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_rtgroup.h"
  28 #include "xfs_dquot_item.h"
  29 #include "xfs_dquot.h"
  30 #include "xfs_reflink.h"
  31 #include "xfs_health.h"
  32 #include "xfs_rtbitmap.h"
  33
  34 #define XFS_ALLOC_ALIGN(mp, off) \
  35         (((off) >> mp->m_allocsize_log) << mp->m_allocsize_log)
  36
  37 static int
  38 xfs_alert_fsblock_zero(
  39         xfs_inode_t     *ip,
  40         xfs_bmbt_irec_t *imap)
  41 {
  42         xfs_alert_tag(ip->i_mount, XFS_PTAG_FSBLOCK_ZERO,
  43                         "Access to block zero in inode %llu "
  44                         "start_block: %llx start_off: %llx "
  45                         "blkcnt: %llx extent-state: %x",
  46                 (unsigned long long)ip->i_ino,
  47                 (unsigned long long)imap->br_startblock,
  48                 (unsigned long long)imap->br_startoff,
  49                 (unsigned long long)imap->br_blockcount,
  50                 imap->br_state);
  51         xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
  52         return -EFSCORRUPTED;
  53 }
  54
  55 u64
  56 xfs_iomap_inode_sequence(
  57         struct xfs_inode        *ip,
  58         u16                     iomap_flags)
  59 {
  60         u64                     cookie = 0;
  61
  62         if (iomap_flags & IOMAP_F_XATTR)
  63                 return READ_ONCE(ip->i_af.if_seq);
  64         if ((iomap_flags & IOMAP_F_SHARED) && ip->i_cowfp)
  65                 cookie = (u64)READ_ONCE(ip->i_cowfp->if_seq) << 32;
  66         return cookie | READ_ONCE(ip->i_df.if_seq);
  67 }
  68
  69 /*
  70  * Check that the iomap passed to us is still valid for the given offset and
  71  * length.
  72  */
  73 static bool
  74 xfs_iomap_valid(
  75         struct inode            *inode,
  76         const struct iomap      *iomap)
  77 {
  78         struct xfs_inode        *ip = XFS_I(inode);
  79
  80         if (iomap->validity_cookie !=
  81                         xfs_iomap_inode_sequence(ip, iomap->flags)) {
  82                 trace_xfs_iomap_invalid(ip, iomap);
  83                 return false;
  84         }
  85
  86         XFS_ERRORTAG_DELAY(ip->i_mount, XFS_ERRTAG_WRITE_DELAY_MS);
  87         return true;
  88 }
  89
  90 static const struct iomap_folio_ops xfs_iomap_folio_ops = {
  91         .iomap_valid            = xfs_iomap_valid,
  92 };
  93
  94 int
  95 xfs_bmbt_to_iomap(
  96         struct xfs_inode        *ip,
  97         struct iomap            *iomap,
  98         struct xfs_bmbt_irec    *imap,
  99         unsigned int            mapping_flags,
 100         u16                     iomap_flags,
 101         u64                     sequence_cookie)
 102 {
 103         struct xfs_mount        *mp = ip->i_mount;
 104         struct xfs_buftarg      *target = xfs_inode_buftarg(ip);
 105
 106         if (unlikely(!xfs_valid_startblock(ip, imap->br_startblock))) {
 107                 xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
 108                 return xfs_alert_fsblock_zero(ip, imap);
 109         }
 110
 111         if (imap->br_startblock == HOLESTARTBLOCK) {
 112                 iomap->addr = IOMAP_NULL_ADDR;
 113                 iomap->type = IOMAP_HOLE;
 114         } else if (imap->br_startblock == DELAYSTARTBLOCK ||
 115                    isnullstartblock(imap->br_startblock)) {
 116                 iomap->addr = IOMAP_NULL_ADDR;
 117                 iomap->type = IOMAP_DELALLOC;
 118         } else {
 119                 xfs_daddr_t     daddr = xfs_fsb_to_db(ip, imap->br_startblock);
 120
 121                 iomap->addr = BBTOB(daddr);
 122                 if (mapping_flags & IOMAP_DAX)
 123                         iomap->addr += target->bt_dax_part_off;
 124
 125                 if (imap->br_state == XFS_EXT_UNWRITTEN)
 126                         iomap->type = IOMAP_UNWRITTEN;
 127                 else
 128                         iomap->type = IOMAP_MAPPED;
 129
 130                 /*
 131                  * Mark iomaps starting at the first sector of a RTG as merge
 132                  * boundary so that each I/O completions is contained to a
 133                  * single RTG.
 134                  */
 135                 if (XFS_IS_REALTIME_INODE(ip) && xfs_has_rtgroups(mp) &&
 136                     xfs_rtbno_is_group_start(mp, imap->br_startblock))
 137                         iomap->flags |= IOMAP_F_BOUNDARY;
 138         }
 139         iomap->offset = XFS_FSB_TO_B(mp, imap->br_startoff);
 140         iomap->length = XFS_FSB_TO_B(mp, imap->br_blockcount);
 141         if (mapping_flags & IOMAP_DAX)
 142                 iomap->dax_dev = target->bt_daxdev;
 143         else
 144                 iomap->bdev = target->bt_bdev;
 145         iomap->flags = iomap_flags;
 146
 147         if (xfs_ipincount(ip) &&
 148             (ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
 149                 iomap->flags |= IOMAP_F_DIRTY;
 150
 151         iomap->validity_cookie = sequence_cookie;
 152         iomap->folio_ops = &xfs_iomap_folio_ops;
 153         return 0;
 154 }
 155
 156 static void
 157 xfs_hole_to_iomap(
 158         struct xfs_inode        *ip,
 159         struct iomap            *iomap,
 160         xfs_fileoff_t           offset_fsb,
 161         xfs_fileoff_t           end_fsb)
 162 {
 163         struct xfs_buftarg      *target = xfs_inode_buftarg(ip);
 164
 165         iomap->addr = IOMAP_NULL_ADDR;
 166         iomap->type = IOMAP_HOLE;
 167         iomap->offset = XFS_FSB_TO_B(ip->i_mount, offset_fsb);
 168         iomap->length = XFS_FSB_TO_B(ip->i_mount, end_fsb - offset_fsb);
 169         iomap->bdev = target->bt_bdev;
 170         iomap->dax_dev = target->bt_daxdev;
 171 }
 172
 173 static inline xfs_fileoff_t
 174 xfs_iomap_end_fsb(
 175         struct xfs_mount        *mp,
 176         loff_t                  offset,
 177         loff_t                  count)
 178 {
 179         ASSERT(offset <= mp->m_super->s_maxbytes);
 180         return min(XFS_B_TO_FSB(mp, offset + count),
 181                    XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes));
 182 }
 183
 184 static xfs_extlen_t
 185 xfs_eof_alignment(
 186         struct xfs_inode        *ip)
 187 {
 188         struct xfs_mount        *mp = ip->i_mount;
 189         xfs_extlen_t            align = 0;
 190
 191         if (!XFS_IS_REALTIME_INODE(ip)) {
 192                 /*
 193                  * Round up the allocation request to a stripe unit
 194                  * (m_dalign) boundary if the file size is >= stripe unit
 195                  * size, and we are allocating past the allocation eof.
 196                  *
 197                  * If mounted with the "-o swalloc" option the alignment is
 198                  * increased from the strip unit size to the stripe width.
 199                  */
 200                 if (mp->m_swidth && xfs_has_swalloc(mp))
 201                         align = mp->m_swidth;
 202                 else if (mp->m_dalign)
 203                         align = mp->m_dalign;
 204
 205                 if (align && XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, align))
 206                         align = 0;
 207         }
 208
 209         return align;
 210 }
 211
 212 /*
 213  * Check if last_fsb is outside the last extent, and if so grow it to the next
 214  * stripe unit boundary.
