fs/xfs/xfs_iomap.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   4  * Copyright (c) 2016-2018 Christoph Hellwig.
   5  * All Rights Reserved.
   6  */
   7 #include "xfs.h"
   8 #include "xfs_fs.h"
   9 #include "xfs_shared.h"
  10 #include "xfs_format.h"
  11 #include "xfs_log_format.h"
  12 #include "xfs_trans_resv.h"
  13 #include "xfs_mount.h"
  14 #include "xfs_inode.h"
  15 #include "xfs_btree.h"
  16 #include "xfs_bmap_btree.h"
  17 #include "xfs_bmap.h"
  18 #include "xfs_bmap_util.h"
  19 #include "xfs_errortag.h"
  20 #include "xfs_error.h"
  21 #include "xfs_trans.h"
  22 #include "xfs_trans_space.h"
  23 #include "xfs_inode_item.h"
  24 #include "xfs_iomap.h"
  25 #include "xfs_trace.h"
  26 #include "xfs_quota.h"
  27 #include "xfs_dquot_item.h"
  28 #include "xfs_dquot.h"
  29 #include "xfs_reflink.h"
  30 #include "xfs_health.h"
  31
  32 #define XFS_ALLOC_ALIGN(mp, off) \
  33         (((off) >> mp->m_allocsize_log) << mp->m_allocsize_log)
  34
  35 static int
  36 xfs_alert_fsblock_zero(
  37         xfs_inode_t     *ip,
  38         xfs_bmbt_irec_t *imap)
  39 {
  40         xfs_alert_tag(ip->i_mount, XFS_PTAG_FSBLOCK_ZERO,
  41                         "Access to block zero in inode %llu "
  42                         "start_block: %llx start_off: %llx "
  43                         "blkcnt: %llx extent-state: %x",
  44                 (unsigned long long)ip->i_ino,
  45                 (unsigned long long)imap->br_startblock,
  46                 (unsigned long long)imap->br_startoff,
  47                 (unsigned long long)imap->br_blockcount,
  48                 imap->br_state);
  49         xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
  50         return -EFSCORRUPTED;
  51 }
  52
  53 u64
  54 xfs_iomap_inode_sequence(
  55         struct xfs_inode        *ip,
  56         u16                     iomap_flags)
  57 {
  58         u64                     cookie = 0;
  59
  60         if (iomap_flags & IOMAP_F_XATTR)
  61                 return READ_ONCE(ip->i_af.if_seq);
  62         if ((iomap_flags & IOMAP_F_SHARED) && ip->i_cowfp)
  63                 cookie = (u64)READ_ONCE(ip->i_cowfp->if_seq) << 32;
  64         return cookie | READ_ONCE(ip->i_df.if_seq);
  65 }
  66
  67 /*
  68  * Check that the iomap passed to us is still valid for the given offset and
  69  * length.
  70  */
  71 static bool
  72 xfs_iomap_valid(
  73         struct inode            *inode,
  74         const struct iomap      *iomap)
  75 {
  76         struct xfs_inode        *ip = XFS_I(inode);
  77
  78         if (iomap->validity_cookie !=
  79                         xfs_iomap_inode_sequence(ip, iomap->flags)) {
  80                 trace_xfs_iomap_invalid(ip, iomap);
  81                 return false;
  82         }
  83
  84         XFS_ERRORTAG_DELAY(ip->i_mount, XFS_ERRTAG_WRITE_DELAY_MS);
  85         return true;
  86 }
  87
  88 static const struct iomap_folio_ops xfs_iomap_folio_ops = {
  89         .iomap_valid            = xfs_iomap_valid,
  90 };
  91
  92 int
  93 xfs_bmbt_to_iomap(
  94         struct xfs_inode        *ip,
  95         struct iomap            *iomap,
  96         struct xfs_bmbt_irec    *imap,
  97         unsigned int            mapping_flags,
  98         u16                     iomap_flags,
  99         u64                     sequence_cookie)
 100 {
 101         struct xfs_mount        *mp = ip->i_mount;
 102         struct xfs_buftarg      *target = xfs_inode_buftarg(ip);
 103
 104         if (unlikely(!xfs_valid_startblock(ip, imap->br_startblock))) {
 105                 xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
 106                 return xfs_alert_fsblock_zero(ip, imap);
 107         }
 108
 109         if (imap->br_startblock == HOLESTARTBLOCK) {
 110                 iomap->addr = IOMAP_NULL_ADDR;
 111                 iomap->type = IOMAP_HOLE;
 112         } else if (imap->br_startblock == DELAYSTARTBLOCK ||
 113                    isnullstartblock(imap->br_startblock)) {
 114                 iomap->addr = IOMAP_NULL_ADDR;
 115                 iomap->type = IOMAP_DELALLOC;
 116         } else {
 117                 iomap->addr = BBTOB(xfs_fsb_to_db(ip, imap->br_startblock));
 118                 if (mapping_flags & IOMAP_DAX)
 119                         iomap->addr += target->bt_dax_part_off;
 120
 121                 if (imap->br_state == XFS_EXT_UNWRITTEN)
 122                         iomap->type = IOMAP_UNWRITTEN;
 123                 else
 124                         iomap->type = IOMAP_MAPPED;
 125
 126         }
 127         iomap->offset = XFS_FSB_TO_B(mp, imap->br_startoff);
 128         iomap->length = XFS_FSB_TO_B(mp, imap->br_blockcount);
 129         if (mapping_flags & IOMAP_DAX)
 130                 iomap->dax_dev = target->bt_daxdev;
 131         else
 132                 iomap->bdev = target->bt_bdev;
 133         iomap->flags = iomap_flags;
 134
 135         if (xfs_ipincount(ip) &&
 136             (ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
 137                 iomap->flags |= IOMAP_F_DIRTY;
 138
 139         iomap->validity_cookie = sequence_cookie;
 140         iomap->folio_ops = &xfs_iomap_folio_ops;
 141         return 0;
 142 }
 143
 144 static void
 145 xfs_hole_to_iomap(
 146         struct xfs_inode        *ip,
 147         struct iomap            *iomap,
 148         xfs_fileoff_t           offset_fsb,
 149         xfs_fileoff_t           end_fsb)
 150 {
 151         struct xfs_buftarg      *target = xfs_inode_buftarg(ip);
 152
 153         iomap->addr = IOMAP_NULL_ADDR;
 154         iomap->type = IOMAP_HOLE;
 155         iomap->offset = XFS_FSB_TO_B(ip->i_mount, offset_fsb);
 156         iomap->length = XFS_FSB_TO_B(ip->i_mount, end_fsb - offset_fsb);
 157         iomap->bdev = target->bt_bdev;
 158         iomap->dax_dev = target->bt_daxdev;
 159 }
 160
 161 static inline xfs_fileoff_t
 162 xfs_iomap_end_fsb(
 163         struct xfs_mount        *mp,
 164         loff_t                  offset,
 165         loff_t                  count)
 166 {
 167         ASSERT(offset <= mp->m_super->s_maxbytes);
 168         return min(XFS_B_TO_FSB(mp, offset + count),
 169                    XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes));
 170 }
 171
 172 static xfs_extlen_t
 173 xfs_eof_alignment(
 174         struct xfs_inode        *ip)
 175 {
 176         struct xfs_mount        *mp = ip->i_mount;
 177         xfs_extlen_t            align = 0;
 178
 179         if (!XFS_IS_REALTIME_INODE(ip)) {
 180                 /*
 181                  * Round up the allocation request to a stripe unit
 182                  * (m_dalign) boundary if the file size is >= stripe unit
 183                  * size, and we are allocating past the allocation eof.
