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>
22 #include "btrfs_inode.h"
23 #include "compression.h"
26 #define ZSTD_BTRFS_MAX_WINDOWLOG 17
27 #define ZSTD_BTRFS_MAX_INPUT (1 << ZSTD_BTRFS_MAX_WINDOWLOG)
28 #define ZSTD_BTRFS_DEFAULT_LEVEL 3
29 #define ZSTD_BTRFS_MAX_LEVEL 15
30 /* 307s to avoid pathologically clashing with transaction commit */
31 #define ZSTD_BTRFS_RECLAIM_JIFFIES (307 * HZ)
33 static zstd_parameters
zstd_get_btrfs_parameters(unsigned int level
,
36 zstd_parameters params
= zstd_get_params(level
, src_len
);
38 if (params
.cParams
.windowLog
> ZSTD_BTRFS_MAX_WINDOWLOG
)
39 params
.cParams
.windowLog
= ZSTD_BTRFS_MAX_WINDOWLOG
;
40 WARN_ON(src_len
> ZSTD_BTRFS_MAX_INPUT
);
49 unsigned int req_level
;
50 unsigned long last_used
; /* jiffies */
51 struct list_head list
;
52 struct list_head lru_list
;
53 zstd_in_buffer in_buf
;
54 zstd_out_buffer out_buf
;
58 * Zstd Workspace Management
60 * Zstd workspaces have different memory requirements depending on the level.
61 * The zstd workspaces are managed by having individual lists for each level
62 * and a global lru. Forward progress is maintained by protecting a max level
65 * Getting a workspace is done by using the bitmap to identify the levels that
66 * have available workspaces and scans up. This lets us recycle higher level
67 * workspaces because of the monotonic memory guarantee. A workspace's
68 * last_used is only updated if it is being used by the corresponding memory
69 * level. Putting a workspace involves adding it back to the appropriate places
70 * and adding it back to the lru if necessary.
72 * A timer is used to reclaim workspaces if they have not been used for
73 * ZSTD_BTRFS_RECLAIM_JIFFIES. This helps keep only active workspaces around.
74 * The upper bound is provided by the workqueue limit which is 2 (percpu limit).
77 struct zstd_workspace_manager
{
78 const struct btrfs_compress_op
*ops
;
80 struct list_head lru_list
;
81 struct list_head idle_ws
[ZSTD_BTRFS_MAX_LEVEL
];
82 unsigned long active_map
;
83 wait_queue_head_t wait
;
84 struct timer_list timer
;
87 static struct zstd_workspace_manager wsm
;
89 static size_t zstd_ws_mem_sizes
[ZSTD_BTRFS_MAX_LEVEL
];
91 static inline struct workspace
*list_to_workspace(struct list_head
*list
)
93 return container_of(list
, struct workspace
, list
);
96 void zstd_free_workspace(struct list_head
*ws
);
97 struct list_head
*zstd_alloc_workspace(unsigned int level
);
100 * Timer callback to free unused workspaces.
104 * This scans the lru_list and attempts to reclaim any workspace that hasn't
105 * been used for ZSTD_BTRFS_RECLAIM_JIFFIES.
107 * The context is softirq and does not need the _bh locking primitives.
109 static void zstd_reclaim_timer_fn(struct timer_list
*timer
)
111 unsigned long reclaim_threshold
= jiffies
- ZSTD_BTRFS_RECLAIM_JIFFIES
;
112 struct list_head
*pos
, *next
;
114 ASSERT(timer
== &wsm
.timer
);
116 spin_lock(&wsm
.lock
);
118 if (list_empty(&wsm
.lru_list
)) {
119 spin_unlock(&wsm
.lock
);
123 list_for_each_prev_safe(pos
, next
, &wsm
.lru_list
) {
124 struct workspace
*victim
= container_of(pos
, struct workspace
,
128 if (time_after(victim
->last_used
, reclaim_threshold
))
131 /* workspace is in use */
132 if (victim
->req_level
)
135 level
= victim
->level
;
136 list_del(&victim
->lru_list
);
137 list_del(&victim
->list
);
138 zstd_free_workspace(&victim
->list
);
140 if (list_empty(&wsm
.idle_ws
[level
- 1]))
141 clear_bit(level
- 1, &wsm
.active_map
);
145 if (!list_empty(&wsm
.lru_list
))
146 mod_timer(&wsm
.timer
, jiffies
+ ZSTD_BTRFS_RECLAIM_JIFFIES
);
148 spin_unlock(&wsm
.lock
);
152 * Calculate monotonic memory bounds.
