From bb82ab88dfdb12948af58989c75bfe904bc1b09d Mon Sep 17 00:00:00 2001
From: Arne Jansen <sensille@gmx.net>
Date: Fri, 10 Jun 2011 14:06:53 +0200
Subject: btrfs: add an extra wait mode to read_extent_buffer_pages

read_extent_buffer_pages currently has two modes, either trigger a read
without waiting for anything, or wait for the I/O to finish. The former
also bails when it's unable to lock the page. This patch now adds an
additional parameter to allow it to block on page lock, but don't wait
for completion.

Changes v5:
 - merge the 2 wait parameters into one and define WAIT_NONE, WAIT_COMPLETE and
   WAIT_PAGE_LOCK

Change v6:
 - fix bug introduced in v5

Signed-off-by: Arne Jansen <sensille@gmx.net>
---
 fs/btrfs/disk-io.c   | 5 +++--
 fs/btrfs/extent_io.c | 7 +++----
 fs/btrfs/extent_io.h | 3 +++
 3 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 07b3ac662e193e..e6170e142cad65 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -367,7 +367,8 @@ static int btree_read_extent_buffer_pages(struct btrfs_root *root,
 	clear_bit(EXTENT_BUFFER_CORRUPT, &eb->bflags);
 	io_tree = &BTRFS_I(root->fs_info->btree_inode)->io_tree;
 	while (1) {
-		ret = read_extent_buffer_pages(io_tree, eb, start, 1,
+		ret = read_extent_buffer_pages(io_tree, eb, start,
+					       WAIT_COMPLETE,
 					       btree_get_extent, mirror_num);
 		if (!ret &&
 		    !verify_parent_transid(io_tree, eb, parent_transid))
@@ -974,7 +975,7 @@ int readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize,
 	if (!buf)
 		return 0;
 	read_extent_buffer_pages(&BTRFS_I(btree_inode)->io_tree,
-				 buf, 0, 0, btree_get_extent, 0);
+				 buf, 0, WAIT_NONE, btree_get_extent, 0);
 	free_extent_buffer(buf);
 	return ret;
 }
diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index d418164a35f134..823028e73cf813 100644
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -3349,8 +3349,7 @@ int extent_buffer_uptodate(struct extent_io_tree *tree,
 }
 
 int read_extent_buffer_pages(struct extent_io_tree *tree,
-			     struct extent_buffer *eb,
-			     u64 start, int wait,
+			     struct extent_buffer *eb, u64 start, int wait,
 			     get_extent_t *get_extent, int mirror_num)
 {
 	unsigned long i;
@@ -3386,7 +3385,7 @@ int read_extent_buffer_pages(struct extent_io_tree *tree,
 	num_pages = num_extent_pages(eb->start, eb->len);
 	for (i = start_i; i < num_pages; i++) {
 		page = extent_buffer_page(eb, i);
-		if (!wait) {
+		if (wait == WAIT_NONE) {
 			if (!trylock_page(page))
 				goto unlock_exit;
 		} else {
@@ -3430,7 +3429,7 @@ int read_extent_buffer_pages(struct extent_io_tree *tree,
 	if (bio)
 		submit_one_bio(READ, bio, mirror_num, bio_flags);
 
-	if (ret || !wait)
+	if (ret || wait != WAIT_COMPLETE)
 		return ret;
 
 	for (i = start_i; i < num_pages; i++) {
diff --git a/fs/btrfs/extent_io.h b/fs/btrfs/extent_io.h
index 7b2f0c3e79292a..6d74c6b3469145 100644
--- a/fs/btrfs/extent_io.h
+++ b/fs/btrfs/extent_io.h
@@ -248,6 +248,9 @@ struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
 					 u64 start, unsigned long len);
 void free_extent_buffer(struct extent_buffer *eb);
+#define WAIT_NONE	0
+#define WAIT_COMPLETE	1
+#define WAIT_PAGE_LOCK	2
 int read_extent_buffer_pages(struct extent_io_tree *tree,
 			     struct extent_buffer *eb, u64 start, int wait,
 			     get_extent_t *get_extent, int mirror_num);
-- 
cgit 1.2.3-korg


From ab0fff03055d2d1b01a7581badeba18db9c4f55c Mon Sep 17 00:00:00 2001
From: Arne Jansen <sensille@gmx.net>
Date: Mon, 23 May 2011 14:25:41 +0200
Subject: btrfs: add READAHEAD extent buffer flag

Add a READAHEAD extent buffer flag.
Add a function to trigger a read with this flag set.

Changes v2:
 - use extent buffer flags instead of extent state flags

Changes v5:
 - adapt to changed read_extent_buffer_pages interface
 - don't return eb from reada_tree_block_flagged if it has CORRUPT flag set

Signed-off-by: Arne Jansen <sensille@gmx.net>
---
 fs/btrfs/disk-io.c   | 32 ++++++++++++++++++++++++++++++++
 fs/btrfs/disk-io.h   |  2 ++
 fs/btrfs/extent_io.h |  1 +
 3 files changed, 35 insertions(+)

diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index e6170e142cad65..1220d04072c89a 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -980,6 +980,38 @@ int readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize,
 	return ret;
 }
 
+int reada_tree_block_flagged(struct btrfs_root *root, u64 bytenr, u32 blocksize,
+			 int mirror_num, struct extent_buffer **eb)
+{
+	struct extent_buffer *buf = NULL;
+	struct inode *btree_inode = root->fs_info->btree_inode;
+	struct extent_io_tree *io_tree = &BTRFS_I(btree_inode)->io_tree;
+	int ret;
+
+	buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
+	if (!buf)
+		return 0;
+
+	set_bit(EXTENT_BUFFER_READAHEAD, &buf->bflags);
+
+	ret = read_extent_buffer_pages(io_tree, buf, 0, WAIT_PAGE_LOCK,
+				       btree_get_extent, mirror_num);
+	if (ret) {
+		free_extent_buffer(buf);
+		return ret;
+	}
+
+	if (test_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags)) {
+		free_extent_buffer(buf);
+		return -EIO;
+	} else if (extent_buffer_uptodate(io_tree, buf, NULL)) {
+		*eb = buf;
+	} else {
+		free_extent_buffer(buf);
+	}
+	return 0;
+}
+
 struct extent_buffer *btrfs_find_tree_block(struct btrfs_root *root,
 					    u64 bytenr, u32 blocksize)
 {
diff --git a/fs/btrfs/disk-io.h b/fs/btrfs/disk-io.h
index bec3ea4bd67fd4..b3bdb5c1390f90 100644
--- a/fs/btrfs/disk-io.h
+++ b/fs/btrfs/disk-io.h
@@ -40,6 +40,8 @@ struct extent_buffer *read_tree_block(struct btrfs_root *root, u64 bytenr,
 				      u32 blocksize, u64 parent_transid);
 int readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize,
 			 u64 parent_transid);
+int reada_tree_block_flagged(struct btrfs_root *root, u64 bytenr, u32 blocksize,
+			 int mirror_num, struct extent_buffer **eb);
 struct extent_buffer *btrfs_find_create_tree_block(struct btrfs_root *root,
 						   u64 bytenr, u32 blocksize);
 int clean_tree_block(struct btrfs_trans_handle *trans,
diff --git a/fs/btrfs/extent_io.h b/fs/btrfs/extent_io.h
index 6d74c6b3469145..fcaf49bcb88096 100644
--- a/fs/btrfs/extent_io.h
+++ b/fs/btrfs/extent_io.h
@@ -32,6 +32,7 @@
 #define EXTENT_BUFFER_BLOCKING 1
 #define EXTENT_BUFFER_DIRTY 2
 #define EXTENT_BUFFER_CORRUPT 3
+#define EXTENT_BUFFER_READAHEAD 4	/* this got triggered by readahead */
 