 184                  *
 185                  * If mounted with the "-o swalloc" option the alignment is
 186                  * increased from the strip unit size to the stripe width.
 187                  */
 188                 if (mp->m_swidth && xfs_has_swalloc(mp))
 189                         align = mp->m_swidth;
 190                 else if (mp->m_dalign)
 191                         align = mp->m_dalign;
 192
 193                 if (align && XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, align))
 194                         align = 0;
 195         }
 196
 197         return align;
 198 }
 199
 200 /*
 201  * Check if last_fsb is outside the last extent, and if so grow it to the next
 202  * stripe unit boundary.
 203  */
 204 xfs_fileoff_t
 205 xfs_iomap_eof_align_last_fsb(
 206         struct xfs_inode        *ip,
 207         xfs_fileoff_t           end_fsb)
 208 {
 209         struct xfs_ifork        *ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
 210         xfs_extlen_t            extsz = xfs_get_extsz_hint(ip);
 211         xfs_extlen_t            align = xfs_eof_alignment(ip);
 212         struct xfs_bmbt_irec    irec;
 213         struct xfs_iext_cursor  icur;
 214
 215         ASSERT(!xfs_need_iread_extents(ifp));
 216
 217         /*
 218          * Always round up the allocation request to the extent hint boundary.
 219          */
 220         if (extsz) {
 221                 if (align)
 222                         align = roundup_64(align, extsz);
 223                 else
 224                         align = extsz;
 225         }
 226
 227         if (align) {
 228                 xfs_fileoff_t   aligned_end_fsb = roundup_64(end_fsb, align);
 229
 230                 xfs_iext_last(ifp, &icur);
 231                 if (!xfs_iext_get_extent(ifp, &icur, &irec) ||
 232                     aligned_end_fsb >= irec.br_startoff + irec.br_blockcount)
 233                         return aligned_end_fsb;
 234         }
 235
 236         return end_fsb;
 237 }
 238
 239 int
 240 xfs_iomap_write_direct(
 241         struct xfs_inode        *ip,
 242         xfs_fileoff_t           offset_fsb,
 243         xfs_fileoff_t           count_fsb,
 244         unsigned int            flags,
 245         struct xfs_bmbt_irec    *imap,
 246         u64                     *seq)
 247 {
 248         struct xfs_mount        *mp = ip->i_mount;
 249         struct xfs_trans        *tp;
 250         xfs_filblks_t           resaligned;
 251         int                     nimaps;
 252         unsigned int            dblocks, rblocks;
 253         bool                    force = false;
 254         int                     error;
 255         int                     bmapi_flags = XFS_BMAPI_PREALLOC;
 256         int                     nr_exts = XFS_IEXT_ADD_NOSPLIT_CNT;
 257
 258         ASSERT(count_fsb > 0);
 259
 260         resaligned = xfs_aligned_fsb_count(offset_fsb, count_fsb,
 261                                            xfs_get_extsz_hint(ip));
 262         if (unlikely(XFS_IS_REALTIME_INODE(ip))) {
 263                 dblocks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
 264                 rblocks = resaligned;
 265         } else {
 266                 dblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
 267                 rblocks = 0;
 268         }
 269
 270         error = xfs_qm_dqattach(ip);
 271         if (error)
 272                 return error;
 273
 274         /*
 275          * For DAX, we do not allocate unwritten extents, but instead we zero
 276          * the block before we commit the transaction.  Ideally we'd like to do
 277          * this outside the transaction context, but if we commit and then crash
 278          * we may not have zeroed the blocks and this will be exposed on
 279          * recovery of the allocation. Hence we must zero before commit.
 280          *
 281          * Further, if we are mapping unwritten extents here, we need to zero
 282          * and convert them to written so that we don't need an unwritten extent
 283          * callback for DAX. This also means that we need to be able to dip into
 284          * the reserve block pool for bmbt block allocation if there is no space
 285          * left but we need to do unwritten extent conversion.
 286          */
 287         if (flags & IOMAP_DAX) {
 288                 bmapi_flags = XFS_BMAPI_CONVERT | XFS_BMAPI_ZERO;
 289                 if (imap->br_state == XFS_EXT_UNWRITTEN) {
 290                         force = true;
 291                         nr_exts = XFS_IEXT_WRITE_UNWRITTEN_CNT;
 292                         dblocks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
 293                 }
 294         }
 295
 296         error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, dblocks,
 297                         rblocks, force, &tp);
 298         if (error)
 299                 return error;
 300
 301         error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK, nr_exts);
 302         if (error == -EFBIG)
 303                 error = xfs_iext_count_upgrade(tp, ip, nr_exts);
 304         if (error)
 305                 goto out_trans_cancel;
 306
 307         /*
 308          * From this point onwards we overwrite the imap pointer that the
 309          * caller gave to us.
 310          */
 311         nimaps = 1;
 312         error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb, bmapi_flags, 0,
 313                                 imap, &nimaps);
 314         if (error)
 315                 goto out_trans_cancel;
 316
 317         /*
 318          * Complete the transaction
 319          */
 320         error = xfs_trans_commit(tp);
 321         if (error)
 322                 goto out_unlock;
 323
 324         /*
 325          * Copy any maps to caller's array and return any error.
 326          */
 327         if (nimaps == 0) {
 328                 error = -ENOSPC;
 329                 goto out_unlock;
 330         }
 331
 332         if (unlikely(!xfs_valid_startblock(ip, imap->br_startblock))) {
 333                 xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
 334                 error = xfs_alert_fsblock_zero(ip, imap);
 335         }
 336
 337 out_unlock:
 338         *seq = xfs_iomap_inode_sequence(ip, 0);
 339         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 340         return error;
 341
 342 out_trans_cancel:
 343         xfs_trans_cancel(tp);
 344         goto out_unlock;
 345 }
 346
 347 STATIC bool
 348 xfs_quota_need_throttle(
 349         struct xfs_inode        *ip,
 350         xfs_dqtype_t            type,
 351         xfs_fsblock_t           alloc_blocks)
 352 {
 353         struct xfs_dquot        *dq = xfs_inode_dquot(ip, type);
 354
 355         if (!dq || !xfs_this_quota_on(ip->i_mount, type))
 356                 return false;
 357
 358         /* no hi watermark, no throttle */
 359         if (!dq->q_prealloc_hi_wmark)
 360                 return false;
 361
 362         /* under the lo watermark, no throttle */
 363         if (dq->q_blk.reserved + alloc_blocks < dq->q_prealloc_lo_wmark)
 364                 return false;
 365
 366         return true;
 367 }
 368
 369 STATIC void
 370 xfs_quota_calc_throttle(
 371         struct xfs_inode        *ip,
 372         xfs_dqtype_t            type,
 373         xfs_fsblock_t           *qblocks,
 374         int                     *qshift,
 375         int64_t                 *qfreesp)
 376 {
 377         struct xfs_dquot        *dq = xfs_inode_dquot(ip, type);
 378         int64_t                 freesp;
 379         int                     shift = 0;
 380
 381         /* no dq, or over hi wmark, squash the prealloc completely */
 382         if (!dq || dq->q_blk.reserved >= dq->q_prealloc_hi_wmark) {
 383                 *qblocks = 0;
 384                 *qfreesp = 0;
 385                 return;
 386         }
 387
 388         freesp = dq->q_prealloc_hi_wmark - dq->q_blk.reserved;
 389         if (freesp < dq->q_low_space[XFS_QLOWSP_5_PCNT]) {
 390                 shift = 2;
 391                 if (freesp < dq->q_low_space[XFS_QLOWSP_3_PCNT])
 392                         shift += 2;
 393                 if (freesp < dq->q_low_space[XFS_QLOWSP_1_PCNT])
 394                         shift += 2;
 395         }
 396
 397         if (freesp < *qfreesp)
 398                 *qfreesp = freesp;
 399
 400         /* only overwrite the throttle values if we are more aggressive */
 401         if ((freesp >> shift) < (*qblocks >> *qshift)) {
 402                 *qblocks = freesp;
 403                 *qshift = shift;
 404         }
 405 }
 406
 407 /*
 408  * If we don't have a user specified preallocation size, dynamically increase
 409  * the preallocation size as the size of the file grows.  Cap the maximum size
 410  * at a single extent or less if the filesystem is near full. The closer the
 411  * filesystem is to being full, the smaller the maximum preallocation.
