fs/btrfs/zlib.c

   1 /*
   2  * Copyright (C) 2008 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  *
  18  * Based on jffs2 zlib code:
  19  * Copyright © 2001-2007 Red Hat, Inc.
  20  * Created by David Woodhouse <dwmw2@infradead.org>
  21  */
  22
  23 #include <linux/kernel.h>
  24 #include <linux/slab.h>
  25 #include <linux/zlib.h>
  26 #include <linux/zutil.h>
  27 #include <linux/vmalloc.h>
  28 #include <linux/init.h>
  29 #include <linux/err.h>
  30 #include <linux/sched.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/bio.h>
  33 #include "compression.h"
  34
  35 /* Plan: call deflate() with avail_in == *sourcelen,
  36         avail_out = *dstlen - 12 and flush == Z_FINISH.
  37         If it doesn't manage to finish, call it again with
  38         avail_in == 0 and avail_out set to the remaining 12
  39         bytes for it to clean up.
  40    Q: Is 12 bytes sufficient?
  41 */
  42 #define STREAM_END_SPACE 12
  43
  44 struct workspace {
  45         z_stream inf_strm;
  46         z_stream def_strm;
  47         char *buf;
  48         struct list_head list;
  49 };
  50
  51 static LIST_HEAD(idle_workspace);
  52 static DEFINE_SPINLOCK(workspace_lock);
  53 static unsigned long num_workspace;
  54 static atomic_t alloc_workspace = ATOMIC_INIT(0);
  55 static DECLARE_WAIT_QUEUE_HEAD(workspace_wait);
  56
  57 /*
  58  * this finds an available zlib workspace or allocates a new one
  59  * NULL or an ERR_PTR is returned if things go bad.
  60  */
  61 static struct workspace *find_zlib_workspace(void)
  62 {
  63         struct workspace *workspace;
  64         int ret;
  65         int cpus = num_online_cpus();
  66
  67 again:
  68         spin_lock(&workspace_lock);
  69         if (!list_empty(&idle_workspace)) {
  70                 workspace = list_entry(idle_workspace.next, struct workspace,
  71                                        list);
  72                 list_del(&workspace->list);
  73                 num_workspace--;
  74                 spin_unlock(&workspace_lock);
  75                 return workspace;
  76
  77         }
  78         if (atomic_read(&alloc_workspace) > cpus) {
  79                 DEFINE_WAIT(wait);
  80
  81                 spin_unlock(&workspace_lock);
  82                 prepare_to_wait(&workspace_wait, &wait, TASK_UNINTERRUPTIBLE);
  83                 if (atomic_read(&alloc_workspace) > cpus && !num_workspace)
  84                         schedule();
  85                 finish_wait(&workspace_wait, &wait);
  86                 goto again;
  87         }
  88         atomic_inc(&alloc_workspace);
  89         spin_unlock(&workspace_lock);
  90
  91         workspace = kzalloc(sizeof(*workspace), GFP_NOFS);
  92         if (!workspace) {
  93                 ret = -ENOMEM;
  94                 goto fail;
  95         }
  96
  97         workspace->def_strm.workspace = vmalloc(zlib_deflate_workspacesize());
  98         if (!workspace->def_strm.workspace) {
  99                 ret = -ENOMEM;
 100                 goto fail;
 101         }
 102         workspace->inf_strm.workspace = vmalloc(zlib_inflate_workspacesize());
 103         if (!workspace->inf_strm.workspace) {
 104                 ret = -ENOMEM;
 105                 goto fail_inflate;
 106         }
 107         workspace->buf = kmalloc(PAGE_CACHE_SIZE, GFP_NOFS);
 108         if (!workspace->buf) {
 109                 ret = -ENOMEM;
 110                 goto fail_kmalloc;
 111         }
 112         return workspace;
 113
 114 fail_kmalloc:
 115         vfree(workspace->inf_strm.workspace);
 116 fail_inflate:
 117         vfree(workspace->def_strm.workspace);
 118 fail:
 119         kfree(workspace);
 120         atomic_dec(&alloc_workspace);
 121         wake_up(&workspace_wait);
 122         return ERR_PTR(ret);
 123 }
 124
 125 /*
 126  * put a workspace struct back on the list or free it if we have enough
 127  * idle ones sitting around
 128  */
 129 static int free_workspace(struct workspace *workspace)
 130 {
 131         spin_lock(&workspace_lock);
