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_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
  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_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
  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 /**
  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  */
 107 static 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  * zstd_calc_ws_mem_sizes - 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  */
 156 static 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(&params.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
 174 void 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
 200 void 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  * zstd_find_workspace - find workspace
 222  * @level: compression level
 223  *
 224  * This iterates over the set bits in the active_map beginning at the requested
 225  * compression level.  This lets us utilize already allocated workspaces before
 226  * allocating a new one.  If the workspace is of a larger size, it is used, but
 227  * the place in the lru_list and last_used times are not updated.  This is to
 228  * offer the opportunity to reclaim the workspace in favor of allocating an
 229  * appropriately sized one in the future.
 230  */
 231 static struct list_head *zstd_find_workspace(unsigned int level)
 232 {
 233         struct list_head *ws;
 234         struct workspace *workspace;
 235         int i = level - 1;
 236
 237         spin_lock_bh(&wsm.lock);
 238         for_each_set_bit_from(i, &wsm.active_map, ZSTD_BTRFS_MAX_LEVEL) {
 239                 if (!list_empty(&wsm.idle_ws[i])) {
 240                         ws = wsm.idle_ws[i].next;
 241                         workspace = list_to_workspace(ws);
 242                         list_del_init(ws);
 243                         /* keep its place if it's a lower level using this */
 244                         workspace->req_level = level;
 245                         if (level == workspace->level)
 246                                 list_del(&workspace->lru_list);
 247                         if (list_empty(&wsm.idle_ws[i]))
 248                                 clear_bit(i, &wsm.active_map);
 249                         spin_unlock_bh(&wsm.lock);
 250                         return ws;
 251                 }
 252         }
 253         spin_unlock_bh(&wsm.lock);
 254
 255         return NULL;
 256 }
 257
 258 /*
 259  * zstd_get_workspace - zstd's get_workspace
 260  * @level: compression level
 261  *
 262  * If @level is 0, then any compression level can be used.  Therefore, we begin
 263  * scanning from 1.  We first scan through possible workspaces and then after
 264  * attempt to allocate a new workspace.  If we fail to allocate one due to
 265  * memory pressure, go to sleep waiting for the max level workspace to free up.
 266  */
 267 struct list_head *zstd_get_workspace(unsigned int level)
 268 {
 269         struct list_head *ws;
 270         unsigned int nofs_flag;
 271
 272         /* level == 0 means we can use any workspace */
 273         if (!level)
 274                 level = 1;
 275
 276 again:
 277         ws = zstd_find_workspace(level);
 278         if (ws)
 279                 return ws;
 280
 281         nofs_flag = memalloc_nofs_save();
 282         ws = zstd_alloc_workspace(level);
 283         memalloc_nofs_restore(nofs_flag);
 284
 285         if (IS_ERR(ws)) {
 286                 DEFINE_WAIT(wait);
 287
 288                 prepare_to_wait(&wsm.wait, &wait, TASK_UNINTERRUPTIBLE);
 289                 schedule();
 290                 finish_wait(&wsm.wait, &wait);
 291
 292                 goto again;
 293         }
 294
 295         return ws;
 296 }
 297
 298 /*
 299  * zstd_put_workspace - zstd put_workspace
 300  * @ws: list_head for the workspace
 301  *
 302  * When putting back a workspace, we only need to update the LRU if we are of
 303  * the requested compression level.  Here is where we continue to protect the
 304  * max level workspace or update last_used accordingly.  If the reclaim timer
 305  * isn't set, it is also set here.  Only the max level workspace tries and wakes
 306  * up waiting workspaces.
