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 "compression.h"
  21 #include "ctree.h"
  22
  23 #define ZSTD_BTRFS_MAX_WINDOWLOG 17
  24 #define ZSTD_BTRFS_MAX_INPUT (1 << ZSTD_BTRFS_MAX_WINDOWLOG)
  25 #define ZSTD_BTRFS_DEFAULT_LEVEL 3
  26 #define ZSTD_BTRFS_MAX_LEVEL 15
  27 /* 307s to avoid pathologically clashing with transaction commit */
  28 #define ZSTD_BTRFS_RECLAIM_JIFFIES (307 * HZ)
  29
  30 static ZSTD_parameters zstd_get_btrfs_parameters(unsigned int level,
  31                                                  size_t src_len)
  32 {
  33         ZSTD_parameters params = ZSTD_getParams(level, src_len, 0);
  34
  35         if (params.cParams.windowLog > ZSTD_BTRFS_MAX_WINDOWLOG)
  36                 params.cParams.windowLog = ZSTD_BTRFS_MAX_WINDOWLOG;
  37         WARN_ON(src_len > ZSTD_BTRFS_MAX_INPUT);
  38         return params;
  39 }
  40
  41 struct workspace {
  42         void *mem;
  43         size_t size;
  44         char *buf;
  45         unsigned int level;
  46         unsigned int req_level;
  47         unsigned long last_used; /* jiffies */
  48         struct list_head list;
  49         struct list_head lru_list;
  50         ZSTD_inBuffer in_buf;
  51         ZSTD_outBuffer out_buf;
  52 };
  53
  54 /*
  55  * Zstd Workspace Management
  56  *
  57  * Zstd workspaces have different memory requirements depending on the level.
  58  * The zstd workspaces are managed by having individual lists for each level
  59  * and a global lru.  Forward progress is maintained by protecting a max level
  60  * workspace.
  61  *
  62  * Getting a workspace is done by using the bitmap to identify the levels that
  63  * have available workspaces and scans up.  This lets us recycle higher level
  64  * workspaces because of the monotonic memory guarantee.  A workspace's
  65  * last_used is only updated if it is being used by the corresponding memory
  66  * level.  Putting a workspace involves adding it back to the appropriate places
  67  * and adding it back to the lru if necessary.
  68  *
  69  * A timer is used to reclaim workspaces if they have not been used for
  70  * ZSTD_BTRFS_RECLAIM_JIFFIES.  This helps keep only active workspaces around.
  71  * The upper bound is provided by the workqueue limit which is 2 (percpu limit).
  72  */
  73
  74 struct zstd_workspace_manager {
  75         const struct btrfs_compress_op *ops;
  76         spinlock_t lock;
  77         struct list_head lru_list;
  78         struct list_head idle_ws[ZSTD_BTRFS_MAX_LEVEL];
  79         unsigned long active_map;
  80         wait_queue_head_t wait;
  81         struct timer_list timer;
  82 };
  83
  84 static struct zstd_workspace_manager wsm;
  85
  86 static size_t zstd_ws_mem_sizes[ZSTD_BTRFS_MAX_LEVEL];
  87
  88 static inline struct workspace *list_to_workspace(struct list_head *list)
  89 {
  90         return container_of(list, struct workspace, list);
  91 }
  92
  93 /*
  94  * zstd_reclaim_timer_fn - reclaim timer
  95  * @t: timer
  96  *
  97  * This scans the lru_list and attempts to reclaim any workspace that hasn't
  98  * been used for ZSTD_BTRFS_RECLAIM_JIFFIES.