 215  */
 216 xfs_fileoff_t
 217 xfs_iomap_eof_align_last_fsb(
 218         struct xfs_inode        *ip,
 219         xfs_fileoff_t           end_fsb)
 220 {
 221         struct xfs_ifork        *ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
 222         xfs_extlen_t            extsz = xfs_get_extsz_hint(ip);
 223         xfs_extlen_t            align = xfs_eof_alignment(ip);
 224         struct xfs_bmbt_irec    irec;
 225         struct xfs_iext_cursor  icur;
 226
 227         ASSERT(!xfs_need_iread_extents(ifp));
 228
 229         /*
 230          * Always round up the allocation request to the extent hint boundary.
 231          */
 232         if (extsz) {
 233                 if (align)
 234                         align = roundup_64(align, extsz);
 235                 else
 236                         align = extsz;
 237         }
 238
 239         if (align) {
 240                 xfs_fileoff_t   aligned_end_fsb = roundup_64(end_fsb, align);
 241
 242                 xfs_iext_last(ifp, &icur);
 243                 if (!xfs_iext_get_extent(ifp, &icur, &irec) ||
 244                     aligned_end_fsb >= irec.br_startoff + irec.br_blockcount)
 245                         return aligned_end_fsb;
 246         }
 247
 248         return end_fsb;
 249 }
 250
 251 int
 252 xfs_iomap_write_direct(
 253         struct xfs_inode        *ip,
 254         xfs_fileoff_t           offset_fsb,
 255         xfs_fileoff_t           count_fsb,
 256         unsigned int            flags,
 257         struct xfs_bmbt_irec    *imap,
 258         u64                     *seq)
 259 {
 260         struct xfs_mount        *mp = ip->i_mount;
 261         struct xfs_trans        *tp;
 262         xfs_filblks_t           resaligned;
 263         int                     nimaps;
 264         unsigned int            dblocks, rblocks;
 265         bool                    force = false;
 266         int                     error;
 267         int                     bmapi_flags = XFS_BMAPI_PREALLOC;
 268         int                     nr_exts = XFS_IEXT_ADD_NOSPLIT_CNT;
 269
 270         ASSERT(count_fsb > 0);
 271
 272         resaligned = xfs_aligned_fsb_count(offset_fsb, count_fsb,
 273                                            xfs_get_extsz_hint(ip));
 274         if (unlikely(XFS_IS_REALTIME_INODE(ip))) {
 275                 dblocks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
 276                 rblocks = resaligned;
 277         } else {
 278                 dblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
 279                 rblocks = 0;
 280         }
 281
 282         error = xfs_qm_dqattach(ip);
 283         if (error)
 284                 return error;
 285
 286         /*
 287          * For DAX, we do not allocate unwritten extents, but instead we zero
 288          * the block before we commit the transaction.  Ideally we'd like to do
 289          * this outside the transaction context, but if we commit and then crash
 290          * we may not have zeroed the blocks and this will be exposed on
 291          * recovery of the allocation. Hence we must zero before commit.
 292          *
 293          * Further, if we are mapping unwritten extents here, we need to zero
 294          * and convert them to written so that we don't need an unwritten extent
 295          * callback for DAX. This also means that we need to be able to dip into
 296          * the reserve block pool for bmbt block allocation if there is no space
 297          * left but we need to do unwritten extent conversion.
 298          */
 299         if (flags & IOMAP_DAX) {
 300                 bmapi_flags = XFS_BMAPI_CONVERT | XFS_BMAPI_ZERO;
 301                 if (imap->br_state == XFS_EXT_UNWRITTEN) {
 302                         force = true;
 303                         nr_exts = XFS_IEXT_WRITE_UNWRITTEN_CNT;
 304                         dblocks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
 305                 }
 306         }
 307
 308         error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, dblocks,
 309                         rblocks, force, &tp);
 310         if (error)
 311                 return error;
 312
 313         error = xfs_iext_count_extend(tp, ip, XFS_DATA_FORK, nr_exts);
 314         if (error)
 315                 goto out_trans_cancel;
 316
 317         /*
 318          * From this point onwards we overwrite the imap pointer that the
 319          * caller gave to us.
 320          */
 321         nimaps = 1;
 322         error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb, bmapi_flags, 0,
 323                                 imap, &nimaps);
 324         if (error)
 325                 goto out_trans_cancel;
 326
 327         /*
 328          * Complete the transaction
 329          */
 330         error = xfs_trans_commit(tp);
 331         if (error)
 332                 goto out_unlock;
 333
 334         if (unlikely(!xfs_valid_startblock(ip, imap->br_startblock))) {
 335                 xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
 336                 error = xfs_alert_fsblock_zero(ip, imap);
 337         }
 338
 339 out_unlock:
 340         *seq = xfs_iomap_inode_sequence(ip, 0);
 341         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 342         return error;
 343
 344 out_trans_cancel:
 345         xfs_trans_cancel(tp);
 346         goto out_unlock;
 347 }
 348
 349 STATIC bool
 350 xfs_quota_need_throttle(
 351         struct xfs_inode        *ip,
 352         xfs_dqtype_t            type,
 353         xfs_fsblock_t           alloc_blocks)
 354 {
 355         struct xfs_dquot        *dq = xfs_inode_dquot(ip, type);
 356         struct xfs_dquot_res    *res;
 357         struct xfs_dquot_pre    *pre;
 358
 359         if (!dq || !xfs_this_quota_on(ip->i_mount, type))
 360                 return false;
 361
 362         if (XFS_IS_REALTIME_INODE(ip)) {
 363                 res = &dq->q_rtb;
 364                 pre = &dq->q_rtb_prealloc;
 365         } else {
 366                 res = &dq->q_blk;
 367                 pre = &dq->q_blk_prealloc;
 368         }
 369
 370         /* no hi watermark, no throttle */
 371         if (!pre->q_prealloc_hi_wmark)
 372                 return false;
 373
 374         /* under the lo watermark, no throttle */
 375         if (res->reserved + alloc_blocks < pre->q_prealloc_lo_wmark)
 376                 return false;
 377
 378         return true;
 379 }
 380
 381 STATIC void
 382 xfs_quota_calc_throttle(
 383         struct xfs_inode        *ip,
 384         xfs_dqtype_t            type,
 385         xfs_fsblock_t           *qblocks,
 386         int                     *qshift,
 387         int64_t                 *qfreesp)
 388 {
 389         struct xfs_dquot        *dq = xfs_inode_dquot(ip, type);
 390         struct xfs_dquot_res    *res;
 391         struct xfs_dquot_pre    *pre;
 392         int64_t                 freesp;
 393         int                     shift = 0;
 394
 395         if (!dq) {
 396                 res = NULL;
 397                 pre = NULL;
 398         } else if (XFS_IS_REALTIME_INODE(ip)) {
 399                 res = &dq->q_rtb;
 400                 pre = &dq->q_rtb_prealloc;
 401         } else {
 402                 res = &dq->q_blk;
 403                 pre = &dq->q_blk_prealloc;
 404         }
 405
 406         /* no dq, or over hi wmark, squash the prealloc completely */
 407         if (!res || res->reserved >= pre->q_prealloc_hi_wmark) {
 408                 *qblocks = 0;
 409                 *qfreesp = 0;
 410                 return;
 411         }
 412
 413         freesp = pre->q_prealloc_hi_wmark - res->reserved;
 414         if (freesp < pre->q_low_space[XFS_QLOWSP_5_PCNT]) {
 415                 shift = 2;
 416                 if (freesp < pre->q_low_space[XFS_QLOWSP_3_PCNT])
 417                         shift += 2;
 418                 if (freesp < pre->q_low_space[XFS_QLOWSP_1_PCNT])
 419                         shift += 2;
 420         }
 421
 422         if (freesp < *qfreesp)
 423                 *qfreesp = freesp;
 424
 425         /* only overwrite the throttle values if we are more aggressive */
 426         if ((freesp >> shift) < (*qblocks >> *qshift)) {
 427                 *qblocks = freesp;
 428                 *qshift = shift;
 429         }
 430 }
 431
 432 static int64_t
 433 xfs_iomap_freesp(
 434         struct percpu_counter   *counter,
 435         uint64_t                low_space[XFS_LOWSP_MAX],
 436         int                     *shift)
 437 {
 438         int64_t                 freesp;
 439
 440         freesp = percpu_counter_read_positive(counter);
 441         if (freesp < low_space[XFS_LOWSP_5_PCNT]) {
 442                 *shift = 2;
 443                 if (freesp < low_space[XFS_LOWSP_4_PCNT])
 444                         (*shift)++;
 445                 if (freesp < low_space[XFS_LOWSP_3_PCNT])
 446                         (*shift)++;
 447                 if (freesp < low_space[XFS_LOWSP_2_PCNT])
 448                         (*shift)++;
 449                 if (freesp < low_space[XFS_LOWSP_1_PCNT])
 450                         (*shift)++;
 451         }
 452         return freesp;
 453 }
 454
 455 /*
 456  * If we don't have a user specified preallocation size, dynamically increase
 457  * the preallocation size as the size of the file grows.  Cap the maximum size
 458  * at a single extent or less if the filesystem is near full. The closer the
 459  * filesystem is to being full, the smaller the maximum preallocation.
