Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2016-present, Facebook, Inc.
4 * All rights reserved.
5 *
6 */
7
8#include <linux/bio.h>
9#include <linux/bitmap.h>
10#include <linux/err.h>
11#include <linux/init.h>
12#include <linux/kernel.h>
13#include <linux/mm.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>
20#include "misc.h"
21#include "compression.h"
22#include "ctree.h"
23
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)
30
31static zstd_parameters zstd_get_btrfs_parameters(unsigned int level,
32 size_t src_len)
33{
34 zstd_parameters params = zstd_get_params(level, src_len);
35
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);
39 return params;
40}
41
42struct workspace {
43 void *mem;
44 size_t size;
45 char *buf;
46 unsigned int level;
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;
53};
54
55/*
56 * Zstd Workspace Management
57 *
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
61 * workspace.
62 *
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.
69 *
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).
73 */
74
75struct zstd_workspace_manager {
76 const struct btrfs_compress_op *ops;
77 spinlock_t lock;
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;
83};
84
85static struct zstd_workspace_manager wsm;
86
87static size_t zstd_ws_mem_sizes[ZSTD_BTRFS_MAX_LEVEL];
88
89static inline struct workspace *list_to_workspace(struct list_head *list)
90{
91 return container_of(list, struct workspace, list);
92}
93
94void zstd_free_workspace(struct list_head *ws);
95struct list_head *zstd_alloc_workspace(unsigned int level);
96
97/*
98 * Timer callback to free unused workspaces.
99 *
100 * @t: timer
101 *
102 * This scans the lru_list and attempts to reclaim any workspace that hasn't
103 * been used for ZSTD_BTRFS_RECLAIM_JIFFIES.
104 *
105 * The context is softirq and does not need the _bh locking primitives.
106 */
107static void zstd_reclaim_timer_fn(struct timer_list *timer)
108{
109 unsigned long reclaim_threshold = jiffies - ZSTD_BTRFS_RECLAIM_JIFFIES;
110 struct list_head *pos, *next;
111
112 spin_lock(&wsm.lock);
113
114 if (list_empty(&wsm.lru_list)) {
115 spin_unlock(&wsm.lock);
116 return;
117 }
118
119 list_for_each_prev_safe(pos, next, &wsm.lru_list) {
120 struct workspace *victim = container_of(pos, struct workspace,
121 lru_list);
122 unsigned int level;
123
124 if (time_after(victim->last_used, reclaim_threshold))
125 break;
126
127 /* workspace is in use */
128 if (victim->req_level)
129 continue;
130
131 level = victim->level;
132 list_del(&victim->lru_list);
133 list_del(&victim->list);
134 zstd_free_workspace(&victim->list);
135
136 if (list_empty(&wsm.idle_ws[level - 1]))
137 clear_bit(level - 1, &wsm.active_map);
138
139 }
140
141 if (!list_empty(&wsm.lru_list))
142 mod_timer(&wsm.timer, jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
143
144 spin_unlock(&wsm.lock);
145}
146
147/*
148 * Calculate monotonic memory bounds.
149 *
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.
155 */
156static void zstd_calc_ws_mem_sizes(void)
157{
158 size_t max_size = 0;
159 unsigned int level;
160
161 for (level = 1; level <= ZSTD_BTRFS_MAX_LEVEL; level++) {
162 zstd_parameters params =
163 zstd_get_btrfs_parameters(level, ZSTD_BTRFS_MAX_INPUT);
164 size_t level_size =
165 max_t(size_t,
166 zstd_cstream_workspace_bound(¶ms.cParams),
167 zstd_dstream_workspace_bound(ZSTD_BTRFS_MAX_INPUT));
168
169 max_size = max_t(size_t, max_size, level_size);
170 zstd_ws_mem_sizes[level - 1] = max_size;
171 }
172}
173
174void zstd_init_workspace_manager(void)
175{
176 struct list_head *ws;
177 int i;
178
179 zstd_calc_ws_mem_sizes();
180
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);
185
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]);
189
190 ws = zstd_alloc_workspace(ZSTD_BTRFS_MAX_LEVEL);
191 if (IS_ERR(ws)) {
192 pr_warn(
193 "BTRFS: cannot preallocate zstd compression workspace\n");
194 } else {
195 set_bit(ZSTD_BTRFS_MAX_LEVEL - 1, &wsm.active_map);
196 list_add(ws, &wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1]);
197 }
198}
199
200void zstd_cleanup_workspace_manager(void)
201{
202 struct workspace *workspace;
203 int i;
204
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);
213 }
214 }
215 spin_unlock_bh(&wsm.lock);
216
217 del_timer_sync(&wsm.timer);
218}
219
220/*
221 * Find workspace for given level.
222 *
223 * @level: compression level
224 *
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.