 412  */
 413 STATIC xfs_fsblock_t
 414 xfs_iomap_prealloc_size(
 415         struct xfs_inode        *ip,
 416         int                     whichfork,
 417         loff_t                  offset,
 418         loff_t                  count,
 419         struct xfs_iext_cursor  *icur)
 420 {
 421         struct xfs_iext_cursor  ncur = *icur;
 422         struct xfs_bmbt_irec    prev, got;
 423         struct xfs_mount        *mp = ip->i_mount;
 424         struct xfs_ifork        *ifp = xfs_ifork_ptr(ip, whichfork);
 425         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
 426         int64_t                 freesp;
 427         xfs_fsblock_t           qblocks;
 428         xfs_fsblock_t           alloc_blocks = 0;
 429         xfs_extlen_t            plen;
 430         int                     shift = 0;
 431         int                     qshift = 0;
 432
 433         /*
 434          * As an exception we don't do any preallocation at all if the file is
 435          * smaller than the minimum preallocation and we are using the default
 436          * dynamic preallocation scheme, as it is likely this is the only write
 437          * to the file that is going to be done.
 438          */
 439         if (XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, mp->m_allocsize_blocks))
 440                 return 0;
 441
 442         /*
 443          * Use the minimum preallocation size for small files or if we are
 444          * writing right after a hole.
 445          */
 446         if (XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, mp->m_dalign) ||
 447             !xfs_iext_prev_extent(ifp, &ncur, &prev) ||
 448             prev.br_startoff + prev.br_blockcount < offset_fsb)
 449                 return mp->m_allocsize_blocks;
 450
 451         /*
 452          * Take the size of the preceding data extents as the basis for the
 453          * preallocation size. Note that we don't care if the previous extents
 454          * are written or not.
 455          */
 456         plen = prev.br_blockcount;
 457         while (xfs_iext_prev_extent(ifp, &ncur, &got)) {
 458                 if (plen > XFS_MAX_BMBT_EXTLEN / 2 ||
 459                     isnullstartblock(got.br_startblock) ||
 460                     got.br_startoff + got.br_blockcount != prev.br_startoff ||
 461                     got.br_startblock + got.br_blockcount != prev.br_startblock)
 462                         break;
 463                 plen += got.br_blockcount;
 464                 prev = got;
 465         }
 466
 467         /*
 468          * If the size of the extents is greater than half the maximum extent
 469          * length, then use the current offset as the basis.  This ensures that
 470          * for large files the preallocation size always extends to
 471          * XFS_BMBT_MAX_EXTLEN rather than falling short due to things like stripe
 472          * unit/width alignment of real extents.
 473          */
 474         alloc_blocks = plen * 2;
 475         if (alloc_blocks > XFS_MAX_BMBT_EXTLEN)
 476                 alloc_blocks = XFS_B_TO_FSB(mp, offset);
 477         qblocks = alloc_blocks;
 478
 479         /*
 480          * XFS_BMBT_MAX_EXTLEN is not a power of two value but we round the prealloc
 481          * down to the nearest power of two value after throttling. To prevent
 482          * the round down from unconditionally reducing the maximum supported
 483          * prealloc size, we round up first, apply appropriate throttling, round
 484          * down and cap the value to XFS_BMBT_MAX_EXTLEN.
 485          */
 486         alloc_blocks = XFS_FILEOFF_MIN(roundup_pow_of_two(XFS_MAX_BMBT_EXTLEN),
 487                                        alloc_blocks);
 488
 489         freesp = percpu_counter_read_positive(&mp->m_fdblocks);
 490         if (freesp < mp->m_low_space[XFS_LOWSP_5_PCNT]) {
 491                 shift = 2;
 492                 if (freesp < mp->m_low_space[XFS_LOWSP_4_PCNT])
 493                         shift++;
 494                 if (freesp < mp->m_low_space[XFS_LOWSP_3_PCNT])
 495                         shift++;
 496                 if (freesp < mp->m_low_space[XFS_LOWSP_2_PCNT])
 497                         shift++;
 498                 if (freesp < mp->m_low_space[XFS_LOWSP_1_PCNT])
 499                         shift++;
 500         }
 501
 502         /*
 503          * Check each quota to cap the prealloc size, provide a shift value to
 504          * throttle with and adjust amount of available space.
 505          */
 506         if (xfs_quota_need_throttle(ip, XFS_DQTYPE_USER, alloc_blocks))
 507                 xfs_quota_calc_throttle(ip, XFS_DQTYPE_USER, &qblocks, &qshift,
 508                                         &freesp);
 509         if (xfs_quota_need_throttle(ip, XFS_DQTYPE_GROUP, alloc_blocks))
 510                 xfs_quota_calc_throttle(ip, XFS_DQTYPE_GROUP, &qblocks, &qshift,
 511                                         &freesp);
 512         if (xfs_quota_need_throttle(ip, XFS_DQTYPE_PROJ, alloc_blocks))
 513                 xfs_quota_calc_throttle(ip, XFS_DQTYPE_PROJ, &qblocks, &qshift,
 514                                         &freesp);
 515
 516         /*
 517          * The final prealloc size is set to the minimum of free space available
 518          * in each of the quotas and the overall filesystem.
 519          *
 520          * The shift throttle value is set to the maximum value as determined by
 521          * the global low free space values and per-quota low free space values.
 522          */
 523         alloc_blocks = min(alloc_blocks, qblocks);
 524         shift = max(shift, qshift);
 525
 526         if (shift)
 527                 alloc_blocks >>= shift;
 528         /*
 529          * rounddown_pow_of_two() returns an undefined result if we pass in
 530          * alloc_blocks = 0.
 531          */
 532         if (alloc_blocks)
 533                 alloc_blocks = rounddown_pow_of_two(alloc_blocks);
 534         if (alloc_blocks > XFS_MAX_BMBT_EXTLEN)
 535                 alloc_blocks = XFS_MAX_BMBT_EXTLEN;
 536
 537         /*
 538          * If we are still trying to allocate more space than is
 539          * available, squash the prealloc hard. This can happen if we
 540          * have a large file on a small filesystem and the above
 541          * lowspace thresholds are smaller than XFS_BMBT_MAX_EXTLEN.