 460  */
 461 STATIC xfs_fsblock_t
 462 xfs_iomap_prealloc_size(
 463         struct xfs_inode        *ip,
 464         int                     whichfork,
 465         loff_t                  offset,
 466         loff_t                  count,
 467         struct xfs_iext_cursor  *icur)
 468 {
 469         struct xfs_iext_cursor  ncur = *icur;
 470         struct xfs_bmbt_irec    prev, got;
 471         struct xfs_mount        *mp = ip->i_mount;
 472         struct xfs_ifork        *ifp = xfs_ifork_ptr(ip, whichfork);
 473         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
 474         int64_t                 freesp;
 475         xfs_fsblock_t           qblocks;
 476         xfs_fsblock_t           alloc_blocks = 0;
 477         xfs_extlen_t            plen;
 478         int                     shift = 0;
 479         int                     qshift = 0;
 480
 481         /*
 482          * As an exception we don't do any preallocation at all if the file is
 483          * smaller than the minimum preallocation and we are using the default
 484          * dynamic preallocation scheme, as it is likely this is the only write
 485          * to the file that is going to be done.
 486          */
 487         if (XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, mp->m_allocsize_blocks))
 488                 return 0;
 489
 490         /*
 491          * Use the minimum preallocation size for small files or if we are
 492          * writing right after a hole.
 493          */
 494         if (XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, mp->m_dalign) ||
 495             !xfs_iext_prev_extent(ifp, &ncur, &prev) ||
 496             prev.br_startoff + prev.br_blockcount < offset_fsb)
 497                 return mp->m_allocsize_blocks;
 498
 499         /*
 500          * Take the size of the preceding data extents as the basis for the
 501          * preallocation size. Note that we don't care if the previous extents
 502          * are written or not.
 503          */
 504         plen = prev.br_blockcount;
 505         while (xfs_iext_prev_extent(ifp, &ncur, &got)) {
 506                 if (plen > XFS_MAX_BMBT_EXTLEN / 2 ||
 507                     isnullstartblock(got.br_startblock) ||
 508                     got.br_startoff + got.br_blockcount != prev.br_startoff ||
 509                     got.br_startblock + got.br_blockcount != prev.br_startblock)
 510                         break;
 511                 plen += got.br_blockcount;
 512                 prev = got;
 513         }
 514
 515         /*
 516          * If the size of the extents is greater than half the maximum extent
 517          * length, then use the current offset as the basis.  This ensures that
 518          * for large files the preallocation size always extends to
 519          * XFS_BMBT_MAX_EXTLEN rather than falling short due to things like stripe
 520          * unit/width alignment of real extents.
 521          */
 522         alloc_blocks = plen * 2;
 523         if (alloc_blocks > XFS_MAX_BMBT_EXTLEN)
 524                 alloc_blocks = XFS_B_TO_FSB(mp, offset);
 525         qblocks = alloc_blocks;
 526
 527         /*
 528          * XFS_BMBT_MAX_EXTLEN is not a power of two value but we round the prealloc
 529          * down to the nearest power of two value after throttling. To prevent
 530          * the round down from unconditionally reducing the maximum supported
 531          * prealloc size, we round up first, apply appropriate throttling, round
 532          * down and cap the value to XFS_BMBT_MAX_EXTLEN.
 533          */
 534         alloc_blocks = XFS_FILEOFF_MIN(roundup_pow_of_two(XFS_MAX_BMBT_EXTLEN),
 535                                        alloc_blocks);
 536
 537         if (unlikely(XFS_IS_REALTIME_INODE(ip)))
 538                 freesp = xfs_rtbxlen_to_blen(mp,
 539                                 xfs_iomap_freesp(&mp->m_frextents,
 540                                         mp->m_low_rtexts, &shift));
 541         else
 542                 freesp = xfs_iomap_freesp(&mp->m_fdblocks, mp->m_low_space,
 543                                 &shift);
 544
 545         /*
 546          * Check each quota to cap the prealloc size, provide a shift value to
 547          * throttle with and adjust amount of available space.
 548          */
 549         if (xfs_quota_need_throttle(ip, XFS_DQTYPE_USER, alloc_blocks))
 550                 xfs_quota_calc_throttle(ip, XFS_DQTYPE_USER, &qblocks, &qshift,
 551                                         &freesp);
 552         if (xfs_quota_need_throttle(ip, XFS_DQTYPE_GROUP, alloc_blocks))
 553                 xfs_quota_calc_throttle(ip, XFS_DQTYPE_GROUP, &qblocks, &qshift,
 554                                         &freesp);
 555         if (xfs_quota_need_throttle(ip, XFS_DQTYPE_PROJ, alloc_blocks))
 556                 xfs_quota_calc_throttle(ip, XFS_DQTYPE_PROJ, &qblocks, &qshift,
 557                                         &freesp);
 558
 559         /*
 560          * The final prealloc size is set to the minimum of free space available
 561          * in each of the quotas and the overall filesystem.
 562          *
 563          * The shift throttle value is set to the maximum value as determined by
 564          * the global low free space values and per-quota low free space values.
 565          */
 566         alloc_blocks = min(alloc_blocks, qblocks);
 567         shift = max(shift, qshift);
 568
 569         if (shift)
 570                 alloc_blocks >>= shift;
 571         /*
 572          * rounddown_pow_of_two() returns an undefined result if we pass in
 573          * alloc_blocks = 0.
 574          */
 575         if (alloc_blocks)
 576                 alloc_blocks = rounddown_pow_of_two(alloc_blocks);
 577         if (alloc_blocks > XFS_MAX_BMBT_EXTLEN)
 578                 alloc_blocks = XFS_MAX_BMBT_EXTLEN;
 579
 580         /*
 581          * If we are still trying to allocate more space than is
 582          * available, squash the prealloc hard. This can happen if we
 583          * have a large file on a small filesystem and the above
 584          * lowspace thresholds are smaller than XFS_BMBT_MAX_EXTLEN.