154 * It is possible based on the level configurations that a higher level
155 * workspace uses less memory than a lower level workspace. In order to reuse
156 * workspaces, this must be made a monotonic relationship. This precomputes
157 * the required memory for each level and enforces the monotonicity between
158 * level and memory required.
160 static void zstd_calc_ws_mem_sizes(void)
165 for (level
= 1; level
<= ZSTD_BTRFS_MAX_LEVEL
; level
++) {
166 zstd_parameters params
=
167 zstd_get_btrfs_parameters(level
, ZSTD_BTRFS_MAX_INPUT
);
170 zstd_cstream_workspace_bound(¶ms
.cParams
),
171 zstd_dstream_workspace_bound(ZSTD_BTRFS_MAX_INPUT
));
173 max_size
= max_t(size_t, max_size
, level_size
);
174 zstd_ws_mem_sizes
[level
- 1] = max_size
;
178 void zstd_init_workspace_manager(void)
180 struct list_head
*ws
;
183 zstd_calc_ws_mem_sizes();
185 wsm
.ops
= &btrfs_zstd_compress
;
186 spin_lock_init(&wsm
.lock
);
187 init_waitqueue_head(&wsm
.wait
);
188 timer_setup(&wsm
.timer
, zstd_reclaim_timer_fn
, 0);
190 INIT_LIST_HEAD(&wsm
.lru_list
);
191 for (i
= 0; i
< ZSTD_BTRFS_MAX_LEVEL
; i
++)
192 INIT_LIST_HEAD(&wsm
.idle_ws
[i
]);
194 ws
= zstd_alloc_workspace(ZSTD_BTRFS_MAX_LEVEL
);
197 "BTRFS: cannot preallocate zstd compression workspace\n");
199 set_bit(ZSTD_BTRFS_MAX_LEVEL
- 1, &wsm
.active_map
);
200 list_add(ws
, &wsm
.idle_ws
[ZSTD_BTRFS_MAX_LEVEL
- 1]);
204 void zstd_cleanup_workspace_manager(void)
206 struct workspace
*workspace
;
209 spin_lock_bh(&wsm
.lock
);
210 for (i
= 0; i
< ZSTD_BTRFS_MAX_LEVEL
; i
++) {
211 while (!list_empty(&wsm
.idle_ws
[i
])) {
212 workspace
= container_of(wsm
.idle_ws
[i
].next
,
213 struct workspace
, list
);
214 list_del(&workspace
->list
);
215 list_del(&workspace
->lru_list
);
216 zstd_free_workspace(&workspace
->list
);
219 spin_unlock_bh(&wsm
.lock
);
221 del_timer_sync(&wsm
.timer
);
225 * Find workspace for given level.
227 * @level: compression level
229 * This iterates over the set bits in the active_map beginning at the requested
230 * compression level. This lets us utilize already allocated workspaces before
231 * allocating a new one. If the workspace is of a larger size, it is used, but
232 * the place in the lru_list and last_used times are not updated. This is to
233 * offer the opportunity to reclaim the workspace in favor of allocating an
234 * appropriately sized one in the future.
236 static struct list_head
*zstd_find_workspace(unsigned int level
)
238 struct list_head
*ws
;
239 struct workspace
*workspace
;
242 spin_lock_bh(&wsm
.lock
);
243 for_each_set_bit_from(i
, &wsm
.active_map
, ZSTD_BTRFS_MAX_LEVEL
) {
244 if (!list_empty(&wsm
.idle_ws
[i
])) {
245 ws
= wsm
.idle_ws
[i
].next
;
246 workspace
= list_to_workspace(ws
);
248 /* keep its place if it's a lower level using this */
249 workspace
->req_level
= level
;
250 if (level
== workspace
->level
)
251 list_del(&workspace
->lru_list
);
252 if (list_empty(&wsm
.idle_ws
[i
]))
253 clear_bit(i
, &wsm
.active_map
);
254 spin_unlock_bh(&wsm
.lock
);
258 spin_unlock_bh(&wsm
.lock
);
264 * Zstd get_workspace for level.
