dm vdo: add the compressed block bin packer

When blocks do not deduplicate, vdo will attempt to compress them. Up to 14
compressed blocks may be packed into a single data block (this limitation
is imposed by the block map). The packer implements a simple best-fit
packing algorithm and also manages the formatting and writing of compressed
blocks when bins fill.

Co-developed-by: J. corwin Coburn <corwin@hurlbutnet.net>
Signed-off-by: J. corwin Coburn <corwin@hurlbutnet.net>
Co-developed-by: Michael Sclafani <dm-devel@lists.linux.dev>
Signed-off-by: Michael Sclafani <dm-devel@lists.linux.dev>
Co-developed-by: Sweet Tea Dorminy <sweettea-kernel@dorminy.me>
Signed-off-by: Sweet Tea Dorminy <sweettea-kernel@dorminy.me>
Signed-off-by: Matthew Sakai <msakai@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@kernel.org>

2 files changed, 906 insertions, 0 deletions

diff --git a/drivers/md/dm-vdo/packer.c b/drivers/md/dm-vdo/packer.c
new file mode 100644
index 00000000000000..f09eb1ca490259
--- /dev/null
+++ b/drivers/md/dm-vdo/packer.c
@@ -0,0 +1,784 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright 2023 Red Hat
+ */
+
+#include "packer.h"
+
+#include <linux/atomic.h>
+#include <linux/blkdev.h>
+
+#include "logger.h"
+#include "memory-alloc.h"
+#include "permassert.h"
+#include "string-utils.h"
+
+#include "admin-state.h"
+#include "completion.h"
+#include "constants.h"
+#include "data-vio.h"
+#include "dedupe.h"
+#include "encodings.h"
+#include "io-submitter.h"
+#include "physical-zone.h"
+#include "status-codes.h"
+#include "vdo.h"
+#include "vio.h"
+
+static const struct version_number COMPRESSED_BLOCK_1_0 = {
+	.major_version = 1,
+	.minor_version = 0,
+};
+
+enum {
+	COMPRESSED_BLOCK_1_0_SIZE = 4 + 4 + (2 * VDO_MAX_COMPRESSION_SLOTS),
+};
+
+/**
+ * vdo_get_compressed_block_fragment() - Get a reference to a compressed fragment from a compressed
+ *                                       block.
+ * @mapping_state [in] The mapping state for the look up.
+ * @compressed_block [in] The compressed block that was read from disk.
+ * @fragment_offset [out] The offset of the fragment within a compressed block.
+ * @fragment_size [out] The size of the fragment.
+ *
+ * Return: If a valid compressed fragment is found, VDO_SUCCESS; otherwise, VDO_INVALID_FRAGMENT if
+ *         the fragment is invalid.
+ */
+int vdo_get_compressed_block_fragment(enum block_mapping_state mapping_state,
+				      struct compressed_block *block,
+				      u16 *fragment_offset, u16 *fragment_size)
+{
+	u16 compressed_size;
+	u16 offset = 0;
+	unsigned int i;
+	u8 slot;
+	struct version_number version;
+
+	if (!vdo_is_state_compressed(mapping_state))
+		return VDO_INVALID_FRAGMENT;
+
+	version = vdo_unpack_version_number(block->header.version);
+	if (!vdo_are_same_version(version, COMPRESSED_BLOCK_1_0))
+		return VDO_INVALID_FRAGMENT;
+
+	slot = mapping_state - VDO_MAPPING_STATE_COMPRESSED_BASE;
+	if (slot >= VDO_MAX_COMPRESSION_SLOTS)
+		return VDO_INVALID_FRAGMENT;
+
+	compressed_size = __le16_to_cpu(block->header.sizes[slot]);
+	for (i = 0; i < slot; i++) {
+		offset += __le16_to_cpu(block->header.sizes[i]);
+		if (offset >= VDO_COMPRESSED_BLOCK_DATA_SIZE)
+			return VDO_INVALID_FRAGMENT;
+	}
+
+	if ((offset + compressed_size) > VDO_COMPRESSED_BLOCK_DATA_SIZE)
+		return VDO_INVALID_FRAGMENT;
+
+	*fragment_offset = offset;
+	*fragment_size = compressed_size;
+	return VDO_SUCCESS;
+}
+
+/**
+ * assert_on_packer_thread() - Check that we are on the packer thread.
+ * @packer: The packer.
+ * @caller: The function which is asserting.
