xref: /openbmc/linux/fs/btrfs/zstd.c (revision bef7a78d)
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 
31 static ZSTD_parameters zstd_get_btrfs_parameters(unsigned int level,
32 						 size_t src_len)
33 {
34 	ZSTD_parameters params = ZSTD_getParams(level, src_len, 0);
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 
42 struct 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_inBuffer in_buf;
52 	ZSTD_outBuffer 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 
75 struct 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 
85 static struct zstd_workspace_manager wsm;
86 
87 static size_t zstd_ws_mem_sizes[ZSTD_BTRFS_MAX_LEVEL];
88 
89 static inline struct workspace *list_to_workspace(struct list_head *list)
90 {
91 	return container_of(list, struct workspace, list);
92 }
93 
94 void zstd_free_workspace(struct list_head *ws);
95 struct list_head *zstd_alloc_workspace(unsigned int level);
96 /*
97  * zstd_reclaim_timer_fn - reclaim timer
98  * @t: timer
99  *
100  * This scans the lru_list and attempts to reclaim any workspace that hasn't
101  * been used for ZSTD_BTRFS_RECLAIM_JIFFIES.
102  */
103 static void zstd_reclaim_timer_fn(struct timer_list *timer)
104 {
105 	unsigned long reclaim_threshold = jiffies - ZSTD_BTRFS_RECLAIM_JIFFIES;
106 	struct list_head *pos, *next;
107 
108 	spin_lock_bh(&wsm.lock);
109 
110 	if (list_empty(&wsm.lru_list)) {
111 		spin_unlock_bh(&wsm.lock);
112 		return;
113 	}
114 
115 	list_for_each_prev_safe(pos, next, &wsm.lru_list) {
116 		struct workspace *victim = container_of(pos, struct workspace,
117 							lru_list);
118 		unsigned int level;
119 
120 		if (time_after(victim->last_used, reclaim_threshold))
121 			break;
122 
123 		/* workspace is in use */
124 		if (victim->req_level)
125 			continue;
126 
127 		level = victim->level;
128 		list_del(&victim->lru_list);
129 		list_del(&victim->list);
130 		zstd_free_workspace(&victim->list);
131 
132 		if (list_empty(&wsm.idle_ws[level - 1]))
133 			clear_bit(level - 1, &wsm.active_map);
134 
135 	}
136 
137 	if (!list_empty(&wsm.lru_list))
138 		mod_timer(&wsm.timer, jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
139 
140 	spin_unlock_bh(&wsm.lock);
141 }
142 
143 /*
144  * zstd_calc_ws_mem_sizes - calculate monotonic memory bounds
145  *
146  * It is possible based on the level configurations that a higher level
147  * workspace uses less memory than a lower level workspace.  In order to reuse
148  * workspaces, this must be made a monotonic relationship.  This precomputes
149  * the required memory for each level and enforces the monotonicity between
150  * level and memory required.
