fs/xfs/libxfs/xfs_refcount.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Copyright (C) 2016 Oracle.  All Rights Reserved.
   4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
   5  */
   6 #include "xfs.h"
   7 #include "xfs_fs.h"
   8 #include "xfs_shared.h"
   9 #include "xfs_format.h"
  10 #include "xfs_log_format.h"
  11 #include "xfs_trans_resv.h"
  12 #include "xfs_sb.h"
  13 #include "xfs_mount.h"
  14 #include "xfs_defer.h"
  15 #include "xfs_btree.h"
  16 #include "xfs_bmap.h"
  17 #include "xfs_refcount_btree.h"
  18 #include "xfs_alloc.h"
  19 #include "xfs_errortag.h"
  20 #include "xfs_error.h"
  21 #include "xfs_trace.h"
  22 #include "xfs_cksum.h"
  23 #include "xfs_trans.h"
  24 #include "xfs_bit.h"
  25 #include "xfs_refcount.h"
  26 #include "xfs_rmap.h"
  27
  28 /* Allowable refcount adjustment amounts. */
  29 enum xfs_refc_adjust_op {
  30         XFS_REFCOUNT_ADJUST_INCREASE    = 1,
  31         XFS_REFCOUNT_ADJUST_DECREASE    = -1,
  32         XFS_REFCOUNT_ADJUST_COW_ALLOC   = 0,
  33         XFS_REFCOUNT_ADJUST_COW_FREE    = -1,
  34 };
  35
  36 STATIC int __xfs_refcount_cow_alloc(struct xfs_btree_cur *rcur,
  37                 xfs_agblock_t agbno, xfs_extlen_t aglen,
  38                 struct xfs_defer_ops *dfops);
  39 STATIC int __xfs_refcount_cow_free(struct xfs_btree_cur *rcur,
  40                 xfs_agblock_t agbno, xfs_extlen_t aglen,
  41                 struct xfs_defer_ops *dfops);
  42
  43 /*
  44  * Look up the first record less than or equal to [bno, len] in the btree
  45  * given by cur.
  46  */
  47 int
  48 xfs_refcount_lookup_le(
  49         struct xfs_btree_cur    *cur,
  50         xfs_agblock_t           bno,
  51         int                     *stat)
  52 {
  53         trace_xfs_refcount_lookup(cur->bc_mp, cur->bc_private.a.agno, bno,
  54                         XFS_LOOKUP_LE);
  55         cur->bc_rec.rc.rc_startblock = bno;
  56         cur->bc_rec.rc.rc_blockcount = 0;
  57         return xfs_btree_lookup(cur, XFS_LOOKUP_LE, stat);
  58 }
  59
  60 /*
  61  * Look up the first record greater than or equal to [bno, len] in the btree
  62  * given by cur.
  63  */
  64 int
  65 xfs_refcount_lookup_ge(
  66         struct xfs_btree_cur    *cur,
  67         xfs_agblock_t           bno,
  68         int                     *stat)
  69 {
  70         trace_xfs_refcount_lookup(cur->bc_mp, cur->bc_private.a.agno, bno,
  71                         XFS_LOOKUP_GE);
  72         cur->bc_rec.rc.rc_startblock = bno;
  73         cur->bc_rec.rc.rc_blockcount = 0;
  74         return xfs_btree_lookup(cur, XFS_LOOKUP_GE, stat);
  75 }
  76
  77 /*
  78  * Look up the first record equal to [bno, len] in the btree
  79  * given by cur.
  80  */
  81 int
  82 xfs_refcount_lookup_eq(
  83         struct xfs_btree_cur    *cur,
  84         xfs_agblock_t           bno,
  85         int                     *stat)
  86 {
  87         trace_xfs_refcount_lookup(cur->bc_mp, cur->bc_private.a.agno, bno,
  88                         XFS_LOOKUP_LE);
  89         cur->bc_rec.rc.rc_startblock = bno;
  90         cur->bc_rec.rc.rc_blockcount = 0;
  91         return xfs_btree_lookup(cur, XFS_LOOKUP_EQ, stat);
  92 }
  93
  94 /* Convert on-disk record to in-core format. */
  95 void
  96 xfs_refcount_btrec_to_irec(
  97         union xfs_btree_rec             *rec,
  98         struct xfs_refcount_irec        *irec)
  99 {
 100         irec->rc_startblock = be32_to_cpu(rec->refc.rc_startblock);
 101         irec->rc_blockcount = be32_to_cpu(rec->refc.rc_blockcount);
 102         irec->rc_refcount = be32_to_cpu(rec->refc.rc_refcount);
 103 }
 104
 105 /*
 106  * Get the data from the pointed-to record.
 107  */
 108 int
 109 xfs_refcount_get_rec(
 110         struct xfs_btree_cur            *cur,
 111         struct xfs_refcount_irec        *irec,
 112         int                             *stat)
 113 {
 114         struct xfs_mount                *mp = cur->bc_mp;
 115         xfs_agnumber_t                  agno = cur->bc_private.a.agno;
 116         union xfs_btree_rec             *rec;
 117         int                             error;
 118         xfs_agblock_t                   realstart;
 119
 120         error = xfs_btree_get_rec(cur, &rec, stat);
 121         if (error || !*stat)
 122                 return error;
 123
 124         xfs_refcount_btrec_to_irec(rec, irec);
 125
 126         agno = cur->bc_private.a.agno;
 127         if (irec->rc_blockcount == 0 || irec->rc_blockcount > MAXREFCEXTLEN)
 128                 goto out_bad_rec;
 129
 130         /* handle special COW-staging state */
 131         realstart = irec->rc_startblock;
 132         if (realstart & XFS_REFC_COW_START) {
 133                 if (irec->rc_refcount != 1)
 134                         goto out_bad_rec;
 135                 realstart &= ~XFS_REFC_COW_START;
 136         } else if (irec->rc_refcount < 2) {
 137                 goto out_bad_rec;
 138         }
 139
 140         /* check for valid extent range, including overflow */
 141         if (!xfs_verify_agbno(mp, agno, realstart))
 142                 goto out_bad_rec;
 143         if (realstart > realstart + irec->rc_blockcount)
 144                 goto out_bad_rec;
 145         if (!xfs_verify_agbno(mp, agno, realstart + irec->rc_blockcount - 1))
 146                 goto out_bad_rec;
 147
 148         if (irec->rc_refcount == 0 || irec->rc_refcount > MAXREFCOUNT)
 149                 goto out_bad_rec;
 150
 151         trace_xfs_refcount_get(cur->bc_mp, cur->bc_private.a.agno, irec);
 152         return 0;
 153
 154 out_bad_rec:
 155         xfs_warn(mp,
 156                 "Refcount BTree record corruption in AG %d detected!", agno);
 157         xfs_warn(mp,
 158                 "Start block 0x%x, block count 0x%x, references 0x%x",
 159                 irec->rc_startblock, irec->rc_blockcount, irec->rc_refcount);
 160         return -EFSCORRUPTED;
 161 }
 162
 163 /*
 164  * Update the record referred to by cur to the value given
 165  * by [bno, len, refcount].
 166  * This either works (return 0) or gets an EFSCORRUPTED error.
 167  */
 168 STATIC int
 169 xfs_refcount_update(
 170         struct xfs_btree_cur            *cur,
 171         struct xfs_refcount_irec        *irec)
 172 {
 173         union xfs_btree_rec     rec;
 174         int                     error;
 175
 176         trace_xfs_refcount_update(cur->bc_mp, cur->bc_private.a.agno, irec);
 177         rec.refc.rc_startblock = cpu_to_be32(irec->rc_startblock);
 178         rec.refc.rc_blockcount = cpu_to_be32(irec->rc_blockcount);
 179         rec.refc.rc_refcount = cpu_to_be32(irec->rc_refcount);
 180         error = xfs_btree_update(cur, &rec);
 181         if (error)
 182                 trace_xfs_refcount_update_error(cur->bc_mp,
 183                                 cur->bc_private.a.agno, error, _RET_IP_);
 184         return error;
 185 }
 186
 187 /*
 188  * Insert the record referred to by cur to the value given
 189  * by [bno, len, refcount].
 190  * This either works (return 0) or gets an EFSCORRUPTED error.