 585          */
 586         while (alloc_blocks && alloc_blocks >= freesp)
 587                 alloc_blocks >>= 4;
 588         if (alloc_blocks < mp->m_allocsize_blocks)
 589                 alloc_blocks = mp->m_allocsize_blocks;
 590         trace_xfs_iomap_prealloc_size(ip, alloc_blocks, shift,
 591                                       mp->m_allocsize_blocks);
 592         return alloc_blocks;
 593 }
 594
 595 int
 596 xfs_iomap_write_unwritten(
 597         xfs_inode_t     *ip,
 598         xfs_off_t       offset,
 599         xfs_off_t       count,
 600         bool            update_isize)
 601 {
 602         xfs_mount_t     *mp = ip->i_mount;
 603         xfs_fileoff_t   offset_fsb;
 604         xfs_filblks_t   count_fsb;
 605         xfs_filblks_t   numblks_fsb;
 606         int             nimaps;
 607         xfs_trans_t     *tp;
 608         xfs_bmbt_irec_t imap;
 609         struct inode    *inode = VFS_I(ip);
 610         xfs_fsize_t     i_size;
 611         uint            resblks;
 612         int             error;
 613
 614         trace_xfs_unwritten_convert(ip, offset, count);
 615
 616         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 617         count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
 618         count_fsb = (xfs_filblks_t)(count_fsb - offset_fsb);
 619
 620         /*
 621          * Reserve enough blocks in this transaction for two complete extent
 622          * btree splits.  We may be converting the middle part of an unwritten
 623          * extent and in this case we will insert two new extents in the btree
 624          * each of which could cause a full split.
 625          *
 626          * This reservation amount will be used in the first call to
 627          * xfs_bmbt_split() to select an AG with enough space to satisfy the
 628          * rest of the operation.
 629          */
 630         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
 631
 632         /* Attach dquots so that bmbt splits are accounted correctly. */
 633         error = xfs_qm_dqattach(ip);
 634         if (error)
 635                 return error;
 636
 637         do {
 638                 /*
 639                  * Set up a transaction to convert the range of extents
 640                  * from unwritten to real. Do allocations in a loop until
 641                  * we have covered the range passed in.
 642                  *
 643                  * Note that we can't risk to recursing back into the filesystem
 644                  * here as we might be asked to write out the same inode that we
 645                  * complete here and might deadlock on the iolock.
 646                  */
 647                 error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, resblks,
 648                                 0, true, &tp);
 649                 if (error)
 650                         return error;
 651
 652                 error = xfs_iext_count_extend(tp, ip, XFS_DATA_FORK,
 653                                 XFS_IEXT_WRITE_UNWRITTEN_CNT);
 654                 if (error)
 655                         goto error_on_bmapi_transaction;
 656
 657                 /*
 658                  * Modify the unwritten extent state of the buffer.
 659                  */
 660                 nimaps = 1;
 661                 error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb,
 662                                         XFS_BMAPI_CONVERT, resblks, &imap,
 663                                         &nimaps);
 664                 if (error)
 665                         goto error_on_bmapi_transaction;
 666
 667                 /*
 668                  * Log the updated inode size as we go.  We have to be careful
 669                  * to only log it up to the actual write offset if it is
 670                  * halfway into a block.
 671                  */
 672                 i_size = XFS_FSB_TO_B(mp, offset_fsb + count_fsb);
 673                 if (i_size > offset + count)
 674                         i_size = offset + count;
 675                 if (update_isize && i_size > i_size_read(inode))
 676                         i_size_write(inode, i_size);
 677                 i_size = xfs_new_eof(ip, i_size);
 678                 if (i_size) {
 679                         ip->i_disk_size = i_size;
 680                         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 681                 }
 682
 683                 error = xfs_trans_commit(tp);
 684                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 685                 if (error)
 686                         return error;
 687
 688                 if (unlikely(!xfs_valid_startblock(ip, imap.br_startblock))) {
 689                         xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
 690                         return xfs_alert_fsblock_zero(ip, &imap);
 691                 }
 692
 693                 if ((numblks_fsb = imap.br_blockcount) == 0) {
 694                         /*
 695                          * The numblks_fsb value should always get
 696                          * smaller, otherwise the loop is stuck.
 697                          */
 698                         ASSERT(imap.br_blockcount);
 699                         break;
 700                 }
 701                 offset_fsb += numblks_fsb;
 702                 count_fsb -= numblks_fsb;
 703         } while (count_fsb > 0);
 704
 705         return 0;
 706
 707 error_on_bmapi_transaction:
 708         xfs_trans_cancel(tp);
 709         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 710         return error;
 711 }
 712
 713 static inline bool
 714 imap_needs_alloc(
 715         struct inode            *inode,
 716         unsigned                flags,
 717         struct xfs_bmbt_irec    *imap,
 718         int                     nimaps)
 719 {
 720         /* don't allocate blocks when just zeroing */
 721         if (flags & IOMAP_ZERO)
 722                 return false;
 723         if (!nimaps ||
 724             imap->br_startblock == HOLESTARTBLOCK ||
 725             imap->br_startblock == DELAYSTARTBLOCK)
 726                 return true;
 727         /* we convert unwritten extents before copying the data for DAX */
 728         if ((flags & IOMAP_DAX) && imap->br_state == XFS_EXT_UNWRITTEN)
 729                 return true;
 730         return false;
 731 }
 732
 733 static inline bool
 734 imap_needs_cow(
 735         struct xfs_inode        *ip,
 736         unsigned int            flags,
 737         struct xfs_bmbt_irec    *imap,
 738         int                     nimaps)
 739 {
 740         if (!xfs_is_cow_inode(ip))
 741                 return false;
 742
 743         /* when zeroing we don't have to COW holes or unwritten extents */
 744         if (flags & (IOMAP_UNSHARE | IOMAP_ZERO)) {
 745                 if (!nimaps ||
 746                     imap->br_startblock == HOLESTARTBLOCK ||
 747                     imap->br_state == XFS_EXT_UNWRITTEN)
 748                         return false;
 749         }
 750
 751         return true;
 752 }
 753
 754 /*
 755  * Extents not yet cached requires exclusive access, don't block for
 756  * IOMAP_NOWAIT.
 757  *
 758  * This is basically an opencoded xfs_ilock_data_map_shared() call, but with
 759  * support for IOMAP_NOWAIT.
 760  */
 761 static int
 762 xfs_ilock_for_iomap(
 763         struct xfs_inode        *ip,
 764         unsigned                flags,
 765         unsigned                *lockmode)
 766 {
 767         if (flags & IOMAP_NOWAIT) {
 768                 if (xfs_need_iread_extents(&ip->i_df))
 769                         return -EAGAIN;
 770                 if (!xfs_ilock_nowait(ip, *lockmode))
 771                         return -EAGAIN;
 772         } else {
 773                 if (xfs_need_iread_extents(&ip->i_df))
 774                         *lockmode = XFS_ILOCK_EXCL;
 775                 xfs_ilock(ip, *lockmode);
 776         }
 777
 778         return 0;
 779 }
 780
 781 /*
 782  * Check that the imap we are going to return to the caller spans the entire
 783  * range that the caller requested for the IO.
