xref: /spdk/lib/reduce/reduce.c (revision 9d92f20ddfa95ca48373e6ce609e0c7829e9e304)

/*   SPDX-License-Identifier: BSD-3-Clause
 *   Copyright (C) 2018 Intel Corporation.
 *   All rights reserved.
 *   Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 */

#include "spdk/stdinc.h"

#include "spdk/reduce.h"
#include "spdk/env.h"
#include "spdk/string.h"
#include "spdk/bit_array.h"
#include "spdk/util.h"
#include "spdk/log.h"
#include "spdk/memory.h"

#include "libpmem.h"

/* Always round up the size of the PM region to the nearest cacheline. */
#define REDUCE_PM_SIZE_ALIGNMENT	64

/* Offset into the backing device where the persistent memory file's path is stored. */
#define REDUCE_BACKING_DEV_PATH_OFFSET	4096

#define REDUCE_EMPTY_MAP_ENTRY	-1ULL

#define REDUCE_NUM_VOL_REQUESTS	256

/* Structure written to offset 0 of both the pm file and the backing device. */
struct spdk_reduce_vol_superblock {
	uint8_t				signature[8];
	struct spdk_reduce_vol_params	params;
	uint8_t				reserved[4048];
};
SPDK_STATIC_ASSERT(sizeof(struct spdk_reduce_vol_superblock) == 4096, "size incorrect");

#define SPDK_REDUCE_SIGNATURE "SPDKREDU"
/* null terminator counts one */
SPDK_STATIC_ASSERT(sizeof(SPDK_REDUCE_SIGNATURE) - 1 ==
		   SPDK_SIZEOF_MEMBER(struct spdk_reduce_vol_superblock, signature), "size incorrect");

#define REDUCE_PATH_MAX 4096

#define REDUCE_ZERO_BUF_SIZE 0x100000

/**
 * Describes a persistent memory file used to hold metadata associated with a
 *  compressed volume.
 */
struct spdk_reduce_pm_file {
	char			path[REDUCE_PATH_MAX];
	void			*pm_buf;
	int			pm_is_pmem;
	uint64_t		size;
};

#define REDUCE_IO_READV		1
#define REDUCE_IO_WRITEV	2

struct spdk_reduce_chunk_map {
	uint32_t		compressed_size;
	uint32_t		reserved;
	uint64_t		io_unit_index[0];
};

struct spdk_reduce_vol_request {
	/**
	 *  Scratch buffer used for uncompressed chunk.  This is used for:
	 *   1) source buffer for compression operations
	 *   2) destination buffer for decompression operations
	 *   3) data buffer when writing uncompressed chunk to disk
	 *   4) data buffer when reading uncompressed chunk from disk
	 */
	uint8_t					*decomp_buf;
	struct iovec				*decomp_buf_iov;

	/**
	 * These are used to construct the iovecs that are sent to
	 *  the decomp engine, they point to a mix of the scratch buffer
	 *  and user buffer
	 */
	struct iovec				decomp_iov[REDUCE_MAX_IOVECS + 2];
	int					decomp_iovcnt;

	/**
	 *  Scratch buffer used for compressed chunk.  This is used for:
	 *   1) destination buffer for compression operations
	 *   2) source buffer for decompression operations
	 *   3) data buffer when writing compressed chunk to disk
	 *   4) data buffer when reading compressed chunk from disk
	 */
	uint8_t					*comp_buf;
	struct iovec				*comp_buf_iov;
	struct iovec				*iov;
	bool					rmw;
	struct spdk_reduce_vol			*vol;
	int					type;
	int					reduce_errno;
	int					iovcnt;
	int					num_backing_ops;
	uint32_t				num_io_units;
	bool					chunk_is_compressed;
	bool					copy_after_decompress;
	uint64_t				offset;
	uint64_t				logical_map_index;
	uint64_t				length;
	uint64_t				chunk_map_index;
	struct spdk_reduce_chunk_map		*chunk;
	spdk_reduce_vol_op_complete		cb_fn;
	void					*cb_arg;
	TAILQ_ENTRY(spdk_reduce_vol_request)	tailq;
	struct spdk_reduce_vol_cb_args		backing_cb_args;
};

struct spdk_reduce_vol {
	struct spdk_reduce_vol_params		params;
	uint32_t				backing_io_units_per_chunk;
	uint32_t				backing_lba_per_io_unit;
	uint32_t				logical_blocks_per_chunk;
	struct spdk_reduce_pm_file		pm_file;
	struct spdk_reduce_backing_dev		*backing_dev;
	struct spdk_reduce_vol_superblock	*backing_super;
	struct spdk_reduce_vol_superblock	*pm_super;
	uint64_t				*pm_logical_map;
	uint64_t				*pm_chunk_maps;

	struct spdk_bit_array			*allocated_chunk_maps;
	struct spdk_bit_array			*allocated_backing_io_units;

	struct spdk_reduce_vol_request		*request_mem;
	TAILQ_HEAD(, spdk_reduce_vol_request)	free_requests;
	TAILQ_HEAD(, spdk_reduce_vol_request)	executing_requests;
	TAILQ_HEAD(, spdk_reduce_vol_request)	queued_requests;

	/* Single contiguous buffer used for all request buffers for this volume. */
	uint8_t					*buf_mem;
	struct iovec				*buf_iov_mem;
};

static void _start_readv_request(struct spdk_reduce_vol_request *req);
static void _start_writev_request(struct spdk_reduce_vol_request *req);
static uint8_t *g_zero_buf;
static int g_vol_count = 0;

/*
 * Allocate extra metadata chunks and corresponding backing io units to account for
 *  outstanding IO in worst case scenario where logical map is completely allocated
 *  and no data can be compressed.  We need extra chunks in this case to handle
 *  in-flight writes since reduce never writes data in place.
 */
#define REDUCE_NUM_EXTRA_CHUNKS 128

static void
_reduce_persist(struct spdk_reduce_vol *vol, const void *addr, size_t len)
{
	if (vol->pm_file.pm_is_pmem) {
		pmem_persist(addr, len);
	} else {
		pmem_msync(addr, len);
	}
}

static uint64_t
_get_pm_logical_map_size(uint64_t vol_size, uint64_t chunk_size)
{
	uint64_t chunks_in_logical_map, logical_map_size;

	chunks_in_logical_map = vol_size / chunk_size;
	logical_map_size = chunks_in_logical_map * sizeof(uint64_t);

	/* Round up to next cacheline. */
	return spdk_divide_round_up(logical_map_size, REDUCE_PM_SIZE_ALIGNMENT) *
	       REDUCE_PM_SIZE_ALIGNMENT;
}

static uint64_t
_get_total_chunks(uint64_t vol_size, uint64_t chunk_size)
{
	uint64_t num_chunks;

	num_chunks = vol_size / chunk_size;
	num_chunks += REDUCE_NUM_EXTRA_CHUNKS;

	return num_chunks;
}

static inline uint32_t
_reduce_vol_get_chunk_struct_size(uint64_t backing_io_units_per_chunk)
{
	return sizeof(struct spdk_reduce_chunk_map) + sizeof(uint64_t) * backing_io_units_per_chunk;
}

static uint64_t
_get_pm_total_chunks_size(uint64_t vol_size, uint64_t chunk_size, uint64_t backing_io_unit_size)
{
	uint64_t io_units_per_chunk, num_chunks, total_chunks_size;

	num_chunks = _get_total_chunks(vol_size, chunk_size);
	io_units_per_chunk = chunk_size / backing_io_unit_size;

	total_chunks_size = num_chunks * _reduce_vol_get_chunk_struct_size(io_units_per_chunk);

	return spdk_divide_round_up(total_chunks_size, REDUCE_PM_SIZE_ALIGNMENT) *
	       REDUCE_PM_SIZE_ALIGNMENT;
}

static struct spdk_reduce_chunk_map *
_reduce_vol_get_chunk_map(struct spdk_reduce_vol *vol, uint64_t chunk_map_index)
{
	uintptr_t chunk_map_addr;

	assert(chunk_map_index < _get_total_chunks(vol->params.vol_size, vol->params.chunk_size));

	chunk_map_addr = (uintptr_t)vol->pm_chunk_maps;
	chunk_map_addr += chunk_map_index *
			  _reduce_vol_get_chunk_struct_size(vol->backing_io_units_per_chunk);

	return (struct spdk_reduce_chunk_map *)chunk_map_addr;
}

static int
_validate_vol_params(struct spdk_reduce_vol_params *params)
{
	if (params->vol_size > 0) {
		/**
		 * User does not pass in the vol size - it gets calculated by libreduce from
		 *  values in this structure plus the size of the backing device.
		 */
		return -EINVAL;
	}

	if (params->chunk_size == 0 || params->backing_io_unit_size == 0 ||
	    params->logical_block_size == 0) {
		return -EINVAL;
	}