+ */
+static inline void assert_on_packer_thread(struct packer *packer, const char *caller)
+{
+	ASSERT_LOG_ONLY((vdo_get_callback_thread_id() == packer->thread_id),
+			"%s() called from packer thread", caller);
+}
+
+/**
+ * insert_in_sorted_list() - Insert a bin to the list.
+ * @packer: The packer.
+ * @bin: The bin to move to its sorted position.
+ *
+ * The list is in ascending order of free space. Since all bins are already in the list, this
+ * actually moves the bin to the correct position in the list.
+ */
+static void insert_in_sorted_list(struct packer *packer, struct packer_bin *bin)
+{
+	struct packer_bin *active_bin;
+
+	list_for_each_entry(active_bin, &packer->bins, list)
+		if (active_bin->free_space > bin->free_space) {
+			list_move_tail(&bin->list, &active_bin->list);
+			return;
+		}
+
+	list_move_tail(&bin->list, &packer->bins);
+}
+
+/**
+ * make_bin() - Allocate a bin and put it into the packer's list.
+ * @packer: The packer.
+ */
+static int __must_check make_bin(struct packer *packer)
+{
+	struct packer_bin *bin;
+	int result;
+
+	result = uds_allocate_extended(struct packer_bin, VDO_MAX_COMPRESSION_SLOTS,
+				       struct vio *, __func__, &bin);
+	if (result != VDO_SUCCESS)
+		return result;
+
+	bin->free_space = VDO_COMPRESSED_BLOCK_DATA_SIZE;
+	INIT_LIST_HEAD(&bin->list);
+	list_add_tail(&bin->list, &packer->bins);
+	return VDO_SUCCESS;
+}
+
+/**
+ * vdo_make_packer() - Make a new block packer.
+ *
+ * @vdo: The vdo to which this packer belongs.
+ * @bin_count: The number of partial bins to keep in memory.
+ * @packer_ptr: A pointer to hold the new packer.
+ *
+ * Return: VDO_SUCCESS or an error
+ */
+int vdo_make_packer(struct vdo *vdo, block_count_t bin_count, struct packer **packer_ptr)
+{
+	struct packer *packer;
+	block_count_t i;
+	int result;
+
+	result = uds_allocate(1, struct packer, __func__, &packer);
+	if (result != VDO_SUCCESS)
+		return result;
+
+	packer->thread_id = vdo->thread_config.packer_thread;
+	packer->size = bin_count;
+	INIT_LIST_HEAD(&packer->bins);
+	vdo_set_admin_state_code(&packer->state, VDO_ADMIN_STATE_NORMAL_OPERATION);
+
+	for (i = 0; i < bin_count; i++) {
+		result = make_bin(packer);
+		if (result != VDO_SUCCESS) {
+			vdo_free_packer(packer);
+			return result;
+		}
+	}
+
+	/*
+	 * The canceled bin can hold up to half the number of user vios. Every canceled vio in the
+	 * bin must have a canceler for which it is waiting, and any canceler will only have
+	 * canceled one lock holder at a time.
+	 */
+	result = uds_allocate_extended(struct packer_bin, MAXIMUM_VDO_USER_VIOS / 2,
+				       struct vio *, __func__, &packer->canceled_bin);
+	if (result != VDO_SUCCESS) {
+		vdo_free_packer(packer);
+		return result;
+	}
+
+	result = vdo_make_default_thread(vdo, packer->thread_id);
+	if (result != VDO_SUCCESS) {
+		vdo_free_packer(packer);
+		return result;
+	}
+
+	*packer_ptr = packer;
+	return VDO_SUCCESS;
+}
+
+/**
+ * vdo_free_packer() - Free a block packer.
+ * @packer: The packer to free.
+ */
+void vdo_free_packer(struct packer *packer)
+{
+	struct packer_bin *bin, *tmp;
+
+	if (packer == NULL)
+		return;
+
+	list_for_each_entry_safe(bin, tmp, &packer->bins, list) {
+		list_del_init(&bin->list);
+		uds_free(bin);
+	}
+
+	uds_free(uds_forget(packer->canceled_bin));
+	uds_free(packer);
+}
+
+/**
+ * get_packer_from_data_vio() - Get the packer from a data_vio.
+ * @data_vio: The data_vio.
+ *
+ * Return: The packer from the VDO to which the data_vio belongs.
