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_get_params(level, src_len);
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;
51 zstd_in_buffer in_buf;
52 zstd_out_buffer 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 void zstd_free_workspace(struct list_head *ws);
95 struct list_head *zstd_alloc_workspace(unsigned int level);
98 * Timer callback to free unused workspaces.
102 * This scans the lru_list and attempts to reclaim any workspace that hasn't
103 * been used for ZSTD_BTRFS_RECLAIM_JIFFIES.
105 * The context is softirq and does not need the _bh locking primitives.
107 static void zstd_reclaim_timer_fn(struct timer_list *timer)
109 unsigned long reclaim_threshold = jiffies - ZSTD_BTRFS_RECLAIM_JIFFIES;
110 struct list_head *pos, *next;
112 spin_lock(&wsm.lock);
114 if (list_empty(&wsm.lru_list)) {
115 spin_unlock(&wsm.lock);
119 list_for_each_prev_safe(pos, next, &wsm.lru_list) {
120 struct workspace *victim = container_of(pos, struct workspace,
124 if (time_after(victim->last_used, reclaim_threshold))
127 /* workspace is in use */
128 if (victim->req_level)
131 level = victim->level;
132 list_del(&victim->lru_list);
133 list_del(&victim->list);
134 zstd_free_workspace(&victim->list);
136 if (list_empty(&wsm.idle_ws[level - 1]))
137 clear_bit(level - 1, &wsm.active_map);
141 if (!list_empty(&wsm.lru_list))
142 mod_timer(&wsm.timer, jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
144 spin_unlock(&wsm.lock);
148 * Calculate monotonic memory bounds.
150 * It is possible based on the level configurations that a higher level
151 * workspace uses less memory than a lower level workspace. In order to reuse
152 * workspaces, this must be made a monotonic relationship. This precomputes
153 * the required memory for each level and enforces the monotonicity between
154 * level and memory required.
156 static void zstd_calc_ws_mem_sizes(void)
161 for (level = 1; level <= ZSTD_BTRFS_MAX_LEVEL; level++) {
162 zstd_parameters params =
163 zstd_get_btrfs_parameters(level, ZSTD_BTRFS_MAX_INPUT);
166 zstd_cstream_workspace_bound(¶ms.cParams),
167 zstd_dstream_workspace_bound(ZSTD_BTRFS_MAX_INPUT));
169 max_size = max_t(size_t, max_size, level_size);
170 zstd_ws_mem_sizes[level - 1] = max_size;
174 void zstd_init_workspace_manager(void)
176 struct list_head *ws;
179 zstd_calc_ws_mem_sizes();
181 wsm.ops = &btrfs_zstd_compress;
182 spin_lock_init(&wsm.lock);
183 init_waitqueue_head(&wsm.wait);
184 timer_setup(&wsm.timer, zstd_reclaim_timer_fn, 0);
186 INIT_LIST_HEAD(&wsm.lru_list);
187 for (i = 0; i < ZSTD_BTRFS_MAX_LEVEL; i++)
188 INIT_LIST_HEAD(&wsm.idle_ws[i]);
190 ws = zstd_alloc_workspace(ZSTD_BTRFS_MAX_LEVEL);
193 "BTRFS: cannot preallocate zstd compression workspace\n");
195 set_bit(ZSTD_BTRFS_MAX_LEVEL - 1, &wsm.active_map);
196 list_add(ws, &wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1]);
200 void zstd_cleanup_workspace_manager(void)
202 struct workspace *workspace;
205 spin_lock_bh(&wsm.lock);
206 for (i = 0; i < ZSTD_BTRFS_MAX_LEVEL; i++) {
207 while (!list_empty(&wsm.idle_ws[i])) {
208 workspace = container_of(wsm.idle_ws[i].next,
209 struct workspace, list);
210 list_del(&workspace->list);
211 list_del(&workspace->lru_list);
212 zstd_free_workspace(&workspace->list);
215 spin_unlock_bh(&wsm.lock);
217 del_timer_sync(&wsm.timer);
221 * Find workspace for given level.
