1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2016-present, Facebook, Inc.
9 #include <linux/bitmap.h>
10 #include <linux/err.h>
11 #include <linux/init.h>
12 #include <linux/kernel.h>
14 #include <linux/sched/mm.h>
15 #include <linux/pagemap.h>
16 #include <linux/refcount.h>
17 #include <linux/sched.h>
18 #include <linux/slab.h>
19 #include <linux/zstd.h>
21 #include "compression.h"
24 #define ZSTD_BTRFS_MAX_WINDOWLOG 17
25 #define ZSTD_BTRFS_MAX_INPUT (1 << ZSTD_BTRFS_MAX_WINDOWLOG)
26 #define ZSTD_BTRFS_DEFAULT_LEVEL 3
27 #define ZSTD_BTRFS_MAX_LEVEL 15
28 /* 307s to avoid pathologically clashing with transaction commit */
29 #define ZSTD_BTRFS_RECLAIM_JIFFIES (307 * HZ)
31 static ZSTD_parameters zstd_get_btrfs_parameters(unsigned int level,
34 ZSTD_parameters params = ZSTD_getParams(level, src_len, 0);
36 if (params.cParams.windowLog > ZSTD_BTRFS_MAX_WINDOWLOG)
37 params.cParams.windowLog = ZSTD_BTRFS_MAX_WINDOWLOG;
38 WARN_ON(src_len > ZSTD_BTRFS_MAX_INPUT);
47 unsigned int req_level;
48 unsigned long last_used; /* jiffies */
49 struct list_head list;
50 struct list_head lru_list;
52 ZSTD_outBuffer out_buf;
56 * Zstd Workspace Management
58 * Zstd workspaces have different memory requirements depending on the level.
59 * The zstd workspaces are managed by having individual lists for each level
60 * and a global lru. Forward progress is maintained by protecting a max level
63 * Getting a workspace is done by using the bitmap to identify the levels that
64 * have available workspaces and scans up. This lets us recycle higher level
65 * workspaces because of the monotonic memory guarantee. A workspace's
66 * last_used is only updated if it is being used by the corresponding memory
67 * level. Putting a workspace involves adding it back to the appropriate places
68 * and adding it back to the lru if necessary.
70 * A timer is used to reclaim workspaces if they have not been used for
71 * ZSTD_BTRFS_RECLAIM_JIFFIES. This helps keep only active workspaces around.
72 * The upper bound is provided by the workqueue limit which is 2 (percpu limit).
75 struct zstd_workspace_manager {
76 const struct btrfs_compress_op *ops;
78 struct list_head lru_list;
79 struct list_head idle_ws[ZSTD_BTRFS_MAX_LEVEL];
80 unsigned long active_map;
81 wait_queue_head_t wait;
82 struct timer_list timer;
85 static struct zstd_workspace_manager wsm;
87 static size_t zstd_ws_mem_sizes[ZSTD_BTRFS_MAX_LEVEL];
89 static inline struct workspace *list_to_workspace(struct list_head *list)
91 return container_of(list, struct workspace, list);
94 static void zstd_free_workspace(struct list_head *ws);
95 static struct list_head *zstd_alloc_workspace(unsigned int level);
98 * zstd_reclaim_timer_fn - reclaim timer
101 * This scans the lru_list and attempts to reclaim any workspace that hasn't
102 * been used for ZSTD_BTRFS_RECLAIM_JIFFIES.
104 static void zstd_reclaim_timer_fn(struct timer_list *timer)
106 unsigned long reclaim_threshold = jiffies - ZSTD_BTRFS_RECLAIM_JIFFIES;
107 struct list_head *pos, *next;
109 spin_lock_bh(&wsm.lock);
111 if (list_empty(&wsm.lru_list)) {
112 spin_unlock_bh(&wsm.lock);
116 list_for_each_prev_safe(pos, next, &wsm.lru_list) {
117 struct workspace *victim = container_of(pos, struct workspace,
121 if (time_after(victim->last_used, reclaim_threshold))
124 /* workspace is in use */
125 if (victim->req_level)
128 level = victim->level;
129 list_del(&victim->lru_list);
130 list_del(&victim->list);
131 zstd_free_workspace(&victim->list);
133 if (list_empty(&wsm.idle_ws[level - 1]))
134 clear_bit(level - 1, &wsm.active_map);
138 if (!list_empty(&wsm.lru_list))
139 mod_timer(&wsm.timer, jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
141 spin_unlock_bh(&wsm.lock);
145 * zstd_calc_ws_mem_sizes - calculate monotonic memory bounds
147 * It is possible based on the level configurations that a higher level
148 * workspace uses less memory than a lower level workspace. In order to reuse
149 * workspaces, this must be made a monotonic relationship. This precomputes
150 * the required memory for each level and enforces the monotonicity between
151 * level and memory required.