+ */
+static inline struct packer *get_packer_from_data_vio(struct data_vio *data_vio)
+{
+	return vdo_from_data_vio(data_vio)->packer;
+}
+
+/**
+ * vdo_get_packer_statistics() - Get the current statistics from the packer.
+ * @packer: The packer to query.
+ *
+ * Return: a copy of the current statistics for the packer.
+ */
+struct packer_statistics vdo_get_packer_statistics(const struct packer *packer)
+{
+	const struct packer_statistics *stats = &packer->statistics;
+
+	return (struct packer_statistics) {
+		.compressed_fragments_written = READ_ONCE(stats->compressed_fragments_written),
+		.compressed_blocks_written = READ_ONCE(stats->compressed_blocks_written),
+		.compressed_fragments_in_packer = READ_ONCE(stats->compressed_fragments_in_packer),
+	};
+}
+
+/**
+ * abort_packing() - Abort packing a data_vio.
+ * @data_vio: The data_vio to abort.
+ */
+static void abort_packing(struct data_vio *data_vio)
+{
+	struct packer *packer = get_packer_from_data_vio(data_vio);
+
+	WRITE_ONCE(packer->statistics.compressed_fragments_in_packer,
+		   packer->statistics.compressed_fragments_in_packer - 1);
+
+	write_data_vio(data_vio);
+}
+
+/**
+ * release_compressed_write_waiter() - Update a data_vio for which a successful compressed write
+ *                                     has completed and send it on its way.
+
+ * @data_vio: The data_vio to release.
+ * @allocation: The allocation to which the compressed block was written.
+ */
+static void release_compressed_write_waiter(struct data_vio *data_vio,
+					    struct allocation *allocation)
+{
+	data_vio->new_mapped = (struct zoned_pbn) {
+		.pbn = allocation->pbn,
+		.zone = allocation->zone,
+		.state = data_vio->compression.slot + VDO_MAPPING_STATE_COMPRESSED_BASE,
+	};
+
+	vdo_share_compressed_write_lock(data_vio, allocation->lock);
+	update_metadata_for_data_vio_write(data_vio, allocation->lock);
+}
+
+/**
+ * finish_compressed_write() - Finish a compressed block write.
+ * @completion: The compressed write completion.
+ *
+ * This callback is registered in continue_after_allocation().
+ */
+static void finish_compressed_write(struct vdo_completion *completion)
+{
+	struct data_vio *agent = as_data_vio(completion);
+	struct data_vio *client, *next;
+
+	assert_data_vio_in_allocated_zone(agent);
+
+	/*
+	 * Process all the non-agent waiters first to ensure that the pbn lock can not be released
+	 * until all of them have had a chance to journal their increfs.
+	 */
+	for (client = agent->compression.next_in_batch; client != NULL; client = next) {
+		next = client->compression.next_in_batch;
+		release_compressed_write_waiter(client, &agent->allocation);
+	}
+
+	completion->error_handler = handle_data_vio_error;
+	release_compressed_write_waiter(agent, &agent->allocation);
+}
+
+static void handle_compressed_write_error(struct vdo_completion *completion)
+{
+	struct data_vio *agent = as_data_vio(completion);
+	struct allocation *allocation = &agent->allocation;
+	struct data_vio *client, *next;
+
+	if (vdo_requeue_completion_if_needed(completion, allocation->zone->thread_id))
+		return;
+
+	update_vio_error_stats(as_vio(completion),
+			       "Completing compressed write vio for physical block %llu with error",
+			       (unsigned long long) allocation->pbn);
+
+	for (client = agent->compression.next_in_batch; client != NULL; client = next) {
+		next = client->compression.next_in_batch;
+		write_data_vio(client);
+	}
+
+	/* Now that we've released the batch from the packer, forget the error and continue on. */
+	vdo_reset_completion(completion);
+	completion->error_handler = handle_data_vio_error;
+	write_data_vio(agent);
+}
+
+/**
+ * add_to_bin() - Put a data_vio in a specific packer_bin in which it will definitely fit.
+ * @bin: The bin in which to put the data_vio.
+ * @data_vio: The data_vio to add.
+ */
+static void add_to_bin(struct packer_bin *bin, struct data_vio *data_vio)
+{
+	data_vio->compression.bin = bin;
+	data_vio->compression.slot = bin->slots_used;
+	bin->incoming[bin->slots_used++] = data_vio;
+}
+
+/**
+ * remove_from_bin() - Get the next data_vio whose compression has not been canceled from a bin.
+ * @packer: The packer.
+ * @bin: The bin from which to get a data_vio.
