
+		bch2_i_sectors_acct(c, inode, NULL, i_sectors_delta);
 
 		if (ret)
 			return ret;
@@ -3410,6 +804,10 @@ static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
 		}
 
 		if (!(mode & FALLOC_FL_ZERO_RANGE)) {
+			/*
+			 * Lock ordering - can't be holding btree locks while
+			 * blocking on a folio lock:
+			 */
 			if (bch2_clamp_data_hole(&inode->v,
 						 &hole_start,
 						 &hole_end,
@@ -3443,10 +841,10 @@ static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
 		if (ret)
 			goto bkey_err;
 
-		i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
+		bch2_i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
 
 		drop_locks_do(&trans,
-			(mark_pagecache_reserved(inode, hole_start, iter.pos.offset), 0));
+			(bch2_mark_pagecache_reserved(inode, hole_start, iter.pos.offset), 0));
 bkey_err:
 		bch2_quota_reservation_put(c, inode, &quota_res);
 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
@@ -3459,7 +857,7 @@ bkey_err:
 
 		bch2_fpunch_at(&trans, &iter, inode_inum(inode),
 			       end_sector, &i_sectors_delta);
-		i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
+		bch2_i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
 		bch2_quota_reservation_put(c, inode, &quota_res);
 	}
 
@@ -3653,7 +1051,7 @@ loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
 
 	aligned_len = round_up((u64) len, block_bytes(c));
 
-	ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
+	ret = bch2_write_invalidate_inode_pages_range(dst->v.i_mapping,
 				pos_dst, pos_dst + len - 1);
 	if (ret)
 		goto err;
@@ -3665,7 +1063,7 @@ loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
 
 	file_update_time(file_dst);
 
-	mark_pagecache_unallocated(src, pos_src >> 9,
+	bch2_mark_pagecache_unallocated(src, pos_src >> 9,
 				   (pos_src + aligned_len) >> 9);
 
 	ret = bch2_remap_range(c,
@@ -3681,7 +1079,7 @@ loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
 	 */
 	ret = min((u64) ret << 9, (u64) len);
 
-	i_sectors_acct(c, dst, &quota_res, i_sectors_delta);
+	bch2_i_sectors_acct(c, dst, &quota_res, i_sectors_delta);
 
 	spin_lock(&dst->v.i_lock);
 	if (pos_dst + ret > dst->v.i_size)
@@ -3700,68 +1098,6 @@ err:
 
 /* fseek: */
 
-static int folio_data_offset(struct folio *folio, loff_t pos,
-			     unsigned min_replicas)
-{
-	struct bch_folio *s = bch2_folio(folio);
-	unsigned i, sectors = folio_sectors(folio);
-
-	if (s)
-		for (i = folio_pos_to_s(folio, pos); i < sectors; i++)
-			if (s->s[i].state >= SECTOR_dirty &&
-			    s->s[i].nr_replicas + s->s[i].replicas_reserved >= min_replicas)
-				return i << SECTOR_SHIFT;
-
-	return -1;
-}
-
-static loff_t bch2_seek_pagecache_data(struct inode *vinode,
-				       loff_t start_offset,
-				       loff_t end_offset,
-				       unsigned min_replicas,
-				       bool nonblock)
-{
-	struct folio_batch fbatch;
-	pgoff_t start_index	= start_offset >> PAGE_SHIFT;
-	pgoff_t end_index	= end_offset >> PAGE_SHIFT;
-	pgoff_t index		= start_index;
-	unsigned i;
-	loff_t ret;
-	int offset;
-
-	folio_batch_init(&fbatch);
-
-	while (filemap_get_folios(vinode->i_mapping,
-				  &index, end_index, &fbatch)) {
-		for (i = 0; i < folio_batch_count(&fbatch); i++) {
-			struct folio *folio = fbatch.folios[i];
-
-			if (!nonblock) {
-				folio_lock(folio);
-			} else if (!folio_trylock(folio)) {
-				folio_batch_release(&fbatch);
-				return -EAGAIN;
-			}
-
-			offset = folio_data_offset(folio,
-					max(folio_pos(folio), start_offset),
-					min_replicas);
-			if (offset >= 0) {
-				ret = clamp(folio_pos(folio) + offset,
-					    start_offset, end_offset);
-				folio_unlock(folio);
-				folio_batch_release(&fbatch);
-				return ret;
-			}
-			folio_unlock(folio);
-		}
-		folio_batch_release(&fbatch);
-		cond_resched();
-	}
-
-	return end_offset;
-}
-
 static loff_t bch2_seek_data(struct file *file, u64 offset)
 {
 	struct bch_inode_info *inode = file_bch_inode(file);
@@ -3815,88 +1151,6 @@ err:
 	return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
 }
 
-static int folio_hole_offset(struct address_space *mapping, loff_t *offset,
-			      unsigned min_replicas, bool nonblock)
-{
-	struct folio *folio;
-	struct bch_folio *s;
-	unsigned i, sectors;
-	bool ret = true;
-
-	folio = __filemap_get_folio(mapping, *offset >> PAGE_SHIFT,
-				    !nonblock ? FGP_LOCK : 0, 0);
-	if (IS_ERR_OR_NULL(folio))
-		return true;
-
-	if (nonblock && !folio_trylock(folio)) {
-		folio_put(folio);
-		return -EAGAIN;
-	}
-
-	s = bch2_folio(folio);
-	if (!s)
-		goto unlock;
-
-	sectors = folio_sectors(folio);
-	for (i = folio_pos_to_s(folio, *offset); i < sectors; i++)
-		if (s->s[i].state < SECTOR_dirty ||
-		    s->s[i].nr_replicas + s->s[i].replicas_reserved < min_replicas) {
-			*offset = max(*offset,
-				      folio_pos(folio) + (i << SECTOR_SHIFT));
-			goto unlock;
-		}
-
-	*offset = folio_end_pos(folio);
-	ret = false;
-unlock:
-	folio_unlock(folio);
-	folio_put(folio);
-	return ret;
-}
-
-static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
-				       loff_t start_offset,
-				       loff_t end_offset,
-				       unsigned min_replicas,
-				       bool nonblock)
-{
-	struct address_space *mapping = vinode->i_mapping;
-	loff_t offset = start_offset;
-
-	while (offset < end_offset &&
-	       !folio_hole_offset(mapping, &offset, min_replicas, nonblock))
-		;
-
-	return min(offset, end_offset);
-}
-
-static int bch2_clamp_data_hole(struct inode *inode,
-				u64 *hole_start,
-				u64 *hole_end,
-				unsigned min_replicas,
-				bool nonblock)
-{
-	loff_t ret;
-
-	ret = bch2_seek_pagecache_hole(inode,
-		*hole_start << 9, *hole_end << 9, min_replicas, nonblock) >> 9;
-	if (ret < 0)
-		return ret;
-
-	*hole_start = ret;
-
-	if (*hole_start == *hole_end)
-		return 0;
-
-	ret = bch2_seek_pagecache_data(inode,
-		*hole_start << 9, *hole_end << 9, min_replicas, nonblock) >> 9;
-	if (ret < 0)
-		return ret;
-
-	*hole_end = ret;
-	return 0;
-}
-
 static loff_t bch2_seek_hole(struct file *file, u64 offset)
 {
 	struct bch_inode_info *inode = file_bch_inode(file);
@@ -3981,28 +1235,10 @@ loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
 void bch2_fs_fsio_exit(struct bch_fs *c)
 {
 	bioset_exit(&c->nocow_flush_bioset);
-	bioset_exit(&c->dio_write_bioset);
-	bioset_exit(&c->dio_read_bioset);
-	bioset_exit(&c->writepage_bioset);
 }
 
 int bch2_fs_fsio_init(struct bch_fs *c)
 {
-	if (bioset_init(&c->writepage_bioset,
-			4, offsetof(struct bch_writepage_io, op.wbio.bio),
-			BIOSET_NEED_BVECS))
-		return -BCH_ERR_ENOMEM_writepage_bioset_init;
-
-	if (bioset_init(&c->dio_read_bioset,
-			4, offsetof(struct dio_read, rbio.bio),
-			BIOSET_NEED_BVECS))
-		return -BCH_ERR_ENOMEM_dio_read_bioset_init;
-
-	if (bioset_init(&c->dio_write_bioset,
-			4, offsetof(struct dio_write, op.wbio.bio),
-			BIOSET_NEED_BVECS))
-		return -BCH_ERR_ENOMEM_dio_write_bioset_init;
-
 	if (bioset_init(&c->nocow_flush_bioset,
 			1, offsetof(struct nocow_flush, bio), 0))
 		return -BCH_ERR_ENOMEM_nocow_flush_bioset_init;
diff --git a/libbcachefs/fs-io.h b/libbcachefs/fs-io.h
index af905331..c6704618 100644
--- a/libbcachefs/fs-io.h
+++ b/libbcachefs/fs-io.h
@@ -5,28 +5,163 @@
 #ifndef NO_BCACHEFS_FS
 
 #include "buckets.h"
+#include "fs.h"
 #include "io_types.h"
+#include "quota.h"
 
 #include <linux/uio.h>
 
-struct quota_res;
+struct folio_vec {
+	struct folio	*fv_folio;
+	size_t		fv_offset;
+	size_t		fv_len;
+};
+
+static inline struct folio_vec biovec_to_foliovec(struct bio_vec bv)
+{
+
+	struct folio *folio	= page_folio(bv.bv_page);
+	size_t offset		= (folio_page_idx(folio, bv.bv_page) << PAGE_SHIFT) +
+		bv.bv_offset;
+	size_t len = min_t(size_t, folio_size(folio) - offset, bv.bv_len);
+
+	return (struct folio_vec) {
+		.fv_folio	= folio,
+		.fv_offset	= offset,
+		.fv_len		= len,
+	};
+}
+
+static inline struct folio_vec bio_iter_iovec_folio(struct bio *bio,
+						    struct bvec_iter iter)
+{
+	return biovec_to_foliovec(bio_iter_iovec(bio, iter));
+}
+
+#define __bio_for_each_folio(bvl, bio, iter, start)			\
+	for (iter = (start);						\
+	     (iter).bi_size &&						\
+		((bvl = bio_iter_iovec_folio((bio), (iter))), 1);	\
+	     bio_advance_iter_single((bio), &(iter), (bvl).fv_len))
+
+/**
+ * bio_for_each_folio - iterate over folios within a bio
+ *
+ * Like other non-_all versions, this iterates over what bio->bi_iter currently
+ * points to. This version is for drivers, where the bio may have previously
+ * been split or cloned.
+ */
+#define bio_for_each_folio(bvl, bio, iter)				\
+	__bio_for_each_folio(bvl, bio, iter, (bio)->bi_iter)
+
+struct quota_res {
+	u64				sectors;
+};
+
+#ifdef CONFIG_BCACHEFS_QUOTA
+
+static inline void __bch2_quota_reservation_put(struct bch_fs *c,
+					 struct bch_inode_info *inode,
+					 struct quota_res *res)
+{
+	BUG_ON(res->sectors > inode->ei_quota_reserved);
+
+	bch2_quota_acct(c, inode->ei_qid, Q_SPC,
+			-((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
+	inode->ei_quota_reserved -= res->sectors;
+	res->sectors = 0;
+}
+
+static inline void bch2_quota_reservation_put(struct bch_fs *c,
+				       struct bch_inode_info *inode,
+				       struct quota_res *res)
+{
+	if (res->sectors) {
+		mutex_lock(&inode->ei_quota_lock);
+		__bch2_quota_reservation_put(c, inode, res);
+		mutex_unlock(&inode->ei_quota_lock);
+	}
+}
+
+static inline int bch2_quota_reservation_add(struct bch_fs *c,
+				      struct bch_inode_info *inode,
+				      struct quota_res *res,
+				      u64 sectors,
+				      bool check_enospc)
+{
+	int ret;
+
+	if (test_bit(EI_INODE_SNAPSHOT, &inode->ei_flags))
+		return 0;
+
+	mutex_lock(&inode->ei_quota_lock);
+	ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
+			      check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
+	if (likely(!ret)) {
+		inode->ei_quota_reserved += sectors;
+		res->sectors += sectors;
+	}
+	mutex_unlock(&inode->ei_quota_lock);
+
+	return ret;
+}
 
-int __must_check bch2_write_inode_size(struct bch_fs *,
-				       struct bch_inode_info *,
-				       loff_t, unsigned);
+#else
+
+static void __bch2_quota_reservation_put(struct bch_fs *c,
+					 struct bch_inode_info *inode,
+					 struct quota_res *res) {}
 
-int bch2_read_folio(struct file *, struct folio *);
+static void bch2_quota_reservation_put(struct bch_fs *c,
+				       struct bch_inode_info *inode,
+				       struct quota_res *res) {}
 
-int bch2_writepages(struct address_space *, struct writeback_control *);
-void bch2_readahead(struct readahead_control *);
+static int bch2_quota_reservation_add(struct bch_fs *c,
+				      struct bch_inode_info *inode,
+				      struct quota_res *res,
+				      unsigned sectors,
+				      bool check_enospc)
+{
+	return 0;
+}
 
-int bch2_write_begin(struct file *, struct address_space *, loff_t,
-		     unsigned, struct page **, void **);
-int bch2_write_end(struct file *, struct address_space *, loff_t,
-		   unsigned, unsigned, struct page *, void *);
+#endif
 
-ssize_t bch2_read_iter(struct kiocb *, struct iov_iter *);
-ssize_t bch2_write_iter(struct kiocb *, struct iov_iter *);
+void __bch2_i_sectors_acct(struct bch_fs *, struct bch_inode_info *,
+			   struct quota_res *, s64);
+
+static inline void bch2_i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
+				       struct quota_res *quota_res, s64 sectors)
+{
+	if (sectors) {
+		mutex_lock(&inode->ei_quota_lock);
+		__bch2_i_sectors_acct(c, inode, quota_res, sectors);
+		mutex_unlock(&inode->ei_quota_lock);
+	}
+}
+
+static inline struct address_space *faults_disabled_mapping(void)
+{
+	return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
+}
+
+static inline void set_fdm_dropped_locks(void)
+{
+	current->faults_disabled_mapping =
+		(void *) (((unsigned long) current->faults_disabled_mapping)|1);
+}
+
+static inline bool fdm_dropped_locks(void)
+{
+	return ((unsigned long) current->faults_disabled_mapping) & 1;
+}
+
+void bch2_inode_flush_nocow_writes_async(struct bch_fs *,
+			struct bch_inode_info *, struct closure *);
+
+int __must_check bch2_write_inode_size(struct bch_fs *,
+				       struct bch_inode_info *,
+				       loff_t, unsigned);
 
 int bch2_fsync(struct file *, loff_t, loff_t, int);
 
@@ -39,11 +174,6 @@ loff_t bch2_remap_file_range(struct file *, loff_t, struct file *,
 
 loff_t bch2_llseek(struct file *, loff_t, int);
 
-vm_fault_t bch2_page_fault(struct vm_fault *);
-vm_fault_t bch2_page_mkwrite(struct vm_fault *);
-void bch2_invalidate_folio(struct folio *, size_t, size_t);
-bool bch2_release_folio(struct folio *, gfp_t);
-
 void bch2_fs_fsio_exit(struct bch_fs *);
 int bch2_fs_fsio_init(struct bch_fs *);
 #else
diff --git a/libbcachefs/fs.c b/libbcachefs/fs.c
index 8d2f388b..0b550a97 100644
--- a/libbcachefs/fs.c
+++ b/libbcachefs/fs.c
@@ -14,6 +14,8 @@
 #include "fs-common.h"
 #include "fs-io.h"
 #include "fs-ioctl.h"
+#include "fs-io-buffered.h"
+#include "fs-io-pagecache.h"
 #include "fsck.h"
 #include "inode.h"
 #include "io.h"
@@ -203,7 +205,7 @@ struct inode *bch2_vfs_inode_get(struct bch_fs *c, subvol_inum inum)
 
 	if (ret) {
 		iget_failed(&inode->v);
-		return ERR_PTR(ret);
+		return ERR_PTR(bch2_err_class(ret));
 	}
 
 	mutex_lock(&c->vfs_inodes_lock);
@@ -1000,11 +1002,16 @@ static int bch2_vfs_readdir(struct file *file, struct dir_context *ctx)
 {
 	struct bch_inode_info *inode = file_bch_inode(file);
 	struct bch_fs *c = inode->v.i_sb->s_fs_info;
+	int ret;
 
 	if (!dir_emit_dots(file, ctx))
 		return 0;
 
