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_bit.h"
  21 #include "xfs_log.h"
  22 #include "xfs_inum.h"
  23 #include "xfs_trans.h"
  24 #include "xfs_sb.h"
  25 #include "xfs_ag.h"
  26 #include "xfs_alloc.h"
  27 #include "xfs_quota.h"
  28 #include "xfs_mount.h"
  29 #include "xfs_bmap_btree.h"
  30 #include "xfs_alloc_btree.h"
  31 #include "xfs_ialloc_btree.h"
  32 #include "xfs_dinode.h"
  33 #include "xfs_inode.h"
  34 #include "xfs_inode_item.h"
  35 #include "xfs_btree.h"
  36 #include "xfs_bmap.h"
  37 #include "xfs_rtalloc.h"
  38 #include "xfs_error.h"
  39 #include "xfs_itable.h"
  40 #include "xfs_rw.h"
  41 #include "xfs_attr.h"
  42 #include "xfs_buf_item.h"
  43 #include "xfs_trans_space.h"
  44 #include "xfs_utils.h"
  45 #include "xfs_iomap.h"
  46 #include "xfs_trace.h"
  47
  48
  49 #define XFS_WRITEIO_ALIGN(mp,off)       (((off) >> mp->m_writeio_log) \
  50                                                 << mp->m_writeio_log)
  51 #define XFS_WRITE_IMAPS         XFS_BMAP_MAX_NMAP
  52
  53 STATIC int
  54 xfs_iomap_eof_align_last_fsb(
  55         xfs_mount_t     *mp,
  56         xfs_inode_t     *ip,
  57         xfs_extlen_t    extsize,
  58         xfs_fileoff_t   *last_fsb)
  59 {
  60         xfs_fileoff_t   new_last_fsb = 0;
  61         xfs_extlen_t    align = 0;
  62         int             eof, error;
  63
  64         if (!XFS_IS_REALTIME_INODE(ip)) {
  65                 /*
  66                  * Round up the allocation request to a stripe unit
  67                  * (m_dalign) boundary if the file size is >= stripe unit
  68                  * size, and we are allocating past the allocation eof.
  69                  *
  70                  * If mounted with the "-o swalloc" option the alignment is
  71                  * increased from the strip unit size to the stripe width.
  72                  */
  73                 if (mp->m_swidth && (mp->m_flags & XFS_MOUNT_SWALLOC))
  74                         align = mp->m_swidth;
  75                 else if (mp->m_dalign)
  76                         align = mp->m_dalign;
  77
  78                 if (align && XFS_ISIZE(ip) >= XFS_FSB_TO_B(mp, align))
  79                         new_last_fsb = roundup_64(*last_fsb, align);
  80         }
  81
  82         /*
  83          * Always round up the allocation request to an extent boundary
  84          * (when file on a real-time subvolume or has di_extsize hint).
  85          */
  86         if (extsize) {
  87                 if (new_last_fsb)
  88                         align = roundup_64(new_last_fsb, extsize);
  89                 else
  90                         align = extsize;
  91                 new_last_fsb = roundup_64(*last_fsb, align);
  92         }
  93
  94         if (new_last_fsb) {
  95                 error = xfs_bmap_eof(ip, new_last_fsb, XFS_DATA_FORK, &eof);
  96                 if (error)
  97                         return error;
  98                 if (eof)
  99                         *last_fsb = new_last_fsb;
 100         }
 101         return 0;
 102 }
 103
 104 STATIC int
 105 xfs_alert_fsblock_zero(
 106         xfs_inode_t     *ip,
 107         xfs_bmbt_irec_t *imap)
 108 {
 109         xfs_alert_tag(ip->i_mount, XFS_PTAG_FSBLOCK_ZERO,
 110                         "Access to block zero in inode %llu "
 111                         "start_block: %llx start_off: %llx "
 112                         "blkcnt: %llx extent-state: %x\n",
 113                 (unsigned long long)ip->i_ino,
 114                 (unsigned long long)imap->br_startblock,
 115                 (unsigned long long)imap->br_startoff,
 116                 (unsigned long long)imap->br_blockcount,
 117                 imap->br_state);
 118         return EFSCORRUPTED;
 119 }
 120
 121 int
 122 xfs_iomap_write_direct(
 123         xfs_inode_t     *ip,
 124         xfs_off_t       offset,
 125         size_t          count,
 126         xfs_bmbt_irec_t *imap,
 127         int             nmaps)
 128 {
 129         xfs_mount_t     *mp = ip->i_mount;
 130         xfs_fileoff_t   offset_fsb;
 131         xfs_fileoff_t   last_fsb;