  99  */
 100 static void zstd_reclaim_timer_fn(struct timer_list *timer)
 101 {
 102         unsigned long reclaim_threshold = jiffies - ZSTD_BTRFS_RECLAIM_JIFFIES;
 103         struct list_head *pos, *next;
 104
 105         spin_lock(&wsm.lock);
 106
 107         if (list_empty(&wsm.lru_list)) {
 108                 spin_unlock(&wsm.lock);
 109                 return;
 110         }
 111
 112         list_for_each_prev_safe(pos, next, &wsm.lru_list) {
 113                 struct workspace *victim = container_of(pos, struct workspace,
 114                                                         lru_list);
 115                 unsigned int level;
 116
 117                 if (time_after(victim->last_used, reclaim_threshold))
 118                         break;
 119
 120                 /* workspace is in use */
 121                 if (victim->req_level)
 122                         continue;
 123
 124                 level = victim->level;
 125                 list_del(&victim->lru_list);
 126                 list_del(&victim->list);
 127                 wsm.ops->free_workspace(&victim->list);
 128
 129                 if (list_empty(&wsm.idle_ws[level - 1]))
 130                         clear_bit(level - 1, &wsm.active_map);
 131
 132         }
 133
 134         if (!list_empty(&wsm.lru_list))
 135                 mod_timer(&wsm.timer, jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
 136
 137         spin_unlock(&wsm.lock);
 138 }
 139
 140 /*
 141  * zstd_calc_ws_mem_sizes - calculate monotonic memory bounds
 142  *
 143  * It is possible based on the level configurations that a higher level
 144  * workspace uses less memory than a lower level workspace.  In order to reuse
 145  * workspaces, this must be made a monotonic relationship.  This precomputes
 146  * the required memory for each level and enforces the monotonicity between
 147  * level and memory required.
 148  */
 149 static void zstd_calc_ws_mem_sizes(void)
 150 {
 151         size_t max_size = 0;
 152         unsigned int level;
 153
 154         for (level = 1; level <= ZSTD_BTRFS_MAX_LEVEL; level++) {
 155                 ZSTD_parameters params =
 156                         zstd_get_btrfs_parameters(level, ZSTD_BTRFS_MAX_INPUT);
 157                 size_t level_size =
 158                         max_t(size_t,
 159                               ZSTD_CStreamWorkspaceBound(params.cParams),
 160                               ZSTD_DStreamWorkspaceBound(ZSTD_BTRFS_MAX_INPUT));
 161
 162                 max_size = max_t(size_t, max_size, level_size);
 163                 zstd_ws_mem_sizes[level - 1] = max_size;
 164         }
 165 }
 166
 167 static void zstd_init_workspace_manager(void)
 168 {
 169         struct list_head *ws;
 170         int i;
 171
 172         zstd_calc_ws_mem_sizes();
 173
 174         wsm.ops = &btrfs_zstd_compress;
 175         spin_lock_init(&wsm.lock);
 176         init_waitqueue_head(&wsm.wait);
 177         timer_setup(&wsm.timer, zstd_reclaim_timer_fn, 0);
 178
 179         INIT_LIST_HEAD(&wsm.lru_list);
 180         for (i = 0; i < ZSTD_BTRFS_MAX_LEVEL; i++)
 181                 INIT_LIST_HEAD(&wsm.idle_ws[i]);
 182
 183         ws = wsm.ops->alloc_workspace(ZSTD_BTRFS_MAX_LEVEL);
 184         if (IS_ERR(ws)) {
 185                 pr_warn(
 186                 "BTRFS: cannot preallocate zstd compression workspace\n");
 187         } else {
 188                 set_bit(ZSTD_BTRFS_MAX_LEVEL - 1, &wsm.active_map);
 189                 list_add(ws, &wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1]);
 190         }
 191 }
 192
 193 static void zstd_cleanup_workspace_manager(void)
 194 {
 195         struct workspace *workspace;
 196         int i;
 197
 198         del_timer(&wsm.timer);
 199
 200         for (i = 0; i < ZSTD_BTRFS_MAX_LEVEL; i++) {
 201                 while (!list_empty(&wsm.idle_ws[i])) {
 202                         workspace = container_of(wsm.idle_ws[i].next,
 203                                                  struct workspace, list);
 204                         list_del(&workspace->list);
 205                         list_del(&workspace->lru_list);
 206                         wsm.ops->free_workspace(&workspace->list);
 207                 }
 208         }
 209 }
 210
 211 /*
 212  * zstd_find_workspace - find workspace
 213  * @level: compression level
 214  *
 215  * This iterates over the set bits in the active_map beginning at the requested
 216  * compression level.  This lets us utilize already allocated workspaces before
 217  * allocating a new one.  If the workspace is of a larger size, it is used, but
 218  * the place in the lru_list and last_used times are not updated.  This is to
 219  * offer the opportunity to reclaim the workspace in favor of allocating an
 220  * appropriately sized one in the future.