223 * @level: compression level
225 * This iterates over the set bits in the active_map beginning at the requested
226 * compression level. This lets us utilize already allocated workspaces before
227 * allocating a new one. If the workspace is of a larger size, it is used, but
228 * the place in the lru_list and last_used times are not updated. This is to
229 * offer the opportunity to reclaim the workspace in favor of allocating an
230 * appropriately sized one in the future.
232 static struct list_head *zstd_find_workspace(unsigned int level)
234 struct list_head *ws;
235 struct workspace *workspace;
238 spin_lock_bh(&wsm.lock);
239 for_each_set_bit_from(i, &wsm.active_map, ZSTD_BTRFS_MAX_LEVEL) {
240 if (!list_empty(&wsm.idle_ws[i])) {
241 ws = wsm.idle_ws[i].next;
242 workspace = list_to_workspace(ws);
244 /* keep its place if it's a lower level using this */
245 workspace->req_level = level;
246 if (level == workspace->level)
247 list_del(&workspace->lru_list);
248 if (list_empty(&wsm.idle_ws[i]))
249 clear_bit(i, &wsm.active_map);
250 spin_unlock_bh(&wsm.lock);
254 spin_unlock_bh(&wsm.lock);
260 * Zstd get_workspace for level.
262 * @level: compression level
264 * If @level is 0, then any compression level can be used. Therefore, we begin
265 * scanning from 1. We first scan through possible workspaces and then after
266 * attempt to allocate a new workspace. If we fail to allocate one due to
267 * memory pressure, go to sleep waiting for the max level workspace to free up.
269 struct list_head *zstd_get_workspace(unsigned int level)
271 struct list_head *ws;
272 unsigned int nofs_flag;
274 /* level == 0 means we can use any workspace */
279 ws = zstd_find_workspace(level);
283 nofs_flag = memalloc_nofs_save();
284 ws = zstd_alloc_workspace(level);
285 memalloc_nofs_restore(nofs_flag);
290 prepare_to_wait(&wsm.wait, &wait, TASK_UNINTERRUPTIBLE);
292 finish_wait(&wsm.wait, &wait);
301 * Zstd put_workspace.
303 * @ws: list_head for the workspace
305 * When putting back a workspace, we only need to update the LRU if we are of
306 * the requested compression level. Here is where we continue to protect the
307 * max level workspace or update last_used accordingly. If the reclaim timer
308 * isn't set, it is also set here. Only the max level workspace tries and wakes
309 * up waiting workspaces.
311 void zstd_put_workspace(struct list_head *ws)
313 struct workspace *workspace = list_to_workspace(ws);
315 spin_lock_bh(&wsm.lock);
317 /* A node is only taken off the lru if we are the corresponding level */
318 if (workspace->req_level == workspace->level) {
319 /* Hide a max level workspace from reclaim */
320 if (list_empty(&wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1])) {
321 INIT_LIST_HEAD(&workspace->lru_list);
323 workspace->last_used = jiffies;
324 list_add(&workspace->lru_list, &wsm.lru_list);
325 if (!timer_pending(&wsm.timer))
326 mod_timer(&wsm.timer,
327 jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
331 set_bit(workspace->level - 1, &wsm.active_map);
332 list_add(&workspace->list, &wsm.idle_ws[workspace->level - 1]);
333 workspace->req_level = 0;
335 spin_unlock_bh(&wsm.lock);
337 if (workspace->level == ZSTD_BTRFS_MAX_LEVEL)
338 cond_wake_up(&wsm.wait);
341 void zstd_free_workspace(struct list_head *ws)
343 struct workspace *workspace = list_entry(ws, struct workspace, list);
345 kvfree(workspace->mem);
346 kfree(workspace->buf);
350 struct list_head *zstd_alloc_workspace(unsigned int level)
352 struct workspace *workspace;
354 workspace = kzalloc(sizeof(*workspace), GFP_KERNEL);
356 return ERR_PTR(-ENOMEM);
358 workspace->size = zstd_ws_mem_sizes[level - 1];
359 workspace->level = level;