	/* Chunk size must be an even multiple of the backing io unit size. */
	if ((params->chunk_size % params->backing_io_unit_size) != 0) {
		return -EINVAL;
	}

	/* Chunk size must be an even multiple of the logical block size. */
	if ((params->chunk_size % params->logical_block_size) != 0) {
		return -1;
	}

	return 0;
}

static uint64_t
_get_vol_size(uint64_t chunk_size, uint64_t backing_dev_size)
{
	uint64_t num_chunks;

	num_chunks = backing_dev_size / chunk_size;
	if (num_chunks <= REDUCE_NUM_EXTRA_CHUNKS) {
		return 0;
	}

	num_chunks -= REDUCE_NUM_EXTRA_CHUNKS;
	return num_chunks * chunk_size;
}

static uint64_t
_get_pm_file_size(struct spdk_reduce_vol_params *params)
{
	uint64_t total_pm_size;

	total_pm_size = sizeof(struct spdk_reduce_vol_superblock);
	total_pm_size += _get_pm_logical_map_size(params->vol_size, params->chunk_size);
	total_pm_size += _get_pm_total_chunks_size(params->vol_size, params->chunk_size,
			 params->backing_io_unit_size);
	return total_pm_size;
}

const struct spdk_uuid *
spdk_reduce_vol_get_uuid(struct spdk_reduce_vol *vol)
{
	return &vol->params.uuid;
}

static void
_initialize_vol_pm_pointers(struct spdk_reduce_vol *vol)
{
	uint64_t logical_map_size;

	/* Superblock is at the beginning of the pm file. */
	vol->pm_super = (struct spdk_reduce_vol_superblock *)vol->pm_file.pm_buf;

	/* Logical map immediately follows the super block. */
	vol->pm_logical_map = (uint64_t *)(vol->pm_super + 1);

	/* Chunks maps follow the logical map. */
	logical_map_size = _get_pm_logical_map_size(vol->params.vol_size, vol->params.chunk_size);
	vol->pm_chunk_maps = (uint64_t *)((uint8_t *)vol->pm_logical_map + logical_map_size);
}

/* We need 2 iovs during load - one for the superblock, another for the path */
#define LOAD_IOV_COUNT	2

struct reduce_init_load_ctx {
	struct spdk_reduce_vol			*vol;
	struct spdk_reduce_vol_cb_args		backing_cb_args;
	spdk_reduce_vol_op_with_handle_complete	cb_fn;
	void					*cb_arg;
	struct iovec				iov[LOAD_IOV_COUNT];
	void					*path;
};

static inline bool
_addr_crosses_huge_page(const void *addr, size_t *size)
{
	size_t _size;
	uint64_t rc;

	assert(size);

	_size = *size;
	rc = spdk_vtophys(addr, size);

	return rc == SPDK_VTOPHYS_ERROR || _size != *size;
}

static inline int
_set_buffer(uint8_t **vol_buffer, uint8_t **_addr, uint8_t *addr_range, size_t buffer_size)
{
	uint8_t *addr;
	size_t size_tmp = buffer_size;

	addr = *_addr;

	/* Verify that addr + buffer_size doesn't cross huge page boundary */
	if (_addr_crosses_huge_page(addr, &size_tmp)) {
		/* Memory start is aligned on 2MiB, so buffer should be located at the end of the page.
		 * Skip remaining bytes and continue from the beginning of the next page */
		addr += size_tmp;
	}

	if (addr + buffer_size > addr_range) {
		SPDK_ERRLOG("Vol buffer %p out of range %p\n", addr, addr_range);
		return -ERANGE;
	}

	*vol_buffer = addr;
	*_addr = addr + buffer_size;

	return 0;
}

static int
_allocate_vol_requests(struct spdk_reduce_vol *vol)
{
	struct spdk_reduce_vol_request *req;
	uint32_t reqs_in_2mb_page, huge_pages_needed;
	uint8_t *buffer, *buffer_end;
	int i = 0;
	int rc = 0;

	/* It is needed to allocate comp and decomp buffers so that they do not cross physical
	* page boundaries. Assume that the system uses default 2MiB pages and chunk_size is not
	* necessarily power of 2
	* Allocate 2x since we need buffers for both read/write and compress/decompress
	* intermediate buffers. */
	reqs_in_2mb_page = VALUE_2MB / (vol->params.chunk_size * 2);
	if (!reqs_in_2mb_page) {
		return -EINVAL;
	}
	huge_pages_needed = SPDK_CEIL_DIV(REDUCE_NUM_VOL_REQUESTS, reqs_in_2mb_page);

	vol->buf_mem = spdk_dma_malloc(VALUE_2MB * huge_pages_needed, VALUE_2MB, NULL);
	if (vol->buf_mem == NULL) {
		return -ENOMEM;
	}

	vol->request_mem = calloc(REDUCE_NUM_VOL_REQUESTS, sizeof(*req));
	if (vol->request_mem == NULL) {
		spdk_free(vol->buf_mem);
		vol->buf_mem = NULL;
		return -ENOMEM;
	}

	/* Allocate 2x since we need iovs for both read/write and compress/decompress intermediate
	 *  buffers.
	 */
	vol->buf_iov_mem = calloc(REDUCE_NUM_VOL_REQUESTS,
				  2 * sizeof(struct iovec) * vol->backing_io_units_per_chunk);
	if (vol->buf_iov_mem == NULL) {
		free(vol->request_mem);
		spdk_free(vol->buf_mem);
		vol->request_mem = NULL;
		vol->buf_mem = NULL;
		return -ENOMEM;
	}

	buffer = vol->buf_mem;
	buffer_end = buffer + VALUE_2MB * huge_pages_needed;

	for (i = 0; i < REDUCE_NUM_VOL_REQUESTS; i++) {
		req = &vol->request_mem[i];
		TAILQ_INSERT_HEAD(&vol->free_requests, req, tailq);
		req->decomp_buf_iov = &vol->buf_iov_mem[(2 * i) * vol->backing_io_units_per_chunk];
		req->comp_buf_iov = &vol->buf_iov_mem[(2 * i + 1) * vol->backing_io_units_per_chunk];

		rc = _set_buffer(&req->comp_buf, &buffer, buffer_end, vol->params.chunk_size);
		if (rc) {
			SPDK_ERRLOG("Failed to set comp buffer for req idx %u, addr %p, start %p, end %p\n", i, buffer,
				    vol->buf_mem, buffer_end);
			break;
		}
		rc = _set_buffer(&req->decomp_buf, &buffer, buffer_end, vol->params.chunk_size);
		if (rc) {
			SPDK_ERRLOG("Failed to set decomp buffer for req idx %u, addr %p, start %p, end %p\n", i, buffer,
				    vol->buf_mem, buffer_end);
			break;
		}
	}

	if (rc) {
		free(vol->buf_iov_mem);
		free(vol->request_mem);
		spdk_free(vol->buf_mem);
		vol->buf_mem = NULL;
		vol->buf_iov_mem = NULL;
		vol->request_mem = NULL;
	}

	return rc;
}

static void
_init_load_cleanup(struct spdk_reduce_vol *vol, struct reduce_init_load_ctx *ctx)
{
	if (ctx != NULL) {
		spdk_free(ctx->path);
		free(ctx);
	}

	if (vol != NULL) {
		if (vol->pm_file.pm_buf != NULL) {
			pmem_unmap(vol->pm_file.pm_buf, vol->pm_file.size);
		}

		spdk_free(vol->backing_super);
		spdk_bit_array_free(&vol->allocated_chunk_maps);
		spdk_bit_array_free(&vol->allocated_backing_io_units);
		free(vol->request_mem);
		free(vol->buf_iov_mem);
		spdk_free(vol->buf_mem);
		free(vol);
	}
}

static int
_alloc_zero_buff(void)
{
	int rc = 0;