231 */
232static struct list_head *zstd_find_workspace(unsigned int level)
233{
234 struct list_head *ws;
235 struct workspace *workspace;
236 int i = level - 1;
237
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);
243 list_del_init(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);
251 return ws;
252 }
253 }
254 spin_unlock_bh(&wsm.lock);
255
256 return NULL;
257}
258
259/*
260 * Zstd get_workspace for level.
261 *
262 * @level: compression level
263 *
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.
268 */
269struct list_head *zstd_get_workspace(unsigned int level)
270{
271 struct list_head *ws;
272 unsigned int nofs_flag;
273
274 /* level == 0 means we can use any workspace */
275 if (!level)
276 level = 1;
277
278again:
279 ws = zstd_find_workspace(level);
280 if (ws)
281 return ws;
282
283 nofs_flag = memalloc_nofs_save();
284 ws = zstd_alloc_workspace(level);
285 memalloc_nofs_restore(nofs_flag);
286
287 if (IS_ERR(ws)) {
288 DEFINE_WAIT(wait);
289
290 prepare_to_wait(&wsm.wait, &wait, TASK_UNINTERRUPTIBLE);
291 schedule();
292 finish_wait(&wsm.wait, &wait);
293
294 goto again;
295 }
296
297 return ws;
298}
299
300/*
301 * Zstd put_workspace.
302 *
303 * @ws: list_head for the workspace
304 *
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.
310 */
311void zstd_put_workspace(struct list_head *ws)
312{
313 struct workspace *workspace = list_to_workspace(ws);
314
315 spin_lock_bh(&wsm.lock);
316
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);
322 } else {
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);
328 }
329 }
330
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;
334
335 spin_unlock_bh(&wsm.lock);
336
337 if (workspace->level == ZSTD_BTRFS_MAX_LEVEL)
338 cond_wake_up(&wsm.wait);
339}
340
341void zstd_free_workspace(struct list_head *ws)
342{
343 struct workspace *workspace = list_entry(ws, struct workspace, list);
344
345 kvfree(workspace->mem);
346 kfree(workspace->buf);
347 kfree(workspace);
348}
349
350struct list_head *zstd_alloc_workspace(unsigned int level)
351{
352 struct workspace *workspace;
353
354 workspace = kzalloc(sizeof(*workspace), GFP_KERNEL);
355 if (!workspace)
356 return ERR_PTR(-ENOMEM);
357
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)
365 goto fail;
366
367 INIT_LIST_HEAD(&workspace->list);
368 INIT_LIST_HEAD(&workspace->lru_list);
369
370 return &workspace->list;
371fail:
372 zstd_free_workspace(&workspace->list);
373 return ERR_PTR(-ENOMEM);
374}
375
376int 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)
379{
380 struct workspace *workspace = list_entry(ws, struct workspace, list);
381 zstd_cstream *stream;
382 int ret = 0;
383 int nr_pages = 0;
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,
392 len);
393
394 *out_pages = 0;
395 *total_out = 0;
396 *total_in = 0;
397
398 /* Initialize the stream */
399 stream = zstd_init_cstream(¶ms, len, workspace->mem,
400 workspace->size);
401 if (!stream) {
402 pr_warn("BTRFS: zstd_init_cstream failed\n");
403 ret = -EIO;
404 goto out;
405 }
406
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);
412
413 /* Allocate and map in the output buffer */
414 out_page = btrfs_alloc_compr_page();
415 if (out_page == NULL) {
416 ret = -ENOMEM;
417 goto out;
418 }
419 pages[nr_pages++] = out_page;
420 workspace->out_buf.dst = page_address(out_page);
421 workspace->out_buf.pos = 0;
422 workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
423
424 while (1) {
425 size_t ret2;
426
427 ret2 = zstd_compress_stream(stream, &workspace->out_buf,
428 &workspace->in_buf);
429 if (zstd_is_error(ret2)) {
430 pr_debug("BTRFS: zstd_compress_stream returned %d\n",
431 zstd_get_error_code(ret2));
432 ret = -EIO;
433 goto out;
434 }
435
436 /* Check to see if we are making it bigger */
437 if (tot_in + workspace->in_buf.pos > 8192 &&
438 tot_in + workspace->in_buf.pos <
439 tot_out + workspace->out_buf.pos) {
440 ret = -E2BIG;
441 goto out;
442 }
443
444 /* We've reached the end of our output range */
445 if (workspace->out_buf.pos >= max_out) {
446 tot_out += workspace->out_buf.pos;
447 ret = -E2BIG;
448 goto out;
449 }
450