 191  */
 192 int
 193 xfs_refcount_insert(
 194         struct xfs_btree_cur            *cur,
 195         struct xfs_refcount_irec        *irec,
 196         int                             *i)
 197 {
 198         int                             error;
 199
 200         trace_xfs_refcount_insert(cur->bc_mp, cur->bc_private.a.agno, irec);
 201         cur->bc_rec.rc.rc_startblock = irec->rc_startblock;
 202         cur->bc_rec.rc.rc_blockcount = irec->rc_blockcount;
 203         cur->bc_rec.rc.rc_refcount = irec->rc_refcount;
 204         error = xfs_btree_insert(cur, i);
 205         if (error)
 206                 goto out_error;
 207         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, *i == 1, out_error);
 208
 209 out_error:
 210         if (error)
 211                 trace_xfs_refcount_insert_error(cur->bc_mp,
 212                                 cur->bc_private.a.agno, error, _RET_IP_);
 213         return error;
 214 }
 215
 216 /*
 217  * Remove the record referred to by cur, then set the pointer to the spot
 218  * where the record could be re-inserted, in case we want to increment or
 219  * decrement the cursor.
 220  * This either works (return 0) or gets an EFSCORRUPTED error.
 221  */
 222 STATIC int
 223 xfs_refcount_delete(
 224         struct xfs_btree_cur    *cur,
 225         int                     *i)
 226 {
 227         struct xfs_refcount_irec        irec;
 228         int                     found_rec;
 229         int                     error;
 230
 231         error = xfs_refcount_get_rec(cur, &irec, &found_rec);
 232         if (error)
 233                 goto out_error;
 234         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 235         trace_xfs_refcount_delete(cur->bc_mp, cur->bc_private.a.agno, &irec);
 236         error = xfs_btree_delete(cur, i);
 237         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, *i == 1, out_error);
 238         if (error)
 239                 goto out_error;
 240         error = xfs_refcount_lookup_ge(cur, irec.rc_startblock, &found_rec);
 241 out_error:
 242         if (error)
 243                 trace_xfs_refcount_delete_error(cur->bc_mp,
 244                                 cur->bc_private.a.agno, error, _RET_IP_);
 245         return error;
 246 }
 247
 248 /*
 249  * Adjusting the Reference Count
 250  *
 251  * As stated elsewhere, the reference count btree (refcbt) stores
 252  * >1 reference counts for extents of physical blocks.  In this
 253  * operation, we're either raising or lowering the reference count of
 254  * some subrange stored in the tree:
 255  *
 256  *      <------ adjustment range ------>
 257  * ----+   +---+-----+ +--+--------+---------
 258  *  2  |   | 3 |  4  | |17|   55   |   10
 259  * ----+   +---+-----+ +--+--------+---------
 260  * X axis is physical blocks number;
 261  * reference counts are the numbers inside the rectangles
 262  *
 263  * The first thing we need to do is to ensure that there are no
 264  * refcount extents crossing either boundary of the range to be
 265  * adjusted.  For any extent that does cross a boundary, split it into
 266  * two extents so that we can increment the refcount of one of the
 267  * pieces later:
 268  *
 269  *      <------ adjustment range ------>
 270  * ----+   +---+-----+ +--+--------+----+----
 271  *  2  |   | 3 |  2  | |17|   55   | 10 | 10
 272  * ----+   +---+-----+ +--+--------+----+----
 273  *
 274  * For this next step, let's assume that all the physical blocks in
 275  * the adjustment range are mapped to a file and are therefore in use
 276  * at least once.  Therefore, we can infer that any gap in the
 277  * refcount tree within the adjustment range represents a physical
 278  * extent with refcount == 1:
 279  *
 280  *      <------ adjustment range ------>
 281  * ----+---+---+-----+-+--+--------+----+----
 282  *  2  |"1"| 3 |  2  |1|17|   55   | 10 | 10
 283  * ----+---+---+-----+-+--+--------+----+----
 284  *      ^
 285  *
 286  * For each extent that falls within the interval range, figure out
 287  * which extent is to the left or the right of that extent.  Now we
 288  * have a left, current, and right extent.  If the new reference count
 289  * of the center extent enables us to merge left, center, and right
 290  * into one record covering all three, do so.  If the center extent is
 291  * at the left end of the range, abuts the left extent, and its new
 292  * reference count matches the left extent's record, then merge them.
 293  * If the center extent is at the right end of the range, abuts the
 294  * right extent, and the reference counts match, merge those.  In the
 295  * example, we can left merge (assuming an increment operation):
 296  *
 297  *      <------ adjustment range ------>
 298  * --------+---+-----+-+--+--------+----+----
 299  *    2    | 3 |  2  |1|17|   55   | 10 | 10
 300  * --------+---+-----+-+--+--------+----+----
 301  *          ^
 302  *
 303  * For all other extents within the range, adjust the reference count
 304  * or delete it if the refcount falls below 2.  If we were
 305  * incrementing, the end result looks like this:
 306  *
 307  *      <------ adjustment range ------>
 308  * --------+---+-----+-+--+--------+----+----
 309  *    2    | 4 |  3  |2|18|   56   | 11 | 10
 310  * --------+---+-----+-+--+--------+----+----
 311  *
 312  * The result of a decrement operation looks as such:
 313  *
 314  *      <------ adjustment range ------>
 315  * ----+   +---+       +--+--------+----+----
 316  *  2  |   | 2 |       |16|   54   |  9 | 10
 317  * ----+   +---+       +--+--------+----+----
 318  *      DDDD    111111DD
 319  *
 320  * The blocks marked "D" are freed; the blocks marked "1" are only
 321  * referenced once and therefore the record is removed from the
 322  * refcount btree.
 323  */
 324
 325 /* Next block after this extent. */
 326 static inline xfs_agblock_t
 327 xfs_refc_next(
 328         struct xfs_refcount_irec        *rc)
 329 {
 330         return rc->rc_startblock + rc->rc_blockcount;
 331 }
 332
 333 /*
 334  * Split a refcount extent that crosses agbno.
 335  */
 336 STATIC int
 337 xfs_refcount_split_extent(
 338         struct xfs_btree_cur            *cur,
 339         xfs_agblock_t                   agbno,
 340         bool                            *shape_changed)
 341 {
 342         struct xfs_refcount_irec        rcext, tmp;
 343         int                             found_rec;
 344         int                             error;
 345
 346         *shape_changed = false;
 347         error = xfs_refcount_lookup_le(cur, agbno, &found_rec);
 348         if (error)
 349                 goto out_error;
 350         if (!found_rec)
 351                 return 0;
 352
 353         error = xfs_refcount_get_rec(cur, &rcext, &found_rec);
 354         if (error)
 355                 goto out_error;
 356         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 357         if (rcext.rc_startblock == agbno || xfs_refc_next(&rcext) <= agbno)
 358                 return 0;
 359
 360         *shape_changed = true;
 361         trace_xfs_refcount_split_extent(cur->bc_mp, cur->bc_private.a.agno,
 362                         &rcext, agbno);
 363
 364         /* Establish the right extent. */
 365         tmp = rcext;
 366         tmp.rc_startblock = agbno;
 367         tmp.rc_blockcount -= (agbno - rcext.rc_startblock);
 368         error = xfs_refcount_update(cur, &tmp);
 369         if (error)
 370                 goto out_error;
 371
 372         /* Insert the left extent. */
 373         tmp = rcext;
 374         tmp.rc_blockcount = agbno - rcext.rc_startblock;
 375         error = xfs_refcount_insert(cur, &tmp, &found_rec);
 376         if (error)
 377                 goto out_error;
 378         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 379         return error;
 380
 381 out_error:
 382         trace_xfs_refcount_split_extent_error(cur->bc_mp,
 383                         cur->bc_private.a.agno, error, _RET_IP_);
 384         return error;
 385 }
 386
 387 /*
 388  * Merge the left, center, and right extents.
 389  */
 390 STATIC int
 391 xfs_refcount_merge_center_extents(
 392         struct xfs_btree_cur            *cur,
 393         struct xfs_refcount_irec        *left,
 394         struct xfs_refcount_irec        *center,
 395         struct xfs_refcount_irec        *right,
 396         unsigned long long              extlen,
 397         xfs_extlen_t                    *aglen)
 398 {
 399         int                             error;
 400         int                             found_rec;
 401
 402         trace_xfs_refcount_merge_center_extents(cur->bc_mp,
 403                         cur->bc_private.a.agno, left, center, right);
 404
 405         /*
 406          * Make sure the center and right extents are not in the btree.
 407          * If the center extent was synthesized, the first delete call
 408          * removes the right extent and we skip the second deletion.
 409          * If center and right were in the btree, then the first delete
 410          * call removes the center and the second one removes the right
 411          * extent.
