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 #include "xfs_icache.h"
  45
  46
  47 #define XFS_WRITEIO_ALIGN(mp,off)       (((off) >> mp->m_writeio_log) \
  48                                                 << mp->m_writeio_log)
  49 #define XFS_WRITE_IMAPS         XFS_BMAP_MAX_NMAP
  50
  51 STATIC int
  52 xfs_iomap_eof_align_last_fsb(
  53         xfs_mount_t     *mp,
  54         xfs_inode_t     *ip,
  55         xfs_extlen_t    extsize,
  56         xfs_fileoff_t   *last_fsb)
  57 {
  58         xfs_fileoff_t   new_last_fsb = 0;
  59         xfs_extlen_t    align = 0;
  60         int             eof, error;
  61
  62         if (!XFS_IS_REALTIME_INODE(ip)) {
  63                 /*
  64                  * Round up the allocation request to a stripe unit
  65                  * (m_dalign) boundary if the file size is >= stripe unit
  66                  * size, and we are allocating past the allocation eof.
  67                  *
  68                  * If mounted with the "-o swalloc" option the alignment is
  69                  * increased from the strip unit size to the stripe width.
  70                  */
  71                 if (mp->m_swidth && (mp->m_flags & XFS_MOUNT_SWALLOC))
  72                         align = mp->m_swidth;
  73                 else if (mp->m_dalign)
  74                         align = mp->m_dalign;
  75
  76                 if (align && XFS_ISIZE(ip) >= XFS_FSB_TO_B(mp, align))
  77                         new_last_fsb = roundup_64(*last_fsb, align);
  78         }
  79
  80         /*
  81          * Always round up the allocation request to an extent boundary
  82          * (when file on a real-time subvolume or has di_extsize hint).
  83          */
  84         if (extsize) {
  85                 if (new_last_fsb)
  86                         align = roundup_64(new_last_fsb, extsize);
  87                 else
  88                         align = extsize;
  89                 new_last_fsb = roundup_64(*last_fsb, align);
  90         }
  91
  92         if (new_last_fsb) {
  93                 error = xfs_bmap_eof(ip, new_last_fsb, XFS_DATA_FORK, &eof);
  94                 if (error)
  95                         return error;
  96                 if (eof)
  97                         *last_fsb = new_last_fsb;
  98         }
  99         return 0;
 100 }
 101
 102 STATIC int
 103 xfs_alert_fsblock_zero(
 104         xfs_inode_t     *ip,
 105         xfs_bmbt_irec_t *imap)
 106 {
 107         xfs_alert_tag(ip->i_mount, XFS_PTAG_FSBLOCK_ZERO,
 108                         "Access to block zero in inode %llu "
 109                         "start_block: %llx start_off: %llx "
 110                         "blkcnt: %llx extent-state: %x\n",
 111                 (unsigned long long)ip->i_ino,
 112                 (unsigned long long)imap->br_startblock,
 113                 (unsigned long long)imap->br_startoff,
 114                 (unsigned long long)imap->br_blockcount,
 115                 imap->br_state);
 116         return EFSCORRUPTED;
 117 }
 118
 119 int
 120 xfs_iomap_write_direct(
 121         xfs_inode_t     *ip,
 122         xfs_off_t       offset,
 123         size_t          count,
 124         xfs_bmbt_irec_t *imap,
 125         int             nmaps)
 126 {
 127         xfs_mount_t     *mp = ip->i_mount;
 128         xfs_fileoff_t   offset_fsb;
 129         xfs_fileoff_t   last_fsb;
 130         xfs_filblks_t   count_fsb, resaligned;
 131         xfs_fsblock_t   firstfsb;
 132         xfs_extlen_t    extsz, temp;
 133         int             nimaps;
 134         int             bmapi_flag;
 135         int             quota_flag;
 136         int             rt;
 137         xfs_trans_t     *tp;
 138         xfs_bmap_free_t free_list;
 139         uint            qblocks, resblks, resrtextents;
 140         int             committed;
 141         int             error;
 142
 143         error = xfs_qm_dqattach(ip, 0);
 144         if (error)
 145                 return XFS_ERROR(error);
 146
 147         rt = XFS_IS_REALTIME_INODE(ip);
 148         extsz = xfs_get_extsz_hint(ip);
 149
 150         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 151         last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