 /* these are flags for extent_clear_unlock_delalloc */
 #define EXTENT_CLEAR_UNLOCK_PAGE 0x1
-- 
cgit 1.2.3-korg


From 90519d66abbccc251d14719ac76f191f70826e40 Mon Sep 17 00:00:00 2001
From: Arne Jansen <sensille@gmx.net>
Date: Mon, 23 May 2011 14:30:00 +0200
Subject: btrfs: state information for readahead

Add state information for readahead to btrfs_fs_info and btrfs_device

Changes v2:
 - don't wait in radix_trees
 - add own set of workers for readahead

Reviewed-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Arne Jansen <sensille@gmx.net>
---
 fs/btrfs/ctree.h   |  5 +++++
 fs/btrfs/disk-io.c | 10 ++++++++++
 fs/btrfs/volumes.c |  8 ++++++++
 fs/btrfs/volumes.h |  8 ++++++++
 4 files changed, 31 insertions(+)

diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
index 03912c5c6f4982..f71fd24cc152d4 100644
--- a/fs/btrfs/ctree.h
+++ b/fs/btrfs/ctree.h
@@ -1036,6 +1036,7 @@ struct btrfs_fs_info {
 	struct btrfs_workers endio_freespace_worker;
 	struct btrfs_workers submit_workers;
 	struct btrfs_workers caching_workers;
+	struct btrfs_workers readahead_workers;
 
 	/*
 	 * fixup workers take dirty pages that didn't properly go through
@@ -1119,6 +1120,10 @@ struct btrfs_fs_info {
 	u64 fs_state;
 
 	struct btrfs_delayed_root *delayed_root;
+
+	/* readahead tree */
+	spinlock_t reada_lock;
+	struct radix_tree_root reada_tree;
 };
 
 /*
diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 1220d04072c89a..fd28c40e9ec353 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -1711,6 +1711,10 @@ struct btrfs_root *open_ctree(struct super_block *sb,
 	fs_info->defrag_inodes = RB_ROOT;
 	fs_info->trans_no_join = 0;
 
+	/* readahead state */
+	INIT_RADIX_TREE(&fs_info->reada_tree, GFP_NOFS & ~__GFP_WAIT);
+	spin_lock_init(&fs_info->reada_lock);
+
 	fs_info->thread_pool_size = min_t(unsigned long,
 					  num_online_cpus() + 2, 8);
 
@@ -1903,6 +1907,9 @@ struct btrfs_root *open_ctree(struct super_block *sb,
 	btrfs_init_workers(&fs_info->delayed_workers, "delayed-meta",
 			   fs_info->thread_pool_size,
 			   &fs_info->generic_worker);
+	btrfs_init_workers(&fs_info->readahead_workers, "readahead",
+			   fs_info->thread_pool_size,
+			   &fs_info->generic_worker);
 
 	/*
 	 * endios are largely parallel and should have a very
@@ -1913,6 +1920,7 @@ struct btrfs_root *open_ctree(struct super_block *sb,
 
 	fs_info->endio_write_workers.idle_thresh = 2;
 	fs_info->endio_meta_write_workers.idle_thresh = 2;
+	fs_info->readahead_workers.idle_thresh = 2;
 
 	btrfs_start_workers(&fs_info->workers, 1);
 	btrfs_start_workers(&fs_info->generic_worker, 1);
@@ -1926,6 +1934,7 @@ struct btrfs_root *open_ctree(struct super_block *sb,
 	btrfs_start_workers(&fs_info->endio_freespace_worker, 1);
 	btrfs_start_workers(&fs_info->delayed_workers, 1);
 	btrfs_start_workers(&fs_info->caching_workers, 1);
+	btrfs_start_workers(&fs_info->readahead_workers, 1);
 
 	fs_info->bdi.ra_pages *= btrfs_super_num_devices(disk_super);
 	fs_info->bdi.ra_pages = max(fs_info->bdi.ra_pages,
@@ -2650,6 +2659,7 @@ int close_ctree(struct btrfs_root *root)
 	btrfs_stop_workers(&fs_info->submit_workers);
 	btrfs_stop_workers(&fs_info->delayed_workers);
 	btrfs_stop_workers(&fs_info->caching_workers);
+	btrfs_stop_workers(&fs_info->readahead_workers);
 
 	btrfs_close_devices(fs_info->fs_devices);
 	btrfs_mapping_tree_free(&fs_info->mapping_tree);
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index f2a4cc79da61da..1dccce5bc93d05 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -366,6 +366,14 @@ static noinline int device_list_add(const char *path,
 		}
 		INIT_LIST_HEAD(&device->dev_alloc_list);
 
+		/* init readahead state */
+		spin_lock_init(&device->reada_lock);
+		device->reada_curr_zone = NULL;
+		atomic_set(&device->reada_in_flight, 0);
+		device->reada_next = 0;
+		INIT_RADIX_TREE(&device->reada_zones, GFP_NOFS & ~__GFP_WAIT);
+		INIT_RADIX_TREE(&device->reada_extents, GFP_NOFS & ~__GFP_WAIT);
+
 		mutex_lock(&fs_devices->device_list_mutex);
 		list_add_rcu(&device->dev_list, &fs_devices->devices);
 		mutex_unlock(&fs_devices->device_list_mutex);
diff --git a/fs/btrfs/volumes.h b/fs/btrfs/volumes.h
index 6d866db4e1774a..2a751246188ad7 100644
--- a/fs/btrfs/volumes.h
+++ b/fs/btrfs/volumes.h
@@ -92,6 +92,14 @@ struct btrfs_device {
 	struct btrfs_work work;
 	struct rcu_head rcu;
 	struct work_struct rcu_work;
+
+	/* readahead state */
+	spinlock_t reada_lock;
+	atomic_t reada_in_flight;
+	u64 reada_next;
+	struct reada_zone *reada_curr_zone;
+	struct radix_tree_root reada_zones;
+	struct radix_tree_root reada_extents;
 };
 
 struct btrfs_fs_devices {
-- 
cgit 1.2.3-korg


From 7414a03fbf9e75fbbf2a3c16828cd862e572aa44 Mon Sep 17 00:00:00 2001
From: Arne Jansen <sensille@gmx.net>
Date: Mon, 23 May 2011 14:33:49 +0200
Subject: btrfs: initial readahead code and prototypes

This is the implementation for the generic read ahead framework.

To trigger a readahead, btrfs_reada_add must be called. It will start
a read ahead for the given range [start, end) on tree root. The returned
handle can either be used to wait on the readahead to finish
(btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).