151  */
152 static void zstd_calc_ws_mem_sizes(void)
153 {
154 	size_t max_size = 0;
155 	unsigned int level;
156 
157 	for (level = 1; level <= ZSTD_BTRFS_MAX_LEVEL; level++) {
158 		ZSTD_parameters params =
159 			zstd_get_btrfs_parameters(level, ZSTD_BTRFS_MAX_INPUT);
160 		size_t level_size =
161 			max_t(size_t,
162 			      ZSTD_CStreamWorkspaceBound(params.cParams),
163 			      ZSTD_DStreamWorkspaceBound(ZSTD_BTRFS_MAX_INPUT));
164 
165 		max_size = max_t(size_t, max_size, level_size);
166 		zstd_ws_mem_sizes[level - 1] = max_size;
167 	}
168 }
169 
170 void zstd_init_workspace_manager(void)
171 {
172 	struct list_head *ws;
173 	int i;
174 
175 	zstd_calc_ws_mem_sizes();
176 
177 	wsm.ops = &btrfs_zstd_compress;
178 	spin_lock_init(&wsm.lock);
179 	init_waitqueue_head(&wsm.wait);
180 	timer_setup(&wsm.timer, zstd_reclaim_timer_fn, 0);
181 
182 	INIT_LIST_HEAD(&wsm.lru_list);
183 	for (i = 0; i < ZSTD_BTRFS_MAX_LEVEL; i++)
184 		INIT_LIST_HEAD(&wsm.idle_ws[i]);
185 
186 	ws = zstd_alloc_workspace(ZSTD_BTRFS_MAX_LEVEL);
187 	if (IS_ERR(ws)) {
188 		pr_warn(
189 		"BTRFS: cannot preallocate zstd compression workspace\n");
190 	} else {
191 		set_bit(ZSTD_BTRFS_MAX_LEVEL - 1, &wsm.active_map);
192 		list_add(ws, &wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1]);
193 	}
194 }
195 
196 void zstd_cleanup_workspace_manager(void)
197 {
198 	struct workspace *workspace;
199 	int i;
200 
201 	spin_lock_bh(&wsm.lock);
202 	for (i = 0; i < ZSTD_BTRFS_MAX_LEVEL; i++) {
203 		while (!list_empty(&wsm.idle_ws[i])) {
204 			workspace = container_of(wsm.idle_ws[i].next,
205 						 struct workspace, list);
206 			list_del(&workspace->list);
207 			list_del(&workspace->lru_list);
208 			zstd_free_workspace(&workspace->list);
209 		}
210 	}
211 	spin_unlock_bh(&wsm.lock);
212 
213 	del_timer_sync(&wsm.timer);
214 }
215 
216 /*
217  * zstd_find_workspace - find workspace
218  * @level: compression level
219  *
220  * This iterates over the set bits in the active_map beginning at the requested
221  * compression level.  This lets us utilize already allocated workspaces before
222  * allocating a new one.  If the workspace is of a larger size, it is used, but
223  * the place in the lru_list and last_used times are not updated.  This is to
224  * offer the opportunity to reclaim the workspace in favor of allocating an
225  * appropriately sized one in the future.
226  */
227 static struct list_head *zstd_find_workspace(unsigned int level)
228 {
229 	struct list_head *ws;
230 	struct workspace *workspace;
231 	int i = level - 1;
232 
233 	spin_lock_bh(&wsm.lock);
234 	for_each_set_bit_from(i, &wsm.active_map, ZSTD_BTRFS_MAX_LEVEL) {
235 		if (!list_empty(&wsm.idle_ws[i])) {
236 			ws = wsm.idle_ws[i].next;
237 			workspace = list_to_workspace(ws);
238 			list_del_init(ws);
239 			/* keep its place if it's a lower level using this */
240 			workspace->req_level = level;
241 			if (level == workspace->level)
242 				list_del(&workspace->lru_list);
243 			if (list_empty(&wsm.idle_ws[i]))
244 				clear_bit(i, &wsm.active_map);
245 			spin_unlock_bh(&wsm.lock);
246 			return ws;
247 		}
248 	}
249 	spin_unlock_bh(&wsm.lock);
250 
251 	return NULL;
252 }
253 
254 /*
255  * zstd_get_workspace - zstd's get_workspace
256  * @level: compression level
257  *
258  * If @level is 0, then any compression level can be used.  Therefore, we begin
259  * scanning from 1.  We first scan through possible workspaces and then after
260  * attempt to allocate a new workspace.  If we fail to allocate one due to
261  * memory pressure, go to sleep waiting for the max level workspace to free up.
