SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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_dir2.h"
  27 #include "xfs_alloc.h"
  28 #include "xfs_dmapi.h"
  29 #include "xfs_quota.h"
  30 #include "xfs_mount.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_alloc_btree.h"
  33 #include "xfs_ialloc_btree.h"
  34 #include "xfs_dir2_sf.h"
  35 #include "xfs_attr_sf.h"
  36 #include "xfs_dinode.h"
  37 #include "xfs_inode.h"
  38 #include "xfs_ialloc.h"
  39 #include "xfs_btree.h"
  40 #include "xfs_bmap.h"
  41 #include "xfs_rtalloc.h"
  42 #include "xfs_error.h"
  43 #include "xfs_itable.h"
  44 #include "xfs_rw.h"
  45 #include "xfs_attr.h"
  46 #include "xfs_buf_item.h"
  47 #include "xfs_trans_space.h"
  48 #include "xfs_utils.h"
  49 #include "xfs_iomap.h"
  50 #include "xfs_trace.h"
  51
  52
  53 #define XFS_WRITEIO_ALIGN(mp,off)       (((off) >> mp->m_writeio_log) \
  54                                                 << mp->m_writeio_log)
  55 #define XFS_STRAT_WRITE_IMAPS   2
  56 #define XFS_WRITE_IMAPS         XFS_BMAP_MAX_NMAP
  57
  58 STATIC void
  59 xfs_imap_to_bmap(
  60         xfs_inode_t     *ip,
  61         xfs_off_t       offset,
  62         xfs_bmbt_irec_t *imap,
  63         xfs_iomap_t     *iomapp,
  64         int             imaps,                  /* Number of imap entries */
  65         int             flags)
  66 {
  67         xfs_fsblock_t   start_block;
  68
  69         iomapp->iomap_offset = imap->br_startoff;
  70         iomapp->iomap_bsize = imap->br_blockcount;
  71         iomapp->iomap_flags = flags;
  72
  73         start_block = imap->br_startblock;
  74         if (start_block == HOLESTARTBLOCK) {
  75                 iomapp->iomap_bn = IOMAP_DADDR_NULL;
  76                 iomapp->iomap_flags |= IOMAP_HOLE;
  77         } else if (start_block == DELAYSTARTBLOCK) {
  78                 iomapp->iomap_bn = IOMAP_DADDR_NULL;
  79                 iomapp->iomap_flags |= IOMAP_DELAY;
  80         } else {
  81                 iomapp->iomap_bn = xfs_fsb_to_db(ip, start_block);
  82                 if (ISUNWRITTEN(imap))
  83                         iomapp->iomap_flags |= IOMAP_UNWRITTEN;
  84         }
  85 }
  86
  87 int
  88 xfs_iomap(
  89         xfs_inode_t     *ip,
  90         xfs_off_t       offset,
  91         ssize_t         count,
  92         int             flags,
  93         xfs_iomap_t     *iomapp,
  94         int             *niomaps)
  95 {
  96         xfs_mount_t     *mp = ip->i_mount;
  97         xfs_fileoff_t   offset_fsb, end_fsb;
  98         int             error = 0;
  99         int             lockmode = 0;
 100         xfs_bmbt_irec_t imap;
 101         int             nimaps = 1;
 102         int             bmapi_flags = 0;
 103         int             iomap_flags = 0;
 104
 105         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
 106         ASSERT(niomaps && *niomaps == 1);
 107
 108         if (XFS_FORCED_SHUTDOWN(mp))
 109                 return XFS_ERROR(EIO);
 110
 111         trace_xfs_iomap_enter(ip, offset, count, flags, NULL);
 112
 113         switch (flags & (BMAPI_READ | BMAPI_WRITE | BMAPI_ALLOCATE)) {
 114         case BMAPI_READ:
 115                 lockmode = xfs_ilock_map_shared(ip);
 116                 bmapi_flags = XFS_BMAPI_ENTIRE;
 117                 break;
 118         case BMAPI_WRITE:
 119                 lockmode = XFS_ILOCK_EXCL;
 120                 if (flags & BMAPI_IGNSTATE)
 121                         bmapi_flags |= XFS_BMAPI_IGSTATE|XFS_BMAPI_ENTIRE;
 122                 xfs_ilock(ip, lockmode);
 123                 break;
 124         case BMAPI_ALLOCATE:
 125                 lockmode = XFS_ILOCK_SHARED;
 126                 bmapi_flags = XFS_BMAPI_ENTIRE;