 132         xfs_filblks_t   count_fsb, resaligned;
 133         xfs_fsblock_t   firstfsb;
 134         xfs_extlen_t    extsz, temp;
 135         int             nimaps;
 136         int             bmapi_flag;
 137         int             quota_flag;
 138         int             rt;
 139         xfs_trans_t     *tp;
 140         xfs_bmap_free_t free_list;
 141         uint            qblocks, resblks, resrtextents;
 142         int             committed;
 143         int             error;
 144
 145         error = xfs_qm_dqattach(ip, 0);
 146         if (error)
 147                 return XFS_ERROR(error);
 148
 149         rt = XFS_IS_REALTIME_INODE(ip);
 150         extsz = xfs_get_extsz_hint(ip);
 151
 152         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 153         last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
 154         if ((offset + count) > XFS_ISIZE(ip)) {
 155                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
 156                 if (error)
 157                         return XFS_ERROR(error);
 158         } else {
 159                 if (nmaps && (imap->br_startblock == HOLESTARTBLOCK))
 160                         last_fsb = MIN(last_fsb, (xfs_fileoff_t)
 161                                         imap->br_blockcount +
 162                                         imap->br_startoff);
 163         }
 164         count_fsb = last_fsb - offset_fsb;
 165         ASSERT(count_fsb > 0);
 166
 167         resaligned = count_fsb;
 168         if (unlikely(extsz)) {
 169                 if ((temp = do_mod(offset_fsb, extsz)))
 170                         resaligned += temp;
 171                 if ((temp = do_mod(resaligned, extsz)))
 172                         resaligned += extsz - temp;
 173         }
 174
 175         if (unlikely(rt)) {
 176                 resrtextents = qblocks = resaligned;
 177                 resrtextents /= mp->m_sb.sb_rextsize;
 178                 resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
 179                 quota_flag = XFS_QMOPT_RES_RTBLKS;
 180         } else {
 181                 resrtextents = 0;
 182                 resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
 183                 quota_flag = XFS_QMOPT_RES_REGBLKS;
 184         }
 185
 186         /*
 187          * Allocate and setup the transaction
 188          */
 189         tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
 190         error = xfs_trans_reserve(tp, resblks,
 191                         XFS_WRITE_LOG_RES(mp), resrtextents,
 192                         XFS_TRANS_PERM_LOG_RES,
 193                         XFS_WRITE_LOG_COUNT);
 194         /*
 195          * Check for running out of space, note: need lock to return
 196          */
 197         if (error) {
 198                 xfs_trans_cancel(tp, 0);
 199                 return XFS_ERROR(error);
 200         }
 201
 202         xfs_ilock(ip, XFS_ILOCK_EXCL);
 203
 204         error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
 205         if (error)
 206                 goto out_trans_cancel;
 207
 208         xfs_trans_ijoin(tp, ip, 0);
 209
 210         bmapi_flag = 0;
 211         if (offset < XFS_ISIZE(ip) || extsz)
 212                 bmapi_flag |= XFS_BMAPI_PREALLOC;
 213
 214         /*
 215          * From this point onwards we overwrite the imap pointer that the
 216          * caller gave to us.
 217          */
 218         xfs_bmap_init(&free_list, &firstfsb);
 219         nimaps = 1;
 220         error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb, bmapi_flag,
 221                                 &firstfsb, 0, imap, &nimaps, &free_list);
 222         if (error)
 223                 goto out_bmap_cancel;
 224
 225         /*
 226          * Complete the transaction
 227          */
 228         error = xfs_bmap_finish(&tp, &free_list, &committed);
 229         if (error)
 230                 goto out_bmap_cancel;
 231         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 232         if (error)
 233                 goto out_unlock;
 234
 235         /*
 236          * Copy any maps to caller's array and return any error.