262  */
263 struct list_head *zstd_get_workspace(unsigned int level)
264 {
265 	struct list_head *ws;
266 	unsigned int nofs_flag;
267 
268 	/* level == 0 means we can use any workspace */
269 	if (!level)
270 		level = 1;
271 
272 again:
273 	ws = zstd_find_workspace(level);
274 	if (ws)
275 		return ws;
276 
277 	nofs_flag = memalloc_nofs_save();
278 	ws = zstd_alloc_workspace(level);
279 	memalloc_nofs_restore(nofs_flag);
280 
281 	if (IS_ERR(ws)) {
282 		DEFINE_WAIT(wait);
283 
284 		prepare_to_wait(&wsm.wait, &wait, TASK_UNINTERRUPTIBLE);
285 		schedule();
286 		finish_wait(&wsm.wait, &wait);
287 
288 		goto again;
289 	}
290 
291 	return ws;
292 }
293 
294 /*
295  * zstd_put_workspace - zstd put_workspace
296  * @ws: list_head for the workspace
297  *
298  * When putting back a workspace, we only need to update the LRU if we are of
299  * the requested compression level.  Here is where we continue to protect the
300  * max level workspace or update last_used accordingly.  If the reclaim timer
301  * isn't set, it is also set here.  Only the max level workspace tries and wakes
302  * up waiting workspaces.
303  */
304 void zstd_put_workspace(struct list_head *ws)
305 {
306 	struct workspace *workspace = list_to_workspace(ws);
307 
308 	spin_lock_bh(&wsm.lock);
309 
310 	/* A node is only taken off the lru if we are the corresponding level */
311 	if (workspace->req_level == workspace->level) {
312 		/* Hide a max level workspace from reclaim */
313 		if (list_empty(&wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1])) {
314 			INIT_LIST_HEAD(&workspace->lru_list);
315 		} else {
316 			workspace->last_used = jiffies;
317 			list_add(&workspace->lru_list, &wsm.lru_list);
318 			if (!timer_pending(&wsm.timer))
319 				mod_timer(&wsm.timer,
320 					  jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
321 		}
322 	}
323 
324 	set_bit(workspace->level - 1, &wsm.active_map);
325 	list_add(&workspace->list, &wsm.idle_ws[workspace->level - 1]);
326 	workspace->req_level = 0;
327 
328 	spin_unlock_bh(&wsm.lock);
329 
330 	if (workspace->level == ZSTD_BTRFS_MAX_LEVEL)
331 		cond_wake_up(&wsm.wait);
332 }
333 
334 void zstd_free_workspace(struct list_head *ws)
335 {
336 	struct workspace *workspace = list_entry(ws, struct workspace, list);
337 
338 	kvfree(workspace->mem);
339 	kfree(workspace->buf);
340 	kfree(workspace);
341 }
342 
343 struct list_head *zstd_alloc_workspace(unsigned int level)
344 {
345 	struct workspace *workspace;
346 
347 	workspace = kzalloc(sizeof(*workspace), GFP_KERNEL);
348 	if (!workspace)
349 		return ERR_PTR(-ENOMEM);
350 
351 	workspace->size = zstd_ws_mem_sizes[level - 1];
352 	workspace->level = level;
353 	workspace->req_level = level;
354 	workspace->last_used = jiffies;
355 	workspace->mem = kvmalloc(workspace->size, GFP_KERNEL);
356 	workspace->buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
357 	if (!workspace->mem || !workspace->buf)
358 		goto fail;
359 
360 	INIT_LIST_HEAD(&workspace->list);
361 	INIT_LIST_HEAD(&workspace->lru_list);
362 
363 	return &workspace->list;
364 fail:
365 	zstd_free_workspace(&workspace->list);
366 	return ERR_PTR(-ENOMEM);
367 }
368 
369 int zstd_compress_pages(struct list_head *ws, struct address_space *mapping,
370 		u64 start, struct page **pages, unsigned long *out_pages,
