fs/xfs/xfs_iomap.c

   1 /*
   2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_log.h"
  21 #include "xfs_trans.h"
  22 #include "xfs_sb.h"
  23 #include "xfs_ag.h"
  24 #include "xfs_alloc.h"
  25 #include "xfs_quota.h"
  26 #include "xfs_mount.h"
  27 #include "xfs_bmap_btree.h"
  28 #include "xfs_alloc_btree.h"
  29 #include "xfs_ialloc_btree.h"
  30 #include "xfs_dinode.h"
  31 #include "xfs_inode.h"
  32 #include "xfs_inode_item.h"
  33 #include "xfs_btree.h"
  34 #include "xfs_bmap.h"
  35 #include "xfs_rtalloc.h"
  36 #include "xfs_error.h"
  37 #include "xfs_itable.h"
  38 #include "xfs_attr.h"
  39 #include "xfs_buf_item.h"
  40 #include "xfs_trans_space.h"
  41 #include "xfs_utils.h"
  42 #include "xfs_iomap.h"
  43 #include "xfs_trace.h"
  44
  45
  46 #define XFS_WRITEIO_ALIGN(mp,off)       (((off) >> mp->m_writeio_log) \
  47                                                 << mp->m_writeio_log)
  48 #define XFS_WRITE_IMAPS         XFS_BMAP_MAX_NMAP
  49
  50 STATIC int
  51 xfs_iomap_eof_align_last_fsb(
  52         xfs_mount_t     *mp,
  53         xfs_inode_t     *ip,
  54         xfs_extlen_t    extsize,
  55         xfs_fileoff_t   *last_fsb)
  56 {
  57         xfs_fileoff_t   new_last_fsb = 0;
  58         xfs_extlen_t    align = 0;
  59         int             eof, error;
  60
  61         if (!XFS_IS_REALTIME_INODE(ip)) {
  62                 /*
  63                  * Round up the allocation request to a stripe unit
  64                  * (m_dalign) boundary if the file size is >= stripe unit
  65                  * size, and we are allocating past the allocation eof.
  66                  *
  67                  * If mounted with the "-o swalloc" option the alignment is
  68                  * increased from the strip unit size to the stripe width.
  69                  */
  70                 if (mp->m_swidth && (mp->m_flags & XFS_MOUNT_SWALLOC))
  71                         align = mp->m_swidth;
  72                 else if (mp->m_dalign)
  73                         align = mp->m_dalign;
  74
  75                 if (align && XFS_ISIZE(ip) >= XFS_FSB_TO_B(mp, align))
  76                         new_last_fsb = roundup_64(*last_fsb, align);
  77         }
  78
  79         /*
  80          * Always round up the allocation request to an extent boundary
  81          * (when file on a real-time subvolume or has di_extsize hint).
  82          */
  83         if (extsize) {
  84                 if (new_last_fsb)
  85                         align = roundup_64(new_last_fsb, extsize);
  86                 else
  87                         align = extsize;
  88                 new_last_fsb = roundup_64(*last_fsb, align);
  89         }
  90
  91         if (new_last_fsb) {
  92                 error = xfs_bmap_eof(ip, new_last_fsb, XFS_DATA_FORK, &eof);
  93                 if (error)
  94                         return error;
  95                 if (eof)
  96                         *last_fsb = new_last_fsb;
  97         }
  98         return 0;
  99 }
 100
 101 STATIC int
 102 xfs_alert_fsblock_zero(
 103         xfs_inode_t     *ip,
 104         xfs_bmbt_irec_t *imap)
 105 {
 106         xfs_alert_tag(ip->i_mount, XFS_PTAG_FSBLOCK_ZERO,
 107                         "Access to block zero in inode %llu "
 108                         "start_block: %llx start_off: %llx "
 109                         "blkcnt: %llx extent-state: %x\n",
 110                 (unsigned long long)ip->i_ino,
 111                 (unsigned long long)imap->br_startblock,
 112                 (unsigned long long)imap->br_startoff,
 113                 (unsigned long long)imap->br_blockcount,
 114                 imap->br_state);
 115         return EFSCORRUPTED;
 116 }
 117
 118 int