	/* The zero buffer is shared between all volumes and just used
	 * for reads so allocate one global instance here if not already
	 * allocated when another vol init'd or loaded.
	 */
	if (g_vol_count++ == 0) {
		g_zero_buf = spdk_zmalloc(REDUCE_ZERO_BUF_SIZE,
					  64, NULL, SPDK_ENV_LCORE_ID_ANY,
					  SPDK_MALLOC_DMA);
		if (g_zero_buf == NULL) {
			g_vol_count--;
			rc = -ENOMEM;
		}
	}
	return rc;
}

static void
_init_write_super_cpl(void *cb_arg, int reduce_errno)
{
	struct reduce_init_load_ctx *init_ctx = cb_arg;
	int rc;

	rc = _allocate_vol_requests(init_ctx->vol);
	if (rc != 0) {
		init_ctx->cb_fn(init_ctx->cb_arg, NULL, rc);
		_init_load_cleanup(init_ctx->vol, init_ctx);
		return;
	}

	rc = _alloc_zero_buff();
	if (rc != 0) {
		init_ctx->cb_fn(init_ctx->cb_arg, NULL, rc);
		_init_load_cleanup(init_ctx->vol, init_ctx);
		return;
	}

	init_ctx->cb_fn(init_ctx->cb_arg, init_ctx->vol, reduce_errno);
	/* Only clean up the ctx - the vol has been passed to the application
	 *  for use now that initialization was successful.
	 */
	_init_load_cleanup(NULL, init_ctx);
}

static void
_init_write_path_cpl(void *cb_arg, int reduce_errno)
{
	struct reduce_init_load_ctx *init_ctx = cb_arg;
	struct spdk_reduce_vol *vol = init_ctx->vol;

	init_ctx->iov[0].iov_base = vol->backing_super;
	init_ctx->iov[0].iov_len = sizeof(*vol->backing_super);
	init_ctx->backing_cb_args.cb_fn = _init_write_super_cpl;
	init_ctx->backing_cb_args.cb_arg = init_ctx;
	vol->backing_dev->writev(vol->backing_dev, init_ctx->iov, 1,
				 0, sizeof(*vol->backing_super) / vol->backing_dev->blocklen,
				 &init_ctx->backing_cb_args);
}

static int
_allocate_bit_arrays(struct spdk_reduce_vol *vol)
{
	uint64_t total_chunks, total_backing_io_units;
	uint32_t i, num_metadata_io_units;

	total_chunks = _get_total_chunks(vol->params.vol_size, vol->params.chunk_size);
	vol->allocated_chunk_maps = spdk_bit_array_create(total_chunks);
	total_backing_io_units = total_chunks * (vol->params.chunk_size / vol->params.backing_io_unit_size);
	vol->allocated_backing_io_units = spdk_bit_array_create(total_backing_io_units);

	if (vol->allocated_chunk_maps == NULL || vol->allocated_backing_io_units == NULL) {
		return -ENOMEM;
	}

	/* Set backing io unit bits associated with metadata. */
	num_metadata_io_units = (sizeof(*vol->backing_super) + REDUCE_PATH_MAX) /
				vol->params.backing_io_unit_size;
	for (i = 0; i < num_metadata_io_units; i++) {
		spdk_bit_array_set(vol->allocated_backing_io_units, i);
	}

	return 0;
}

void
spdk_reduce_vol_init(struct spdk_reduce_vol_params *params,
		     struct spdk_reduce_backing_dev *backing_dev,
		     const char *pm_file_dir,
		     spdk_reduce_vol_op_with_handle_complete cb_fn, void *cb_arg)
{
	struct spdk_reduce_vol *vol;
	struct reduce_init_load_ctx *init_ctx;
	uint64_t backing_dev_size;
	size_t mapped_len;
	int dir_len, max_dir_len, rc;

	/* We need to append a path separator and the UUID to the supplied
	 * path.
	 */
	max_dir_len = REDUCE_PATH_MAX - SPDK_UUID_STRING_LEN - 1;
	dir_len = strnlen(pm_file_dir, max_dir_len);
	/* Strip trailing slash if the user provided one - we will add it back
	 * later when appending the filename.
	 */
	if (pm_file_dir[dir_len - 1] == '/') {
		dir_len--;
	}
	if (dir_len == max_dir_len) {
		SPDK_ERRLOG("pm_file_dir (%s) too long\n", pm_file_dir);
		cb_fn(cb_arg, NULL, -EINVAL);
		return;
	}

	rc = _validate_vol_params(params);
	if (rc != 0) {
		SPDK_ERRLOG("invalid vol params\n");
		cb_fn(cb_arg, NULL, rc);
		return;
	}

	backing_dev_size = backing_dev->blockcnt * backing_dev->blocklen;
	params->vol_size = _get_vol_size(params->chunk_size, backing_dev_size);
	if (params->vol_size == 0) {
		SPDK_ERRLOG("backing device is too small\n");
		cb_fn(cb_arg, NULL, -EINVAL);
		return;
	}

	if (backing_dev->readv == NULL || backing_dev->writev == NULL ||
	    backing_dev->unmap == NULL) {
		SPDK_ERRLOG("backing_dev function pointer not specified\n");
		cb_fn(cb_arg, NULL, -EINVAL);
		return;
	}

	vol = calloc(1, sizeof(*vol));
	if (vol == NULL) {
		cb_fn(cb_arg, NULL, -ENOMEM);
		return;
	}

	TAILQ_INIT(&vol->free_requests);
	TAILQ_INIT(&vol->executing_requests);
	TAILQ_INIT(&vol->queued_requests);

	vol->backing_super = spdk_zmalloc(sizeof(*vol->backing_super), 0, NULL,
					  SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
	if (vol->backing_super == NULL) {
		cb_fn(cb_arg, NULL, -ENOMEM);
		_init_load_cleanup(vol, NULL);
		return;
	}

	init_ctx = calloc(1, sizeof(*init_ctx));
	if (init_ctx == NULL) {
		cb_fn(cb_arg, NULL, -ENOMEM);
		_init_load_cleanup(vol, NULL);
		return;
	}

	init_ctx->path = spdk_zmalloc(REDUCE_PATH_MAX, 0, NULL,
				      SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
	if (init_ctx->path == NULL) {
		cb_fn(cb_arg, NULL, -ENOMEM);
		_init_load_cleanup(vol, init_ctx);
		return;
	}

	if (spdk_uuid_is_null(&params->uuid)) {
		spdk_uuid_generate(&params->uuid);
	}

	memcpy(vol->pm_file.path, pm_file_dir, dir_len);
	vol->pm_file.path[dir_len] = '/';
	spdk_uuid_fmt_lower(&vol->pm_file.path[dir_len + 1], SPDK_UUID_STRING_LEN,
			    &params->uuid);
	vol->pm_file.size = _get_pm_file_size(params);
	vol->pm_file.pm_buf = pmem_map_file(vol->pm_file.path, vol->pm_file.size,
					    PMEM_FILE_CREATE | PMEM_FILE_EXCL, 0600,
					    &mapped_len, &vol->pm_file.pm_is_pmem);
	if (vol->pm_file.pm_buf == NULL) {
		SPDK_ERRLOG("could not pmem_map_file(%s): %s\n",
			    vol->pm_file.path, strerror(errno));
		cb_fn(cb_arg, NULL, -errno);
		_init_load_cleanup(vol, init_ctx);
		return;
	}

	if (vol->pm_file.size != mapped_len) {
		SPDK_ERRLOG("could not map entire pmem file (size=%" PRIu64 " mapped=%" PRIu64 ")\n",
			    vol->pm_file.size, mapped_len);
		cb_fn(cb_arg, NULL, -ENOMEM);
		_init_load_cleanup(vol, init_ctx);
		return;
	}

	vol->backing_io_units_per_chunk = params->chunk_size / params->backing_io_unit_size;
	vol->logical_blocks_per_chunk = params->chunk_size / params->logical_block_size;
	vol->backing_lba_per_io_unit = params->backing_io_unit_size / backing_dev->blocklen;
	memcpy(&vol->params, params, sizeof(*params));

	vol->backing_dev = backing_dev;

	rc = _allocate_bit_arrays(vol);
	if (rc != 0) {
		cb_fn(cb_arg, NULL, rc);
		_init_load_cleanup(vol, init_ctx);
		return;
	}

	memcpy(vol->backing_super->signature, SPDK_REDUCE_SIGNATURE,
	       sizeof(vol->backing_super->signature));
	memcpy(&vol->backing_super->params, params, sizeof(*params));

	_initialize_vol_pm_pointers(vol);

	memcpy(vol->pm_super, vol->backing_super, sizeof(*vol->backing_super));
	/* Writing 0xFF's is equivalent of filling it all with SPDK_EMPTY_MAP_ENTRY.
	 * Note that this writes 0xFF to not just the logical map but the chunk maps as well.
	 */
	memset(vol->pm_logical_map, 0xFF, vol->pm_file.size - sizeof(*vol->backing_super));
	_reduce_persist(vol, vol->pm_file.pm_buf, vol->pm_file.size);

	init_ctx->vol = vol;
	init_ctx->cb_fn = cb_fn;
	init_ctx->cb_arg = cb_arg;

	memcpy(init_ctx->path, vol->pm_file.path, REDUCE_PATH_MAX);
	init_ctx->iov[0].iov_base = init_ctx->path;
	init_ctx->iov[0].iov_len = REDUCE_PATH_MAX;
	init_ctx->backing_cb_args.cb_fn = _init_write_path_cpl;
	init_ctx->backing_cb_args.cb_arg = init_ctx;
	/* Write path to offset 4K on backing device - just after where the super
	 *  block will be written.  We wait until this is committed before writing the
	 *  super block to guarantee we don't get the super block written without the
	 *  the path if the system crashed in the middle of a write operation.
	 */
	vol->backing_dev->writev(vol->backing_dev, init_ctx->iov, 1,
				 REDUCE_BACKING_DEV_PATH_OFFSET / vol->backing_dev->blocklen,
				 REDUCE_PATH_MAX / vol->backing_dev->blocklen,
				 &init_ctx->backing_cb_args);
}

static void destroy_load_cb(void *cb_arg, struct spdk_reduce_vol *vol, int reduce_errno);

static void
_load_read_super_and_path_cpl(void *cb_arg, int reduce_errno)
{
	struct reduce_init_load_ctx *load_ctx = cb_arg;
	struct spdk_reduce_vol *vol = load_ctx->vol;
	uint64_t backing_dev_size;
	uint64_t i, num_chunks, logical_map_index;
	struct spdk_reduce_chunk_map *chunk;
	size_t mapped_len;
	uint32_t j;
	int rc;

	rc = _alloc_zero_buff();
	if (rc) {
		goto error;
	}

	if (memcmp(vol->backing_super->signature,
		   SPDK_REDUCE_SIGNATURE,
		   sizeof(vol->backing_super->signature)) != 0) {
		/* This backing device isn't a libreduce backing device. */
		rc = -EILSEQ;
		goto error;
	}