451 /* Check if we need more output space */
452 if (workspace->out_buf.pos == workspace->out_buf.size) {
453 tot_out += PAGE_SIZE;
454 max_out -= PAGE_SIZE;
455 if (nr_pages == nr_dest_pages) {
456 ret = -E2BIG;
457 goto out;
458 }
459 out_page = btrfs_alloc_compr_page();
460 if (out_page == NULL) {
461 ret = -ENOMEM;
462 goto out;
463 }
464 pages[nr_pages++] = out_page;
465 workspace->out_buf.dst = page_address(out_page);
466 workspace->out_buf.pos = 0;
467 workspace->out_buf.size = min_t(size_t, max_out,
468 PAGE_SIZE);
469 }
470
471 /* We've reached the end of the input */
472 if (workspace->in_buf.pos >= len) {
473 tot_in += workspace->in_buf.pos;
474 break;
475 }
476
477 /* Check if we need more input */
478 if (workspace->in_buf.pos == workspace->in_buf.size) {
479 tot_in += PAGE_SIZE;
480 kunmap_local(workspace->in_buf.src);
481 put_page(in_page);
482 start += PAGE_SIZE;
483 len -= PAGE_SIZE;
484 in_page = find_get_page(mapping, start >> PAGE_SHIFT);
485 workspace->in_buf.src = kmap_local_page(in_page);
486 workspace->in_buf.pos = 0;
487 workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
488 }
489 }
490 while (1) {
491 size_t ret2;
492
493 ret2 = zstd_end_stream(stream, &workspace->out_buf);
494 if (zstd_is_error(ret2)) {
495 pr_debug("BTRFS: zstd_end_stream returned %d\n",
496 zstd_get_error_code(ret2));
497 ret = -EIO;
498 goto out;
499 }
500 if (ret2 == 0) {
501 tot_out += workspace->out_buf.pos;
502 break;
503 }
504 if (workspace->out_buf.pos >= max_out) {
505 tot_out += workspace->out_buf.pos;
506 ret = -E2BIG;
507 goto out;
508 }
509
510 tot_out += PAGE_SIZE;
511 max_out -= PAGE_SIZE;
512 if (nr_pages == nr_dest_pages) {
513 ret = -E2BIG;
514 goto out;
515 }
516 out_page = btrfs_alloc_compr_page();
517 if (out_page == NULL) {
518 ret = -ENOMEM;
519 goto out;
520 }
521 pages[nr_pages++] = out_page;
522 workspace->out_buf.dst = page_address(out_page);
523 workspace->out_buf.pos = 0;
524 workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
525 }
526
527 if (tot_out >= tot_in) {
528 ret = -E2BIG;
529 goto out;
530 }
531
532 ret = 0;
533 *total_in = tot_in;
534 *total_out = tot_out;
535out:
536 *out_pages = nr_pages;
537 if (workspace->in_buf.src) {
538 kunmap_local(workspace->in_buf.src);
539 put_page(in_page);
540 }
541 return ret;
542}
543
544int zstd_decompress_bio(struct list_head *ws, struct compressed_bio *cb)
545{
546 struct workspace *workspace = list_entry(ws, struct workspace, list);
547 struct page **pages_in = cb->compressed_pages;
548 size_t srclen = cb->compressed_len;
549 zstd_dstream *stream;
550 int ret = 0;
551 unsigned long page_in_index = 0;
552 unsigned long total_pages_in = DIV_ROUND_UP(srclen, PAGE_SIZE);
553 unsigned long buf_start;
554 unsigned long total_out = 0;
555
556 stream = zstd_init_dstream(
557 ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
558 if (!stream) {
559 pr_debug("BTRFS: zstd_init_dstream failed\n");
560 ret = -EIO;
561 goto done;
562 }
563
564 workspace->in_buf.src = kmap_local_page(pages_in[page_in_index]);
565 workspace->in_buf.pos = 0;
566 workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
567
568 workspace->out_buf.dst = workspace->buf;
569 workspace->out_buf.pos = 0;
570 workspace->out_buf.size = PAGE_SIZE;
571
572 while (1) {
573 size_t ret2;
574
575 ret2 = zstd_decompress_stream(stream, &workspace->out_buf,
576 &workspace->in_buf);
577 if (zstd_is_error(ret2)) {
578 pr_debug("BTRFS: zstd_decompress_stream returned %d\n",
579 zstd_get_error_code(ret2));
580 ret = -EIO;
581 goto done;
582 }
583 buf_start = total_out;
584 total_out += workspace->out_buf.pos;
585 workspace->out_buf.pos = 0;
586
587 ret = btrfs_decompress_buf2page(workspace->out_buf.dst,
588 total_out - buf_start, cb, buf_start);
589 if (ret == 0)
590 break;
591
592 if (workspace->in_buf.pos >= srclen)
593 break;
594
595 /* Check if we've hit the end of a frame */
596 if (ret2 == 0)
597 break;
598
599 if (workspace->in_buf.pos == workspace->in_buf.size) {
600 kunmap_local(workspace->in_buf.src);
601 page_in_index++;
602 if (page_in_index >= total_pages_in) {
603 workspace->in_buf.src = NULL;
604 ret = -EIO;
605 goto done;