 119 xfs_iomap_write_direct(
 120         xfs_inode_t     *ip,
 121         xfs_off_t       offset,
 122         size_t          count,
 123         xfs_bmbt_irec_t *imap,
 124         int             nmaps)
 125 {
 126         xfs_mount_t     *mp = ip->i_mount;
 127         xfs_fileoff_t   offset_fsb;
 128         xfs_fileoff_t   last_fsb;
 129         xfs_filblks_t   count_fsb, resaligned;
 130         xfs_fsblock_t   firstfsb;
 131         xfs_extlen_t    extsz, temp;
 132         int             nimaps;
 133         int             bmapi_flag;
 134         int             quota_flag;
 135         int             rt;
 136         xfs_trans_t     *tp;
 137         xfs_bmap_free_t free_list;
 138         uint            qblocks, resblks, resrtextents;
 139         int             committed;
 140         int             error;
 141
 142         error = xfs_qm_dqattach(ip, 0);
 143         if (error)
 144                 return XFS_ERROR(error);
 145
 146         rt = XFS_IS_REALTIME_INODE(ip);
 147         extsz = xfs_get_extsz_hint(ip);
 148
 149         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 150         last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
 151         if ((offset + count) > XFS_ISIZE(ip)) {
 152                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
 153                 if (error)
 154                         return XFS_ERROR(error);
 155         } else {
 156                 if (nmaps && (imap->br_startblock == HOLESTARTBLOCK))
 157                         last_fsb = MIN(last_fsb, (xfs_fileoff_t)
 158                                         imap->br_blockcount +
 159                                         imap->br_startoff);
 160         }
 161         count_fsb = last_fsb - offset_fsb;
 162         ASSERT(count_fsb > 0);
 163
 164         resaligned = count_fsb;
 165         if (unlikely(extsz)) {
 166                 if ((temp = do_mod(offset_fsb, extsz)))
 167                         resaligned += temp;
 168                 if ((temp = do_mod(resaligned, extsz)))
 169                         resaligned += extsz - temp;
 170         }
 171
 172         if (unlikely(rt)) {
 173                 resrtextents = qblocks = resaligned;
 174                 resrtextents /= mp->m_sb.sb_rextsize;
 175                 resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
 176                 quota_flag = XFS_QMOPT_RES_RTBLKS;
 177         } else {
 178                 resrtextents = 0;
 179                 resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
 180                 quota_flag = XFS_QMOPT_RES_REGBLKS;
 181         }
 182
 183         /*
 184          * Allocate and setup the transaction
 185          */
 186         tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
 187         error = xfs_trans_reserve(tp, resblks,
 188                         XFS_WRITE_LOG_RES(mp), resrtextents,
 189                         XFS_TRANS_PERM_LOG_RES,
 190                         XFS_WRITE_LOG_COUNT);
 191         /*
 192          * Check for running out of space, note: need lock to return
 193          */
 194         if (error) {
 195                 xfs_trans_cancel(tp, 0);
 196                 return XFS_ERROR(error);
 197         }
 198
 199         xfs_ilock(ip, XFS_ILOCK_EXCL);
 200
 201         error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
 202         if (error)
 203                 goto out_trans_cancel;
 204
 205         xfs_trans_ijoin(tp, ip, 0);
 206
 207         bmapi_flag = 0;
 208         if (offset < XFS_ISIZE(ip) || extsz)
 209                 bmapi_flag |= XFS_BMAPI_PREALLOC;
 210
 211         /*
 212          * From this point onwards we overwrite the imap pointer that the
 213          * caller gave to us.
