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