 542          */
 543         while (alloc_blocks && alloc_blocks >= freesp)
 544                 alloc_blocks >>= 4;
 545         if (alloc_blocks < mp->m_allocsize_blocks)
 546                 alloc_blocks = mp->m_allocsize_blocks;
 547         trace_xfs_iomap_prealloc_size(ip, alloc_blocks, shift,
 548                                       mp->m_allocsize_blocks);
 549         return alloc_blocks;
 550 }
 551
 552 int
 553 xfs_iomap_write_unwritten(
 554         xfs_inode_t     *ip,
 555         xfs_off_t       offset,
 556         xfs_off_t       count,
 557         bool            update_isize)
 558 {
 559         xfs_mount_t     *mp = ip->i_mount;
 560         xfs_fileoff_t   offset_fsb;
 561         xfs_filblks_t   count_fsb;
 562         xfs_filblks_t   numblks_fsb;
 563         int             nimaps;
 564         xfs_trans_t     *tp;
 565         xfs_bmbt_irec_t imap;
 566         struct inode    *inode = VFS_I(ip);
 567         xfs_fsize_t     i_size;
 568         uint            resblks;
 569         int             error;
 570
 571         trace_xfs_unwritten_convert(ip, offset, count);
 572
 573         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 574         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
 575         count_fsb = (xfs_filblks_t)(count_fsb - offset_fsb);
 576
 577         /*
 578          * Reserve enough blocks in this transaction for two complete extent
 579          * btree splits.  We may be converting the middle part of an unwritten
 580          * extent and in this case we will insert two new extents in the btree
 581          * each of which could cause a full split.
 582          *
 583          * This reservation amount will be used in the first call to
 584          * xfs_bmbt_split() to select an AG with enough space to satisfy the
 585          * rest of the operation.
 586          */
 587         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
 588
 589         /* Attach dquots so that bmbt splits are accounted correctly. */
 590         error = xfs_qm_dqattach(ip);
 591         if (error)
 592                 return error;
 593
 594         do {
 595                 /*
 596                  * Set up a transaction to convert the range of extents
 597                  * from unwritten to real. Do allocations in a loop until
 598                  * we have covered the range passed in.
 599                  *
 600                  * Note that we can't risk to recursing back into the filesystem
 601                  * here as we might be asked to write out the same inode that we
 602                  * complete here and might deadlock on the iolock.
 603                  */
 604                 error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, resblks,
 605                                 0, true, &tp);
 606                 if (error)
 607                         return error;
 608
 609                 error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK,
 610                                 XFS_IEXT_WRITE_UNWRITTEN_CNT);
 611                 if (error == -EFBIG)
 612                         error = xfs_iext_count_upgrade(tp, ip,
 613                                         XFS_IEXT_WRITE_UNWRITTEN_CNT);
 614                 if (error)
 615                         goto error_on_bmapi_transaction;
 616
 617                 /*
 618                  * Modify the unwritten extent state of the buffer.
 619                  */
 620                 nimaps = 1;
 621                 error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb,
 622                                         XFS_BMAPI_CONVERT, resblks, &imap,
 623                                         &nimaps);
 624                 if (error)
 625                         goto error_on_bmapi_transaction;
 626
 627                 /*
 628                  * Log the updated inode size as we go.  We have to be careful
 629                  * to only log it up to the actual write offset if it is
 630                  * halfway into a block.
 631                  */
 632                 i_size = XFS_FSB_TO_B(mp, offset_fsb + count_fsb);
 633                 if (i_size > offset + count)
 634                         i_size = offset + count;
 635                 if (update_isize && i_size > i_size_read(inode))
 636                         i_size_write(inode, i_size);
 637                 i_size = xfs_new_eof(ip, i_size);
 638                 if (i_size) {
 639                         ip->i_disk_size = i_size;
 640                         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 641                 }
 642
 643                 error = xfs_trans_commit(tp);
 644                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 645                 if (error)
 646                         return error;
 647
 648                 if (unlikely(!xfs_valid_startblock(ip, imap.br_startblock))) {
 649                         xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
 650                         return xfs_alert_fsblock_zero(ip, &imap);
 651                 }
 652
 653                 if ((numblks_fsb = imap.br_blockcount) == 0) {
 654                         /*
 655                          * The numblks_fsb value should always get
 656                          * smaller, otherwise the loop is stuck.
 657                          */
 658                         ASSERT(imap.br_blockcount);
 659                         break;
 660                 }
 661                 offset_fsb += numblks_fsb;
 662                 count_fsb -= numblks_fsb;
 663         } while (count_fsb > 0);
 664
 665         return 0;
 666
 667 error_on_bmapi_transaction:
 668         xfs_trans_cancel(tp);
 669         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 670         return error;
 671 }
 672
 673 static inline bool
 674 imap_needs_alloc(
 675         struct inode            *inode,
 676         unsigned                flags,
 677         struct xfs_bmbt_irec    *imap,
 678         int                     nimaps)
 679 {
 680         /* don't allocate blocks when just zeroing */
 681         if (flags & IOMAP_ZERO)
 682                 return false;
 683         if (!nimaps ||
 684             imap->br_startblock == HOLESTARTBLOCK ||
 685             imap->br_startblock == DELAYSTARTBLOCK)
 686                 return true;
 687         /* we convert unwritten extents before copying the data for DAX */
 688         if ((flags & IOMAP_DAX) && imap->br_state == XFS_EXT_UNWRITTEN)
 689                 return true;
 690         return false;
 691 }
 692
 693 static inline bool
 694 imap_needs_cow(
 695         struct xfs_inode        *ip,
 696         unsigned int            flags,
 697         struct xfs_bmbt_irec    *imap,
 698         int                     nimaps)
 699 {
 700         if (!xfs_is_cow_inode(ip))
 701                 return false;
 702
 703         /* when zeroing we don't have to COW holes or unwritten extents */
 704         if (flags & IOMAP_ZERO) {
 705                 if (!nimaps ||
 706                     imap->br_startblock == HOLESTARTBLOCK ||
 707                     imap->br_state == XFS_EXT_UNWRITTEN)
 708                         return false;
 709         }
 710
 711         return true;
 712 }
 713
 714 static int
 715 xfs_ilock_for_iomap(
 716         struct xfs_inode        *ip,
 717         unsigned                flags,
 718         unsigned                *lockmode)
 719 {
 720         unsigned int            mode = *lockmode;
 721         bool                    is_write = flags & (IOMAP_WRITE | IOMAP_ZERO);
 722
 723         /*
 724          * COW writes may allocate delalloc space or convert unwritten COW
 725          * extents, so we need to make sure to take the lock exclusively here.
 726          */
 727         if (xfs_is_cow_inode(ip) && is_write)
 728                 mode = XFS_ILOCK_EXCL;
 729
 730         /*
 731          * Extents not yet cached requires exclusive access, don't block.  This
 732          * is an opencoded xfs_ilock_data_map_shared() call but with
 733          * non-blocking behaviour.