 214          */
 215         xfs_bmap_init(&free_list, &firstfsb);
 216         nimaps = 1;
 217         error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb, bmapi_flag,
 218                                 &firstfsb, 0, imap, &nimaps, &free_list);
 219         if (error)
 220                 goto out_bmap_cancel;
 221
 222         /*
 223          * Complete the transaction
 224          */
 225         error = xfs_bmap_finish(&tp, &free_list, &committed);
 226         if (error)
 227                 goto out_bmap_cancel;
 228         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 229         if (error)
 230                 goto out_unlock;
 231
 232         /*
 233          * Copy any maps to caller's array and return any error.
 234          */
 235         if (nimaps == 0) {
 236                 error = XFS_ERROR(ENOSPC);
 237                 goto out_unlock;
 238         }
 239
 240         if (!(imap->br_startblock || XFS_IS_REALTIME_INODE(ip)))
 241                 error = xfs_alert_fsblock_zero(ip, imap);
 242
 243 out_unlock:
 244         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 245         return error;
 246
 247 out_bmap_cancel:
 248         xfs_bmap_cancel(&free_list);
 249         xfs_trans_unreserve_quota_nblks(tp, ip, (long)qblocks, 0, quota_flag);
 250 out_trans_cancel:
 251         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 252         goto out_unlock;
 253 }
 254
 255 /*
 256  * If the caller is doing a write at the end of the file, then extend the
 257  * allocation out to the file system's write iosize.  We clean up any extra
 258  * space left over when the file is closed in xfs_inactive().
 259  *
 260  * If we find we already have delalloc preallocation beyond EOF, don't do more
 261  * preallocation as it it not needed.
 262  */
 263 STATIC int
 264 xfs_iomap_eof_want_preallocate(
 265         xfs_mount_t     *mp,
 266         xfs_inode_t     *ip,
 267         xfs_off_t       offset,
 268         size_t          count,
 269         xfs_bmbt_irec_t *imap,
 270         int             nimaps,
 271         int             *prealloc)
 272 {
 273         xfs_fileoff_t   start_fsb;
 274         xfs_filblks_t   count_fsb;
 275         xfs_fsblock_t   firstblock;
 276         int             n, error, imaps;
 277         int             found_delalloc = 0;
 278
 279         *prealloc = 0;
 280         if (offset + count <= XFS_ISIZE(ip))
 281                 return 0;
 282
 283         /*
 284          * If there are any real blocks past eof, then don't
 285          * do any speculative allocation.
 286          */
 287         start_fsb = XFS_B_TO_FSBT(mp, ((xfs_ufsize_t)(offset + count - 1)));
 288         count_fsb = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
 289         while (count_fsb > 0) {
 290                 imaps = nimaps;
 291                 firstblock = NULLFSBLOCK;
 292                 error = xfs_bmapi_read(ip, start_fsb, count_fsb, imap, &imaps,
 293                                        0);
 294                 if (error)
 295                         return error;
 296                 for (n = 0; n < imaps; n++) {
 297                         if ((imap[n].br_startblock != HOLESTARTBLOCK) &&
 298                             (imap[n].br_startblock != DELAYSTARTBLOCK))
 299                                 return 0;
 300                         start_fsb += imap[n].br_blockcount;
 301                         count_fsb -= imap[n].br_blockcount;
 302
 303                         if (imap[n].br_startblock == DELAYSTARTBLOCK)
 304                                 found_delalloc = 1;
 305                 }
 306         }
 307         if (!found_delalloc)
 308                 *prealloc = 1;
 309         return 0;
 310 }
 311
 312 /*
 313  * If we don't have a user specified preallocation size, dynamically increase
 314  * the preallocation size as the size of the file grows. Cap the maximum size
 315  * at a single extent or less if the filesystem is near full. The closer the
 316  * filesystem is to full, the smaller the maximum prealocation.
 317  */
 318 STATIC xfs_fsblock_t
 319 xfs_iomap_prealloc_size(
 320         struct xfs_mount        *mp,
 321         struct xfs_inode        *ip)
 322 {
 323         xfs_fsblock_t           alloc_blocks = 0;
 324
 325         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
 326                 int shift = 0;
 327                 int64_t freesp;
 328
 329                 /*
 330                  * rounddown_pow_of_two() returns an undefined result
 331                  * if we pass in alloc_blocks = 0. Hence the "+ 1" to
 332                  * ensure we always pass in a non-zero value.