266 * @level: compression level
268 * If @level is 0, then any compression level can be used. Therefore, we begin
269 * scanning from 1. We first scan through possible workspaces and then after
270 * attempt to allocate a new workspace. If we fail to allocate one due to
271 * memory pressure, go to sleep waiting for the max level workspace to free up.
273 struct list_head
*zstd_get_workspace(unsigned int level
)
275 struct list_head
*ws
;
276 unsigned int nofs_flag
;
278 /* level == 0 means we can use any workspace */
283 ws
= zstd_find_workspace(level
);
287 nofs_flag
= memalloc_nofs_save();
288 ws
= zstd_alloc_workspace(level
);
289 memalloc_nofs_restore(nofs_flag
);
294 prepare_to_wait(&wsm
.wait
, &wait
, TASK_UNINTERRUPTIBLE
);
296 finish_wait(&wsm
.wait
, &wait
);
305 * Zstd put_workspace.
307 * @ws: list_head for the workspace
309 * When putting back a workspace, we only need to update the LRU if we are of
310 * the requested compression level. Here is where we continue to protect the
311 * max level workspace or update last_used accordingly. If the reclaim timer
312 * isn't set, it is also set here. Only the max level workspace tries and wakes
313 * up waiting workspaces.
315 void zstd_put_workspace(struct list_head
*ws
)
317 struct workspace
*workspace
= list_to_workspace(ws
);
319 spin_lock_bh(&wsm
.lock
);
321 /* A node is only taken off the lru if we are the corresponding level */
322 if (workspace
->req_level
== workspace
->level
) {
323 /* Hide a max level workspace from reclaim */
324 if (list_empty(&wsm
.idle_ws
[ZSTD_BTRFS_MAX_LEVEL
- 1])) {
325 INIT_LIST_HEAD(&workspace
->lru_list
);
327 workspace
->last_used
= jiffies
;
328 list_add(&workspace
->lru_list
, &wsm
.lru_list
);
329 if (!timer_pending(&wsm
.timer
))
330 mod_timer(&wsm
.timer
,
331 jiffies
+ ZSTD_BTRFS_RECLAIM_JIFFIES
);
335 set_bit(workspace
->level
- 1, &wsm
.active_map
);
336 list_add(&workspace
->list
, &wsm
.idle_ws
[workspace
->level
- 1]);
337 workspace
->req_level
= 0;
339 spin_unlock_bh(&wsm
.lock
);
341 if (workspace
->level
== ZSTD_BTRFS_MAX_LEVEL
)
342 cond_wake_up(&wsm
.wait
);
345 void zstd_free_workspace(struct list_head
*ws
)
347 struct workspace
*workspace
= list_entry(ws
, struct workspace
, list
);
349 kvfree(workspace
->mem
);
350 kfree(workspace
->buf
);
354 struct list_head
*zstd_alloc_workspace(unsigned int level
)
356 struct workspace
*workspace
;
358 workspace
= kzalloc(sizeof(*workspace
), GFP_KERNEL
);
360 return ERR_PTR(-ENOMEM
);
362 workspace
->size
= zstd_ws_mem_sizes
[level
- 1];
363 workspace
->level
= level
;
364 workspace
->req_level
= level
;
365 workspace
->last_used
= jiffies
;
366 workspace
->mem
= kvmalloc(workspace
->size
, GFP_KERNEL
| __GFP_NOWARN
);
367 workspace
->buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
368 if (!workspace
->mem
|| !workspace
->buf
)
371 INIT_LIST_HEAD(&workspace
->list
);
372 INIT_LIST_HEAD(&workspace
->lru_list
);
374 return &workspace
->list
;
376 zstd_free_workspace(&workspace
->list
);
377 return ERR_PTR(-ENOMEM
);