 412          */
 413         error = xfs_refcount_lookup_ge(cur, center->rc_startblock,
 414                         &found_rec);
 415         if (error)
 416                 goto out_error;
 417         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 418
 419         error = xfs_refcount_delete(cur, &found_rec);
 420         if (error)
 421                 goto out_error;
 422         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 423
 424         if (center->rc_refcount > 1) {
 425                 error = xfs_refcount_delete(cur, &found_rec);
 426                 if (error)
 427                         goto out_error;
 428                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 429                                 out_error);
 430         }
 431
 432         /* Enlarge the left extent. */
 433         error = xfs_refcount_lookup_le(cur, left->rc_startblock,
 434                         &found_rec);
 435         if (error)
 436                 goto out_error;
 437         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 438
 439         left->rc_blockcount = extlen;
 440         error = xfs_refcount_update(cur, left);
 441         if (error)
 442                 goto out_error;
 443
 444         *aglen = 0;
 445         return error;
 446
 447 out_error:
 448         trace_xfs_refcount_merge_center_extents_error(cur->bc_mp,
 449                         cur->bc_private.a.agno, error, _RET_IP_);
 450         return error;
 451 }
 452
 453 /*
 454  * Merge with the left extent.
 455  */
 456 STATIC int
 457 xfs_refcount_merge_left_extent(
 458         struct xfs_btree_cur            *cur,
 459         struct xfs_refcount_irec        *left,
 460         struct xfs_refcount_irec        *cleft,
 461         xfs_agblock_t                   *agbno,
 462         xfs_extlen_t                    *aglen)
 463 {
 464         int                             error;
 465         int                             found_rec;
 466
 467         trace_xfs_refcount_merge_left_extent(cur->bc_mp,
 468                         cur->bc_private.a.agno, left, cleft);
 469
 470         /* If the extent at agbno (cleft) wasn't synthesized, remove it. */
 471         if (cleft->rc_refcount > 1) {
 472                 error = xfs_refcount_lookup_le(cur, cleft->rc_startblock,
 473                                 &found_rec);
 474                 if (error)
 475                         goto out_error;
 476                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 477                                 out_error);
 478
 479                 error = xfs_refcount_delete(cur, &found_rec);
 480                 if (error)
 481                         goto out_error;
 482                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 483                                 out_error);
 484         }
 485
 486         /* Enlarge the left extent. */
 487         error = xfs_refcount_lookup_le(cur, left->rc_startblock,
 488                         &found_rec);
 489         if (error)
 490                 goto out_error;
 491         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 492
 493         left->rc_blockcount += cleft->rc_blockcount;
 494         error = xfs_refcount_update(cur, left);
 495         if (error)
 496                 goto out_error;
 497
 498         *agbno += cleft->rc_blockcount;
 499         *aglen -= cleft->rc_blockcount;
 500         return error;
 501
 502 out_error:
 503         trace_xfs_refcount_merge_left_extent_error(cur->bc_mp,
 504                         cur->bc_private.a.agno, error, _RET_IP_);
 505         return error;
 506 }
 507
 508 /*
 509  * Merge with the right extent.
 510  */
 511 STATIC int
 512 xfs_refcount_merge_right_extent(
 513         struct xfs_btree_cur            *cur,
 514         struct xfs_refcount_irec        *right,
 515         struct xfs_refcount_irec        *cright,
 516         xfs_extlen_t                    *aglen)
 517 {
 518         int                             error;
 519         int                             found_rec;
 520
 521         trace_xfs_refcount_merge_right_extent(cur->bc_mp,
 522                         cur->bc_private.a.agno, cright, right);
 523
 524         /*
 525          * If the extent ending at agbno+aglen (cright) wasn't synthesized,
 526          * remove it.
 527          */
 528         if (cright->rc_refcount > 1) {
 529                 error = xfs_refcount_lookup_le(cur, cright->rc_startblock,
 530                         &found_rec);
 531                 if (error)
 532                         goto out_error;
 533                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 534                                 out_error);
 535
 536                 error = xfs_refcount_delete(cur, &found_rec);
 537                 if (error)
 538                         goto out_error;
 539                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 540                                 out_error);
 541         }
 542
 543         /* Enlarge the right extent. */
 544         error = xfs_refcount_lookup_le(cur, right->rc_startblock,
 545                         &found_rec);
 546         if (error)
 547                 goto out_error;
 548         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 549
 550         right->rc_startblock -= cright->rc_blockcount;
 551         right->rc_blockcount += cright->rc_blockcount;
 552         error = xfs_refcount_update(cur, right);
 553         if (error)
 554                 goto out_error;
 555
 556         *aglen -= cright->rc_blockcount;
 557         return error;
 558
 559 out_error:
 560         trace_xfs_refcount_merge_right_extent_error(cur->bc_mp,
 561                         cur->bc_private.a.agno, error, _RET_IP_);
 562         return error;
 563 }
 564
 565 #define XFS_FIND_RCEXT_SHARED   1
 566 #define XFS_FIND_RCEXT_COW      2
 567 /*
 568  * Find the left extent and the one after it (cleft).  This function assumes
 569  * that we've already split any extent crossing agbno.
 570  */
 571 STATIC int
 572 xfs_refcount_find_left_extents(
 573         struct xfs_btree_cur            *cur,
 574         struct xfs_refcount_irec        *left,
 575         struct xfs_refcount_irec        *cleft,
 576         xfs_agblock_t                   agbno,
 577         xfs_extlen_t                    aglen,
 578         int                             flags)
 579 {
 580         struct xfs_refcount_irec        tmp;
 581         int                             error;
 582         int                             found_rec;
 583
 584         left->rc_startblock = cleft->rc_startblock = NULLAGBLOCK;
 585         error = xfs_refcount_lookup_le(cur, agbno - 1, &found_rec);
 586         if (error)
 587                 goto out_error;
 588         if (!found_rec)
 589                 return 0;
 590
 591         error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
 592         if (error)
 593                 goto out_error;
 594         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 595
 596         if (xfs_refc_next(&tmp) != agbno)
 597                 return 0;
 598         if ((flags & XFS_FIND_RCEXT_SHARED) && tmp.rc_refcount < 2)
 599                 return 0;
 600         if ((flags & XFS_FIND_RCEXT_COW) && tmp.rc_refcount > 1)
 601                 return 0;
 602         /* We have a left extent; retrieve (or invent) the next right one */
 603         *left = tmp;
 604
 605         error = xfs_btree_increment(cur, 0, &found_rec);
 606         if (error)
 607                 goto out_error;
 608         if (found_rec) {
 609                 error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
 610                 if (error)
 611                         goto out_error;
 612                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 613                                 out_error);
 614
 615                 /* if tmp starts at the end of our range, just use that */
 616                 if (tmp.rc_startblock == agbno)
 617                         *cleft = tmp;
 618                 else {
 619                         /*
 620                          * There's a gap in the refcntbt at the start of the
 621                          * range we're interested in (refcount == 1) so
 622                          * synthesize the implied extent and pass it back.
 623                          * We assume here that the agbno/aglen range was
 624                          * passed in from a data fork extent mapping and
 625                          * therefore is allocated to exactly one owner.
 626                          */
 627                         cleft->rc_startblock = agbno;
 628                         cleft->rc_blockcount = min(aglen,
 629                                         tmp.rc_startblock - agbno);
 630                         cleft->rc_refcount = 1;
 631                 }
 632         } else {
 633                 /*
 634                  * No extents, so pretend that there's one covering the whole
 635                  * range.
 636                  */
 637                 cleft->rc_startblock = agbno;
 638                 cleft->rc_blockcount = aglen;
 639                 cleft->rc_refcount = 1;
 640         }
 641         trace_xfs_refcount_find_left_extent(cur->bc_mp, cur->bc_private.a.agno,
 642                         left, cleft, agbno);
 643         return error;
 644
 645 out_error:
 646         trace_xfs_refcount_find_left_extent_error(cur->bc_mp,
 647                         cur->bc_private.a.agno, error, _RET_IP_);
 648         return error;
 649 }
 650
 651 /*
 652  * Find the right extent and the one before it (cright).  This function
 653  * assumes that we've already split any extents crossing agbno + aglen.