 734          */
 735         if (xfs_need_iread_extents(&ip->i_df)) {
 736                 if (flags & IOMAP_NOWAIT)
 737                         return -EAGAIN;
 738                 mode = XFS_ILOCK_EXCL;
 739         }
 740
 741 relock:
 742         if (flags & IOMAP_NOWAIT) {
 743                 if (!xfs_ilock_nowait(ip, mode))
 744                         return -EAGAIN;
 745         } else {
 746                 xfs_ilock(ip, mode);
 747         }
 748
 749         /*
 750          * The reflink iflag could have changed since the earlier unlocked
 751          * check, so if we got ILOCK_SHARED for a write and but we're now a
 752          * reflink inode we have to switch to ILOCK_EXCL and relock.
 753          */
 754         if (mode == XFS_ILOCK_SHARED && is_write && xfs_is_cow_inode(ip)) {
 755                 xfs_iunlock(ip, mode);
 756                 mode = XFS_ILOCK_EXCL;
 757                 goto relock;
 758         }
 759
 760         *lockmode = mode;
 761         return 0;
 762 }
 763
 764 /*
 765  * Check that the imap we are going to return to the caller spans the entire
 766  * range that the caller requested for the IO.
 767  */
 768 static bool
 769 imap_spans_range(
 770         struct xfs_bmbt_irec    *imap,
 771         xfs_fileoff_t           offset_fsb,
 772         xfs_fileoff_t           end_fsb)
 773 {
 774         if (imap->br_startoff > offset_fsb)
 775                 return false;
 776         if (imap->br_startoff + imap->br_blockcount < end_fsb)
 777                 return false;
 778         return true;
 779 }
 780
 781 static int
 782 xfs_direct_write_iomap_begin(
 783         struct inode            *inode,
 784         loff_t                  offset,
 785         loff_t                  length,
 786         unsigned                flags,
 787         struct iomap            *iomap,
 788         struct iomap            *srcmap)
 789 {
 790         struct xfs_inode        *ip = XFS_I(inode);
 791         struct xfs_mount        *mp = ip->i_mount;
 792         struct xfs_bmbt_irec    imap, cmap;
 793         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
 794         xfs_fileoff_t           end_fsb = xfs_iomap_end_fsb(mp, offset, length);
 795         int                     nimaps = 1, error = 0;
 796         bool                    shared = false;
 797         u16                     iomap_flags = 0;
 798         unsigned int            lockmode = XFS_ILOCK_SHARED;
 799         u64                     seq;
 800
 801         ASSERT(flags & (IOMAP_WRITE | IOMAP_ZERO));
 802
 803         if (xfs_is_shutdown(mp))
 804                 return -EIO;
 805
 806         /*
 807          * Writes that span EOF might trigger an IO size update on completion,
 808          * so consider them to be dirty for the purposes of O_DSYNC even if
 809          * there is no other metadata changes pending or have been made here.
 810          */
 811         if (offset + length > i_size_read(inode))
 812                 iomap_flags |= IOMAP_F_DIRTY;
 813
 814         error = xfs_ilock_for_iomap(ip, flags, &lockmode);
 815         if (error)
 816                 return error;
 817
 818         error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
 819                                &nimaps, 0);
 820         if (error)
 821                 goto out_unlock;
 822
 823         if (imap_needs_cow(ip, flags, &imap, nimaps)) {
 824                 error = -EAGAIN;
 825                 if (flags & IOMAP_NOWAIT)
 826                         goto out_unlock;
 827
 828                 /* may drop and re-acquire the ilock */
 829                 error = xfs_reflink_allocate_cow(ip, &imap, &cmap, &shared,
 830                                 &lockmode,
 831                                 (flags & IOMAP_DIRECT) || IS_DAX(inode));
 832                 if (error)
 833                         goto out_unlock;
 834                 if (shared)
 835                         goto out_found_cow;
 836                 end_fsb = imap.br_startoff + imap.br_blockcount;
 837                 length = XFS_FSB_TO_B(mp, end_fsb) - offset;
 838         }
 839
 840         if (imap_needs_alloc(inode, flags, &imap, nimaps))
 841                 goto allocate_blocks;
 842
 843         /*
 844          * NOWAIT and OVERWRITE I/O needs to span the entire requested I/O with
 845          * a single map so that we avoid partial IO failures due to the rest of
 846          * the I/O range not covered by this map triggering an EAGAIN condition
 847          * when it is subsequently mapped and aborting the I/O.
 848          */
 849         if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY)) {
 850                 error = -EAGAIN;
 851                 if (!imap_spans_range(&imap, offset_fsb, end_fsb))
 852                         goto out_unlock;
 853         }
 854
 855         /*
 856          * For overwrite only I/O, we cannot convert unwritten extents without
 857          * requiring sub-block zeroing.  This can only be done under an
 858          * exclusive IOLOCK, hence return -EAGAIN if this is not a written
 859          * extent to tell the caller to try again.
 860          */
 861         if (flags & IOMAP_OVERWRITE_ONLY) {
 862                 error = -EAGAIN;
 863                 if (imap.br_state != XFS_EXT_NORM &&
 864                     ((offset | length) & mp->m_blockmask))
 865                         goto out_unlock;
 866         }
 867
 868         seq = xfs_iomap_inode_sequence(ip, iomap_flags);
 869         xfs_iunlock(ip, lockmode);
 870         trace_xfs_iomap_found(ip, offset, length, XFS_DATA_FORK, &imap);
 871         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, iomap_flags, seq);
 872
 873 allocate_blocks:
 874         error = -EAGAIN;
 875         if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY))
 876                 goto out_unlock;
 877
 878         /*
 879          * We cap the maximum length we map to a sane size  to keep the chunks
 880          * of work done where somewhat symmetric with the work writeback does.
 881          * This is a completely arbitrary number pulled out of thin air as a
 882          * best guess for initial testing.
 883          *
 884          * Note that the values needs to be less than 32-bits wide until the
 885          * lower level functions are updated.