606 }
607 srclen -= PAGE_SIZE;
608 workspace->in_buf.src = kmap_local_page(pages_in[page_in_index]);
609 workspace->in_buf.pos = 0;
610 workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
611 }
612 }
613 ret = 0;
614done:
615 if (workspace->in_buf.src)
616 kunmap_local(workspace->in_buf.src);
617 return ret;
618}
619
620int zstd_decompress(struct list_head *ws, const u8 *data_in,
621 struct page *dest_page, unsigned long start_byte, size_t srclen,
622 size_t destlen)
623{
624 struct workspace *workspace = list_entry(ws, struct workspace, list);
625 zstd_dstream *stream;
626 int ret = 0;
627 size_t ret2;
628 unsigned long total_out = 0;
629 unsigned long pg_offset = 0;
630
631 stream = zstd_init_dstream(
632 ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
633 if (!stream) {
634 pr_warn("BTRFS: zstd_init_dstream failed\n");
635 ret = -EIO;
636 goto finish;
637 }
638
639 destlen = min_t(size_t, destlen, PAGE_SIZE);
640
641 workspace->in_buf.src = data_in;
642 workspace->in_buf.pos = 0;
643 workspace->in_buf.size = srclen;
644
645 workspace->out_buf.dst = workspace->buf;
646 workspace->out_buf.pos = 0;
647 workspace->out_buf.size = PAGE_SIZE;
648
649 ret2 = 1;
650 while (pg_offset < destlen
651 && workspace->in_buf.pos < workspace->in_buf.size) {
652 unsigned long buf_start;
653 unsigned long buf_offset;
654 unsigned long bytes;
655
656 /* Check if the frame is over and we still need more input */
657 if (ret2 == 0) {
658 pr_debug("BTRFS: zstd_decompress_stream ended early\n");
659 ret = -EIO;
660 goto finish;
661 }
662 ret2 = zstd_decompress_stream(stream, &workspace->out_buf,
663 &workspace->in_buf);
664 if (zstd_is_error(ret2)) {
665 pr_debug("BTRFS: zstd_decompress_stream returned %d\n",
666 zstd_get_error_code(ret2));
667 ret = -EIO;
668 goto finish;
669 }
670
671 buf_start = total_out;
672 total_out += workspace->out_buf.pos;
673 workspace->out_buf.pos = 0;
674
675 if (total_out <= start_byte)
676 continue;
677
678 if (total_out > start_byte && buf_start < start_byte)
679 buf_offset = start_byte - buf_start;
680 else
681 buf_offset = 0;
682
683 bytes = min_t(unsigned long, destlen - pg_offset,
684 workspace->out_buf.size - buf_offset);
685
686 memcpy_to_page(dest_page, pg_offset,
687 workspace->out_buf.dst + buf_offset, bytes);
688
689 pg_offset += bytes;
690 }
691 ret = 0;
692finish:
693 if (pg_offset < destlen) {
694 memzero_page(dest_page, pg_offset, destlen - pg_offset);
695 }
696 return ret;
697}
698
699const struct btrfs_compress_op btrfs_zstd_compress = {
700 /* ZSTD uses own workspace manager */
701 .workspace_manager = NULL,
702 .max_level = ZSTD_BTRFS_MAX_LEVEL,
703 .default_level = ZSTD_BTRFS_DEFAULT_LEVEL,
704};
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2016-present, Facebook, Inc.
4 * All rights reserved.
5 *
6 */
7
8#include <linux/bio.h>
9#include <linux/bitmap.h>
10#include <linux/err.h>
11#include <linux/init.h>
12#include <linux/kernel.h>
13#include <linux/mm.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>
20#include "misc.h"
21#include "fs.h"
22#include "compression.h"
23#include "super.h"
24
25#define ZSTD_BTRFS_MAX_WINDOWLOG 17
26#define ZSTD_BTRFS_MAX_INPUT (1 << ZSTD_BTRFS_MAX_WINDOWLOG)
27#define ZSTD_BTRFS_DEFAULT_LEVEL 3
28#define ZSTD_BTRFS_MAX_LEVEL 15
29/* 307s to avoid pathologically clashing with transaction commit */
30#define ZSTD_BTRFS_RECLAIM_JIFFIES (307 * HZ)
31
32static zstd_parameters zstd_get_btrfs_parameters(unsigned int level,
33 size_t src_len)
34{
35 zstd_parameters params = zstd_get_params(level, src_len);
36
37 if (params.cParams.windowLog > ZSTD_BTRFS_MAX_WINDOWLOG)
38 params.cParams.windowLog = ZSTD_BTRFS_MAX_WINDOWLOG;
39 WARN_ON(src_len > ZSTD_BTRFS_MAX_INPUT);
40 return params;
41}
42
43struct workspace {
44 void *mem;
45 size_t size;
46 char *buf;
47 unsigned int level;
48 unsigned int req_level;
49 unsigned long last_used; /* jiffies */
50 struct list_head list;
51 struct list_head lru_list;
52 zstd_in_buffer in_buf;
53 zstd_out_buffer out_buf;
54};
55
56/*
57 * Zstd Workspace Management
58 *
59 * Zstd workspaces have different memory requirements depending on the level.