 654  */
 655 STATIC int
 656 xfs_refcount_find_right_extents(
 657         struct xfs_btree_cur            *cur,
 658         struct xfs_refcount_irec        *right,
 659         struct xfs_refcount_irec        *cright,
 660         xfs_agblock_t                   agbno,
 661         xfs_extlen_t                    aglen,
 662         int                             flags)
 663 {
 664         struct xfs_refcount_irec        tmp;
 665         int                             error;
 666         int                             found_rec;
 667
 668         right->rc_startblock = cright->rc_startblock = NULLAGBLOCK;
 669         error = xfs_refcount_lookup_ge(cur, agbno + aglen, &found_rec);
 670         if (error)
 671                 goto out_error;
 672         if (!found_rec)
 673                 return 0;
 674
 675         error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
 676         if (error)
 677                 goto out_error;
 678         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 679
 680         if (tmp.rc_startblock != agbno + aglen)
 681                 return 0;
 682         if ((flags & XFS_FIND_RCEXT_SHARED) && tmp.rc_refcount < 2)
 683                 return 0;
 684         if ((flags & XFS_FIND_RCEXT_COW) && tmp.rc_refcount > 1)
 685                 return 0;
 686         /* We have a right extent; retrieve (or invent) the next left one */
 687         *right = tmp;
 688
 689         error = xfs_btree_decrement(cur, 0, &found_rec);
 690         if (error)
 691                 goto out_error;
 692         if (found_rec) {
 693                 error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
 694                 if (error)
 695                         goto out_error;
 696                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 697                                 out_error);
 698
 699                 /* if tmp ends at the end of our range, just use that */
 700                 if (xfs_refc_next(&tmp) == agbno + aglen)
 701                         *cright = tmp;
 702                 else {
 703                         /*
 704                          * There's a gap in the refcntbt at the end of the
 705                          * range we're interested in (refcount == 1) so
 706                          * create the implied extent and pass it back.
 707                          * We assume here that the agbno/aglen range was
 708                          * passed in from a data fork extent mapping and
 709                          * therefore is allocated to exactly one owner.
 710                          */
 711                         cright->rc_startblock = max(agbno, xfs_refc_next(&tmp));
 712                         cright->rc_blockcount = right->rc_startblock -
 713                                         cright->rc_startblock;
 714                         cright->rc_refcount = 1;
 715                 }
 716         } else {
 717                 /*
 718                  * No extents, so pretend that there's one covering the whole
 719                  * range.
 720                  */
 721                 cright->rc_startblock = agbno;
 722                 cright->rc_blockcount = aglen;
 723                 cright->rc_refcount = 1;
 724         }
 725         trace_xfs_refcount_find_right_extent(cur->bc_mp, cur->bc_private.a.agno,
 726                         cright, right, agbno + aglen);
 727         return error;
 728
 729 out_error:
 730         trace_xfs_refcount_find_right_extent_error(cur->bc_mp,
 731                         cur->bc_private.a.agno, error, _RET_IP_);
 732         return error;
 733 }
 734
 735 /* Is this extent valid? */
 736 static inline bool
 737 xfs_refc_valid(
 738         struct xfs_refcount_irec        *rc)
 739 {
 740         return rc->rc_startblock != NULLAGBLOCK;
 741 }
 742
 743 /*
 744  * Try to merge with any extents on the boundaries of the adjustment range.
 745  */
 746 STATIC int
 747 xfs_refcount_merge_extents(
 748         struct xfs_btree_cur    *cur,
 749         xfs_agblock_t           *agbno,
 750         xfs_extlen_t            *aglen,
 751         enum xfs_refc_adjust_op adjust,
 752         int                     flags,
 753         bool                    *shape_changed)
 754 {
 755         struct xfs_refcount_irec        left = {0}, cleft = {0};
 756         struct xfs_refcount_irec        cright = {0}, right = {0};
 757         int                             error;
 758         unsigned long long              ulen;
 759         bool                            cequal;
 760
 761         *shape_changed = false;
 762         /*
 763          * Find the extent just below agbno [left], just above agbno [cleft],
 764          * just below (agbno + aglen) [cright], and just above (agbno + aglen)
 765          * [right].
 766          */
 767         error = xfs_refcount_find_left_extents(cur, &left, &cleft, *agbno,
 768                         *aglen, flags);
 769         if (error)
 770                 return error;
 771         error = xfs_refcount_find_right_extents(cur, &right, &cright, *agbno,
 772                         *aglen, flags);
 773         if (error)
 774                 return error;
 775
 776         /* No left or right extent to merge; exit. */
 777         if (!xfs_refc_valid(&left) && !xfs_refc_valid(&right))
 778                 return 0;
 779
 780         cequal = (cleft.rc_startblock == cright.rc_startblock) &&
 781                  (cleft.rc_blockcount == cright.rc_blockcount);
 782
 783         /* Try to merge left, cleft, and right.  cleft must == cright. */
 784         ulen = (unsigned long long)left.rc_blockcount + cleft.rc_blockcount +
 785                         right.rc_blockcount;
 786         if (xfs_refc_valid(&left) && xfs_refc_valid(&right) &&
 787             xfs_refc_valid(&cleft) && xfs_refc_valid(&cright) && cequal &&
 788             left.rc_refcount == cleft.rc_refcount + adjust &&
 789             right.rc_refcount == cleft.rc_refcount + adjust &&
 790             ulen < MAXREFCEXTLEN) {
 791                 *shape_changed = true;
 792                 return xfs_refcount_merge_center_extents(cur, &left, &cleft,
 793                                 &right, ulen, aglen);
 794         }
 795
 796         /* Try to merge left and cleft. */
 797         ulen = (unsigned long long)left.rc_blockcount + cleft.rc_blockcount;
 798         if (xfs_refc_valid(&left) && xfs_refc_valid(&cleft) &&
 799             left.rc_refcount == cleft.rc_refcount + adjust &&
 800             ulen < MAXREFCEXTLEN) {
 801                 *shape_changed = true;
 802                 error = xfs_refcount_merge_left_extent(cur, &left, &cleft,
 803                                 agbno, aglen);
 804                 if (error)
 805                         return error;
 806
 807                 /*
 808                  * If we just merged left + cleft and cleft == cright,
 809                  * we no longer have a cright to merge with right.  We're done.
 810                  */
 811                 if (cequal)
 812                         return 0;
 813         }
 814
 815         /* Try to merge cright and right. */
 816         ulen = (unsigned long long)right.rc_blockcount + cright.rc_blockcount;
 817         if (xfs_refc_valid(&right) && xfs_refc_valid(&cright) &&
 818             right.rc_refcount == cright.rc_refcount + adjust &&
 819             ulen < MAXREFCEXTLEN) {
 820                 *shape_changed = true;
 821                 return xfs_refcount_merge_right_extent(cur, &right, &cright,
 822                                 aglen);
 823         }
 824
 825         return error;
 826 }
 827
 828 /*
 829  * XXX: This is a pretty hand-wavy estimate.  The penalty for guessing
 830  * true incorrectly is a shutdown FS; the penalty for guessing false
 831  * incorrectly is more transaction rolls than might be necessary.
 832  * Be conservative here.
 833  */
 834 static bool
 835 xfs_refcount_still_have_space(
 836         struct xfs_btree_cur            *cur)
 837 {
 838         unsigned long                   overhead;
 839
 840         overhead = cur->bc_private.a.priv.refc.shape_changes *
 841                         xfs_allocfree_log_count(cur->bc_mp, 1);
 842         overhead *= cur->bc_mp->m_sb.sb_blocksize;
 843
 844         /*
 845          * Only allow 2 refcount extent updates per transaction if the
 846          * refcount continue update "error" has been injected.
 847          */
 848         if (cur->bc_private.a.priv.refc.nr_ops > 2 &&
 849             XFS_TEST_ERROR(false, cur->bc_mp,
 850                         XFS_ERRTAG_REFCOUNT_CONTINUE_UPDATE))
 851                 return false;
 852
 853         if (cur->bc_private.a.priv.refc.nr_ops == 0)
 854                 return true;
 855         else if (overhead > cur->bc_tp->t_log_res)
 856                 return false;
 857         return  cur->bc_tp->t_log_res - overhead >
 858                 cur->bc_private.a.priv.refc.nr_ops * XFS_REFCOUNT_ITEM_OVERHEAD;
 859 }
 860
 861 /*
 862  * Adjust the refcounts of middle extents.  At this point we should have
 863  * split extents that crossed the adjustment range; merged with adjacent
 864  * extents; and updated agbno/aglen to reflect the merges.  Therefore,
 865  * all we have to do is update the extents inside [agbno, agbno + aglen].
 866  */
 867 STATIC int
 868 xfs_refcount_adjust_extents(
 869         struct xfs_btree_cur    *cur,
 870         xfs_agblock_t           *agbno,
 871         xfs_extlen_t            *aglen,
 872         enum xfs_refc_adjust_op adj,
 873         struct xfs_defer_ops    *dfops,
 874         struct xfs_owner_info   *oinfo)
 875 {
 876         struct xfs_refcount_irec        ext, tmp;
 877         int                             error;
 878         int                             found_rec, found_tmp;
 879         xfs_fsblock_t                   fsbno;
 880
 881         /* Merging did all the work already. */
 882         if (*aglen == 0)
 883                 return 0;
 884
 885         error = xfs_refcount_lookup_ge(cur, *agbno, &found_rec);
 886         if (error)
 887                 goto out_error;
 888
 889         while (*aglen > 0 && xfs_refcount_still_have_space(cur)) {
 890                 error = xfs_refcount_get_rec(cur, &ext, &found_rec);
 891                 if (error)
 892                         goto out_error;
 893                 if (!found_rec) {
 894                         ext.rc_startblock = cur->bc_mp->m_sb.sb_agblocks;
 895                         ext.rc_blockcount = 0;
 896                         ext.rc_refcount = 0;
 897                 }
 898
 899                 /*
 900                  * Deal with a hole in the refcount tree; if a file maps to
 901                  * these blocks and there's no refcountbt record, pretend that
 902                  * there is one with refcount == 1.