 886          */
 887         length = min_t(loff_t, length, 1024 * PAGE_SIZE);
 888         end_fsb = xfs_iomap_end_fsb(mp, offset, length);
 889
 890         if (offset + length > XFS_ISIZE(ip))
 891                 end_fsb = xfs_iomap_eof_align_last_fsb(ip, end_fsb);
 892         else if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
 893                 end_fsb = min(end_fsb, imap.br_startoff + imap.br_blockcount);
 894         xfs_iunlock(ip, lockmode);
 895
 896         error = xfs_iomap_write_direct(ip, offset_fsb, end_fsb - offset_fsb,
 897                         flags, &imap, &seq);
 898         if (error)
 899                 return error;
 900
 901         trace_xfs_iomap_alloc(ip, offset, length, XFS_DATA_FORK, &imap);
 902         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags,
 903                                  iomap_flags | IOMAP_F_NEW, seq);
 904
 905 out_found_cow:
 906         length = XFS_FSB_TO_B(mp, cmap.br_startoff + cmap.br_blockcount);
 907         trace_xfs_iomap_found(ip, offset, length - offset, XFS_COW_FORK, &cmap);
 908         if (imap.br_startblock != HOLESTARTBLOCK) {
 909                 seq = xfs_iomap_inode_sequence(ip, 0);
 910                 error = xfs_bmbt_to_iomap(ip, srcmap, &imap, flags, 0, seq);
 911                 if (error)
 912                         goto out_unlock;
 913         }
 914         seq = xfs_iomap_inode_sequence(ip, IOMAP_F_SHARED);
 915         xfs_iunlock(ip, lockmode);
 916         return xfs_bmbt_to_iomap(ip, iomap, &cmap, flags, IOMAP_F_SHARED, seq);
 917
 918 out_unlock:
 919         if (lockmode)
 920                 xfs_iunlock(ip, lockmode);
 921         return error;
 922 }
 923
 924 const struct iomap_ops xfs_direct_write_iomap_ops = {
 925         .iomap_begin            = xfs_direct_write_iomap_begin,
 926 };
 927
 928 static int
 929 xfs_dax_write_iomap_end(
 930         struct inode            *inode,
 931         loff_t                  pos,
 932         loff_t                  length,
 933         ssize_t                 written,
 934         unsigned                flags,
 935         struct iomap            *iomap)
 936 {
 937         struct xfs_inode        *ip = XFS_I(inode);
 938
 939         if (!xfs_is_cow_inode(ip))
 940                 return 0;
 941
 942         if (!written) {
 943                 xfs_reflink_cancel_cow_range(ip, pos, length, true);
 944                 return 0;
 945         }
 946
 947         return xfs_reflink_end_cow(ip, pos, written);
 948 }
 949
 950 const struct iomap_ops xfs_dax_write_iomap_ops = {
 951         .iomap_begin    = xfs_direct_write_iomap_begin,
 952         .iomap_end      = xfs_dax_write_iomap_end,
 953 };
 954
 955 static int
 956 xfs_buffered_write_iomap_begin(
 957         struct inode            *inode,
 958         loff_t                  offset,
 959         loff_t                  count,
 960         unsigned                flags,
 961         struct iomap            *iomap,
 962         struct iomap            *srcmap)
 963 {
 964         struct xfs_inode        *ip = XFS_I(inode);
 965         struct xfs_mount        *mp = ip->i_mount;
 966         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
 967         xfs_fileoff_t           end_fsb = xfs_iomap_end_fsb(mp, offset, count);
 968         struct xfs_bmbt_irec    imap, cmap;
 969         struct xfs_iext_cursor  icur, ccur;
 970         xfs_fsblock_t           prealloc_blocks = 0;
 971         bool                    eof = false, cow_eof = false, shared = false;
 972         int                     allocfork = XFS_DATA_FORK;
 973         int                     error = 0;
 974         unsigned int            lockmode = XFS_ILOCK_EXCL;
 975         u64                     seq;
 976
 977         if (xfs_is_shutdown(mp))
 978                 return -EIO;
 979
 980         /* we can't use delayed allocations when using extent size hints */
 981         if (xfs_get_extsz_hint(ip))
 982                 return xfs_direct_write_iomap_begin(inode, offset, count,
 983                                 flags, iomap, srcmap);
 984
 985         ASSERT(!XFS_IS_REALTIME_INODE(ip));
 986
 987         error = xfs_qm_dqattach(ip);
 988         if (error)
 989                 return error;
 990
 991         error = xfs_ilock_for_iomap(ip, flags, &lockmode);
 992         if (error)
 993                 return error;
 994
 995         if (XFS_IS_CORRUPT(mp, !xfs_ifork_has_extents(&ip->i_df)) ||
 996             XFS_TEST_ERROR(false, mp, XFS_ERRTAG_BMAPIFORMAT)) {
 997                 xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
 998                 error = -EFSCORRUPTED;
 999                 goto out_unlock;
1000         }
1001
1002         XFS_STATS_INC(mp, xs_blk_mapw);
1003
1004         error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK);
1005         if (error)
1006                 goto out_unlock;
1007
1008         /*
1009          * Search the data fork first to look up our source mapping.  We
1010          * always need the data fork map, as we have to return it to the
1011          * iomap code so that the higher level write code can read data in to
1012          * perform read-modify-write cycles for unaligned writes.
1013          */
1014         eof = !xfs_iext_lookup_extent(ip, &ip->i_df, offset_fsb, &icur, &imap);
1015         if (eof)
1016                 imap.br_startoff = end_fsb; /* fake hole until the end */
1017
1018         /* We never need to allocate blocks for zeroing or unsharing a hole. */
1019         if ((flags & (IOMAP_UNSHARE | IOMAP_ZERO)) &&
1020             imap.br_startoff > offset_fsb) {
1021                 xfs_hole_to_iomap(ip, iomap, offset_fsb, imap.br_startoff);
1022                 goto out_unlock;
1023         }
1024
1025         /*
1026          * Search the COW fork extent list even if we did not find a data fork
1027          * extent.  This serves two purposes: first this implements the
1028          * speculative preallocation using cowextsize, so that we also unshare
1029          * block adjacent to shared blocks instead of just the shared blocks
1030          * themselves.  Second the lookup in the extent list is generally faster
1031          * than going out to the shared extent tree.
1032          */
1033         if (xfs_is_cow_inode(ip)) {
1034                 if (!ip->i_cowfp) {
1035                         ASSERT(!xfs_is_reflink_inode(ip));
1036                         xfs_ifork_init_cow(ip);
1037                 }
1038                 cow_eof = !xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb,
1039                                 &ccur, &cmap);
1040                 if (!cow_eof && cmap.br_startoff <= offset_fsb) {
1041                         trace_xfs_reflink_cow_found(ip, &cmap);
1042                         goto found_cow;
1043                 }
1044         }
1045
1046         if (imap.br_startoff <= offset_fsb) {
1047                 /*
1048                  * For reflink files we may need a delalloc reservation when
1049                  * overwriting shared extents.   This includes zeroing of
1050                  * existing extents that contain data.
1051                  */
1052                 if (!xfs_is_cow_inode(ip) ||
1053                     ((flags & IOMAP_ZERO) && imap.br_state != XFS_EXT_NORM)) {
1054                         trace_xfs_iomap_found(ip, offset, count, XFS_DATA_FORK,
1055                                         &imap);
1056                         goto found_imap;
1057                 }
1058
1059                 xfs_trim_extent(&imap, offset_fsb, end_fsb - offset_fsb);
1060
1061                 /* Trim the mapping to the nearest shared extent boundary. */
1062                 error = xfs_bmap_trim_cow(ip, &imap, &shared);
1063                 if (error)
1064                         goto out_unlock;
1065
1066                 /* Not shared?  Just report the (potentially capped) extent. */
1067                 if (!shared) {
1068                         trace_xfs_iomap_found(ip, offset, count, XFS_DATA_FORK,
1069                                         &imap);
1070                         goto found_imap;
1071                 }
1072
1073                 /*
1074                  * Fork all the shared blocks from our write offset until the
1075                  * end of the extent.
1076                  */
1077                 allocfork = XFS_COW_FORK;
1078                 end_fsb = imap.br_startoff + imap.br_blockcount;
1079         } else {
1080                 /*
1081                  * We cap the maximum length we map here to MAX_WRITEBACK_PAGES
1082                  * pages to keep the chunks of work done where somewhat
1083                  * symmetric with the work writeback does.  This is a completely
1084                  * arbitrary number pulled out of thin air.
1085                  *
1086                  * Note that the values needs to be less than 32-bits wide until
1087                  * the lower level functions are updated.