-	return bch2_readdir(c, inode_inum(inode), ctx);
+	ret = bch2_readdir(c, inode_inum(inode), ctx);
+	if (ret)
+		bch_err_fn(c, ret);
+
+	return bch2_err_class(ret);
 }
 
 static const struct file_operations bch_file_operations = {
diff --git a/libbcachefs/fsck.c b/libbcachefs/fsck.c
index 0852dbe9..d99c04af 100644
--- a/libbcachefs/fsck.c
+++ b/libbcachefs/fsck.c
@@ -11,6 +11,7 @@
 #include "fsck.h"
 #include "inode.h"
 #include "keylist.h"
+#include "recovery.h"
 #include "subvolume.h"
 #include "super.h"
 #include "xattr.h"
diff --git a/libbcachefs/inode.c b/libbcachefs/inode.c
index fea21e1e..e0d41655 100644
--- a/libbcachefs/inode.c
+++ b/libbcachefs/inode.c
@@ -348,6 +348,8 @@ int bch2_inode_peek(struct btree_trans *trans,
 	return 0;
 err:
 	bch2_trans_iter_exit(trans, iter);
+	if (!bch2_err_matches(ret, BCH_ERR_transaction_restart))
+		bch_err_msg(trans->c, ret, "looking up inum %u:%llu:", inum.subvol, inum.inum);
 	return ret;
 }
 
@@ -520,23 +522,25 @@ void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c
 	__bch2_inode_unpacked_to_text(out, &inode);
 }
 
-static inline bool bkey_is_deleted_inode(struct bkey_s_c k)
+static inline u64 bkey_inode_flags(struct bkey_s_c k)
 {
 	switch (k.k->type) {
 	case KEY_TYPE_inode:
-		return bkey_s_c_to_inode(k).v->bi_flags &
-			cpu_to_le32(BCH_INODE_UNLINKED);
+		return le32_to_cpu(bkey_s_c_to_inode(k).v->bi_flags);
 	case KEY_TYPE_inode_v2:
-		return bkey_s_c_to_inode_v2(k).v->bi_flags &
-			cpu_to_le32(BCH_INODE_UNLINKED);
+		return le64_to_cpu(bkey_s_c_to_inode_v2(k).v->bi_flags);
 	case KEY_TYPE_inode_v3:
-		return bkey_s_c_to_inode_v3(k).v->bi_flags &
-			cpu_to_le64(BCH_INODE_UNLINKED);
+		return le64_to_cpu(bkey_s_c_to_inode_v3(k).v->bi_flags);
 	default:
-		return false;
+		return 0;
 	}
 }
 
+static inline bool bkey_is_deleted_inode(struct bkey_s_c k)
+{
+	return bkey_inode_flags(k) & BCH_INODE_UNLINKED;
+}
+
 int bch2_trans_mark_inode(struct btree_trans *trans,
 			  enum btree_id btree_id, unsigned level,
 			  struct bkey_s_c old,
diff --git a/libbcachefs/journal_io.c b/libbcachefs/journal_io.c
index f861ae2f..d141c749 100644
--- a/libbcachefs/journal_io.c
+++ b/libbcachefs/journal_io.c
@@ -14,6 +14,7 @@
 #include "journal_reclaim.h"
 #include "journal_seq_blacklist.h"
 #include "replicas.h"
+#include "sb-clean.h"
 #include "trace.h"
 
 static struct nonce journal_nonce(const struct jset *jset)
@@ -208,33 +209,41 @@ static void journal_entry_null_range(void *start, void *end)
 #define JOURNAL_ENTRY_BAD	7
 
 static void journal_entry_err_msg(struct printbuf *out,
+				  u32 version,
 				  struct jset *jset,
 				  struct jset_entry *entry)
 {
-	prt_str(out, "invalid journal entry ");
-	if (entry)
-		prt_printf(out, "%s ", bch2_jset_entry_types[entry->type]);
-
-	if (!jset)
-		prt_printf(out, "in superblock");
-	else if (!entry)
-		prt_printf(out, "at seq %llu", le64_to_cpu(jset->seq));
-	else
-		prt_printf(out, "at offset %zi/%u seq %llu",
-			   (u64 *) entry - jset->_data,
-			   le32_to_cpu(jset->u64s),
-			   le64_to_cpu(jset->seq));
+	prt_str(out, "invalid journal entry, version=");
+	bch2_version_to_text(out, version);
+
+	if (entry) {
+		prt_str(out, " type=");
+		prt_str(out, bch2_jset_entry_types[entry->type]);
+	}
+
+	if (!jset) {
+		prt_printf(out, " in superblock");
+	} else {
+
+		prt_printf(out, " seq=%llu", le64_to_cpu(jset->seq));
+
+		if (entry)
+			prt_printf(out, " offset=%zi/%u",
+				   (u64 *) entry - jset->_data,
+				   le32_to_cpu(jset->u64s));
+	}
+
 	prt_str(out, ": ");
 }
 
-#define journal_entry_err(c, jset, entry, msg, ...)			\
+#define journal_entry_err(c, version, jset, entry, msg, ...)		\
 ({									\
 	struct printbuf buf = PRINTBUF;					\
 									\
-	journal_entry_err_msg(&buf, jset, entry);			\
+	journal_entry_err_msg(&buf, version, jset, entry);		\
 	prt_printf(&buf, msg, ##__VA_ARGS__);				\
 									\
-	switch (write) {						\
+	switch (flags & BKEY_INVALID_WRITE) {				\
 	case READ:							\
 		mustfix_fsck_err(c, "%s", buf.buf);			\
 		break;							\
@@ -251,8 +260,8 @@ static void journal_entry_err_msg(struct printbuf *out,
 	true;								\
 })
 
-#define journal_entry_err_on(cond, c, jset, entry, msg, ...)		\
-	((cond) ? journal_entry_err(c, jset, entry, msg, ##__VA_ARGS__) : false)
+#define journal_entry_err_on(cond, c, version, jset, entry, msg, ...)	\
+	((cond) ? journal_entry_err(c, version, jset, entry, msg, ##__VA_ARGS__) : false)
 
 #define FSCK_DELETED_KEY	5
 
@@ -261,13 +270,15 @@ static int journal_validate_key(struct bch_fs *c,
 				struct jset_entry *entry,
 				unsigned level, enum btree_id btree_id,
 				struct bkey_i *k,
-				unsigned version, int big_endian, int write)
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
+	int write = flags & BKEY_INVALID_WRITE;
 	void *next = vstruct_next(entry);
 	struct printbuf buf = PRINTBUF;
 	int ret = 0;
 
-	if (journal_entry_err_on(!k->k.u64s, c, jset, entry, "k->u64s 0")) {
+	if (journal_entry_err_on(!k->k.u64s, c, version, jset, entry, "k->u64s 0")) {
 		entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 		journal_entry_null_range(vstruct_next(entry), next);
 		return FSCK_DELETED_KEY;
@@ -275,7 +286,7 @@ static int journal_validate_key(struct bch_fs *c,
 
 	if (journal_entry_err_on((void *) bkey_next(k) >
 				 (void *) vstruct_next(entry),
-				 c, jset, entry,
+				 c, version, jset, entry,
 				 "extends past end of journal entry")) {
 		entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 		journal_entry_null_range(vstruct_next(entry), next);
@@ -283,7 +294,7 @@ static int journal_validate_key(struct bch_fs *c,
 	}
 
 	if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT,
-				 c, jset, entry,
+				 c, version, jset, entry,
 				 "bad format %u", k->k.format)) {
 		le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 		memmove(k, bkey_next(k), next - (void *) bkey_next(k));
@@ -298,11 +309,7 @@ static int journal_validate_key(struct bch_fs *c,
 	if (bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 			      __btree_node_type(level, btree_id), write, &buf)) {
 		printbuf_reset(&buf);
-		prt_printf(&buf, "invalid journal entry %s at offset %zi/%u seq %llu:",
-			   bch2_jset_entry_types[entry->type],
-			   (u64 *) entry - jset->_data,
-			   le32_to_cpu(jset->u64s),
-			   le64_to_cpu(jset->seq));
+		journal_entry_err_msg(&buf, version, jset, entry);
 		prt_newline(&buf);
 		printbuf_indent_add(&buf, 2);
 
@@ -312,6 +319,7 @@ static int journal_validate_key(struct bch_fs *c,
 				  __btree_node_type(level, btree_id), write, &buf);
 
 		mustfix_fsck_err(c, "%s", buf.buf);
+		BUG();
 
 		le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 		memmove(k, bkey_next(k), next - (void *) bkey_next(k));
@@ -330,9 +338,10 @@ fsck_err:
 }
 
 static int journal_entry_btree_keys_validate(struct bch_fs *c,
-					     struct jset *jset,
-					     struct jset_entry *entry,
-					     unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct bkey_i *k = entry->start;
 
@@ -341,7 +350,7 @@ static int journal_entry_btree_keys_validate(struct bch_fs *c,
 					       entry->level,
 					       entry->btree_id,
 					       k, version, big_endian,
-					       write|BKEY_INVALID_JOURNAL);
+					       flags|BKEY_INVALID_JOURNAL);
 		if (ret == FSCK_DELETED_KEY)
 			continue;
 
@@ -369,16 +378,17 @@ static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs
 }
 
 static int journal_entry_btree_root_validate(struct bch_fs *c,
-					     struct jset *jset,
-					     struct jset_entry *entry,
-					     unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct bkey_i *k = entry->start;
 	int ret = 0;
 
 	if (journal_entry_err_on(!entry->u64s ||
 				 le16_to_cpu(entry->u64s) != k->k.u64s,
-				 c, jset, entry,
+				 c, version, jset, entry,
 				 "invalid btree root journal entry: wrong number of keys")) {
 		void *next = vstruct_next(entry);
 		/*
@@ -392,7 +402,7 @@ static int journal_entry_btree_root_validate(struct bch_fs *c,
 	}
 
 	return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
-				    version, big_endian, write);
+				    version, big_endian, flags);
 fsck_err:
 	return ret;
 }
@@ -404,9 +414,10 @@ static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs
 }
 
 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
-					    struct jset *jset,
-					    struct jset_entry *entry,
-					    unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	/* obsolete, don't care: */
 	return 0;
@@ -418,14 +429,15 @@ static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs
 }
 
 static int journal_entry_blacklist_validate(struct bch_fs *c,
-					    struct jset *jset,
-					    struct jset_entry *entry,
-					    unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	int ret = 0;
 
 	if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1,
-				 c, jset, entry,
+				 c, version, jset, entry,
 		"invalid journal seq blacklist entry: bad size")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 	}
@@ -443,15 +455,16 @@ static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs
 }
 
 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
-					       struct jset *jset,
-					       struct jset_entry *entry,
-					       unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct jset_entry_blacklist_v2 *bl_entry;
 	int ret = 0;
 
 	if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2,
-				 c, jset, entry,
+				 c, version, jset, entry,
 		"invalid journal seq blacklist entry: bad size")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		goto out;
@@ -461,7 +474,7 @@ static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 
 	if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 				 le64_to_cpu(bl_entry->end),
-				 c, jset, entry,
+				 c, version, jset, entry,
 		"invalid journal seq blacklist entry: start > end")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 	}
@@ -482,9 +495,10 @@ static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_
 }
 
 static int journal_entry_usage_validate(struct bch_fs *c,
-					struct jset *jset,
-					struct jset_entry *entry,
-					unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct jset_entry_usage *u =
 		container_of(entry, struct jset_entry_usage, entry);
@@ -492,7 +506,7 @@ static int journal_entry_usage_validate(struct bch_fs *c,
 	int ret = 0;
 
 	if (journal_entry_err_on(bytes < sizeof(*u),
-				 c, jset, entry,
+				 c, version, jset, entry,
 				 "invalid journal entry usage: bad size")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
@@ -514,9 +528,10 @@ static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 static int journal_entry_data_usage_validate(struct bch_fs *c,
-					struct jset *jset,
-					struct jset_entry *entry,
-					unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct jset_entry_data_usage *u =
 		container_of(entry, struct jset_entry_data_usage, entry);
@@ -525,7 +540,7 @@ static int journal_entry_data_usage_validate(struct bch_fs *c,
 
 	if (journal_entry_err_on(bytes < sizeof(*u) ||
 				 bytes < sizeof(*u) + u->r.nr_devs,
-				 c, jset, entry,
+				 c, version, jset, entry,
 				 "invalid journal entry usage: bad size")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
@@ -546,9 +561,10 @@ static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs
 }
 
 static int journal_entry_clock_validate(struct bch_fs *c,
-					struct jset *jset,
-					struct jset_entry *entry,
-					unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct jset_entry_clock *clock =
 		container_of(entry, struct jset_entry_clock, entry);
@@ -556,13 +572,13 @@ static int journal_entry_clock_validate(struct bch_fs *c,
 	int ret = 0;
 
 	if (journal_entry_err_on(bytes != sizeof(*clock),
-				 c, jset, entry, "bad size")) {
+				 c, version, jset, entry, "bad size")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
 	}
 
 	if (journal_entry_err_on(clock->rw > 1,
-				 c, jset, entry, "bad rw")) {
+				 c, version, jset, entry, "bad rw")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
 	}
@@ -581,9 +597,10 @@ static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 static int journal_entry_dev_usage_validate(struct bch_fs *c,
-					    struct jset *jset,
-					    struct jset_entry *entry,
-					    unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct jset_entry_dev_usage *u =
 		container_of(entry, struct jset_entry_dev_usage, entry);
@@ -593,7 +610,7 @@ static int journal_entry_dev_usage_validate(struct bch_fs *c,
 	int ret = 0;
 
 	if (journal_entry_err_on(bytes < expected,
-				 c, jset, entry, "bad size (%u < %u)",
+				 c, version, jset, entry, "bad size (%u < %u)",
 				 bytes, expected)) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
@@ -602,13 +619,13 @@ static int journal_entry_dev_usage_validate(struct bch_fs *c,
 	dev = le32_to_cpu(u->dev);
 
 	if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
-				 c, jset, entry, "bad dev")) {
+				 c, version, jset, entry, "bad dev")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
 	}
 
 	if (journal_entry_err_on(u->pad,
-				 c, jset, entry, "bad pad")) {
+				 c, version, jset, entry, "bad pad")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
 	}
@@ -641,9 +658,10 @@ static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs
 }
 
 static int journal_entry_log_validate(struct bch_fs *c,
-				      struct jset *jset,
-				      struct jset_entry *entry,
-				      unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	return 0;
 }
@@ -658,9 +676,10 @@ static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 static int journal_entry_overwrite_validate(struct bch_fs *c,
-				      struct jset *jset,
-				      struct jset_entry *entry,
-				      unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	return journal_entry_btree_keys_validate(c, jset, entry,
 				version, big_endian, READ);
@@ -674,7 +693,8 @@ static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs
 
 struct jset_entry_ops {
 	int (*validate)(struct bch_fs *, struct jset *,
-			struct jset_entry *, unsigned, int, int);
+			struct jset_entry *, unsigned, int,
+			enum bkey_invalid_flags);
 	void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 };
 
@@ -691,11 +711,12 @@ static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 int bch2_journal_entry_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
-				unsigned version, int big_endian, int write)
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	return entry->type < BCH_JSET_ENTRY_NR
 		? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
-				version, big_endian, write)
+				version, big_endian, flags)
 		: 0;
 }
 