371 		unsigned long *total_in, unsigned long *total_out)
372 {
373 	struct workspace *workspace = list_entry(ws, struct workspace, list);
374 	ZSTD_CStream *stream;
375 	int ret = 0;
376 	int nr_pages = 0;
377 	struct page *in_page = NULL;  /* The current page to read */
378 	struct page *out_page = NULL; /* The current page to write to */
379 	unsigned long tot_in = 0;
380 	unsigned long tot_out = 0;
381 	unsigned long len = *total_out;
382 	const unsigned long nr_dest_pages = *out_pages;
383 	unsigned long max_out = nr_dest_pages * PAGE_SIZE;
384 	ZSTD_parameters params = zstd_get_btrfs_parameters(workspace->req_level,
385 							   len);
386 
387 	*out_pages = 0;
388 	*total_out = 0;
389 	*total_in = 0;
390 
391 	/* Initialize the stream */
392 	stream = ZSTD_initCStream(params, len, workspace->mem,
393 			workspace->size);
394 	if (!stream) {
395 		pr_warn("BTRFS: ZSTD_initCStream failed\n");
396 		ret = -EIO;
397 		goto out;
398 	}
399 
400 	/* map in the first page of input data */
401 	in_page = find_get_page(mapping, start >> PAGE_SHIFT);
402 	workspace->in_buf.src = kmap(in_page);
403 	workspace->in_buf.pos = 0;
404 	workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
405 
406 
407 	/* Allocate and map in the output buffer */
408 	out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
409 	if (out_page == NULL) {
410 		ret = -ENOMEM;
411 		goto out;
412 	}
413 	pages[nr_pages++] = out_page;
414 	workspace->out_buf.dst = kmap(out_page);
415 	workspace->out_buf.pos = 0;
416 	workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
417 
418 	while (1) {
419 		size_t ret2;
420 
421 		ret2 = ZSTD_compressStream(stream, &workspace->out_buf,
422 				&workspace->in_buf);
423 		if (ZSTD_isError(ret2)) {
424 			pr_debug("BTRFS: ZSTD_compressStream returned %d\n",
425 					ZSTD_getErrorCode(ret2));
426 			ret = -EIO;
427 			goto out;
428 		}
429 
430 		/* Check to see if we are making it bigger */
431 		if (tot_in + workspace->in_buf.pos > 8192 &&
432 				tot_in + workspace->in_buf.pos <
433 				tot_out + workspace->out_buf.pos) {
434 			ret = -E2BIG;
435 			goto out;
436 		}
437 
438 		/* We've reached the end of our output range */
439 		if (workspace->out_buf.pos >= max_out) {
440 			tot_out += workspace->out_buf.pos;
441 			ret = -E2BIG;
442 			goto out;
443 		}
444 
445 		/* Check if we need more output space */
446 		if (workspace->out_buf.pos == workspace->out_buf.size) {
447 			tot_out += PAGE_SIZE;
448 			max_out -= PAGE_SIZE;
449 			kunmap(out_page);
450 			if (nr_pages == nr_dest_pages) {
451 				out_page = NULL;
452 				ret = -E2BIG;
453 				goto out;
454 			}
455 			out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
456 			if (out_page == NULL) {
457 				ret = -ENOMEM;
458 				goto out;
459 			}
460 			pages[nr_pages++] = out_page;
461 			workspace->out_buf.dst = kmap(out_page);
462 			workspace->out_buf.pos = 0;
463 			workspace->out_buf.size = min_t(size_t, max_out,
464 							PAGE_SIZE);
465 		}
466 
467 		/* We've reached the end of the input */
468 		if (workspace->in_buf.pos >= len) {
469 			tot_in += workspace->in_buf.pos;
470 			break;
471 		}
472 
473 		/* Check if we need more input */
474 		if (workspace->in_buf.pos == workspace->in_buf.size) {
475 			tot_in += PAGE_SIZE;
476 			kunmap(in_page);
477 			put_page(in_page);
478 