60 * The zstd workspaces are managed by having individual lists for each level
61 * and a global lru. Forward progress is maintained by protecting a max level
62 * workspace.
63 *
64 * Getting a workspace is done by using the bitmap to identify the levels that
65 * have available workspaces and scans up. This lets us recycle higher level
66 * workspaces because of the monotonic memory guarantee. A workspace's
67 * last_used is only updated if it is being used by the corresponding memory
68 * level. Putting a workspace involves adding it back to the appropriate places
69 * and adding it back to the lru if necessary.
70 *
71 * A timer is used to reclaim workspaces if they have not been used for
72 * ZSTD_BTRFS_RECLAIM_JIFFIES. This helps keep only active workspaces around.
73 * The upper bound is provided by the workqueue limit which is 2 (percpu limit).
74 */
75
76struct zstd_workspace_manager {
77 const struct btrfs_compress_op *ops;
78 spinlock_t lock;
79 struct list_head lru_list;
80 struct list_head idle_ws[ZSTD_BTRFS_MAX_LEVEL];
81 unsigned long active_map;
82 wait_queue_head_t wait;
83 struct timer_list timer;
84};
85
86static struct zstd_workspace_manager wsm;
87
88static size_t zstd_ws_mem_sizes[ZSTD_BTRFS_MAX_LEVEL];
89
90static inline struct workspace *list_to_workspace(struct list_head *list)
91{
92 return container_of(list, struct workspace, list);
93}
94
95void zstd_free_workspace(struct list_head *ws);
96struct list_head *zstd_alloc_workspace(unsigned int level);
97
98/*
99 * Timer callback to free unused workspaces.
100 *
101 * @t: timer
102 *
103 * This scans the lru_list and attempts to reclaim any workspace that hasn't
104 * been used for ZSTD_BTRFS_RECLAIM_JIFFIES.
105 *
106 * The context is softirq and does not need the _bh locking primitives.
107 */
108static void zstd_reclaim_timer_fn(struct timer_list *timer)
109{
110 unsigned long reclaim_threshold = jiffies - ZSTD_BTRFS_RECLAIM_JIFFIES;
111 struct list_head *pos, *next;
112
113 spin_lock(&wsm.lock);
114
115 if (list_empty(&wsm.lru_list)) {
116 spin_unlock(&wsm.lock);
117 return;
118 }
119
120 list_for_each_prev_safe(pos, next, &wsm.lru_list) {
121 struct workspace *victim = container_of(pos, struct workspace,
122 lru_list);
123 unsigned int level;
124
125 if (time_after(victim->last_used, reclaim_threshold))
126 break;
127
128 /* workspace is in use */
129 if (victim->req_level)
130 continue;
131
132 level = victim->level;
133 list_del(&victim->lru_list);
134 list_del(&victim->list);
135 zstd_free_workspace(&victim->list);
136
137 if (list_empty(&wsm.idle_ws[level - 1]))
138 clear_bit(level - 1, &wsm.active_map);
139
140 }
141
142 if (!list_empty(&wsm.lru_list))
143 mod_timer(&wsm.timer, jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
144
145 spin_unlock(&wsm.lock);
146}
147
148/*
149 * Calculate monotonic memory bounds.
150 *
151 * It is possible based on the level configurations that a higher level
152 * workspace uses less memory than a lower level workspace. In order to reuse
153 * workspaces, this must be made a monotonic relationship. This precomputes
154 * the required memory for each level and enforces the monotonicity between
155 * level and memory required.
156 */
157static void zstd_calc_ws_mem_sizes(void)
158{
159 size_t max_size = 0;
160 unsigned int level;
161
162 for (level = 1; level <= ZSTD_BTRFS_MAX_LEVEL; level++) {
163 zstd_parameters params =
164 zstd_get_btrfs_parameters(level, ZSTD_BTRFS_MAX_INPUT);
165 size_t level_size =
166 max_t(size_t,
167 zstd_cstream_workspace_bound(¶ms.cParams),
168 zstd_dstream_workspace_bound(ZSTD_BTRFS_MAX_INPUT));
169
170 max_size = max_t(size_t, max_size, level_size);
171 zstd_ws_mem_sizes[level - 1] = max_size;
172 }
173}
174
175void zstd_init_workspace_manager(void)
176{
177 struct list_head *ws;
178 int i;
179
180 zstd_calc_ws_mem_sizes();
181
182 wsm.ops = &btrfs_zstd_compress;
183 spin_lock_init(&wsm.lock);
184 init_waitqueue_head(&wsm.wait);
185 timer_setup(&wsm.timer, zstd_reclaim_timer_fn, 0);
186
187 INIT_LIST_HEAD(&wsm.lru_list);
188 for (i = 0; i < ZSTD_BTRFS_MAX_LEVEL; i++)
189 INIT_LIST_HEAD(&wsm.idle_ws[i]);
190
191 ws = zstd_alloc_workspace(ZSTD_BTRFS_MAX_LEVEL);
192 if (IS_ERR(ws)) {
193 pr_warn(
194 "BTRFS: cannot preallocate zstd compression workspace\n");
195 } else {
196 set_bit(ZSTD_BTRFS_MAX_LEVEL - 1, &wsm.active_map);
197 list_add(ws, &wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1]);
198 }
199}
200
201void zstd_cleanup_workspace_manager(void)
202{
203 struct workspace *workspace;
204 int i;
205
206 spin_lock_bh(&wsm.lock);
207 for (i = 0; i < ZSTD_BTRFS_MAX_LEVEL; i++) {
208 while (!list_empty(&wsm.idle_ws[i])) {
209 workspace = container_of(wsm.idle_ws[i].next,
210 struct workspace, list);
211 list_del(&workspace->list);
212 list_del(&workspace->lru_list);
213 zstd_free_workspace(&workspace->list);
214 }
215 }
216 spin_unlock_bh(&wsm.lock);
217
218 del_timer_sync(&wsm.timer);
219}
220
221/*
222 * Find workspace for given level.