	/* If the cb_fn is destroy_load_cb, it means we are wanting to destroy this compress bdev.
	 *  So don't bother getting the volume ready to use - invoke the callback immediately
	 *  so destroy_load_cb can delete the metadata off of the block device and delete the
	 *  persistent memory file if it exists.
	 */
	memcpy(vol->pm_file.path, load_ctx->path, sizeof(vol->pm_file.path));
	if (load_ctx->cb_fn == (*destroy_load_cb)) {
		load_ctx->cb_fn(load_ctx->cb_arg, vol, 0);
		_init_load_cleanup(NULL, load_ctx);
		return;
	}

	memcpy(&vol->params, &vol->backing_super->params, sizeof(vol->params));
	vol->backing_io_units_per_chunk = vol->params.chunk_size / vol->params.backing_io_unit_size;
	vol->logical_blocks_per_chunk = vol->params.chunk_size / vol->params.logical_block_size;
	vol->backing_lba_per_io_unit = vol->params.backing_io_unit_size / vol->backing_dev->blocklen;

	rc = _allocate_bit_arrays(vol);
	if (rc != 0) {
		goto error;
	}

	backing_dev_size = vol->backing_dev->blockcnt * vol->backing_dev->blocklen;
	if (_get_vol_size(vol->params.chunk_size, backing_dev_size) < vol->params.vol_size) {
		SPDK_ERRLOG("backing device size %" PRIi64 " smaller than expected\n",
			    backing_dev_size);
		rc = -EILSEQ;
		goto error;
	}

	vol->pm_file.size = _get_pm_file_size(&vol->params);
	vol->pm_file.pm_buf = pmem_map_file(vol->pm_file.path, 0, 0, 0, &mapped_len,
					    &vol->pm_file.pm_is_pmem);
	if (vol->pm_file.pm_buf == NULL) {
		SPDK_ERRLOG("could not pmem_map_file(%s): %s\n", vol->pm_file.path, strerror(errno));
		rc = -errno;
		goto error;
	}

	if (vol->pm_file.size != mapped_len) {
		SPDK_ERRLOG("could not map entire pmem file (size=%" PRIu64 " mapped=%" PRIu64 ")\n",
			    vol->pm_file.size, mapped_len);
		rc = -ENOMEM;
		goto error;
	}

	rc = _allocate_vol_requests(vol);
	if (rc != 0) {
		goto error;
	}

	_initialize_vol_pm_pointers(vol);

	num_chunks = vol->params.vol_size / vol->params.chunk_size;
	for (i = 0; i < num_chunks; i++) {
		logical_map_index = vol->pm_logical_map[i];
		if (logical_map_index == REDUCE_EMPTY_MAP_ENTRY) {
			continue;
		}
		spdk_bit_array_set(vol->allocated_chunk_maps, logical_map_index);
		chunk = _reduce_vol_get_chunk_map(vol, logical_map_index);
		for (j = 0; j < vol->backing_io_units_per_chunk; j++) {
			if (chunk->io_unit_index[j] != REDUCE_EMPTY_MAP_ENTRY) {
				spdk_bit_array_set(vol->allocated_backing_io_units, chunk->io_unit_index[j]);
			}
		}
	}

	load_ctx->cb_fn(load_ctx->cb_arg, vol, 0);
	/* Only clean up the ctx - the vol has been passed to the application
	 *  for use now that volume load was successful.
	 */
	_init_load_cleanup(NULL, load_ctx);
	return;

error:
	load_ctx->cb_fn(load_ctx->cb_arg, NULL, rc);
	_init_load_cleanup(vol, load_ctx);
}

void
spdk_reduce_vol_load(struct spdk_reduce_backing_dev *backing_dev,
		     spdk_reduce_vol_op_with_handle_complete cb_fn, void *cb_arg)
{
	struct spdk_reduce_vol *vol;
	struct reduce_init_load_ctx *load_ctx;

	if (backing_dev->readv == NULL || backing_dev->writev == NULL ||
	    backing_dev->unmap == NULL) {
		SPDK_ERRLOG("backing_dev function pointer not specified\n");
		cb_fn(cb_arg, NULL, -EINVAL);
		return;
	}

	vol = calloc(1, sizeof(*vol));
	if (vol == NULL) {
		cb_fn(cb_arg, NULL, -ENOMEM);
		return;
	}

	TAILQ_INIT(&vol->free_requests);
	TAILQ_INIT(&vol->executing_requests);
	TAILQ_INIT(&vol->queued_requests);

	vol->backing_super = spdk_zmalloc(sizeof(*vol->backing_super), 64, NULL,
					  SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
	if (vol->backing_super == NULL) {
		_init_load_cleanup(vol, NULL);
		cb_fn(cb_arg, NULL, -ENOMEM);
		return;
	}

	vol->backing_dev = backing_dev;

	load_ctx = calloc(1, sizeof(*load_ctx));
	if (load_ctx == NULL) {
		_init_load_cleanup(vol, NULL);
		cb_fn(cb_arg, NULL, -ENOMEM);
		return;
	}

	load_ctx->path = spdk_zmalloc(REDUCE_PATH_MAX, 64, NULL,
				      SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
	if (load_ctx->path == NULL) {
		_init_load_cleanup(vol, load_ctx);
		cb_fn(cb_arg, NULL, -ENOMEM);
		return;
	}

	load_ctx->vol = vol;
	load_ctx->cb_fn = cb_fn;
	load_ctx->cb_arg = cb_arg;

	load_ctx->iov[0].iov_base = vol->backing_super;
	load_ctx->iov[0].iov_len = sizeof(*vol->backing_super);
	load_ctx->iov[1].iov_base = load_ctx->path;
	load_ctx->iov[1].iov_len = REDUCE_PATH_MAX;
	load_ctx->backing_cb_args.cb_fn = _load_read_super_and_path_cpl;
	load_ctx->backing_cb_args.cb_arg = load_ctx;
	vol->backing_dev->readv(vol->backing_dev, load_ctx->iov, LOAD_IOV_COUNT, 0,
				(sizeof(*vol->backing_super) + REDUCE_PATH_MAX) /
				vol->backing_dev->blocklen,
				&load_ctx->backing_cb_args);
}

void
spdk_reduce_vol_unload(struct spdk_reduce_vol *vol,
		       spdk_reduce_vol_op_complete cb_fn, void *cb_arg)
{
	if (vol == NULL) {
		/* This indicates a programming error. */
		assert(false);
		cb_fn(cb_arg, -EINVAL);
		return;
	}

	if (--g_vol_count == 0) {
		spdk_free(g_zero_buf);
	}
	assert(g_vol_count >= 0);
	_init_load_cleanup(vol, NULL);
	cb_fn(cb_arg, 0);
}

struct reduce_destroy_ctx {
	spdk_reduce_vol_op_complete		cb_fn;
	void					*cb_arg;
	struct spdk_reduce_vol			*vol;
	struct spdk_reduce_vol_superblock	*super;
	struct iovec				iov;
	struct spdk_reduce_vol_cb_args		backing_cb_args;
	int					reduce_errno;
	char					pm_path[REDUCE_PATH_MAX];
};

static void
destroy_unload_cpl(void *cb_arg, int reduce_errno)
{
	struct reduce_destroy_ctx *destroy_ctx = cb_arg;

	if (destroy_ctx->reduce_errno == 0) {
		if (unlink(destroy_ctx->pm_path)) {
			SPDK_ERRLOG("%s could not be unlinked: %s\n",
				    destroy_ctx->pm_path, strerror(errno));
		}
	}