 152         if ((offset + count) > XFS_ISIZE(ip)) {
 153                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
 154                 if (error)
 155                         return XFS_ERROR(error);
 156         } else {
 157                 if (nmaps && (imap->br_startblock == HOLESTARTBLOCK))
 158                         last_fsb = MIN(last_fsb, (xfs_fileoff_t)
 159                                         imap->br_blockcount +
 160                                         imap->br_startoff);
 161         }
 162         count_fsb = last_fsb - offset_fsb;
 163         ASSERT(count_fsb > 0);
 164
 165         resaligned = count_fsb;
 166         if (unlikely(extsz)) {
 167                 if ((temp = do_mod(offset_fsb, extsz)))
 168                         resaligned += temp;
 169                 if ((temp = do_mod(resaligned, extsz)))
 170                         resaligned += extsz - temp;
 171         }
 172
 173         if (unlikely(rt)) {
 174                 resrtextents = qblocks = resaligned;
 175                 resrtextents /= mp->m_sb.sb_rextsize;
 176                 resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
 177                 quota_flag = XFS_QMOPT_RES_RTBLKS;
 178         } else {
 179                 resrtextents = 0;
 180                 resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
 181                 quota_flag = XFS_QMOPT_RES_REGBLKS;
 182         }
 183
 184         /*
 185          * Allocate and setup the transaction
 186          */
 187         tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
 188         error = xfs_trans_reserve(tp, resblks,
 189                         XFS_WRITE_LOG_RES(mp), resrtextents,
 190                         XFS_TRANS_PERM_LOG_RES,
 191                         XFS_WRITE_LOG_COUNT);
 192         /*
 193          * Check for running out of space, note: need lock to return
 194          */
 195         if (error) {
 196                 xfs_trans_cancel(tp, 0);
 197                 return XFS_ERROR(error);
 198         }
 199
 200         xfs_ilock(ip, XFS_ILOCK_EXCL);
 201
 202         error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
 203         if (error)
 204                 goto out_trans_cancel;
 205
 206         xfs_trans_ijoin(tp, ip, 0);
 207
 208         bmapi_flag = 0;
 209         if (offset < XFS_ISIZE(ip) || extsz)
 210                 bmapi_flag |= XFS_BMAPI_PREALLOC;
 211
 212         /*
 213          * From this point onwards we overwrite the imap pointer that the
 214          * caller gave to us.
 215          */
 216         xfs_bmap_init(&free_list, &firstfsb);
 217         nimaps = 1;
 218         error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb, bmapi_flag,
 219                                 &firstfsb, 0, imap, &nimaps, &free_list);
 220         if (error)
 221                 goto out_bmap_cancel;
 222
 223         /*
 224          * Complete the transaction
 225          */
 226         error = xfs_bmap_finish(&tp, &free_list, &committed);
 227         if (error)
 228                 goto out_bmap_cancel;
 229         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 230         if (error)
 231                 goto out_unlock;
 232
 233         /*
 234          * Copy any maps to caller's array and return any error.
 235          */
 236         if (nimaps == 0) {
 237                 error = XFS_ERROR(ENOSPC);
 238                 goto out_unlock;
 239         }
 240
 241         if (!(imap->br_startblock || XFS_IS_REALTIME_INODE(ip)))
 242                 error = xfs_alert_fsblock_zero(ip, imap);
 243
 244 out_unlock:
 245         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 246         return error;
 247
 248 out_bmap_cancel:
 249         xfs_bmap_cancel(&free_list);
 250         xfs_trans_unreserve_quota_nblks(tp, ip, (long)qblocks, 0, quota_flag);
 251 out_trans_cancel:
 252         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 253         goto out_unlock;
 254 }
 255
 256 /*
 257  * If the caller is doing a write at the end of the file, then extend the
 258  * allocation out to the file system's write iosize.  We clean up any extra
 259  * space left over when the file is closed in xfs_inactive().