The read ahead works as follows:
On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
reada_start_machine will then search for extents to prefetch and trigger
some reads. When a read finishes for a node, all contained node/leaf
pointers that lie in the given range will also be enqueued. The reads will
be triggered in sequential order, thus giving a big win over a naive
enumeration. It will also make use of multi-device layouts. Each disk
will have its on read pointer and all disks will by utilized in parallel.
Also will no two disks read both sides of a mirror simultaneously, as this
would waste seeking capacity. Instead both disks will read different parts
of the filesystem.
Any number of readaheads can be started in parallel. The read order will be
determined globally, i.e. 2 parallel readaheads will normally finish faster
than the 2 started one after another.

Changes v2:
 - protect root->node by transaction instead of node_lock
 - fix missed branches:
    The readahead had a too simple check to determine if a branch from
    a node should be checked or not. It now also records the upper bound
    of each node to see if the requested RA range lies within.
 - use KERN_CONT to debug output, to avoid line breaks
 - defer reada_start_machine to worker to avoid deadlock

Changes v3:
 - protect root->node by rcu

Changes v5:
 - changed EIO-semantics of reada_tree_block_flagged
 - remove spin_lock from reada_control and make elems an atomic_t
 - remove unused read_total from reada_control
 - kill reada_key_cmp, use btrfs_comp_cpu_keys instead
 - use kref-style release functions where possible
 - return struct reada_control * instead of void * from btrfs_reada_add

Signed-off-by: Arne Jansen <sensille@gmx.net>
---
 fs/btrfs/Makefile |   3 +-
 fs/btrfs/ctree.h  |  16 +
 fs/btrfs/reada.c  | 949 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 967 insertions(+), 1 deletion(-)
 create mode 100644 fs/btrfs/reada.c

diff --git a/fs/btrfs/Makefile b/fs/btrfs/Makefile
index 40e6ac08c21ffb..bdd6fb238ce16a 100644
--- a/fs/btrfs/Makefile
+++ b/fs/btrfs/Makefile
@@ -7,6 +7,7 @@ btrfs-y += super.o ctree.o extent-tree.o print-tree.o root-tree.o dir-item.o \
 	   extent_map.o sysfs.o struct-funcs.o xattr.o ordered-data.o \
 	   extent_io.o volumes.o async-thread.o ioctl.o locking.o orphan.o \
 	   export.o tree-log.o free-space-cache.o zlib.o lzo.o \
-	   compression.o delayed-ref.o relocation.o delayed-inode.o scrub.o
+	   compression.o delayed-ref.o relocation.o delayed-inode.o scrub.o \
+	   reada.o
 
 btrfs-$(CONFIG_BTRFS_FS_POSIX_ACL) += acl.o
diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
index f71fd24cc152d4..370af767440d66 100644
--- a/fs/btrfs/ctree.h
+++ b/fs/btrfs/ctree.h
@@ -2702,4 +2702,20 @@ int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid);
 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
 			 struct btrfs_scrub_progress *progress);
 