153 static void zstd_calc_ws_mem_sizes(void)
158 for (level = 1; level <= ZSTD_BTRFS_MAX_LEVEL; level++) {
159 ZSTD_parameters params =
160 zstd_get_btrfs_parameters(level, ZSTD_BTRFS_MAX_INPUT);
163 ZSTD_CStreamWorkspaceBound(params.cParams),
164 ZSTD_DStreamWorkspaceBound(ZSTD_BTRFS_MAX_INPUT));
166 max_size = max_t(size_t, max_size, level_size);
167 zstd_ws_mem_sizes[level - 1] = max_size;
171 static void zstd_init_workspace_manager(void)
173 struct list_head *ws;
176 zstd_calc_ws_mem_sizes();
178 wsm.ops = &btrfs_zstd_compress;
179 spin_lock_init(&wsm.lock);
180 init_waitqueue_head(&wsm.wait);
181 timer_setup(&wsm.timer, zstd_reclaim_timer_fn, 0);
183 INIT_LIST_HEAD(&wsm.lru_list);
184 for (i = 0; i < ZSTD_BTRFS_MAX_LEVEL; i++)
185 INIT_LIST_HEAD(&wsm.idle_ws[i]);
187 ws = zstd_alloc_workspace(ZSTD_BTRFS_MAX_LEVEL);
190 "BTRFS: cannot preallocate zstd compression workspace\n");
192 set_bit(ZSTD_BTRFS_MAX_LEVEL - 1, &wsm.active_map);
193 list_add(ws, &wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1]);
197 static void zstd_cleanup_workspace_manager(void)
199 struct workspace *workspace;
202 spin_lock_bh(&wsm.lock);
203 for (i = 0; i < ZSTD_BTRFS_MAX_LEVEL; i++) {
204 while (!list_empty(&wsm.idle_ws[i])) {
205 workspace = container_of(wsm.idle_ws[i].next,
206 struct workspace, list);
207 list_del(&workspace->list);
208 list_del(&workspace->lru_list);
209 zstd_free_workspace(&workspace->list);
212 spin_unlock_bh(&wsm.lock);
214 del_timer_sync(&wsm.timer);
218 * zstd_find_workspace - find workspace
219 * @level: compression level
221 * This iterates over the set bits in the active_map beginning at the requested
222 * compression level. This lets us utilize already allocated workspaces before
223 * allocating a new one. If the workspace is of a larger size, it is used, but
224 * the place in the lru_list and last_used times are not updated. This is to
225 * offer the opportunity to reclaim the workspace in favor of allocating an
226 * appropriately sized one in the future.
228 static struct list_head *zstd_find_workspace(unsigned int level)
230 struct list_head *ws;
231 struct workspace *workspace;
234 spin_lock_bh(&wsm.lock);
235 for_each_set_bit_from(i, &wsm.active_map, ZSTD_BTRFS_MAX_LEVEL) {
236 if (!list_empty(&wsm.idle_ws[i])) {
237 ws = wsm.idle_ws[i].next;
238 workspace = list_to_workspace(ws);
240 /* keep its place if it's a lower level using this */
241 workspace->req_level = level;
242 if (level == workspace->level)
243 list_del(&workspace->lru_list);
244 if (list_empty(&wsm.idle_ws[i]))
245 clear_bit(i, &wsm.active_map);
246 spin_unlock_bh(&wsm.lock);
250 spin_unlock_bh(&wsm.lock);
256 * zstd_get_workspace - zstd's get_workspace
257 * @level: compression level
259 * If @level is 0, then any compression level can be used. Therefore, we begin
260 * scanning from 1. We first scan through possible workspaces and then after
261 * attempt to allocate a new workspace. If we fail to allocate one due to
262 * memory pressure, go to sleep waiting for the max level workspace to free up.