 784  */
 785 static bool
 786 imap_spans_range(
 787         struct xfs_bmbt_irec    *imap,
 788         xfs_fileoff_t           offset_fsb,
 789         xfs_fileoff_t           end_fsb)
 790 {
 791         if (imap->br_startoff > offset_fsb)
 792                 return false;
 793         if (imap->br_startoff + imap->br_blockcount < end_fsb)
 794                 return false;
 795         return true;
 796 }
 797
 798 static int
 799 xfs_direct_write_iomap_begin(
 800         struct inode            *inode,
 801         loff_t                  offset,
 802         loff_t                  length,
 803         unsigned                flags,
 804         struct iomap            *iomap,
 805         struct iomap            *srcmap)
 806 {
 807         struct xfs_inode        *ip = XFS_I(inode);
 808         struct xfs_mount        *mp = ip->i_mount;
 809         struct xfs_bmbt_irec    imap, cmap;
 810         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
 811         xfs_fileoff_t           end_fsb = xfs_iomap_end_fsb(mp, offset, length);
 812         int                     nimaps = 1, error = 0;
 813         bool                    shared = false;
 814         u16                     iomap_flags = 0;
 815         unsigned int            lockmode;
 816         u64                     seq;
 817
 818         ASSERT(flags & (IOMAP_WRITE | IOMAP_ZERO));
 819
 820         if (xfs_is_shutdown(mp))
 821                 return -EIO;
 822
 823         /*
 824          * Writes that span EOF might trigger an IO size update on completion,
 825          * so consider them to be dirty for the purposes of O_DSYNC even if
 826          * there is no other metadata changes pending or have been made here.
 827          */
 828         if (offset + length > i_size_read(inode))
 829                 iomap_flags |= IOMAP_F_DIRTY;
 830
 831         /*
 832          * COW writes may allocate delalloc space or convert unwritten COW
 833          * extents, so we need to make sure to take the lock exclusively here.
 834          */
 835         if (xfs_is_cow_inode(ip))
 836                 lockmode = XFS_ILOCK_EXCL;
 837         else
 838                 lockmode = XFS_ILOCK_SHARED;
 839
 840 relock:
 841         error = xfs_ilock_for_iomap(ip, flags, &lockmode);
 842         if (error)
 843                 return error;
 844
 845         /*
 846          * The reflink iflag could have changed since the earlier unlocked
 847          * check, check if it again and relock if needed.
 848          */
 849         if (xfs_is_cow_inode(ip) && lockmode == XFS_ILOCK_SHARED) {
 850                 xfs_iunlock(ip, lockmode);
 851                 lockmode = XFS_ILOCK_EXCL;
 852                 goto relock;
 853         }
 854
 855         error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
 856                                &nimaps, 0);
 857         if (error)
 858                 goto out_unlock;
 859
 860         if (imap_needs_cow(ip, flags, &imap, nimaps)) {
 861                 error = -EAGAIN;
 862                 if (flags & IOMAP_NOWAIT)
 863                         goto out_unlock;
 864
 865                 /* may drop and re-acquire the ilock */
 866                 error = xfs_reflink_allocate_cow(ip, &imap, &cmap, &shared,
 867                                 &lockmode,
 868                                 (flags & IOMAP_DIRECT) || IS_DAX(inode));
 869                 if (error)
 870                         goto out_unlock;
 871                 if (shared)
 872                         goto out_found_cow;
 873                 end_fsb = imap.br_startoff + imap.br_blockcount;
 874                 length = XFS_FSB_TO_B(mp, end_fsb) - offset;
 875         }
 876
 877         if (imap_needs_alloc(inode, flags, &imap, nimaps))
 878                 goto allocate_blocks;
 879
 880         /*
 881          * NOWAIT and OVERWRITE I/O needs to span the entire requested I/O with
 882          * a single map so that we avoid partial IO failures due to the rest of
 883          * the I/O range not covered by this map triggering an EAGAIN condition
 884          * when it is subsequently mapped and aborting the I/O.
 885          */
 886         if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY)) {
 887                 error = -EAGAIN;
 888                 if (!imap_spans_range(&imap, offset_fsb, end_fsb))
 889                         goto out_unlock;
 890         }
 891
 892         /*
 893          * For overwrite only I/O, we cannot convert unwritten extents without
 894          * requiring sub-block zeroing.  This can only be done under an
 895          * exclusive IOLOCK, hence return -EAGAIN if this is not a written
 896          * extent to tell the caller to try again.
 897          */
 898         if (flags & IOMAP_OVERWRITE_ONLY) {
 899                 error = -EAGAIN;
 900                 if (imap.br_state != XFS_EXT_NORM &&
 901                     ((offset | length) & mp->m_blockmask))
 902                         goto out_unlock;
 903         }
 904
 905         seq = xfs_iomap_inode_sequence(ip, iomap_flags);
 906         xfs_iunlock(ip, lockmode);
 907         trace_xfs_iomap_found(ip, offset, length, XFS_DATA_FORK, &imap);
 908         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, iomap_flags, seq);
 909
 910 allocate_blocks:
 911         error = -EAGAIN;
 912         if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY))
 913                 goto out_unlock;
 914
 915         /*
 916          * We cap the maximum length we map to a sane size  to keep the chunks
 917          * of work done where somewhat symmetric with the work writeback does.
 918          * This is a completely arbitrary number pulled out of thin air as a
 919          * best guess for initial testing.
 920          *
 921          * Note that the values needs to be less than 32-bits wide until the
 922          * lower level functions are updated.
 923          */
 924         length = min_t(loff_t, length, 1024 * PAGE_SIZE);
 925         end_fsb = xfs_iomap_end_fsb(mp, offset, length);
 926
 927         if (offset + length > XFS_ISIZE(ip))
 928                 end_fsb = xfs_iomap_eof_align_last_fsb(ip, end_fsb);
 929         else if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
 930                 end_fsb = min(end_fsb, imap.br_startoff + imap.br_blockcount);
 931         xfs_iunlock(ip, lockmode);
 932
 933         error = xfs_iomap_write_direct(ip, offset_fsb, end_fsb - offset_fsb,
 934                         flags, &imap, &seq);
 935         if (error)
 936                 return error;
 937
 938         trace_xfs_iomap_alloc(ip, offset, length, XFS_DATA_FORK, &imap);
 939         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags,
 940                                  iomap_flags | IOMAP_F_NEW, seq);
 941
 942 out_found_cow:
 943         length = XFS_FSB_TO_B(mp, cmap.br_startoff + cmap.br_blockcount);
 944         trace_xfs_iomap_found(ip, offset, length - offset, XFS_COW_FORK, &cmap);
 945         if (imap.br_startblock != HOLESTARTBLOCK) {
 946                 seq = xfs_iomap_inode_sequence(ip, 0);
 947                 error = xfs_bmbt_to_iomap(ip, srcmap, &imap, flags, 0, seq);
 948                 if (error)
 949                         goto out_unlock;
 950         }
 951         seq = xfs_iomap_inode_sequence(ip, IOMAP_F_SHARED);
 952         xfs_iunlock(ip, lockmode);
 953         return xfs_bmbt_to_iomap(ip, iomap, &cmap, flags, IOMAP_F_SHARED, seq);
 954
 955 out_unlock:
 956         if (lockmode)
 957                 xfs_iunlock(ip, lockmode);
 958         return error;
 959 }
 960
 961 const struct iomap_ops xfs_direct_write_iomap_ops = {
 962         .iomap_begin            = xfs_direct_write_iomap_begin,
 963 };
 964
 965 static int
 966 xfs_dax_write_iomap_end(
 967         struct inode            *inode,
 968         loff_t                  pos,
 969         loff_t                  length,
 970         ssize_t                 written,
 971         unsigned                flags,
 972         struct iomap            *iomap)
 973 {