 221  */
 222 static struct list_head *zstd_find_workspace(unsigned int level)
 223 {
 224         struct list_head *ws;
 225         struct workspace *workspace;
 226         int i = level - 1;
 227
 228         spin_lock(&wsm.lock);
 229         for_each_set_bit_from(i, &wsm.active_map, ZSTD_BTRFS_MAX_LEVEL) {
 230                 if (!list_empty(&wsm.idle_ws[i])) {
 231                         ws = wsm.idle_ws[i].next;
 232                         workspace = list_to_workspace(ws);
 233                         list_del_init(ws);
 234                         /* keep its place if it's a lower level using this */
 235                         workspace->req_level = level;
 236                         if (level == workspace->level)
 237                                 list_del(&workspace->lru_list);
 238                         if (list_empty(&wsm.idle_ws[i]))
 239                                 clear_bit(i, &wsm.active_map);
 240                         spin_unlock(&wsm.lock);
 241                         return ws;
 242                 }
 243         }
 244         spin_unlock(&wsm.lock);
 245
 246         return NULL;
 247 }
 248
 249 /*
 250  * zstd_get_workspace - zstd's get_workspace
 251  * @level: compression level
 252  *
 253  * If @level is 0, then any compression level can be used.  Therefore, we begin
 254  * scanning from 1.  We first scan through possible workspaces and then after
 255  * attempt to allocate a new workspace.  If we fail to allocate one due to
 256  * memory pressure, go to sleep waiting for the max level workspace to free up.
 257  */
 258 static struct list_head *zstd_get_workspace(unsigned int level)
 259 {
 260         struct list_head *ws;
 261         unsigned int nofs_flag;
 262
 263         /* level == 0 means we can use any workspace */
 264         if (!level)
 265                 level = 1;
 266
 267 again:
 268         ws = zstd_find_workspace(level);
 269         if (ws)
 270                 return ws;
 271
 272         nofs_flag = memalloc_nofs_save();
 273         ws = wsm.ops->alloc_workspace(level);
 274         memalloc_nofs_restore(nofs_flag);
 275
 276         if (IS_ERR(ws)) {
 277                 DEFINE_WAIT(wait);
 278
 279                 prepare_to_wait(&wsm.wait, &wait, TASK_UNINTERRUPTIBLE);
 280                 schedule();
 281                 finish_wait(&wsm.wait, &wait);
 282
 283                 goto again;
 284         }
 285
 286         return ws;
 287 }
 288
 289 /*
 290  * zstd_put_workspace - zstd put_workspace
 291  * @ws: list_head for the workspace
 292  *
 293  * When putting back a workspace, we only need to update the LRU if we are of
 294  * the requested compression level.  Here is where we continue to protect the
 295  * max level workspace or update last_used accordingly.  If the reclaim timer
 296  * isn't set, it is also set here.  Only the max level workspace tries and wakes
 297  * up waiting workspaces.