@@ -711,22 +732,22 @@ void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
-				 int write)
+				 enum bkey_invalid_flags flags)
 {
 	struct jset_entry *entry;
+	unsigned version = le32_to_cpu(jset->version);
 	int ret = 0;
 
 	vstruct_for_each(jset, entry) {
-		if (journal_entry_err_on(vstruct_next(entry) >
-					 vstruct_last(jset), c, jset, entry,
+		if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset),
+					 c, version, jset, entry,
 				"journal entry extends past end of jset")) {
 			jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 			break;
 		}
 
 		ret = bch2_journal_entry_validate(c, jset, entry,
-					le32_to_cpu(jset->version),
-					JSET_BIG_ENDIAN(jset), write);
+					version, JSET_BIG_ENDIAN(jset), flags);
 		if (ret)
 			break;
 	}
@@ -737,7 +758,7 @@ fsck_err:
 static int jset_validate(struct bch_fs *c,
 			 struct bch_dev *ca,
 			 struct jset *jset, u64 sector,
-			 int write)
+			 enum bkey_invalid_flags flags)
 {
 	unsigned version;
 	int ret = 0;
@@ -746,7 +767,8 @@ static int jset_validate(struct bch_fs *c,
 		return JOURNAL_ENTRY_NONE;
 
 	version = le32_to_cpu(jset->version);
-	if (journal_entry_err_on(!bch2_version_compatible(version), c, jset, NULL,
+	if (journal_entry_err_on(!bch2_version_compatible(version),
+			c, version, jset, NULL,
 			"%s sector %llu seq %llu: incompatible journal entry version %u.%u",
 			ca ? ca->name : c->name,
 			sector, le64_to_cpu(jset->seq),
@@ -757,7 +779,7 @@ static int jset_validate(struct bch_fs *c,
 	}
 
 	if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
-				 c, jset, NULL,
+				 c, version, jset, NULL,
 			"%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 			ca ? ca->name : c->name,
 			sector, le64_to_cpu(jset->seq),
@@ -767,7 +789,7 @@ static int jset_validate(struct bch_fs *c,
 	/* last_seq is ignored when JSET_NO_FLUSH is true */
 	if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 				 le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
-				 c, jset, NULL,
+				 c, version, jset, NULL,
 				 "invalid journal entry: last_seq > seq (%llu > %llu)",
 				 le64_to_cpu(jset->last_seq),
 				 le64_to_cpu(jset->seq))) {
@@ -775,7 +797,7 @@ static int jset_validate(struct bch_fs *c,
 		return JOURNAL_ENTRY_BAD;
 	}
 
-	ret = jset_validate_entries(c, jset, write);
+	ret = jset_validate_entries(c, jset, flags);
 fsck_err:
 	return ret;
 }
@@ -788,14 +810,15 @@ static int jset_validate_early(struct bch_fs *c,
 {
 	size_t bytes = vstruct_bytes(jset);
 	unsigned version;
-	int write = READ;
+	enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
 	int ret = 0;
 
 	if (le64_to_cpu(jset->magic) != jset_magic(c))
 		return JOURNAL_ENTRY_NONE;
 
 	version = le32_to_cpu(jset->version);
-	if (journal_entry_err_on(!bch2_version_compatible(version), c, jset, NULL,
+	if (journal_entry_err_on(!bch2_version_compatible(version),
+			 c, version, jset, NULL,
 			"%s sector %llu seq %llu: unknown journal entry version %u.%u",
 			ca ? ca->name : c->name,
 			sector, le64_to_cpu(jset->seq),
@@ -810,7 +833,7 @@ static int jset_validate_early(struct bch_fs *c,
 		return JOURNAL_ENTRY_REREAD;
 
 	if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
-				 c, jset, NULL,
+			 c, version, jset, NULL,
 			"%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 			ca ? ca->name : c->name,
 			sector, le64_to_cpu(jset->seq), bytes))
@@ -1127,7 +1150,7 @@ int bch2_journal_read(struct bch_fs *c,
 	 * those entries will be blacklisted:
 	 */
 	genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
-		int write = READ;
+		enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
 
 		i = *_i;
 
@@ -1149,7 +1172,7 @@ int bch2_journal_read(struct bch_fs *c,
 		}
 
 		if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
-					 c, &i->j, NULL,
+					 c, le32_to_cpu(i->j.version), &i->j, NULL,
 					 "invalid journal entry: last_seq > seq (%llu > %llu)",
 					 le64_to_cpu(i->j.last_seq),
 					 le64_to_cpu(i->j.seq)))
diff --git a/libbcachefs/journal_io.h b/libbcachefs/journal_io.h
index 8801e981..a88d097b 100644
--- a/libbcachefs/journal_io.h
+++ b/libbcachefs/journal_io.h
@@ -50,7 +50,8 @@ static inline struct jset_entry *__jset_entry_type_next(struct jset *jset,
 		jset_entry_for_each_key(entry, k)
 
 int bch2_journal_entry_validate(struct bch_fs *, struct jset *,
-				struct jset_entry *, unsigned, int, int);
+				struct jset_entry *, unsigned, int,
+				enum bkey_invalid_flags);
 void bch2_journal_entry_to_text(struct printbuf *, struct bch_fs *,
 				struct jset_entry *);
 
diff --git a/libbcachefs/journal_reclaim.c b/libbcachefs/journal_reclaim.c
index 8de83e10..7a3139b5 100644
--- a/libbcachefs/journal_reclaim.c
+++ b/libbcachefs/journal_reclaim.c
@@ -3,13 +3,14 @@
 #include "bcachefs.h"
 #include "btree_key_cache.h"
 #include "btree_update.h"
+#include "buckets.h"
 #include "errcode.h"
 #include "error.h"
 #include "journal.h"
 #include "journal_io.h"
 #include "journal_reclaim.h"
 #include "replicas.h"
-#include "super.h"
+#include "sb-members.h"
 #include "trace.h"
 
 #include <linux/kthread.h>
diff --git a/libbcachefs/movinggc.c b/libbcachefs/movinggc.c
index 5242f20b..256431a6 100644
--- a/libbcachefs/movinggc.c
+++ b/libbcachefs/movinggc.c
@@ -220,8 +220,10 @@ static int bch2_copygc(struct btree_trans *trans,
 
 		f = move_bucket_in_flight_add(buckets_in_flight, *i);
 		ret = PTR_ERR_OR_ZERO(f);
-		if (ret == -EEXIST) /* rare race: copygc_get_buckets returned same bucket more than once */
+		if (ret == -EEXIST) { /* rare race: copygc_get_buckets returned same bucket more than once */
+			ret = 0;
 			continue;
+		}
 		if (ret == -ENOMEM) { /* flush IO, continue later */
 			ret = 0;
 			break;
diff --git a/libbcachefs/recovery.c b/libbcachefs/recovery.c
index 55a233c2..cb2186a7 100644
--- a/libbcachefs/recovery.c
+++ b/libbcachefs/recovery.c
@@ -5,6 +5,7 @@
 #include "bkey_buf.h"
 #include "alloc_background.h"
 #include "btree_gc.h"
+#include "btree_journal_iter.h"
 #include "btree_update.h"
 #include "btree_update_interior.h"
 #include "btree_io.h"
@@ -23,6 +24,7 @@
 #include "quota.h"
 #include "recovery.h"
 #include "replicas.h"
+#include "sb-clean.h"
 #include "subvolume.h"
 #include "super-io.h"
 
@@ -57,524 +59,6 @@ static void zero_out_btree_mem_ptr(struct journal_keys *keys)
 			bkey_i_to_btree_ptr_v2(i->k)->v.mem_ptr = 0;
 }
 
-/* iterate over keys read from the journal: */
-
-static int __journal_key_cmp(enum btree_id	l_btree_id,
-			     unsigned		l_level,
-			     struct bpos	l_pos,
-			     const struct journal_key *r)
-{
-	return (cmp_int(l_btree_id,	r->btree_id) ?:
-		cmp_int(l_level,	r->level) ?:
-		bpos_cmp(l_pos,	r->k->k.p));
-}
-
-static int journal_key_cmp(const struct journal_key *l, const struct journal_key *r)
-{
-	return __journal_key_cmp(l->btree_id, l->level, l->k->k.p, r);
-}
-
-static inline size_t idx_to_pos(struct journal_keys *keys, size_t idx)
-{
-	size_t gap_size = keys->size - keys->nr;
-
-	if (idx >= keys->gap)
-		idx += gap_size;
-	return idx;
-}
-
-static inline struct journal_key *idx_to_key(struct journal_keys *keys, size_t idx)
-{
-	return keys->d + idx_to_pos(keys, idx);
-}
-
-static size_t __bch2_journal_key_search(struct journal_keys *keys,
-					enum btree_id id, unsigned level,
-					struct bpos pos)
-{
-	size_t l = 0, r = keys->nr, m;
-
-	while (l < r) {
-		m = l + ((r - l) >> 1);
-		if (__journal_key_cmp(id, level, pos, idx_to_key(keys, m)) > 0)
-			l = m + 1;
-		else
-			r = m;
-	}
-
-	BUG_ON(l < keys->nr &&
-	       __journal_key_cmp(id, level, pos, idx_to_key(keys, l)) > 0);
-
-	BUG_ON(l &&
-	       __journal_key_cmp(id, level, pos, idx_to_key(keys, l - 1)) <= 0);
-
-	return l;
-}
-
-static size_t bch2_journal_key_search(struct journal_keys *keys,
-				      enum btree_id id, unsigned level,
-				      struct bpos pos)
-{
-	return idx_to_pos(keys, __bch2_journal_key_search(keys, id, level, pos));
-}
-
-struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *c, enum btree_id btree_id,
-					   unsigned level, struct bpos pos,
-					   struct bpos end_pos, size_t *idx)
-{
-	struct journal_keys *keys = &c->journal_keys;
-	unsigned iters = 0;
-	struct journal_key *k;
-search:
-	if (!*idx)
-		*idx = __bch2_journal_key_search(keys, btree_id, level, pos);
-
-	while ((k = *idx < keys->nr ? idx_to_key(keys, *idx) : NULL)) {
-		if (__journal_key_cmp(btree_id, level, end_pos, k) < 0)
-			return NULL;
-
-		if (__journal_key_cmp(btree_id, level, pos, k) <= 0 &&
-		    !k->overwritten)
-			return k->k;
-
-		(*idx)++;
-		iters++;
-		if (iters == 10) {
-			*idx = 0;
-			goto search;
-		}
-	}
-
-	return NULL;
-}
-
-struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *c, enum btree_id btree_id,
-					   unsigned level, struct bpos pos)
-{
-	size_t idx = 0;
-
-	return bch2_journal_keys_peek_upto(c, btree_id, level, pos, pos, &idx);
-}
-
-static void journal_iters_fix(struct bch_fs *c)
-{
-	struct journal_keys *keys = &c->journal_keys;
-	/* The key we just inserted is immediately before the gap: */
-	size_t gap_end = keys->gap + (keys->size - keys->nr);
-	struct btree_and_journal_iter *iter;
-
-	/*
-	 * If an iterator points one after the key we just inserted, decrement
-	 * the iterator so it points at the key we just inserted - if the
-	 * decrement was unnecessary, bch2_btree_and_journal_iter_peek() will
-	 * handle that:
-	 */
-	list_for_each_entry(iter, &c->journal_iters, journal.list)
-		if (iter->journal.idx == gap_end)
-			iter->journal.idx = keys->gap - 1;
-}
-
-static void journal_iters_move_gap(struct bch_fs *c, size_t old_gap, size_t new_gap)
-{
-	struct journal_keys *keys = &c->journal_keys;
-	struct journal_iter *iter;
-	size_t gap_size = keys->size - keys->nr;
-
-	list_for_each_entry(iter, &c->journal_iters, list) {
-		if (iter->idx > old_gap)
-			iter->idx -= gap_size;
-		if (iter->idx >= new_gap)
-			iter->idx += gap_size;
-	}
-}
-
-int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
-				 unsigned level, struct bkey_i *k)
-{
-	struct journal_key n = {
-		.btree_id	= id,
-		.level		= level,
-		.k		= k,
-		.allocated	= true,
-		/*
-		 * Ensure these keys are done last by journal replay, to unblock
-		 * journal reclaim:
-		 */
-		.journal_seq	= U32_MAX,
-	};
-	struct journal_keys *keys = &c->journal_keys;
-	size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
-
-	BUG_ON(test_bit(BCH_FS_RW, &c->flags));
-
-	if (idx < keys->size &&
-	    journal_key_cmp(&n, &keys->d[idx]) == 0) {
-		if (keys->d[idx].allocated)
-			kfree(keys->d[idx].k);
-		keys->d[idx] = n;
-		return 0;
-	}
-
-	if (idx > keys->gap)
-		idx -= keys->size - keys->nr;
-
-	if (keys->nr == keys->size) {
-		struct journal_keys new_keys = {
-			.nr			= keys->nr,
-			.size			= max_t(size_t, keys->size, 8) * 2,
-		};
-
-		new_keys.d = kvmalloc_array(new_keys.size, sizeof(new_keys.d[0]), GFP_KERNEL);
-		if (!new_keys.d) {
-			bch_err(c, "%s: error allocating new key array (size %zu)",
-				__func__, new_keys.size);
-			return -BCH_ERR_ENOMEM_journal_key_insert;
-		}
-
-		/* Since @keys was full, there was no gap: */
-		memcpy(new_keys.d, keys->d, sizeof(keys->d[0]) * keys->nr);
-		kvfree(keys->d);
-		*keys = new_keys;
-
-		/* And now the gap is at the end: */
-		keys->gap = keys->nr;
-	}
-
-	journal_iters_move_gap(c, keys->gap, idx);
-
-	move_gap(keys->d, keys->nr, keys->size, keys->gap, idx);
-	keys->gap = idx;
-
-	keys->nr++;
-	keys->d[keys->gap++] = n;
-
-	journal_iters_fix(c);
-
-	return 0;
-}
-
-/*
- * Can only be used from the recovery thread while we're still RO - can't be
- * used once we've got RW, as journal_keys is at that point used by multiple
- * threads:
- */
-int bch2_journal_key_insert(struct bch_fs *c, enum btree_id id,
-			    unsigned level, struct bkey_i *k)
-{
-	struct bkey_i *n;
-	int ret;
-
-	n = kmalloc(bkey_bytes(&k->k), GFP_KERNEL);
-	if (!n)
-		return -BCH_ERR_ENOMEM_journal_key_insert;
-
-	bkey_copy(n, k);
-	ret = bch2_journal_key_insert_take(c, id, level, n);
-	if (ret)
-		kfree(n);
-	return ret;
-}
-
-int bch2_journal_key_delete(struct bch_fs *c, enum btree_id id,
-			    unsigned level, struct bpos pos)
-{
-	struct bkey_i whiteout;
-
-	bkey_init(&whiteout.k);
-	whiteout.k.p = pos;
-
-	return bch2_journal_key_insert(c, id, level, &whiteout);
-}
-
-void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
-				  unsigned level, struct bpos pos)
-{
-	struct journal_keys *keys = &c->journal_keys;
-	size_t idx = bch2_journal_key_search(keys, btree, level, pos);
-
-	if (idx < keys->size &&
-	    keys->d[idx].btree_id	== btree &&
-	    keys->d[idx].level		== level &&
-	    bpos_eq(keys->d[idx].k->k.p, pos))
-		keys->d[idx].overwritten = true;
-}
-
-static void bch2_journal_iter_advance(struct journal_iter *iter)
-{
-	if (iter->idx < iter->keys->size) {
-		iter->idx++;
-		if (iter->idx == iter->keys->gap)
-			iter->idx += iter->keys->size - iter->keys->nr;
-	}
-}
-
-static struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
-{
-	struct journal_key *k = iter->keys->d + iter->idx;
-
-	while (k < iter->keys->d + iter->keys->size &&
-	       k->btree_id	== iter->btree_id &&
-	       k->level		== iter->level) {
-		if (!k->overwritten)
-			return bkey_i_to_s_c(k->k);
-
-		bch2_journal_iter_advance(iter);
-		k = iter->keys->d + iter->idx;
-	}
-
-	return bkey_s_c_null;
-}
-
-static void bch2_journal_iter_exit(struct journal_iter *iter)
-{
-	list_del(&iter->list);
-}
-
-static void bch2_journal_iter_init(struct bch_fs *c,
-				   struct journal_iter *iter,
-				   enum btree_id id, unsigned level,
-				   struct bpos pos)
-{
-	iter->btree_id	= id;
-	iter->level	= level;
-	iter->keys	= &c->journal_keys;
-	iter->idx	= bch2_journal_key_search(&c->journal_keys, id, level, pos);
-}
-
-static struct bkey_s_c bch2_journal_iter_peek_btree(struct btree_and_journal_iter *iter)
-{
-	return bch2_btree_node_iter_peek_unpack(&iter->node_iter,
-						iter->b, &iter->unpacked);
-}
-
-static void bch2_journal_iter_advance_btree(struct btree_and_journal_iter *iter)
-{
-	bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
-}
-
-void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
-{
-	if (bpos_eq(iter->pos, SPOS_MAX))
-		iter->at_end = true;
-	else
-		iter->pos = bpos_successor(iter->pos);
-}
-
-struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *iter)
-{
-	struct bkey_s_c btree_k, journal_k, ret;
-again:
-	if (iter->at_end)
-		return bkey_s_c_null;
-
-	while ((btree_k = bch2_journal_iter_peek_btree(iter)).k &&
-	       bpos_lt(btree_k.k->p, iter->pos))
-		bch2_journal_iter_advance_btree(iter);
-
-	while ((journal_k = bch2_journal_iter_peek(&iter->journal)).k &&
-	       bpos_lt(journal_k.k->p, iter->pos))
-		bch2_journal_iter_advance(&iter->journal);
-
-	ret = journal_k.k &&
-		(!btree_k.k || bpos_le(journal_k.k->p, btree_k.k->p))
-		? journal_k
-		: btree_k;
-
-	if (ret.k && iter->b && bpos_gt(ret.k->p, iter->b->data->max_key))
-		ret = bkey_s_c_null;
-
-	if (ret.k) {
-		iter->pos = ret.k->p;
-		if (bkey_deleted(ret.k)) {
-			bch2_btree_and_journal_iter_advance(iter);
-			goto again;
-		}
-	} else {
-		iter->pos = SPOS_MAX;
-		iter->at_end = true;
-	}
-
-	return ret;
-}
-
-void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *iter)
-{
-	bch2_journal_iter_exit(&iter->journal);
-}
-
-void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
-						  struct bch_fs *c,
-						  struct btree *b,
-						  struct btree_node_iter node_iter,
-						  struct bpos pos)
-{
-	memset(iter, 0, sizeof(*iter));
-
-	iter->b = b;
-	iter->node_iter = node_iter;
-	bch2_journal_iter_init(c, &iter->journal, b->c.btree_id, b->c.level, pos);
-	INIT_LIST_HEAD(&iter->journal.list);
-	iter->pos = b->data->min_key;
-	iter->at_end = false;
-}
-
-/*
- * this version is used by btree_gc before filesystem has gone RW and
- * multithreaded, so uses the journal_iters list:
- */
-void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
-						struct bch_fs *c,
-						struct btree *b)
-{
-	struct btree_node_iter node_iter;
-
-	bch2_btree_node_iter_init_from_start(&node_iter, b);
-	__bch2_btree_and_journal_iter_init_node_iter(iter, c, b, node_iter, b->data->min_key);
-	list_add(&iter->journal.list, &c->journal_iters);
-}
-
-/* sort and dedup all keys in the journal: */
-
-void bch2_journal_entries_free(struct bch_fs *c)
-{
-	struct journal_replay **i;
-	struct genradix_iter iter;
-
-	genradix_for_each(&c->journal_entries, iter, i)
-		if (*i)
-			kvpfree(*i, offsetof(struct journal_replay, j) +
-				vstruct_bytes(&(*i)->j));
-	genradix_free(&c->journal_entries);
-}
-
-/*
- * When keys compare equal, oldest compares first:
- */
-static int journal_sort_key_cmp(const void *_l, const void *_r)
-{
-	const struct journal_key *l = _l;
-	const struct journal_key *r = _r;
-
-	return  journal_key_cmp(l, r) ?:
-		cmp_int(l->journal_seq, r->journal_seq) ?:
-		cmp_int(l->journal_offset, r->journal_offset);
-}
-
-void bch2_journal_keys_free(struct journal_keys *keys)
-{
-	struct journal_key *i;
-
-	move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
-	keys->gap = keys->nr;
-
-	for (i = keys->d; i < keys->d + keys->nr; i++)
-		if (i->allocated)
-			kfree(i->k);
-
-	kvfree(keys->d);
-	keys->d = NULL;
-	keys->nr = keys->gap = keys->size = 0;
-}
-
-static void __journal_keys_sort(struct journal_keys *keys)
-{
-	struct journal_key *src, *dst;
-
-	sort(keys->d, keys->nr, sizeof(keys->d[0]), journal_sort_key_cmp, NULL);
-
-	src = dst = keys->d;
-	while (src < keys->d + keys->nr) {
-		while (src + 1 < keys->d + keys->nr &&
-		       src[0].btree_id	== src[1].btree_id &&
-		       src[0].level	== src[1].level &&
-		       bpos_eq(src[0].k->k.p, src[1].k->k.p))
-			src++;
-
-		*dst++ = *src++;
-	}
-
-	keys->nr = dst - keys->d;
-}
-
-static int journal_keys_sort(struct bch_fs *c)
-{
-	struct genradix_iter iter;
-	struct journal_replay *i, **_i;
-	struct jset_entry *entry;
-	struct bkey_i *k;
-	struct journal_keys *keys = &c->journal_keys;
-	size_t nr_keys = 0, nr_read = 0;
-
-	genradix_for_each(&c->journal_entries, iter, _i) {
-		i = *_i;
-
-		if (!i || i->ignore)
-			continue;
-
-		for_each_jset_key(k, entry, &i->j)
-			nr_keys++;
-	}
-
-	if (!nr_keys)
-		return 0;
-
-	keys->size = roundup_pow_of_two(nr_keys);
-
-	keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
-	if (!keys->d) {
-		bch_err(c, "Failed to allocate buffer for sorted journal keys (%zu keys); trying slowpath",
-			nr_keys);
-
-		do {
-			keys->size >>= 1;
-			keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
-		} while (!keys->d && keys->size > nr_keys / 8);
-
-		if (!keys->d) {
-			bch_err(c, "Failed to allocate %zu size buffer for sorted journal keys; exiting",
-				keys->size);
-			return -BCH_ERR_ENOMEM_journal_keys_sort;
-		}
-	}
-
-	genradix_for_each(&c->journal_entries, iter, _i) {
-		i = *_i;
-
-		if (!i || i->ignore)
-			continue;
-
-		cond_resched();
-
-		for_each_jset_key(k, entry, &i->j) {
-			if (keys->nr == keys->size) {
-				__journal_keys_sort(keys);
-
-				if (keys->nr > keys->size * 7 / 8) {
-					bch_err(c, "Too many journal keys for slowpath; have %zu compacted, buf size %zu, processed %zu/%zu",
-						keys->nr, keys->size, nr_read, nr_keys);
-					return -BCH_ERR_ENOMEM_journal_keys_sort;
-				}
-			}
-
-			keys->d[keys->nr++] = (struct journal_key) {
-				.btree_id	= entry->btree_id,
-				.level		= entry->level,
-				.k		= k,
-				.journal_seq	= le64_to_cpu(i->j.seq),
-				.journal_offset	= k->_data - i->j._data,
-			};
-
-			nr_read++;
-		}
-	}
-
-	__journal_keys_sort(keys);
-	keys->gap = keys->nr;
-
-	bch_verbose(c, "Journal keys: %zu read, %zu after sorting and compacting", nr_keys, keys->nr);
-	return 0;
-}
-
 /* journal replay: */
 