 974         struct xfs_inode        *ip = XFS_I(inode);
 975
 976         if (!xfs_is_cow_inode(ip))
 977                 return 0;
 978
 979         if (!written) {
 980                 xfs_reflink_cancel_cow_range(ip, pos, length, true);
 981                 return 0;
 982         }
 983
 984         return xfs_reflink_end_cow(ip, pos, written);
 985 }
 986
 987 const struct iomap_ops xfs_dax_write_iomap_ops = {
 988         .iomap_begin    = xfs_direct_write_iomap_begin,
 989         .iomap_end      = xfs_dax_write_iomap_end,
 990 };
 991
 992 static int
 993 xfs_buffered_write_iomap_begin(
 994         struct inode            *inode,
 995         loff_t                  offset,
 996         loff_t                  count,
 997         unsigned                flags,
 998         struct iomap            *iomap,
 999         struct iomap            *srcmap)
1000 {
1001         struct xfs_inode        *ip = XFS_I(inode);
1002         struct xfs_mount        *mp = ip->i_mount;
1003         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
1004         xfs_fileoff_t           end_fsb = xfs_iomap_end_fsb(mp, offset, count);
1005         struct xfs_bmbt_irec    imap, cmap;
1006         struct xfs_iext_cursor  icur, ccur;
1007         xfs_fsblock_t           prealloc_blocks = 0;
1008         bool                    eof = false, cow_eof = false, shared = false;
1009         int                     allocfork = XFS_DATA_FORK;
1010         int                     error = 0;
1011         unsigned int            lockmode = XFS_ILOCK_EXCL;
1012         unsigned int            iomap_flags = 0;
1013         u64                     seq;
1014
1015         if (xfs_is_shutdown(mp))
1016                 return -EIO;
1017
1018         /* we can't use delayed allocations when using extent size hints */
1019         if (xfs_get_extsz_hint(ip))
1020                 return xfs_direct_write_iomap_begin(inode, offset, count,
1021                                 flags, iomap, srcmap);
1022
1023         error = xfs_qm_dqattach(ip);
1024         if (error)
1025                 return error;
1026
1027         error = xfs_ilock_for_iomap(ip, flags, &lockmode);
1028         if (error)
1029                 return error;
1030
1031         if (XFS_IS_CORRUPT(mp, !xfs_ifork_has_extents(&ip->i_df)) ||
1032             XFS_TEST_ERROR(false, mp, XFS_ERRTAG_BMAPIFORMAT)) {
1033                 xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
1034                 error = -EFSCORRUPTED;
1035                 goto out_unlock;
1036         }
1037
1038         XFS_STATS_INC(mp, xs_blk_mapw);
1039
1040         error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK);
1041         if (error)
1042                 goto out_unlock;
1043
1044         /*
1045          * Search the data fork first to look up our source mapping.  We
1046          * always need the data fork map, as we have to return it to the
1047          * iomap code so that the higher level write code can read data in to
1048          * perform read-modify-write cycles for unaligned writes.
1049          */
1050         eof = !xfs_iext_lookup_extent(ip, &ip->i_df, offset_fsb, &icur, &imap);
1051         if (eof)
1052                 imap.br_startoff = end_fsb; /* fake hole until the end */
1053
1054         /* We never need to allocate blocks for zeroing or unsharing a hole. */
1055         if ((flags & (IOMAP_UNSHARE | IOMAP_ZERO)) &&
1056             imap.br_startoff > offset_fsb) {
1057                 xfs_hole_to_iomap(ip, iomap, offset_fsb, imap.br_startoff);
1058                 goto out_unlock;
1059         }
1060
1061         /*
1062          * For zeroing, trim a delalloc extent that extends beyond the EOF
1063          * block.  If it starts beyond the EOF block, convert it to an
1064          * unwritten extent.
1065          */
1066         if ((flags & IOMAP_ZERO) && imap.br_startoff <= offset_fsb &&
1067             isnullstartblock(imap.br_startblock)) {
1068                 xfs_fileoff_t eof_fsb = XFS_B_TO_FSB(mp, XFS_ISIZE(ip));
1069
1070                 if (offset_fsb >= eof_fsb)
1071                         goto convert_delay;
1072                 if (end_fsb > eof_fsb) {
1073                         end_fsb = eof_fsb;
1074                         xfs_trim_extent(&imap, offset_fsb,
1075                                         end_fsb - offset_fsb);
1076                 }
1077         }
1078
1079         /*
1080          * Search the COW fork extent list even if we did not find a data fork
1081          * extent.  This serves two purposes: first this implements the
1082          * speculative preallocation using cowextsize, so that we also unshare
1083          * block adjacent to shared blocks instead of just the shared blocks
1084          * themselves.  Second the lookup in the extent list is generally faster
1085          * than going out to the shared extent tree.
1086          */
1087         if (xfs_is_cow_inode(ip)) {
1088                 if (!ip->i_cowfp) {
1089                         ASSERT(!xfs_is_reflink_inode(ip));
1090                         xfs_ifork_init_cow(ip);
1091                 }
1092                 cow_eof = !xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb,
1093                                 &ccur, &cmap);
1094                 if (!cow_eof && cmap.br_startoff <= offset_fsb) {
1095                         trace_xfs_reflink_cow_found(ip, &cmap);
1096                         goto found_cow;
1097                 }
1098         }
1099
1100         if (imap.br_startoff <= offset_fsb) {
1101                 /*
1102                  * For reflink files we may need a delalloc reservation when
1103                  * overwriting shared extents.   This includes zeroing of
1104                  * existing extents that contain data.
1105                  */
1106                 if (!xfs_is_cow_inode(ip) ||
1107                     ((flags & IOMAP_ZERO) && imap.br_state != XFS_EXT_NORM)) {
1108                         trace_xfs_iomap_found(ip, offset, count, XFS_DATA_FORK,
1109                                         &imap);
1110                         goto found_imap;
1111                 }
1112
1113                 xfs_trim_extent(&imap, offset_fsb, end_fsb - offset_fsb);
1114
1115                 /* Trim the mapping to the nearest shared extent boundary. */
1116                 error = xfs_bmap_trim_cow(ip, &imap, &shared);
1117                 if (error)
1118                         goto out_unlock;
1119
1120                 /* Not shared?  Just report the (potentially capped) extent. */
1121                 if (!shared) {
1122                         trace_xfs_iomap_found(ip, offset, count, XFS_DATA_FORK,
1123                                         &imap);
1124                         goto found_imap;
1125                 }
1126
1127                 /*
1128                  * Fork all the shared blocks from our write offset until the
1129                  * end of the extent.
1130                  */
1131                 allocfork = XFS_COW_FORK;
1132                 end_fsb = imap.br_startoff + imap.br_blockcount;
1133         } else {
1134                 /*
1135                  * We cap the maximum length we map here to MAX_WRITEBACK_PAGES
1136                  * pages to keep the chunks of work done where somewhat
1137                  * symmetric with the work writeback does.  This is a completely
1138                  * arbitrary number pulled out of thin air.
1139                  *
1140                  * Note that the values needs to be less than 32-bits wide until
1141                  * the lower level functions are updated.
1142                  */
1143                 count = min_t(loff_t, count, 1024 * PAGE_SIZE);
1144                 end_fsb = xfs_iomap_end_fsb(mp, offset, count);
1145
1146                 if (xfs_is_always_cow_inode(ip))
1147                         allocfork = XFS_COW_FORK;
1148         }
1149
1150         if (eof && offset + count > XFS_ISIZE(ip)) {
1151                 /*
1152                  * Determine the initial size of the preallocation.