 127
 128                 /* Attempt non-blocking lock */
 129                 if (flags & BMAPI_TRYLOCK) {
 130                         if (!xfs_ilock_nowait(ip, lockmode))
 131                                 return XFS_ERROR(EAGAIN);
 132                 } else {
 133                         xfs_ilock(ip, lockmode);
 134                 }
 135                 break;
 136         default:
 137                 BUG();
 138         }
 139
 140         ASSERT(offset <= mp->m_maxioffset);
 141         if ((xfs_fsize_t)offset + count > mp->m_maxioffset)
 142                 count = mp->m_maxioffset - offset;
 143         end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
 144         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 145
 146         error = xfs_bmapi(NULL, ip, offset_fsb,
 147                         (xfs_filblks_t)(end_fsb - offset_fsb),
 148                         bmapi_flags,  NULL, 0, &imap,
 149                         &nimaps, NULL, NULL);
 150
 151         if (error)
 152                 goto out;
 153
 154         switch (flags & (BMAPI_WRITE|BMAPI_ALLOCATE)) {
 155         case BMAPI_WRITE:
 156                 /* If we found an extent, return it */
 157                 if (nimaps &&
 158                     (imap.br_startblock != HOLESTARTBLOCK) &&
 159                     (imap.br_startblock != DELAYSTARTBLOCK)) {
 160                         trace_xfs_iomap_found(ip, offset, count, flags, &imap);
 161                         break;
 162                 }
 163
 164                 if (flags & (BMAPI_DIRECT|BMAPI_MMAP)) {
 165                         error = xfs_iomap_write_direct(ip, offset, count, flags,
 166                                                        &imap, &nimaps, nimaps);
 167                 } else {
 168                         error = xfs_iomap_write_delay(ip, offset, count, flags,
 169                                                       &imap, &nimaps);
 170                 }
 171                 if (!error) {
 172                         trace_xfs_iomap_alloc(ip, offset, count, flags, &imap);
 173                 }
 174                 iomap_flags = IOMAP_NEW;
 175                 break;
 176         case BMAPI_ALLOCATE:
 177                 /* If we found an extent, return it */
 178                 xfs_iunlock(ip, lockmode);
 179                 lockmode = 0;
 180
 181                 if (nimaps && !isnullstartblock(imap.br_startblock)) {
 182                         trace_xfs_iomap_found(ip, offset, count, flags, &imap);
 183                         break;
 184                 }
 185
 186                 error = xfs_iomap_write_allocate(ip, offset, count,
 187                                                  &imap, &nimaps);
 188                 break;
 189         }
 190
 191         ASSERT(nimaps <= 1);
 192
 193         if (nimaps)
 194                 xfs_imap_to_bmap(ip, offset, &imap, iomapp, nimaps, iomap_flags);
 195         *niomaps = nimaps;
 196
 197 out:
 198         if (lockmode)
 199                 xfs_iunlock(ip, lockmode);
 200         return XFS_ERROR(error);
 201 }
 202
 203
 204 STATIC int
 205 xfs_iomap_eof_align_last_fsb(
 206         xfs_mount_t     *mp,
 207         xfs_inode_t     *ip,
 208         xfs_extlen_t    extsize,
 209         xfs_fileoff_t   *last_fsb)
 210 {
 211         xfs_fileoff_t   new_last_fsb = 0;
 212         xfs_extlen_t    align;
 213         int             eof, error;
 214
 215         if (XFS_IS_REALTIME_INODE(ip))
 216                 ;
 217         /*
 218          * If mounted with the "-o swalloc" option, roundup the allocation
 219          * request to a stripe width boundary if the file size is >=
 220          * stripe width and we are allocating past the allocation eof.