 static void replay_now_at(struct journal *j, u64 seq)
@@ -846,134 +330,6 @@ static int journal_replay_early(struct bch_fs *c,
 
 /* sb clean section: */
 
-static struct bkey_i *btree_root_find(struct bch_fs *c,
-				      struct bch_sb_field_clean *clean,
-				      struct jset *j,
-				      enum btree_id id, unsigned *level)
-{
-	struct bkey_i *k;
-	struct jset_entry *entry, *start, *end;
-
-	if (clean) {
-		start = clean->start;
-		end = vstruct_end(&clean->field);
-	} else {
-		start = j->start;
-		end = vstruct_last(j);
-	}
-
-	for (entry = start; entry < end; entry = vstruct_next(entry))
-		if (entry->type == BCH_JSET_ENTRY_btree_root &&
-		    entry->btree_id == id)
-			goto found;
-
-	return NULL;
-found:
-	if (!entry->u64s)
-		return ERR_PTR(-EINVAL);
-
-	k = entry->start;
-	*level = entry->level;
-	return k;
-}
-
-static int verify_superblock_clean(struct bch_fs *c,
-				   struct bch_sb_field_clean **cleanp,
-				   struct jset *j)
-{
-	unsigned i;
-	struct bch_sb_field_clean *clean = *cleanp;
-	struct printbuf buf1 = PRINTBUF;
-	struct printbuf buf2 = PRINTBUF;
-	int ret = 0;
-
-	if (mustfix_fsck_err_on(j->seq != clean->journal_seq, c,
-			"superblock journal seq (%llu) doesn't match journal (%llu) after clean shutdown",
-			le64_to_cpu(clean->journal_seq),
-			le64_to_cpu(j->seq))) {
-		kfree(clean);
-		*cleanp = NULL;
-		return 0;
-	}
-
-	for (i = 0; i < BTREE_ID_NR; i++) {
-		struct bkey_i *k1, *k2;
-		unsigned l1 = 0, l2 = 0;
-
-		k1 = btree_root_find(c, clean, NULL, i, &l1);
-		k2 = btree_root_find(c, NULL, j, i, &l2);
-
-		if (!k1 && !k2)
-			continue;
-
-		printbuf_reset(&buf1);
-		printbuf_reset(&buf2);
-
-		if (k1)
-			bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(k1));
-		else
-			prt_printf(&buf1, "(none)");
-
-		if (k2)
-			bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(k2));
-		else
-			prt_printf(&buf2, "(none)");
-
-		mustfix_fsck_err_on(!k1 || !k2 ||
-				    IS_ERR(k1) ||
-				    IS_ERR(k2) ||
-				    k1->k.u64s != k2->k.u64s ||
-				    memcmp(k1, k2, bkey_bytes(&k1->k)) ||
-				    l1 != l2, c,
-			"superblock btree root %u doesn't match journal after clean shutdown\n"
-			"sb:      l=%u %s\n"
-			"journal: l=%u %s\n", i,
-			l1, buf1.buf,
-			l2, buf2.buf);
-	}
-fsck_err:
-	printbuf_exit(&buf2);
-	printbuf_exit(&buf1);
-	return ret;
-}
-
-static struct bch_sb_field_clean *read_superblock_clean(struct bch_fs *c)
-{
-	struct bch_sb_field_clean *clean, *sb_clean;
-	int ret;
-
-	mutex_lock(&c->sb_lock);
-	sb_clean = bch2_sb_get_clean(c->disk_sb.sb);
-
-	if (fsck_err_on(!sb_clean, c,
-			"superblock marked clean but clean section not present")) {
-		SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
-		c->sb.clean = false;
-		mutex_unlock(&c->sb_lock);
-		return NULL;
-	}
-
-	clean = kmemdup(sb_clean, vstruct_bytes(&sb_clean->field),
-			GFP_KERNEL);
-	if (!clean) {
-		mutex_unlock(&c->sb_lock);
-		return ERR_PTR(-BCH_ERR_ENOMEM_read_superblock_clean);
-	}
-
-	ret = bch2_sb_clean_validate_late(c, clean, READ);
-	if (ret) {
-		mutex_unlock(&c->sb_lock);
-		return ERR_PTR(ret);
-	}
-
-	mutex_unlock(&c->sb_lock);
-
-	return clean;
-fsck_err:
-	mutex_unlock(&c->sb_lock);
-	return ERR_PTR(ret);
-}
-
 static bool btree_id_is_alloc(enum btree_id id)
 {
 	switch (id) {
@@ -1120,6 +476,35 @@ static int bch2_fs_upgrade_for_subvolumes(struct bch_fs *c)
 	return ret;
 }
 
+const char * const bch2_recovery_passes[] = {
+#define x(_fn, _when)	#_fn,
+	BCH_RECOVERY_PASSES()
+#undef x
+	NULL
+};
+
+static int bch2_check_allocations(struct bch_fs *c)
+{
+	return bch2_gc(c, true, c->opts.norecovery);
+}
+
+static int bch2_set_may_go_rw(struct bch_fs *c)
+{
+	set_bit(BCH_FS_MAY_GO_RW, &c->flags);
+	return 0;
+}
+
+struct recovery_pass_fn {
+	int		(*fn)(struct bch_fs *);
+	unsigned	when;
+};
+
+static struct recovery_pass_fn recovery_pass_fns[] = {
+#define x(_fn, _when)	{ .fn = bch2_##_fn, .when = _when },
+	BCH_RECOVERY_PASSES()
+#undef x
+};
+
 static void check_version_upgrade(struct bch_fs *c)
 {
 	unsigned latest_compatible = bch2_version_compatible(c->sb.version);
@@ -1172,7 +557,12 @@ static void check_version_upgrade(struct bch_fs *c)
 
 		recovery_passes = bch2_upgrade_recovery_passes(c, old_version, new_version);
 		if (recovery_passes) {
-			prt_str(&buf, "fsck required");
+			if ((recovery_passes & RECOVERY_PASS_ALL_FSCK) == RECOVERY_PASS_ALL_FSCK)
+				prt_str(&buf, "fsck required");
+			else {
+				prt_str(&buf, "running recovery passses: ");
+				prt_bitflags(&buf, bch2_recovery_passes, recovery_passes);
+			}
 
 			c->recovery_passes_explicit |= recovery_passes;
 			c->opts.fix_errors = FSCK_FIX_yes;
@@ -1188,42 +578,19 @@ static void check_version_upgrade(struct bch_fs *c)
 	}
 }
 
-static int bch2_check_allocations(struct bch_fs *c)
-{
-	return bch2_gc(c, true, c->opts.norecovery);
-}
-
-static int bch2_set_may_go_rw(struct bch_fs *c)
-{
-	set_bit(BCH_FS_MAY_GO_RW, &c->flags);
-	return 0;
-}
-
-struct recovery_pass_fn {
-	int		(*fn)(struct bch_fs *);
-	const char	*name;
-	unsigned	when;
-};
-
-static struct recovery_pass_fn recovery_passes[] = {
-#define x(_fn, _when)	{ .fn = bch2_##_fn, .name = #_fn, .when = _when },
-	BCH_RECOVERY_PASSES()
-#undef x
-};
-
 u64 bch2_fsck_recovery_passes(void)
 {
 	u64 ret = 0;
 
-	for (unsigned i = 0; i < ARRAY_SIZE(recovery_passes); i++)
-		if (recovery_passes[i].when & PASS_FSCK)
+	for (unsigned i = 0; i < ARRAY_SIZE(recovery_pass_fns); i++)
+		if (recovery_pass_fns[i].when & PASS_FSCK)
 			ret |= BIT_ULL(i);
 	return ret;
 }
 
 static bool should_run_recovery_pass(struct bch_fs *c, enum bch_recovery_pass pass)
 {
-	struct recovery_pass_fn *p = recovery_passes + c->curr_recovery_pass;
+	struct recovery_pass_fn *p = recovery_pass_fns + c->curr_recovery_pass;
 
 	if (c->opts.norecovery && pass > BCH_RECOVERY_PASS_snapshots_read)
 		return false;
@@ -1245,15 +612,18 @@ static int bch2_run_recovery_pass(struct bch_fs *c, enum bch_recovery_pass pass)
 	c->curr_recovery_pass = pass;
 
 	if (should_run_recovery_pass(c, pass)) {
-		struct recovery_pass_fn *p = recovery_passes + pass;
+		struct recovery_pass_fn *p = recovery_pass_fns + pass;
 
 		if (!(p->when & PASS_SILENT))
-			printk(KERN_INFO bch2_log_msg(c, "%s..."), p->name);
+			printk(KERN_INFO bch2_log_msg(c, "%s..."),
+			       bch2_recovery_passes[pass]);
 		ret = p->fn(c);
 		if (ret)
 			return ret;
 		if (!(p->when & PASS_SILENT))
 			printk(KERN_CONT " done\n");
+
+		c->recovery_passes_complete |= BIT_ULL(pass);
 	}
 
 	return 0;
@@ -1263,7 +633,7 @@ static int bch2_run_recovery_passes(struct bch_fs *c)
 {
 	int ret = 0;
 