	/* Even if the unload somehow failed, we still pass the destroy_ctx
	 * reduce_errno since that indicates whether or not the volume was
	 * actually destroyed.
	 */
	destroy_ctx->cb_fn(destroy_ctx->cb_arg, destroy_ctx->reduce_errno);
	spdk_free(destroy_ctx->super);
	free(destroy_ctx);
}

static void
_destroy_zero_super_cpl(void *cb_arg, int reduce_errno)
{
	struct reduce_destroy_ctx *destroy_ctx = cb_arg;
	struct spdk_reduce_vol *vol = destroy_ctx->vol;

	destroy_ctx->reduce_errno = reduce_errno;
	spdk_reduce_vol_unload(vol, destroy_unload_cpl, destroy_ctx);
}

static void
destroy_load_cb(void *cb_arg, struct spdk_reduce_vol *vol, int reduce_errno)
{
	struct reduce_destroy_ctx *destroy_ctx = cb_arg;

	if (reduce_errno != 0) {
		destroy_ctx->cb_fn(destroy_ctx->cb_arg, reduce_errno);
		spdk_free(destroy_ctx->super);
		free(destroy_ctx);
		return;
	}

	destroy_ctx->vol = vol;
	memcpy(destroy_ctx->pm_path, vol->pm_file.path, sizeof(destroy_ctx->pm_path));
	destroy_ctx->iov.iov_base = destroy_ctx->super;
	destroy_ctx->iov.iov_len = sizeof(*destroy_ctx->super);
	destroy_ctx->backing_cb_args.cb_fn = _destroy_zero_super_cpl;
	destroy_ctx->backing_cb_args.cb_arg = destroy_ctx;
	vol->backing_dev->writev(vol->backing_dev, &destroy_ctx->iov, 1, 0,
				 sizeof(*destroy_ctx->super) / vol->backing_dev->blocklen,
				 &destroy_ctx->backing_cb_args);
}

void
spdk_reduce_vol_destroy(struct spdk_reduce_backing_dev *backing_dev,
			spdk_reduce_vol_op_complete cb_fn, void *cb_arg)
{
	struct reduce_destroy_ctx *destroy_ctx;

	destroy_ctx = calloc(1, sizeof(*destroy_ctx));
	if (destroy_ctx == NULL) {
		cb_fn(cb_arg, -ENOMEM);
		return;
	}

	destroy_ctx->super = spdk_zmalloc(sizeof(*destroy_ctx->super), 64, NULL,
					  SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
	if (destroy_ctx->super == NULL) {
		free(destroy_ctx);
		cb_fn(cb_arg, -ENOMEM);
		return;
	}
	destroy_ctx->cb_fn = cb_fn;
	destroy_ctx->cb_arg = cb_arg;
	spdk_reduce_vol_load(backing_dev, destroy_load_cb, destroy_ctx);
}

static bool
_request_spans_chunk_boundary(struct spdk_reduce_vol *vol, uint64_t offset, uint64_t length)
{
	uint64_t start_chunk, end_chunk;

	start_chunk = offset / vol->logical_blocks_per_chunk;
	end_chunk = (offset + length - 1) / vol->logical_blocks_per_chunk;

	return (start_chunk != end_chunk);
}

typedef void (*reduce_request_fn)(void *_req, int reduce_errno);

static void
_reduce_vol_complete_req(struct spdk_reduce_vol_request *req, int reduce_errno)
{
	struct spdk_reduce_vol_request *next_req;
	struct spdk_reduce_vol *vol = req->vol;

	req->cb_fn(req->cb_arg, reduce_errno);
	TAILQ_REMOVE(&vol->executing_requests, req, tailq);

	TAILQ_FOREACH(next_req, &vol->queued_requests, tailq) {
		if (next_req->logical_map_index == req->logical_map_index) {
			TAILQ_REMOVE(&vol->queued_requests, next_req, tailq);
			if (next_req->type == REDUCE_IO_READV) {
				_start_readv_request(next_req);
			} else {
				assert(next_req->type == REDUCE_IO_WRITEV);
				_start_writev_request(next_req);
			}
			break;
		}
	}

	TAILQ_INSERT_HEAD(&vol->free_requests, req, tailq);
}

static void
_reduce_vol_reset_chunk(struct spdk_reduce_vol *vol, uint64_t chunk_map_index)
{
	struct spdk_reduce_chunk_map *chunk;
	uint32_t i;

	chunk = _reduce_vol_get_chunk_map(vol, chunk_map_index);
	for (i = 0; i < vol->backing_io_units_per_chunk; i++) {
		if (chunk->io_unit_index[i] == REDUCE_EMPTY_MAP_ENTRY) {
			break;
		}
		assert(spdk_bit_array_get(vol->allocated_backing_io_units,
					  chunk->io_unit_index[i]) == true);
		spdk_bit_array_clear(vol->allocated_backing_io_units, chunk->io_unit_index[i]);
		chunk->io_unit_index[i] = REDUCE_EMPTY_MAP_ENTRY;
	}
	spdk_bit_array_clear(vol->allocated_chunk_maps, chunk_map_index);
}

static void
_write_write_done(void *_req, int reduce_errno)
{
	struct spdk_reduce_vol_request *req = _req;
	struct spdk_reduce_vol *vol = req->vol;
	uint64_t old_chunk_map_index;

	if (reduce_errno != 0) {
		req->reduce_errno = reduce_errno;
	}

	assert(req->num_backing_ops > 0);
	if (--req->num_backing_ops > 0) {
		return;
	}

	if (req->reduce_errno != 0) {
		_reduce_vol_complete_req(req, req->reduce_errno);
		return;
	}

	old_chunk_map_index = vol->pm_logical_map[req->logical_map_index];
	if (old_chunk_map_index != REDUCE_EMPTY_MAP_ENTRY) {
		_reduce_vol_reset_chunk(vol, old_chunk_map_index);
	}

	/*
	 * We don't need to persist the clearing of the old chunk map here.  The old chunk map
	 * becomes invalid after we update the logical map, since the old chunk map will no
	 * longer have a reference to it in the logical map.
	 */

	/* Persist the new chunk map.  This must be persisted before we update the logical map. */
	_reduce_persist(vol, req->chunk,
			_reduce_vol_get_chunk_struct_size(vol->backing_io_units_per_chunk));

	vol->pm_logical_map[req->logical_map_index] = req->chunk_map_index;

	_reduce_persist(vol, &vol->pm_logical_map[req->logical_map_index], sizeof(uint64_t));

	_reduce_vol_complete_req(req, 0);
}

struct reduce_merged_io_desc {
	uint64_t io_unit_index;
	uint32_t num_io_units;
};

static void
_issue_backing_ops_without_merge(struct spdk_reduce_vol_request *req, struct spdk_reduce_vol *vol,
				 reduce_request_fn next_fn, bool is_write)
{
	struct iovec *iov;
	uint8_t *buf;
	uint32_t i;

	if (req->chunk_is_compressed) {
		iov = req->comp_buf_iov;
		buf = req->comp_buf;
	} else {
		iov = req->decomp_buf_iov;
		buf = req->decomp_buf;
	}

	req->num_backing_ops = req->num_io_units;
	req->backing_cb_args.cb_fn = next_fn;
	req->backing_cb_args.cb_arg = req;
	for (i = 0; i < req->num_io_units; i++) {
		iov[i].iov_base = buf + i * vol->params.backing_io_unit_size;
		iov[i].iov_len = vol->params.backing_io_unit_size;
		if (is_write) {
			vol->backing_dev->writev(vol->backing_dev, &iov[i], 1,
						 req->chunk->io_unit_index[i] * vol->backing_lba_per_io_unit,
						 vol->backing_lba_per_io_unit, &req->backing_cb_args);
		} else {
			vol->backing_dev->readv(vol->backing_dev, &iov[i], 1,
						req->chunk->io_unit_index[i] * vol->backing_lba_per_io_unit,
						vol->backing_lba_per_io_unit, &req->backing_cb_args);
		}
	}
}

static void
_issue_backing_ops(struct spdk_reduce_vol_request *req, struct spdk_reduce_vol *vol,
		   reduce_request_fn next_fn, bool is_write)
{
	struct iovec *iov;
	struct reduce_merged_io_desc merged_io_desc[4];
	uint8_t *buf;
	bool merge = false;
	uint32_t num_io = 0;
	uint32_t io_unit_counts = 0;
	uint32_t merged_io_idx = 0;
	uint32_t i;