 221          */
 222         else if (mp->m_swidth && (mp->m_flags & XFS_MOUNT_SWALLOC) &&
 223                 (ip->i_size >= XFS_FSB_TO_B(mp, mp->m_swidth)))
 224                 new_last_fsb = roundup_64(*last_fsb, mp->m_swidth);
 225         /*
 226          * Roundup the allocation request to a stripe unit (m_dalign) boundary
 227          * if the file size is >= stripe unit size, and we are allocating past
 228          * the allocation eof.
 229          */
 230         else if (mp->m_dalign && (ip->i_size >= XFS_FSB_TO_B(mp, mp->m_dalign)))
 231                 new_last_fsb = roundup_64(*last_fsb, mp->m_dalign);
 232
 233         /*
 234          * Always round up the allocation request to an extent boundary
 235          * (when file on a real-time subvolume or has di_extsize hint).
 236          */
 237         if (extsize) {
 238                 if (new_last_fsb)
 239                         align = roundup_64(new_last_fsb, extsize);
 240                 else
 241                         align = extsize;
 242                 new_last_fsb = roundup_64(*last_fsb, align);
 243         }
 244
 245         if (new_last_fsb) {
 246                 error = xfs_bmap_eof(ip, new_last_fsb, XFS_DATA_FORK, &eof);
 247                 if (error)
 248                         return error;
 249                 if (eof)
 250                         *last_fsb = new_last_fsb;
 251         }
 252         return 0;
 253 }
 254
 255 STATIC int
 256 xfs_cmn_err_fsblock_zero(
 257         xfs_inode_t     *ip,
 258         xfs_bmbt_irec_t *imap)
 259 {
 260         xfs_cmn_err(XFS_PTAG_FSBLOCK_ZERO, CE_ALERT, ip->i_mount,
 261                         "Access to block zero in inode %llu "
 262                         "start_block: %llx start_off: %llx "
 263                         "blkcnt: %llx extent-state: %x\n",
 264                 (unsigned long long)ip->i_ino,
 265                 (unsigned long long)imap->br_startblock,
 266                 (unsigned long long)imap->br_startoff,
 267                 (unsigned long long)imap->br_blockcount,
 268                 imap->br_state);
 269         return EFSCORRUPTED;
 270 }
 271
 272 int
 273 xfs_iomap_write_direct(
 274         xfs_inode_t     *ip,
 275         xfs_off_t       offset,
 276         size_t          count,
 277         int             flags,
 278         xfs_bmbt_irec_t *ret_imap,
 279         int             *nmaps,
 280         int             found)
 281 {
 282         xfs_mount_t     *mp = ip->i_mount;
 283         xfs_fileoff_t   offset_fsb;
 284         xfs_fileoff_t   last_fsb;
 285         xfs_filblks_t   count_fsb, resaligned;
 286         xfs_fsblock_t   firstfsb;
 287         xfs_extlen_t    extsz, temp;
 288         int             nimaps;
 289         int             bmapi_flag;
 290         int             quota_flag;
 291         int             rt;
 292         xfs_trans_t     *tp;
 293         xfs_bmbt_irec_t imap;
 294         xfs_bmap_free_t free_list;
 295         uint            qblocks, resblks, resrtextents;
 296         int             committed;
 297         int             error;
 298
 299         /*
 300          * Make sure that the dquots are there. This doesn't hold
 301          * the ilock across a disk read.