+ *
+ * Any canceled data_vios will be moved to the canceled bin.
+ * Return: An uncanceled data_vio from the bin or NULL if there are none.
+ */
+static struct data_vio *remove_from_bin(struct packer *packer, struct packer_bin *bin)
+{
+	while (bin->slots_used > 0) {
+		struct data_vio *data_vio = bin->incoming[--bin->slots_used];
+
+		if (!advance_data_vio_compression_stage(data_vio).may_not_compress) {
+			data_vio->compression.bin = NULL;
+			return data_vio;
+		}
+
+		add_to_bin(packer->canceled_bin, data_vio);
+	}
+
+	/* The bin is now empty. */
+	bin->free_space = VDO_COMPRESSED_BLOCK_DATA_SIZE;
+	return NULL;
+}
+
+/**
+ * initialize_compressed_block() - Initialize a compressed block.
+ * @block: The compressed block to initialize.
+ * @size: The size of the agent's fragment.
+ *
+ * This method initializes the compressed block in the compressed write agent. Because the
+ * compressor already put the agent's compressed fragment at the start of the compressed block's
+ * data field, it needn't be copied. So all we need do is initialize the header and set the size of
+ * the agent's fragment.
+ */
+static void initialize_compressed_block(struct compressed_block *block, u16 size)
+{
+	/*
+	 * Make sure the block layout isn't accidentally changed by changing the length of the
+	 * block header.
+	 */
+	BUILD_BUG_ON(sizeof(struct compressed_block_header) != COMPRESSED_BLOCK_1_0_SIZE);
+
+	block->header.version = vdo_pack_version_number(COMPRESSED_BLOCK_1_0);
+	block->header.sizes[0] = __cpu_to_le16(size);
+}
+
+/**
+ * pack_fragment() - Pack a data_vio's fragment into the compressed block in which it is already
+ *                   known to fit.
+ * @compression: The agent's compression_state to pack in to.
+ * @data_vio: The data_vio to pack.
+ * @offset: The offset into the compressed block at which to pack the fragment.
+ * @compressed_block: The compressed block which will be written out when batch is fully packed.
+ *
+ * Return: The new amount of space used.
+ */
+static block_size_t __must_check pack_fragment(struct compression_state *compression,
+					       struct data_vio *data_vio,
+					       block_size_t offset, slot_number_t slot,
+					       struct compressed_block *block)
+{
+	struct compression_state *to_pack = &data_vio->compression;
+	char *fragment = to_pack->block->data;
+
+	to_pack->next_in_batch = compression->next_in_batch;
+	compression->next_in_batch = data_vio;
+	to_pack->slot = slot;
+	block->header.sizes[slot] = __cpu_to_le16(to_pack->size);
+	memcpy(&block->data[offset], fragment, to_pack->size);
+	return (offset + to_pack->size);
+}
+
+/**
+ * compressed_write_end_io() - The bio_end_io for a compressed block write.
+ * @bio: The bio for the compressed write.
+ */
+static void compressed_write_end_io(struct bio *bio)
+{
+	struct data_vio *data_vio = vio_as_data_vio(bio->bi_private);
+
+	vdo_count_completed_bios(bio);
+	set_data_vio_allocated_zone_callback(data_vio, finish_compressed_write);
+	continue_data_vio_with_error(data_vio, blk_status_to_errno(bio->bi_status));
+}
+
+/**
+ * write_bin() - Write out a bin.
+ * @packer: The packer.
+ * @bin: The bin to write.
+ */
+static void write_bin(struct packer *packer, struct packer_bin *bin)
+{
+	int result;
+	block_size_t offset;
+	slot_number_t slot = 1;
+	struct compression_state *compression;
+	struct compressed_block *block;
+	struct data_vio *agent = remove_from_bin(packer, bin);
+	struct data_vio *client;
+	struct packer_statistics *stats;
+
+	if (agent == NULL)
+		return;
+
+	compression = &agent->compression;
+	compression->slot = 0;
+	block = compression->block;
+	initialize_compressed_block(block, compression->size);
+	offset = compression->size;
+
+	while ((client = remove_from_bin(packer, bin)) != NULL)
+		offset = pack_fragment(compression, client, offset, slot++, block);
+
+	/*
+	 * If the batch contains only a single vio, then we save nothing by saving the compressed
+	 * form. Continue processing the single vio in the batch.