	/* The merged_io_desc value is defined here to contain four elements,
	 * and the chunk size must be four times the maximum of the io unit.
	 * if chunk size is too big, don't merge IO.
	 */
	if (vol->backing_io_units_per_chunk > 4) {
		_issue_backing_ops_without_merge(req, vol, next_fn, is_write);
		return;
	}

	if (req->chunk_is_compressed) {
		iov = req->comp_buf_iov;
		buf = req->comp_buf;
	} else {
		iov = req->decomp_buf_iov;
		buf = req->decomp_buf;
	}

	for (i = 0; i < req->num_io_units; i++) {
		if (!merge) {
			merged_io_desc[merged_io_idx].io_unit_index = req->chunk->io_unit_index[i];
			merged_io_desc[merged_io_idx].num_io_units = 1;
			num_io++;
		}

		if (i + 1 == req->num_io_units) {
			break;
		}

		if (req->chunk->io_unit_index[i] + 1 == req->chunk->io_unit_index[i + 1]) {
			merged_io_desc[merged_io_idx].num_io_units += 1;
			merge = true;
			continue;
		}
		merge = false;
		merged_io_idx++;
	}

	req->num_backing_ops = num_io;
	req->backing_cb_args.cb_fn = next_fn;
	req->backing_cb_args.cb_arg = req;
	for (i = 0; i < num_io; i++) {
		iov[i].iov_base = buf + io_unit_counts * vol->params.backing_io_unit_size;
		iov[i].iov_len = vol->params.backing_io_unit_size * merged_io_desc[i].num_io_units;
		if (is_write) {
			vol->backing_dev->writev(vol->backing_dev, &iov[i], 1,
						 merged_io_desc[i].io_unit_index * vol->backing_lba_per_io_unit,
						 vol->backing_lba_per_io_unit * merged_io_desc[i].num_io_units,
						 &req->backing_cb_args);
		} else {
			vol->backing_dev->readv(vol->backing_dev, &iov[i], 1,
						merged_io_desc[i].io_unit_index * vol->backing_lba_per_io_unit,
						vol->backing_lba_per_io_unit * merged_io_desc[i].num_io_units,
						&req->backing_cb_args);
		}

		/* Collects the number of processed I/O. */
		io_unit_counts += merged_io_desc[i].num_io_units;
	}
}

static void
_reduce_vol_write_chunk(struct spdk_reduce_vol_request *req, reduce_request_fn next_fn,
			uint32_t compressed_size)
{
	struct spdk_reduce_vol *vol = req->vol;
	uint32_t i;
	uint64_t chunk_offset, remainder, total_len = 0;
	uint8_t *buf;
	int j;

	req->chunk_map_index = spdk_bit_array_find_first_clear(vol->allocated_chunk_maps, 0);

	/* TODO: fail if no chunk map found - but really this should not happen if we
	 * size the number of requests similarly to number of extra chunk maps
	 */
	assert(req->chunk_map_index != UINT32_MAX);
	spdk_bit_array_set(vol->allocated_chunk_maps, req->chunk_map_index);

	req->chunk = _reduce_vol_get_chunk_map(vol, req->chunk_map_index);
	req->num_io_units = spdk_divide_round_up(compressed_size,
			    vol->params.backing_io_unit_size);
	req->chunk_is_compressed = (req->num_io_units != vol->backing_io_units_per_chunk);
	req->chunk->compressed_size =
		req->chunk_is_compressed ? compressed_size : vol->params.chunk_size;

	/* if the chunk is uncompressed we need to copy the data from the host buffers. */
	if (req->chunk_is_compressed == false) {
		chunk_offset = req->offset % vol->logical_blocks_per_chunk;
		buf = req->decomp_buf;
		total_len = chunk_offset * vol->params.logical_block_size;

		/* zero any offset into chunk */
		if (req->rmw == false && chunk_offset) {
			memset(buf, 0, total_len);
		}
		buf += total_len;

		/* copy the data */
		for (j = 0; j < req->iovcnt; j++) {
			memcpy(buf, req->iov[j].iov_base, req->iov[j].iov_len);
			buf += req->iov[j].iov_len;
			total_len += req->iov[j].iov_len;
		}

		/* zero any remainder */
		remainder = vol->params.chunk_size - total_len;
		total_len += remainder;
		if (req->rmw == false && remainder) {
			memset(buf, 0, remainder);
		}
		assert(total_len == vol->params.chunk_size);
	}

	for (i = 0; i < req->num_io_units; i++) {
		req->chunk->io_unit_index[i] = spdk_bit_array_find_first_clear(vol->allocated_backing_io_units, 0);
		/* TODO: fail if no backing block found - but really this should also not
		 * happen (see comment above).
		 */
		assert(req->chunk->io_unit_index[i] != UINT32_MAX);
		spdk_bit_array_set(vol->allocated_backing_io_units, req->chunk->io_unit_index[i]);
	}

	_issue_backing_ops(req, vol, next_fn, true /* write */);
}

static void
_write_compress_done(void *_req, int reduce_errno)
{
	struct spdk_reduce_vol_request *req = _req;

	/* Negative reduce_errno indicates failure for compression operations.
	 * Just write the uncompressed data instead.  Force this to happen
	 * by just passing the full chunk size to _reduce_vol_write_chunk.
	 * When it sees the data couldn't be compressed, it will just write
	 * the uncompressed buffer to disk.
	 */
	if (reduce_errno < 0) {
		req->backing_cb_args.output_size = req->vol->params.chunk_size;
	}

	_reduce_vol_write_chunk(req, _write_write_done, req->backing_cb_args.output_size);
}

static void
_reduce_vol_compress_chunk(struct spdk_reduce_vol_request *req, reduce_request_fn next_fn)
{
	struct spdk_reduce_vol *vol = req->vol;

	req->backing_cb_args.cb_fn = next_fn;
	req->backing_cb_args.cb_arg = req;
	req->comp_buf_iov[0].iov_base = req->comp_buf;
	req->comp_buf_iov[0].iov_len = vol->params.chunk_size;
	vol->backing_dev->compress(vol->backing_dev,
				   req->decomp_iov, req->decomp_iovcnt, req->comp_buf_iov, 1,
				   &req->backing_cb_args);
}

static void
_reduce_vol_decompress_chunk_scratch(struct spdk_reduce_vol_request *req, reduce_request_fn next_fn)
{
	struct spdk_reduce_vol *vol = req->vol;

	req->backing_cb_args.cb_fn = next_fn;
	req->backing_cb_args.cb_arg = req;
	req->comp_buf_iov[0].iov_base = req->comp_buf;
	req->comp_buf_iov[0].iov_len = req->chunk->compressed_size;
	req->decomp_buf_iov[0].iov_base = req->decomp_buf;
	req->decomp_buf_iov[0].iov_len = vol->params.chunk_size;
	vol->backing_dev->decompress(vol->backing_dev,
				     req->comp_buf_iov, 1, req->decomp_buf_iov, 1,
				     &req->backing_cb_args);
}

static void
_reduce_vol_decompress_chunk(struct spdk_reduce_vol_request *req, reduce_request_fn next_fn)
{
	struct spdk_reduce_vol *vol = req->vol;
	uint64_t chunk_offset, remainder = 0;
	uint64_t ttl_len = 0;
	size_t iov_len;
	int i;

	req->decomp_iovcnt = 0;
	chunk_offset = req->offset % vol->logical_blocks_per_chunk;

	/* If backing device doesn't support SGL output then we should copy the result of decompression to user's buffer
	 * if at least one of the conditions below is true:
	 * 1. User's buffer is fragmented
	 * 2. Length of the user's buffer is less than the chunk
	 * 3. User's buffer is contig, equals chunk_size but crosses huge page boundary */
	iov_len = req->iov[0].iov_len;
	req->copy_after_decompress = !vol->backing_dev->sgl_out && (req->iovcnt > 1 ||
				     req->iov[0].iov_len < vol->params.chunk_size ||
				     _addr_crosses_huge_page(req->iov[0].iov_base, &iov_len));
	if (req->copy_after_decompress) {
		req->decomp_iov[0].iov_base = req->decomp_buf;
		req->decomp_iov[0].iov_len = vol->params.chunk_size;
		req->decomp_iovcnt = 1;
		goto decompress;
	}

	if (chunk_offset) {
		/* first iov point to our scratch buffer for any offset into the chunk */
		req->decomp_iov[0].iov_base = req->decomp_buf;
		req->decomp_iov[0].iov_len = chunk_offset * vol->params.logical_block_size;
		ttl_len += req->decomp_iov[0].iov_len;
		req->decomp_iovcnt = 1;
	}

	/* now the user data iov, direct to the user buffer */
	for (i = 0; i < req->iovcnt; i++) {
		req->decomp_iov[i + req->decomp_iovcnt].iov_base = req->iov[i].iov_base;
		req->decomp_iov[i + req->decomp_iovcnt].iov_len = req->iov[i].iov_len;
		ttl_len += req->decomp_iov[i + req->decomp_iovcnt].iov_len;
	}
	req->decomp_iovcnt += req->iovcnt;