264 static struct list_head *zstd_get_workspace(unsigned int level)
266 struct list_head *ws;
267 unsigned int nofs_flag;
269 /* level == 0 means we can use any workspace */
274 ws = zstd_find_workspace(level);
278 nofs_flag = memalloc_nofs_save();
279 ws = zstd_alloc_workspace(level);
280 memalloc_nofs_restore(nofs_flag);
285 prepare_to_wait(&wsm.wait, &wait, TASK_UNINTERRUPTIBLE);
287 finish_wait(&wsm.wait, &wait);
296 * zstd_put_workspace - zstd put_workspace
297 * @ws: list_head for the workspace
299 * When putting back a workspace, we only need to update the LRU if we are of
300 * the requested compression level. Here is where we continue to protect the
301 * max level workspace or update last_used accordingly. If the reclaim timer
302 * isn't set, it is also set here. Only the max level workspace tries and wakes
303 * up waiting workspaces.
305 static void zstd_put_workspace(struct list_head *ws)
307 struct workspace *workspace = list_to_workspace(ws);
309 spin_lock_bh(&wsm.lock);
311 /* A node is only taken off the lru if we are the corresponding level */
312 if (workspace->req_level == workspace->level) {
313 /* Hide a max level workspace from reclaim */
314 if (list_empty(&wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1])) {
315 INIT_LIST_HEAD(&workspace->lru_list);
317 workspace->last_used = jiffies;
318 list_add(&workspace->lru_list, &wsm.lru_list);
319 if (!timer_pending(&wsm.timer))
320 mod_timer(&wsm.timer,
321 jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
325 set_bit(workspace->level - 1, &wsm.active_map);
326 list_add(&workspace->list, &wsm.idle_ws[workspace->level - 1]);
327 workspace->req_level = 0;
329 spin_unlock_bh(&wsm.lock);
331 if (workspace->level == ZSTD_BTRFS_MAX_LEVEL)
332 cond_wake_up(&wsm.wait);
335 static void zstd_free_workspace(struct list_head *ws)
337 struct workspace *workspace = list_entry(ws, struct workspace, list);
339 kvfree(workspace->mem);
340 kfree(workspace->buf);
344 static struct list_head *zstd_alloc_workspace(unsigned int level)
346 struct workspace *workspace;
348 workspace = kzalloc(sizeof(*workspace), GFP_KERNEL);
350 return ERR_PTR(-ENOMEM);
352 workspace->size = zstd_ws_mem_sizes[level - 1];
353 workspace->level = level;
354 workspace->req_level = level;
355 workspace->last_used = jiffies;
356 workspace->mem = kvmalloc(workspace->size, GFP_KERNEL);
357 workspace->buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
358 if (!workspace->mem || !workspace->buf)
361 INIT_LIST_HEAD(&workspace->list);
362 INIT_LIST_HEAD(&workspace->lru_list);
364 return &workspace->list;
366 zstd_free_workspace(&workspace->list);
367 return ERR_PTR(-ENOMEM);
370 static int zstd_compress_pages(struct list_head *ws,
371 struct address_space *mapping,
374 unsigned long *out_pages,
375 unsigned long *total_in,
376 unsigned long *total_out)
378 struct workspace *workspace = list_entry(ws, struct workspace, list);
379 ZSTD_CStream *stream;
382 struct page *in_page = NULL; /* The current page to read */
383 struct page *out_page = NULL; /* The current page to write to */
384 unsigned long tot_in = 0;
385 unsigned long tot_out = 0;
386 unsigned long len = *total_out;
387 const unsigned long nr_dest_pages = *out_pages;
388 unsigned long max_out = nr_dest_pages * PAGE_SIZE;
389 ZSTD_parameters params = zstd_get_btrfs_parameters(workspace->req_level,
396 /* Initialize the stream */
397 stream = ZSTD_initCStream(params, len, workspace->mem,