 903                  */
 904                 if (ext.rc_startblock != *agbno) {
 905                         tmp.rc_startblock = *agbno;
 906                         tmp.rc_blockcount = min(*aglen,
 907                                         ext.rc_startblock - *agbno);
 908                         tmp.rc_refcount = 1 + adj;
 909                         trace_xfs_refcount_modify_extent(cur->bc_mp,
 910                                         cur->bc_private.a.agno, &tmp);
 911
 912                         /*
 913                          * Either cover the hole (increment) or
 914                          * delete the range (decrement).
 915                          */
 916                         if (tmp.rc_refcount) {
 917                                 error = xfs_refcount_insert(cur, &tmp,
 918                                                 &found_tmp);
 919                                 if (error)
 920                                         goto out_error;
 921                                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
 922                                                 found_tmp == 1, out_error);
 923                                 cur->bc_private.a.priv.refc.nr_ops++;
 924                         } else {
 925                                 fsbno = XFS_AGB_TO_FSB(cur->bc_mp,
 926                                                 cur->bc_private.a.agno,
 927                                                 tmp.rc_startblock);
 928                                 xfs_bmap_add_free(cur->bc_mp, dfops, fsbno,
 929                                                 tmp.rc_blockcount, oinfo);
 930                         }
 931
 932                         (*agbno) += tmp.rc_blockcount;
 933                         (*aglen) -= tmp.rc_blockcount;
 934
 935                         error = xfs_refcount_lookup_ge(cur, *agbno,
 936                                         &found_rec);
 937                         if (error)
 938                                 goto out_error;
 939                 }
 940
 941                 /* Stop if there's nothing left to modify */
 942                 if (*aglen == 0 || !xfs_refcount_still_have_space(cur))
 943                         break;
 944
 945                 /*
 946                  * Adjust the reference count and either update the tree
 947                  * (incr) or free the blocks (decr).
 948                  */
 949                 if (ext.rc_refcount == MAXREFCOUNT)
 950                         goto skip;
 951                 ext.rc_refcount += adj;
 952                 trace_xfs_refcount_modify_extent(cur->bc_mp,
 953                                 cur->bc_private.a.agno, &ext);
 954                 if (ext.rc_refcount > 1) {
 955                         error = xfs_refcount_update(cur, &ext);
 956                         if (error)
 957                                 goto out_error;
 958                         cur->bc_private.a.priv.refc.nr_ops++;
 959                 } else if (ext.rc_refcount == 1) {
 960                         error = xfs_refcount_delete(cur, &found_rec);
 961                         if (error)
 962                                 goto out_error;
 963                         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
 964                                         found_rec == 1, out_error);
 965                         cur->bc_private.a.priv.refc.nr_ops++;
 966                         goto advloop;
 967                 } else {
 968                         fsbno = XFS_AGB_TO_FSB(cur->bc_mp,
 969                                         cur->bc_private.a.agno,
 970                                         ext.rc_startblock);
 971                         xfs_bmap_add_free(cur->bc_mp, dfops, fsbno,
 972                                         ext.rc_blockcount, oinfo);
 973                 }
 974
 975 skip:
 976                 error = xfs_btree_increment(cur, 0, &found_rec);
 977                 if (error)
 978                         goto out_error;
 979
 980 advloop:
 981                 (*agbno) += ext.rc_blockcount;
 982                 (*aglen) -= ext.rc_blockcount;
 983         }
 984
 985         return error;
 986 out_error:
 987         trace_xfs_refcount_modify_extent_error(cur->bc_mp,
 988                         cur->bc_private.a.agno, error, _RET_IP_);
 989         return error;
 990 }
 991
 992 /* Adjust the reference count of a range of AG blocks. */
 993 STATIC int
 994 xfs_refcount_adjust(
 995         struct xfs_btree_cur    *cur,
 996         xfs_agblock_t           agbno,
 997         xfs_extlen_t            aglen,
 998         xfs_agblock_t           *new_agbno,
 999         xfs_extlen_t            *new_aglen,
1000         enum xfs_refc_adjust_op adj,
1001         struct xfs_defer_ops    *dfops,
1002         struct xfs_owner_info   *oinfo)
1003 {
1004         bool                    shape_changed;
1005         int                     shape_changes = 0;
1006         int                     error;
1007
1008         *new_agbno = agbno;
1009         *new_aglen = aglen;
1010         if (adj == XFS_REFCOUNT_ADJUST_INCREASE)
1011                 trace_xfs_refcount_increase(cur->bc_mp, cur->bc_private.a.agno,
1012                                 agbno, aglen);
1013         else
1014                 trace_xfs_refcount_decrease(cur->bc_mp, cur->bc_private.a.agno,
1015                                 agbno, aglen);
1016
1017         /*
1018          * Ensure that no rcextents cross the boundary of the adjustment range.
1019          */
1020         error = xfs_refcount_split_extent(cur, agbno, &shape_changed);
1021         if (error)
1022                 goto out_error;
1023         if (shape_changed)
1024                 shape_changes++;
1025
1026         error = xfs_refcount_split_extent(cur, agbno + aglen, &shape_changed);
1027         if (error)
1028                 goto out_error;
1029         if (shape_changed)
1030                 shape_changes++;
1031
1032         /*
1033          * Try to merge with the left or right extents of the range.
1034          */
1035         error = xfs_refcount_merge_extents(cur, new_agbno, new_aglen, adj,
1036                         XFS_FIND_RCEXT_SHARED, &shape_changed);
1037         if (error)
1038                 goto out_error;
1039         if (shape_changed)
1040                 shape_changes++;
1041         if (shape_changes)
1042                 cur->bc_private.a.priv.refc.shape_changes++;
1043
1044         /* Now that we've taken care of the ends, adjust the middle extents */
1045         error = xfs_refcount_adjust_extents(cur, new_agbno, new_aglen,
1046                         adj, dfops, oinfo);
1047         if (error)
1048                 goto out_error;
1049
1050         return 0;
1051
1052 out_error:
1053         trace_xfs_refcount_adjust_error(cur->bc_mp, cur->bc_private.a.agno,
1054                         error, _RET_IP_);
1055         return error;
1056 }
1057
1058 /* Clean up after calling xfs_refcount_finish_one. */
1059 void
1060 xfs_refcount_finish_one_cleanup(
1061         struct xfs_trans        *tp,
1062         struct xfs_btree_cur    *rcur,
1063         int                     error)
1064 {
1065         struct xfs_buf          *agbp;
1066
1067         if (rcur == NULL)
1068                 return;
1069         agbp = rcur->bc_private.a.agbp;
1070         xfs_btree_del_cursor(rcur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
1071         if (error)
1072                 xfs_trans_brelse(tp, agbp);
1073 }
1074
1075 /*
1076  * Process one of the deferred refcount operations.  We pass back the
1077  * btree cursor to maintain our lock on the btree between calls.
1078  * This saves time and eliminates a buffer deadlock between the
1079  * superblock and the AGF because we'll always grab them in the same
1080  * order.
1081  */
1082 int
1083 xfs_refcount_finish_one(
1084         struct xfs_trans                *tp,
1085         struct xfs_defer_ops            *dfops,
1086         enum xfs_refcount_intent_type   type,
1087         xfs_fsblock_t                   startblock,
1088         xfs_extlen_t                    blockcount,
1089         xfs_fsblock_t                   *new_fsb,
1090         xfs_extlen_t                    *new_len,
1091         struct xfs_btree_cur            **pcur)
1092 {
1093         struct xfs_mount                *mp = tp->t_mountp;
1094         struct xfs_btree_cur            *rcur;
1095         struct xfs_buf                  *agbp = NULL;
1096         int                             error = 0;
1097         xfs_agnumber_t                  agno;
1098         xfs_agblock_t                   bno;
1099         xfs_agblock_t                   new_agbno;
1100         unsigned long                   nr_ops = 0;
1101         int                             shape_changes = 0;
1102
1103         agno = XFS_FSB_TO_AGNO(mp, startblock);
1104         ASSERT(agno != NULLAGNUMBER);
1105         bno = XFS_FSB_TO_AGBNO(mp, startblock);
1106
1107         trace_xfs_refcount_deferred(mp, XFS_FSB_TO_AGNO(mp, startblock),
1108                         type, XFS_FSB_TO_AGBNO(mp, startblock),
1109                         blockcount);
1110
1111         if (XFS_TEST_ERROR(false, mp,
1112                         XFS_ERRTAG_REFCOUNT_FINISH_ONE))
1113                 return -EIO;
1114
1115         /*
1116          * If we haven't gotten a cursor or the cursor AG doesn't match
1117          * the startblock, get one now.