 302          */
 303         error = xfs_qm_dqattach_locked(ip, 0);
 304         if (error)
 305                 return XFS_ERROR(error);
 306
 307         rt = XFS_IS_REALTIME_INODE(ip);
 308         extsz = xfs_get_extsz_hint(ip);
 309
 310         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 311         last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
 312         if ((offset + count) > ip->i_size) {
 313                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
 314                 if (error)
 315                         goto error_out;
 316         } else {
 317                 if (found && (ret_imap->br_startblock == HOLESTARTBLOCK))
 318                         last_fsb = MIN(last_fsb, (xfs_fileoff_t)
 319                                         ret_imap->br_blockcount +
 320                                         ret_imap->br_startoff);
 321         }
 322         count_fsb = last_fsb - offset_fsb;
 323         ASSERT(count_fsb > 0);
 324
 325         resaligned = count_fsb;
 326         if (unlikely(extsz)) {
 327                 if ((temp = do_mod(offset_fsb, extsz)))
 328                         resaligned += temp;
 329                 if ((temp = do_mod(resaligned, extsz)))
 330                         resaligned += extsz - temp;
 331         }
 332
 333         if (unlikely(rt)) {
 334                 resrtextents = qblocks = resaligned;
 335                 resrtextents /= mp->m_sb.sb_rextsize;
 336                 resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
 337                 quota_flag = XFS_QMOPT_RES_RTBLKS;
 338         } else {
 339                 resrtextents = 0;
 340                 resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
 341                 quota_flag = XFS_QMOPT_RES_REGBLKS;
 342         }
 343
 344         /*
 345          * Allocate and setup the transaction
 346          */
 347         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 348         tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
 349         error = xfs_trans_reserve(tp, resblks,
 350                         XFS_WRITE_LOG_RES(mp), resrtextents,
 351                         XFS_TRANS_PERM_LOG_RES,
 352                         XFS_WRITE_LOG_COUNT);
 353         /*
 354          * Check for running out of space, note: need lock to return
 355          */
 356         if (error)
 357                 xfs_trans_cancel(tp, 0);
 358         xfs_ilock(ip, XFS_ILOCK_EXCL);
 359         if (error)
 360                 goto error_out;
 361
 362         error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
 363         if (error)
 364                 goto error1;
 365
 366         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 367         xfs_trans_ihold(tp, ip);
 368
 369         bmapi_flag = XFS_BMAPI_WRITE;
 370         if ((flags & BMAPI_DIRECT) && (offset < ip->i_size || extsz))
 371                 bmapi_flag |= XFS_BMAPI_PREALLOC;
 372
 373         /*
 374          * Issue the xfs_bmapi() call to allocate the blocks
 375          */
 376         xfs_bmap_init(&free_list, &firstfsb);
 377         nimaps = 1;
 378         error = xfs_bmapi(tp, ip, offset_fsb, count_fsb, bmapi_flag,
 379                 &firstfsb, 0, &imap, &nimaps, &free_list, NULL);
 380         if (error)
 381                 goto error0;
 382
 383         /*
 384          * Complete the transaction
 385          */
 386         error = xfs_bmap_finish(&tp, &free_list, &committed);
 387         if (error)
 388                 goto error0;
 389         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 390         if (error)
 391                 goto error_out;
 392
 393         /*
 394          * Copy any maps to caller's array and return any error.
 395          */
 396         if (nimaps == 0) {
 397                 error = ENOSPC;
 398                 goto error_out;
 399         }
 400
 401         if (!(imap.br_startblock || XFS_IS_REALTIME_INODE(ip))) {
 402                 error = xfs_cmn_err_fsblock_zero(ip, &imap);
 403                 goto error_out;
 404         }
 405
 406         *ret_imap = imap;
 407         *nmaps = 1;
 408         return 0;
 409
 410 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
 411         xfs_bmap_cancel(&free_list);
 412         xfs_trans_unreserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
 413
 414 error1: /* Just cancel transaction */
 415         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 416         *nmaps = 0;     /* nothing set-up here */
 417
 418 error_out:
 419         return XFS_ERROR(error);
 420 }
 421
 422 /*
 423  * If the caller is doing a write at the end of the file, then extend the
 424  * allocation out to the file system's write iosize.  We clean up any extra
 425  * space left over when the file is closed in xfs_inactive().