+	 */
+	if (slot == 1) {
+		abort_packing(agent);
+		return;
+	}
+
+	if (slot < VDO_MAX_COMPRESSION_SLOTS) {
+		/* Clear out the sizes of the unused slots */
+		memset(&block->header.sizes[slot], 0,
+		       (VDO_MAX_COMPRESSION_SLOTS - slot) * sizeof(__le16));
+	}
+
+	agent->vio.completion.error_handler = handle_compressed_write_error;
+	if (vdo_is_read_only(vdo_from_data_vio(agent))) {
+		continue_data_vio_with_error(agent, VDO_READ_ONLY);
+		return;
+	}
+
+	result = vio_reset_bio(&agent->vio, (char *) block, compressed_write_end_io,
+			       REQ_OP_WRITE, agent->allocation.pbn);
+	if (result != VDO_SUCCESS) {
+		continue_data_vio_with_error(agent, result);
+		return;
+	}
+
+	/*
+	 * Once the compressed write is submitted, the fragments are no longer in the packer, so
+	 * update stats now.
+	 */
+	stats = &packer->statistics;
+	WRITE_ONCE(stats->compressed_fragments_in_packer,
+		   (stats->compressed_fragments_in_packer - slot));
+	WRITE_ONCE(stats->compressed_fragments_written,
+		   (stats->compressed_fragments_written + slot));
+	WRITE_ONCE(stats->compressed_blocks_written,
+		   stats->compressed_blocks_written + 1);
+
+	submit_data_vio_io(agent);
+}
+
+/**
+ * add_data_vio_to_packer_bin() - Add a data_vio to a bin's incoming queue
+ * @packer: The packer.
+ * @bin: The bin to which to add the data_vio.
+ * @data_vio: The data_vio to add to the bin's queue.
+ *
+ * Adds a data_vio to a bin's incoming queue, handles logical space change, and calls physical
+ * space processor.
+ */
+static void add_data_vio_to_packer_bin(struct packer *packer, struct packer_bin *bin,
+				       struct data_vio *data_vio)
+{
+	/* If the selected bin doesn't have room, start a new batch to make room. */
+	if (bin->free_space < data_vio->compression.size)
+		write_bin(packer, bin);
+
+	add_to_bin(bin, data_vio);
+	bin->free_space -= data_vio->compression.size;
+
+	/* If we happen to exactly fill the bin, start a new batch. */
+	if ((bin->slots_used == VDO_MAX_COMPRESSION_SLOTS) ||
+	    (bin->free_space == 0))
+		write_bin(packer, bin);
+
+	/* Now that we've finished changing the free space, restore the sort order. */
+	insert_in_sorted_list(packer, bin);
+}
+
+/**
+ * select_bin() - Select the bin that should be used to pack the compressed data in a data_vio with
+ *                other data_vios.
+ * @packer: The packer.
+ * @data_vio: The data_vio.
+ */
+static struct packer_bin * __must_check select_bin(struct packer *packer,
+						   struct data_vio *data_vio)
+{
+	/*
+	 * First best fit: select the bin with the least free space that has enough room for the
+	 * compressed data in the data_vio.
+	 */
+	struct packer_bin *bin, *fullest_bin;
+
+	list_for_each_entry(bin, &packer->bins, list) {
+		if (bin->free_space >= data_vio->compression.size)
+			return bin;
+	}
+
+	/*
+	 * None of the bins have enough space for the data_vio. We're not allowed to create new
+	 * bins, so we have to overflow one of the existing bins. It's pretty intuitive to select
+	 * the fullest bin, since that "wastes" the least amount of free space in the compressed
+	 * block. But if the space currently used in the fullest bin is smaller than the compressed
+	 * size of the incoming block, it seems wrong to force that bin to write when giving up on
+	 * compressing the incoming data_vio would likewise "waste" the least amount of free space.
+	 */
+	fullest_bin = list_first_entry(&packer->bins, struct packer_bin, list);
+	if (data_vio->compression.size >=
+	    (VDO_COMPRESSED_BLOCK_DATA_SIZE - fullest_bin->free_space))
+		return NULL;
+
+	/*
+	 * The fullest bin doesn't have room, but writing it out and starting a new batch with the
+	 * incoming data_vio will increase the packer's free space.
+	 */
+	return fullest_bin;
+}
+
+/**
+ * vdo_attempt_packing() - Attempt to rewrite the data in this data_vio as part of a compressed
+ *                         block.
+ * @data_vio: The data_vio to pack.