 132         if (num_workspace < num_online_cpus()) {
 133                 list_add_tail(&workspace->list, &idle_workspace);
 134                 num_workspace++;
 135                 spin_unlock(&workspace_lock);
 136                 if (waitqueue_active(&workspace_wait))
 137                         wake_up(&workspace_wait);
 138                 return 0;
 139         }
 140         spin_unlock(&workspace_lock);
 141         vfree(workspace->def_strm.workspace);
 142         vfree(workspace->inf_strm.workspace);
 143         kfree(workspace->buf);
 144         kfree(workspace);
 145
 146         atomic_dec(&alloc_workspace);
 147         if (waitqueue_active(&workspace_wait))
 148                 wake_up(&workspace_wait);
 149         return 0;
 150 }
 151
 152 /*
 153  * cleanup function for module exit
 154  */
 155 static void free_workspaces(void)
 156 {
 157         struct workspace *workspace;
 158         while (!list_empty(&idle_workspace)) {
 159                 workspace = list_entry(idle_workspace.next, struct workspace,
 160                                        list);
 161                 list_del(&workspace->list);
 162                 vfree(workspace->def_strm.workspace);
 163                 vfree(workspace->inf_strm.workspace);
 164                 kfree(workspace->buf);
 165                 kfree(workspace);
 166                 atomic_dec(&alloc_workspace);
 167         }
 168 }
 169
 170 /*
 171  * given an address space and start/len, compress the bytes.
 172  *
 173  * pages are allocated to hold the compressed result and stored
 174  * in 'pages'
 175  *
 176  * out_pages is used to return the number of pages allocated.  There
 177  * may be pages allocated even if we return an error
 178  *
 179  * total_in is used to return the number of bytes actually read.  It
 180  * may be smaller then len if we had to exit early because we
 181  * ran out of room in the pages array or because we cross the
 182  * max_out threshold.
 183  *
 184  * total_out is used to return the total number of compressed bytes
 185  *
 186  * max_out tells us the max number of bytes that we're allowed to
 187  * stuff into pages
 188  */
 189 int btrfs_zlib_compress_pages(struct address_space *mapping,
 190                               u64 start, unsigned long len,
 191                               struct page **pages,
 192                               unsigned long nr_dest_pages,
 193                               unsigned long *out_pages,
 194                               unsigned long *total_in,
 195                               unsigned long *total_out,
 196                               unsigned long max_out)
 197 {
 198         int ret;
 199         struct workspace *workspace;
 200         char *data_in;
 201         char *cpage_out;
 202         int nr_pages = 0;
 203         struct page *in_page = NULL;
 204         struct page *out_page = NULL;
 205         unsigned long bytes_left;
 206
 207         *out_pages = 0;
 208         *total_out = 0;
 209         *total_in = 0;
 210
 211         workspace = find_zlib_workspace();
 212         if (IS_ERR(workspace))
 213                 return -1;
 214
 215         if (Z_OK != zlib_deflateInit(&workspace->def_strm, 3)) {
 216                 printk(KERN_WARNING "deflateInit failed\n");
 217                 ret = -1;
 218                 goto out;
 219         }
 220
 221         workspace->def_strm.total_in = 0;
 222         workspace->def_strm.total_out = 0;
 223
 224         in_page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
 225         data_in = kmap(in_page);
 226
 227         out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
 228         if (out_page == NULL) {
 229                 ret = -1;
 230                 goto out;
 231         }
 232         cpage_out = kmap(out_page);
 233         pages[0] = out_page;
 234         nr_pages = 1;
 235
 236         workspace->def_strm.next_in = data_in;
 237         workspace->def_strm.next_out = cpage_out;
 238         workspace->def_strm.avail_out = PAGE_CACHE_SIZE;
 239         workspace->def_strm.avail_in = min(len, PAGE_CACHE_SIZE);