 426  */
 427 STATIC int
 428 xfs_iomap_eof_want_preallocate(
 429         xfs_mount_t     *mp,
 430         xfs_inode_t     *ip,
 431         xfs_off_t       offset,
 432         size_t          count,
 433         int             ioflag,
 434         xfs_bmbt_irec_t *imap,
 435         int             nimaps,
 436         int             *prealloc)
 437 {
 438         xfs_fileoff_t   start_fsb;
 439         xfs_filblks_t   count_fsb;
 440         xfs_fsblock_t   firstblock;
 441         int             n, error, imaps;
 442
 443         *prealloc = 0;
 444         if ((offset + count) <= ip->i_size)
 445                 return 0;
 446
 447         /*
 448          * If there are any real blocks past eof, then don't
 449          * do any speculative allocation.
 450          */
 451         start_fsb = XFS_B_TO_FSBT(mp, ((xfs_ufsize_t)(offset + count - 1)));
 452         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
 453         while (count_fsb > 0) {
 454                 imaps = nimaps;
 455                 firstblock = NULLFSBLOCK;
 456                 error = xfs_bmapi(NULL, ip, start_fsb, count_fsb, 0,
 457                                   &firstblock, 0, imap, &imaps, NULL, NULL);
 458                 if (error)
 459                         return error;
 460                 for (n = 0; n < imaps; n++) {
 461                         if ((imap[n].br_startblock != HOLESTARTBLOCK) &&
 462                             (imap[n].br_startblock != DELAYSTARTBLOCK))
 463                                 return 0;
 464                         start_fsb += imap[n].br_blockcount;
 465                         count_fsb -= imap[n].br_blockcount;
 466                 }
 467         }
 468         *prealloc = 1;
 469         return 0;
 470 }
 471
 472 int
 473 xfs_iomap_write_delay(
 474         xfs_inode_t     *ip,
 475         xfs_off_t       offset,
 476         size_t          count,
 477         int             ioflag,
 478         xfs_bmbt_irec_t *ret_imap,
 479         int             *nmaps)
 480 {
 481         xfs_mount_t     *mp = ip->i_mount;
 482         xfs_fileoff_t   offset_fsb;
 483         xfs_fileoff_t   last_fsb;
 484         xfs_off_t       aligned_offset;
 485         xfs_fileoff_t   ioalign;
 486         xfs_fsblock_t   firstblock;
 487         xfs_extlen_t    extsz;
 488         int             nimaps;
 489         xfs_bmbt_irec_t imap[XFS_WRITE_IMAPS];
 490         int             prealloc, flushed = 0;
 491         int             error;
 492
 493         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 494
 495         /*
 496          * Make sure that the dquots are there. This doesn't hold
 497          * the ilock across a disk read.
 498          */
 499         error = xfs_qm_dqattach_locked(ip, 0);
 500         if (error)
 501                 return XFS_ERROR(error);
 502
 503         extsz = xfs_get_extsz_hint(ip);
 504         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 505
 506         error = xfs_iomap_eof_want_preallocate(mp, ip, offset, count,
 507                                 ioflag, imap, XFS_WRITE_IMAPS, &prealloc);
 508         if (error)
 509                 return error;
 510
 511 retry:
 512         if (prealloc) {
 513                 aligned_offset = XFS_WRITEIO_ALIGN(mp, (offset + count - 1));
 514                 ioalign = XFS_B_TO_FSBT(mp, aligned_offset);
 515                 last_fsb = ioalign + mp->m_writeio_blocks;
 516         } else {
 517                 last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
 518         }
 519
 520         if (prealloc || extsz) {
 521                 error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
 522                 if (error)
 523                         return error;
 524         }
 525
 526         nimaps = XFS_WRITE_IMAPS;
 527         firstblock = NULLFSBLOCK;
 528         error = xfs_bmapi(NULL, ip, offset_fsb,
 529                           (xfs_filblks_t)(last_fsb - offset_fsb),
 530                           XFS_BMAPI_DELAY | XFS_BMAPI_WRITE |
 531                           XFS_BMAPI_ENTIRE, &firstblock, 1, imap,
 532                           &nimaps, NULL, NULL);
 533         if (error && (error != ENOSPC))
 534                 return XFS_ERROR(error);
 535
 536         /*
 537          * If bmapi returned us nothing, and if we didn't get back EDQUOT,
 538          * then we must have run out of space - flush all other inodes with
 539          * delalloc blocks and retry without EOF preallocation.