1118          */
1119         rcur = *pcur;
1120         if (rcur != NULL && rcur->bc_private.a.agno != agno) {
1121                 nr_ops = rcur->bc_private.a.priv.refc.nr_ops;
1122                 shape_changes = rcur->bc_private.a.priv.refc.shape_changes;
1123                 xfs_refcount_finish_one_cleanup(tp, rcur, 0);
1124                 rcur = NULL;
1125                 *pcur = NULL;
1126         }
1127         if (rcur == NULL) {
1128                 error = xfs_alloc_read_agf(tp->t_mountp, tp, agno,
1129                                 XFS_ALLOC_FLAG_FREEING, &agbp);
1130                 if (error)
1131                         return error;
1132                 if (!agbp)
1133                         return -EFSCORRUPTED;
1134
1135                 rcur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno, dfops);
1136                 if (!rcur) {
1137                         error = -ENOMEM;
1138                         goto out_cur;
1139                 }
1140                 rcur->bc_private.a.priv.refc.nr_ops = nr_ops;
1141                 rcur->bc_private.a.priv.refc.shape_changes = shape_changes;
1142         }
1143         *pcur = rcur;
1144
1145         switch (type) {
1146         case XFS_REFCOUNT_INCREASE:
1147                 error = xfs_refcount_adjust(rcur, bno, blockcount, &new_agbno,
1148                         new_len, XFS_REFCOUNT_ADJUST_INCREASE, dfops, NULL);
1149                 *new_fsb = XFS_AGB_TO_FSB(mp, agno, new_agbno);
1150                 break;
1151         case XFS_REFCOUNT_DECREASE:
1152                 error = xfs_refcount_adjust(rcur, bno, blockcount, &new_agbno,
1153                         new_len, XFS_REFCOUNT_ADJUST_DECREASE, dfops, NULL);
1154                 *new_fsb = XFS_AGB_TO_FSB(mp, agno, new_agbno);
1155                 break;
1156         case XFS_REFCOUNT_ALLOC_COW:
1157                 *new_fsb = startblock + blockcount;
1158                 *new_len = 0;
1159                 error = __xfs_refcount_cow_alloc(rcur, bno, blockcount, dfops);
1160                 break;
1161         case XFS_REFCOUNT_FREE_COW:
1162                 *new_fsb = startblock + blockcount;
1163                 *new_len = 0;
1164                 error = __xfs_refcount_cow_free(rcur, bno, blockcount, dfops);
1165                 break;
1166         default:
1167                 ASSERT(0);
1168                 error = -EFSCORRUPTED;
1169         }
1170         if (!error && *new_len > 0)
1171                 trace_xfs_refcount_finish_one_leftover(mp, agno, type,
1172                                 bno, blockcount, new_agbno, *new_len);
1173         return error;
1174
1175 out_cur:
1176         xfs_trans_brelse(tp, agbp);
1177
1178         return error;
1179 }
1180
1181 /*
1182  * Record a refcount intent for later processing.
1183  */
1184 static int
1185 __xfs_refcount_add(
1186         struct xfs_mount                *mp,
1187         struct xfs_defer_ops            *dfops,
1188         enum xfs_refcount_intent_type   type,
1189         xfs_fsblock_t                   startblock,
1190         xfs_extlen_t                    blockcount)
1191 {
1192         struct xfs_refcount_intent      *ri;
1193
1194         trace_xfs_refcount_defer(mp, XFS_FSB_TO_AGNO(mp, startblock),
1195                         type, XFS_FSB_TO_AGBNO(mp, startblock),
1196                         blockcount);
1197
1198         ri = kmem_alloc(sizeof(struct xfs_refcount_intent),
1199                         KM_SLEEP | KM_NOFS);
1200         INIT_LIST_HEAD(&ri->ri_list);
1201         ri->ri_type = type;
1202         ri->ri_startblock = startblock;
1203         ri->ri_blockcount = blockcount;
1204
1205         xfs_defer_add(dfops, XFS_DEFER_OPS_TYPE_REFCOUNT, &ri->ri_list);
1206         return 0;
1207 }
1208
1209 /*
1210  * Increase the reference count of the blocks backing a file's extent.
1211  */
1212 int
1213 xfs_refcount_increase_extent(
1214         struct xfs_mount                *mp,
1215         struct xfs_defer_ops            *dfops,
1216         struct xfs_bmbt_irec            *PREV)
1217 {
1218         if (!xfs_sb_version_hasreflink(&mp->m_sb))
1219                 return 0;
1220
1221         return __xfs_refcount_add(mp, dfops, XFS_REFCOUNT_INCREASE,
1222                         PREV->br_startblock, PREV->br_blockcount);
1223 }
1224
1225 /*
1226  * Decrease the reference count of the blocks backing a file's extent.
1227  */
1228 int
1229 xfs_refcount_decrease_extent(
1230         struct xfs_mount                *mp,
1231         struct xfs_defer_ops            *dfops,
1232         struct xfs_bmbt_irec            *PREV)
1233 {
1234         if (!xfs_sb_version_hasreflink(&mp->m_sb))
1235                 return 0;
1236
1237         return __xfs_refcount_add(mp, dfops, XFS_REFCOUNT_DECREASE,
1238                         PREV->br_startblock, PREV->br_blockcount);
1239 }
1240
1241 /*
1242  * Given an AG extent, find the lowest-numbered run of shared blocks
1243  * within that range and return the range in fbno/flen.  If
1244  * find_end_of_shared is set, return the longest contiguous extent of
1245  * shared blocks; if not, just return the first extent we find.  If no
1246  * shared blocks are found, fbno and flen will be set to NULLAGBLOCK
1247  * and 0, respectively.
1248  */
1249 int
1250 xfs_refcount_find_shared(
1251         struct xfs_btree_cur            *cur,
1252         xfs_agblock_t                   agbno,
1253         xfs_extlen_t                    aglen,
1254         xfs_agblock_t                   *fbno,
1255         xfs_extlen_t                    *flen,
1256         bool                            find_end_of_shared)
1257 {
1258         struct xfs_refcount_irec        tmp;
1259         int                             i;
1260         int                             have;
1261         int                             error;
1262
1263         trace_xfs_refcount_find_shared(cur->bc_mp, cur->bc_private.a.agno,
1264                         agbno, aglen);
1265
1266         /* By default, skip the whole range */
1267         *fbno = NULLAGBLOCK;
1268         *flen = 0;
1269
1270         /* Try to find a refcount extent that crosses the start */
1271         error = xfs_refcount_lookup_le(cur, agbno, &have);
1272         if (error)
1273                 goto out_error;
1274         if (!have) {
1275                 /* No left extent, look at the next one */
1276                 error = xfs_btree_increment(cur, 0, &have);
1277                 if (error)
1278                         goto out_error;
1279                 if (!have)
1280                         goto done;
1281         }
1282         error = xfs_refcount_get_rec(cur, &tmp, &i);
1283         if (error)
1284                 goto out_error;
1285         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, i == 1, out_error);
1286
1287         /* If the extent ends before the start, look at the next one */
1288         if (tmp.rc_startblock + tmp.rc_blockcount <= agbno) {
1289                 error = xfs_btree_increment(cur, 0, &have);
1290                 if (error)
1291                         goto out_error;
1292                 if (!have)
1293                         goto done;
1294                 error = xfs_refcount_get_rec(cur, &tmp, &i);
1295                 if (error)
1296                         goto out_error;
1297                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, i == 1, out_error);
1298         }
1299
1300         /* If the extent starts after the range we want, bail out */
1301         if (tmp.rc_startblock >= agbno + aglen)
1302                 goto done;
1303
1304         /* We found the start of a shared extent! */
1305         if (tmp.rc_startblock < agbno) {
1306                 tmp.rc_blockcount -= (agbno - tmp.rc_startblock);
1307                 tmp.rc_startblock = agbno;
1308         }
1309
1310         *fbno = tmp.rc_startblock;
1311         *flen = min(tmp.rc_blockcount, agbno + aglen - *fbno);
1312         if (!find_end_of_shared)
1313                 goto done;
1314
1315         /* Otherwise, find the end of this shared extent */
1316         while (*fbno + *flen < agbno + aglen) {
1317                 error = xfs_btree_increment(cur, 0, &have);
1318                 if (error)
1319                         goto out_error;
1320                 if (!have)
1321                         break;
1322                 error = xfs_refcount_get_rec(cur, &tmp, &i);
1323                 if (error)
1324                         goto out_error;
1325                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, i == 1, out_error);
1326                 if (tmp.rc_startblock >= agbno + aglen ||
1327                     tmp.rc_startblock != *fbno + *flen)
1328                         break;
1329                 *flen = min(*flen + tmp.rc_blockcount, agbno + aglen - *fbno);
1330         }
1331
1332 done:
1333         trace_xfs_refcount_find_shared_result(cur->bc_mp,
1334                         cur->bc_private.a.agno, *fbno, *flen);
1335
1336 out_error:
1337         if (error)
1338                 trace_xfs_refcount_find_shared_error(cur->bc_mp,
1339                                 cur->bc_private.a.agno, error, _RET_IP_);
1340         return error;
1341 }
1342
1343 /*
1344  * Recovering CoW Blocks After a Crash
1345  *
1346  * Due to the way that the copy on write mechanism works, there's a window of
1347  * opportunity in which we can lose track of allocated blocks during a crash.