 333                  */
 334                 alloc_blocks = XFS_B_TO_FSB(mp, XFS_ISIZE(ip)) + 1;
 335                 alloc_blocks = XFS_FILEOFF_MIN(MAXEXTLEN,
 336                                         rounddown_pow_of_two(alloc_blocks));
 337
 338                 xfs_icsb_sync_counters(mp, XFS_ICSB_LAZY_COUNT);
 339                 freesp = mp->m_sb.sb_fdblocks;
 340                 if (freesp < mp->m_low_space[XFS_LOWSP_5_PCNT]) {
 341                         shift = 2;
 342                         if (freesp < mp->m_low_space[XFS_LOWSP_4_PCNT])
 343                                 shift++;
 344                         if (freesp < mp->m_low_space[XFS_LOWSP_3_PCNT])
 345                                 shift++;
 346                         if (freesp < mp->m_low_space[XFS_LOWSP_2_PCNT])
 347                                 shift++;
 348                         if (freesp < mp->m_low_space[XFS_LOWSP_1_PCNT])
 349                                 shift++;
 350                 }
 351                 if (shift)
 352                         alloc_blocks >>= shift;
 353         }
 354
 355         if (alloc_blocks < mp->m_writeio_blocks)
 356                 alloc_blocks = mp->m_writeio_blocks;
 357
 358         return alloc_blocks;
 359 }
 360
 361 int
 362 xfs_iomap_write_delay(
 363         xfs_inode_t     *ip,
 364         xfs_off_t       offset,
 365         size_t          count,
 366         xfs_bmbt_irec_t *ret_imap)
 367 {
 368         xfs_mount_t     *mp = ip->i_mount;
 369         xfs_fileoff_t   offset_fsb;
 370         xfs_fileoff_t   last_fsb;
 371         xfs_off_t       aligned_offset;
 372         xfs_fileoff_t   ioalign;
 373         xfs_extlen_t    extsz;
 374         int             nimaps;
 375         xfs_bmbt_irec_t imap[XFS_WRITE_IMAPS];
 376         int             prealloc;
 377         int             error;
 378
 379         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 380
 381         /*
 382          * Make sure that the dquots are there. This doesn't hold
 383          * the ilock across a disk read.
 384          */
 385         error = xfs_qm_dqattach_locked(ip, 0);
 386         if (error)
 387                 return XFS_ERROR(error);
 388
 389         extsz = xfs_get_extsz_hint(ip);
 390         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 391
 392
 393         error = xfs_iomap_eof_want_preallocate(mp, ip, offset, count,
 394                                 imap, XFS_WRITE_IMAPS, &prealloc);
 395         if (error)
 396                 return error;
 397
 398 retry:
 399         if (prealloc) {
 400                 xfs_fsblock_t   alloc_blocks = xfs_iomap_prealloc_size(mp, ip);
 401
 402                 aligned_offset = XFS_WRITEIO_ALIGN(mp, (offset + count - 1));
 403                 ioalign = XFS_B_TO_FSBT(mp, aligned_offset);
 404                 last_fsb = ioalign + alloc_blocks;
 405         } else {
 406                 last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
 407         }
 408
 409         if (prealloc || extsz) {
 410                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
 411                 if (error)
 412                         return error;
 413         }
 414
 415         /*
 416          * Make sure preallocation does not create extents beyond the range we
 417          * actually support in this filesystem.
 418          */
 419         if (last_fsb > XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes))
 420                 last_fsb = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
 421
 422         ASSERT(last_fsb > offset_fsb);
 423
 424         nimaps = XFS_WRITE_IMAPS;
 425         error = xfs_bmapi_delay(ip, offset_fsb, last_fsb - offset_fsb,
 426                                 imap, &nimaps, XFS_BMAPI_ENTIRE);
 427         switch (error) {
 428         case 0:
 429         case ENOSPC:
 430         case EDQUOT:
 431                 break;
 432         default:
 433                 return XFS_ERROR(error);
 434         }
 435
 436         /*
 437          * If bmapi returned us nothing, we got either ENOSPC or EDQUOT. Retry
 438          * without EOF preallocation.