-	while (c->curr_recovery_pass < ARRAY_SIZE(recovery_passes)) {
+	while (c->curr_recovery_pass < ARRAY_SIZE(recovery_pass_fns)) {
 		ret = bch2_run_recovery_pass(c, c->curr_recovery_pass);
 		if (bch2_err_matches(ret, BCH_ERR_restart_recovery))
 			continue;
@@ -1283,17 +653,17 @@ int bch2_fs_recovery(struct bch_fs *c)
 	bool write_sb = false;
 	int ret = 0;
 
-	if (c->sb.clean)
-		clean = read_superblock_clean(c);
-	ret = PTR_ERR_OR_ZERO(clean);
-	if (ret)
-		goto err;
+	if (c->sb.clean) {
+		clean = bch2_read_superblock_clean(c);
+		ret = PTR_ERR_OR_ZERO(clean);
+		if (ret)
+			goto err;
 
-	if (c->sb.clean)
 		bch_info(c, "recovering from clean shutdown, journal seq %llu",
 			 le64_to_cpu(clean->journal_seq));
-	else
+	} else {
 		bch_info(c, "recovering from unclean shutdown");
+	}
 
 	if (!(c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))) {
 		bch_err(c, "feature new_extent_overwrite not set, filesystem no longer supported");
@@ -1308,12 +678,6 @@ int bch2_fs_recovery(struct bch_fs *c)
 		goto err;
 	}
 
-	if (!(c->sb.compat & (1ULL << BCH_COMPAT_bformat_overflow_done))) {
-		bch_err(c, "filesystem may have incompatible bkey formats; run fsck from the compat branch to fix");
-		ret = -EINVAL;
-		goto err;
-	}
-
 	if (c->opts.fsck || !(c->opts.nochanges && c->opts.norecovery))
 		check_version_upgrade(c);
 
@@ -1373,12 +737,12 @@ int bch2_fs_recovery(struct bch_fs *c)
 				}
 		}
 
-		ret = journal_keys_sort(c);
+		ret = bch2_journal_keys_sort(c);
 		if (ret)
 			goto err;
 
 		if (c->sb.clean && last_journal_entry) {
-			ret = verify_superblock_clean(c, &clean,
+			ret = bch2_verify_superblock_clean(c, &clean,
 						      last_journal_entry);
 			if (ret)
 				goto err;
@@ -1513,7 +877,6 @@ use_clean:
 	mutex_unlock(&c->sb_lock);
 
 	if (!(c->sb.compat & (1ULL << BCH_COMPAT_extents_above_btree_updates_done)) ||
-	    !(c->sb.compat & (1ULL << BCH_COMPAT_bformat_overflow_done)) ||
 	    c->sb.version_min < bcachefs_metadata_version_btree_ptr_sectors_written) {
 		struct bch_move_stats stats;
 
@@ -1581,7 +944,7 @@ int bch2_fs_initialize(struct bch_fs *c)
 	}
 	mutex_unlock(&c->sb_lock);
 
-	c->curr_recovery_pass = ARRAY_SIZE(recovery_passes);
+	c->curr_recovery_pass = ARRAY_SIZE(recovery_pass_fns);
 	set_bit(BCH_FS_MAY_GO_RW, &c->flags);
 	set_bit(BCH_FS_FSCK_DONE, &c->flags);
 
diff --git a/libbcachefs/recovery.h b/libbcachefs/recovery.h
index f8e796c0..852d3056 100644
--- a/libbcachefs/recovery.h
+++ b/libbcachefs/recovery.h
@@ -2,55 +2,28 @@
 #ifndef _BCACHEFS_RECOVERY_H
 #define _BCACHEFS_RECOVERY_H
 
-struct journal_iter {
-	struct list_head	list;
-	enum btree_id		btree_id;
-	unsigned		level;
-	size_t			idx;
-	struct journal_keys	*keys;
-};
+extern const char * const bch2_recovery_passes[];
 
 /*
- * Iterate over keys in the btree, with keys from the journal overlaid on top:
+ * For when we need to rewind recovery passes and run a pass we skipped:
  */
-
-struct btree_and_journal_iter {
-	struct btree		*b;
-	struct btree_node_iter	node_iter;
-	struct bkey		unpacked;
-
-	struct journal_iter	journal;
-	struct bpos		pos;
-	bool			at_end;
-};
-
-struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *, enum btree_id,
-				unsigned, struct bpos, struct bpos, size_t *);
-struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *, enum btree_id,
-					   unsigned, struct bpos);
-
-int bch2_journal_key_insert_take(struct bch_fs *, enum btree_id,
-				 unsigned, struct bkey_i *);
-int bch2_journal_key_insert(struct bch_fs *, enum btree_id,
-			    unsigned, struct bkey_i *);
-int bch2_journal_key_delete(struct bch_fs *, enum btree_id,
-			    unsigned, struct bpos);
-void bch2_journal_key_overwritten(struct bch_fs *, enum btree_id,
-				  unsigned, struct bpos);
-
-void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *);
-struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *);
-
-void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *);
-void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *,
-				struct bch_fs *, struct btree *,
-				struct btree_node_iter, struct bpos);
-void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *,
-						struct bch_fs *,
-						struct btree *);
-
-void bch2_journal_keys_free(struct journal_keys *);
-void bch2_journal_entries_free(struct bch_fs *);
+static inline int bch2_run_explicit_recovery_pass(struct bch_fs *c,
+						  enum bch_recovery_pass pass)
+{
+	bch_info(c, "running explicit recovery pass %s (%u), currently at %s (%u)",
+		 bch2_recovery_passes[pass], pass,
+		 bch2_recovery_passes[c->curr_recovery_pass], c->curr_recovery_pass);
+
+	c->recovery_passes_explicit |= BIT_ULL(pass);
+
+	if (c->curr_recovery_pass >= pass) {
+		c->curr_recovery_pass = pass;
+		c->recovery_passes_complete &= (1ULL << pass) >> 1;
+		return -BCH_ERR_restart_recovery;
+	} else {
+		return 0;
+	}
+}
 
 u64 bch2_fsck_recovery_passes(void);
 