 260  *
 261  * If we find we already have delalloc preallocation beyond EOF, don't do more
 262  * preallocation as it it not needed.
 263  */
 264 STATIC int
 265 xfs_iomap_eof_want_preallocate(
 266         xfs_mount_t     *mp,
 267         xfs_inode_t     *ip,
 268         xfs_off_t       offset,
 269         size_t          count,
 270         xfs_bmbt_irec_t *imap,
 271         int             nimaps,
 272         int             *prealloc)
 273 {
 274         xfs_fileoff_t   start_fsb;
 275         xfs_filblks_t   count_fsb;
 276         xfs_fsblock_t   firstblock;
 277         int             n, error, imaps;
 278         int             found_delalloc = 0;
 279
 280         *prealloc = 0;
 281         if (offset + count <= XFS_ISIZE(ip))
 282                 return 0;
 283
 284         /*
 285          * If there are any real blocks past eof, then don't
 286          * do any speculative allocation.
 287          */
 288         start_fsb = XFS_B_TO_FSBT(mp, ((xfs_ufsize_t)(offset + count - 1)));
 289         count_fsb = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
 290         while (count_fsb > 0) {
 291                 imaps = nimaps;
 292                 firstblock = NULLFSBLOCK;
 293                 error = xfs_bmapi_read(ip, start_fsb, count_fsb, imap, &imaps,
 294                                        0);
 295                 if (error)
 296                         return error;
 297                 for (n = 0; n < imaps; n++) {
 298                         if ((imap[n].br_startblock != HOLESTARTBLOCK) &&
 299                             (imap[n].br_startblock != DELAYSTARTBLOCK))
 300                                 return 0;
 301                         start_fsb += imap[n].br_blockcount;
 302                         count_fsb -= imap[n].br_blockcount;
 303
 304                         if (imap[n].br_startblock == DELAYSTARTBLOCK)
 305                                 found_delalloc = 1;
 306                 }
 307         }
 308         if (!found_delalloc)
 309                 *prealloc = 1;
 310         return 0;
 311 }
 312
 313 /*
 314  * If we don't have a user specified preallocation size, dynamically increase
 315  * the preallocation size as the size of the file grows. Cap the maximum size
 316  * at a single extent or less if the filesystem is near full. The closer the
 317  * filesystem is to full, the smaller the maximum prealocation.
 318  */
 319 STATIC xfs_fsblock_t
 320 xfs_iomap_prealloc_size(
 321         struct xfs_mount        *mp,
 322         struct xfs_inode        *ip)
 323 {
 324         xfs_fsblock_t           alloc_blocks = 0;
 325
 326         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
 327                 int shift = 0;
 328                 int64_t freesp;
 329
 330                 /*
 331                  * rounddown_pow_of_two() returns an undefined result
 332                  * if we pass in alloc_blocks = 0. Hence the "+ 1" to
 333                  * ensure we always pass in a non-zero value.
 334                  */
 335                 alloc_blocks = XFS_B_TO_FSB(mp, XFS_ISIZE(ip)) + 1;
 336                 alloc_blocks = XFS_FILEOFF_MIN(MAXEXTLEN,
 337                                         rounddown_pow_of_two(alloc_blocks));
 338
 339                 xfs_icsb_sync_counters(mp, XFS_ICSB_LAZY_COUNT);
 340                 freesp = mp->m_sb.sb_fdblocks;
 341                 if (freesp < mp->m_low_space[XFS_LOWSP_5_PCNT]) {
 342                         shift = 2;
 343                         if (freesp < mp->m_low_space[XFS_LOWSP_4_PCNT])
 344                                 shift++;
 345                         if (freesp < mp->m_low_space[XFS_LOWSP_3_PCNT])
 346                                 shift++;
 347                         if (freesp < mp->m_low_space[XFS_LOWSP_2_PCNT])
 348                                 shift++;
 349                         if (freesp < mp->m_low_space[XFS_LOWSP_1_PCNT])
 350                                 shift++;
 351                 }
 352                 if (shift)
 353                         alloc_blocks >>= shift;
 354
 355                 /*
 356                  * If we are still trying to allocate more space than is
 357                  * available, squash the prealloc hard. This can happen if we
 358                  * have a large file on a small filesystem and the above
 359                  * lowspace thresholds are smaller than MAXEXTLEN.