+/* reada.c */
+struct reada_control {
+	struct btrfs_root	*root;		/* tree to prefetch */
+	struct btrfs_key	key_start;
+	struct btrfs_key	key_end;	/* exclusive */
+	atomic_t		elems;
+	struct kref		refcnt;
+	wait_queue_head_t	wait;
+};
+struct reada_control *btrfs_reada_add(struct btrfs_root *root,
+			      struct btrfs_key *start, struct btrfs_key *end);
+int btrfs_reada_wait(void *handle);
+void btrfs_reada_detach(void *handle);
+int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
+			 u64 start, int err);
+
 #endif
diff --git a/fs/btrfs/reada.c b/fs/btrfs/reada.c
new file mode 100644
index 00000000000000..2b701d08222749
--- /dev/null
+++ b/fs/btrfs/reada.c
@@ -0,0 +1,949 @@
+/*
+ * Copyright (C) 2011 STRATO.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License v2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public
+ * License along with this program; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 021110-1307, USA.
+ */
+
+#include <linux/sched.h>
+#include <linux/pagemap.h>
+#include <linux/writeback.h>
+#include <linux/blkdev.h>
+#include <linux/rbtree.h>
+#include <linux/slab.h>
+#include <linux/workqueue.h>
+#include "ctree.h"
+#include "volumes.h"
+#include "disk-io.h"
+#include "transaction.h"
+
+#undef DEBUG
+
+/*
+ * This is the implementation for the generic read ahead framework.
+ *
+ * To trigger a readahead, btrfs_reada_add must be called. It will start
+ * a read ahead for the given range [start, end) on tree root. The returned
+ * handle can either be used to wait on the readahead to finish
+ * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
+ *
+ * The read ahead works as follows:
+ * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
+ * reada_start_machine will then search for extents to prefetch and trigger
+ * some reads. When a read finishes for a node, all contained node/leaf
+ * pointers that lie in the given range will also be enqueued. The reads will
+ * be triggered in sequential order, thus giving a big win over a naive
+ * enumeration. It will also make use of multi-device layouts. Each disk
+ * will have its on read pointer and all disks will by utilized in parallel.
+ * Also will no two disks read both sides of a mirror simultaneously, as this
+ * would waste seeking capacity. Instead both disks will read different parts
+ * of the filesystem.
+ * Any number of readaheads can be started in parallel. The read order will be
+ * determined globally, i.e. 2 parallel readaheads will normally finish faster
+ * than the 2 started one after another.
+ */
+
+#define MAX_MIRRORS 2
+#define MAX_IN_FLIGHT 6
+
+struct reada_extctl {
+	struct list_head	list;
+	struct reada_control	*rc;
+	u64			generation;
+};
+
+struct reada_extent {
+	u64			logical;
+	struct btrfs_key	top;
+	u32			blocksize;
+	int			err;
+	struct list_head	extctl;
+	struct kref		refcnt;
+	spinlock_t		lock;
+	struct reada_zone	*zones[MAX_MIRRORS];
+	int			nzones;
+	struct btrfs_device	*scheduled_for;
+};
+
+struct reada_zone {
+	u64			start;
+	u64			end;
+	u64			elems;
+	struct list_head	list;
+	spinlock_t		lock;
+	int			locked;
+	struct btrfs_device	*device;
+	struct btrfs_device	*devs[MAX_MIRRORS]; /* full list, incl self */
+	int			ndevs;
+	struct kref		refcnt;
+};
+
+struct reada_machine_work {
+	struct btrfs_work	work;
+	struct btrfs_fs_info	*fs_info;
+};
+
+static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
+static void reada_control_release(struct kref *kref);
+static void reada_zone_release(struct kref *kref);
+static void reada_start_machine(struct btrfs_fs_info *fs_info);
+static void __reada_start_machine(struct btrfs_fs_info *fs_info);
+
+static int reada_add_block(struct reada_control *rc, u64 logical,
+			   struct btrfs_key *top, int level, u64 generation);
+
+/* recurses */
+/* in case of err, eb might be NULL */
+static int __readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
+			    u64 start, int err)
+{
+	int level = 0;
+	int nritems;
+	int i;
+	u64 bytenr;
+	u64 generation;
+	struct reada_extent *re;
+	struct btrfs_fs_info *fs_info = root->fs_info;
+	struct list_head list;
+	unsigned long index = start >> PAGE_CACHE_SHIFT;
+	struct btrfs_device *for_dev;
+
+	if (eb)
+		level = btrfs_header_level(eb);
+
+	/* find extent */
+	spin_lock(&fs_info->reada_lock);
+	re = radix_tree_lookup(&fs_info->reada_tree, index);
+	if (re)
+		kref_get(&re->refcnt);
+	spin_unlock(&fs_info->reada_lock);
+
+	if (!re)
+		return -1;
+
+	spin_lock(&re->lock);
+	/*
+	 * just take the full list from the extent. afterwards we
+	 * don't need the lock anymore
+	 */
+	list_replace_init(&re->extctl, &list);
+	for_dev = re->scheduled_for;
+	re->scheduled_for = NULL;
+	spin_unlock(&re->lock);
+
+	if (err == 0) {
+		nritems = level ? btrfs_header_nritems(eb) : 0;
+		generation = btrfs_header_generation(eb);
+		/*
+		 * FIXME: currently we just set nritems to 0 if this is a leaf,
+		 * effectively ignoring the content. In a next step we could
+		 * trigger more readahead depending from the content, e.g.
+		 * fetch the checksums for the extents in the leaf.
+		 */
+	} else {
+		/*
+		 * this is the error case, the extent buffer has not been
+		 * read correctly. We won't access anything from it and
+		 * just cleanup our data structures. Effectively this will
+		 * cut the branch below this node from read ahead.
+		 */
+		nritems = 0;
+		generation = 0;
+	}
+
+	for (i = 0; i < nritems; i++) {
+		struct reada_extctl *rec;
+		u64 n_gen;
+		struct btrfs_key key;
+		struct btrfs_key next_key;
+
+		btrfs_node_key_to_cpu(eb, &key, i);
+		if (i + 1 < nritems)
+			btrfs_node_key_to_cpu(eb, &next_key, i + 1);
+		else
+			next_key = re->top;
+		bytenr = btrfs_node_blockptr(eb, i);
+		n_gen = btrfs_node_ptr_generation(eb, i);
+
+		list_for_each_entry(rec, &list, list) {
+			struct reada_control *rc = rec->rc;
+
+			/*
+			 * if the generation doesn't match, just ignore this
+			 * extctl. This will probably cut off a branch from
+			 * prefetch. Alternatively one could start a new (sub-)
+			 * prefetch for this branch, starting again from root.
+			 * FIXME: move the generation check out of this loop
+			 */
+#ifdef DEBUG
+			if (rec->generation != generation) {
+				printk(KERN_DEBUG "generation mismatch for "
+						"(%llu,%d,%llu) %llu != %llu\n",
+				       key.objectid, key.type, key.offset,
+				       rec->generation, generation);
+			}
+#endif
+			if (rec->generation == generation &&
+			    btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
+			    btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
+				reada_add_block(rc, bytenr, &next_key,
+						level - 1, n_gen);
+		}
+	}
+	/*
+	 * free extctl records
+	 */
+	while (!list_empty(&list)) {
+		struct reada_control *rc;
+		struct reada_extctl *rec;
+
+		rec = list_first_entry(&list, struct reada_extctl, list);
+		list_del(&rec->list);
+		rc = rec->rc;
+		kfree(rec);
+
+		kref_get(&rc->refcnt);
+		if (atomic_dec_and_test(&rc->elems)) {
+			kref_put(&rc->refcnt, reada_control_release);
+			wake_up(&rc->wait);
+		}
+		kref_put(&rc->refcnt, reada_control_release);
+
+		reada_extent_put(fs_info, re);	/* one ref for each entry */
+	}
+	reada_extent_put(fs_info, re);	/* our ref */
+	if (for_dev)
+		atomic_dec(&for_dev->reada_in_flight);
+
+	return 0;
+}
+
+/*
+ * start is passed separately in case eb in NULL, which may be the case with
+ * failed I/O
+ */
+int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
+			 u64 start, int err)
+{
+	int ret;
+
+	ret = __readahead_hook(root, eb, start, err);
+
+	reada_start_machine(root->fs_info);
+
+	return ret;
+}
+
+static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
+					  struct btrfs_device *dev, u64 logical,
+					  struct btrfs_multi_bio *multi)
+{
+	int ret;
+	int looped = 0;
+	struct reada_zone *zone;
+	struct btrfs_block_group_cache *cache = NULL;
+	u64 start;
+	u64 end;
+	int i;
+
+again:
+	zone = NULL;
+	spin_lock(&fs_info->reada_lock);
+	ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
+				     logical >> PAGE_CACHE_SHIFT, 1);
+	if (ret == 1)
+		kref_get(&zone->refcnt);
+	spin_unlock(&fs_info->reada_lock);
+
+	if (ret == 1) {
+		if (logical >= zone->start && logical < zone->end)
+			return zone;
+		spin_lock(&fs_info->reada_lock);