 298  */
 299 static void zstd_put_workspace(struct list_head *ws)
 300 {
 301         struct workspace *workspace = list_to_workspace(ws);
 302
 303         spin_lock(&wsm.lock);
 304
 305         /* A node is only taken off the lru if we are the corresponding level */
 306         if (workspace->req_level == workspace->level) {
 307                 /* Hide a max level workspace from reclaim */
 308                 if (list_empty(&wsm.idle_ws[ZSTD_BTRFS_MAX_LEVEL - 1])) {
 309                         INIT_LIST_HEAD(&workspace->lru_list);
 310                 } else {
 311                         workspace->last_used = jiffies;
 312                         list_add(&workspace->lru_list, &wsm.lru_list);
 313                         if (!timer_pending(&wsm.timer))
 314                                 mod_timer(&wsm.timer,
 315                                           jiffies + ZSTD_BTRFS_RECLAIM_JIFFIES);
 316                 }
 317         }
 318
 319         set_bit(workspace->level - 1, &wsm.active_map);
 320         list_add(&workspace->list, &wsm.idle_ws[workspace->level - 1]);
 321         workspace->req_level = 0;
 322
 323         spin_unlock(&wsm.lock);
 324
 325         if (workspace->level == ZSTD_BTRFS_MAX_LEVEL)
 326                 cond_wake_up(&wsm.wait);
 327 }
 328
 329 static void zstd_free_workspace(struct list_head *ws)
 330 {
 331         struct workspace *workspace = list_entry(ws, struct workspace, list);
 332
 333         kvfree(workspace->mem);
 334         kfree(workspace->buf);
 335         kfree(workspace);
 336 }
 337
 338 static struct list_head *zstd_alloc_workspace(unsigned int level)
 339 {
 340         struct workspace *workspace;
 341
 342         workspace = kzalloc(sizeof(*workspace), GFP_KERNEL);
 343         if (!workspace)
 344                 return ERR_PTR(-ENOMEM);
 345
 346         workspace->size = zstd_ws_mem_sizes[level - 1];
 347         workspace->level = level;
 348         workspace->req_level = level;
 349         workspace->last_used = jiffies;
 350         workspace->mem = kvmalloc(workspace->size, GFP_KERNEL);
 351         workspace->buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
 352         if (!workspace->mem || !workspace->buf)
 353                 goto fail;
 354
 355         INIT_LIST_HEAD(&workspace->list);
 356         INIT_LIST_HEAD(&workspace->lru_list);
 357
 358         return &workspace->list;
 359 fail:
 360         zstd_free_workspace(&workspace->list);
 361         return ERR_PTR(-ENOMEM);
 362 }
 363
 364 static int zstd_compress_pages(struct list_head *ws,
 365                 struct address_space *mapping,
 366                 u64 start,
 367                 struct page **pages,
 368                 unsigned long *out_pages,
 369                 unsigned long *total_in,
 370                 unsigned long *total_out)
 371 {
 372         struct workspace *workspace = list_entry(ws, struct workspace, list);
 373         ZSTD_CStream *stream;
 374         int ret = 0;
 375         int nr_pages = 0;
 376         struct page *in_page = NULL;  /* The current page to read */
 377         struct page *out_page = NULL; /* The current page to write to */
 378         unsigned long tot_in = 0;
 379         unsigned long tot_out = 0;
 380         unsigned long len = *total_out;
 381         const unsigned long nr_dest_pages = *out_pages;
 382         unsigned long max_out = nr_dest_pages * PAGE_SIZE;
 383         ZSTD_parameters params = zstd_get_btrfs_parameters(workspace->req_level,
 384                                                            len);
 385
 386         *out_pages = 0;
 387         *total_out = 0;
 388         *total_in = 0;
 389
 390         /* Initialize the stream */
 391         stream = ZSTD_initCStream(params, len, workspace->mem,
 392                         workspace->size);
 393         if (!stream) {
 394                 pr_warn("BTRFS: ZSTD_initCStream failed\n");
 395                 ret = -EIO;
 396                 goto out;
 397         }
 398
 399         /* map in the first page of input data */
 400         in_page = find_get_page(mapping, start >> PAGE_SHIFT);
 401         workspace->in_buf.src = kmap(in_page);
 402         workspace->in_buf.pos = 0;
 403         workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
 404
 405
 406         /* Allocate and map in the output buffer */
 407         out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
 408         if (out_page == NULL) {
 409                 ret = -ENOMEM;
 410                 goto out;
 411         }
 412         pages[nr_pages++] = out_page;
 413         workspace->out_buf.dst = kmap(out_page);
 414         workspace->out_buf.pos = 0;
 415         workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
 416
 417         while (1) {
 418                 size_t ret2;
 419
 420                 ret2 = ZSTD_compressStream(stream, &workspace->out_buf,