400 pr_warn("BTRFS: ZSTD_initCStream failed\n");
405 /* map in the first page of input data */
406 in_page = find_get_page(mapping, start >> PAGE_SHIFT);
407 workspace->in_buf.src = kmap(in_page);
408 workspace->in_buf.pos = 0;
409 workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
412 /* Allocate and map in the output buffer */
413 out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
414 if (out_page == NULL) {
418 pages[nr_pages++] = out_page;
419 workspace->out_buf.dst = kmap(out_page);
420 workspace->out_buf.pos = 0;
421 workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
426 ret2 = ZSTD_compressStream(stream, &workspace->out_buf,
428 if (ZSTD_isError(ret2)) {
429 pr_debug("BTRFS: ZSTD_compressStream returned %d\n",
430 ZSTD_getErrorCode(ret2));
435 /* Check to see if we are making it bigger */
436 if (tot_in + workspace->in_buf.pos > 8192 &&
437 tot_in + workspace->in_buf.pos <
438 tot_out + workspace->out_buf.pos) {
443 /* We've reached the end of our output range */
444 if (workspace->out_buf.pos >= max_out) {
445 tot_out += workspace->out_buf.pos;
450 /* Check if we need more output space */
451 if (workspace->out_buf.pos == workspace->out_buf.size) {
452 tot_out += PAGE_SIZE;
453 max_out -= PAGE_SIZE;
455 if (nr_pages == nr_dest_pages) {
460 out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
461 if (out_page == NULL) {
465 pages[nr_pages++] = out_page;
466 workspace->out_buf.dst = kmap(out_page);
467 workspace->out_buf.pos = 0;
468 workspace->out_buf.size = min_t(size_t, max_out,
472 /* We've reached the end of the input */
473 if (workspace->in_buf.pos >= len) {
474 tot_in += workspace->in_buf.pos;
478 /* Check if we need more input */
479 if (workspace->in_buf.pos == workspace->in_buf.size) {
486 in_page = find_get_page(mapping, start >> PAGE_SHIFT);
487 workspace->in_buf.src = kmap(in_page);
488 workspace->in_buf.pos = 0;
489 workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
495 ret2 = ZSTD_endStream(stream, &workspace->out_buf);
496 if (ZSTD_isError(ret2)) {
497 pr_debug("BTRFS: ZSTD_endStream returned %d\n",
498 ZSTD_getErrorCode(ret2));
503 tot_out += workspace->out_buf.pos;
506 if (workspace->out_buf.pos >= max_out) {
507 tot_out += workspace->out_buf.pos;
512 tot_out += PAGE_SIZE;
513 max_out -= PAGE_SIZE;
515 if (nr_pages == nr_dest_pages) {
520 out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
521 if (out_page == NULL) {
525 pages[nr_pages++] = out_page;
526 workspace->out_buf.dst = kmap(out_page);
527 workspace->out_buf.pos = 0;
528 workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
531 if (tot_out >= tot_in) {
538 *total_out = tot_out;
540 *out_pages = nr_pages;
551 static int zstd_decompress_bio(struct list_head *ws, struct compressed_bio *cb)
553 struct workspace *workspace = list_entry(ws, struct workspace, list);
554 struct page **pages_in = cb->compressed_pages;
555 u64 disk_start = cb->start;
556 struct bio *orig_bio = cb->orig_bio;
557 size_t srclen = cb->compressed_len;
558 ZSTD_DStream *stream;
560 unsigned long page_in_index = 0;
561 unsigned long total_pages_in = DIV_ROUND_UP(srclen, PAGE_SIZE);
562 unsigned long buf_start;
563 unsigned long total_out = 0;
565 stream = ZSTD_initDStream(
566 ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
568 pr_debug("BTRFS: ZSTD_initDStream failed\n");
573 workspace->in_buf.src = kmap(pages_in[page_in_index]);
574 workspace->in_buf.pos = 0;