 360                  */
 361                 while (alloc_blocks >= freesp)
 362                         alloc_blocks >>= 4;
 363         }
 364
 365         if (alloc_blocks < mp->m_writeio_blocks)
 366                 alloc_blocks = mp->m_writeio_blocks;
 367
 368         return alloc_blocks;
 369 }
 370
 371 int
 372 xfs_iomap_write_delay(
 373         xfs_inode_t     *ip,
 374         xfs_off_t       offset,
 375         size_t          count,
 376         xfs_bmbt_irec_t *ret_imap)
 377 {
 378         xfs_mount_t     *mp = ip->i_mount;
 379         xfs_fileoff_t   offset_fsb;
 380         xfs_fileoff_t   last_fsb;
 381         xfs_off_t       aligned_offset;
 382         xfs_fileoff_t   ioalign;
 383         xfs_extlen_t    extsz;
 384         int             nimaps;
 385         xfs_bmbt_irec_t imap[XFS_WRITE_IMAPS];
 386         int             prealloc;
 387         int             error;
 388
 389         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 390
 391         /*
 392          * Make sure that the dquots are there. This doesn't hold
 393          * the ilock across a disk read.
 394          */
 395         error = xfs_qm_dqattach_locked(ip, 0);
 396         if (error)
 397                 return XFS_ERROR(error);
 398
 399         extsz = xfs_get_extsz_hint(ip);
 400         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 401
 402
 403         error = xfs_iomap_eof_want_preallocate(mp, ip, offset, count,
 404                                 imap, XFS_WRITE_IMAPS, &prealloc);
 405         if (error)
 406                 return error;
 407
 408 retry:
 409         if (prealloc) {
 410                 xfs_fsblock_t   alloc_blocks = xfs_iomap_prealloc_size(mp, ip);
 411
 412                 aligned_offset = XFS_WRITEIO_ALIGN(mp, (offset + count - 1));
 413                 ioalign = XFS_B_TO_FSBT(mp, aligned_offset);
 414                 last_fsb = ioalign + alloc_blocks;
 415         } else {
 416                 last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
 417         }
 418
 419         if (prealloc || extsz) {
 420                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
 421                 if (error)
 422                         return error;
 423         }
 424
 425         /*
 426          * Make sure preallocation does not create extents beyond the range we
 427          * actually support in this filesystem.
 428          */
 429         if (last_fsb > XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes))
 430                 last_fsb = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
 431
 432         ASSERT(last_fsb > offset_fsb);
 433
 434         nimaps = XFS_WRITE_IMAPS;
 435         error = xfs_bmapi_delay(ip, offset_fsb, last_fsb - offset_fsb,
 436                                 imap, &nimaps, XFS_BMAPI_ENTIRE);
 437         switch (error) {
 438         case 0:
 439         case ENOSPC:
 440         case EDQUOT:
 441                 break;
 442         default:
 443                 return XFS_ERROR(error);
 444         }
 445
 446         /*
 447          * If bmapi returned us nothing, we got either ENOSPC or EDQUOT. Retry
 448          * without EOF preallocation.