 307  */
 308 void zstd_put_workspace(struct list_head *ws)
 309 {
 310         struct workspace *workspace = list_to_workspace(ws);
 311
 312         spin_lock_bh(&wsm.lock);
 313
 314         /* A node is only taken off the lru if we are the corresponding level */
 315         if (workspace->req_level == workspace->level) {
 316                 /* Hide a max level workspace from reclaim */
 317                 if (list_empty(&wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1])) {
 318                         INIT_LIST_HEAD(&workspace->lru_list);
 319                 } else {
 320                         workspace->last_used = jiffies;
 321                         list_add(&workspace->lru_list, &wsm.lru_list);
 322                         if (!timer_pending(&wsm.timer))
 323                                 mod_timer(&wsm.timer,
 324                                           jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
 325                 }
 326         }
 327
 328         set_bit(workspace->level - 1, &wsm.active_map);
 329         list_add(&workspace->list, &wsm.idle_ws[workspace->level - 1]);
 330         workspace->req_level = 0;
 331
 332         spin_unlock_bh(&wsm.lock);
 333
 334         if (workspace->level == ZSTD_BTRFS_MAX_LEVEL)
 335                 cond_wake_up(&wsm.wait);
 336 }
 337
 338 void zstd_free_workspace(struct list_head *ws)
 339 {
 340         struct workspace *workspace = list_entry(ws, struct workspace, list);
 341
 342         kvfree(workspace->mem);
 343         kfree(workspace->buf);
 344         kfree(workspace);
 345 }
 346
 347 struct list_head *zstd_alloc_workspace(unsigned int level)
 348 {
 349         struct workspace *workspace;
 350
 351         workspace = kzalloc(sizeof(*workspace), GFP_KERNEL);
 352         if (!workspace)
 353                 return ERR_PTR(-ENOMEM);
 354
 355         workspace->size = zstd_ws_mem_sizes[level - 1];
 356         workspace->level = level;
 357         workspace->req_level = level;
 358         workspace->last_used = jiffies;
 359         workspace->mem = kvmalloc(workspace->size, GFP_KERNEL);
 360         workspace->buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
 361         if (!workspace->mem || !workspace->buf)
 362                 goto fail;
 363
 364         INIT_LIST_HEAD(&workspace->list);
 365         INIT_LIST_HEAD(&workspace->lru_list);
 366
 367         return &workspace->list;
 368 fail:
 369         zstd_free_workspace(&workspace->list);
 370         return ERR_PTR(-ENOMEM);
 371 }
 372
 373 int zstd_compress_pages(struct list_head *ws, struct address_space *mapping,
 374                 u64 start, struct page **pages, unsigned long *out_pages,
 375                 unsigned long *total_in, unsigned long *total_out)
 376 {
 377         struct workspace *workspace = list_entry(ws, struct workspace, list);
 378         zstd_cstream *stream;
 379         int ret = 0;
 380         int nr_pages = 0;
 381         struct page *in_page = NULL;  /* The current page to read */
 382         struct page *out_page = NULL; /* The current page to write to */
 383         unsigned long tot_in = 0;
 384         unsigned long tot_out = 0;
 385         unsigned long len = *total_out;
 386         const unsigned long nr_dest_pages = *out_pages;
 387         unsigned long max_out = nr_dest_pages * PAGE_SIZE;
 388         zstd_parameters params = zstd_get_btrfs_parameters(workspace->req_level,
 389                                                            len);
 390
 391         *out_pages = 0;
 392         *total_out = 0;
 393         *total_in = 0;
 394
 395         /* Initialize the stream */
 396         stream = zstd_init_cstream(&params, len, workspace->mem,
 397                         workspace->size);
 398         if (!stream) {
 399                 pr_warn("BTRFS: zstd_init_cstream failed\n");
 400                 ret = -EIO;
 401                 goto out;
 402         }
 403
 404         /* map in the first page of input data */
 405         in_page = find_get_page(mapping, start >> PAGE_SHIFT);
 406         workspace->in_buf.src = kmap(in_page);
 407         workspace->in_buf.pos = 0;
 408         workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
 409
 410
 411         /* Allocate and map in the output buffer */
 412         out_page = alloc_page(GFP_NOFS);
 413         if (out_page == NULL) {
 414                 ret = -ENOMEM;
 415                 goto out;
 416         }
 417         pages[nr_pages++] = out_page;
 418         workspace->out_buf.dst = kmap(out_page);
 419         workspace->out_buf.pos = 0;
 420         workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
 421
 422         while (1) {
 423                 size_t ret2;
 424
 425                 ret2 = zstd_compress_stream(stream, &workspace->out_buf,
 426                                 &workspace->in_buf);
 427                 if (zstd_is_error(ret2)) {
 428                         pr_debug("BTRFS: zstd_compress_stream returned %d\n",