+		kref_put(&zone->refcnt, reada_zone_release);
+		spin_unlock(&fs_info->reada_lock);
+	}
+
+	if (looped)
+		return NULL;
+
+	cache = btrfs_lookup_block_group(fs_info, logical);
+	if (!cache)
+		return NULL;
+
+	start = cache->key.objectid;
+	end = start + cache->key.offset - 1;
+	btrfs_put_block_group(cache);
+
+	zone = kzalloc(sizeof(*zone), GFP_NOFS);
+	if (!zone)
+		return NULL;
+
+	zone->start = start;
+	zone->end = end;
+	INIT_LIST_HEAD(&zone->list);
+	spin_lock_init(&zone->lock);
+	zone->locked = 0;
+	kref_init(&zone->refcnt);
+	zone->elems = 0;
+	zone->device = dev; /* our device always sits at index 0 */
+	for (i = 0; i < multi->num_stripes; ++i) {
+		/* bounds have already been checked */
+		zone->devs[i] = multi->stripes[i].dev;
+	}
+	zone->ndevs = multi->num_stripes;
+
+	spin_lock(&fs_info->reada_lock);
+	ret = radix_tree_insert(&dev->reada_zones,
+				(unsigned long)zone->end >> PAGE_CACHE_SHIFT,
+				zone);
+	spin_unlock(&fs_info->reada_lock);
+
+	if (ret) {
+		kfree(zone);
+		looped = 1;
+		goto again;
+	}
+
+	return zone;
+}
+
+static struct reada_extent *reada_find_extent(struct btrfs_root *root,
+					      u64 logical,
+					      struct btrfs_key *top, int level)
+{
+	int ret;
+	int looped = 0;
+	struct reada_extent *re = NULL;
+	struct btrfs_fs_info *fs_info = root->fs_info;
+	struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
+	struct btrfs_multi_bio *multi = NULL;
+	struct btrfs_device *dev;
+	u32 blocksize;
+	u64 length;
+	int nzones = 0;
+	int i;
+	unsigned long index = logical >> PAGE_CACHE_SHIFT;
+
+again:
+	spin_lock(&fs_info->reada_lock);
+	re = radix_tree_lookup(&fs_info->reada_tree, index);
+	if (re)
+		kref_get(&re->refcnt);
+	spin_unlock(&fs_info->reada_lock);
+
+	if (re || looped)
+		return re;
+
+	re = kzalloc(sizeof(*re), GFP_NOFS);
+	if (!re)
+		return NULL;
+
+	blocksize = btrfs_level_size(root, level);
+	re->logical = logical;
+	re->blocksize = blocksize;
+	re->top = *top;
+	INIT_LIST_HEAD(&re->extctl);
+	spin_lock_init(&re->lock);
+	kref_init(&re->refcnt);
+
+	/*
+	 * map block
+	 */
+	length = blocksize;
+	ret = btrfs_map_block(map_tree, REQ_WRITE, logical, &length, &multi, 0);
+	if (ret || !multi || length < blocksize)
+		goto error;
+
+	if (multi->num_stripes > MAX_MIRRORS) {
+		printk(KERN_ERR "btrfs readahead: more than %d copies not "
+				"supported", MAX_MIRRORS);
+		goto error;
+	}
+
+	for (nzones = 0; nzones < multi->num_stripes; ++nzones) {
+		struct reada_zone *zone;
+
+		dev = multi->stripes[nzones].dev;
+		zone = reada_find_zone(fs_info, dev, logical, multi);
+		if (!zone)
+			break;
+
+		re->zones[nzones] = zone;
+		spin_lock(&zone->lock);
+		if (!zone->elems)
+			kref_get(&zone->refcnt);
+		++zone->elems;
+		spin_unlock(&zone->lock);
+		spin_lock(&fs_info->reada_lock);
+		kref_put(&zone->refcnt, reada_zone_release);
+		spin_unlock(&fs_info->reada_lock);
+	}
+	re->nzones = nzones;
+	if (nzones == 0) {
+		/* not a single zone found, error and out */
+		goto error;
+	}
+
+	/* insert extent in reada_tree + all per-device trees, all or nothing */
+	spin_lock(&fs_info->reada_lock);
+	ret = radix_tree_insert(&fs_info->reada_tree, index, re);
+	if (ret) {
+		spin_unlock(&fs_info->reada_lock);
+		if (ret != -ENOMEM) {
+			/* someone inserted the extent in the meantime */
+			looped = 1;
+		}
+		goto error;
+	}
+	for (i = 0; i < nzones; ++i) {
+		dev = multi->stripes[i].dev;
+		ret = radix_tree_insert(&dev->reada_extents, index, re);
+		if (ret) {
+			while (--i >= 0) {
+				dev = multi->stripes[i].dev;
+				BUG_ON(dev == NULL);
+				radix_tree_delete(&dev->reada_extents, index);
+			}
+			BUG_ON(fs_info == NULL);
+			radix_tree_delete(&fs_info->reada_tree, index);
+			spin_unlock(&fs_info->reada_lock);
+			goto error;
+		}
+	}
+	spin_unlock(&fs_info->reada_lock);
+
+	return re;
+
+error:
+	while (nzones) {
+		struct reada_zone *zone;
+
+		--nzones;
+		zone = re->zones[nzones];
+		kref_get(&zone->refcnt);
+		spin_lock(&zone->lock);
+		--zone->elems;
+		if (zone->elems == 0) {
+			/*
+			 * no fs_info->reada_lock needed, as this can't be
+			 * the last ref
+			 */
+			kref_put(&zone->refcnt, reada_zone_release);
+		}
+		spin_unlock(&zone->lock);
+
+		spin_lock(&fs_info->reada_lock);
+		kref_put(&zone->refcnt, reada_zone_release);
+		spin_unlock(&fs_info->reada_lock);
+	}
+	kfree(re);
+	if (looped)
+		goto again;
+	return NULL;
+}
+
+static void reada_kref_dummy(struct kref *kr)
+{
+}
+
+static void reada_extent_put(struct btrfs_fs_info *fs_info,
+			     struct reada_extent *re)
+{
+	int i;
+	unsigned long index = re->logical >> PAGE_CACHE_SHIFT;
+
+	spin_lock(&fs_info->reada_lock);
+	if (!kref_put(&re->refcnt, reada_kref_dummy)) {
+		spin_unlock(&fs_info->reada_lock);
+		return;
+	}
+
+	radix_tree_delete(&fs_info->reada_tree, index);
+	for (i = 0; i < re->nzones; ++i) {
+		struct reada_zone *zone = re->zones[i];
+
+		radix_tree_delete(&zone->device->reada_extents, index);
+	}
+
+	spin_unlock(&fs_info->reada_lock);
+
+	for (i = 0; i < re->nzones; ++i) {
+		struct reada_zone *zone = re->zones[i];
+
+		kref_get(&zone->refcnt);
+		spin_lock(&zone->lock);
+		--zone->elems;
+		if (zone->elems == 0) {
+			/* no fs_info->reada_lock needed, as this can't be
+			 * the last ref */
+			kref_put(&zone->refcnt, reada_zone_release);
+		}
+		spin_unlock(&zone->lock);
+
+		spin_lock(&fs_info->reada_lock);
+		kref_put(&zone->refcnt, reada_zone_release);
+		spin_unlock(&fs_info->reada_lock);
+	}
+	if (re->scheduled_for)
+		atomic_dec(&re->scheduled_for->reada_in_flight);
+
+	kfree(re);
+}
+
+static void reada_zone_release(struct kref *kref)
+{
+	struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
+
+	radix_tree_delete(&zone->device->reada_zones,
+			  zone->end >> PAGE_CACHE_SHIFT);
+
+	kfree(zone);
+}
+
+static void reada_control_release(struct kref *kref)
+{
+	struct reada_control *rc = container_of(kref, struct reada_control,
+						refcnt);
+
+	kfree(rc);
+}
+
+static int reada_add_block(struct reada_control *rc, u64 logical,
+			   struct btrfs_key *top, int level, u64 generation)
+{
+	struct btrfs_root *root = rc->root;
+	struct reada_extent *re;
+	struct reada_extctl *rec;
+
+	re = reada_find_extent(root, logical, top, level); /* takes one ref */
+	if (!re)
+		return -1;
+
+	rec = kzalloc(sizeof(*rec), GFP_NOFS);
+	if (!rec) {
+		reada_extent_put(root->fs_info, re);
+		return -1;
+	}
+
+	rec->rc = rc;
+	rec->generation = generation;
+	atomic_inc(&rc->elems);
+
+	spin_lock(&re->lock);
+	list_add_tail(&rec->list, &re->extctl);
+	spin_unlock(&re->lock);
+
+	/* leave the ref on the extent */
+
+	return 0;
+}
+
+/*
+ * called with fs_info->reada_lock held
+ */
+static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
+{
+	int i;
+	unsigned long index = zone->end >> PAGE_CACHE_SHIFT;
+
+	for (i = 0; i < zone->ndevs; ++i) {
+		struct reada_zone *peer;
+		peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
+		if (peer && peer->device != zone->device)
+			peer->locked = lock;
+	}
+}
+
+/*
+ * called with fs_info->reada_lock held
+ */
+static int reada_pick_zone(struct btrfs_device *dev)
+{
+	struct reada_zone *top_zone = NULL;
+	struct reada_zone *top_locked_zone = NULL;
+	u64 top_elems = 0;
+	u64 top_locked_elems = 0;
+	unsigned long index = 0;
+	int ret;
+
+	if (dev->reada_curr_zone) {
+		reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
+		kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
+		dev->reada_curr_zone = NULL;
+	}
+	/* pick the zone with the most elements */
+	while (1) {
+		struct reada_zone *zone;
+
+		ret = radix_tree_gang_lookup(&dev->reada_zones,
+					     (void **)&zone, index, 1);
+		if (ret == 0)
+			break;
+		index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
+		if (zone->locked) {
+			if (zone->elems > top_locked_elems) {
+				top_locked_elems = zone->elems;
+				top_locked_zone = zone;
+			}
+		} else {
+			if (zone->elems > top_elems) {
+				top_elems = zone->elems;
+				top_zone = zone;
+			}
+		}
+	}
+	if (top_zone)
+		dev->reada_curr_zone = top_zone;
+	else if (top_locked_zone)
+		dev->reada_curr_zone = top_locked_zone;
+	else
+		return 0;
+
+	dev->reada_next = dev->reada_curr_zone->start;
+	kref_get(&dev->reada_curr_zone->refcnt);
+	reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
+