380 int zstd_compress_folios(struct list_head
*ws
, struct address_space
*mapping
,
381 u64 start
, struct folio
**folios
, unsigned long *out_folios
,
382 unsigned long *total_in
, unsigned long *total_out
)
384 struct workspace
*workspace
= list_entry(ws
, struct workspace
, list
);
385 zstd_cstream
*stream
;
388 struct folio
*in_folio
= NULL
; /* The current folio to read. */
389 struct folio
*out_folio
= NULL
; /* The current folio to write to. */
390 unsigned long tot_in
= 0;
391 unsigned long tot_out
= 0;
392 unsigned long len
= *total_out
;
393 const unsigned long nr_dest_folios
= *out_folios
;
394 const u64 orig_end
= start
+ len
;
395 unsigned long max_out
= nr_dest_folios
* PAGE_SIZE
;
397 unsigned int cur_len
;
398 zstd_parameters params
= zstd_get_btrfs_parameters(workspace
->req_level
,
405 /* Initialize the stream */
406 stream
= zstd_init_cstream(¶ms
, len
, workspace
->mem
,
408 if (unlikely(!stream
)) {
409 struct btrfs_inode
*inode
= BTRFS_I(mapping
->host
);
411 btrfs_err(inode
->root
->fs_info
,
412 "zstd compression init level %d failed, root %llu inode %llu offset %llu",
413 workspace
->req_level
, btrfs_root_id(inode
->root
),
414 btrfs_ino(inode
), start
);
419 /* map in the first page of input data */
420 ret
= btrfs_compress_filemap_get_folio(mapping
, start
, &in_folio
);
423 pg_off
= offset_in_page(start
);
424 cur_len
= btrfs_calc_input_length(orig_end
, start
);
425 workspace
->in_buf
.src
= kmap_local_folio(in_folio
, pg_off
);
426 workspace
->in_buf
.pos
= 0;
427 workspace
->in_buf
.size
= cur_len
;
429 /* Allocate and map in the output buffer */
430 out_folio
= btrfs_alloc_compr_folio();
431 if (out_folio
== NULL
) {
435 folios
[nr_folios
++] = out_folio
;
436 workspace
->out_buf
.dst
= folio_address(out_folio
);
437 workspace
->out_buf
.pos
= 0;
438 workspace
->out_buf
.size
= min_t(size_t, max_out
, PAGE_SIZE
);
443 ret2
= zstd_compress_stream(stream
, &workspace
->out_buf
,
445 if (unlikely(zstd_is_error(ret2
))) {
446 struct btrfs_inode
*inode
= BTRFS_I(mapping
->host
);
448 btrfs_warn(inode
->root
->fs_info
,
449 "zstd compression level %d failed, error %d root %llu inode %llu offset %llu",
450 workspace
->req_level
, zstd_get_error_code(ret2
),
451 btrfs_root_id(inode
->root
), btrfs_ino(inode
),
457 /* Check to see if we are making it bigger */
458 if (tot_in
+ workspace
->in_buf
.pos
> 8192 &&
459 tot_in
+ workspace
->in_buf
.pos
<
460 tot_out
+ workspace
->out_buf
.pos
) {
465 /* We've reached the end of our output range */
466 if (workspace
->out_buf
.pos
>= max_out
) {
467 tot_out
+= workspace
->out_buf
.pos
;
472 /* Check if we need more output space */
473 if (workspace
->out_buf
.pos
== workspace
->out_buf
.size
) {
474 tot_out
+= PAGE_SIZE
;
475 max_out
-= PAGE_SIZE
;
476 if (nr_folios
== nr_dest_folios
) {
480 out_folio
= btrfs_alloc_compr_folio();
481 if (out_folio
== NULL
) {
485 folios
[nr_folios
++] = out_folio
;
486 workspace
->out_buf
.dst
= folio_address(out_folio
);
487 workspace
->out_buf
.pos
= 0;
488 workspace
->out_buf