	/* send the rest of the chunk to our scratch buffer */
	remainder = vol->params.chunk_size - ttl_len;
	if (remainder) {
		req->decomp_iov[req->decomp_iovcnt].iov_base = req->decomp_buf + ttl_len;
		req->decomp_iov[req->decomp_iovcnt].iov_len = remainder;
		ttl_len += req->decomp_iov[req->decomp_iovcnt].iov_len;
		req->decomp_iovcnt++;
	}
	assert(ttl_len == vol->params.chunk_size);

decompress:
	assert(!req->copy_after_decompress || (req->copy_after_decompress && req->decomp_iovcnt == 1));
	req->backing_cb_args.cb_fn = next_fn;
	req->backing_cb_args.cb_arg = req;
	req->comp_buf_iov[0].iov_base = req->comp_buf;
	req->comp_buf_iov[0].iov_len = req->chunk->compressed_size;
	vol->backing_dev->decompress(vol->backing_dev,
				     req->comp_buf_iov, 1, req->decomp_iov, req->decomp_iovcnt,
				     &req->backing_cb_args);
}

static inline void
_prepare_compress_chunk_copy_user_buffers(struct spdk_reduce_vol_request *req, bool zero_paddings)
{
	struct spdk_reduce_vol *vol = req->vol;
	char *padding_buffer = zero_paddings ? g_zero_buf : req->decomp_buf;
	uint64_t chunk_offset, ttl_len = 0;
	uint64_t remainder = 0;
	char *copy_offset = NULL;
	uint32_t lbsize = vol->params.logical_block_size;
	int i;

	req->decomp_iov[0].iov_base = req->decomp_buf;
	req->decomp_iov[0].iov_len = vol->params.chunk_size;
	req->decomp_iovcnt = 1;
	copy_offset = req->decomp_iov[0].iov_base;
	chunk_offset = req->offset % vol->logical_blocks_per_chunk;

	if (chunk_offset) {
		ttl_len += chunk_offset * lbsize;
		/* copy_offset already points to padding buffer if zero_paddings=false */
		if (zero_paddings) {
			memcpy(copy_offset, padding_buffer, ttl_len);
		}
		copy_offset += ttl_len;
	}

	/* now the user data iov, direct from the user buffer */
	for (i = 0; i < req->iovcnt; i++) {
		memcpy(copy_offset, req->iov[i].iov_base, req->iov[i].iov_len);
		copy_offset += req->iov[i].iov_len;
		ttl_len += req->iov[i].iov_len;
	}

	remainder = vol->params.chunk_size - ttl_len;
	if (remainder) {
		/* copy_offset already points to padding buffer if zero_paddings=false */
		if (zero_paddings) {
			memcpy(copy_offset, padding_buffer + ttl_len, remainder);
		}
		ttl_len += remainder;
	}

	assert(ttl_len == req->vol->params.chunk_size);
}

/* This function can be called when we are compressing a new data or in case of read-modify-write
 * In the first case possible paddings should be filled with zeroes, in the second case the paddings
 * should point to already read and decompressed buffer */
static inline void
_prepare_compress_chunk(struct spdk_reduce_vol_request *req, bool zero_paddings)
{
	struct spdk_reduce_vol *vol = req->vol;
	char *padding_buffer = zero_paddings ? g_zero_buf : req->decomp_buf;
	uint64_t chunk_offset, ttl_len = 0;
	uint64_t remainder = 0;
	uint32_t lbsize = vol->params.logical_block_size;
	size_t iov_len;
	int i;

	/* If backing device doesn't support SGL input then we should copy user's buffer into decomp_buf
	 * if at least one of the conditions below is true:
	 * 1. User's buffer is fragmented
	 * 2. Length of the user's buffer is less than the chunk
	 * 3. User's buffer is contig, equals chunk_size but crosses huge page boundary */
	iov_len = req->iov[0].iov_len;
	if (!vol->backing_dev->sgl_in && (req->iovcnt > 1 ||
					  req->iov[0].iov_len < vol->params.chunk_size ||
					  _addr_crosses_huge_page(req->iov[0].iov_base, &iov_len))) {
		_prepare_compress_chunk_copy_user_buffers(req, zero_paddings);
		return;
	}

	req->decomp_iovcnt = 0;
	chunk_offset = req->offset % vol->logical_blocks_per_chunk;

	if (chunk_offset != 0) {
		ttl_len += chunk_offset * lbsize;
		req->decomp_iov[0].iov_base = padding_buffer;
		req->decomp_iov[0].iov_len = ttl_len;
		req->decomp_iovcnt = 1;
	}

	/* now the user data iov, direct from the user buffer */
	for (i = 0; i < req->iovcnt; i++) {
		req->decomp_iov[i + req->decomp_iovcnt].iov_base = req->iov[i].iov_base;
		req->decomp_iov[i + req->decomp_iovcnt].iov_len = req->iov[i].iov_len;
		ttl_len += req->iov[i].iov_len;
	}
	req->decomp_iovcnt += req->iovcnt;

	remainder = vol->params.chunk_size - ttl_len;
	if (remainder) {
		req->decomp_iov[req->decomp_iovcnt].iov_base = padding_buffer + ttl_len;
		req->decomp_iov[req->decomp_iovcnt].iov_len = remainder;
		req->decomp_iovcnt++;
		ttl_len += remainder;
	}
	assert(ttl_len == req->vol->params.chunk_size);
}

static void
_write_decompress_done(void *_req, int reduce_errno)
{
	struct spdk_reduce_vol_request *req = _req;

	/* Negative reduce_errno indicates failure for compression operations. */
	if (reduce_errno < 0) {
		_reduce_vol_complete_req(req, reduce_errno);
		return;
	}

	/* Positive reduce_errno indicates that the output size field in the backing_cb_args
	 * represents the output_size.
	 */
	if (req->backing_cb_args.output_size != req->vol->params.chunk_size) {
		_reduce_vol_complete_req(req, -EIO);
		return;
	}

	_prepare_compress_chunk(req, false);
	_reduce_vol_compress_chunk(req, _write_compress_done);
}

static void
_write_read_done(void *_req, int reduce_errno)
{
	struct spdk_reduce_vol_request *req = _req;

	if (reduce_errno != 0) {
		req->reduce_errno = reduce_errno;
	}

	assert(req->num_backing_ops > 0);
	if (--req->num_backing_ops > 0) {
		return;
	}

	if (req->reduce_errno != 0) {
		_reduce_vol_complete_req(req, req->reduce_errno);
		return;
	}

	if (req->chunk_is_compressed) {
		_reduce_vol_decompress_chunk_scratch(req, _write_decompress_done);
	} else {
		req->backing_cb_args.output_size = req->chunk->compressed_size;

		_write_decompress_done(req, 0);
	}
}

static void
_read_decompress_done(void *_req, int reduce_errno)
{
	struct spdk_reduce_vol_request *req = _req;
	struct spdk_reduce_vol *vol = req->vol;

	/* Negative reduce_errno indicates failure for compression operations. */
	if (reduce_errno < 0) {
		_reduce_vol_complete_req(req, reduce_errno);
		return;
	}

	/* Positive reduce_errno indicates that the output size field in the backing_cb_args
	 * represents the output_size.
	 */
	if (req->backing_cb_args.output_size != vol->params.chunk_size) {
		_reduce_vol_complete_req(req, -EIO);
		return;
	}

	if (req->copy_after_decompress) {
		uint64_t chunk_offset = req->offset % vol->logical_blocks_per_chunk;
		char *decomp_buffer = (char *)req->decomp_buf + chunk_offset * vol->params.logical_block_size;
		int i;

		for (i = 0; i < req->iovcnt; i++) {
			memcpy(req->iov[i].iov_base, decomp_buffer, req->iov[i].iov_len);
			decomp_buffer += req->iov[i].iov_len;
			assert(decomp_buffer <= (char *)req->decomp_buf + vol->params.chunk_size);
		}
	}

	_reduce_vol_complete_req(req, 0);
}

static void
_read_read_done(void *_req, int reduce_errno)
{
	struct spdk_reduce_vol_request *req = _req;
	uint64_t chunk_offset;
	uint8_t *buf;
	int i;

	if (reduce_errno != 0) {
		req->reduce_errno = reduce_errno;
	}

	assert(req->num_backing_ops > 0);
	if (--req->num_backing_ops > 0) {
		return;
	}

	if (req->reduce_errno != 0) {
		_reduce_vol_complete_req(req, req->reduce_errno);
		return;
	}

	if (req->chunk_is_compressed) {
		_reduce_vol_decompress_chunk(req, _read_decompress_done);
	} else {

		/* If the chunk was compressed, the data would have been sent to the
		 *  host buffers by the decompression operation, if not we need to memcpy here.
		 */
		chunk_offset = req->offset % req->vol->logical_blocks_per_chunk;
		buf = req->decomp_buf + chunk_offset * req->vol->params.logical_block_size;
		for (i = 0; i < req->iovcnt; i++) {
			memcpy(req->iov[i].iov_base, buf, req->iov[i].iov_len);
			buf += req->iov[i].iov_len;
		}

		req->backing_cb_args.output_size = req->chunk->compressed_size;

		_read_decompress_done(req, 0);
	}
}

static void
_reduce_vol_read_chunk(struct spdk_reduce_vol_request *req, reduce_request_fn next_fn)
{
	struct spdk_reduce_vol *vol = req->vol;

	req->chunk_map_index = vol->pm_logical_map[req->logical_map_index];
	assert(req->chunk_map_index != UINT32_MAX);

	req->chunk = _reduce_vol_get_chunk_map(vol, req->chunk_map_index);
	req->num_io_units = spdk_divide_round_up(req->chunk->compressed_size,
			    vol->params.backing_io_unit_size);
	req->chunk_is_compressed = (req->num_io_units != vol->backing_io_units_per_chunk);