 237          */
 238         if (nimaps == 0) {
 239                 error = XFS_ERROR(ENOSPC);
 240                 goto out_unlock;
 241         }
 242
 243         if (!(imap->br_startblock || XFS_IS_REALTIME_INODE(ip)))
 244                 error = xfs_alert_fsblock_zero(ip, imap);
 245
 246 out_unlock:
 247         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 248         return error;
 249
 250 out_bmap_cancel:
 251         xfs_bmap_cancel(&free_list);
 252         xfs_trans_unreserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
 253 out_trans_cancel:
 254         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 255         goto out_unlock;
 256 }
 257
 258 /*
 259  * If the caller is doing a write at the end of the file, then extend the
 260  * allocation out to the file system's write iosize.  We clean up any extra
 261  * space left over when the file is closed in xfs_inactive().
 262  *
 263  * If we find we already have delalloc preallocation beyond EOF, don't do more
 264  * preallocation as it it not needed.
 265  */
 266 STATIC int
 267 xfs_iomap_eof_want_preallocate(
 268         xfs_mount_t     *mp,
 269         xfs_inode_t     *ip,
 270         xfs_off_t       offset,
 271         size_t          count,
 272         xfs_bmbt_irec_t *imap,
 273         int             nimaps,
 274         int             *prealloc)
 275 {
 276         xfs_fileoff_t   start_fsb;
 277         xfs_filblks_t   count_fsb;
 278         xfs_fsblock_t   firstblock;
 279         int             n, error, imaps;
 280         int             found_delalloc = 0;
 281
 282         *prealloc = 0;
 283         if (offset + count <= XFS_ISIZE(ip))
 284                 return 0;
 285
 286         /*
 287          * If there are any real blocks past eof, then don't
 288          * do any speculative allocation.
 289          */
 290         start_fsb = XFS_B_TO_FSBT(mp, ((xfs_ufsize_t)(offset + count - 1)));
 291         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
 292         while (count_fsb > 0) {
 293                 imaps = nimaps;
 294                 firstblock = NULLFSBLOCK;
 295                 error = xfs_bmapi_read(ip, start_fsb, count_fsb, imap, &imaps,
 296                                        0);
 297                 if (error)
 298                         return error;
 299                 for (n = 0; n < imaps; n++) {
 300                         if ((imap[n].br_startblock != HOLESTARTBLOCK) &&
 301                             (imap[n].br_startblock != DELAYSTARTBLOCK))
 302                                 return 0;
 303                         start_fsb += imap[n].br_blockcount;
 304                         count_fsb -= imap[n].br_blockcount;
 305
 306                         if (imap[n].br_startblock == DELAYSTARTBLOCK)
 307                                 found_delalloc = 1;
 308                 }
 309         }
 310         if (!found_delalloc)
 311                 *prealloc = 1;
 312         return 0;
 313 }
 314
 315 /*
 316  * If we don't have a user specified preallocation size, dynamically increase
 317  * the preallocation size as the size of the file grows. Cap the maximum size
 318  * at a single extent or less if the filesystem is near full. The closer the
 319  * filesystem is to full, the smaller the maximum prealocation.
 320  */
 321 STATIC xfs_fsblock_t
 322 xfs_iomap_prealloc_size(
 323         struct xfs_mount        *mp,
 324         struct xfs_inode        *ip)
 325 {
 326         xfs_fsblock_t           alloc_blocks = 0;
 327
 328         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
 329                 int shift = 0;
 330                 int64_t freesp;
 331
 332                 /*
 333                  * rounddown_pow_of_two() returns an undefined result
 334                  * if we pass in alloc_blocks = 0. Hence the "+ 1" to
 335                  * ensure we always pass in a non-zero value.