 540          */
 541         if (nimaps == 0) {
 542                 trace_xfs_delalloc_enospc(ip, offset, count);
 543                 if (flushed)
 544                         return XFS_ERROR(ENOSPC);
 545
 546                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 547                 xfs_flush_inodes(ip);
 548                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 549
 550                 flushed = 1;
 551                 error = 0;
 552                 prealloc = 0;
 553                 goto retry;
 554         }
 555
 556         if (!(imap[0].br_startblock || XFS_IS_REALTIME_INODE(ip)))
 557                 return xfs_cmn_err_fsblock_zero(ip, &imap[0]);
 558
 559         *ret_imap = imap[0];
 560         *nmaps = 1;
 561
 562         return 0;
 563 }
 564
 565 /*
 566  * Pass in a delayed allocate extent, convert it to real extents;
 567  * return to the caller the extent we create which maps on top of
 568  * the originating callers request.
 569  *
 570  * Called without a lock on the inode.
 571  *
 572  * We no longer bother to look at the incoming map - all we have to
 573  * guarantee is that whatever we allocate fills the required range.
 574  */
 575 int
 576 xfs_iomap_write_allocate(
 577         xfs_inode_t     *ip,
 578         xfs_off_t       offset,
 579         size_t          count,
 580         xfs_bmbt_irec_t *map,
 581         int             *retmap)
 582 {
 583         xfs_mount_t     *mp = ip->i_mount;
 584         xfs_fileoff_t   offset_fsb, last_block;
 585         xfs_fileoff_t   end_fsb, map_start_fsb;
 586         xfs_fsblock_t   first_block;
 587         xfs_bmap_free_t free_list;
 588         xfs_filblks_t   count_fsb;
 589         xfs_bmbt_irec_t imap;
 590         xfs_trans_t     *tp;
 591         int             nimaps, committed;
 592         int             error = 0;
 593         int             nres;
 594
 595         *retmap = 0;
 596
 597         /*
 598          * Make sure that the dquots are there.
 599          */
 600         error = xfs_qm_dqattach(ip, 0);
 601         if (error)
 602                 return XFS_ERROR(error);
 603
 604         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 605         count_fsb = map->br_blockcount;
 606         map_start_fsb = map->br_startoff;
 607
 608         XFS_STATS_ADD(xs_xstrat_bytes, XFS_FSB_TO_B(mp, count_fsb));
 609
 610         while (count_fsb != 0) {
 611                 /*
 612                  * Set up a transaction with which to allocate the
 613                  * backing store for the file.  Do allocations in a
 614                  * loop until we get some space in the range we are
 615                  * interested in.  The other space that might be allocated
 616                  * is in the delayed allocation extent on which we sit
 617                  * but before our buffer starts.
 618                  */
 619
 620                 nimaps = 0;
 621                 while (nimaps == 0) {
 622                         tp = xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE);
 623                         tp->t_flags |= XFS_TRANS_RESERVE;
 624                         nres = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
 625                         error = xfs_trans_reserve(tp, nres,
 626                                         XFS_WRITE_LOG_RES(mp),
 627                                         0, XFS_TRANS_PERM_LOG_RES,
 628                                         XFS_WRITE_LOG_COUNT);
 629                         if (error) {
 630                                 xfs_trans_cancel(tp, 0);
 631                                 return XFS_ERROR(error);
 632                         }
 633                         xfs_ilock(ip, XFS_ILOCK_EXCL);
 634                         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 635                         xfs_trans_ihold(tp, ip);
 636
 637                         xfs_bmap_init(&free_list, &first_block);
 638
 639                         /*
 640                          * it is possible that the extents have changed since
 641                          * we did the read call as we dropped the ilock for a
 642                          * while. We have to be careful about truncates or hole
 643                          * punchs here - we are not allowed to allocate
 644                          * non-delalloc blocks here.