diff --git a/libbcachefs/sb-clean.c b/libbcachefs/sb-clean.c
new file mode 100644
index 00000000..a3695e56
--- /dev/null
+++ b/libbcachefs/sb-clean.c
@@ -0,0 +1,395 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "btree_update_interior.h"
+#include "buckets.h"
+#include "error.h"
+#include "journal_io.h"
+#include "replicas.h"
+#include "sb-clean.h"
+#include "super-io.h"
+
+/*
+ * BCH_SB_FIELD_clean:
+ *
+ * Btree roots, and a few other things, are recovered from the journal after an
+ * unclean shutdown - but after a clean shutdown, to avoid having to read the
+ * journal, we can store them in the superblock.
+ *
+ * bch_sb_field_clean simply contains a list of journal entries, stored exactly
+ * as they would be in the journal:
+ */
+
+int bch2_sb_clean_validate_late(struct bch_fs *c, struct bch_sb_field_clean *clean,
+				int write)
+{
+	struct jset_entry *entry;
+	int ret;
+
+	for (entry = clean->start;
+	     entry < (struct jset_entry *) vstruct_end(&clean->field);
+	     entry = vstruct_next(entry)) {
+		ret = bch2_journal_entry_validate(c, NULL, entry,
+						  le16_to_cpu(c->disk_sb.sb->version),
+						  BCH_SB_BIG_ENDIAN(c->disk_sb.sb),
+						  write);
+		if (ret)
+			return ret;
+	}
+
+	return 0;
+}
+
+static struct bkey_i *btree_root_find(struct bch_fs *c,
+				      struct bch_sb_field_clean *clean,
+				      struct jset *j,
+				      enum btree_id id, unsigned *level)
+{
+	struct bkey_i *k;
+	struct jset_entry *entry, *start, *end;
+
+	if (clean) {
+		start = clean->start;
+		end = vstruct_end(&clean->field);
+	} else {
+		start = j->start;
+		end = vstruct_last(j);
+	}
+
+	for (entry = start; entry < end; entry = vstruct_next(entry))
+		if (entry->type == BCH_JSET_ENTRY_btree_root &&
+		    entry->btree_id == id)
+			goto found;
+
+	return NULL;
+found:
+	if (!entry->u64s)
+		return ERR_PTR(-EINVAL);
+
+	k = entry->start;
+	*level = entry->level;
+	return k;
+}
+
+int bch2_verify_superblock_clean(struct bch_fs *c,
+				 struct bch_sb_field_clean **cleanp,
+				 struct jset *j)
+{
+	unsigned i;
+	struct bch_sb_field_clean *clean = *cleanp;
+	struct printbuf buf1 = PRINTBUF;
+	struct printbuf buf2 = PRINTBUF;
+	int ret = 0;
+
+	if (mustfix_fsck_err_on(j->seq != clean->journal_seq, c,
+			"superblock journal seq (%llu) doesn't match journal (%llu) after clean shutdown",
+			le64_to_cpu(clean->journal_seq),
+			le64_to_cpu(j->seq))) {
+		kfree(clean);
+		*cleanp = NULL;
+		return 0;
+	}
+
+	for (i = 0; i < BTREE_ID_NR; i++) {
+		struct bkey_i *k1, *k2;
+		unsigned l1 = 0, l2 = 0;
+
+		k1 = btree_root_find(c, clean, NULL, i, &l1);
+		k2 = btree_root_find(c, NULL, j, i, &l2);
+
+		if (!k1 && !k2)
+			continue;
+
+		printbuf_reset(&buf1);
+		printbuf_reset(&buf2);
+
+		if (k1)
+			bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(k1));
+		else
+			prt_printf(&buf1, "(none)");
+
+		if (k2)
+			bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(k2));
+		else
+			prt_printf(&buf2, "(none)");
+
+		mustfix_fsck_err_on(!k1 || !k2 ||
+				    IS_ERR(k1) ||
+				    IS_ERR(k2) ||
+				    k1->k.u64s != k2->k.u64s ||
+				    memcmp(k1, k2, bkey_bytes(&k1->k)) ||
+				    l1 != l2, c,
+			"superblock btree root %u doesn't match journal after clean shutdown\n"
+			"sb:      l=%u %s\n"
+			"journal: l=%u %s\n", i,
+			l1, buf1.buf,
+			l2, buf2.buf);
+	}
+fsck_err:
+	printbuf_exit(&buf2);
+	printbuf_exit(&buf1);
+	return ret;
+}
+
+struct bch_sb_field_clean *bch2_read_superblock_clean(struct bch_fs *c)
+{
+	struct bch_sb_field_clean *clean, *sb_clean;
+	int ret;
+
+	mutex_lock(&c->sb_lock);
+	sb_clean = bch2_sb_get_clean(c->disk_sb.sb);
+
+	if (fsck_err_on(!sb_clean, c,
+			"superblock marked clean but clean section not present")) {
+		SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
+		c->sb.clean = false;
+		mutex_unlock(&c->sb_lock);
+		return NULL;
+	}
+
+	clean = kmemdup(sb_clean, vstruct_bytes(&sb_clean->field),
+			GFP_KERNEL);
+	if (!clean) {
+		mutex_unlock(&c->sb_lock);
+		return ERR_PTR(-BCH_ERR_ENOMEM_read_superblock_clean);
+	}
+
+	ret = bch2_sb_clean_validate_late(c, clean, READ);
+	if (ret) {
+		mutex_unlock(&c->sb_lock);
+		return ERR_PTR(ret);
+	}
+
+	mutex_unlock(&c->sb_lock);
+
+	return clean;
+fsck_err:
+	mutex_unlock(&c->sb_lock);
+	return ERR_PTR(ret);
+}
+
+static struct jset_entry *jset_entry_init(struct jset_entry **end, size_t size)
+{
+	struct jset_entry *entry = *end;
+	unsigned u64s = DIV_ROUND_UP(size, sizeof(u64));
+
+	memset(entry, 0, u64s * sizeof(u64));
+	/*
+	 * The u64s field counts from the start of data, ignoring the shared
+	 * fields.
+	 */
+	entry->u64s = cpu_to_le16(u64s - 1);
+
+	*end = vstruct_next(*end);
+	return entry;
+}
+
+void bch2_journal_super_entries_add_common(struct bch_fs *c,
+					   struct jset_entry **end,
+					   u64 journal_seq)
+{
+	struct bch_dev *ca;
+	unsigned i, dev;
+
+	percpu_down_read(&c->mark_lock);
+
+	if (!journal_seq) {
+		for (i = 0; i < ARRAY_SIZE(c->usage); i++)
+			bch2_fs_usage_acc_to_base(c, i);
+	} else {
+		bch2_fs_usage_acc_to_base(c, journal_seq & JOURNAL_BUF_MASK);
+	}
+
+	{
+		struct jset_entry_usage *u =
+			container_of(jset_entry_init(end, sizeof(*u)),
+				     struct jset_entry_usage, entry);
+
+		u->entry.type	= BCH_JSET_ENTRY_usage;
+		u->entry.btree_id = BCH_FS_USAGE_inodes;
+		u->v		= cpu_to_le64(c->usage_base->nr_inodes);
+	}
+
+	{
+		struct jset_entry_usage *u =
+			container_of(jset_entry_init(end, sizeof(*u)),
+				     struct jset_entry_usage, entry);
+
+		u->entry.type	= BCH_JSET_ENTRY_usage;
+		u->entry.btree_id = BCH_FS_USAGE_key_version;
+		u->v		= cpu_to_le64(atomic64_read(&c->key_version));
+	}
+
+	for (i = 0; i < BCH_REPLICAS_MAX; i++) {
+		struct jset_entry_usage *u =
+			container_of(jset_entry_init(end, sizeof(*u)),
+				     struct jset_entry_usage, entry);
+
+		u->entry.type	= BCH_JSET_ENTRY_usage;
+		u->entry.btree_id = BCH_FS_USAGE_reserved;
+		u->entry.level	= i;
+		u->v		= cpu_to_le64(c->usage_base->persistent_reserved[i]);
+	}
+
+	for (i = 0; i < c->replicas.nr; i++) {
+		struct bch_replicas_entry *e =
+			cpu_replicas_entry(&c->replicas, i);
+		struct jset_entry_data_usage *u =
+			container_of(jset_entry_init(end, sizeof(*u) + e->nr_devs),
+				     struct jset_entry_data_usage, entry);
+
+		u->entry.type	= BCH_JSET_ENTRY_data_usage;
+		u->v		= cpu_to_le64(c->usage_base->replicas[i]);
+		unsafe_memcpy(&u->r, e, replicas_entry_bytes(e),
+			      "embedded variable length struct");
+	}
+
+	for_each_member_device(ca, c, dev) {
+		unsigned b = sizeof(struct jset_entry_dev_usage) +
+			sizeof(struct jset_entry_dev_usage_type) * BCH_DATA_NR;
+		struct jset_entry_dev_usage *u =
+			container_of(jset_entry_init(end, b),
+				     struct jset_entry_dev_usage, entry);
+
+		u->entry.type = BCH_JSET_ENTRY_dev_usage;
+		u->dev = cpu_to_le32(dev);
+		u->buckets_ec		= cpu_to_le64(ca->usage_base->buckets_ec);
+
+		for (i = 0; i < BCH_DATA_NR; i++) {
+			u->d[i].buckets = cpu_to_le64(ca->usage_base->d[i].buckets);
+			u->d[i].sectors	= cpu_to_le64(ca->usage_base->d[i].sectors);
+			u->d[i].fragmented = cpu_to_le64(ca->usage_base->d[i].fragmented);
+		}
+	}
+
+	percpu_up_read(&c->mark_lock);
+
+	for (i = 0; i < 2; i++) {
+		struct jset_entry_clock *clock =
+			container_of(jset_entry_init(end, sizeof(*clock)),
+				     struct jset_entry_clock, entry);
+
+		clock->entry.type = BCH_JSET_ENTRY_clock;
+		clock->rw	= i;
+		clock->time	= cpu_to_le64(atomic64_read(&c->io_clock[i].now));
+	}
+}
+
+static int bch2_sb_clean_validate(struct bch_sb *sb,
+				  struct bch_sb_field *f,
+				  struct printbuf *err)
+{
+	struct bch_sb_field_clean *clean = field_to_type(f, clean);
+
+	if (vstruct_bytes(&clean->field) < sizeof(*clean)) {
+		prt_printf(err, "wrong size (got %zu should be %zu)",
+		       vstruct_bytes(&clean->field), sizeof(*clean));
+		return -BCH_ERR_invalid_sb_clean;
+	}
+
+	return 0;
+}
+
+static void bch2_sb_clean_to_text(struct printbuf *out, struct bch_sb *sb,
+				  struct bch_sb_field *f)
+{
+	struct bch_sb_field_clean *clean = field_to_type(f, clean);
+	struct jset_entry *entry;
+
+	prt_printf(out, "flags:          %x",	le32_to_cpu(clean->flags));
+	prt_newline(out);
+	prt_printf(out, "journal_seq:    %llu",	le64_to_cpu(clean->journal_seq));
+	prt_newline(out);
+
+	for (entry = clean->start;
+	     entry != vstruct_end(&clean->field);
+	     entry = vstruct_next(entry)) {
+		if (entry->type == BCH_JSET_ENTRY_btree_keys &&
+		    !entry->u64s)
+			continue;
+
+		bch2_journal_entry_to_text(out, NULL, entry);
+		prt_newline(out);
+	}
+}
+
+const struct bch_sb_field_ops bch_sb_field_ops_clean = {
+	.validate	= bch2_sb_clean_validate,
+	.to_text	= bch2_sb_clean_to_text,
+};
+
+int bch2_fs_mark_dirty(struct bch_fs *c)
+{
+	int ret;
+
+	/*
+	 * Unconditionally write superblock, to verify it hasn't changed before
+	 * we go rw:
+	 */
+
+	mutex_lock(&c->sb_lock);
+	SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
+
+	bch2_sb_maybe_downgrade(c);
+	c->disk_sb.sb->features[0] |= cpu_to_le64(BCH_SB_FEATURES_ALWAYS);
+
+	ret = bch2_write_super(c);
+	mutex_unlock(&c->sb_lock);
+
+	return ret;
+}
+
+void bch2_fs_mark_clean(struct bch_fs *c)
+{
+	struct bch_sb_field_clean *sb_clean;
+	struct jset_entry *entry;
+	unsigned u64s;
+	int ret;
+
+	mutex_lock(&c->sb_lock);
+	if (BCH_SB_CLEAN(c->disk_sb.sb))
+		goto out;
+
+	SET_BCH_SB_CLEAN(c->disk_sb.sb, true);
+
+	c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_alloc_info);
+	c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_alloc_metadata);
+	c->disk_sb.sb->features[0] &= cpu_to_le64(~(1ULL << BCH_FEATURE_extents_above_btree_updates));
+	c->disk_sb.sb->features[0] &= cpu_to_le64(~(1ULL << BCH_FEATURE_btree_updates_journalled));
+
+	u64s = sizeof(*sb_clean) / sizeof(u64) + c->journal.entry_u64s_reserved;
+
+	sb_clean = bch2_sb_resize_clean(&c->disk_sb, u64s);
+	if (!sb_clean) {
+		bch_err(c, "error resizing superblock while setting filesystem clean");
+		goto out;
+	}
+
+	sb_clean->flags		= 0;
+	sb_clean->journal_seq	= cpu_to_le64(atomic64_read(&c->journal.seq));
+
+	/* Trying to catch outstanding bug: */
+	BUG_ON(le64_to_cpu(sb_clean->journal_seq) > S64_MAX);
+
+	entry = sb_clean->start;
+	bch2_journal_super_entries_add_common(c, &entry, 0);
+	entry = bch2_btree_roots_to_journal_entries(c, entry, entry);
+	BUG_ON((void *) entry > vstruct_end(&sb_clean->field));
+
+	memset(entry, 0,
+	       vstruct_end(&sb_clean->field) - (void *) entry);
+
+	/*
+	 * this should be in the write path, and we should be validating every
+	 * superblock section:
+	 */
+	ret = bch2_sb_clean_validate_late(c, sb_clean, WRITE);
+	if (ret) {
+		bch_err(c, "error writing marking filesystem clean: validate error");
+		goto out;
+	}
+
+	bch2_write_super(c);
+out:
+	mutex_unlock(&c->sb_lock);
+}
diff --git a/libbcachefs/sb-clean.h b/libbcachefs/sb-clean.h
new file mode 100644
index 00000000..71caef28
--- /dev/null
+++ b/libbcachefs/sb-clean.h
@@ -0,0 +1,16 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_SB_CLEAN_H
+#define _BCACHEFS_SB_CLEAN_H
+
+int bch2_sb_clean_validate_late(struct bch_fs *, struct bch_sb_field_clean *, int);
+int bch2_verify_superblock_clean(struct bch_fs *, struct bch_sb_field_clean **,
+				 struct jset *);
+struct bch_sb_field_clean *bch2_read_superblock_clean(struct bch_fs *);
+void bch2_journal_super_entries_add_common(struct bch_fs *, struct jset_entry **, u64);
+
+extern const struct bch_sb_field_ops bch_sb_field_ops_clean;
+
+int bch2_fs_mark_dirty(struct bch_fs *);
+void bch2_fs_mark_clean(struct bch_fs *);
+
+#endif /* _BCACHEFS_SB_CLEAN_H */
diff --git a/libbcachefs/sb-members.c b/libbcachefs/sb-members.c
new file mode 100644
index 00000000..16a2b338
--- /dev/null
+++ b/libbcachefs/sb-members.c
@@ -0,0 +1,173 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "disk_groups.h"
+#include "replicas.h"
+#include "sb-members.h"
+#include "super-io.h"
+
+/* Code for bch_sb_field_members: */
+
+static int bch2_sb_members_validate(struct bch_sb *sb,
+				    struct bch_sb_field *f,
+				    struct printbuf *err)
+{
+	struct bch_sb_field_members *mi = field_to_type(f, members);
+	unsigned i;
+
+	if ((void *) (mi->members + sb->nr_devices) >
+	    vstruct_end(&mi->field)) {
+		prt_printf(err, "too many devices for section size");
+		return -BCH_ERR_invalid_sb_members;
+	}
+
+	for (i = 0; i < sb->nr_devices; i++) {
+		struct bch_member *m = mi->members + i;
+
+		if (!bch2_member_exists(m))
+			continue;
+
+		if (le64_to_cpu(m->nbuckets) > LONG_MAX) {
+			prt_printf(err, "device %u: too many buckets (got %llu, max %lu)",
+			       i, le64_to_cpu(m->nbuckets), LONG_MAX);
+			return -BCH_ERR_invalid_sb_members;
+		}
+
+		if (le64_to_cpu(m->nbuckets) -
+		    le16_to_cpu(m->first_bucket) < BCH_MIN_NR_NBUCKETS) {
+			prt_printf(err, "device %u: not enough buckets (got %llu, max %u)",
+			       i, le64_to_cpu(m->nbuckets), BCH_MIN_NR_NBUCKETS);
+			return -BCH_ERR_invalid_sb_members;
+		}
+
+		if (le16_to_cpu(m->bucket_size) <
+		    le16_to_cpu(sb->block_size)) {
+			prt_printf(err, "device %u: bucket size %u smaller than block size %u",
+			       i, le16_to_cpu(m->bucket_size), le16_to_cpu(sb->block_size));
+			return -BCH_ERR_invalid_sb_members;
+		}
+
+		if (le16_to_cpu(m->bucket_size) <
+		    BCH_SB_BTREE_NODE_SIZE(sb)) {
+			prt_printf(err, "device %u: bucket size %u smaller than btree node size %llu",
+			       i, le16_to_cpu(m->bucket_size), BCH_SB_BTREE_NODE_SIZE(sb));
+			return -BCH_ERR_invalid_sb_members;
+		}
+	}
+
+	return 0;
+}
+
+static void bch2_sb_members_to_text(struct printbuf *out, struct bch_sb *sb,
+				    struct bch_sb_field *f)
+{
+	struct bch_sb_field_members *mi = field_to_type(f, members);
+	struct bch_sb_field_disk_groups *gi = bch2_sb_get_disk_groups(sb);
+	unsigned i;
+
+	for (i = 0; i < sb->nr_devices; i++) {
+		struct bch_member *m = mi->members + i;
+		unsigned data_have = bch2_sb_dev_has_data(sb, i);
+		u64 bucket_size = le16_to_cpu(m->bucket_size);
+		u64 device_size = le64_to_cpu(m->nbuckets) * bucket_size;
+
+		if (!bch2_member_exists(m))
+			continue;
+
+		prt_printf(out, "Device:");
+		prt_tab(out);
+		prt_printf(out, "%u", i);
+		prt_newline(out);
+
+		printbuf_indent_add(out, 2);
+
+		prt_printf(out, "UUID:");
+		prt_tab(out);
+		pr_uuid(out, m->uuid.b);
+		prt_newline(out);
+
+		prt_printf(out, "Size:");
+		prt_tab(out);
+		prt_units_u64(out, device_size << 9);
+		prt_newline(out);
+
+		prt_printf(out, "Bucket size:");
+		prt_tab(out);
+		prt_units_u64(out, bucket_size << 9);
+		prt_newline(out);
+
+		prt_printf(out, "First bucket:");
+		prt_tab(out);
+		prt_printf(out, "%u", le16_to_cpu(m->first_bucket));
+		prt_newline(out);
+
+		prt_printf(out, "Buckets:");
+		prt_tab(out);
+		prt_printf(out, "%llu", le64_to_cpu(m->nbuckets));
+		prt_newline(out);
+
+		prt_printf(out, "Last mount:");
+		prt_tab(out);
+		if (m->last_mount)
+			pr_time(out, le64_to_cpu(m->last_mount));
+		else
+			prt_printf(out, "(never)");
+		prt_newline(out);
+
+		prt_printf(out, "State:");
+		prt_tab(out);
+		prt_printf(out, "%s",
+		       BCH_MEMBER_STATE(m) < BCH_MEMBER_STATE_NR
+		       ? bch2_member_states[BCH_MEMBER_STATE(m)]
+		       : "unknown");
+		prt_newline(out);
+
+		prt_printf(out, "Label:");
+		prt_tab(out);
+		if (BCH_MEMBER_GROUP(m)) {
+			unsigned idx = BCH_MEMBER_GROUP(m) - 1;
+
+			if (idx < disk_groups_nr(gi))
+				prt_printf(out, "%s (%u)",
+				       gi->entries[idx].label, idx);
+			else
+				prt_printf(out, "(bad disk labels section)");
+		} else {
+			prt_printf(out, "(none)");
+		}
+		prt_newline(out);
+
+		prt_printf(out, "Data allowed:");
+		prt_tab(out);
+		if (BCH_MEMBER_DATA_ALLOWED(m))
+			prt_bitflags(out, bch2_data_types, BCH_MEMBER_DATA_ALLOWED(m));
+		else
+			prt_printf(out, "(none)");
+		prt_newline(out);
+
+		prt_printf(out, "Has data:");
+		prt_tab(out);
+		if (data_have)
+			prt_bitflags(out, bch2_data_types, data_have);
+		else
+			prt_printf(out, "(none)");
+		prt_newline(out);
+
+		prt_printf(out, "Discard:");
+		prt_tab(out);
+		prt_printf(out, "%llu", BCH_MEMBER_DISCARD(m));
+		prt_newline(out);
+
+		prt_printf(out, "Freespace initialized:");
+		prt_tab(out);
+		prt_printf(out, "%llu", BCH_MEMBER_FREESPACE_INITIALIZED(m));
+		prt_newline(out);
+
+		printbuf_indent_sub(out, 2);
+	}
+}
+
+const struct bch_sb_field_ops bch_sb_field_ops_members = {
+	.validate	= bch2_sb_members_validate,
+	.to_text	= bch2_sb_members_to_text,
+};
diff --git a/libbcachefs/sb-members.h b/libbcachefs/sb-members.h
new file mode 100644
index 00000000..34e1cf60
--- /dev/null
+++ b/libbcachefs/sb-members.h
@@ -0,0 +1,176 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_SB_MEMBERS_H
+#define _BCACHEFS_SB_MEMBERS_H
+
+static inline bool bch2_dev_is_online(struct bch_dev *ca)
+{
+	return !percpu_ref_is_zero(&ca->io_ref);
+}
+
+static inline bool bch2_dev_is_readable(struct bch_dev *ca)
+{
+	return bch2_dev_is_online(ca) &&
+		ca->mi.state != BCH_MEMBER_STATE_failed;
+}
+
+static inline bool bch2_dev_get_ioref(struct bch_dev *ca, int rw)
+{
+	if (!percpu_ref_tryget(&ca->io_ref))
+		return false;
+
+	if (ca->mi.state == BCH_MEMBER_STATE_rw ||
+	    (ca->mi.state == BCH_MEMBER_STATE_ro && rw == READ))
+		return true;
+
+	percpu_ref_put(&ca->io_ref);
+	return false;
+}
+
+static inline unsigned dev_mask_nr(const struct bch_devs_mask *devs)
+{
+	return bitmap_weight(devs->d, BCH_SB_MEMBERS_MAX);
+}
+
+static inline bool bch2_dev_list_has_dev(struct bch_devs_list devs,
+					 unsigned dev)
+{
+	unsigned i;
+
+	for (i = 0; i < devs.nr; i++)
+		if (devs.devs[i] == dev)
+			return true;
+
+	return false;
+}
+
+static inline void bch2_dev_list_drop_dev(struct bch_devs_list *devs,
+					  unsigned dev)
+{
+	unsigned i;
+
+	for (i = 0; i < devs->nr; i++)
+		if (devs->devs[i] == dev) {
+			array_remove_item(devs->devs, devs->nr, i);
+			return;
+		}
+}
+
+static inline void bch2_dev_list_add_dev(struct bch_devs_list *devs,
+					 unsigned dev)
+{
+	if (!bch2_dev_list_has_dev(*devs, dev)) {
+		BUG_ON(devs->nr >= ARRAY_SIZE(devs->devs));
+		devs->devs[devs->nr++] = dev;
+	}
+}
+
+static inline struct bch_devs_list bch2_dev_list_single(unsigned dev)
+{
+	return (struct bch_devs_list) { .nr = 1, .devs[0] = dev };
+}
+
+static inline struct bch_dev *__bch2_next_dev(struct bch_fs *c, unsigned *iter,
+					      const struct bch_devs_mask *mask)
+{
+	struct bch_dev *ca = NULL;
+
+	while ((*iter = mask
+		? find_next_bit(mask->d, c->sb.nr_devices, *iter)
+		: *iter) < c->sb.nr_devices &&
+	       !(ca = rcu_dereference_check(c->devs[*iter],
+					    lockdep_is_held(&c->state_lock))))
+		(*iter)++;
+
+	return ca;
+}
+
+#define for_each_member_device_rcu(ca, c, iter, mask)			\
+	for ((iter) = 0; ((ca) = __bch2_next_dev((c), &(iter), mask)); (iter)++)
+
+static inline struct bch_dev *bch2_get_next_dev(struct bch_fs *c, unsigned *iter)
+{
+	struct bch_dev *ca;
+
+	rcu_read_lock();
+	if ((ca = __bch2_next_dev(c, iter, NULL)))
+		percpu_ref_get(&ca->ref);
+	rcu_read_unlock();
+
+	return ca;
+}
+
+/*
+ * If you break early, you must drop your ref on the current device
+ */
+#define for_each_member_device(ca, c, iter)				\
+	for ((iter) = 0;						\
+	     (ca = bch2_get_next_dev(c, &(iter)));			\
+	     percpu_ref_put(&ca->ref), (iter)++)
+
+static inline struct bch_dev *bch2_get_next_online_dev(struct bch_fs *c,
+						      unsigned *iter,
+						      int state_mask)
+{
+	struct bch_dev *ca;
+
+	rcu_read_lock();
+	while ((ca = __bch2_next_dev(c, iter, NULL)) &&
+	       (!((1 << ca->mi.state) & state_mask) ||
+		!percpu_ref_tryget(&ca->io_ref)))
+		(*iter)++;
+	rcu_read_unlock();
+
+	return ca;
+}
+
+#define __for_each_online_member(ca, c, iter, state_mask)		\
+	for ((iter) = 0;						\
+	     (ca = bch2_get_next_online_dev(c, &(iter), state_mask));	\
+	     percpu_ref_put(&ca->io_ref), (iter)++)
+
+#define for_each_online_member(ca, c, iter)				\
+	__for_each_online_member(ca, c, iter, ~0)
+
+#define for_each_rw_member(ca, c, iter)					\
+	__for_each_online_member(ca, c, iter, 1 << BCH_MEMBER_STATE_rw)
+
+#define for_each_readable_member(ca, c, iter)				\
+	__for_each_online_member(ca, c, iter,				\
+		(1 << BCH_MEMBER_STATE_rw)|(1 << BCH_MEMBER_STATE_ro))
+
+/*
+ * If a key exists that references a device, the device won't be going away and
+ * we can omit rcu_read_lock():
+ */
+static inline struct bch_dev *bch_dev_bkey_exists(const struct bch_fs *c, unsigned idx)
+{
+	EBUG_ON(idx >= c->sb.nr_devices || !c->devs[idx]);
+
+	return rcu_dereference_check(c->devs[idx], 1);
+}
+
+static inline struct bch_dev *bch_dev_locked(struct bch_fs *c, unsigned idx)
+{
+	EBUG_ON(idx >= c->sb.nr_devices || !c->devs[idx]);
+
+	return rcu_dereference_protected(c->devs[idx],
+					 lockdep_is_held(&c->sb_lock) ||
+					 lockdep_is_held(&c->state_lock));
+}
+
+/* XXX kill, move to struct bch_fs */
+static inline struct bch_devs_mask bch2_online_devs(struct bch_fs *c)
+{
+	struct bch_devs_mask devs;
+	struct bch_dev *ca;
+	unsigned i;
+
+	memset(&devs, 0, sizeof(devs));
+	for_each_online_member(ca, c, i)
+		__set_bit(ca->dev_idx, devs.d);
+	return devs;
+}
+
+extern const struct bch_sb_field_ops bch_sb_field_ops_members;
+
+#endif /* _BCACHEFS_SB_MEMBERS_H */
diff --git a/libbcachefs/super-io.c b/libbcachefs/super-io.c
index c9a5a7cb..86270288 100644
--- a/libbcachefs/super-io.c
+++ b/libbcachefs/super-io.c
@@ -1,8 +1,6 @@
 // SPDX-License-Identifier: GPL-2.0
 