360 workspace->req_level = level;
361 workspace->last_used = jiffies;
362 workspace->mem = kvmalloc(workspace->size, GFP_KERNEL | __GFP_NOWARN);
363 workspace->buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
364 if (!workspace->mem || !workspace->buf)
367 INIT_LIST_HEAD(&workspace->list);
368 INIT_LIST_HEAD(&workspace->lru_list);
370 return &workspace->list;
372 zstd_free_workspace(&workspace->list);
373 return ERR_PTR(-ENOMEM);
376 int zstd_compress_pages(struct list_head *ws, struct address_space *mapping,
377 u64 start, struct page **pages, unsigned long *out_pages,
378 unsigned long *total_in, unsigned long *total_out)
380 struct workspace *workspace = list_entry(ws, struct workspace, list);
381 zstd_cstream *stream;
384 struct page *in_page = NULL; /* The current page to read */
385 struct page *out_page = NULL; /* The current page to write to */
386 unsigned long tot_in = 0;
387 unsigned long tot_out = 0;
388 unsigned long len = *total_out;
389 const unsigned long nr_dest_pages = *out_pages;
390 unsigned long max_out = nr_dest_pages * PAGE_SIZE;
391 zstd_parameters params = zstd_get_btrfs_parameters(workspace->req_level,
398 /* Initialize the stream */
399 stream = zstd_init_cstream(¶ms, len, workspace->mem,
402 pr_warn("BTRFS: zstd_init_cstream failed\n");
407 /* map in the first page of input data */
408 in_page = find_get_page(mapping, start >> PAGE_SHIFT);
409 workspace->in_buf.src = kmap_local_page(in_page);
410 workspace->in_buf.pos = 0;
411 workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
414 /* Allocate and map in the output buffer */
415 out_page = alloc_page(GFP_NOFS);
416 if (out_page == NULL) {
420 pages[nr_pages++] = out_page;
421 workspace->out_buf.dst = page_address(out_page);
422 workspace->out_buf.pos = 0;
423 workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
428 ret2 = zstd_compress_stream(stream, &workspace->out_buf,
430 if (zstd_is_error(ret2)) {
431 pr_debug("BTRFS: zstd_compress_stream returned %d\n",
432 zstd_get_error_code(ret2));
437 /* Check to see if we are making it bigger */
438 if (tot_in + workspace->in_buf.pos > 8192 &&
439 tot_in + workspace->in_buf.pos <
440 tot_out + workspace->out_buf.pos) {
445 /* We've reached the end of our output range */
446 if (workspace->out_buf.pos >= max_out) {
447 tot_out += workspace->out_buf.pos;
452 /* Check if we need more output space */
453 if (workspace->out_buf.pos == workspace->out_buf.size) {
454 tot_out += PAGE_SIZE;
455 max_out -= PAGE_SIZE;
456 if (nr_pages == nr_dest_pages) {
460 out_page = alloc_page(GFP_NOFS);
461 if (out_page == NULL) {
465 pages[nr_pages++] = out_page;
466 workspace->out_buf.dst = page_address(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) {
481 kunmap_local(workspace->in_buf.src);
485 in_page = find_get_page(mapping, start >> PAGE_SHIFT);
486 workspace->in_buf.src = kmap_local_page(in_page);
487 workspace->in_buf.pos = 0;
488 workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
494 ret2 = zstd_end_stream(stream, &workspace->out_buf);
495 if (zstd_is_error(ret2)) {
496 pr_debug("BTRFS: zstd_end_stream returned %d\n",
497 zstd_get_error_code(ret2));
502 tot_out += workspace->out_buf.pos;
505 if (workspace->out_buf.pos >= max_out) {
506 tot_out += workspace->out_buf.pos;
511 tot_out += PAGE_SIZE;
512 max_out -= PAGE_SIZE;
513 if (nr_pages == nr_dest_pages) {
517 out_page = alloc_page(GFP_NOFS);
518 if (out_page == NULL) {
522 pages[nr_pages++] = out_page;
523 workspace->out_buf.dst = page_address(out_page);
524 workspace->out_buf.pos = 0;
525 workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
528 if (tot_out >= tot_in) {
535 *total_out = tot_out;
537 *out_pages = nr_pages;
538 if (workspace->in_buf.src) {