+	return 1;
+}
+
+static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
+				   struct btrfs_device *dev)
+{
+	struct reada_extent *re = NULL;
+	int mirror_num = 0;
+	struct extent_buffer *eb = NULL;
+	u64 logical;
+	u32 blocksize;
+	int ret;
+	int i;
+	int need_kick = 0;
+
+	spin_lock(&fs_info->reada_lock);
+	if (dev->reada_curr_zone == NULL) {
+		ret = reada_pick_zone(dev);
+		if (!ret) {
+			spin_unlock(&fs_info->reada_lock);
+			return 0;
+		}
+	}
+	/*
+	 * FIXME currently we issue the reads one extent at a time. If we have
+	 * a contiguous block of extents, we could also coagulate them or use
+	 * plugging to speed things up
+	 */
+	ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
+				     dev->reada_next >> PAGE_CACHE_SHIFT, 1);
+	if (ret == 0 || re->logical >= dev->reada_curr_zone->end) {
+		ret = reada_pick_zone(dev);
+		if (!ret) {
+			spin_unlock(&fs_info->reada_lock);
+			return 0;
+		}
+		re = NULL;
+		ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
+					dev->reada_next >> PAGE_CACHE_SHIFT, 1);
+	}
+	if (ret == 0) {
+		spin_unlock(&fs_info->reada_lock);
+		return 0;
+	}
+	dev->reada_next = re->logical + re->blocksize;
+	kref_get(&re->refcnt);
+
+	spin_unlock(&fs_info->reada_lock);
+
+	/*
+	 * find mirror num
+	 */
+	for (i = 0; i < re->nzones; ++i) {
+		if (re->zones[i]->device == dev) {
+			mirror_num = i + 1;
+			break;
+		}
+	}
+	logical = re->logical;
+	blocksize = re->blocksize;
+
+	spin_lock(&re->lock);
+	if (re->scheduled_for == NULL) {
+		re->scheduled_for = dev;
+		need_kick = 1;
+	}
+	spin_unlock(&re->lock);
+
+	reada_extent_put(fs_info, re);
+
+	if (!need_kick)
+		return 0;
+
+	atomic_inc(&dev->reada_in_flight);
+	ret = reada_tree_block_flagged(fs_info->extent_root, logical, blocksize,
+			 mirror_num, &eb);
+	if (ret)
+		__readahead_hook(fs_info->extent_root, NULL, logical, ret);
+	else if (eb)
+		__readahead_hook(fs_info->extent_root, eb, eb->start, ret);
+
+	if (eb)
+		free_extent_buffer(eb);
+
+	return 1;
+
+}
+
+static void reada_start_machine_worker(struct btrfs_work *work)
+{
+	struct reada_machine_work *rmw;
+	struct btrfs_fs_info *fs_info;
+
+	rmw = container_of(work, struct reada_machine_work, work);
+	fs_info = rmw->fs_info;
+
+	kfree(rmw);
+
+	__reada_start_machine(fs_info);
+}
+
+static void __reada_start_machine(struct btrfs_fs_info *fs_info)
+{
+	struct btrfs_device *device;
+	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
+	u64 enqueued;
+	u64 total = 0;
+	int i;
+
+	do {
+		enqueued = 0;
+		list_for_each_entry(device, &fs_devices->devices, dev_list) {
+			if (atomic_read(&device->reada_in_flight) <
+			    MAX_IN_FLIGHT)
+				enqueued += reada_start_machine_dev(fs_info,
+								    device);
+		}
+		total += enqueued;
+	} while (enqueued && total < 10000);
+
+	if (enqueued == 0)
+		return;
+
+	/*
+	 * If everything is already in the cache, this is effectively single
+	 * threaded. To a) not hold the caller for too long and b) to utilize
+	 * more cores, we broke the loop above after 10000 iterations and now
+	 * enqueue to workers to finish it. This will distribute the load to
+	 * the cores.
+	 */
+	for (i = 0; i < 2; ++i)
+		reada_start_machine(fs_info);
+}
+
+static void reada_start_machine(struct btrfs_fs_info *fs_info)
+{
+	struct reada_machine_work *rmw;
+
+	rmw = kzalloc(sizeof(*rmw), GFP_NOFS);
+	if (!rmw) {
+		/* FIXME we cannot handle this properly right now */
+		BUG();
+	}
+	rmw->work.func = reada_start_machine_worker;
+	rmw->fs_info = fs_info;
+
+	btrfs_queue_worker(&fs_info->readahead_workers, &rmw->work);
+}
+
+#ifdef DEBUG
+static void dump_devs(struct btrfs_fs_info *fs_info, int all)
+{
+	struct btrfs_device *device;
+	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
+	unsigned long index;
+	int ret;
+	int i;
+	int j;
+	int cnt;
+
+	spin_lock(&fs_info->reada_lock);
+	list_for_each_entry(device, &fs_devices->devices, dev_list) {
+		printk(KERN_DEBUG "dev %lld has %d in flight\n", device->devid,
+			atomic_read(&device->reada_in_flight));
+		index = 0;
+		while (1) {
+			struct reada_zone *zone;
+			ret = radix_tree_gang_lookup(&device->reada_zones,
+						     (void **)&zone, index, 1);
+			if (ret == 0)
+				break;
+			printk(KERN_DEBUG "  zone %llu-%llu elems %llu locked "
+				"%d devs", zone->start, zone->end, zone->elems,
+				zone->locked);
+			for (j = 0; j < zone->ndevs; ++j) {
+				printk(KERN_CONT " %lld",
+					zone->devs[j]->devid);
+			}
+			if (device->reada_curr_zone == zone)
+				printk(KERN_CONT " curr off %llu",
+					device->reada_next - zone->start);
+			printk(KERN_CONT "\n");
+			index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
+		}
+		cnt = 0;
+		index = 0;
+		while (all) {
+			struct reada_extent *re = NULL;
+
+			ret = radix_tree_gang_lookup(&device->reada_extents,
+						     (void **)&re, index, 1);
+			if (ret == 0)
+				break;
+			printk(KERN_DEBUG
+				"  re: logical %llu size %u empty %d for %lld",
+				re->logical, re->blocksize,
+				list_empty(&re->extctl), re->scheduled_for ?
+				re->scheduled_for->devid : -1);
+
+			for (i = 0; i < re->nzones; ++i) {
+				printk(KERN_CONT " zone %llu-%llu devs",
+					re->zones[i]->start,
+					re->zones[i]->end);
+				for (j = 0; j < re->zones[i]->ndevs; ++j) {
+					printk(KERN_CONT " %lld",
+						re->zones[i]->devs[j]->devid);
+				}
+			}
+			printk(KERN_CONT "\n");
+			index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
+			if (++cnt > 15)
+				break;
+		}
+	}
+
+	index = 0;
+	cnt = 0;
+	while (all) {
+		struct reada_extent *re = NULL;
+
+		ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
+					     index, 1);
+		if (ret == 0)
+			break;
+		if (!re->scheduled_for) {
+			index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
+			continue;
+		}
+		printk(KERN_DEBUG
+			"re: logical %llu size %u list empty %d for %lld",
+			re->logical, re->blocksize, list_empty(&re->extctl),
+			re->scheduled_for ? re->scheduled_for->devid : -1);
+		for (i = 0; i < re->nzones; ++i) {
+			printk(KERN_CONT " zone %llu-%llu devs",
+				re->zones[i]->start,
+				re->zones[i]->end);
+			for (i = 0; i < re->nzones; ++i) {
+				printk(KERN_CONT " zone %llu-%llu devs",
+					re->zones[i]->start,
+					re->zones[i]->end);
+				for (j = 0; j < re->zones[i]->ndevs; ++j) {
+					printk(KERN_CONT " %lld",
+						re->zones[i]->devs[j]->devid);
+				}
+			}
+		}
+		printk(KERN_CONT "\n");
+		index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
+	}
+	spin_unlock(&fs_info->reada_lock);
+}
+#endif
+
+/*
+ * interface
+ */
+struct reada_control *btrfs_reada_add(struct btrfs_root *root,
+			struct btrfs_key *key_start, struct btrfs_key *key_end)
+{
+	struct reada_control *rc;
+	u64 start;
+	u64 generation;
+	int level;
+	struct extent_buffer *node;
+	static struct btrfs_key max_key = {
+		.objectid = (u64)-1,
+		.type = (u8)-1,
+		.offset = (u64)-1
+	};
+
+	rc = kzalloc(sizeof(*rc), GFP_NOFS);
+	if (!rc)
+		return ERR_PTR(-ENOMEM);
+
+	rc->root = root;
+	rc->key_start = *key_start;
+	rc->key_end = *key_end;
+	atomic_set(&rc->elems, 0);
+	init_waitqueue_head(&rc->wait);
+	kref_init(&rc->refcnt);
+	kref_get(&rc->refcnt); /* one ref for having elements */
+
+	node = btrfs_root_node(root);
+	start = node->start;
+	level = btrfs_header_level(node);
+	generation = btrfs_header_generation(node);
+	free_extent_buffer(node);
+
+	reada_add_block(rc, start, &max_key, level, generation);
+
+	reada_start_machine(root->fs_info);
+
+	return rc;
+}
+
+#ifdef DEBUG
+int btrfs_reada_wait(void *handle)
+{
+	struct reada_control *rc = handle;
+
+	while (atomic_read(&rc->elems)) {
+		wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
+				   5 * HZ);
+		dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
+	}
+
+	dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
+
+	kref_put(&rc->refcnt, reada_control_release);
+
+	return 0;
+}
+#else
+int btrfs_reada_wait(void *handle)
+{
+	struct reada_control *rc = handle;
+
+	while (atomic_read(&rc->elems)) {
+		wait_event(rc->wait, atomic_read(&rc->elems) == 0);
+	}
+
+	kref_put(&rc->refcnt, reada_control_release);
+
+	return 0;
+}
+#endif
+
+void btrfs_reada_detach(void *handle)
+{
+	struct reada_control *rc = handle;
+
+	kref_put(&rc->refcnt, reada_control_release);
+}
-- 
cgit 1.2.3-korg