 439          */
 440         if (nimaps == 0) {
 441                 trace_xfs_delalloc_enospc(ip, offset, count);
 442                 if (prealloc) {
 443                         prealloc = 0;
 444                         error = 0;
 445                         goto retry;
 446                 }
 447                 return XFS_ERROR(error ? error : ENOSPC);
 448         }
 449
 450         if (!(imap[0].br_startblock || XFS_IS_REALTIME_INODE(ip)))
 451                 return xfs_alert_fsblock_zero(ip, &imap[0]);
 452
 453         *ret_imap = imap[0];
 454         return 0;
 455 }
 456
 457 /*
 458  * Pass in a delayed allocate extent, convert it to real extents;
 459  * return to the caller the extent we create which maps on top of
 460  * the originating callers request.
 461  *
 462  * Called without a lock on the inode.
 463  *
 464  * We no longer bother to look at the incoming map - all we have to
 465  * guarantee is that whatever we allocate fills the required range.
 466  */
 467 int
 468 xfs_iomap_write_allocate(
 469         xfs_inode_t     *ip,
 470         xfs_off_t       offset,
 471         size_t          count,
 472         xfs_bmbt_irec_t *imap)
 473 {
 474         xfs_mount_t     *mp = ip->i_mount;
 475         xfs_fileoff_t   offset_fsb, last_block;
 476         xfs_fileoff_t   end_fsb, map_start_fsb;
 477         xfs_fsblock_t   first_block;
 478         xfs_bmap_free_t free_list;
 479         xfs_filblks_t   count_fsb;
 480         xfs_trans_t     *tp;
 481         int             nimaps, committed;
 482         int             error = 0;
 483         int             nres;
 484
 485         /*
 486          * Make sure that the dquots are there.
 487          */
 488         error = xfs_qm_dqattach(ip, 0);
 489         if (error)
 490                 return XFS_ERROR(error);
 491
 492         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 493         count_fsb = imap->br_blockcount;
 494         map_start_fsb = imap->br_startoff;
 495
 496         XFS_STATS_ADD(xs_xstrat_bytes, XFS_FSB_TO_B(mp, count_fsb));
 497
 498         while (count_fsb != 0) {
 499                 /*
 500                  * Set up a transaction with which to allocate the
 501                  * backing store for the file.  Do allocations in a
 502                  * loop until we get some space in the range we are
 503                  * interested in.  The other space that might be allocated
 504                  * is in the delayed allocation extent on which we sit
 505                  * but before our buffer starts.
 506                  */
 507
 508                 nimaps = 0;
 509                 while (nimaps == 0) {
 510                         tp = xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE);
 511                         tp->t_flags |= XFS_TRANS_RESERVE;
 512                         nres = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
 513                         error = xfs_trans_reserve(tp, nres,
 514                                         XFS_WRITE_LOG_RES(mp),
 515                                         0, XFS_TRANS_PERM_LOG_RES,
 516                                         XFS_WRITE_LOG_COUNT);
 517                         if (error) {
 518                                 xfs_trans_cancel(tp, 0);
 519                                 return XFS_ERROR(error);
 520                         }
 521                         xfs_ilock(ip, XFS_ILOCK_EXCL);
 522                         xfs_trans_ijoin(tp, ip, 0);
 523
 524                         xfs_bmap_init(&free_list, &first_block);
 525
 526                         /*
 527                          * it is possible that the extents have changed since
 528                          * we did the read call as we dropped the ilock for a
 529                          * while. We have to be careful about truncates or hole
 530                          * punchs here - we are not allowed to allocate
 531                          * non-delalloc blocks here.
 532                          *
 533                          * The only protection against truncation is the pages
 534                          * for the range we are being asked to convert are
 535                          * locked and hence a truncate will block on them
 536                          * first.