 240
 241         while (workspace->def_strm.total_in < len) {
 242                 ret = zlib_deflate(&workspace->def_strm, Z_SYNC_FLUSH);
 243                 if (ret != Z_OK) {
 244                         printk(KERN_DEBUG "btrfs deflate in loop returned %d\n",
 245                                ret);
 246                         zlib_deflateEnd(&workspace->def_strm);
 247                         ret = -1;
 248                         goto out;
 249                 }
 250
 251                 /* we're making it bigger, give up */
 252                 if (workspace->def_strm.total_in > 8192 &&
 253                     workspace->def_strm.total_in <
 254                     workspace->def_strm.total_out) {
 255                         ret = -1;
 256                         goto out;
 257                 }
 258                 /* we need another page for writing out.  Test this
 259                  * before the total_in so we will pull in a new page for
 260                  * the stream end if required
 261                  */
 262                 if (workspace->def_strm.avail_out == 0) {
 263                         kunmap(out_page);
 264                         if (nr_pages == nr_dest_pages) {
 265                                 out_page = NULL;
 266                                 ret = -1;
 267                                 goto out;
 268                         }
 269                         out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
 270                         if (out_page == NULL) {
 271                                 ret = -1;
 272                                 goto out;
 273                         }
 274                         cpage_out = kmap(out_page);
 275                         pages[nr_pages] = out_page;
 276                         nr_pages++;
 277                         workspace->def_strm.avail_out = PAGE_CACHE_SIZE;
 278                         workspace->def_strm.next_out = cpage_out;
 279                 }
 280                 /* we're all done */
 281                 if (workspace->def_strm.total_in >= len)
 282                         break;
 283
 284                 /* we've read in a full page, get a new one */
 285                 if (workspace->def_strm.avail_in == 0) {
 286                         if (workspace->def_strm.total_out > max_out)
 287                                 break;
 288
 289                         bytes_left = len - workspace->def_strm.total_in;
 290                         kunmap(in_page);
 291                         page_cache_release(in_page);
 292
 293                         start += PAGE_CACHE_SIZE;
 294                         in_page = find_get_page(mapping,
 295                                                 start >> PAGE_CACHE_SHIFT);
 296                         data_in = kmap(in_page);
 297                         workspace->def_strm.avail_in = min(bytes_left,
 298                                                            PAGE_CACHE_SIZE);
 299                         workspace->def_strm.next_in = data_in;
 300                 }
 301         }
 302         workspace->def_strm.avail_in = 0;
 303         ret = zlib_deflate(&workspace->def_strm, Z_FINISH);
 304         zlib_deflateEnd(&workspace->def_strm);
 305
 306         if (ret != Z_STREAM_END) {
 307                 ret = -1;
 308                 goto out;
 309         }
 310
 311         if (workspace->def_strm.total_out >= workspace->def_strm.total_in) {
 312                 ret = -1;
 313                 goto out;
 314         }
 315
 316         ret = 0;
 317         *total_out = workspace->def_strm.total_out;
 318         *total_in = workspace->def_strm.total_in;
 319 out:
 320         *out_pages = nr_pages;
 321         if (out_page)
 322                 kunmap(out_page);
 323
 324         if (in_page) {
 325                 kunmap(in_page);
 326                 page_cache_release(in_page);
 327         }
 328         free_workspace(workspace);
 329         return ret;
 330 }
 331
 332 /*
 333  * pages_in is an array of pages with compressed data.