From 4bb31e928d1a47f5bd046ecb176b8eff7c589fc0 Mon Sep 17 00:00:00 2001
From: Arne Jansen <sensille@gmx.net>
Date: Fri, 10 Jun 2011 13:55:54 +0200
Subject: btrfs: hooks for readahead

This adds the hooks needed for readahead. In the readpage_end_io_hook,
the extent state is checked for the EXTENT_READAHEAD flag. Only in this
case the readahead hook is called, to keep the impact on non-ra as low
as possible.
Additionally, a hook for a failed IO is added, otherwise readahead would
wait indefinitely for the extent to finish.

Changes for v2:
 - eliminate race condition

Signed-off-by: Arne Jansen <sensille@gmx.net>
---
 fs/btrfs/disk-io.c   | 37 +++++++++++++++++++++++++++++++++++++
 fs/btrfs/extent_io.c |  2 +-
 2 files changed, 38 insertions(+), 1 deletion(-)

diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index fd28c40e9ec353..2151828aa1423f 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -609,11 +609,47 @@ static int btree_readpage_end_io_hook(struct page *page, u64 start, u64 end,
 	end = min_t(u64, eb->len, PAGE_CACHE_SIZE);
 	end = eb->start + end - 1;
 err:
+	if (test_bit(EXTENT_BUFFER_READAHEAD, &eb->bflags)) {
+		clear_bit(EXTENT_BUFFER_READAHEAD, &eb->bflags);
+		btree_readahead_hook(root, eb, eb->start, ret);
+	}
+
 	free_extent_buffer(eb);
 out:
 	return ret;
 }
 
+static int btree_io_failed_hook(struct bio *failed_bio,
+			 struct page *page, u64 start, u64 end,
+			 struct extent_state *state)
+{
+	struct extent_io_tree *tree;
+	unsigned long len;
+	struct extent_buffer *eb;
+	struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
+
+	tree = &BTRFS_I(page->mapping->host)->io_tree;
+	if (page->private == EXTENT_PAGE_PRIVATE)
+		goto out;
+	if (!page->private)
+		goto out;
+
+	len = page->private >> 2;
+	WARN_ON(len == 0);
+
+	eb = alloc_extent_buffer(tree, start, len, page);
+	if (eb == NULL)
+		goto out;
+
+	if (test_bit(EXTENT_BUFFER_READAHEAD, &eb->bflags)) {
+		clear_bit(EXTENT_BUFFER_READAHEAD, &eb->bflags);
+		btree_readahead_hook(root, eb, eb->start, -EIO);
+	}
+
+out:
+	return -EIO;	/* we fixed nothing */
+}
+
 static void end_workqueue_bio(struct bio *bio, int err)
 {
 	struct end_io_wq *end_io_wq = bio->bi_private;
@@ -3166,6 +3202,7 @@ static int btrfs_cleanup_transaction(struct btrfs_root *root)
 static struct extent_io_ops btree_extent_io_ops = {
 	.write_cache_pages_lock_hook = btree_lock_page_hook,
 	.readpage_end_io_hook = btree_readpage_end_io_hook,
+	.readpage_io_failed_hook = btree_io_failed_hook,
 	.submit_bio_hook = btree_submit_bio_hook,
 	/* note we're sharing with inode.c for the merge bio hook */
 	.merge_bio_hook = btrfs_merge_bio_hook,
diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index 823028e73cf813..deba714236d782 100644
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -1731,7 +1731,7 @@ static void end_bio_extent_readpage(struct bio *bio, int err)
 		if (!uptodate && tree->ops &&
 		    tree->ops->readpage_io_failed_hook) {
 			ret = tree->ops->readpage_io_failed_hook(bio, page,
-							 start, end, NULL);
+							 start, end, state);
 			if (ret == 0) {
 				uptodate =
 					test_bit(BIO_UPTODATE, &bio->bi_flags);
-- 
cgit 1.2.3-korg


From 7a26285eea8eb92e0088db011571d887d4551b0f Mon Sep 17 00:00:00 2001
From: Arne Jansen <sensille@gmx.net>
Date: Fri, 10 Jun 2011 12:39:23 +0200
Subject: btrfs: use readahead API for scrub

Scrub uses a simple tree-enumeration to bring the relevant portions
of the extent- and csum-tree into the page cache before starting the
scrub-I/O. This is now replaced by using the new readahead-API.
During readahead the scrub is being accounted as paused, so it won't
hold off transaction commits.