 537                          *
 538                          * As a result, if we go beyond the range we really
 539                          * need and hit an delalloc extent boundary followed by
 540                          * a hole while we have excess blocks in the map, we
 541                          * will fill the hole incorrectly and overrun the
 542                          * transaction reservation.
 543                          *
 544                          * Using a single map prevents this as we are forced to
 545                          * check each map we look for overlap with the desired
 546                          * range and abort as soon as we find it. Also, given
 547                          * that we only return a single map, having one beyond
 548                          * what we can return is probably a bit silly.
 549                          *
 550                          * We also need to check that we don't go beyond EOF;
 551                          * this is a truncate optimisation as a truncate sets
 552                          * the new file size before block on the pages we
 553                          * currently have locked under writeback. Because they
 554                          * are about to be tossed, we don't need to write them
 555                          * back....
 556                          */
 557                         nimaps = 1;
 558                         end_fsb = XFS_B_TO_FSB(mp, XFS_ISIZE(ip));
 559                         error = xfs_bmap_last_offset(NULL, ip, &last_block,
 560                                                         XFS_DATA_FORK);
 561                         if (error)
 562                                 goto trans_cancel;
 563
 564                         last_block = XFS_FILEOFF_MAX(last_block, end_fsb);
 565                         if ((map_start_fsb + count_fsb) > last_block) {
 566                                 count_fsb = last_block - map_start_fsb;
 567                                 if (count_fsb == 0) {
 568                                         error = EAGAIN;
 569                                         goto trans_cancel;
 570                                 }
 571                         }
 572
 573                         /*
 574                          * From this point onwards we overwrite the imap
 575                          * pointer that the caller gave to us.
 576                          */
 577                         error = xfs_bmapi_write(tp, ip, map_start_fsb,
 578                                                 count_fsb,
 579                                                 XFS_BMAPI_STACK_SWITCH,
 580                                                 &first_block, 1,
 581                                                 imap, &nimaps, &free_list);
 582                         if (error)
 583                                 goto trans_cancel;
 584
 585                         error = xfs_bmap_finish(&tp, &free_list, &committed);
 586                         if (error)
 587                                 goto trans_cancel;
 588
 589                         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 590                         if (error)
 591                                 goto error0;
 592
 593                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 594                 }
 595
 596                 /*
 597                  * See if we were able to allocate an extent that
 598                  * covers at least part of the callers request
 599                  */
 600                 if (!(imap->br_startblock || XFS_IS_REALTIME_INODE(ip)))
 601                         return xfs_alert_fsblock_zero(ip, imap);
 602
 603                 if ((offset_fsb >= imap->br_startoff) &&
 604                     (offset_fsb < (imap->br_startoff +
 605                                    imap->br_blockcount))) {
 606                         XFS_STATS_INC(xs_xstrat_quick);
 607                         return 0;
 608                 }
 609
 610                 /*
 611                  * So far we have not mapped the requested part of the
 612                  * file, just surrounding data, try again.
 613                  */
 614                 count_fsb -= imap->br_blockcount;
 615                 map_start_fsb = imap->br_startoff + imap->br_blockcount;
 616         }
 617
 618 trans_cancel:
 619         xfs_bmap_cancel(&free_list);
 620         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 621 error0:
 622         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 623         return XFS_ERROR(error);
 624 }
 625
 626 int
 627 xfs_iomap_write_unwritten(
 628         xfs_inode_t     *ip,
 629         xfs_off_t       offset,
 630         size_t          count)
 631 {
 632         xfs_mount_t     *mp = ip->i_mount;
 633         xfs_fileoff_t   offset_fsb;
 634         xfs_filblks_t   count_fsb;
 635         xfs_filblks_t   numblks_fsb;
 636         xfs_fsblock_t   firstfsb;
 637         int             nimaps;
 638         xfs_trans_t     *tp;
 639         xfs_bmbt_irec_t imap;
 640         xfs_bmap_free_t free_list;
 641         xfs_fsize_t     i_size;
 642         uint            resblks;
 643         int             committed;
 644         int             error;
 645
 646         trace_xfs_unwritten_convert(ip, offset, count);
 647
 648         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 649         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
 650         count_fsb = (xfs_filblks_t)(count_fsb - offset_fsb);
 651
 652         /*
 653          * Reserve enough blocks in this transaction for two complete extent
 654          * btree splits.  We may be converting the middle part of an unwritten
 655          * extent and in this case we will insert two new extents in the btree
 656          * each of which could cause a full split.
