fs/btrfs/zstd.c

   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
 635         stream = ZSTD_initDStream(
 636                         ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
 637         if (!stream) {
 638                 pr_warn("BTRFS: ZSTD_initDStream failed\n");
 639                 ret = -EIO;
 640                 goto finish;
 641         }
 642
 643         destlen = min_t(size_t, destlen, PAGE_SIZE);
 644
 645         workspace->in_buf.src = data_in;
 646         workspace->in_buf.pos = 0;
 647         workspace->in_buf.size = srclen;
 648
 649         workspace->out_buf.dst = workspace->buf;
 650         workspace->out_buf.pos = 0;
 651         workspace->out_buf.size = PAGE_SIZE;
 652
 653         ret2 = 1;
 654         while (pg_offset < destlen
 655                && workspace->in_buf.pos < workspace->in_buf.size) {
 656                 unsigned long buf_start;
 657                 unsigned long buf_offset;
 658                 unsigned long bytes;
 659
 660                 /* Check if the frame is over and we still need more input */
 661                 if (ret2 == 0) {
 662                         pr_debug("BTRFS: ZSTD_decompressStream ended early\n");
 663                         ret = -EIO;
 664                         goto finish;
 665                 }
 666                 ret2 = ZSTD_decompressStream(stream, &workspace->out_buf,
 667                                 &workspace->in_buf);
 668                 if (ZSTD_isError(ret2)) {
 669                         pr_debug("BTRFS: ZSTD_decompressStream returned %d\n",
 670                                         ZSTD_getErrorCode(ret2));
 671                         ret = -EIO;
 672                         goto finish;
 673                 }
 674
 675                 buf_start = total_out;
 676                 total_out += workspace->out_buf.pos;
 677                 workspace->out_buf.pos = 0;
 678
 679                 if (total_out <= start_byte)
 680                         continue;
 681
 682                 if (total_out > start_byte && buf_start < start_byte)
 683                         buf_offset = start_byte - buf_start;
 684                 else
 685                         buf_offset = 0;
 686
 687                 bytes = min_t(unsigned long, destlen - pg_offset,
 688                                 workspace->out_buf.size - buf_offset);
 689
 690                 memcpy_to_page(dest_page, pg_offset,
 691                                workspace->out_buf.dst + buf_offset, bytes);
 692
 693                 pg_offset += bytes;
 694         }
 695         ret = 0;
 696 finish:
 697         if (pg_offset < destlen) {
 698                 memzero_page(dest_page, pg_offset, destlen - pg_offset);
 699         }
 700         return ret;
 701 }
 702
 703 const struct btrfs_compress_op btrfs_zstd_compress = {
 704         /* ZSTD uses own workspace manager */
 705         .workspace_manager = NULL,
 706         .max_level      = ZSTD_BTRFS_MAX_LEVEL,
 707         .default_level  = ZSTD_BTRFS_DEFAULT_LEVEL,
 708 };