1348  * Because CoW uses delayed allocation in the in-core CoW fork, writeback
1349  * causes blocks to be allocated and stored in the CoW fork.  The blocks are
1350  * no longer in the free space btree but are not otherwise recorded anywhere
1351  * until the write completes and the blocks are mapped into the file.  A crash
1352  * in between allocation and remapping results in the replacement blocks being
1353  * lost.  This situation is exacerbated by the CoW extent size hint because
1354  * allocations can hang around for long time.
1355  *
1356  * However, there is a place where we can record these allocations before they
1357  * become mappings -- the reference count btree.  The btree does not record
1358  * extents with refcount == 1, so we can record allocations with a refcount of
1359  * 1.  Blocks being used for CoW writeout cannot be shared, so there should be
1360  * no conflict with shared block records.  These mappings should be created
1361  * when we allocate blocks to the CoW fork and deleted when they're removed
1362  * from the CoW fork.
1363  *
1364  * Minor nit: records for in-progress CoW allocations and records for shared
1365  * extents must never be merged, to preserve the property that (except for CoW
1366  * allocations) there are no refcount btree entries with refcount == 1.  The
1367  * only time this could potentially happen is when unsharing a block that's
1368  * adjacent to CoW allocations, so we must be careful to avoid this.
1369  *
1370  * At mount time we recover lost CoW allocations by searching the refcount
1371  * btree for these refcount == 1 mappings.  These represent CoW allocations
1372  * that were in progress at the time the filesystem went down, so we can free
1373  * them to get the space back.
1374  *
1375  * This mechanism is superior to creating EFIs for unmapped CoW extents for
1376  * several reasons -- first, EFIs pin the tail of the log and would have to be
1377  * periodically relogged to avoid filling up the log.  Second, CoW completions
1378  * will have to file an EFD and create new EFIs for whatever remains in the
1379  * CoW fork; this partially takes care of (1) but extent-size reservations
1380  * will have to periodically relog even if there's no writeout in progress.
1381  * This can happen if the CoW extent size hint is set, which you really want.
1382  * Third, EFIs cannot currently be automatically relogged into newer
1383  * transactions to advance the log tail.  Fourth, stuffing the log full of
1384  * EFIs places an upper bound on the number of CoW allocations that can be
1385  * held filesystem-wide at any given time.  Recording them in the refcount
1386  * btree doesn't require us to maintain any state in memory and doesn't pin
1387  * the log.
1388  */
1389 /*
1390  * Adjust the refcounts of CoW allocations.  These allocations are "magic"
1391  * in that they're not referenced anywhere else in the filesystem, so we
1392  * stash them in the refcount btree with a refcount of 1 until either file
1393  * remapping (or CoW cancellation) happens.
1394  */
1395 STATIC int
1396 xfs_refcount_adjust_cow_extents(
1397         struct xfs_btree_cur    *cur,
1398         xfs_agblock_t           agbno,
1399         xfs_extlen_t            aglen,
1400         enum xfs_refc_adjust_op adj)
1401 {
1402         struct xfs_refcount_irec        ext, tmp;
1403         int                             error;
1404         int                             found_rec, found_tmp;
1405
1406         if (aglen == 0)
1407                 return 0;
1408
1409         /* Find any overlapping refcount records */
1410         error = xfs_refcount_lookup_ge(cur, agbno, &found_rec);
1411         if (error)
1412                 goto out_error;
1413         error = xfs_refcount_get_rec(cur, &ext, &found_rec);
1414         if (error)
1415                 goto out_error;
1416         if (!found_rec) {
1417                 ext.rc_startblock = cur->bc_mp->m_sb.sb_agblocks +
1418                                 XFS_REFC_COW_START;
1419                 ext.rc_blockcount = 0;
1420                 ext.rc_refcount = 0;
1421         }
1422
1423         switch (adj) {
1424         case XFS_REFCOUNT_ADJUST_COW_ALLOC:
1425                 /* Adding a CoW reservation, there should be nothing here. */
1426                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1427                                 ext.rc_startblock >= agbno + aglen, out_error);
1428
1429                 tmp.rc_startblock = agbno;
1430                 tmp.rc_blockcount = aglen;
1431                 tmp.rc_refcount = 1;
1432                 trace_xfs_refcount_modify_extent(cur->bc_mp,
1433                                 cur->bc_private.a.agno, &tmp);
1434
1435                 error = xfs_refcount_insert(cur, &tmp,
1436                                 &found_tmp);
1437                 if (error)
1438                         goto out_error;
1439                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1440                                 found_tmp == 1, out_error);
1441                 break;
1442         case XFS_REFCOUNT_ADJUST_COW_FREE:
1443                 /* Removing a CoW reservation, there should be one extent. */
1444                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1445                         ext.rc_startblock == agbno, out_error);
1446                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1447                         ext.rc_blockcount == aglen, out_error);
1448                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1449                         ext.rc_refcount == 1, out_error);
1450
1451                 ext.rc_refcount = 0;
1452                 trace_xfs_refcount_modify_extent(cur->bc_mp,
1453                                 cur->bc_private.a.agno, &ext);
1454                 error = xfs_refcount_delete(cur, &found_rec);
1455                 if (error)
1456                         goto out_error;
1457                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1458                                 found_rec == 1, out_error);
1459                 break;
1460         default:
1461                 ASSERT(0);
1462         }
1463
1464         return error;
1465 out_error:
1466         trace_xfs_refcount_modify_extent_error(cur->bc_mp,
1467                         cur->bc_private.a.agno, error, _RET_IP_);
1468         return error;
1469 }
1470
1471 /*
1472  * Add or remove refcount btree entries for CoW reservations.
1473  */
1474 STATIC int
1475 xfs_refcount_adjust_cow(
1476         struct xfs_btree_cur    *cur,
1477         xfs_agblock_t           agbno,
1478         xfs_extlen_t            aglen,
1479         enum xfs_refc_adjust_op adj)
1480 {
1481         bool                    shape_changed;
1482         int                     error;
1483
1484         agbno += XFS_REFC_COW_START;
1485
1486         /*
1487          * Ensure that no rcextents cross the boundary of the adjustment range.
1488          */
1489         error = xfs_refcount_split_extent(cur, agbno, &shape_changed);
1490         if (error)
1491                 goto out_error;
1492
1493         error = xfs_refcount_split_extent(cur, agbno + aglen, &shape_changed);
1494         if (error)
1495                 goto out_error;
1496
1497         /*
1498          * Try to merge with the left or right extents of the range.
1499          */
1500         error = xfs_refcount_merge_extents(cur, &agbno, &aglen, adj,
1501                         XFS_FIND_RCEXT_COW, &shape_changed);
1502         if (error)
1503                 goto out_error;
1504
1505         /* Now that we've taken care of the ends, adjust the middle extents */
1506         error = xfs_refcount_adjust_cow_extents(cur, agbno, aglen, adj);
1507         if (error)
1508                 goto out_error;
1509
1510         return 0;
1511
1512 out_error:
1513         trace_xfs_refcount_adjust_cow_error(cur->bc_mp, cur->bc_private.a.agno,
1514                         error, _RET_IP_);
1515         return error;
1516 }
1517
1518 /*
1519  * Record a CoW allocation in the refcount btree.