1153                  * We clean up any extra preallocation when the file is closed.
1154                  */
1155                 if (xfs_has_allocsize(mp))
1156                         prealloc_blocks = mp->m_allocsize_blocks;
1157                 else if (allocfork == XFS_DATA_FORK)
1158                         prealloc_blocks = xfs_iomap_prealloc_size(ip, allocfork,
1159                                                 offset, count, &icur);
1160                 else
1161                         prealloc_blocks = xfs_iomap_prealloc_size(ip, allocfork,
1162                                                 offset, count, &ccur);
1163                 if (prealloc_blocks) {
1164                         xfs_extlen_t    align;
1165                         xfs_off_t       end_offset;
1166                         xfs_fileoff_t   p_end_fsb;
1167
1168                         end_offset = XFS_ALLOC_ALIGN(mp, offset + count - 1);
1169                         p_end_fsb = XFS_B_TO_FSBT(mp, end_offset) +
1170                                         prealloc_blocks;
1171
1172                         align = xfs_eof_alignment(ip);
1173                         if (align)
1174                                 p_end_fsb = roundup_64(p_end_fsb, align);
1175
1176                         p_end_fsb = min(p_end_fsb,
1177                                 XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes));
1178                         ASSERT(p_end_fsb > offset_fsb);
1179                         prealloc_blocks = p_end_fsb - end_fsb;
1180                 }
1181         }
1182
1183         /*
1184          * Flag newly allocated delalloc blocks with IOMAP_F_NEW so we punch
1185          * them out if the write happens to fail.
1186          */
1187         iomap_flags |= IOMAP_F_NEW;
1188         if (allocfork == XFS_COW_FORK) {
1189                 error = xfs_bmapi_reserve_delalloc(ip, allocfork, offset_fsb,
1190                                 end_fsb - offset_fsb, prealloc_blocks, &cmap,
1191                                 &ccur, cow_eof);
1192                 if (error)
1193                         goto out_unlock;
1194
1195                 trace_xfs_iomap_alloc(ip, offset, count, allocfork, &cmap);
1196                 goto found_cow;
1197         }
1198
1199         error = xfs_bmapi_reserve_delalloc(ip, allocfork, offset_fsb,
1200                         end_fsb - offset_fsb, prealloc_blocks, &imap, &icur,
1201                         eof);
1202         if (error)
1203                 goto out_unlock;
1204
1205         trace_xfs_iomap_alloc(ip, offset, count, allocfork, &imap);
1206 found_imap:
1207         seq = xfs_iomap_inode_sequence(ip, iomap_flags);
1208         xfs_iunlock(ip, lockmode);
1209         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, iomap_flags, seq);
1210
1211 convert_delay:
1212         xfs_iunlock(ip, lockmode);
1213         truncate_pagecache(inode, offset);
1214         error = xfs_bmapi_convert_delalloc(ip, XFS_DATA_FORK, offset,
1215                                            iomap, NULL);
1216         if (error)
1217                 return error;
1218
1219         trace_xfs_iomap_alloc(ip, offset, count, XFS_DATA_FORK, &imap);
1220         return 0;
1221
1222 found_cow:
1223         if (imap.br_startoff <= offset_fsb) {
1224                 error = xfs_bmbt_to_iomap(ip, srcmap, &imap, flags, 0,
1225                                 xfs_iomap_inode_sequence(ip, 0));
1226                 if (error)
1227                         goto out_unlock;
1228         } else {
1229                 xfs_trim_extent(&cmap, offset_fsb,
1230                                 imap.br_startoff - offset_fsb);
1231         }
1232
1233         iomap_flags |= IOMAP_F_SHARED;
1234         seq = xfs_iomap_inode_sequence(ip, iomap_flags);
1235         xfs_iunlock(ip, lockmode);
1236         return xfs_bmbt_to_iomap(ip, iomap, &cmap, flags, iomap_flags, seq);
1237
1238 out_unlock:
1239         xfs_iunlock(ip, lockmode);
1240         return error;
1241 }
1242
1243 static void
1244 xfs_buffered_write_delalloc_punch(
1245         struct inode            *inode,
1246         loff_t                  offset,
1247         loff_t                  length,
1248         struct iomap            *iomap)
1249 {
1250         xfs_bmap_punch_delalloc_range(XFS_I(inode),
1251                         (iomap->flags & IOMAP_F_SHARED) ?
1252                                 XFS_COW_FORK : XFS_DATA_FORK,
1253                         offset, offset + length);
1254 }
1255
1256 static int
1257 xfs_buffered_write_iomap_end(
1258         struct inode            *inode,
1259         loff_t                  offset,
1260         loff_t                  length,
1261         ssize_t                 written,
1262         unsigned                flags,
1263         struct iomap            *iomap)
1264 {
1265         loff_t                  start_byte, end_byte;
1266
1267         /* If we didn't reserve the blocks, we're not allowed to punch them. */
1268         if (iomap->type != IOMAP_DELALLOC || !(iomap->flags & IOMAP_F_NEW))
1269                 return 0;
1270
1271         /*
1272          * iomap_page_mkwrite() will never fail in a way that requires delalloc
1273          * extents that it allocated to be revoked.  Hence never try to release
1274          * them here.
1275          */
1276         if (flags & IOMAP_FAULT)
1277                 return 0;
1278
1279         /* Nothing to do if we've written the entire delalloc extent */
1280         start_byte = iomap_last_written_block(inode, offset, written);
1281         end_byte = round_up(offset + length, i_blocksize(inode));
1282         if (start_byte >= end_byte)
1283                 return 0;
1284
1285         /* For zeroing operations the callers already hold invalidate_lock. */
1286         if (flags & (IOMAP_UNSHARE | IOMAP_ZERO)) {
1287                 rwsem_assert_held_write(&inode->i_mapping->invalidate_lock);
1288                 iomap_write_delalloc_release(inode, start_byte, end_byte, flags,
1289                                 iomap, xfs_buffered_write_delalloc_punch);
1290         } else {
1291                 filemap_invalidate_lock(inode->i_mapping);
1292                 iomap_write_delalloc_release(inode, start_byte, end_byte, flags,
1293                                 iomap, xfs_buffered_write_delalloc_punch);
1294                 filemap_invalidate_unlock(inode->i_mapping);
1295         }
1296
1297         return 0;
1298 }
1299
1300 const struct iomap_ops xfs_buffered_write_iomap_ops = {
1301         .iomap_begin            = xfs_buffered_write_iomap_begin,
1302         .iomap_end              = xfs_buffered_write_iomap_end,
1303 };
1304
1305 static int
1306 xfs_read_iomap_begin(
1307         struct inode            *inode,
1308         loff_t                  offset,
1309         loff_t                  length,
1310         unsigned                flags,
1311         struct iomap            *iomap,
1312         struct iomap            *srcmap)
1313 {
1314         struct xfs_inode        *ip = XFS_I(inode);
1315         struct xfs_mount        *mp = ip->i_mount;
1316         struct xfs_bmbt_irec    imap;
1317         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
1318         xfs_fileoff_t           end_fsb = xfs_iomap_end_fsb(mp, offset, length);
1319         int                     nimaps = 1, error = 0;
1320         bool                    shared = false;
1321         unsigned int            lockmode = XFS_ILOCK_SHARED;
1322         u64                     seq;
1323
1324         ASSERT(!(flags & (IOMAP_WRITE | IOMAP_ZERO)));
1325
1326         if (xfs_is_shutdown(mp))
1327                 return -EIO;
1328
1329         error = xfs_ilock_for_iomap(ip, flags, &lockmode);
1330         if (error)
1331                 return error;
1332         error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
1333                                &nimaps, 0);
1334         if (!error && ((flags & IOMAP_REPORT) || IS_DAX(inode)))
1335                 error = xfs_reflink_trim_around_shared(ip, &imap, &shared);
1336         seq = xfs_iomap_inode_sequence(ip, shared ? IOMAP_F_SHARED : 0);
1337         xfs_iunlock(ip, lockmode);
1338
1339         if (error)
1340                 return error;
1341         trace_xfs_iomap_found(ip, offset, length, XFS_DATA_FORK, &imap);
1342         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags,
1343                                  shared ? IOMAP_F_SHARED : 0, seq);
1344 }
1345
1346 const struct iomap_ops xfs_read_iomap_ops = {
1347         .iomap_begin            = xfs_read_iomap_begin,
1348 };
1349
1350 static int
1351 xfs_seek_iomap_begin(
1352         struct inode            *inode,
1353         loff_t                  offset,
1354         loff_t                  length,
1355         unsigned                flags,
1356         struct iomap            *iomap,
1357         struct iomap            *srcmap)
1358 {
1359         struct xfs_inode        *ip = XFS_I(inode);
1360         struct xfs_mount        *mp = ip->i_mount;
1361         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
1362         xfs_fileoff_t           end_fsb = XFS_B_TO_FSB(mp, offset + length);
1363         xfs_fileoff_t           cow_fsb = NULLFILEOFF, data_fsb = NULLFILEOFF;
1364         struct xfs_iext_cursor  icur;
1365         struct xfs_bmbt_irec    imap, cmap;
1366         int                     error = 0;
1367         unsigned                lockmode;
1368         u64                     seq;
1369
1370         if (xfs_is_shutdown(mp))
1371                 return -EIO;
1372
1373         lockmode = xfs_ilock_data_map_shared(ip);
1374         error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK);
1375         if (error)
1376                 goto out_unlock;
1377
1378         if (xfs_iext_lookup_extent(ip, &ip->i_df, offset_fsb, &icur, &imap)) {
1379                 /*
1380                  * If we found a data extent we are done.