 336                  */
 337                 alloc_blocks = XFS_B_TO_FSB(mp, XFS_ISIZE(ip)) + 1;
 338                 alloc_blocks = XFS_FILEOFF_MIN(MAXEXTLEN,
 339                                         rounddown_pow_of_two(alloc_blocks));
 340
 341                 xfs_icsb_sync_counters(mp, XFS_ICSB_LAZY_COUNT);
 342                 freesp = mp->m_sb.sb_fdblocks;
 343                 if (freesp < mp->m_low_space[XFS_LOWSP_5_PCNT]) {
 344                         shift = 2;
 345                         if (freesp < mp->m_low_space[XFS_LOWSP_4_PCNT])
 346                                 shift++;
 347                         if (freesp < mp->m_low_space[XFS_LOWSP_3_PCNT])
 348                                 shift++;
 349                         if (freesp < mp->m_low_space[XFS_LOWSP_2_PCNT])
 350                                 shift++;
 351                         if (freesp < mp->m_low_space[XFS_LOWSP_1_PCNT])
 352                                 shift++;
 353                 }
 354                 if (shift)
 355                         alloc_blocks >>= shift;
 356         }
 357
 358         if (alloc_blocks < mp->m_writeio_blocks)
 359                 alloc_blocks = mp->m_writeio_blocks;
 360
 361         return alloc_blocks;
 362 }
 363
 364 int
 365 xfs_iomap_write_delay(
 366         xfs_inode_t     *ip,
 367         xfs_off_t       offset,
 368         size_t          count,
 369         xfs_bmbt_irec_t *ret_imap)
 370 {
 371         xfs_mount_t     *mp = ip->i_mount;
 372         xfs_fileoff_t   offset_fsb;
 373         xfs_fileoff_t   last_fsb;
 374         xfs_off_t       aligned_offset;
 375         xfs_fileoff_t   ioalign;
 376         xfs_extlen_t    extsz;
 377         int             nimaps;
 378         xfs_bmbt_irec_t imap[XFS_WRITE_IMAPS];
 379         int             prealloc, flushed = 0;
 380         int             error;
 381
 382         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 383
 384         /*
 385          * Make sure that the dquots are there. This doesn't hold
 386          * the ilock across a disk read.
 387          */
 388         error = xfs_qm_dqattach_locked(ip, 0);
 389         if (error)
 390                 return XFS_ERROR(error);
 391
 392         extsz = xfs_get_extsz_hint(ip);
 393         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 394
 395
 396         error = xfs_iomap_eof_want_preallocate(mp, ip, offset, count,
 397                                 imap, XFS_WRITE_IMAPS, &prealloc);
 398         if (error)
 399                 return error;
 400
 401 retry:
 402         if (prealloc) {
 403                 xfs_fsblock_t   alloc_blocks = xfs_iomap_prealloc_size(mp, ip);
 404
 405                 aligned_offset = XFS_WRITEIO_ALIGN(mp, (offset + count - 1));
 406                 ioalign = XFS_B_TO_FSBT(mp, aligned_offset);
 407                 last_fsb = ioalign + alloc_blocks;
 408         } else {
 409                 last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
 410         }
 411
 412         if (prealloc || extsz) {
 413                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
 414                 if (error)
 415                         return error;
 416         }
 417
 418         /*
 419          * Make sure preallocation does not create extents beyond the range we
 420          * actually support in this filesystem.
 421          */
 422         if (last_fsb > XFS_B_TO_FSB(mp, mp->m_maxioffset))
 423                 last_fsb = XFS_B_TO_FSB(mp, mp->m_maxioffset);
 424
 425         ASSERT(last_fsb > offset_fsb);
 426
 427         nimaps = XFS_WRITE_IMAPS;
 428         error = xfs_bmapi_delay(ip, offset_fsb, last_fsb - offset_fsb,
 429                                 imap, &nimaps, XFS_BMAPI_ENTIRE);
 430         switch (error) {
 431         case 0:
 432         case ENOSPC:
 433         case EDQUOT:
 434                 break;
 435         default:
 436                 return XFS_ERROR(error);
 437         }
 438
 439         /*
 440          * If bmapi returned us nothing, we got either ENOSPC or EDQUOT.  For
 441          * ENOSPC, * flush all other inodes with delalloc blocks to free up
 442          * some of the excess reserved metadata space. For both cases, retry
 443          * without EOF preallocation.