223 *
224 * @level: compression level
225 *
226 * This iterates over the set bits in the active_map beginning at the requested
227 * compression level. This lets us utilize already allocated workspaces before
228 * allocating a new one. If the workspace is of a larger size, it is used, but
229 * the place in the lru_list and last_used times are not updated. This is to
230 * offer the opportunity to reclaim the workspace in favor of allocating an
231 * appropriately sized one in the future.
232 */
233static struct list_head *zstd_find_workspace(unsigned int level)
234{
235 struct list_head *ws;
236 struct workspace *workspace;
237 int i = level - 1;
238
239 spin_lock_bh(&wsm.lock);
240 for_each_set_bit_from(i, &wsm.active_map, ZSTD_BTRFS_MAX_LEVEL) {
241 if (!list_empty(&wsm.idle_ws[i])) {
242 ws = wsm.idle_ws[i].next;
243 workspace = list_to_workspace(ws);
244 list_del_init(ws);
245 /* keep its place if it's a lower level using this */
246 workspace->req_level = level;
247 if (level == workspace->level)
248 list_del(&workspace->lru_list);
249 if (list_empty(&wsm.idle_ws[i]))
250 clear_bit(i, &wsm.active_map);
251 spin_unlock_bh(&wsm.lock);
252 return ws;
253 }
254 }
255 spin_unlock_bh(&wsm.lock);
256
257 return NULL;
258}
259
260/*
261 * Zstd get_workspace for level.
262 *
263 * @level: compression level
264 *
265 * If @level is 0, then any compression level can be used. Therefore, we begin
266 * scanning from 1. We first scan through possible workspaces and then after
267 * attempt to allocate a new workspace. If we fail to allocate one due to
268 * memory pressure, go to sleep waiting for the max level workspace to free up.
269 */
270struct list_head *zstd_get_workspace(unsigned int level)
271{
272 struct list_head *ws;
273 unsigned int nofs_flag;
274
275 /* level == 0 means we can use any workspace */
276 if (!level)
277 level = 1;
278
279again:
280 ws = zstd_find_workspace(level);
281 if (ws)
282 return ws;
283
284 nofs_flag = memalloc_nofs_save();
285 ws = zstd_alloc_workspace(level);
286 memalloc_nofs_restore(nofs_flag);
287
288 if (IS_ERR(ws)) {
289 DEFINE_WAIT(wait);
290
291 prepare_to_wait(&wsm.wait, &wait, TASK_UNINTERRUPTIBLE);
292 schedule();
293 finish_wait(&wsm.wait, &wait);
294
295 goto again;
296 }
297
298 return ws;
299}
300
301/*
302 * Zstd put_workspace.
303 *
304 * @ws: list_head for the workspace
305 *
306 * When putting back a workspace, we only need to update the LRU if we are of
307 * the requested compression level. Here is where we continue to protect the
308 * max level workspace or update last_used accordingly. If the reclaim timer
309 * isn't set, it is also set here. Only the max level workspace tries and wakes
310 * up waiting workspaces.