	_issue_backing_ops(req, vol, next_fn, false /* read */);
}

static bool
_iov_array_is_valid(struct spdk_reduce_vol *vol, struct iovec *iov, int iovcnt,
		    uint64_t length)
{
	uint64_t size = 0;
	int i;

	if (iovcnt > REDUCE_MAX_IOVECS) {
		return false;
	}

	for (i = 0; i < iovcnt; i++) {
		size += iov[i].iov_len;
	}

	return size == (length * vol->params.logical_block_size);
}

static bool
_check_overlap(struct spdk_reduce_vol *vol, uint64_t logical_map_index)
{
	struct spdk_reduce_vol_request *req;

	TAILQ_FOREACH(req, &vol->executing_requests, tailq) {
		if (logical_map_index == req->logical_map_index) {
			return true;
		}
	}

	return false;
}

static void
_start_readv_request(struct spdk_reduce_vol_request *req)
{
	TAILQ_INSERT_TAIL(&req->vol->executing_requests, req, tailq);
	_reduce_vol_read_chunk(req, _read_read_done);
}

void
spdk_reduce_vol_readv(struct spdk_reduce_vol *vol,
		      struct iovec *iov, int iovcnt, uint64_t offset, uint64_t length,
		      spdk_reduce_vol_op_complete cb_fn, void *cb_arg)
{
	struct spdk_reduce_vol_request *req;
	uint64_t logical_map_index;
	bool overlapped;
	int i;

	if (length == 0) {
		cb_fn(cb_arg, 0);
		return;
	}

	if (_request_spans_chunk_boundary(vol, offset, length)) {
		cb_fn(cb_arg, -EINVAL);
		return;
	}

	if (!_iov_array_is_valid(vol, iov, iovcnt, length)) {
		cb_fn(cb_arg, -EINVAL);
		return;
	}

	logical_map_index = offset / vol->logical_blocks_per_chunk;
	overlapped = _check_overlap(vol, logical_map_index);

	if (!overlapped && vol->pm_logical_map[logical_map_index] == REDUCE_EMPTY_MAP_ENTRY) {
		/*
		 * This chunk hasn't been allocated.  So treat the data as all
		 * zeroes for this chunk - do the memset and immediately complete
		 * the operation.
		 */
		for (i = 0; i < iovcnt; i++) {
			memset(iov[i].iov_base, 0, iov[i].iov_len);
		}
		cb_fn(cb_arg, 0);
		return;
	}

	req = TAILQ_FIRST(&vol->free_requests);
	if (req == NULL) {
		cb_fn(cb_arg, -ENOMEM);
		return;
	}

	TAILQ_REMOVE(&vol->free_requests, req, tailq);
	req->type = REDUCE_IO_READV;
	req->vol = vol;
	req->iov = iov;
	req->iovcnt = iovcnt;
	req->offset = offset;
	req->logical_map_index = logical_map_index;
	req->length = length;
	req->copy_after_decompress = false;
	req->cb_fn = cb_fn;
	req->cb_arg = cb_arg;

	if (!overlapped) {
		_start_readv_request(req);
	} else {
		TAILQ_INSERT_TAIL(&vol->queued_requests, req, tailq);
	}
}

static void
_start_writev_request(struct spdk_reduce_vol_request *req)
{
	struct spdk_reduce_vol *vol = req->vol;

	TAILQ_INSERT_TAIL(&req->vol->executing_requests, req, tailq);
	if (vol->pm_logical_map[req->logical_map_index] != REDUCE_EMPTY_MAP_ENTRY) {
		if ((req->length * vol->params.logical_block_size) < vol->params.chunk_size) {
			/* Read old chunk, then overwrite with data from this write
			 *  operation.
			 */
			req->rmw = true;
			_reduce_vol_read_chunk(req, _write_read_done);
			return;
		}
	}

	req->rmw = false;

	_prepare_compress_chunk(req, true);
	_reduce_vol_compress_chunk(req, _write_compress_done);
}

void
spdk_reduce_vol_writev(struct spdk_reduce_vol *vol,
		       struct iovec *iov, int iovcnt, uint64_t offset, uint64_t length,
		       spdk_reduce_vol_op_complete cb_fn, void *cb_arg)
{
	struct spdk_reduce_vol_request *req;
	uint64_t logical_map_index;
	bool overlapped;

	if (length == 0) {
		cb_fn(cb_arg, 0);
		return;
	}

	if (_request_spans_chunk_boundary(vol, offset, length)) {
		cb_fn(cb_arg, -EINVAL);
		return;
	}

	if (!_iov_array_is_valid(vol, iov, iovcnt, length)) {
		cb_fn(cb_arg, -EINVAL);
		return;
	}

	logical_map_index = offset / vol->logical_blocks_per_chunk;
	overlapped = _check_overlap(vol, logical_map_index);

	req = TAILQ_FIRST(&vol->free_requests);
	if (req == NULL) {
		cb_fn(cb_arg, -ENOMEM);
		return;
	}

	TAILQ_REMOVE(&vol->free_requests, req, tailq);
	req->type = REDUCE_IO_WRITEV;
	req->vol = vol;
	req->iov = iov;
	req->iovcnt = iovcnt;
	req->offset = offset;
	req->logical_map_index = logical_map_index;
	req->length = length;
	req->copy_after_decompress = false;
	req->cb_fn = cb_fn;
	req->cb_arg = cb_arg;

	if (!overlapped) {
		_start_writev_request(req);
	} else {
		TAILQ_INSERT_TAIL(&vol->queued_requests, req, tailq);
	}
}

const struct spdk_reduce_vol_params *
spdk_reduce_vol_get_params(struct spdk_reduce_vol *vol)
{
	return &vol->params;
}

const char *
spdk_reduce_vol_get_pm_path(const struct spdk_reduce_vol *vol)
{
	return vol->pm_file.path;
}

void
spdk_reduce_vol_print_info(struct spdk_reduce_vol *vol)
{
	uint64_t logical_map_size, num_chunks, ttl_chunk_sz;
	uint32_t struct_size;
	uint64_t chunk_map_size;

	SPDK_NOTICELOG("vol info:\n");
	SPDK_NOTICELOG("\tvol->params.backing_io_unit_size = 0x%x\n", vol->params.backing_io_unit_size);
	SPDK_NOTICELOG("\tvol->params.logical_block_size = 0x%x\n", vol->params.logical_block_size);
	SPDK_NOTICELOG("\tvol->params.chunk_size = 0x%x\n", vol->params.chunk_size);
	SPDK_NOTICELOG("\tvol->params.vol_size = 0x%" PRIx64 "\n", vol->params.vol_size);
	num_chunks = _get_total_chunks(vol->params.vol_size, vol->params.chunk_size);
	SPDK_NOTICELOG("\ttotal chunks (including extra) = 0x%" PRIx64 "\n", num_chunks);
	SPDK_NOTICELOG("\ttotal chunks (excluding extra) = 0x%" PRIx64 "\n",
		       vol->params.vol_size / vol->params.chunk_size);
	ttl_chunk_sz = _get_pm_total_chunks_size(vol->params.vol_size, vol->params.chunk_size,
			vol->params.backing_io_unit_size);
	SPDK_NOTICELOG("\ttotal_chunks_size = 0x%" PRIx64 "\n", ttl_chunk_sz);
	struct_size = _reduce_vol_get_chunk_struct_size(vol->backing_io_units_per_chunk);
	SPDK_NOTICELOG("\tchunk_struct_size = 0x%x\n", struct_size);

	SPDK_NOTICELOG("pmem info:\n");
	SPDK_NOTICELOG("\tvol->pm_file.size = 0x%" PRIx64 "\n", vol->pm_file.size);
	SPDK_NOTICELOG("\tvol->pm_file.pm_buf = %p\n", (void *)vol->pm_file.pm_buf);
	SPDK_NOTICELOG("\tvol->pm_super = %p\n", (void *)vol->pm_super);
	SPDK_NOTICELOG("\tvol->pm_logical_map = %p\n", (void *)vol->pm_logical_map);
	logical_map_size = _get_pm_logical_map_size(vol->params.vol_size,
			   vol->params.chunk_size);
	SPDK_NOTICELOG("\tlogical_map_size = 0x%" PRIx64 "\n", logical_map_size);
	SPDK_NOTICELOG("\tvol->pm_chunk_maps = %p\n", (void *)vol->pm_chunk_maps);
	chunk_map_size = _get_pm_total_chunks_size(vol->params.vol_size, vol->params.chunk_size,
			 vol->params.backing_io_unit_size);
	SPDK_NOTICELOG("\tchunk_map_size = 0x%" PRIx64 "\n", chunk_map_size);
}

SPDK_LOG_REGISTER_COMPONENT(reduce)