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