+ */
+void vdo_attempt_packing(struct data_vio *data_vio)
+{
+	int result;
+	struct packer_bin *bin;
+	struct data_vio_compression_status status = get_data_vio_compression_status(data_vio);
+	struct packer *packer = get_packer_from_data_vio(data_vio);
+
+	assert_on_packer_thread(packer, __func__);
+
+	result = ASSERT((status.stage == DATA_VIO_COMPRESSING),
+			"attempt to pack data_vio not ready for packing, stage: %u",
+			status.stage);
+	if (result != VDO_SUCCESS)
+		return;
+
+	/*
+	 * Increment whether or not this data_vio will be packed or not since abort_packing()
+	 * always decrements the counter.
+	 */
+	WRITE_ONCE(packer->statistics.compressed_fragments_in_packer,
+		   packer->statistics.compressed_fragments_in_packer + 1);
+
+	/*
+	 * If packing of this data_vio is disallowed for administrative reasons, give up before
+	 * making any state changes.
+	 */
+	if (!vdo_is_state_normal(&packer->state) ||
+	    (data_vio->flush_generation < packer->flush_generation)) {
+		abort_packing(data_vio);
+		return;
+	}
+
+	/*
+	 * The check of may_vio_block_in_packer() here will set the data_vio's compression state to
+	 * VIO_PACKING if the data_vio is allowed to be compressed (if it has already been
+	 * canceled, we'll fall out here). Once the data_vio is in the VIO_PACKING state, it must
+	 * be guaranteed to be put in a bin before any more requests can be processed by the packer
+	 * thread. Otherwise, a canceling data_vio could attempt to remove the canceled data_vio
+	 * from the packer and fail to rendezvous with it (VDO-2809). We must also make sure that
+	 * we will actually bin the data_vio and not give up on it as being larger than the space
+	 * used in the fullest bin. Hence we must call select_bin() before calling
+	 * may_vio_block_in_packer() (VDO-2826).
+	 */
+	bin = select_bin(packer, data_vio);
+	if ((bin == NULL) ||
+	    (advance_data_vio_compression_stage(data_vio).stage != DATA_VIO_PACKING)) {
+		abort_packing(data_vio);
+		return;
+	}
+
+	add_data_vio_to_packer_bin(packer, bin, data_vio);
+}
+
+/**
+ * check_for_drain_complete() - Check whether the packer has drained.
+ * @packer: The packer.
+ */
+static void check_for_drain_complete(struct packer *packer)
+{
+	if (vdo_is_state_draining(&packer->state) && (packer->canceled_bin->slots_used == 0))
+		vdo_finish_draining(&packer->state);
+}
+
+/**
+ * write_all_non_empty_bins() - Write out all non-empty bins on behalf of a flush or suspend.
+ * @packer: The packer being flushed.
+ */
+static void write_all_non_empty_bins(struct packer *packer)
+{
+	struct packer_bin *bin;
+
+	list_for_each_entry(bin, &packer->bins, list)
+		write_bin(packer, bin);
+		/*
+		 * We don't need to re-sort the bin here since this loop will make every bin have
+		 * the same amount of free space, so every ordering is sorted.
+		 */
+
+	check_for_drain_complete(packer);
+}
+
+/**
+ * vdo_flush_packer() - Request that the packer flush asynchronously.
+ * @packer: The packer to flush.
+ *
+ * All bins with at least two compressed data blocks will be written out, and any solitary pending
+ * VIOs will be released from the packer. While flushing is in progress, any VIOs submitted to
+ * vdo_attempt_packing() will be continued immediately without attempting to pack them.
+ */
+void vdo_flush_packer(struct packer *packer)
+{
+	assert_on_packer_thread(packer, __func__);
+	if (vdo_is_state_normal(&packer->state))
+		write_all_non_empty_bins(packer);
+}
+
+/**
+ * vdo_remove_lock_holder_from_packer() - Remove a lock holder from the packer.
+ * @completion: The data_vio which needs a lock held by a data_vio in the packer. The data_vio's
+ *              compression.lock_holder field will point to the data_vio to remove.