 421                                 &workspace->in_buf);
 422                 if (ZSTD_isError(ret2)) {
 423                         pr_debug("BTRFS: ZSTD_compressStream returned %d\n",
 424                                         ZSTD_getErrorCode(ret2));
 425                         ret = -EIO;
 426                         goto out;
 427                 }
 428
 429                 /* Check to see if we are making it bigger */
 430                 if (tot_in + workspace->in_buf.pos > 8192 &&
 431                                 tot_in + workspace->in_buf.pos <
 432                                 tot_out + workspace->out_buf.pos) {
 433                         ret = -E2BIG;
 434                         goto out;
 435                 }
 436
 437                 /* We've reached the end of our output range */
 438                 if (workspace->out_buf.pos >= max_out) {
 439                         tot_out += workspace->out_buf.pos;
 440                         ret = -E2BIG;
 441                         goto out;
 442                 }
 443
 444                 /* Check if we need more output space */
 445                 if (workspace->out_buf.pos == workspace->out_buf.size) {
 446                         tot_out += PAGE_SIZE;
 447                         max_out -= PAGE_SIZE;
 448                         kunmap(out_page);
 449                         if (nr_pages == nr_dest_pages) {
 450                                 out_page = NULL;
 451                                 ret = -E2BIG;
 452                                 goto out;
 453                         }
 454                         out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
 455                         if (out_page == NULL) {
 456                                 ret = -ENOMEM;
 457                                 goto out;
 458                         }
 459                         pages[nr_pages++] = out_page;
 460                         workspace->out_buf.dst = kmap(out_page);
 461                         workspace->out_buf.pos = 0;
 462                         workspace->out_buf.size = min_t(size_t, max_out,
 463                                                         PAGE_SIZE);
 464                 }
 465
 466                 /* We've reached the end of the input */
 467                 if (workspace->in_buf.pos >= len) {
 468                         tot_in += workspace->in_buf.pos;
 469                         break;
 470                 }
 471
 472                 /* Check if we need more input */
 473                 if (workspace->in_buf.pos == workspace->in_buf.size) {
 474                         tot_in += PAGE_SIZE;
 475                         kunmap(in_page);
 476                         put_page(in_page);
 477
 478                         start += PAGE_SIZE;
 479                         len -= PAGE_SIZE;
 480                         in_page = find_get_page(mapping, start >> PAGE_SHIFT);
 481                         workspace->in_buf.src = kmap(in_page);
 482                         workspace->in_buf.pos = 0;
 483                         workspace->in_buf.size = min_t(size_t, len, PAGE_SIZE);
 484                 }
 485         }
 486         while (1) {
 487                 size_t ret2;
 488
 489                 ret2 = ZSTD_endStream(stream, &workspace->out_buf);
 490                 if (ZSTD_isError(ret2)) {
 491                         pr_debug("BTRFS: ZSTD_endStream returned %d\n",
 492                                         ZSTD_getErrorCode(ret2));
 493                         ret = -EIO;
 494                         goto out;
 495                 }
 496                 if (ret2 == 0) {
 497                         tot_out += workspace->out_buf.pos;
 498                         break;
 499                 }
 500                 if (workspace->out_buf.pos >= max_out) {
 501                         tot_out += workspace->out_buf.pos;
 502                         ret = -E2BIG;
 503                         goto out;
 504                 }
 505
 506                 tot_out += PAGE_SIZE;
 507                 max_out -= PAGE_SIZE;
 508                 kunmap(out_page);
 509                 if (nr_pages == nr_dest_pages) {
 510                         out_page = NULL;
 511                         ret = -E2BIG;
 512                         goto out;
 513                 }
 514                 out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
 515                 if (out_page == NULL) {
 516                         ret = -ENOMEM;
 517                         goto out;
 518                 }
 519                 pages[nr_pages++] = out_page;
 520                 workspace->out_buf.dst = kmap(out_page);
 521                 workspace->out_buf.pos = 0;
 522                 workspace->out_buf.size = min_t(size_t, max_out, PAGE_SIZE);
 523         }
 524
 525         if (tot_out >= tot_in) {