575 workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
577 workspace->out_buf.dst = workspace->buf;
578 workspace->out_buf.pos = 0;
579 workspace->out_buf.size = PAGE_SIZE;
584 ret2 = ZSTD_decompressStream(stream, &workspace->out_buf,
586 if (ZSTD_isError(ret2)) {
587 pr_debug("BTRFS: ZSTD_decompressStream returned %d\n",
588 ZSTD_getErrorCode(ret2));
592 buf_start = total_out;
593 total_out += workspace->out_buf.pos;
594 workspace->out_buf.pos = 0;
596 ret = btrfs_decompress_buf2page(workspace->out_buf.dst,
597 buf_start, total_out, disk_start, orig_bio);
601 if (workspace->in_buf.pos >= srclen)
604 /* Check if we've hit the end of a frame */
608 if (workspace->in_buf.pos == workspace->in_buf.size) {
609 kunmap(pages_in[page_in_index++]);
610 if (page_in_index >= total_pages_in) {
611 workspace->in_buf.src = NULL;
616 workspace->in_buf.src = kmap(pages_in[page_in_index]);
617 workspace->in_buf.pos = 0;
618 workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
622 zero_fill_bio(orig_bio);
624 if (workspace->in_buf.src)
625 kunmap(pages_in[page_in_index]);
629 static int zstd_decompress(struct list_head *ws, unsigned char *data_in,
630 struct page *dest_page,
631 unsigned long start_byte,
632 size_t srclen, size_t destlen)
634 struct workspace *workspace = list_entry(ws, struct workspace, list);
635 ZSTD_DStream *stream;
638 unsigned long total_out = 0;
639 unsigned long pg_offset = 0;
642 stream = ZSTD_initDStream(
643 ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
645 pr_warn("BTRFS: ZSTD_initDStream failed\n");
650 destlen = min_t(size_t, destlen, PAGE_SIZE);
652 workspace->in_buf.src = data_in;
653 workspace->in_buf.pos = 0;
654 workspace->in_buf.size = srclen;
656 workspace->out_buf.dst = workspace->buf;
657 workspace->out_buf.pos = 0;
658 workspace->out_buf.size = PAGE_SIZE;
661 while (pg_offset < destlen
662 && workspace->in_buf.pos < workspace->in_buf.size) {
663 unsigned long buf_start;
664 unsigned long buf_offset;
667 /* Check if the frame is over and we still need more input */
669 pr_debug("BTRFS: ZSTD_decompressStream ended early\n");
673 ret2 = ZSTD_decompressStream(stream, &workspace->out_buf,
675 if (ZSTD_isError(ret2)) {
676 pr_debug("BTRFS: ZSTD_decompressStream returned %d\n",
677 ZSTD_getErrorCode(ret2));
682 buf_start = total_out;
683 total_out += workspace->out_buf.pos;
684 workspace->out_buf.pos = 0;
686 if (total_out <= start_byte)
689 if (total_out > start_byte && buf_start < start_byte)
690 buf_offset = start_byte - buf_start;
694 bytes = min_t(unsigned long, destlen - pg_offset,
695 workspace->out_buf.size - buf_offset);
697 kaddr = kmap_atomic(dest_page);
698 memcpy(kaddr + pg_offset, workspace->out_buf.dst + buf_offset,
700 kunmap_atomic(kaddr);
706 if (pg_offset < destlen) {
707 kaddr = kmap_atomic(dest_page);
708 memset(kaddr + pg_offset, 0, destlen - pg_offset);
709 kunmap_atomic(kaddr);
714 const struct btrfs_compress_op btrfs_zstd_compress = {
715 .init_workspace_manager = zstd_init_workspace_manager,
716 .cleanup_workspace_manager = zstd_cleanup_workspace_manager,
717 .get_workspace = zstd_get_workspace,
718 .put_workspace = zstd_put_workspace,
719 .alloc_workspace = zstd_alloc_workspace,
720 .free_workspace = zstd_free_workspace,
721 .compress_pages = zstd_compress_pages,
722 .decompress_bio = zstd_decompress_bio,
723 .decompress = zstd_decompress,
724 .max_level = ZSTD_BTRFS_MAX_LEVEL,
725 .default_level = ZSTD_BTRFS_DEFAULT_LEVEL,