479 			start += PAGE_SIZE;
480 			len -= PAGE_SIZE;
481 			in_page = find_get_page(mapping, start >> PAGE_SHIFT);
482 			workspace->in_buf.src = kmap(in_page);
483 			workspace->in_buf.pos = 0;
484 			workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
485 		}
486 	}
487 	while (1) {
488 		size_t ret2;
489 
490 		ret2 = ZSTD_endStream(stream, &workspace->out_buf);
491 		if (ZSTD_isError(ret2)) {
492 			pr_debug("BTRFS: ZSTD_endStream returned %d\n",
493 					ZSTD_getErrorCode(ret2));
494 			ret = -EIO;
495 			goto out;
496 		}
497 		if (ret2 == 0) {
498 			tot_out += workspace->out_buf.pos;
499 			break;
500 		}
501 		if (workspace->out_buf.pos >= max_out) {
502 			tot_out += workspace->out_buf.pos;
503 			ret = -E2BIG;
504 			goto out;
505 		}
506 
507 		tot_out += PAGE_SIZE;
508 		max_out -= PAGE_SIZE;
509 		kunmap(out_page);
510 		if (nr_pages == nr_dest_pages) {
511 			out_page = NULL;
512 			ret = -E2BIG;
513 			goto out;
514 		}
515 		out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
516 		if (out_page == NULL) {
517 			ret = -ENOMEM;
518 			goto out;
519 		}
520 		pages[nr_pages++] = out_page;
521 		workspace->out_buf.dst = kmap(out_page);
522 		workspace->out_buf.pos = 0;
523 		workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
524 	}
525 
526 	if (tot_out >= tot_in) {
527 		ret = -E2BIG;
528 		goto out;
529 	}
530 
531 	ret = 0;
532 	*total_in = tot_in;
533 	*total_out = tot_out;
534 out:
535 	*out_pages = nr_pages;
536 	/* Cleanup */
537 	if (in_page) {
538 		kunmap(in_page);
539 		put_page(in_page);
540 	}
541 	if (out_page)
542 		kunmap(out_page);
543 	return ret;
544 }
545 
546 int zstd_decompress_bio(struct list_head *ws, struct compressed_bio *cb)
547 {
548 	struct workspace *workspace = list_entry(ws, struct workspace, list);
549 	struct page **pages_in = cb->compressed_pages;
550 	u64 disk_start = cb->start;
551 	struct bio *orig_bio = cb->orig_bio;
552 	size_t srclen = cb->compressed_len;
553 	ZSTD_DStream *stream;
554 	int ret = 0;
555 	unsigned long page_in_index = 0;
556 	unsigned long total_pages_in = DIV_ROUND_UP(srclen, PAGE_SIZE);
557 	unsigned long buf_start;
558 	unsigned long total_out = 0;
559 
560 	stream = ZSTD_initDStream(
561 			ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
562 	if (!stream) {
563 		pr_debug("BTRFS: ZSTD_initDStream failed\n");
564 		ret = -EIO;
565 		goto done;
566 	}
567 
568 	workspace->in_buf.src = kmap(pages_in[page_in_index]);
569 	workspace->in_buf.pos = 0;
570 	workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
571 
572 	workspace->out_buf.dst = workspace->buf;
573 	workspace->out_buf.pos = 0;
574 	workspace->out_buf.size = PAGE_SIZE;
575 
576 	while (1) {
577 		size_t ret2;
578 
579 		ret2 = ZSTD_decompressStream(stream, &workspace->out_buf,
580 				&workspace->in_buf);
581 		if (ZSTD_isError(ret2)) {
582 			pr_debug("BTRFS: ZSTD_decompressStream returned %d\n",
583 					ZSTD_getErrorCode(ret2));
584 			ret = -EIO;
585 			goto done;
586 		}
587 		buf_start = total_out;
588 		total_out += workspace->out_buf.pos;
589 		workspace->out_buf.pos = 0;
590 
591 		ret = btrfs_decompress_buf2page(workspace->out_buf.dst,
592 				buf_start, total_out, disk_start, orig_bio);
593 		if (ret == 0)
594 			break;
595 
596 		if (workspace->in_buf.pos >= srclen)
597 			break;
598 