 444          */
 445         if (nimaps == 0) {
 446                 trace_xfs_delalloc_enospc(ip, offset, count);
 447                 if (flushed)
 448                         return XFS_ERROR(error ? error : ENOSPC);
 449
 450                 if (error == ENOSPC) {
 451                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 452                         xfs_flush_inodes(ip);
 453                         xfs_ilock(ip, XFS_ILOCK_EXCL);
 454                 }
 455
 456                 flushed = 1;
 457                 error = 0;
 458                 prealloc = 0;
 459                 goto retry;
 460         }
 461
 462         if (!(imap[0].br_startblock || XFS_IS_REALTIME_INODE(ip)))
 463                 return xfs_alert_fsblock_zero(ip, &imap[0]);
 464
 465         *ret_imap = imap[0];
 466         return 0;
 467 }
 468
 469 /*
 470  * Pass in a delayed allocate extent, convert it to real extents;
 471  * return to the caller the extent we create which maps on top of
 472  * the originating callers request.
 473  *
 474  * Called without a lock on the inode.
 475  *
 476  * We no longer bother to look at the incoming map - all we have to
 477  * guarantee is that whatever we allocate fills the required range.
 478  */
 479 int
 480 xfs_iomap_write_allocate(
 481         xfs_inode_t     *ip,
 482         xfs_off_t       offset,
 483         size_t          count,
 484         xfs_bmbt_irec_t *imap)
 485 {
 486         xfs_mount_t     *mp = ip->i_mount;
 487         xfs_fileoff_t   offset_fsb, last_block;
 488         xfs_fileoff_t   end_fsb, map_start_fsb;
 489         xfs_fsblock_t   first_block;
 490         xfs_bmap_free_t free_list;
 491         xfs_filblks_t   count_fsb;
 492         xfs_trans_t     *tp;
 493         int             nimaps, committed;
 494         int             error = 0;
 495         int             nres;
 496
 497         /*
 498          * Make sure that the dquots are there.
 499          */
 500         error = xfs_qm_dqattach(ip, 0);
 501         if (error)
 502                 return XFS_ERROR(error);
 503
 504         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 505         count_fsb = imap->br_blockcount;
 506         map_start_fsb = imap->br_startoff;
 507
 508         XFS_STATS_ADD(xs_xstrat_bytes, XFS_FSB_TO_B(mp, count_fsb));
 509
 510         while (count_fsb != 0) {
 511                 /*
 512                  * Set up a transaction with which to allocate the
 513                  * backing store for the file.  Do allocations in a
 514                  * loop until we get some space in the range we are
 515                  * interested in.  The other space that might be allocated
 516                  * is in the delayed allocation extent on which we sit
 517                  * but before our buffer starts.
 518                  */
 519
 520                 nimaps = 0;
 521                 while (nimaps == 0) {
 522                         tp = xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE);
 523                         tp->t_flags |= XFS_TRANS_RESERVE;
 524                         nres = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
 525                         error = xfs_trans_reserve(tp, nres,
 526                                         XFS_WRITE_LOG_RES(mp),
 527                                         0, XFS_TRANS_PERM_LOG_RES,
 528                                         XFS_WRITE_LOG_COUNT);
 529                         if (error) {
 530                                 xfs_trans_cancel(tp, 0);
 531                                 return XFS_ERROR(error);
 532                         }
 533                         xfs_ilock(ip, XFS_ILOCK_EXCL);
 534                         xfs_trans_ijoin(tp, ip, 0);
 535
 536                         xfs_bmap_init(&free_list, &first_block);
 537
 538                         /*
 539                          * it is possible that the extents have changed since
 540                          * we did the read call as we dropped the ilock for a
 541                          * while. We have to be careful about truncates or hole
 542                          * punchs here - we are not allowed to allocate
 543                          * non-delalloc blocks here.