 #include "bcachefs.h"
-#include "btree_update_interior.h"
-#include "buckets.h"
 #include "checksum.h"
 #include "counters.h"
 #include "disk_groups.h"
@@ -10,12 +8,13 @@
 #include "error.h"
 #include "io.h"
 #include "journal.h"
-#include "journal_io.h"
 #include "journal_sb.h"
 #include "journal_seq_blacklist.h"
 #include "recovery.h"
 #include "replicas.h"
 #include "quota.h"
+#include "sb-clean.h"
+#include "sb-members.h"
 #include "super-io.h"
 #include "super.h"
 #include "trace.h"
@@ -1005,235 +1004,6 @@ void __bch2_check_set_feature(struct bch_fs *c, unsigned feat)
 	mutex_unlock(&c->sb_lock);
 }
 
-/* BCH_SB_FIELD_members: */
-
-static int bch2_sb_members_validate(struct bch_sb *sb,
-				    struct bch_sb_field *f,
-				    struct printbuf *err)
-{
-	struct bch_sb_field_members *mi = field_to_type(f, members);
-	unsigned i;
-
-	if ((void *) (mi->members + sb->nr_devices) >
-	    vstruct_end(&mi->field)) {
-		prt_printf(err, "too many devices for section size");
-		return -BCH_ERR_invalid_sb_members;
-	}
-
-	for (i = 0; i < sb->nr_devices; i++) {
-		struct bch_member *m = mi->members + i;
-
-		if (!bch2_member_exists(m))
-			continue;
-
-		if (le64_to_cpu(m->nbuckets) > LONG_MAX) {
-			prt_printf(err, "device %u: too many buckets (got %llu, max %lu)",
-			       i, le64_to_cpu(m->nbuckets), LONG_MAX);
-			return -BCH_ERR_invalid_sb_members;
-		}
-
-		if (le64_to_cpu(m->nbuckets) -
-		    le16_to_cpu(m->first_bucket) < BCH_MIN_NR_NBUCKETS) {
-			prt_printf(err, "device %u: not enough buckets (got %llu, max %u)",
-			       i, le64_to_cpu(m->nbuckets), BCH_MIN_NR_NBUCKETS);
-			return -BCH_ERR_invalid_sb_members;
-		}
-
-		if (le16_to_cpu(m->bucket_size) <
-		    le16_to_cpu(sb->block_size)) {
-			prt_printf(err, "device %u: bucket size %u smaller than block size %u",
-			       i, le16_to_cpu(m->bucket_size), le16_to_cpu(sb->block_size));
-			return -BCH_ERR_invalid_sb_members;
-		}
-
-		if (le16_to_cpu(m->bucket_size) <
-		    BCH_SB_BTREE_NODE_SIZE(sb)) {
-			prt_printf(err, "device %u: bucket size %u smaller than btree node size %llu",
-			       i, le16_to_cpu(m->bucket_size), BCH_SB_BTREE_NODE_SIZE(sb));
-			return -BCH_ERR_invalid_sb_members;
-		}
-	}
-
-	return 0;
-}
-
-static void bch2_sb_members_to_text(struct printbuf *out, struct bch_sb *sb,
-				    struct bch_sb_field *f)
-{
-	struct bch_sb_field_members *mi = field_to_type(f, members);
-	struct bch_sb_field_disk_groups *gi = bch2_sb_get_disk_groups(sb);
-	unsigned i;
-
-	for (i = 0; i < sb->nr_devices; i++) {
-		struct bch_member *m = mi->members + i;
-		unsigned data_have = bch2_sb_dev_has_data(sb, i);
-		u64 bucket_size = le16_to_cpu(m->bucket_size);
-		u64 device_size = le64_to_cpu(m->nbuckets) * bucket_size;
-
-		if (!bch2_member_exists(m))
-			continue;
-
-		prt_printf(out, "Device:");
-		prt_tab(out);
-		prt_printf(out, "%u", i);
-		prt_newline(out);
-
-		printbuf_indent_add(out, 2);
-
-		prt_printf(out, "UUID:");
-		prt_tab(out);
-		pr_uuid(out, m->uuid.b);
-		prt_newline(out);
-
-		prt_printf(out, "Size:");
-		prt_tab(out);
-		prt_units_u64(out, device_size << 9);
-		prt_newline(out);
-
-		prt_printf(out, "Bucket size:");
-		prt_tab(out);
-		prt_units_u64(out, bucket_size << 9);
-		prt_newline(out);
-
-		prt_printf(out, "First bucket:");
-		prt_tab(out);
-		prt_printf(out, "%u", le16_to_cpu(m->first_bucket));
-		prt_newline(out);
-
-		prt_printf(out, "Buckets:");
-		prt_tab(out);
-		prt_printf(out, "%llu", le64_to_cpu(m->nbuckets));
-		prt_newline(out);
-
-		prt_printf(out, "Last mount:");
-		prt_tab(out);
-		if (m->last_mount)
-			pr_time(out, le64_to_cpu(m->last_mount));
-		else
-			prt_printf(out, "(never)");
-		prt_newline(out);
-
-		prt_printf(out, "State:");
-		prt_tab(out);
-		prt_printf(out, "%s",
-		       BCH_MEMBER_STATE(m) < BCH_MEMBER_STATE_NR
-		       ? bch2_member_states[BCH_MEMBER_STATE(m)]
-		       : "unknown");
-		prt_newline(out);
-
-		prt_printf(out, "Label:");
-		prt_tab(out);
-		if (BCH_MEMBER_GROUP(m)) {
-			unsigned idx = BCH_MEMBER_GROUP(m) - 1;
-
-			if (idx < disk_groups_nr(gi))
-				prt_printf(out, "%s (%u)",
-				       gi->entries[idx].label, idx);
-			else
-				prt_printf(out, "(bad disk labels section)");
-		} else {
-			prt_printf(out, "(none)");
-		}
-		prt_newline(out);
-
-		prt_printf(out, "Data allowed:");
-		prt_tab(out);
-		if (BCH_MEMBER_DATA_ALLOWED(m))
-			prt_bitflags(out, bch2_data_types, BCH_MEMBER_DATA_ALLOWED(m));
-		else
-			prt_printf(out, "(none)");
-		prt_newline(out);
-
-		prt_printf(out, "Has data:");
-		prt_tab(out);
-		if (data_have)
-			prt_bitflags(out, bch2_data_types, data_have);
-		else
-			prt_printf(out, "(none)");
-		prt_newline(out);
-
-		prt_printf(out, "Discard:");
-		prt_tab(out);
-		prt_printf(out, "%llu", BCH_MEMBER_DISCARD(m));
-		prt_newline(out);
-
-		prt_printf(out, "Freespace initialized:");
-		prt_tab(out);
-		prt_printf(out, "%llu", BCH_MEMBER_FREESPACE_INITIALIZED(m));
-		prt_newline(out);
-
-		printbuf_indent_sub(out, 2);
-	}
-}
-
-static const struct bch_sb_field_ops bch_sb_field_ops_members = {
-	.validate	= bch2_sb_members_validate,
-	.to_text	= bch2_sb_members_to_text,
-};
-
-/* BCH_SB_FIELD_crypt: */
-
-static int bch2_sb_crypt_validate(struct bch_sb *sb,
-				  struct bch_sb_field *f,
-				  struct printbuf *err)
-{
-	struct bch_sb_field_crypt *crypt = field_to_type(f, crypt);
-
-	if (vstruct_bytes(&crypt->field) < sizeof(*crypt)) {
-		prt_printf(err, "wrong size (got %zu should be %zu)",
-		       vstruct_bytes(&crypt->field), sizeof(*crypt));
-		return -BCH_ERR_invalid_sb_crypt;
-	}
-
-	if (BCH_CRYPT_KDF_TYPE(crypt)) {
-		prt_printf(err, "bad kdf type %llu", BCH_CRYPT_KDF_TYPE(crypt));
-		return -BCH_ERR_invalid_sb_crypt;
-	}
-
-	return 0;
-}
-
-static void bch2_sb_crypt_to_text(struct printbuf *out, struct bch_sb *sb,
-				  struct bch_sb_field *f)
-{
-	struct bch_sb_field_crypt *crypt = field_to_type(f, crypt);
-
-	prt_printf(out, "KFD:               %llu", BCH_CRYPT_KDF_TYPE(crypt));
-	prt_newline(out);
-	prt_printf(out, "scrypt n:          %llu", BCH_KDF_SCRYPT_N(crypt));
-	prt_newline(out);
-	prt_printf(out, "scrypt r:          %llu", BCH_KDF_SCRYPT_R(crypt));
-	prt_newline(out);
-	prt_printf(out, "scrypt p:          %llu", BCH_KDF_SCRYPT_P(crypt));
-	prt_newline(out);
-}
-
-static const struct bch_sb_field_ops bch_sb_field_ops_crypt = {
-	.validate	= bch2_sb_crypt_validate,
-	.to_text	= bch2_sb_crypt_to_text,
-};
-
-/* BCH_SB_FIELD_clean: */
-
-int bch2_sb_clean_validate_late(struct bch_fs *c, struct bch_sb_field_clean *clean, int write)
-{
-	struct jset_entry *entry;
-	int ret;
-
-	for (entry = clean->start;
-	     entry < (struct jset_entry *) vstruct_end(&clean->field);
-	     entry = vstruct_next(entry)) {
-		ret = bch2_journal_entry_validate(c, NULL, entry,
-						  le16_to_cpu(c->disk_sb.sb->version),
-						  BCH_SB_BIG_ENDIAN(c->disk_sb.sb),
-						  write);
-		if (ret)
-			return ret;
-	}
-
-	return 0;
-}
-
 /* Downgrade if superblock is at a higher version than currently supported: */
 void bch2_sb_maybe_downgrade(struct bch_fs *c)
 {
@@ -1260,232 +1030,6 @@ void bch2_sb_upgrade(struct bch_fs *c, unsigned new_version)
 	c->disk_sb.sb->features[0] |= cpu_to_le64(BCH_SB_FEATURES_ALL);
 }
 