 334  *
 335  * disk_start is the starting logical offset of this array in the file
 336  *
 337  * bvec is a bio_vec of pages from the file that we want to decompress into
 338  *
 339  * vcnt is the count of pages in the biovec
 340  *
 341  * srclen is the number of bytes in pages_in
 342  *
 343  * The basic idea is that we have a bio that was created by readpages.
 344  * The pages in the bio are for the uncompressed data, and they may not
 345  * be contiguous.  They all correspond to the range of bytes covered by
 346  * the compressed extent.
 347  */
 348 int btrfs_zlib_decompress_biovec(struct page **pages_in,
 349                               u64 disk_start,
 350                               struct bio_vec *bvec,
 351                               int vcnt,
 352                               size_t srclen)
 353 {
 354         int ret = 0;
 355         int wbits = MAX_WBITS;
 356         struct workspace *workspace;
 357         char *data_in;
 358         size_t total_out = 0;
 359         unsigned long page_bytes_left;
 360         unsigned long page_in_index = 0;
 361         unsigned long page_out_index = 0;
 362         struct page *page_out;
 363         unsigned long total_pages_in = (srclen + PAGE_CACHE_SIZE - 1) /
 364                                         PAGE_CACHE_SIZE;
 365         unsigned long buf_start;
 366         unsigned long buf_offset;
 367         unsigned long bytes;
 368         unsigned long working_bytes;
 369         unsigned long pg_offset;
 370         unsigned long start_byte;
 371         unsigned long current_buf_start;
 372         char *kaddr;
 373
 374         workspace = find_zlib_workspace();
 375         if (IS_ERR(workspace))
 376                 return -ENOMEM;
 377
 378         data_in = kmap(pages_in[page_in_index]);
 379         workspace->inf_strm.next_in = data_in;
 380         workspace->inf_strm.avail_in = min_t(size_t, srclen, PAGE_CACHE_SIZE);
 381         workspace->inf_strm.total_in = 0;
 382
 383         workspace->inf_strm.total_out = 0;
 384         workspace->inf_strm.next_out = workspace->buf;
 385         workspace->inf_strm.avail_out = PAGE_CACHE_SIZE;
 386         page_out = bvec[page_out_index].bv_page;
 387         page_bytes_left = PAGE_CACHE_SIZE;
 388         pg_offset = 0;
 389
 390         /* If it's deflate, and it's got no preset dictionary, then
 391            we can tell zlib to skip the adler32 check. */
 392         if (srclen > 2 && !(data_in[1] & PRESET_DICT) &&
 393             ((data_in[0] & 0x0f) == Z_DEFLATED) &&
 394             !(((data_in[0]<<8) + data_in[1]) % 31)) {
 395
 396                 wbits = -((data_in[0] >> 4) + 8);
 397                 workspace->inf_strm.next_in += 2;
 398                 workspace->inf_strm.avail_in -= 2;
 399         }
 400
 401         if (Z_OK != zlib_inflateInit2(&workspace->inf_strm, wbits)) {
 402                 printk(KERN_WARNING "inflateInit failed\n");
 403                 ret = -1;
 404                 goto out;
 405         }
 406         while (workspace->inf_strm.total_in < srclen) {
 407                 ret = zlib_inflate(&workspace->inf_strm, Z_NO_FLUSH);
 408                 if (ret != Z_OK && ret != Z_STREAM_END)
 409                         break;
 410                 /*
 411                  * buf start is the byte offset we're of the start of
 412                  * our workspace buffer
 413                  */
 414                 buf_start = total_out;
 415
 416                 /* total_out is the last byte of the workspace buffer */
 417                 total_out = workspace->inf_strm.total_out;
 418
 419                 working_bytes = total_out - buf_start;
 420
 421                 /*
 422                  * start byte is the first byte of the page we're currently
 423                  * copying into relative to the start of the compressed data.