 645                          *
 646                          * The only protection against truncation is the pages
 647                          * for the range we are being asked to convert are
 648                          * locked and hence a truncate will block on them
 649                          * first.
 650                          *
 651                          * As a result, if we go beyond the range we really
 652                          * need and hit an delalloc extent boundary followed by
 653                          * a hole while we have excess blocks in the map, we
 654                          * will fill the hole incorrectly and overrun the
 655                          * transaction reservation.
 656                          *
 657                          * Using a single map prevents this as we are forced to
 658                          * check each map we look for overlap with the desired
 659                          * range and abort as soon as we find it. Also, given
 660                          * that we only return a single map, having one beyond
 661                          * what we can return is probably a bit silly.
 662                          *
 663                          * We also need to check that we don't go beyond EOF;
 664                          * this is a truncate optimisation as a truncate sets
 665                          * the new file size before block on the pages we
 666                          * currently have locked under writeback. Because they
 667                          * are about to be tossed, we don't need to write them
 668                          * back....
 669                          */
 670                         nimaps = 1;
 671                         end_fsb = XFS_B_TO_FSB(mp, ip->i_size);
 672                         error = xfs_bmap_last_offset(NULL, ip, &last_block,
 673                                                         XFS_DATA_FORK);
 674                         if (error)
 675                                 goto trans_cancel;
 676
 677                         last_block = XFS_FILEOFF_MAX(last_block, end_fsb);
 678                         if ((map_start_fsb + count_fsb) > last_block) {
 679                                 count_fsb = last_block - map_start_fsb;
 680                                 if (count_fsb == 0) {
 681                                         error = EAGAIN;
 682                                         goto trans_cancel;
 683                                 }
 684                         }
 685
 686                         /* Go get the actual blocks */
 687                         error = xfs_bmapi(tp, ip, map_start_fsb, count_fsb,
 688                                         XFS_BMAPI_WRITE, &first_block, 1,
 689                                         &imap, &nimaps, &free_list, NULL);
 690                         if (error)
 691                                 goto trans_cancel;
 692
 693                         error = xfs_bmap_finish(&tp, &free_list, &committed);
 694                         if (error)
 695                                 goto trans_cancel;
 696
 697                         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 698                         if (error)
 699                                 goto error0;
 700
 701                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 702                 }
 703
 704                 /*
 705                  * See if we were able to allocate an extent that
 706                  * covers at least part of the callers request
 707                  */
 708                 if (!(imap.br_startblock || XFS_IS_REALTIME_INODE(ip)))
 709                         return xfs_cmn_err_fsblock_zero(ip, &imap);
 710
 711                 if ((offset_fsb >= imap.br_startoff) &&
 712                     (offset_fsb < (imap.br_startoff +
 713                                    imap.br_blockcount))) {
 714                         *map = imap;
 715                         *retmap = 1;
 716                         XFS_STATS_INC(xs_xstrat_quick);
 717                         return 0;
 718                 }
 719
 720                 /*
 721                  * So far we have not mapped the requested part of the
 722                  * file, just surrounding data, try again.