 526                 ret = -E2BIG;
 527                 goto out;
 528         }
 529
 530         ret = 0;
 531         *total_in = tot_in;
 532         *total_out = tot_out;
 533 out:
 534         *out_pages = nr_pages;
 535         /* Cleanup */
 536         if (in_page) {
 537                 kunmap(in_page);
 538                 put_page(in_page);
 539         }
 540         if (out_page)
 541                 kunmap(out_page);
 542         return ret;
 543 }
 544
 545 static int zstd_decompress_bio(struct list_head *ws, struct compressed_bio *cb)
 546 {
 547         struct workspace *workspace = list_entry(ws, struct workspace, list);
 548         struct page **pages_in = cb->compressed_pages;
 549         u64 disk_start = cb->start;
 550         struct bio *orig_bio = cb->orig_bio;
 551         size_t srclen = cb->compressed_len;
 552         ZSTD_DStream *stream;
 553         int ret = 0;
 554         unsigned long page_in_index = 0;
 555         unsigned long total_pages_in = DIV_ROUND_UP(srclen, PAGE_SIZE);
 556         unsigned long buf_start;
 557         unsigned long total_out = 0;
 558
 559         stream = ZSTD_initDStream(
 560                         ZSTD_BTRFS_MAX_INPUT, workspace->mem, workspace->size);
 561         if (!stream) {
 562                 pr_debug("BTRFS: ZSTD_initDStream failed\n");
 563                 ret = -EIO;
 564                 goto done;
 565         }
 566
 567         workspace->in_buf.src = kmap(pages_in[page_in_index]);
 568         workspace->in_buf.pos = 0;
 569         workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
 570
 571         workspace->out_buf.dst = workspace->buf;
 572         workspace->out_buf.pos = 0;
 573         workspace->out_buf.size = PAGE_SIZE;
 574
 575         while (1) {
 576                 size_t ret2;
 577
 578                 ret2 = ZSTD_decompressStream(stream, &workspace->out_buf,
 579                                 &workspace->in_buf);
 580                 if (ZSTD_isError(ret2)) {
 581                         pr_debug("BTRFS: ZSTD_decompressStream returned %d\n",
 582                                         ZSTD_getErrorCode(ret2));
 583                         ret = -EIO;
 584                         goto done;
 585                 }
 586                 buf_start = total_out;
 587                 total_out += workspace->out_buf.pos;
 588                 workspace->out_buf.pos = 0;
 589
 590                 ret = btrfs_decompress_buf2page(workspace->out_buf.dst,
 591                                 buf_start, total_out, disk_start, orig_bio);
 592                 if (ret == 0)
 593                         break;
 594
 595                 if (workspace->in_buf.pos >= srclen)
 596                         break;
 597
 598                 /* Check if we've hit the end of a frame */
 599                 if (ret2 == 0)
 600                         break;
 601
 602                 if (workspace->in_buf.pos == workspace->in_buf.size) {
 603                         kunmap(pages_in[page_in_index++]);
 604                         if (page_in_index >= total_pages_in) {
 605                                 workspace->in_buf.src = NULL;
 606                                 ret = -EIO;
 607                                 goto done;
 608                         }
 609                         srclen -= PAGE_SIZE;
 610                         workspace->in_buf.src = kmap(pages_in[page_in_index]);
 611                         workspace->in_buf.pos = 0;
 612                         workspace->in_buf.size = min_t(size_t, srclen, PAGE_SIZE);
 613                 }
 614         }
 615         ret = 0;
 616         zero_fill_bio(orig_bio);
 617 done:
 618         if (workspace->in_buf.src)
 619                 kunmap(pages_in[page_in_index]);
 620         return ret;
 621 }
 622
 623 static int zstd_decompress(struct list_head *ws, unsigned char *data_in,
 624                 struct page *dest_page,
 625                 unsigned long start_byte,
 626                 size_t srclen, 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 static unsigned int zstd_set_level(unsigned int level)
 709 {
 710         if (!level)
 711                 return ZSTD_BTRFS_DEFAULT_LEVEL;
 712
 713         return min_t(unsigned int, level, ZSTD_BTRFS_MAX_LEVEL);
 714 }
 715
 716 const struct btrfs_compress_op btrfs_zstd_compress = {
 717         .init_workspace_manager = zstd_init_workspace_manager,
 718         .cleanup_workspace_manager = zstd_cleanup_workspace_manager,
 719         .get_workspace = zstd_get_workspace,
 720         .put_workspace = zstd_put_workspace,
 721         .alloc_workspace = zstd_alloc_workspace,
 722         .free_workspace = zstd_free_workspace,
 723         .compress_pages = zstd_compress_pages,
 724         .decompress_bio = zstd_decompress_bio,
 725         .decompress = zstd_decompress,
 726         .set_level = zstd_set_level,
 727 };