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