 544                          *
 545                          * The only protection against truncation is the pages
 546                          * for the range we are being asked to convert are
 547                          * locked and hence a truncate will block on them
 548                          * first.
 549                          *
 550                          * As a result, if we go beyond the range we really
 551                          * need and hit an delalloc extent boundary followed by
 552                          * a hole while we have excess blocks in the map, we
 553                          * will fill the hole incorrectly and overrun the
 554                          * transaction reservation.
 555                          *
 556                          * Using a single map prevents this as we are forced to
 557                          * check each map we look for overlap with the desired
 558                          * range and abort as soon as we find it. Also, given
 559                          * that we only return a single map, having one beyond
 560                          * what we can return is probably a bit silly.
 561                          *
 562                          * We also need to check that we don't go beyond EOF;
 563                          * this is a truncate optimisation as a truncate sets
 564                          * the new file size before block on the pages we
 565                          * currently have locked under writeback. Because they
 566                          * are about to be tossed, we don't need to write them
 567                          * back....
 568                          */
 569                         nimaps = 1;
 570                         end_fsb = XFS_B_TO_FSB(mp, XFS_ISIZE(ip));
 571                         error = xfs_bmap_last_offset(NULL, ip, &last_block,
 572                                                         XFS_DATA_FORK);
 573                         if (error)
 574                                 goto trans_cancel;
 575
 576                         last_block = XFS_FILEOFF_MAX(last_block, end_fsb);
 577                         if ((map_start_fsb + count_fsb) > last_block) {
 578                                 count_fsb = last_block - map_start_fsb;
 579                                 if (count_fsb == 0) {
 580                                         error = EAGAIN;
 581                                         goto trans_cancel;
 582                                 }
 583                         }
 584
 585                         /*
 586                          * From this point onwards we overwrite the imap
 587                          * pointer that the caller gave to us.
 588                          */
 589                         error = xfs_bmapi_write(tp, ip, map_start_fsb,
 590                                                 count_fsb, 0, &first_block, 1,
 591                                                 imap, &nimaps, &free_list);
 592                         if (error)
 593                                 goto trans_cancel;
 594
 595                         error = xfs_bmap_finish(&tp, &free_list, &committed);
 596                         if (error)
 597                                 goto trans_cancel;
 598
 599                         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 600                         if (error)
 601                                 goto error0;
 602
 603                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 604                 }
 605
 606                 /*
 607                  * See if we were able to allocate an extent that
 608                  * covers at least part of the callers request
 609                  */
 610                 if (!(imap->br_startblock || XFS_IS_REALTIME_INODE(ip)))
 611                         return xfs_alert_fsblock_zero(ip, imap);
 612
 613                 if ((offset_fsb >= imap->br_startoff) &&
 614                     (offset_fsb < (imap->br_startoff +
 615                                    imap->br_blockcount))) {
 616                         XFS_STATS_INC(xs_xstrat_quick);
 617                         return 0;
 618                 }
 619
 620                 /*
 621                  * So far we have not mapped the requested part of the
 622                  * file, just surrounding data, try again.
 623                  */
 624                 count_fsb -= imap->br_blockcount;
 625                 map_start_fsb = imap->br_startoff + imap->br_blockcount;
 626         }
 627
 628 trans_cancel:
 629         xfs_bmap_cancel(&free_list);
 630         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 631 error0:
 632         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 633         return XFS_ERROR(error);
 634 }
 635
 636 int
 637 xfs_iomap_write_unwritten(
 638         xfs_inode_t     *ip,
 639         xfs_off_t       offset,
 640         size_t          count)
 641 {
 642         xfs_mount_t     *mp = ip->i_mount;
 643         xfs_fileoff_t   offset_fsb;
 644         xfs_filblks_t   count_fsb;
 645         xfs_filblks_t   numblks_fsb;
 646         xfs_fsblock_t   firstfsb;
 647         int             nimaps;
 648         xfs_trans_t     *tp;
 649         xfs_bmbt_irec_t imap;
 650         xfs_bmap_free_t free_list;
 651         xfs_fsize_t     i_size;
 652         uint            resblks;
 653         int             committed;
 654         int             error;
 655
 656         trace_xfs_unwritten_convert(ip, offset, count);
 657
 658         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 659         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
 660         count_fsb = (xfs_filblks_t)(count_fsb - offset_fsb);
 661
 662         /*
 663          * Reserve enough blocks in this transaction for two complete extent
 664          * btree splits.  We may be converting the middle part of an unwritten
 665          * extent and in this case we will insert two new extents in the btree
 666          * each of which could cause a full split.