 449          */
 450         if (nimaps == 0) {
 451                 trace_xfs_delalloc_enospc(ip, offset, count);
 452                 if (prealloc) {
 453                         prealloc = 0;
 454                         error = 0;
 455                         goto retry;
 456                 }
 457                 return XFS_ERROR(error ? error : ENOSPC);
 458         }
 459
 460         if (!(imap[0].br_startblock || XFS_IS_REALTIME_INODE(ip)))
 461                 return xfs_alert_fsblock_zero(ip, &imap[0]);
 462
 463         /*
 464          * Tag the inode as speculatively preallocated so we can reclaim this
 465          * space on demand, if necessary.
 466          */
 467         if (prealloc)
 468                 xfs_inode_set_eofblocks_tag(ip);
 469
 470         *ret_imap = imap[0];
 471         return 0;
 472 }
 473
 474 /*
 475  * Pass in a delayed allocate extent, convert it to real extents;
 476  * return to the caller the extent we create which maps on top of
 477  * the originating callers request.
 478  *
 479  * Called without a lock on the inode.
 480  *
 481  * We no longer bother to look at the incoming map - all we have to
 482  * guarantee is that whatever we allocate fills the required range.
 483  */
 484 int
 485 xfs_iomap_write_allocate(
 486         xfs_inode_t     *ip,
 487         xfs_off_t       offset,
 488         size_t          count,
 489         xfs_bmbt_irec_t *imap)
 490 {
 491         xfs_mount_t     *mp = ip->i_mount;
 492         xfs_fileoff_t   offset_fsb, last_block;
 493         xfs_fileoff_t   end_fsb, map_start_fsb;
 494         xfs_fsblock_t   first_block;
 495         xfs_bmap_free_t free_list;
 496         xfs_filblks_t   count_fsb;
 497         xfs_trans_t     *tp;
 498         int             nimaps, committed;
 499         int             error = 0;
 500         int             nres;
 501
 502         /*
 503          * Make sure that the dquots are there.
 504          */
 505         error = xfs_qm_dqattach(ip, 0);
 506         if (error)
 507                 return XFS_ERROR(error);
 508
 509         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 510         count_fsb = imap->br_blockcount;
 511         map_start_fsb = imap->br_startoff;
 512
 513         XFS_STATS_ADD(xs_xstrat_bytes, XFS_FSB_TO_B(mp, count_fsb));
 514
 515         while (count_fsb != 0) {
 516                 /*
 517                  * Set up a transaction with which to allocate the
 518                  * backing store for the file.  Do allocations in a
 519                  * loop until we get some space in the range we are
 520                  * interested in.  The other space that might be allocated
 521                  * is in the delayed allocation extent on which we sit
 522                  * but before our buffer starts.
 523                  */
 524
 525                 nimaps = 0;
 526                 while (nimaps == 0) {
 527                         tp = xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE);
 528                         tp->t_flags |= XFS_TRANS_RESERVE;
 529                         nres = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
 530                         error = xfs_trans_reserve(tp, nres,
 531                                         XFS_WRITE_LOG_RES(mp),
 532                                         0, XFS_TRANS_PERM_LOG_RES,
 533                                         XFS_WRITE_LOG_COUNT);
 534                         if (error) {
 535                                 xfs_trans_cancel(tp, 0);
 536                                 return XFS_ERROR(error);
 537                         }
 538                         xfs_ilock(ip, XFS_ILOCK_EXCL);
 539                         xfs_trans_ijoin(tp, ip, 0);
 540
 541                         xfs_bmap_init(&free_list, &first_block);
 542
 543                         /*
 544                          * it is possible that the extents have changed since
 545                          * we did the read call as we dropped the ilock for a
 546                          * while. We have to be careful about truncates or hole
 547                          * punchs here - we are not allowed to allocate
 548                          * non-delalloc blocks here.
 549                          *
 550                          * The only protection against truncation is the pages
 551                          * for the range we are being asked to convert are
 552                          * locked and hence a truncate will block on them
 553                          * first.