311 */
312void zstd_put_workspace(struct list_head *ws)
313{
314 struct workspace *workspace = list_to_workspace(ws);
315
316 spin_lock_bh(&wsm.lock);
317
318 /* A node is only taken off the lru if we are the corresponding level */
319 if (workspace->req_level == workspace->level) {
320 /* Hide a max level workspace from reclaim */
321 if (list_empty(&wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1])) {
322 INIT_LIST_HEAD(&workspace->lru_list);
323 } else {
324 workspace->last_used = jiffies;
325 list_add(&workspace->lru_list, &wsm.lru_list);
326 if (!timer_pending(&wsm.timer))
327 mod_timer(&wsm.timer,
328 jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
329 }
330 }
331
332 set_bit(workspace->level - 1, &wsm.active_map);
333 list_add(&workspace->list, &wsm.idle_ws[workspace->level - 1]);
334 workspace->req_level = 0;
335
336 spin_unlock_bh(&wsm.lock);
337
338 if (workspace->level == ZSTD_BTRFS_MAX_LEVEL)
339 cond_wake_up(&wsm.wait);
340}
341
342void zstd_free_workspace(struct list_head *ws)
343{
344 struct workspace *workspace = list_entry(ws, struct workspace, list);
345
346 kvfree(workspace->mem);
347 kfree(workspace->buf);
348 kfree(workspace);
349}
350
351struct list_head *zstd_alloc_workspace(unsigned int level)
352{
353 struct workspace *workspace;
354
355 workspace = kzalloc(sizeof(*workspace), GFP_KERNEL);
356 if (!workspace)
357 return ERR_PTR(-ENOMEM);
358
359 workspace->size = zstd_ws_mem_sizes[level - 1];
360 workspace->level = level;
361 workspace->req_level = level;
362 workspace->last_used = jiffies;
363 workspace->mem = kvmalloc(workspace->size, GFP_KERNEL | __GFP_NOWARN);
364 workspace->buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
365 if (!workspace->mem || !workspace->buf)
366 goto fail;
367
368 INIT_LIST_HEAD(&workspace->list);
369 INIT_LIST_HEAD(&workspace->lru_list);
370
371 return &workspace->list;
372fail:
373 zstd_free_workspace(&workspace->list);
374 return ERR_PTR(-ENOMEM);
375}
376
377int zstd_compress_pages(struct list_head *ws, struct address_space *mapping,
378 u64 start, struct page **pages, unsigned long *out_pages,
379 unsigned long *total_in, unsigned long *total_out)
380{
381 struct workspace *workspace = list_entry(ws, struct workspace, list);
382 zstd_cstream *stream;
383 int ret = 0;
384 int nr_pages = 0;
385 struct page *in_page = NULL; /* The current page to read */
386 struct page *out_page = NULL; /* The current page to write to */
387 unsigned long tot_in = 0;
388 unsigned long tot_out = 0;
389 unsigned long len = *total_out;
390 const unsigned long nr_dest_pages = *out_pages;
391 unsigned long max_out = nr_dest_pages * PAGE_SIZE;
392 zstd_parameters params = zstd_get_btrfs_parameters(workspace->req_level,
393 len);
394
395 *out_pages = 0;
396 *total_out = 0;
397 *total_in = 0;
398
399 /* Initialize the stream */
400 stream = zstd_init_cstream(¶ms, len, workspace->mem,
401 workspace->size);
402 if (!stream) {
403 pr_warn("BTRFS: zstd_init_cstream failed\n");
404 ret = -EIO;
405 goto out;
406 }
407
408 /* map in the first page of input data */
409 in_page = find_get_page(mapping, start >> PAGE_SHIFT);
410 workspace->in_buf.src = kmap_local_page(in_page);
411 workspace->in_buf.pos = 0;
412 workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
413
414 /* Allocate and map in the output buffer */
415 out_page = btrfs_alloc_compr_page();
416 if (out_page == NULL) {
417 ret = -ENOMEM;
418 goto out;
419 }
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);
424
425 while (1) {
426 size_t ret2;
427
428 ret2 = zstd_compress_stream(stream, &workspace->out_buf,
429 &workspace->in_buf);
430 if (zstd_is_error(ret2)) {
431 pr_debug("BTRFS: zstd_compress_stream returned %d\n",
432 zstd_get_error_code(ret2));
433 ret = -EIO;
434 goto out;
435 }
436
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) {
441 ret = -E2BIG;
442 goto out;
443 }
444
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;
448 ret = -E2BIG;
449 goto out;
450 }
451
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) {
457 ret = -E2BIG;
458 goto out;
459 }
460 out_page = btrfs_alloc_compr_page();
461 if (out_page == NULL) {
462 ret = -ENOMEM;
463 goto out;
464 }
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,
469 PAGE_SIZE);
470 }
471
472 /* We've reached the end of the input */
473 if (workspace->in_buf.pos >= len) {
474 tot_in += workspace->in_buf.pos;
475 break;
476 }
477
478 /* Check if we need more input */
479 if (workspace->in_buf.pos == workspace->in_buf.size) {
480 tot_in += PAGE_SIZE;
481 kunmap_local(workspace->in_buf.src);