.size
= min_t(size_t, max_out
,
492 /* We've reached the end of the input */
493 if (workspace
->in_buf
.pos
>= len
) {
494 tot_in
+= workspace
->in_buf
.pos
;
498 /* Check if we need more input */
499 if (workspace
->in_buf
.pos
== workspace
->in_buf
.size
) {
500 tot_in
+= workspace
->in_buf
.size
;
501 kunmap_local(workspace
->in_buf
.src
);
502 workspace
->in_buf
.src
= NULL
;
506 ret
= btrfs_compress_filemap_get_folio(mapping
, start
, &in_folio
);
509 pg_off
= offset_in_page(start
);
510 cur_len
= btrfs_calc_input_length(orig_end
, start
);
511 workspace
->in_buf
.src
= kmap_local_folio(in_folio
, pg_off
);
512 workspace
->in_buf
.pos
= 0;
513 workspace
->in_buf
.size
= cur_len
;
519 ret2
= zstd_end_stream(stream
, &workspace
->out_buf
);
520 if (unlikely(zstd_is_error(ret2
))) {
521 struct btrfs_inode
*inode
= BTRFS_I(mapping
->host
);
523 btrfs_err(inode
->root
->fs_info
,
524 "zstd compression end level %d failed, error %d root %llu inode %llu offset %llu",
525 workspace
->req_level
, zstd_get_error_code(ret2
),
526 btrfs_root_id(inode
->root
), btrfs_ino(inode
),
532 tot_out
+= workspace
->out_buf
.pos
;
535 if (workspace
->out_buf
.pos
>= max_out
) {
536 tot_out
+= workspace
->out_buf
.pos
;
541 tot_out
+= PAGE_SIZE
;
542 max_out
-= PAGE_SIZE
;
543 if (nr_folios
== nr_dest_folios
) {
547 out_folio
= btrfs_alloc_compr_folio();
548 if (out_folio
== NULL
) {
552 folios
[nr_folios
++] = out_folio
;
553 workspace
->out_buf
.dst
= folio_address(out_folio
);
554 workspace
->out_buf
.pos
= 0;
555 workspace
->out_buf
.size
= min_t(size_t, max_out
, PAGE_SIZE
);
558 if (tot_out
>= tot_in
) {
565 *total_out
= tot_out
;
567 *out_folios
= nr_folios
;
568 if (workspace
->in_buf
.src
) {
569 kunmap_local(workspace
->in_buf
.src
);
575 int zstd_decompress_bio(struct list_head
*ws
, struct compressed_bio
*cb
)
577 struct workspace
*workspace
= list_entry(ws
, struct workspace
, list
);
578 struct folio
**folios_in
= cb
->compressed_folios
;
579 size_t srclen
= cb
->compressed_len
;
580 zstd_dstream
*stream
;
582 unsigned long folio_in_index
= 0;
583 unsigned long total_folios_in
= DIV_ROUND_UP(srclen
, PAGE_SIZE
);
584 unsigned long buf_start
;
585 unsigned long total_out
= 0;
587 stream
= zstd_init_dstream(
588 ZSTD_BTRFS_MAX_INPUT
, workspace
->mem
, workspace
->size
);
589 if (unlikely(!stream
)) {
590 struct btrfs_inode
*inode
= cb
->bbio
.inode
;
592 btrfs_err(inode
->root
->fs_info
,
593 "zstd decompression init failed, root %llu inode %llu offset %llu",
594 btrfs_root_id(inode
->root
), btrfs_ino(inode
), cb
->start
);
599 workspace
->in_buf
.src
= kmap_local_folio(folios_in
[folio_in_index
], 0);
600 workspace
->in_buf
.pos
= 0;
601 workspace
->in_buf
.size
= min_t(size_t, srclen
, PAGE_SIZE
);
603 workspace
->out_buf
.dst
= workspace
->buf
;
604 workspace
->out_buf
.pos
= 0;
605 workspace
->out_buf
.size
= PAGE_SIZE
;
610 ret2
= zstd_decompress_stream(stream
, &workspace
->out_buf
,
612 if (unlikely(zstd_is_error(ret2
))) {
613 struct btrfs_inode
*inode
= cb
->bbio
.inode
