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