 424                  */
 425                 start_byte = page_offset(page_out) - disk_start;
 426
 427                 if (working_bytes == 0) {
 428                         /* we didn't make progress in this inflate
 429                          * call, we're done
 430                          */
 431                         if (ret != Z_STREAM_END)
 432                                 ret = -1;
 433                         break;
 434                 }
 435
 436                 /* we haven't yet hit data corresponding to this page */
 437                 if (total_out <= start_byte)
 438                         goto next;
 439
 440                 /*
 441                  * the start of the data we care about is offset into
 442                  * the middle of our working buffer
 443                  */
 444                 if (total_out > start_byte && buf_start < start_byte) {
 445                         buf_offset = start_byte - buf_start;
 446                         working_bytes -= buf_offset;
 447                 } else {
 448                         buf_offset = 0;
 449                 }
 450                 current_buf_start = buf_start;
 451
 452                 /* copy bytes from the working buffer into the pages */
 453                 while (working_bytes > 0) {
 454                         bytes = min(PAGE_CACHE_SIZE - pg_offset,
 455                                     PAGE_CACHE_SIZE - buf_offset);
 456                         bytes = min(bytes, working_bytes);
 457                         kaddr = kmap_atomic(page_out, KM_USER0);
 458                         memcpy(kaddr + pg_offset, workspace->buf + buf_offset,
 459                                bytes);
 460                         kunmap_atomic(kaddr, KM_USER0);
 461                         flush_dcache_page(page_out);
 462
 463                         pg_offset += bytes;
 464                         page_bytes_left -= bytes;
 465                         buf_offset += bytes;
 466                         working_bytes -= bytes;
 467                         current_buf_start += bytes;
 468
 469                         /* check if we need to pick another page */
 470                         if (page_bytes_left == 0) {
 471                                 page_out_index++;
 472                                 if (page_out_index >= vcnt) {
 473                                         ret = 0;
 474                                         goto done;
 475                                 }
 476
 477                                 page_out = bvec[page_out_index].bv_page;
 478                                 pg_offset = 0;
 479                                 page_bytes_left = PAGE_CACHE_SIZE;
 480                                 start_byte = page_offset(page_out) - disk_start;
 481
 482                                 /*
 483                                  * make sure our new page is covered by this
 484                                  * working buffer
 485                                  */
 486                                 if (total_out <= start_byte)
 487                                         goto next;
 488
 489                                 /* the next page in the biovec might not
 490                                  * be adjacent to the last page, but it
 491                                  * might still be found inside this working
 492                                  * buffer.  bump our offset pointer
 493                                  */
 494                                 if (total_out > start_byte &&
 495                                     current_buf_start < start_byte) {
 496                                         buf_offset = start_byte - buf_start;
 497                                         working_bytes = total_out - start_byte;
 498                                         current_buf_start = buf_start +
 499                                                 buf_offset;
 500                                 }
 501                         }
 502                 }
 503 next:
 504                 workspace->inf_strm.next_out = workspace->buf;
 505                 workspace->inf_strm.avail_out = PAGE_CACHE_SIZE;
 506
 507                 if (workspace->inf_strm.avail_in == 0) {
 508                         unsigned long tmp;
 509                         kunmap(pages_in[page_in_index]);
 510                         page_in_index++;
 511                         if (page_in_index >= total_pages_in) {
 512                                 data_in = NULL;
 513                                 break;
 514                         }
 515                         data_in = kmap(pages_in[page_in_index]);
 516                         workspace->inf_strm.next_in = data_in;
 517                         tmp = srclen - workspace->inf_strm.total_in;
 518                         workspace->inf_strm.avail_in = min(tmp,
 519                                                            PAGE_CACHE_SIZE);
 520                 }
 521         }
 522         if (ret != Z_STREAM_END)
 523                 ret = -1;
 524         else
 525                 ret = 0;
 526 done:
 527         zlib_inflateEnd(&workspace->inf_strm);
 528         if (data_in)
 529                 kunmap(pages_in[page_in_index]);
 530 out:
 531         free_workspace(workspace);
 532         return ret;
 533 }
 534
 535 /*
 536  * a less complex decompression routine.  Our compressed data fits in a
 537  * single page, and we want to read a single page out of it.