+ */
+void vdo_remove_lock_holder_from_packer(struct vdo_completion *completion)
+{
+	struct data_vio *data_vio = as_data_vio(completion);
+	struct packer *packer = get_packer_from_data_vio(data_vio);
+	struct data_vio *lock_holder;
+	struct packer_bin *bin;
+	slot_number_t slot;
+
+	assert_data_vio_in_packer_zone(data_vio);
+
+	lock_holder = uds_forget(data_vio->compression.lock_holder);
+	bin = lock_holder->compression.bin;
+	ASSERT_LOG_ONLY((bin != NULL), "data_vio in packer has a bin");
+
+	slot = lock_holder->compression.slot;
+	bin->slots_used--;
+	if (slot < bin->slots_used) {
+		bin->incoming[slot] = bin->incoming[bin->slots_used];
+		bin->incoming[slot]->compression.slot = slot;
+	}
+
+	lock_holder->compression.bin = NULL;
+	lock_holder->compression.slot = 0;
+
+	if (bin != packer->canceled_bin) {
+		bin->free_space += lock_holder->compression.size;
+		insert_in_sorted_list(packer, bin);
+	}
+
+	abort_packing(lock_holder);
+	check_for_drain_complete(packer);
+}
+
+/**
+ * vdo_increment_packer_flush_generation() - Increment the flush generation in the packer.
+ * @packer: The packer.
+ *
+ * This will also cause the packer to flush so that any VIOs from previous generations will exit
+ * the packer.
+ */
+void vdo_increment_packer_flush_generation(struct packer *packer)
+{
+	assert_on_packer_thread(packer, __func__);
+	packer->flush_generation++;
+	vdo_flush_packer(packer);
+}
+
+/**
+ * initiate_drain() - Initiate a drain.
+ *
+ * Implements vdo_admin_initiator_fn.
+ */
+static void initiate_drain(struct admin_state *state)
+{
+	struct packer *packer = container_of(state, struct packer, state);
+
+	write_all_non_empty_bins(packer);
+}
+
+/**
+ * vdo_drain_packer() - Drain the packer by preventing any more VIOs from entering the packer and
+ *                      then flushing.
+ * @packer: The packer to drain.
+ * @completion: The completion to finish when the packer has drained.
+ */
+void vdo_drain_packer(struct packer *packer, struct vdo_completion *completion)
+{
+	assert_on_packer_thread(packer, __func__);
+	vdo_start_draining(&packer->state, VDO_ADMIN_STATE_SUSPENDING, completion,
+			   initiate_drain);
+}
+
+/**
+ * vdo_resume_packer() - Resume a packer which has been suspended.
+ * @packer: The packer to resume.
+ * @parent: The completion to finish when the packer has resumed.
+ */
+void vdo_resume_packer(struct packer *packer, struct vdo_completion *parent)
+{
+	assert_on_packer_thread(packer, __func__);
+	vdo_continue_completion(parent, vdo_resume_if_quiescent(&packer->state));
+}
+
+static void dump_packer_bin(const struct packer_bin *bin, bool canceled)
+{
+	if (bin->slots_used == 0)
+		/* Don't dump empty bins. */
+		return;
+
+	uds_log_info("	  %sBin slots_used=%u free_space=%zu",
+		     (canceled ? "Canceled" : ""), bin->slots_used, bin->free_space);
+
+	/*
+	 * FIXME: dump vios in bin->incoming? The vios should have been dumped from the vio pool.
+	 * Maybe just dump their addresses so it's clear they're here?
+	 */
+}
+
+/**
+ * vdo_dump_packer() - Dump the packer.
+ * @packer: The packer.
+ *
+ * Context: dumps in a thread-unsafe fashion.
+ */
+void vdo_dump_packer(const struct packer *packer)
+{
+	struct packer_bin *bin;
+
+	uds_log_info("packer");
+	uds_log_info("	flushGeneration=%llu state %s  packer_bin_count=%llu",
+		     (unsigned long long) packer->flush_generation,
+		     vdo_get_admin_state_code(&packer->state)->name,
+		     (unsigned long long) packer->size);
+
+	list_for_each_entry(bin, &packer->bins, list)
+		dump_packer_bin(bin, false);
+
+	dump_packer_bin(packer->canceled_bin, true);
+}
diff --git a/drivers/md/dm-vdo/packer.h b/drivers/md/dm-vdo/packer.h
new file mode 100644
index 00000000000000..2dcc40bd44172d
--- /dev/null
+++ b/drivers/md/dm-vdo/packer.h
@@ -0,0 +1,122 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright 2023 Red Hat
+ */
+
+#ifndef VDO_PACKER_H
+#define VDO_PACKER_H
+
+#include <linux/list.h>
+
+#include "admin-state.h"
+#include "constants.h"
+#include "encodings.h"
+#include "statistics.h"
+#include "types.h"
+#include "wait-queue.h"
+
+enum {
+	DEFAULT_PACKER_BINS = 16,
+};
+
+/* The header of a compressed block. */
+struct compressed_block_header {
+	/* Unsigned 32-bit major and minor versions, little-endian */
+	struct packed_version_number version;
+
+	/* List of unsigned 16-bit compressed block sizes, little-endian */
+	__le16 sizes[VDO_MAX_COMPRESSION_SLOTS];
+} __packed;
+
+enum {
+	VDO_COMPRESSED_BLOCK_DATA_SIZE = VDO_BLOCK_SIZE - sizeof(struct compressed_block_header),
+
+	/*
+	 * A compressed block is only written if we can pack at least two fragments into it, so a
+	 * fragment which fills the entire data portion of a compressed block is too big.