599 		/* Check if we've hit the end of a frame */
600 		if (ret2 == 0)
601 			break;
602 
603 		if (workspace->in_buf.pos == workspace->in_buf.size) {
604 			kunmap(pages_in[page_in_index++]);
605 			if (page_in_index >= total_pages_in) {
606 				workspace->in_buf.src = NULL;
607 				ret = -EIO;
608 				goto done;
609 			}
610 			srclen -= PAGE_SIZE;
611 			workspace->in_buf.src = kmap(pages_in[page_in_index]);
612 			workspace->in_buf.pos = 0;
613 			workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
614 		}
615 	}
616 	ret = 0;
617 	zero_fill_bio(orig_bio);
618 done:
619 	if (workspace->in_buf.src)
620 		kunmap(pages_in[page_in_index]);
621 	return ret;
622 }
623 
624 int zstd_decompress(struct list_head *ws, unsigned char *data_in,
625 		struct page *dest_page, unsigned long start_byte, size_t srclen,
626 		size_t destlen)
627 {
628 	struct workspace *workspace = list_entry(ws, struct workspace, list);
629 	ZSTD_DStream *stream;
630 	int ret = 0;
631 	size_t ret2;
632 	unsigned long total_out = 0;
633 	unsigned long pg_offset = 0;
634 	char *kaddr;
635 
636 	stream = ZSTD_initDStream(
637 			ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
638 	if (!stream) {
639 		pr_warn("BTRFS: ZSTD_initDStream failed\n");
640 		ret = -EIO;
641 		goto finish;
642 	}
643 
644 	destlen = min_t(size_t, destlen, PAGE_SIZE);
645 
646 	workspace->in_buf.src = data_in;
647 	workspace->in_buf.pos = 0;
648 	workspace->in_buf.size = srclen;
649 
650 	workspace->out_buf.dst = workspace->buf;
651 	workspace->out_buf.pos = 0;
652 	workspace->out_buf.size = PAGE_SIZE;
653 
654 	ret2 = 1;
655 	while (pg_offset < destlen
656 	       && workspace->in_buf.pos < workspace->in_buf.size) {
657 		unsigned long buf_start;
658 		unsigned long buf_offset;
659 		unsigned long bytes;
660 
661 		/* Check if the frame is over and we still need more input */
662 		if (ret2 == 0) {
663 			pr_debug("BTRFS: ZSTD_decompressStream ended early\n");
664 			ret = -EIO;
665 			goto finish;
666 		}
667 		ret2 = ZSTD_decompressStream(stream, &workspace->out_buf,
668 				&workspace->in_buf);
669 		if (ZSTD_isError(ret2)) {
670 			pr_debug("BTRFS: ZSTD_decompressStream returned %d\n",
671 					ZSTD_getErrorCode(ret2));
672 			ret = -EIO;
673 			goto finish;
674 		}
675 
676 		buf_start = total_out;
677 		total_out += workspace->out_buf.pos;
678 		workspace->out_buf.pos = 0;
679 
680 		if (total_out <= start_byte)
681 			continue;
682 
683 		if (total_out > start_byte && buf_start < start_byte)
684 			buf_offset = start_byte - buf_start;
685 		else
686 			buf_offset = 0;
687 
688 		bytes = min_t(unsigned long, destlen - pg_offset,
689 				workspace->out_buf.size - buf_offset);
690 
691 		kaddr = kmap_atomic(dest_page);
692 		memcpy(kaddr + pg_offset, workspace->out_buf.dst + buf_offset,
693 				bytes);
694 		kunmap_atomic(kaddr);
695 
696 		pg_offset += bytes;
697 	}
698 	ret = 0;
699 finish:
700 	if (pg_offset < destlen) {
701 		kaddr = kmap_atomic(dest_page);
702 		memset(kaddr + pg_offset, 0, destlen - pg_offset);
703 		kunmap_atomic(kaddr);
704 	}
705 	return ret;
706 }
707 
708 const struct btrfs_compress_op btrfs_zstd_compress = {
709 	/* ZSTD uses own workspace manager */
710 	.workspace_manager = NULL,
711 	.max_level	= ZSTD_BTRFS_MAX_LEVEL,
712 	.default_level	= ZSTD_BTRFS_DEFAULT_LEVEL,
713 };
714