;
615 btrfs_err(inode
->root
->fs_info
,
616 "zstd decompression failed, error %d root %llu inode %llu offset %llu",
617 zstd_get_error_code(ret2
), btrfs_root_id(inode
->root
),
618 btrfs_ino(inode
), cb
->start
);
622 buf_start
= total_out
;
623 total_out
+= workspace
->out_buf
.pos
;
624 workspace
->out_buf
.pos
= 0;
626 ret
= btrfs_decompress_buf2page(workspace
->out_buf
.dst
,
627 total_out
- buf_start
, cb
, buf_start
);
631 if (workspace
->in_buf
.pos
>= srclen
)
634 /* Check if we've hit the end of a frame */
638 if (workspace
->in_buf
.pos
== workspace
->in_buf
.size
) {
639 kunmap_local(workspace
->in_buf
.src
);
641 if (folio_in_index
>= total_folios_in
) {
642 workspace
->in_buf
.src
= NULL
;
647 workspace
->in_buf
.src
=
648 kmap_local_folio(folios_in
[folio_in_index
], 0);
649 workspace
->in_buf
.pos
= 0;
650 workspace
->in_buf
.size
= min_t(size_t, srclen
, PAGE_SIZE
);
655 if (workspace
->in_buf
.src
)
656 kunmap_local(workspace
->in_buf
.src
);
660 int zstd_decompress(struct list_head
*ws
, const u8
*data_in
,
661 struct folio
*dest_folio
, unsigned long dest_pgoff
, size_t srclen
,
664 struct workspace
*workspace
= list_entry(ws
, struct workspace
, list
);
665 struct btrfs_fs_info
*fs_info
= btrfs_sb(folio_inode(dest_folio
)->i_sb
);
666 const u32 sectorsize
= fs_info
->sectorsize
;
667 zstd_dstream
*stream
;
669 unsigned long to_copy
= 0;
671 stream
= zstd_init_dstream(
672 ZSTD_BTRFS_MAX_INPUT
, workspace
->mem
, workspace
->size
);
673 if (unlikely(!stream
)) {
674 struct btrfs_inode
*inode
= folio_to_inode(dest_folio
);
676 btrfs_err(inode
->root
->fs_info
,
677 "zstd decompression init failed, root %llu inode %llu offset %llu",
678 btrfs_root_id(inode
->root
), btrfs_ino(inode
),
679 folio_pos(dest_folio
));
684 workspace
->in_buf
.src
= data_in
;
685 workspace
->in_buf
.pos
= 0;
686 workspace
->in_buf
.size
= srclen
;
688 workspace
->out_buf
.dst
= workspace
->buf
;
689 workspace
->out_buf
.pos
= 0;
690 workspace
->out_buf
.size
= sectorsize
;
693 * Since both input and output buffers should not exceed one sector,
694 * one call should end the decompression.
696 ret
= zstd_decompress_stream(stream
, &workspace
->out_buf
, &workspace
->in_buf
);
697 if (unlikely(zstd_is_error(ret
))) {
698 struct btrfs_inode
*inode
= folio_to_inode(dest_folio
);
700 btrfs_err(inode
->root
->fs_info
,
701 "zstd decompression failed, error %d root %llu inode %llu offset %llu",
702 zstd_get_error_code(ret
), btrfs_root_id(inode
->root
),
703 btrfs_ino(inode
), folio_pos(dest_folio
));
706 to_copy
= workspace
->out_buf
.pos
;
707 memcpy_to_folio(dest_folio
, dest_pgoff
, workspace
->out_buf
.dst
, to_copy
);
709 /* Error or early end. */
710 if (unlikely(to_copy
< destlen
)) {
712 folio_zero_range(dest_folio
, dest_pgoff
+ to_copy
, destlen
- to_copy
);
717 const struct btrfs_compress_op btrfs_zstd_compress
= {
718 /* ZSTD uses own workspace manager */
719 .workspace_manager
= NULL
,
720 .max_level
= ZSTD_BTRFS_MAX_LEVEL
,
721 .default_level
= ZSTD_BTRFS_DEFAULT_LEVEL
,