1088                  */
1089                 count = min_t(loff_t, count, 1024 * PAGE_SIZE);
1090                 end_fsb = xfs_iomap_end_fsb(mp, offset, count);
1091
1092                 if (xfs_is_always_cow_inode(ip))
1093                         allocfork = XFS_COW_FORK;
1094         }
1095
1096         if (eof && offset + count > XFS_ISIZE(ip)) {
1097                 /*
1098                  * Determine the initial size of the preallocation.
1099                  * We clean up any extra preallocation when the file is closed.
1100                  */
1101                 if (xfs_has_allocsize(mp))
1102                         prealloc_blocks = mp->m_allocsize_blocks;
1103                 else if (allocfork == XFS_DATA_FORK)
1104                         prealloc_blocks = xfs_iomap_prealloc_size(ip, allocfork,
1105                                                 offset, count, &icur);
1106                 else
1107                         prealloc_blocks = xfs_iomap_prealloc_size(ip, allocfork,
1108                                                 offset, count, &ccur);
1109                 if (prealloc_blocks) {
1110                         xfs_extlen_t    align;
1111                         xfs_off_t       end_offset;
1112                         xfs_fileoff_t   p_end_fsb;
1113
1114                         end_offset = XFS_ALLOC_ALIGN(mp, offset + count - 1);
1115                         p_end_fsb = XFS_B_TO_FSBT(mp, end_offset) +
1116                                         prealloc_blocks;
1117
1118                         align = xfs_eof_alignment(ip);
1119                         if (align)
1120                                 p_end_fsb = roundup_64(p_end_fsb, align);
1121
1122                         p_end_fsb = min(p_end_fsb,
1123                                 XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes));
1124                         ASSERT(p_end_fsb > offset_fsb);
1125                         prealloc_blocks = p_end_fsb - end_fsb;
1126                 }
1127         }
1128
1129 retry:
1130         error = xfs_bmapi_reserve_delalloc(ip, allocfork, offset_fsb,
1131                         end_fsb - offset_fsb, prealloc_blocks,
1132                         allocfork == XFS_DATA_FORK ? &imap : &cmap,
1133                         allocfork == XFS_DATA_FORK ? &icur : &ccur,
1134                         allocfork == XFS_DATA_FORK ? eof : cow_eof);
1135         switch (error) {
1136         case 0:
1137                 break;
1138         case -ENOSPC:
1139         case -EDQUOT:
1140                 /* retry without any preallocation */
1141                 trace_xfs_delalloc_enospc(ip, offset, count);
1142                 if (prealloc_blocks) {
1143                         prealloc_blocks = 0;
1144                         goto retry;
1145                 }
1146                 fallthrough;
1147         default:
1148                 goto out_unlock;
1149         }
1150
1151         if (allocfork == XFS_COW_FORK) {
1152                 trace_xfs_iomap_alloc(ip, offset, count, allocfork, &cmap);
1153                 goto found_cow;
1154         }
1155
1156         /*
1157          * Flag newly allocated delalloc blocks with IOMAP_F_NEW so we punch
1158          * them out if the write happens to fail.
1159          */
1160         seq = xfs_iomap_inode_sequence(ip, IOMAP_F_NEW);
1161         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1162         trace_xfs_iomap_alloc(ip, offset, count, allocfork, &imap);
1163         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, IOMAP_F_NEW, seq);
1164
1165 found_imap:
1166         seq = xfs_iomap_inode_sequence(ip, 0);
1167         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1168         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, 0, seq);
1169
1170 found_cow:
1171         seq = xfs_iomap_inode_sequence(ip, 0);
1172         if (imap.br_startoff <= offset_fsb) {
1173                 error = xfs_bmbt_to_iomap(ip, srcmap, &imap, flags, 0, seq);
1174                 if (error)
1175                         goto out_unlock;
1176                 seq = xfs_iomap_inode_sequence(ip, IOMAP_F_SHARED);
1177                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1178                 return xfs_bmbt_to_iomap(ip, iomap, &cmap, flags,
1179                                          IOMAP_F_SHARED, seq);
1180         }
1181
1182         xfs_trim_extent(&cmap, offset_fsb, imap.br_startoff - offset_fsb);
1183         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1184         return xfs_bmbt_to_iomap(ip, iomap, &cmap, flags, 0, seq);
1185
1186 out_unlock:
1187         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1188         return error;
1189 }
1190
1191 static int
1192 xfs_buffered_write_delalloc_punch(
1193         struct inode            *inode,
1194         loff_t                  offset,
1195         loff_t                  length)
1196 {
1197         return xfs_bmap_punch_delalloc_range(XFS_I(inode), offset,
1198                         offset + length);
1199 }
1200
1201 static int
1202 xfs_buffered_write_iomap_end(
1203         struct inode            *inode,
1204         loff_t                  offset,
1205         loff_t                  length,
1206         ssize_t                 written,
1207         unsigned                flags,
1208         struct iomap            *iomap)
1209 {
1210
1211         struct xfs_mount        *mp = XFS_M(inode->i_sb);
1212         int                     error;
1213
1214         error = iomap_file_buffered_write_punch_delalloc(inode, iomap, offset,
1215                         length, written, &xfs_buffered_write_delalloc_punch);
1216         if (error && !xfs_is_shutdown(mp)) {
1217                 xfs_alert(mp, "%s: unable to clean up ino 0x%llx",
1218                         __func__, XFS_I(inode)->i_ino);
1219                 return error;
1220         }
1221         return 0;
1222 }
1223
1224 const struct iomap_ops xfs_buffered_write_iomap_ops = {
1225         .iomap_begin            = xfs_buffered_write_iomap_begin,
1226         .iomap_end              = xfs_buffered_write_iomap_end,
1227 };
1228
1229 /*
1230  * iomap_page_mkwrite() will never fail in a way that requires delalloc extents
1231  * that it allocated to be revoked. Hence we do not need an .iomap_end method
1232  * for this operation.