 429                                         zstd_get_error_code(ret2));
 430                         ret = -EIO;
 431                         goto out;
 432                 }
 433
 434                 /* Check to see if we are making it bigger */
 435                 if (tot_in + workspace->in_buf.pos > 8192 &&
 436                                 tot_in + workspace->in_buf.pos <
 437                                 tot_out + workspace->out_buf.pos) {
 438                         ret = -E2BIG;
 439                         goto out;
 440                 }
 441
 442                 /* We've reached the end of our output range */
 443                 if (workspace->out_buf.pos >= max_out) {
 444                         tot_out += workspace->out_buf.pos;
 445                         ret = -E2BIG;
 446                         goto out;
 447                 }
 448
 449                 /* Check if we need more output space */
 450                 if (workspace->out_buf.pos == workspace->out_buf.size) {
 451                         tot_out += PAGE_SIZE;
 452                         max_out -= PAGE_SIZE;
 453                         kunmap(out_page);
 454                         if (nr_pages == nr_dest_pages) {
 455                                 out_page = NULL;
 456                                 ret = -E2BIG;
 457                                 goto out;
 458                         }
 459                         out_page = alloc_page(GFP_NOFS);
 460                         if (out_page == NULL) {
 461                                 ret = -ENOMEM;
 462                                 goto out;
 463                         }
 464                         pages[nr_pages++] = out_page;
 465                         workspace->out_buf.dst = kmap(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(in_page);
 481                         put_page(in_page);
 482
 483                         start += PAGE_SIZE;
 484                         len -= PAGE_SIZE;
 485                         in_page = find_get_page(mapping, start >> PAGE_SHIFT);
 486                         workspace->in_buf.src = kmap(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                 kunmap(out_page);
 514                 if (nr_pages == nr_dest_pages) {
 515                         out_page = NULL;
 516                         ret = -E2BIG;
 517                         goto out;
 518                 }
 519                 out_page = alloc_page(GFP_NOFS);
 520                 if (out_page == NULL) {
 521                         ret = -ENOMEM;
 522                         goto out;
 523                 }
 524                 pages[nr_pages++] = out_page;
 525                 workspace->out_buf.dst = kmap(out_page);
 526                 workspace->out_buf.pos = 0;
 527                 workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
 528         }
 529
 530         if (tot_out >= tot_in) {
 531                 ret = -E2BIG;
 532                 goto out;
 533         }
 534
 535         ret = 0;
 536         *total_in = tot_in;
 537         *total_out = tot_out;
 538 out:
 539         *out_pages = nr_pages;
 540         /* Cleanup */
 541         if (in_page) {
 542                 kunmap(in_page);
 543                 put_page(in_page);
 544         }
 545         if (out_page)
 546                 kunmap(out_page);
 547         return ret;
 548 }
 549
 550 int zstd_decompress_bio(struct list_head *ws, struct compressed_bio *cb)
 551 {
 552         struct workspace *workspace = list_entry(ws, struct workspace, list);
 553         struct page **pages_in = cb->compressed_pages;
 554         size_t srclen = cb->compressed_len;
 555         zstd_dstream *stream;
 556         int ret = 0;
 557         unsigned long page_in_index = 0;
 558         unsigned long total_pages_in = DIV_ROUND_UP(srclen, PAGE_SIZE);
 559         unsigned long buf_start;
 560         unsigned long total_out = 0;
 561
 562         stream = zstd_init_dstream(
 563                         ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
 564         if (!stream) {
 565                 pr_debug("BTRFS: zstd_init_dstream failed\n");
 566                 ret = -EIO;
 567                 goto done;
 568         }
 569
 570         workspace->in_buf.src = kmap(pages_in[page_in_index]);
 571         workspace->in_buf.pos = 0;
 572         workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
 573
 574         workspace->out_buf.dst = workspace->buf;
 575         workspace->out_buf.pos = 0;
 576         workspace->out_buf.size = PAGE_SIZE;
 577
 578         while (1) {
 579                 size_t ret2;
 580
 581                 ret2 = zstd_decompress_stream(stream, &workspace->out_buf,
 582                                 &workspace->in_buf);
 583                 if (zstd_is_error(ret2)) {
 584                         pr_debug("BTRFS: zstd_decompress_stream returned %d\n",
 585                                         zstd_get_error_code(ret2));
 586                         ret = -EIO;
 587                         goto done;