-int bch2_fs_mark_dirty(struct bch_fs *c)
-{
-	int ret;
-
-	/*
-	 * Unconditionally write superblock, to verify it hasn't changed before
-	 * we go rw:
-	 */
-
-	mutex_lock(&c->sb_lock);
-	SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
-
-	bch2_sb_maybe_downgrade(c);
-	c->disk_sb.sb->features[0] |= cpu_to_le64(BCH_SB_FEATURES_ALWAYS);
-
-	ret = bch2_write_super(c);
-	mutex_unlock(&c->sb_lock);
-
-	return ret;
-}
-
-static struct jset_entry *jset_entry_init(struct jset_entry **end, size_t size)
-{
-	struct jset_entry *entry = *end;
-	unsigned u64s = DIV_ROUND_UP(size, sizeof(u64));
-
-	memset(entry, 0, u64s * sizeof(u64));
-	/*
-	 * The u64s field counts from the start of data, ignoring the shared
-	 * fields.
-	 */
-	entry->u64s = cpu_to_le16(u64s - 1);
-
-	*end = vstruct_next(*end);
-	return entry;
-}
-
-void bch2_journal_super_entries_add_common(struct bch_fs *c,
-					   struct jset_entry **end,
-					   u64 journal_seq)
-{
-	struct bch_dev *ca;
-	unsigned i, dev;
-
-	percpu_down_read(&c->mark_lock);
-
-	if (!journal_seq) {
-		for (i = 0; i < ARRAY_SIZE(c->usage); i++)
-			bch2_fs_usage_acc_to_base(c, i);
-	} else {
-		bch2_fs_usage_acc_to_base(c, journal_seq & JOURNAL_BUF_MASK);
-	}
-
-	{
-		struct jset_entry_usage *u =
-			container_of(jset_entry_init(end, sizeof(*u)),
-				     struct jset_entry_usage, entry);
-
-		u->entry.type	= BCH_JSET_ENTRY_usage;
-		u->entry.btree_id = BCH_FS_USAGE_inodes;
-		u->v		= cpu_to_le64(c->usage_base->nr_inodes);
-	}
-
-	{
-		struct jset_entry_usage *u =
-			container_of(jset_entry_init(end, sizeof(*u)),
-				     struct jset_entry_usage, entry);
-
-		u->entry.type	= BCH_JSET_ENTRY_usage;
-		u->entry.btree_id = BCH_FS_USAGE_key_version;
-		u->v		= cpu_to_le64(atomic64_read(&c->key_version));
-	}
-
-	for (i = 0; i < BCH_REPLICAS_MAX; i++) {
-		struct jset_entry_usage *u =
-			container_of(jset_entry_init(end, sizeof(*u)),
-				     struct jset_entry_usage, entry);
-
-		u->entry.type	= BCH_JSET_ENTRY_usage;
-		u->entry.btree_id = BCH_FS_USAGE_reserved;
-		u->entry.level	= i;
-		u->v		= cpu_to_le64(c->usage_base->persistent_reserved[i]);
-	}
-
-	for (i = 0; i < c->replicas.nr; i++) {
-		struct bch_replicas_entry *e =
-			cpu_replicas_entry(&c->replicas, i);
-		struct jset_entry_data_usage *u =
-			container_of(jset_entry_init(end, sizeof(*u) + e->nr_devs),
-				     struct jset_entry_data_usage, entry);
-
-		u->entry.type	= BCH_JSET_ENTRY_data_usage;
-		u->v		= cpu_to_le64(c->usage_base->replicas[i]);
-		unsafe_memcpy(&u->r, e, replicas_entry_bytes(e),
-			      "embedded variable length struct");
-	}
-
-	for_each_member_device(ca, c, dev) {
-		unsigned b = sizeof(struct jset_entry_dev_usage) +
-			sizeof(struct jset_entry_dev_usage_type) * BCH_DATA_NR;
-		struct jset_entry_dev_usage *u =
-			container_of(jset_entry_init(end, b),
-				     struct jset_entry_dev_usage, entry);
-
-		u->entry.type = BCH_JSET_ENTRY_dev_usage;
-		u->dev = cpu_to_le32(dev);
-		u->buckets_ec		= cpu_to_le64(ca->usage_base->buckets_ec);
-
-		for (i = 0; i < BCH_DATA_NR; i++) {
-			u->d[i].buckets = cpu_to_le64(ca->usage_base->d[i].buckets);
-			u->d[i].sectors	= cpu_to_le64(ca->usage_base->d[i].sectors);
-			u->d[i].fragmented = cpu_to_le64(ca->usage_base->d[i].fragmented);
-		}
-	}
-
-	percpu_up_read(&c->mark_lock);
-
-	for (i = 0; i < 2; i++) {
-		struct jset_entry_clock *clock =
-			container_of(jset_entry_init(end, sizeof(*clock)),
-				     struct jset_entry_clock, entry);
-
-		clock->entry.type = BCH_JSET_ENTRY_clock;
-		clock->rw	= i;
-		clock->time	= cpu_to_le64(atomic64_read(&c->io_clock[i].now));
-	}
-}
-
-void bch2_fs_mark_clean(struct bch_fs *c)
-{
-	struct bch_sb_field_clean *sb_clean;
-	struct jset_entry *entry;
-	unsigned u64s;
-	int ret;
-
-	mutex_lock(&c->sb_lock);
-	if (BCH_SB_CLEAN(c->disk_sb.sb))
-		goto out;
-
-	SET_BCH_SB_CLEAN(c->disk_sb.sb, true);
-
-	c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_alloc_info);
-	c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_alloc_metadata);
-	c->disk_sb.sb->features[0] &= cpu_to_le64(~(1ULL << BCH_FEATURE_extents_above_btree_updates));
-	c->disk_sb.sb->features[0] &= cpu_to_le64(~(1ULL << BCH_FEATURE_btree_updates_journalled));
-
-	u64s = sizeof(*sb_clean) / sizeof(u64) + c->journal.entry_u64s_reserved;
-
-	sb_clean = bch2_sb_resize_clean(&c->disk_sb, u64s);
-	if (!sb_clean) {
-		bch_err(c, "error resizing superblock while setting filesystem clean");
-		goto out;
-	}
-
-	sb_clean->flags		= 0;
-	sb_clean->journal_seq	= cpu_to_le64(atomic64_read(&c->journal.seq));
-
-	/* Trying to catch outstanding bug: */
-	BUG_ON(le64_to_cpu(sb_clean->journal_seq) > S64_MAX);
-
-	entry = sb_clean->start;
-	bch2_journal_super_entries_add_common(c, &entry, 0);
-	entry = bch2_btree_roots_to_journal_entries(c, entry, entry);
-	BUG_ON((void *) entry > vstruct_end(&sb_clean->field));
-
-	memset(entry, 0,
-	       vstruct_end(&sb_clean->field) - (void *) entry);
-
-	/*
-	 * this should be in the write path, and we should be validating every
-	 * superblock section:
-	 */
-	ret = bch2_sb_clean_validate_late(c, sb_clean, WRITE);
-	if (ret) {
-		bch_err(c, "error writing marking filesystem clean: validate error");
-		goto out;
-	}
-
-	bch2_write_super(c);
-out:
-	mutex_unlock(&c->sb_lock);
-}
-
-static int bch2_sb_clean_validate(struct bch_sb *sb,
-				  struct bch_sb_field *f,
-				  struct printbuf *err)
-{
-	struct bch_sb_field_clean *clean = field_to_type(f, clean);
-
-	if (vstruct_bytes(&clean->field) < sizeof(*clean)) {
-		prt_printf(err, "wrong size (got %zu should be %zu)",
-		       vstruct_bytes(&clean->field), sizeof(*clean));
-		return -BCH_ERR_invalid_sb_clean;
-	}
-
-	return 0;
-}
-
-static void bch2_sb_clean_to_text(struct printbuf *out, struct bch_sb *sb,
-				  struct bch_sb_field *f)
-{
-	struct bch_sb_field_clean *clean = field_to_type(f, clean);
-	struct jset_entry *entry;
-
-	prt_printf(out, "flags:          %x",	le32_to_cpu(clean->flags));
-	prt_newline(out);
-	prt_printf(out, "journal_seq:    %llu",	le64_to_cpu(clean->journal_seq));
-	prt_newline(out);
-
-	for (entry = clean->start;
-	     entry != vstruct_end(&clean->field);
-	     entry = vstruct_next(entry)) {
-		if (entry->type == BCH_JSET_ENTRY_btree_keys &&
-		    !entry->u64s)
-			continue;
-
-		bch2_journal_entry_to_text(out, NULL, entry);
-		prt_newline(out);
-	}
-}
-
-static const struct bch_sb_field_ops bch_sb_field_ops_clean = {
-	.validate	= bch2_sb_clean_validate,
-	.to_text	= bch2_sb_clean_to_text,
-};
-
 static const struct bch_sb_field_ops *bch2_sb_field_ops[] = {
 #define x(f, nr)					\
 	[BCH_SB_FIELD_##f] = &bch_sb_field_ops_##f,
diff --git a/libbcachefs/super-io.h b/libbcachefs/super-io.h
index 904adea6..5181dcb3 100644
--- a/libbcachefs/super-io.h
+++ b/libbcachefs/super-io.h
@@ -121,19 +121,9 @@ static inline struct bch_member_cpu bch2_mi_to_cpu(struct bch_member *mi)
 	};
 }
 
-/* BCH_SB_FIELD_clean: */
-
-void bch2_journal_super_entries_add_common(struct bch_fs *,
-					   struct jset_entry **, u64);
-
-int bch2_sb_clean_validate_late(struct bch_fs *, struct bch_sb_field_clean *, int);
-
 void bch2_sb_maybe_downgrade(struct bch_fs *);
 void bch2_sb_upgrade(struct bch_fs *, unsigned);
 
-int bch2_fs_mark_dirty(struct bch_fs *);
-void bch2_fs_mark_clean(struct bch_fs *);
-
 void bch2_sb_field_to_text(struct printbuf *, struct bch_sb *,
 			   struct bch_sb_field *);
 void bch2_sb_layout_to_text(struct printbuf *, struct bch_sb_layout *);
diff --git a/libbcachefs/super.c b/libbcachefs/super.c
index eee56969..82e992b3 100644
--- a/libbcachefs/super.c
+++ b/libbcachefs/super.c
@@ -13,6 +13,7 @@
 #include "bkey_sort.h"
 #include "btree_cache.h"
 #include "btree_gc.h"
+#include "btree_journal_iter.h"
 #include "btree_key_cache.h"
 #include "btree_update_interior.h"
 #include "btree_io.h"
@@ -30,6 +31,8 @@
 #include "error.h"
 #include "fs.h"
 #include "fs-io.h"
+#include "fs-io-buffered.h"
+#include "fs-io-direct.h"
 #include "fsck.h"
 #include "inode.h"
 #include "io.h"
@@ -44,6 +47,7 @@
 #include "rebalance.h"
 #include "recovery.h"
 #include "replicas.h"
+#include "sb-clean.h"
 #include "subvolume.h"
 #include "super.h"
 #include "super-io.h"
@@ -469,6 +473,8 @@ static void __bch2_fs_free(struct bch_fs *c)
 	bch2_fs_counters_exit(c);
 	bch2_fs_snapshots_exit(c);
 	bch2_fs_quota_exit(c);
+	bch2_fs_fs_io_direct_exit(c);
+	bch2_fs_fs_io_buffered_exit(c);
 	bch2_fs_fsio_exit(c);
 	bch2_fs_ec_exit(c);
 	bch2_fs_encryption_exit(c);
@@ -844,7 +850,9 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	    bch2_fs_encryption_init(c) ?:
 	    bch2_fs_compress_init(c) ?:
 	    bch2_fs_ec_init(c) ?:
-	    bch2_fs_fsio_init(c);
+	    bch2_fs_fsio_init(c) ?:
+	    bch2_fs_fs_io_buffered_init(c);
+	    bch2_fs_fs_io_direct_init(c);
 	if (ret)
 		goto err;
 
@@ -2000,6 +2008,7 @@ err:
 BCH_DEBUG_PARAMS()
 #undef BCH_DEBUG_PARAM
 
+__maybe_unused
 static unsigned bch2_metadata_version = bcachefs_metadata_version_current;
 module_param_named(version, bch2_metadata_version, uint, 0400);
 
diff --git a/libbcachefs/super.h b/libbcachefs/super.h
index 36bcb9ec..bf762df1 100644
--- a/libbcachefs/super.h
+++ b/libbcachefs/super.h
@@ -8,220 +8,6 @@
 
 #include <linux/math64.h>
 
-static inline size_t sector_to_bucket(const struct bch_dev *ca, sector_t s)
-{
-	return div_u64(s, ca->mi.bucket_size);
-}
-
-static inline sector_t bucket_to_sector(const struct bch_dev *ca, size_t b)
-{
-	return ((sector_t) b) * ca->mi.bucket_size;
-}
-
-static inline sector_t bucket_remainder(const struct bch_dev *ca, sector_t s)
-{
-	u32 remainder;
-
-	div_u64_rem(s, ca->mi.bucket_size, &remainder);
-	return remainder;
-}
-
-static inline size_t sector_to_bucket_and_offset(const struct bch_dev *ca, sector_t s,
-						 u32 *offset)
-{
-	return div_u64_rem(s, ca->mi.bucket_size, offset);
-}
-
-static inline bool bch2_dev_is_online(struct bch_dev *ca)
-{
-	return !percpu_ref_is_zero(&ca->io_ref);
-}
-
-static inline bool bch2_dev_is_readable(struct bch_dev *ca)
-{
-	return bch2_dev_is_online(ca) &&
-		ca->mi.state != BCH_MEMBER_STATE_failed;
-}
-
-static inline bool bch2_dev_get_ioref(struct bch_dev *ca, int rw)
-{
-	if (!percpu_ref_tryget(&ca->io_ref))
-		return false;
-
-	if (ca->mi.state == BCH_MEMBER_STATE_rw ||
-	    (ca->mi.state == BCH_MEMBER_STATE_ro && rw == READ))
-		return true;
-
-	percpu_ref_put(&ca->io_ref);
-	return false;
-}
-
-static inline unsigned dev_mask_nr(const struct bch_devs_mask *devs)
-{
-	return bitmap_weight(devs->d, BCH_SB_MEMBERS_MAX);
-}
-
-static inline bool bch2_dev_list_has_dev(struct bch_devs_list devs,
-					 unsigned dev)
-{
-	unsigned i;
-
-	for (i = 0; i < devs.nr; i++)
-		if (devs.devs[i] == dev)
-			return true;
-
-	return false;
-}
-
-static inline void bch2_dev_list_drop_dev(struct bch_devs_list *devs,
-					  unsigned dev)
-{
-	unsigned i;
-
-	for (i = 0; i < devs->nr; i++)
-		if (devs->devs[i] == dev) {
-			array_remove_item(devs->devs, devs->nr, i);
-			return;
-		}
-}
-
-static inline void bch2_dev_list_add_dev(struct bch_devs_list *devs,
-					 unsigned dev)
-{
-	if (!bch2_dev_list_has_dev(*devs, dev)) {
-		BUG_ON(devs->nr >= ARRAY_SIZE(devs->devs));
-		devs->devs[devs->nr++] = dev;
-	}
-}
-
-static inline struct bch_devs_list bch2_dev_list_single(unsigned dev)
-{
-	return (struct bch_devs_list) { .nr = 1, .devs[0] = dev };
-}
-
-static inline struct bch_dev *__bch2_next_dev(struct bch_fs *c, unsigned *iter,
-					      const struct bch_devs_mask *mask)
-{
-	struct bch_dev *ca = NULL;
-
-	while ((*iter = mask
-		? find_next_bit(mask->d, c->sb.nr_devices, *iter)
-		: *iter) < c->sb.nr_devices &&
-	       !(ca = rcu_dereference_check(c->devs[*iter],
-					    lockdep_is_held(&c->state_lock))))
-		(*iter)++;
-
-	return ca;
-}
-
-#define for_each_member_device_rcu(ca, c, iter, mask)			\
-	for ((iter) = 0; ((ca) = __bch2_next_dev((c), &(iter), mask)); (iter)++)
-
-static inline struct bch_dev *bch2_get_next_dev(struct bch_fs *c, unsigned *iter)
-{
-	struct bch_dev *ca;
-
-	rcu_read_lock();
-	if ((ca = __bch2_next_dev(c, iter, NULL)))
-		percpu_ref_get(&ca->ref);
-	rcu_read_unlock();
-
-	return ca;
-}
-
-/*
- * If you break early, you must drop your ref on the current device
- */
-#define for_each_member_device(ca, c, iter)				\
-	for ((iter) = 0;						\
-	     (ca = bch2_get_next_dev(c, &(iter)));			\
-	     percpu_ref_put(&ca->ref), (iter)++)
-
-static inline struct bch_dev *bch2_get_next_online_dev(struct bch_fs *c,
-						      unsigned *iter,
-						      int state_mask)
-{
-	struct bch_dev *ca;
-
-	rcu_read_lock();
-	while ((ca = __bch2_next_dev(c, iter, NULL)) &&
-	       (!((1 << ca->mi.state) & state_mask) ||
-		!percpu_ref_tryget(&ca->io_ref)))
-		(*iter)++;
-	rcu_read_unlock();
-
-	return ca;
-}
-
-#define __for_each_online_member(ca, c, iter, state_mask)		\
-	for ((iter) = 0;						\
-	     (ca = bch2_get_next_online_dev(c, &(iter), state_mask));	\
-	     percpu_ref_put(&ca->io_ref), (iter)++)
-
-#define for_each_online_member(ca, c, iter)				\
-	__for_each_online_member(ca, c, iter, ~0)
-
-#define for_each_rw_member(ca, c, iter)					\
-	__for_each_online_member(ca, c, iter, 1 << BCH_MEMBER_STATE_rw)
-
-#define for_each_readable_member(ca, c, iter)				\
-	__for_each_online_member(ca, c, iter,				\
-		(1 << BCH_MEMBER_STATE_rw)|(1 << BCH_MEMBER_STATE_ro))
-
-/*
- * If a key exists that references a device, the device won't be going away and
- * we can omit rcu_read_lock():
- */
-static inline struct bch_dev *bch_dev_bkey_exists(const struct bch_fs *c, unsigned idx)
-{
-	EBUG_ON(idx >= c->sb.nr_devices || !c->devs[idx]);
-
-	return rcu_dereference_check(c->devs[idx], 1);
-}
-
-static inline struct bch_dev *bch_dev_locked(struct bch_fs *c, unsigned idx)
-{
-	EBUG_ON(idx >= c->sb.nr_devices || !c->devs[idx]);
-
-	return rcu_dereference_protected(c->devs[idx],
-					 lockdep_is_held(&c->sb_lock) ||
-					 lockdep_is_held(&c->state_lock));
-}
-
-/* XXX kill, move to struct bch_fs */
-static inline struct bch_devs_mask bch2_online_devs(struct bch_fs *c)
-{
-	struct bch_devs_mask devs;
-	struct bch_dev *ca;
-	unsigned i;
-
-	memset(&devs, 0, sizeof(devs));
-	for_each_online_member(ca, c, i)
-		__set_bit(ca->dev_idx, devs.d);
-	return devs;
-}
-
-static inline bool is_superblock_bucket(struct bch_dev *ca, u64 b)
-{
-	struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
-	u64 b_offset	= bucket_to_sector(ca, b);
-	u64 b_end	= bucket_to_sector(ca, b + 1);
-	unsigned i;
-
-	if (!b)
-		return true;
-
-	for (i = 0; i < layout->nr_superblocks; i++) {
-		u64 offset = le64_to_cpu(layout->sb_offset[i]);
-		u64 end = offset + (1 << layout->sb_max_size_bits);
-
-		if (!(offset >= b_end || end <= b_offset))
-			return true;
-	}
-
-	return false;
-}
-
 struct bch_fs *bch2_dev_to_fs(dev_t);
 struct bch_fs *bch2_uuid_to_fs(__uuid_t);