1520  */
1521 STATIC int
1522 __xfs_refcount_cow_alloc(
1523         struct xfs_btree_cur    *rcur,
1524         xfs_agblock_t           agbno,
1525         xfs_extlen_t            aglen,
1526         struct xfs_defer_ops    *dfops)
1527 {
1528         trace_xfs_refcount_cow_increase(rcur->bc_mp, rcur->bc_private.a.agno,
1529                         agbno, aglen);
1530
1531         /* Add refcount btree reservation */
1532         return xfs_refcount_adjust_cow(rcur, agbno, aglen,
1533                         XFS_REFCOUNT_ADJUST_COW_ALLOC);
1534 }
1535
1536 /*
1537  * Remove a CoW allocation from the refcount btree.
1538  */
1539 STATIC int
1540 __xfs_refcount_cow_free(
1541         struct xfs_btree_cur    *rcur,
1542         xfs_agblock_t           agbno,
1543         xfs_extlen_t            aglen,
1544         struct xfs_defer_ops    *dfops)
1545 {
1546         trace_xfs_refcount_cow_decrease(rcur->bc_mp, rcur->bc_private.a.agno,
1547                         agbno, aglen);
1548
1549         /* Remove refcount btree reservation */
1550         return xfs_refcount_adjust_cow(rcur, agbno, aglen,
1551                         XFS_REFCOUNT_ADJUST_COW_FREE);
1552 }
1553
1554 /* Record a CoW staging extent in the refcount btree. */
1555 int
1556 xfs_refcount_alloc_cow_extent(
1557         struct xfs_mount                *mp,
1558         struct xfs_defer_ops            *dfops,
1559         xfs_fsblock_t                   fsb,
1560         xfs_extlen_t                    len)
1561 {
1562         int                             error;
1563
1564         if (!xfs_sb_version_hasreflink(&mp->m_sb))
1565                 return 0;
1566
1567         error = __xfs_refcount_add(mp, dfops, XFS_REFCOUNT_ALLOC_COW,
1568                         fsb, len);
1569         if (error)
1570                 return error;
1571
1572         /* Add rmap entry */
1573         return xfs_rmap_alloc_extent(mp, dfops, XFS_FSB_TO_AGNO(mp, fsb),
1574                         XFS_FSB_TO_AGBNO(mp, fsb), len, XFS_RMAP_OWN_COW);
1575 }
1576
1577 /* Forget a CoW staging event in the refcount btree. */
1578 int
1579 xfs_refcount_free_cow_extent(
1580         struct xfs_mount                *mp,
1581         struct xfs_defer_ops            *dfops,
1582         xfs_fsblock_t                   fsb,
1583         xfs_extlen_t                    len)
1584 {
1585         int                             error;
1586
1587         if (!xfs_sb_version_hasreflink(&mp->m_sb))
1588                 return 0;
1589
1590         /* Remove rmap entry */
1591         error = xfs_rmap_free_extent(mp, dfops, XFS_FSB_TO_AGNO(mp, fsb),
1592                         XFS_FSB_TO_AGBNO(mp, fsb), len, XFS_RMAP_OWN_COW);
1593         if (error)
1594                 return error;
1595
1596         return __xfs_refcount_add(mp, dfops, XFS_REFCOUNT_FREE_COW,
1597                         fsb, len);
1598 }
1599
1600 struct xfs_refcount_recovery {
1601         struct list_head                rr_list;
1602         struct xfs_refcount_irec        rr_rrec;
1603 };
1604
1605 /* Stuff an extent on the recovery list. */
1606 STATIC int
1607 xfs_refcount_recover_extent(
1608         struct xfs_btree_cur            *cur,
1609         union xfs_btree_rec             *rec,
1610         void                            *priv)
1611 {
1612         struct list_head                *debris = priv;
1613         struct xfs_refcount_recovery    *rr;
1614
1615         if (be32_to_cpu(rec->refc.rc_refcount) != 1)
1616                 return -EFSCORRUPTED;
1617
1618         rr = kmem_alloc(sizeof(struct xfs_refcount_recovery), KM_SLEEP);
1619         xfs_refcount_btrec_to_irec(rec, &rr->rr_rrec);
1620         list_add_tail(&rr->rr_list, debris);
1621
1622         return 0;
1623 }
1624
1625 /* Find and remove leftover CoW reservations. */
1626 int
1627 xfs_refcount_recover_cow_leftovers(
1628         struct xfs_mount                *mp,
1629         xfs_agnumber_t                  agno)
1630 {
1631         struct xfs_trans                *tp;
1632         struct xfs_btree_cur            *cur;
1633         struct xfs_buf                  *agbp;
1634         struct xfs_refcount_recovery    *rr, *n;
1635         struct list_head                debris;
1636         union xfs_btree_irec            low;
1637         union xfs_btree_irec            high;
1638         struct xfs_defer_ops            dfops;
1639         xfs_fsblock_t                   fsb;
1640         xfs_agblock_t                   agbno;
1641         int                             error;
1642
1643         if (mp->m_sb.sb_agblocks >= XFS_REFC_COW_START)
1644                 return -EOPNOTSUPP;
1645
1646         INIT_LIST_HEAD(&debris);
1647
1648         /*
1649          * In this first part, we use an empty transaction to gather up
1650          * all the leftover CoW extents so that we can subsequently
1651          * delete them.  The empty transaction is used to avoid
1652          * a buffer lock deadlock if there happens to be a loop in the
1653          * refcountbt because we're allowed to re-grab a buffer that is
1654          * already attached to our transaction.  When we're done
1655          * recording the CoW debris we cancel the (empty) transaction
1656          * and everything goes away cleanly.
1657          */
1658         error = xfs_trans_alloc_empty(mp, &tp);
1659         if (error)
1660                 return error;
1661
1662         error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
1663         if (error)
1664                 goto out_trans;
1665         if (!agbp) {
1666                 error = -ENOMEM;
1667                 goto out_trans;
1668         }
1669         cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno, NULL);
1670
1671         /* Find all the leftover CoW staging extents. */
1672         memset(&low, 0, sizeof(low));
1673         memset(&high, 0, sizeof(high));
1674         low.rc.rc_startblock = XFS_REFC_COW_START;
1675         high.rc.rc_startblock = -1U;
1676         error = xfs_btree_query_range(cur, &low, &high,
1677                         xfs_refcount_recover_extent, &debris);
1678         if (error)
1679                 goto out_cursor;
1680         xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1681         xfs_trans_brelse(tp, agbp);
1682         xfs_trans_cancel(tp);
1683
1684         /* Now iterate the list to free the leftovers */
1685         list_for_each_entry_safe(rr, n, &debris, rr_list) {
1686                 /* Set up transaction. */
1687                 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1688                 if (error)
1689                         goto out_free;
1690
1691                 trace_xfs_refcount_recover_extent(mp, agno, &rr->rr_rrec);
1692
1693                 /* Free the orphan record */
1694                 xfs_defer_init(&dfops, &fsb);
1695                 agbno = rr->rr_rrec.rc_startblock - XFS_REFC_COW_START;
1696                 fsb = XFS_AGB_TO_FSB(mp, agno, agbno);
1697                 error = xfs_refcount_free_cow_extent(mp, &dfops, fsb,
1698                                 rr->rr_rrec.rc_blockcount);
1699                 if (error)
1700                         goto out_defer;
1701
1702                 /* Free the block. */
1703                 xfs_bmap_add_free(mp, &dfops, fsb,
1704                                 rr->rr_rrec.rc_blockcount, NULL);
1705
1706                 error = xfs_defer_finish(&tp, &dfops);
1707                 if (error)
1708                         goto out_defer;
1709
1710                 error = xfs_trans_commit(tp);
1711                 if (error)
1712                         goto out_free;
1713
1714                 list_del(&rr->rr_list);
1715                 kmem_free(rr);
1716         }
1717
1718         return error;
1719 out_defer:
1720         xfs_defer_cancel(&dfops);
1721 out_trans:
1722         xfs_trans_cancel(tp);
1723 out_free:
1724         /* Free the leftover list */
1725         list_for_each_entry_safe(rr, n, &debris, rr_list) {
1726                 list_del(&rr->rr_list);
1727                 kmem_free(rr);
1728         }
1729         return error;
1730
1731 out_cursor:
1732         xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1733         xfs_trans_brelse(tp, agbp);
1734         goto out_trans;
1735 }
1736
1737 /* Is there a record covering a given extent? */
1738 int
1739 xfs_refcount_has_record(
1740         struct xfs_btree_cur    *cur,
1741         xfs_agblock_t           bno,
1742         xfs_extlen_t            len,
1743         bool                    *exists)
1744 {
1745         union xfs_btree_irec    low;
1746         union xfs_btree_irec    high;
1747
1748         memset(&low, 0, sizeof(low));
1749         low.rc.rc_startblock = bno;
1750         memset(&high, 0xFF, sizeof(high));
1751         high.rc.rc_startblock = bno + len - 1;
1752
1753         return xfs_btree_has_record(cur, &low, &high, exists);
1754 }