1233  */
1234 const struct iomap_ops xfs_page_mkwrite_iomap_ops = {
1235         .iomap_begin            = xfs_buffered_write_iomap_begin,
1236 };
1237
1238 static int
1239 xfs_read_iomap_begin(
1240         struct inode            *inode,
1241         loff_t                  offset,
1242         loff_t                  length,
1243         unsigned                flags,
1244         struct iomap            *iomap,
1245         struct iomap            *srcmap)
1246 {
1247         struct xfs_inode        *ip = XFS_I(inode);
1248         struct xfs_mount        *mp = ip->i_mount;
1249         struct xfs_bmbt_irec    imap;
1250         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
1251         xfs_fileoff_t           end_fsb = xfs_iomap_end_fsb(mp, offset, length);
1252         int                     nimaps = 1, error = 0;
1253         bool                    shared = false;
1254         unsigned int            lockmode = XFS_ILOCK_SHARED;
1255         u64                     seq;
1256
1257         ASSERT(!(flags & (IOMAP_WRITE | IOMAP_ZERO)));
1258
1259         if (xfs_is_shutdown(mp))
1260                 return -EIO;
1261
1262         error = xfs_ilock_for_iomap(ip, flags, &lockmode);
1263         if (error)
1264                 return error;
1265         error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
1266                                &nimaps, 0);
1267         if (!error && ((flags & IOMAP_REPORT) || IS_DAX(inode)))
1268                 error = xfs_reflink_trim_around_shared(ip, &imap, &shared);
1269         seq = xfs_iomap_inode_sequence(ip, shared ? IOMAP_F_SHARED : 0);
1270         xfs_iunlock(ip, lockmode);
1271
1272         if (error)
1273                 return error;
1274         trace_xfs_iomap_found(ip, offset, length, XFS_DATA_FORK, &imap);
1275         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags,
1276                                  shared ? IOMAP_F_SHARED : 0, seq);
1277 }
1278
1279 const struct iomap_ops xfs_read_iomap_ops = {
1280         .iomap_begin            = xfs_read_iomap_begin,
1281 };
1282
1283 static int
1284 xfs_seek_iomap_begin(
1285         struct inode            *inode,
1286         loff_t                  offset,
1287         loff_t                  length,
1288         unsigned                flags,
1289         struct iomap            *iomap,
1290         struct iomap            *srcmap)
1291 {
1292         struct xfs_inode        *ip = XFS_I(inode);
1293         struct xfs_mount        *mp = ip->i_mount;
1294         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
1295         xfs_fileoff_t           end_fsb = XFS_B_TO_FSB(mp, offset + length);
1296         xfs_fileoff_t           cow_fsb = NULLFILEOFF, data_fsb = NULLFILEOFF;
1297         struct xfs_iext_cursor  icur;
1298         struct xfs_bmbt_irec    imap, cmap;
1299         int                     error = 0;
1300         unsigned                lockmode;
1301         u64                     seq;
1302
1303         if (xfs_is_shutdown(mp))
1304                 return -EIO;
1305
1306         lockmode = xfs_ilock_data_map_shared(ip);
1307         error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK);
1308         if (error)
1309                 goto out_unlock;
1310
1311         if (xfs_iext_lookup_extent(ip, &ip->i_df, offset_fsb, &icur, &imap)) {
1312                 /*
1313                  * If we found a data extent we are done.
1314                  */
1315                 if (imap.br_startoff <= offset_fsb)
1316                         goto done;
1317                 data_fsb = imap.br_startoff;
1318         } else {
1319                 /*
1320                  * Fake a hole until the end of the file.
1321                  */
1322                 data_fsb = xfs_iomap_end_fsb(mp, offset, length);
1323         }
1324
1325         /*
1326          * If a COW fork extent covers the hole, report it - capped to the next
1327          * data fork extent:
1328          */
1329         if (xfs_inode_has_cow_data(ip) &&
1330             xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &cmap))
1331                 cow_fsb = cmap.br_startoff;
1332         if (cow_fsb != NULLFILEOFF && cow_fsb <= offset_fsb) {
1333                 if (data_fsb < cow_fsb + cmap.br_blockcount)
1334                         end_fsb = min(end_fsb, data_fsb);
1335                 xfs_trim_extent(&cmap, offset_fsb, end_fsb - offset_fsb);
1336                 seq = xfs_iomap_inode_sequence(ip, IOMAP_F_SHARED);
1337                 error = xfs_bmbt_to_iomap(ip, iomap, &cmap, flags,
1338                                 IOMAP_F_SHARED, seq);
1339                 /*
1340                  * This is a COW extent, so we must probe the page cache
1341                  * because there could be dirty page cache being backed
1342                  * by this extent.
1343                  */
1344                 iomap->type = IOMAP_UNWRITTEN;
1345                 goto out_unlock;
1346         }
1347
1348         /*
1349          * Else report a hole, capped to the next found data or COW extent.
1350          */
1351         if (cow_fsb != NULLFILEOFF && cow_fsb < data_fsb)
1352                 imap.br_blockcount = cow_fsb - offset_fsb;
1353         else
1354                 imap.br_blockcount = data_fsb - offset_fsb;
1355         imap.br_startoff = offset_fsb;
1356         imap.br_startblock = HOLESTARTBLOCK;
1357         imap.br_state = XFS_EXT_NORM;
1358 done:
1359         seq = xfs_iomap_inode_sequence(ip, 0);
1360         xfs_trim_extent(&imap, offset_fsb, end_fsb - offset_fsb);
1361         error = xfs_bmbt_to_iomap(ip, iomap, &imap, flags, 0, seq);
1362 out_unlock:
1363         xfs_iunlock(ip, lockmode);
1364         return error;
1365 }
1366
1367 const struct iomap_ops xfs_seek_iomap_ops = {
1368         .iomap_begin            = xfs_seek_iomap_begin,
1369 };
1370
1371 static int
1372 xfs_xattr_iomap_begin(
1373         struct inode            *inode,
1374         loff_t                  offset,
1375         loff_t                  length,
1376         unsigned                flags,
1377         struct iomap            *iomap,
1378         struct iomap            *srcmap)
1379 {
1380         struct xfs_inode        *ip = XFS_I(inode);
1381         struct xfs_mount        *mp = ip->i_mount;
1382         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
1383         xfs_fileoff_t           end_fsb = XFS_B_TO_FSB(mp, offset + length);
1384         struct xfs_bmbt_irec    imap;
1385         int                     nimaps = 1, error = 0;
1386         unsigned                lockmode;
1387         int                     seq;
1388
1389         if (xfs_is_shutdown(mp))
1390                 return -EIO;
1391
1392         lockmode = xfs_ilock_attr_map_shared(ip);
1393
1394         /* if there are no attribute fork or extents, return ENOENT */
1395         if (!xfs_inode_has_attr_fork(ip) || !ip->i_af.if_nextents) {
1396                 error = -ENOENT;
1397                 goto out_unlock;
1398         }
1399
1400         ASSERT(ip->i_af.if_format != XFS_DINODE_FMT_LOCAL);
1401         error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
1402                                &nimaps, XFS_BMAPI_ATTRFORK);
1403 out_unlock:
1404
1405         seq = xfs_iomap_inode_sequence(ip, IOMAP_F_XATTR);
1406         xfs_iunlock(ip, lockmode);
1407
1408         if (error)
1409                 return error;
1410         ASSERT(nimaps);
1411         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, IOMAP_F_XATTR, seq);
1412 }
1413
1414 const struct iomap_ops xfs_xattr_iomap_ops = {
1415         .iomap_begin            = xfs_xattr_iomap_begin,
1416 };
1417
1418 int
1419 xfs_zero_range(
1420         struct xfs_inode        *ip,
1421         loff_t                  pos,
1422         loff_t                  len,
1423         bool                    *did_zero)
1424 {
1425         struct inode            *inode = VFS_I(ip);
1426
1427         if (IS_DAX(inode))
1428                 return dax_zero_range(inode, pos, len, did_zero,
1429                                       &xfs_dax_write_iomap_ops);
1430         return iomap_zero_range(inode, pos, len, did_zero,
1431                                 &xfs_buffered_write_iomap_ops);
1432 }
1433
1434 int
1435 xfs_truncate_page(
1436         struct xfs_inode        *ip,
1437         loff_t                  pos,
1438         bool                    *did_zero)
1439 {
1440         struct inode            *inode = VFS_I(ip);
1441
1442         if (IS_DAX(inode))
1443                 return dax_truncate_page(inode, pos, did_zero,
1444                                         &xfs_dax_write_iomap_ops);
1445         return iomap_truncate_page(inode, pos, did_zero,
1446                                    &xfs_buffered_write_iomap_ops);
1447 }