539 kunmap_local(workspace->in_buf.src);
545 int zstd_decompress_bio(struct list_head *ws, struct compressed_bio *cb)
547 struct workspace *workspace = list_entry(ws, struct workspace, list);
548 struct page **pages_in = cb->compressed_pages;
549 size_t srclen = cb->compressed_len;
550 zstd_dstream *stream;
552 unsigned long page_in_index = 0;
553 unsigned long total_pages_in = DIV_ROUND_UP(srclen, PAGE_SIZE);
554 unsigned long buf_start;
555 unsigned long total_out = 0;
557 stream = zstd_init_dstream(
558 ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
560 pr_debug("BTRFS: zstd_init_dstream failed\n");
565 workspace->in_buf.src = kmap_local_page(pages_in[page_in_index]);
566 workspace->in_buf.pos = 0;
567 workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
569 workspace->out_buf.dst = workspace->buf;
570 workspace->out_buf.pos = 0;
571 workspace->out_buf.size = PAGE_SIZE;
576 ret2 = zstd_decompress_stream(stream, &workspace->out_buf,
578 if (zstd_is_error(ret2)) {
579 pr_debug("BTRFS: zstd_decompress_stream returned %d\n",
580 zstd_get_error_code(ret2));
584 buf_start = total_out;
585 total_out += workspace->out_buf.pos;
586 workspace->out_buf.pos = 0;
588 ret = btrfs_decompress_buf2page(workspace->out_buf.dst,
589 total_out - buf_start, cb, buf_start);
593 if (workspace->in_buf.pos >= srclen)
596 /* Check if we've hit the end of a frame */
600 if (workspace->in_buf.pos == workspace->in_buf.size) {
601 kunmap_local(workspace->in_buf.src);
603 if (page_in_index >= total_pages_in) {
604 workspace->in_buf.src = NULL;
609 workspace->in_buf.src = kmap_local_page(pages_in[page_in_index]);
610 workspace->in_buf.pos = 0;
611 workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
616 if (workspace->in_buf.src)
617 kunmap_local(workspace->in_buf.src);
621 int zstd_decompress(struct list_head *ws, const u8 *data_in,
622 struct page *dest_page, unsigned long start_byte, size_t srclen,
625 struct workspace *workspace = list_entry(ws, struct workspace, list);
626 zstd_dstream *stream;
629 unsigned long total_out = 0;
630 unsigned long pg_offset = 0;
632 stream = zstd_init_dstream(
633 ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
635 pr_warn("BTRFS: zstd_init_dstream failed\n");
640 destlen = min_t(size_t, destlen, PAGE_SIZE);
642 workspace->in_buf.src = data_in;
643 workspace->in_buf.pos = 0;
644 workspace->in_buf.size = srclen;
646 workspace->out_buf.dst = workspace->buf;
647 workspace->out_buf.pos = 0;
648 workspace->out_buf.size = PAGE_SIZE;
651 while (pg_offset < destlen
652 && workspace->in_buf.pos < workspace->in_buf.size) {
653 unsigned long buf_start;
654 unsigned long buf_offset;
657 /* Check if the frame is over and we still need more input */
659 pr_debug("BTRFS: zstd_decompress_stream ended early\n");
663 ret2 = zstd_decompress_stream(stream, &workspace->out_buf,
665 if (zstd_is_error(ret2)) {
666 pr_debug("BTRFS: zstd_decompress_stream returned %d\n",
667 zstd_get_error_code(ret2));
672 buf_start = total_out;
673 total_out += workspace->out_buf.pos;
674 workspace->out_buf.pos = 0;
676 if (total_out <= start_byte)
679 if (total_out > start_byte && buf_start < start_byte)
680 buf_offset = start_byte - buf_start;
684 bytes = min_t(unsigned long, destlen - pg_offset,
685 workspace->out_buf.size - buf_offset);
687 memcpy_to_page(dest_page, pg_offset,
688 workspace->out_buf.dst + buf_offset, bytes);
694 if (pg_offset < destlen) {
695 memzero_page(dest_page, pg_offset, destlen - pg_offset);
700 const struct btrfs_compress_op btrfs_zstd_compress = {
701 /* ZSTD uses own workspace manager */
702 .workspace_manager = NULL,
703 .max_level = ZSTD_BTRFS_MAX_LEVEL,
704 .default_level = ZSTD_BTRFS_DEFAULT_LEVEL,