 723                  */
 724                 count_fsb -= imap.br_blockcount;
 725                 map_start_fsb = imap.br_startoff + imap.br_blockcount;
 726         }
 727
 728 trans_cancel:
 729         xfs_bmap_cancel(&free_list);
 730         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 731 error0:
 732         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 733         return XFS_ERROR(error);
 734 }
 735
 736 int
 737 xfs_iomap_write_unwritten(
 738         xfs_inode_t     *ip,
 739         xfs_off_t       offset,
 740         size_t          count)
 741 {
 742         xfs_mount_t     *mp = ip->i_mount;
 743         xfs_fileoff_t   offset_fsb;
 744         xfs_filblks_t   count_fsb;
 745         xfs_filblks_t   numblks_fsb;
 746         xfs_fsblock_t   firstfsb;
 747         int             nimaps;
 748         xfs_trans_t     *tp;
 749         xfs_bmbt_irec_t imap;
 750         xfs_bmap_free_t free_list;
 751         uint            resblks;
 752         int             committed;
 753         int             error;
 754
 755         trace_xfs_unwritten_convert(ip, offset, count);
 756
 757         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 758         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
 759         count_fsb = (xfs_filblks_t)(count_fsb - offset_fsb);
 760
 761         /*
 762          * Reserve enough blocks in this transaction for two complete extent
 763          * btree splits.  We may be converting the middle part of an unwritten
 764          * extent and in this case we will insert two new extents in the btree
 765          * each of which could cause a full split.
 766          *
 767          * This reservation amount will be used in the first call to
 768          * xfs_bmbt_split() to select an AG with enough space to satisfy the
 769          * rest of the operation.
 770          */
 771         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
 772
 773         do {
 774                 /*
 775                  * set up a transaction to convert the range of extents
 776                  * from unwritten to real. Do allocations in a loop until
 777                  * we have covered the range passed in.
 778                  *
 779                  * Note that we open code the transaction allocation here
 780                  * to pass KM_NOFS--we can't risk to recursing back into
 781                  * the filesystem here as we might be asked to write out
 782                  * the same inode that we complete here and might deadlock
 783                  * on the iolock.
 784                  */
 785                 xfs_wait_for_freeze(mp, SB_FREEZE_TRANS);
 786                 tp = _xfs_trans_alloc(mp, XFS_TRANS_STRAT_WRITE, KM_NOFS);
 787                 tp->t_flags |= XFS_TRANS_RESERVE;
 788                 error = xfs_trans_reserve(tp, resblks,
 789                                 XFS_WRITE_LOG_RES(mp), 0,
 790                                 XFS_TRANS_PERM_LOG_RES,
 791                                 XFS_WRITE_LOG_COUNT);
 792                 if (error) {
 793                         xfs_trans_cancel(tp, 0);
 794                         return XFS_ERROR(error);
 795                 }
 796
 797                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 798                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 799                 xfs_trans_ihold(tp, ip);
 800
 801                 /*
 802                  * Modify the unwritten extent state of the buffer.
 803                  */
 804                 xfs_bmap_init(&free_list, &firstfsb);
 805                 nimaps = 1;
 806                 error = xfs_bmapi(tp, ip, offset_fsb, count_fsb,
 807                                   XFS_BMAPI_WRITE|XFS_BMAPI_CONVERT, &firstfsb,
 808                                   1, &imap, &nimaps, &free_list, NULL);
 809                 if (error)
 810                         goto error_on_bmapi_transaction;
 811
 812                 error = xfs_bmap_finish(&(tp), &(free_list), &committed);
 813                 if (error)
 814                         goto error_on_bmapi_transaction;
 815
 816                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 817                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 818                 if (error)
 819                         return XFS_ERROR(error);
 820
 821                 if (!(imap.br_startblock || XFS_IS_REALTIME_INODE(ip)))
 822                         return xfs_cmn_err_fsblock_zero(ip, &imap);
 823
 824                 if ((numblks_fsb = imap.br_blockcount) == 0) {
 825                         /*
 826                          * The numblks_fsb value should always get
 827                          * smaller, otherwise the loop is stuck.
 828                          */
 829                         ASSERT(imap.br_blockcount);
 830                         break;
 831                 }
 832                 offset_fsb += numblks_fsb;
 833                 count_fsb -= numblks_fsb;
 834         } while (count_fsb > 0);
 835
 836         return 0;
 837
 838 error_on_bmapi_transaction:
 839         xfs_bmap_cancel(&free_list);
 840         xfs_trans_cancel(tp, (XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT));
 841         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 842         return XFS_ERROR(error);
 843 }