 554                          *
 555                          * As a result, if we go beyond the range we really
 556                          * need and hit an delalloc extent boundary followed by
 557                          * a hole while we have excess blocks in the map, we
 558                          * will fill the hole incorrectly and overrun the
 559                          * transaction reservation.
 560                          *
 561                          * Using a single map prevents this as we are forced to
 562                          * check each map we look for overlap with the desired
 563                          * range and abort as soon as we find it. Also, given
 564                          * that we only return a single map, having one beyond
 565                          * what we can return is probably a bit silly.
 566                          *
 567                          * We also need to check that we don't go beyond EOF;
 568                          * this is a truncate optimisation as a truncate sets
 569                          * the new file size before block on the pages we
 570                          * currently have locked under writeback. Because they
 571                          * are about to be tossed, we don't need to write them
 572                          * back....
 573                          */
 574                         nimaps = 1;
 575                         end_fsb = XFS_B_TO_FSB(mp, XFS_ISIZE(ip));
 576                         error = xfs_bmap_last_offset(NULL, ip, &last_block,
 577                                                         XFS_DATA_FORK);
 578                         if (error)
 579                                 goto trans_cancel;
 580
 581                         last_block = XFS_FILEOFF_MAX(last_block, end_fsb);
 582                         if ((map_start_fsb + count_fsb) > last_block) {
 583                                 count_fsb = last_block - map_start_fsb;
 584                                 if (count_fsb == 0) {
 585                                         error = EAGAIN;
 586                                         goto trans_cancel;
 587                                 }
 588                         }
 589
 590                         /*
 591                          * From this point onwards we overwrite the imap
 592                          * pointer that the caller gave to us.
 593                          */
 594                         error = xfs_bmapi_write(tp, ip, map_start_fsb,
 595                                                 count_fsb,
 596                                                 XFS_BMAPI_STACK_SWITCH,
 597                                                 &first_block, 1,
 598                                                 imap, &nimaps, &free_list);
 599                         if (error)
 600                                 goto trans_cancel;
 601
 602                         error = xfs_bmap_finish(&tp, &free_list, &committed);
 603                         if (error)
 604                                 goto trans_cancel;
 605
 606                         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 607                         if (error)
 608                                 goto error0;
 609
 610                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 611                 }
 612
 613                 /*
 614                  * See if we were able to allocate an extent that
 615                  * covers at least part of the callers request
 616                  */
 617                 if (!(imap->br_startblock || XFS_IS_REALTIME_INODE(ip)))
 618                         return xfs_alert_fsblock_zero(ip, imap);
 619
 620                 if ((offset_fsb >= imap->br_startoff) &&
 621                     (offset_fsb < (imap->br_startoff +
 622                                    imap->br_blockcount))) {
 623                         XFS_STATS_INC(xs_xstrat_quick);
 624                         return 0;
 625                 }
 626
 627                 /*
 628                  * So far we have not mapped the requested part of the
 629                  * file, just surrounding data, try again.
 630                  */
 631                 count_fsb -= imap->br_blockcount;
 632                 map_start_fsb = imap->br_startoff + imap->br_blockcount;
 633         }
 634
 635 trans_cancel:
 636         xfs_bmap_cancel(&free_list);
 637         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 638 error0:
 639         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 640         return XFS_ERROR(error);
 641 }
 642
 643 int
 644 xfs_iomap_write_unwritten(
 645         xfs_inode_t     *ip,
 646         xfs_off_t       offset,
 647         size_t          count)
 648 {
 649         xfs_mount_t     *mp = ip->i_mount;
 650         xfs_fileoff_t   offset_fsb;
 651         xfs_filblks_t   count_fsb;
 652         xfs_filblks_t   numblks_fsb;
 653         xfs_fsblock_t   firstfsb;
 654         int             nimaps;
 655         xfs_trans_t     *tp;
 656         xfs_bmbt_irec_t imap;
 657         xfs_bmap_free_t free_list;
 658         xfs_fsize_t     i_size;
 659         uint            resblks;
 660         int             committed;
 661         int             error;
 662
 663         trace_xfs_unwritten_convert(ip, offset, count);
 664
 665         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 666         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
 667         count_fsb = (xfs_filblks_t)(count_fsb - offset_fsb);
 668
 669         /*
 670          * Reserve enough blocks in this transaction for two complete extent
 671          * btree splits.  We may be converting the middle part of an unwritten
 672          * extent and in this case we will insert two new extents in the btree
 673          * each of which could cause a full split.