 588                 }
 589                 buf_start = total_out;
 590                 total_out += workspace->out_buf.pos;
 591                 workspace->out_buf.pos = 0;
 592
 593                 ret = btrfs_decompress_buf2page(workspace->out_buf.dst,
 594                                 total_out - buf_start, cb, buf_start);
 595                 if (ret == 0)
 596                         break;
 597
 598                 if (workspace->in_buf.pos >= srclen)
 599                         break;
 600
 601                 /* Check if we've hit the end of a frame */
 602                 if (ret2 == 0)
 603                         break;
 604
 605                 if (workspace->in_buf.pos == workspace->in_buf.size) {
 606                         kunmap(pages_in[page_in_index++]);
 607                         if (page_in_index >= total_pages_in) {
 608                                 workspace->in_buf.src = NULL;
 609                                 ret = -EIO;
 610                                 goto done;
 611                         }
 612                         srclen -= PAGE_SIZE;
 613                         workspace->in_buf.src = kmap(pages_in[page_in_index]);
 614                         workspace->in_buf.pos = 0;
 615                         workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
 616                 }
 617         }
 618         ret = 0;
 619         zero_fill_bio(cb->orig_bio);
 620 done:
 621         if (workspace->in_buf.src)
 622                 kunmap(pages_in[page_in_index]);
 623         return ret;
 624 }
 625
 626 int zstd_decompress(struct list_head *ws, unsigned char *data_in,
 627                 struct page *dest_page, unsigned long start_byte, size_t srclen,
 628                 size_t destlen)
 629 {
 630         struct workspace *workspace = list_entry(ws, struct workspace, list);
 631         zstd_dstream *stream;
 632         int ret = 0;
 633         size_t ret2;
 634         unsigned long total_out = 0;
 635         unsigned long pg_offset = 0;
 636
 637         stream = zstd_init_dstream(
 638                         ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
 639         if (!stream) {
 640                 pr_warn("BTRFS: zstd_init_dstream failed\n");
 641                 ret = -EIO;
 642                 goto finish;
 643         }
 644
 645         destlen = min_t(size_t, destlen, PAGE_SIZE);
 646
 647         workspace->in_buf.src = data_in;
 648         workspace->in_buf.pos = 0;
 649         workspace->in_buf.size = srclen;
 650
 651         workspace->out_buf.dst = workspace->buf;
 652         workspace->out_buf.pos = 0;
 653         workspace->out_buf.size = PAGE_SIZE;
 654
 655         ret2 = 1;
 656         while (pg_offset < destlen
 657                && workspace->in_buf.pos < workspace->in_buf.size) {
 658                 unsigned long buf_start;
 659                 unsigned long buf_offset;
 660                 unsigned long bytes;
 661
 662                 /* Check if the frame is over and we still need more input */
 663                 if (ret2 == 0) {
 664                         pr_debug("BTRFS: zstd_decompress_stream ended early\n");
 665                         ret = -EIO;
 666                         goto finish;
 667                 }
 668                 ret2 = zstd_decompress_stream(stream, &workspace->out_buf,
 669                                 &workspace->in_buf);
 670                 if (zstd_is_error(ret2)) {
 671                         pr_debug("BTRFS: zstd_decompress_stream returned %d\n",
 672                                         zstd_get_error_code(ret2));
 673                         ret = -EIO;
 674                         goto finish;
 675                 }
 676
 677                 buf_start = total_out;
 678                 total_out += workspace->out_buf.pos;
 679                 workspace->out_buf.pos = 0;
 680
 681                 if (total_out <= start_byte)
 682                         continue;
 683
 684                 if (total_out > start_byte && buf_start < start_byte)
 685                         buf_offset = start_byte - buf_start;
 686                 else
 687                         buf_offset = 0;
 688
 689                 bytes = min_t(unsigned long, destlen - pg_offset,
 690                                 workspace->out_buf.size - buf_offset);
 691
 692                 memcpy_to_page(dest_page, pg_offset,
 693                                workspace->out_buf.dst + buf_offset, bytes);
 694
 695                 pg_offset += bytes;
 696         }
 697         ret = 0;
 698 finish:
 699         if (pg_offset < destlen) {
 700                 memzero_page(dest_page, pg_offset, destlen - pg_offset);
 701         }
 702         return ret;
 703 }
 704
 705 const struct btrfs_compress_op btrfs_zstd_compress = {
 706         /* ZSTD uses own workspace manager */
 707         .workspace_manager = NULL,
 708         .max_level      = ZSTD_BTRFS_MAX_LEVEL,
 709         .default_level  = ZSTD_BTRFS_DEFAULT_LEVEL,
 710 };