482 put_page(in_page);
483 start += PAGE_SIZE;
484 len -= PAGE_SIZE;
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);
489 }
490 }
491 while (1) {
492 size_t ret2;
493
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));
498 ret = -EIO;
499 goto out;
500 }
501 if (ret2 == 0) {
502 tot_out += workspace->out_buf.pos;
503 break;
504 }
505 if (workspace->out_buf.pos >= max_out) {
506 tot_out += workspace->out_buf.pos;
507 ret = -E2BIG;
508 goto out;
509 }
510
511 tot_out += PAGE_SIZE;
512 max_out -= PAGE_SIZE;
513 if (nr_pages == nr_dest_pages) {
514 ret = -E2BIG;
515 goto out;
516 }
517 out_page = btrfs_alloc_compr_page();
518 if (out_page == NULL) {
519 ret = -ENOMEM;
520 goto out;
521 }
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);
526 }
527
528 if (tot_out >= tot_in) {
529 ret = -E2BIG;
530 goto out;
531 }
532
533 ret = 0;
534 *total_in = tot_in;
535 *total_out = tot_out;
536out:
537 *out_pages = nr_pages;
538 if (workspace->in_buf.src) {
539 kunmap_local(workspace->in_buf.src);
540 put_page(in_page);
541 }
542 return ret;
543}
544
545int zstd_decompress_bio(struct list_head *ws, struct compressed_bio *cb)
546{
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;
551 int ret = 0;
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;
556
557 stream = zstd_init_dstream(
558 ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
559 if (!stream) {
560 pr_debug("BTRFS: zstd_init_dstream failed\n");
561 ret = -EIO;
562 goto done;
563 }
564
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);
568
569 workspace->out_buf.dst = workspace->buf;
570 workspace->out_buf.pos = 0;
571 workspace->out_buf.size = PAGE_SIZE;
572
573 while (1) {
574 size_t ret2;
575
576 ret2 = zstd_decompress_stream(stream, &workspace->out_buf,
577 &workspace->in_buf);
578 if (zstd_is_error(ret2)) {
579 pr_debug("BTRFS: zstd_decompress_stream returned %d\n",
580 zstd_get_error_code(ret2));
581 ret = -EIO;
582 goto done;
583 }
584 buf_start = total_out;
585 total_out += workspace->out_buf.pos;
586 workspace->out_buf.pos = 0;
587
588 ret = btrfs_decompress_buf2page(workspace->out_buf.dst,
589 total_out - buf_start, cb, buf_start);
590 if (ret == 0)
591 break;
592
593 if (workspace->in_buf.pos >= srclen)
594 break;
595
596 /* Check if we've hit the end of a frame */
597 if (ret2 == 0)
598 break;
599
600 if (workspace->in_buf.pos == workspace->in_buf.size) {
601 kunmap_local(workspace->in_buf.src);
602 page_in_index++;
603 if (page_in_index >= total_pages_in) {
604 workspace->in_buf.src = NULL;
605 ret = -EIO;
606 goto done;
607 }
608 srclen -= PAGE_SIZE;
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);
612 }
613 }
614 ret = 0;
615done:
616 if (workspace->in_buf.src)
617 kunmap_local(workspace->in_buf.src);
618 return ret;
619}
620
621int zstd_decompress(struct list_head *ws, const u8 *data_in,
622 struct page *dest_page, unsigned long dest_pgoff, size_t srclen,
623 size_t destlen)
624{
625 struct workspace *workspace = list_entry(ws, struct workspace, list);
626 struct btrfs_fs_info *fs_info = btrfs_sb(dest_page->mapping->host->i_sb);
627 const u32 sectorsize = fs_info->sectorsize;
628 zstd_dstream *stream;
629 int ret = 0;
630 unsigned long to_copy = 0;
631
632 stream = zstd_init_dstream(
633 ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
634 if (!stream) {
635 pr_warn("BTRFS: zstd_init_dstream failed\n");
636 goto finish;
637 }
638
639 workspace->in_buf.src = data_in;
640 workspace->in_buf.pos = 0;
641 workspace->in_buf.size = srclen;
642
643 workspace->out_buf.dst = workspace->buf;
644 workspace->out_buf.pos = 0;
645 workspace->out_buf.size = sectorsize;
646
647 /*
648 * Since both input and output buffers should not exceed one sector,
649 * one call should end the decompression.
650 */
651 ret = zstd_decompress_stream(stream, &workspace->out_buf, &workspace->in_buf);
652 if (zstd_is_error(ret)) {
653 pr_warn_ratelimited("BTRFS: zstd_decompress_stream return %d\n",
654 zstd_get_error_code(ret));
655 goto finish;
656 }
657 to_copy = workspace->out_buf.pos;
658 memcpy_to_page(dest_page, dest_pgoff, workspace->out_buf.dst, to_copy);
659finish:
660 /* Error or early end. */
661 if (unlikely(to_copy < destlen)) {
662 ret = -EIO;
663 memzero_page(dest_page, dest_pgoff + to_copy, destlen - to_copy);
664 }
665 return ret;
666}
667
668const struct btrfs_compress_op btrfs_zstd_compress = {
669 /* ZSTD uses own workspace manager */
670 .workspace_manager = NULL,
671 .max_level = ZSTD_BTRFS_MAX_LEVEL,
672 .default_level = ZSTD_BTRFS_DEFAULT_LEVEL,
673};