 674          *
 675          * This reservation amount will be used in the first call to
 676          * xfs_bmbt_split() to select an AG with enough space to satisfy the
 677          * rest of the operation.
 678          */
 679         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
 680
 681         do {
 682                 /*
 683                  * set up a transaction to convert the range of extents
 684                  * from unwritten to real. Do allocations in a loop until
 685                  * we have covered the range passed in.
 686                  *
 687                  * Note that we open code the transaction allocation here
 688                  * to pass KM_NOFS--we can't risk to recursing back into
 689                  * the filesystem here as we might be asked to write out
 690                  * the same inode that we complete here and might deadlock
 691                  * on the iolock.
 692                  */
 693                 sb_start_intwrite(mp->m_super);
 694                 tp = _xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE, KM_NOFS);
 695                 tp->t_flags |= XFS_TRANS_RESERVE | XFS_TRANS_FREEZE_PROT;
 696                 error = xfs_trans_reserve(tp, resblks,
 697                                 XFS_WRITE_LOG_RES(mp), 0,
 698                                 XFS_TRANS_PERM_LOG_RES,
 699                                 XFS_WRITE_LOG_COUNT);
 700                 if (error) {
 701                         xfs_trans_cancel(tp, 0);
 702                         return XFS_ERROR(error);
 703                 }
 704
 705                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 706                 xfs_trans_ijoin(tp, ip, 0);
 707
 708                 /*
 709                  * Modify the unwritten extent state of the buffer.
 710                  */
 711                 xfs_bmap_init(&free_list, &firstfsb);
 712                 nimaps = 1;
 713                 error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb,
 714                                   XFS_BMAPI_CONVERT, &firstfsb,
 715                                   1, &imap, &nimaps, &free_list);
 716                 if (error)
 717                         goto error_on_bmapi_transaction;
 718
 719                 /*
 720                  * Log the updated inode size as we go.  We have to be careful
 721                  * to only log it up to the actual write offset if it is
 722                  * halfway into a block.
 723                  */
 724                 i_size = XFS_FSB_TO_B(mp, offset_fsb + count_fsb);
 725                 if (i_size > offset + count)
 726                         i_size = offset + count;
 727
 728                 i_size = xfs_new_eof(ip, i_size);
 729                 if (i_size) {
 730                         ip->i_d.di_size = i_size;
 731                         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 732                 }
 733
 734                 error = xfs_bmap_finish(&tp, &free_list, &committed);
 735                 if (error)
 736                         goto error_on_bmapi_transaction;
 737
 738                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 739                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 740                 if (error)
 741                         return XFS_ERROR(error);
 742
 743                 if (!(imap.br_startblock || XFS_IS_REALTIME_INODE(ip)))
 744                         return xfs_alert_fsblock_zero(ip, &imap);
 745
 746                 if ((numblks_fsb = imap.br_blockcount) == 0) {
 747                         /*
 748                          * The numblks_fsb value should always get
 749                          * smaller, otherwise the loop is stuck.
 750                          */
 751                         ASSERT(imap.br_blockcount);
 752                         break;
 753                 }
 754                 offset_fsb += numblks_fsb;
 755                 count_fsb -= numblks_fsb;
 756         } while (count_fsb > 0);
 757
 758         return 0;
 759
 760 error_on_bmapi_transaction:
 761         xfs_bmap_cancel(&free_list);
 762         xfs_trans_cancel(tp, (XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT));
 763         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 764         return XFS_ERROR(error);
 765 }