+	 */
+	VDO_MAX_COMPRESSED_FRAGMENT_SIZE = VDO_COMPRESSED_BLOCK_DATA_SIZE - 1,
+};
+
+/* * The compressed block overlay. */
+struct compressed_block {
+	struct compressed_block_header header;
+	char data[VDO_COMPRESSED_BLOCK_DATA_SIZE];
+} __packed;
+
+/*
+ * Each packer_bin holds an incomplete batch of data_vios that only partially fill a compressed
+ * block. The bins are kept in a ring sorted by the amount of unused space so the first bin with
+ * enough space to hold a newly-compressed data_vio can easily be found. When the bin fills up or
+ * is flushed, the first uncanceled data_vio in the bin is selected to be the agent for that bin.
+ * Upon entering the packer, each data_vio already has its compressed data in the first slot of the
+ * data_vio's compressed_block (overlaid on the data_vio's scratch_block). So the agent's fragment
+ * is already in place. The fragments for the other uncanceled data_vios in the bin are packed into
+ * the agent's compressed block. The agent then writes out the compressed block. If the write is
+ * successful, the agent shares its pbn lock which each of the other data_vios in its compressed
+ * block and sends each on its way. Finally the agent itself continues on the write path as before.
+ *
+ * There is one special bin which is used to hold data_vios which have been canceled and removed
+ * from their bin by the packer. These data_vios need to wait for the canceller to rendezvous with
+ * them (VDO-2809) and so they sit in this special bin.
+ */
+struct packer_bin {
+	/* List links for packer.packer_bins */
+	struct list_head list;
+	/* The number of items in the bin */
+	slot_number_t slots_used;
+	/* The number of compressed block bytes remaining in the current batch */
+	size_t free_space;
+	/* The current partial batch of data_vios, waiting for more */
+	struct data_vio *incoming[];
+};
+
+struct packer {
+	/* The ID of the packer's callback thread */
+	thread_id_t thread_id;
+	/* The number of bins */
+	block_count_t size;
+	/* A list of all packer_bins, kept sorted by free_space */
+	struct list_head bins;
+	/*
+	 * A bin to hold data_vios which were canceled out of the packer and are waiting to
+	 * rendezvous with the canceling data_vio.
+	 */
+	struct packer_bin *canceled_bin;
+
+	/* The current flush generation */
+	sequence_number_t flush_generation;
+
+	/* The administrative state of the packer */
+	struct admin_state state;
+
+	/* Statistics are only updated on the packer thread, but are accessed from other threads */
+	struct packer_statistics statistics;
+};
+
+int vdo_get_compressed_block_fragment(enum block_mapping_state mapping_state,
+				      struct compressed_block *block,
+				      u16 *fragment_offset, u16 *fragment_size);
+
+int __must_check vdo_make_packer(struct vdo *vdo, block_count_t bin_count,
+				 struct packer **packer_ptr);
+
+void vdo_free_packer(struct packer *packer);
+
+struct packer_statistics __must_check vdo_get_packer_statistics(const struct packer *packer);
+
+void vdo_attempt_packing(struct data_vio *data_vio);
+
+void vdo_flush_packer(struct packer *packer);
+
+void vdo_remove_lock_holder_from_packer(struct vdo_completion *completion);
+
+void vdo_increment_packer_flush_generation(struct packer *packer);
+
+void vdo_drain_packer(struct packer *packer, struct vdo_completion *completion);
+
+void vdo_resume_packer(struct packer *packer, struct vdo_completion *parent);
+
+void vdo_dump_packer(const struct packer *packer);
+
+#endif /* VDO_PACKER_H */