 538  * start_byte tells us the offset into the compressed data we're interested in
 539  */
 540 int btrfs_zlib_decompress(unsigned char *data_in,
 541                           struct page *dest_page,
 542                           unsigned long start_byte,
 543                           size_t srclen, size_t destlen)
 544 {
 545         int ret = 0;
 546         int wbits = MAX_WBITS;
 547         struct workspace *workspace;
 548         unsigned long bytes_left = destlen;
 549         unsigned long total_out = 0;
 550         char *kaddr;
 551
 552         if (destlen > PAGE_CACHE_SIZE)
 553                 return -ENOMEM;
 554
 555         workspace = find_zlib_workspace();
 556         if (IS_ERR(workspace))
 557                 return -ENOMEM;
 558
 559         workspace->inf_strm.next_in = data_in;
 560         workspace->inf_strm.avail_in = srclen;
 561         workspace->inf_strm.total_in = 0;
 562
 563         workspace->inf_strm.next_out = workspace->buf;
 564         workspace->inf_strm.avail_out = PAGE_CACHE_SIZE;
 565         workspace->inf_strm.total_out = 0;
 566         /* If it's deflate, and it's got no preset dictionary, then
 567            we can tell zlib to skip the adler32 check. */
 568         if (srclen > 2 && !(data_in[1] & PRESET_DICT) &&
 569             ((data_in[0] & 0x0f) == Z_DEFLATED) &&
 570             !(((data_in[0]<<8) + data_in[1]) % 31)) {
 571
 572                 wbits = -((data_in[0] >> 4) + 8);
 573                 workspace->inf_strm.next_in += 2;
 574                 workspace->inf_strm.avail_in -= 2;
 575         }
 576
 577         if (Z_OK != zlib_inflateInit2(&workspace->inf_strm, wbits)) {
 578                 printk(KERN_WARNING "inflateInit failed\n");
 579                 ret = -1;
 580                 goto out;
 581         }
 582
 583         while (bytes_left > 0) {
 584                 unsigned long buf_start;
 585                 unsigned long buf_offset;
 586                 unsigned long bytes;
 587                 unsigned long pg_offset = 0;
 588
 589                 ret = zlib_inflate(&workspace->inf_strm, Z_NO_FLUSH);
 590                 if (ret != Z_OK && ret != Z_STREAM_END)
 591                         break;
 592
 593                 buf_start = total_out;
 594                 total_out = workspace->inf_strm.total_out;
 595
 596                 if (total_out == buf_start) {
 597                         ret = -1;
 598                         break;
 599                 }
 600
 601                 if (total_out <= start_byte)
 602                         goto next;
 603
 604                 if (total_out > start_byte && buf_start < start_byte)
 605                         buf_offset = start_byte - buf_start;
 606                 else
 607                         buf_offset = 0;
 608
 609                 bytes = min(PAGE_CACHE_SIZE - pg_offset,
 610                             PAGE_CACHE_SIZE - buf_offset);
 611                 bytes = min(bytes, bytes_left);
 612
 613                 kaddr = kmap_atomic(dest_page, KM_USER0);
 614                 memcpy(kaddr + pg_offset, workspace->buf + buf_offset, bytes);
 615                 kunmap_atomic(kaddr, KM_USER0);
 616
 617                 pg_offset += bytes;
 618                 bytes_left -= bytes;
 619 next:
 620                 workspace->inf_strm.next_out = workspace->buf;
 621                 workspace->inf_strm.avail_out = PAGE_CACHE_SIZE;
 622         }
 623
 624         if (ret != Z_STREAM_END && bytes_left != 0)
 625                 ret = -1;
 626         else
 627                 ret = 0;
 628
 629         zlib_inflateEnd(&workspace->inf_strm);
 630 out:
 631         free_workspace(workspace);
 632         return ret;
 633 }
 634
 635 void btrfs_zlib_exit(void)
 636 {
 637     free_workspaces();
 638 }