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