This change raises the average disk bandwith utilisation on my test
volume from 70% to 90%. On another volume, the time for a test run
went down from 89s to 43s.

Changes v5:
 - reada1/2 are now of type struct reada_control *

Signed-off-by: Arne Jansen <sensille@gmx.net>
---
 fs/btrfs/scrub.c | 112 +++++++++++++++++++++++++------------------------------
 1 file changed, 50 insertions(+), 62 deletions(-)

diff --git a/fs/btrfs/scrub.c b/fs/btrfs/scrub.c
index a8d03d5efb5df3..f930f2776589c9 100644
--- a/fs/btrfs/scrub.c
+++ b/fs/btrfs/scrub.c
@@ -29,15 +29,12 @@
  * any can be found.
  *
  * Future enhancements:
- *  - To enhance the performance, better read-ahead strategies for the
- *    extent-tree can be employed.
  *  - In case an unrepairable extent is encountered, track which files are
  *    affected and report them
  *  - In case of a read error on files with nodatasum, map the file and read
  *    the extent to trigger a writeback of the good copy
  *  - track and record media errors, throw out bad devices
  *  - add a mode to also read unallocated space
- *  - make the prefetch cancellable
  */
 
 struct scrub_bio;
@@ -741,13 +738,16 @@ static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
 	int slot;
 	int i;
 	u64 nstripes;
-	int start_stripe;
 	struct extent_buffer *l;
 	struct btrfs_key key;
 	u64 physical;
 	u64 logical;
 	u64 generation;
 	u64 mirror_num;
+	struct reada_control *reada1;
+	struct reada_control *reada2;
+	struct btrfs_key key_start;
+	struct btrfs_key key_end;
 
 	u64 increment = map->stripe_len;
 	u64 offset;
@@ -779,81 +779,67 @@ static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
 	if (!path)
 		return -ENOMEM;
 
-	path->reada = 2;
 	path->search_commit_root = 1;
 	path->skip_locking = 1;
 
 	/*
-	 * find all extents for each stripe and just read them to get
-	 * them into the page cache
-	 * FIXME: we can do better. build a more intelligent prefetching
+	 * trigger the readahead for extent tree csum tree and wait for
+	 * completion. During readahead, the scrub is officially paused
+	 * to not hold off transaction commits
 	 */
 	logical = base + offset;
-	physical = map->stripes[num].physical;
-	ret = 0;
-	for (i = 0; i < nstripes; ++i) {
-		key.objectid = logical;
-		key.type = BTRFS_EXTENT_ITEM_KEY;
-		key.offset = (u64)0;
-
-		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
-		if (ret < 0)
-			goto out_noplug;
-
-		/*
-		 * we might miss half an extent here, but that doesn't matter,
-		 * as it's only the prefetch
-		 */
-		while (1) {
-			l = path->nodes[0];
-			slot = path->slots[0];
-			if (slot >= btrfs_header_nritems(l)) {
-				ret = btrfs_next_leaf(root, path);
-				if (ret == 0)
-					continue;
-				if (ret < 0)
-					goto out_noplug;
 
-				break;
-			}
-			btrfs_item_key_to_cpu(l, &key, slot);
+	wait_event(sdev->list_wait,
+		   atomic_read(&sdev->in_flight) == 0);
+	atomic_inc(&fs_info->scrubs_paused);
+	wake_up(&fs_info->scrub_pause_wait);
 
-			if (key.objectid >= logical + map->stripe_len)
-				break;
+	/* FIXME it might be better to start readahead at commit root */
+	key_start.objectid = logical;
+	key_start.type = BTRFS_EXTENT_ITEM_KEY;
+	key_start.offset = (u64)0;
+	key_end.objectid = base + offset + nstripes * increment;
+	key_end.type = BTRFS_EXTENT_ITEM_KEY;
+	key_end.offset = (u64)0;
+	reada1 = btrfs_reada_add(root, &key_start, &key_end);
+
+	key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
+	key_start.type = BTRFS_EXTENT_CSUM_KEY;
+	key_start.offset = logical;
+	key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
+	key_end.type = BTRFS_EXTENT_CSUM_KEY;
+	key_end.offset = base + offset + nstripes * increment;
+	reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
+
+	if (!IS_ERR(reada1))
+		btrfs_reada_wait(reada1);
+	if (!IS_ERR(reada2))
+		btrfs_reada_wait(reada2);
 
-			path->slots[0]++;
-		}
-		btrfs_release_path(path);
-		logical += increment;
-		physical += map->stripe_len;
-		cond_resched();
+	mutex_lock(&fs_info->scrub_lock);
+	while (atomic_read(&fs_info->scrub_pause_req)) {
+		mutex_unlock(&fs_info->scrub_lock);
+		wait_event(fs_info->scrub_pause_wait,
+		   atomic_read(&fs_info->scrub_pause_req) == 0);
+		mutex_lock(&fs_info->scrub_lock);
 	}
+	atomic_dec(&fs_info->scrubs_paused);
+	mutex_unlock(&fs_info->scrub_lock);
+	wake_up(&fs_info->scrub_pause_wait);
 
 	/*
 	 * collect all data csums for the stripe to avoid seeking during
 	 * the scrub. This might currently (crc32) end up to be about 1MB
 	 */
-	start_stripe = 0;
 	blk_start_plug(&plug);
-again:
-	logical = base + offset + start_stripe * increment;
-	for (i = start_stripe; i < nstripes; ++i) {
-		ret = btrfs_lookup_csums_range(csum_root, logical,
-					       logical + map->stripe_len - 1,
-					       &sdev->csum_list, 1);
-		if (ret)
-			goto out;
 
-		logical += increment;
-		cond_resched();
-	}
 	/*
 	 * now find all extents for each stripe and scrub them
 	 */
-	logical = base + offset + start_stripe * increment;
-	physical = map->stripes[num].physical + start_stripe * map->stripe_len;
+	logical = base + offset;
+	physical = map->stripes[num].physical;
 	ret = 0;
-	for (i = start_stripe; i < nstripes; ++i) {
+	for (i = 0; i < nstripes; ++i) {
 		/*
 		 * canceled?
 		 */
@@ -882,11 +868,14 @@ again:
 			atomic_dec(&fs_info->scrubs_paused);
 			mutex_unlock(&fs_info->scrub_lock);
 			wake_up(&fs_info->scrub_pause_wait);
-			scrub_free_csums(sdev);
-			start_stripe = i;
-			goto again;
 		}
 
+		ret = btrfs_lookup_csums_range(csum_root, logical,
+					       logical + map->stripe_len - 1,
+					       &sdev->csum_list, 1);
+		if (ret)
+			goto out;
+
 		key.objectid = logical;
 		key.type = BTRFS_EXTENT_ITEM_KEY;
 		key.offset = (u64)0;
@@ -982,7 +971,6 @@ next:
 
 out:
 	blk_finish_plug(&plug);
-out_noplug:
 	btrfs_free_path(path);
 	return ret < 0 ? ret : 0;
 }
-- 
cgit 1.2.3-korg