 667          *
 668          * This reservation amount will be used in the first call to
 669          * xfs_bmbt_split() to select an AG with enough space to satisfy the
 670          * rest of the operation.
 671          */
 672         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
 673
 674         do {
 675                 /*
 676                  * set up a transaction to convert the range of extents
 677                  * from unwritten to real. Do allocations in a loop until
 678                  * we have covered the range passed in.
 679                  *
 680                  * Note that we open code the transaction allocation here
 681                  * to pass KM_NOFS--we can't risk to recursing back into
 682                  * the filesystem here as we might be asked to write out
 683                  * the same inode that we complete here and might deadlock
 684                  * on the iolock.
 685                  */
 686                 xfs_wait_for_freeze(mp, SB_FREEZE_TRANS);
 687                 tp = _xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE, KM_NOFS);
 688                 tp->t_flags |= XFS_TRANS_RESERVE;
 689                 error = xfs_trans_reserve(tp, resblks,
 690                                 XFS_WRITE_LOG_RES(mp), 0,
 691                                 XFS_TRANS_PERM_LOG_RES,
 692                                 XFS_WRITE_LOG_COUNT);
 693                 if (error) {
 694                         xfs_trans_cancel(tp, 0);
 695                         return XFS_ERROR(error);
 696                 }
 697
 698                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 699                 xfs_trans_ijoin(tp, ip, 0);
 700
 701                 /*
 702                  * Modify the unwritten extent state of the buffer.
 703                  */
 704                 xfs_bmap_init(&free_list, &firstfsb);
 705                 nimaps = 1;
 706                 error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb,
 707                                   XFS_BMAPI_CONVERT, &firstfsb,
 708                                   1, &imap, &nimaps, &free_list);
 709                 if (error)
 710                         goto error_on_bmapi_transaction;
 711
 712                 /*
 713                  * Log the updated inode size as we go.  We have to be careful
 714                  * to only log it up to the actual write offset if it is
 715                  * halfway into a block.
 716                  */
 717                 i_size = XFS_FSB_TO_B(mp, offset_fsb + count_fsb);
 718                 if (i_size > offset + count)
 719                         i_size = offset + count;
 720
 721                 i_size = xfs_new_eof(ip, i_size);
 722                 if (i_size) {
 723                         ip->i_d.di_size = i_size;
 724                         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 725                 }
 726
 727                 error = xfs_bmap_finish(&tp, &free_list, &committed);
 728                 if (error)
 729                         goto error_on_bmapi_transaction;
 730
 731                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 732                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 733                 if (error)
 734                         return XFS_ERROR(error);
 735
 736                 if (!(imap.br_startblock || XFS_IS_REALTIME_INODE(ip)))
 737                         return xfs_alert_fsblock_zero(ip, &imap);
 738
 739                 if ((numblks_fsb = imap.br_blockcount) == 0) {
 740                         /*
 741                          * The numblks_fsb value should always get
 742                          * smaller, otherwise the loop is stuck.
 743                          */
 744                         ASSERT(imap.br_blockcount);
 745                         break;
 746                 }
 747                 offset_fsb += numblks_fsb;
 748                 count_fsb -= numblks_fsb;
 749         } while (count_fsb > 0);
 750
 751         return 0;
 752
 753 error_on_bmapi_transaction:
 754         xfs_bmap_cancel(&free_list);
 755         xfs_trans_cancel(tp, (XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT));
 756         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 757         return XFS_ERROR(error);
 758 }