1381                  */
1382                 if (imap.br_startoff <= offset_fsb)
1383                         goto done;
1384                 data_fsb = imap.br_startoff;
1385         } else {
1386                 /*
1387                  * Fake a hole until the end of the file.
1388                  */
1389                 data_fsb = xfs_iomap_end_fsb(mp, offset, length);
1390         }
1391
1392         /*
1393          * If a COW fork extent covers the hole, report it - capped to the next
1394          * data fork extent:
1395          */
1396         if (xfs_inode_has_cow_data(ip) &&
1397             xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &cmap))
1398                 cow_fsb = cmap.br_startoff;
1399         if (cow_fsb != NULLFILEOFF && cow_fsb <= offset_fsb) {
1400                 if (data_fsb < cow_fsb + cmap.br_blockcount)
1401                         end_fsb = min(end_fsb, data_fsb);
1402                 xfs_trim_extent(&cmap, offset_fsb, end_fsb - offset_fsb);
1403                 seq = xfs_iomap_inode_sequence(ip, IOMAP_F_SHARED);
1404                 error = xfs_bmbt_to_iomap(ip, iomap, &cmap, flags,
1405                                 IOMAP_F_SHARED, seq);
1406                 /*
1407                  * This is a COW extent, so we must probe the page cache
1408                  * because there could be dirty page cache being backed
1409                  * by this extent.
1410                  */
1411                 iomap->type = IOMAP_UNWRITTEN;
1412                 goto out_unlock;
1413         }
1414
1415         /*
1416          * Else report a hole, capped to the next found data or COW extent.
1417          */
1418         if (cow_fsb != NULLFILEOFF && cow_fsb < data_fsb)
1419                 imap.br_blockcount = cow_fsb - offset_fsb;
1420         else
1421                 imap.br_blockcount = data_fsb - offset_fsb;
1422         imap.br_startoff = offset_fsb;
1423         imap.br_startblock = HOLESTARTBLOCK;
1424         imap.br_state = XFS_EXT_NORM;
1425 done:
1426         seq = xfs_iomap_inode_sequence(ip, 0);
1427         xfs_trim_extent(&imap, offset_fsb, end_fsb - offset_fsb);
1428         error = xfs_bmbt_to_iomap(ip, iomap, &imap, flags, 0, seq);
1429 out_unlock:
1430         xfs_iunlock(ip, lockmode);
1431         return error;
1432 }
1433
1434 const struct iomap_ops xfs_seek_iomap_ops = {
1435         .iomap_begin            = xfs_seek_iomap_begin,
1436 };
1437
1438 static int
1439 xfs_xattr_iomap_begin(
1440         struct inode            *inode,
1441         loff_t                  offset,
1442         loff_t                  length,
1443         unsigned                flags,
1444         struct iomap            *iomap,
1445         struct iomap            *srcmap)
1446 {
1447         struct xfs_inode        *ip = XFS_I(inode);
1448         struct xfs_mount        *mp = ip->i_mount;
1449         xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset);
1450         xfs_fileoff_t           end_fsb = XFS_B_TO_FSB(mp, offset + length);
1451         struct xfs_bmbt_irec    imap;
1452         int                     nimaps = 1, error = 0;
1453         unsigned                lockmode;
1454         int                     seq;
1455
1456         if (xfs_is_shutdown(mp))
1457                 return -EIO;
1458
1459         lockmode = xfs_ilock_attr_map_shared(ip);
1460
1461         /* if there are no attribute fork or extents, return ENOENT */
1462         if (!xfs_inode_has_attr_fork(ip) || !ip->i_af.if_nextents) {
1463                 error = -ENOENT;
1464                 goto out_unlock;
1465         }
1466
1467         ASSERT(ip->i_af.if_format != XFS_DINODE_FMT_LOCAL);
1468         error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
1469                                &nimaps, XFS_BMAPI_ATTRFORK);
1470 out_unlock:
1471
1472         seq = xfs_iomap_inode_sequence(ip, IOMAP_F_XATTR);
1473         xfs_iunlock(ip, lockmode);
1474
1475         if (error)
1476                 return error;
1477         ASSERT(nimaps);
1478         return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, IOMAP_F_XATTR, seq);
1479 }
1480
1481 const struct iomap_ops xfs_xattr_iomap_ops = {
1482         .iomap_begin            = xfs_xattr_iomap_begin,
1483 };
1484
1485 int
1486 xfs_zero_range(
1487         struct xfs_inode        *ip,
1488         loff_t                  pos,
1489         loff_t                  len,
1490         bool                    *did_zero)
1491 {
1492         struct inode            *inode = VFS_I(ip);
1493
1494         xfs_assert_ilocked(ip, XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL);
1495
1496         if (IS_DAX(inode))
1497                 return dax_zero_range(inode, pos, len, did_zero,
1498                                       &xfs_dax_write_iomap_ops);
1499         return iomap_zero_range(inode, pos, len, did_zero,
1500                                 &xfs_buffered_write_iomap_ops);
1501 }
1502
1503 int
1504 xfs_truncate_page(
1505         struct xfs_inode        *ip,
1506         loff_t                  pos,
1507         bool                    *did_zero)
1508 {
1509         struct inode            *inode = VFS_I(ip);
1510
1511         if (IS_DAX(inode))
1512                 return dax_truncate_page(inode, pos, did_zero,
1513                                         &xfs_dax_write_iomap_ops);
1514         return iomap_truncate_page(inode, pos, did_zero,
1515                                    &xfs_buffered_write_iomap_ops);
1516 }