 657          *
 658          * This reservation amount will be used in the first call to
 659          * xfs_bmbt_split() to select an AG with enough space to satisfy the
 660          * rest of the operation.
 661          */
 662         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
 663
 664         do {
 665                 /*
 666                  * set up a transaction to convert the range of extents
 667                  * from unwritten to real. Do allocations in a loop until
 668                  * we have covered the range passed in.
 669                  *
 670                  * Note that we open code the transaction allocation here
 671                  * to pass KM_NOFS--we can't risk to recursing back into
 672                  * the filesystem here as we might be asked to write out
 673                  * the same inode that we complete here and might deadlock
 674                  * on the iolock.
 675                  */
 676                 sb_start_intwrite(mp->m_super);
 677                 tp = _xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE, KM_NOFS);
 678                 tp->t_flags |= XFS_TRANS_RESERVE | XFS_TRANS_FREEZE_PROT;
 679                 error = xfs_trans_reserve(tp, resblks,
 680                                 XFS_WRITE_LOG_RES(mp), 0,
 681                                 XFS_TRANS_PERM_LOG_RES,
 682                                 XFS_WRITE_LOG_COUNT);
 683                 if (error) {
 684                         xfs_trans_cancel(tp, 0);
 685                         return XFS_ERROR(error);
 686                 }
 687
 688                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 689                 xfs_trans_ijoin(tp, ip, 0);
 690
 691                 /*
 692                  * Modify the unwritten extent state of the buffer.
 693                  */
 694                 xfs_bmap_init(&free_list, &firstfsb);
 695                 nimaps = 1;
 696                 error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb,
 697                                   XFS_BMAPI_CONVERT, &firstfsb,
 698                                   1, &imap, &nimaps, &free_list);
 699                 if (error)
 700                         goto error_on_bmapi_transaction;
 701
 702                 /*
 703                  * Log the updated inode size as we go.  We have to be careful
 704                  * to only log it up to the actual write offset if it is
 705                  * halfway into a block.
 706                  */
 707                 i_size = XFS_FSB_TO_B(mp, offset_fsb + count_fsb);
 708                 if (i_size > offset + count)
 709                         i_size = offset + count;
 710
 711                 i_size = xfs_new_eof(ip, i_size);
 712                 if (i_size) {
 713                         ip->i_d.di_size = i_size;
 714                         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 715                 }
 716
 717                 error = xfs_bmap_finish(&tp, &free_list, &committed);
 718                 if (error)
 719                         goto error_on_bmapi_transaction;
 720
 721                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 722                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 723                 if (error)
 724                         return XFS_ERROR(error);
 725
 726                 if (!(imap.br_startblock || XFS_IS_REALTIME_INODE(ip)))
 727                         return xfs_alert_fsblock_zero(ip, &imap);
 728
 729                 if ((numblks_fsb = imap.br_blockcount) == 0) {
 730                         /*
 731                          * The numblks_fsb value should always get
 732                          * smaller, otherwise the loop is stuck.
 733                          */
 734                         ASSERT(imap.br_blockcount);
 735                         break;
 736                 }
 737                 offset_fsb += numblks_fsb;
 738                 count_fsb -= numblks_fsb;
 739         } while (count_fsb > 0);
 740
 741         return 0;
 742
 743 error_on_bmapi_transaction:
 744         xfs_bmap_cancel(&free_list);
 745         xfs_trans_cancel(tp, (XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT));
 746         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 747         return XFS_ERROR(error);
 748 }