fs/xfs/libxfs/xfs_refcount.c

   1 /*
   2  * Copyright (C) 2016 Oracle.  All Rights Reserved.
   3  *
   4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version 2
   9  * of the License, or (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it would be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write the Free Software Foundation,
  18  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301, USA.
  19  */
  20 #include "xfs.h"
  21 #include "xfs_fs.h"
  22 #include "xfs_shared.h"
  23 #include "xfs_format.h"
  24 #include "xfs_log_format.h"
  25 #include "xfs_trans_resv.h"
  26 #include "xfs_sb.h"
  27 #include "xfs_mount.h"
  28 #include "xfs_defer.h"
  29 #include "xfs_btree.h"
  30 #include "xfs_bmap.h"
  31 #include "xfs_refcount_btree.h"
  32 #include "xfs_alloc.h"
  33 #include "xfs_errortag.h"
  34 #include "xfs_error.h"
  35 #include "xfs_trace.h"
  36 #include "xfs_cksum.h"
  37 #include "xfs_trans.h"
  38 #include "xfs_bit.h"
  39 #include "xfs_refcount.h"
  40 #include "xfs_rmap.h"
  41
  42 /* Allowable refcount adjustment amounts. */
  43 enum xfs_refc_adjust_op {
  44         XFS_REFCOUNT_ADJUST_INCREASE    = 1,
  45         XFS_REFCOUNT_ADJUST_DECREASE    = -1,
  46         XFS_REFCOUNT_ADJUST_COW_ALLOC   = 0,
  47         XFS_REFCOUNT_ADJUST_COW_FREE    = -1,
  48 };
  49
  50 STATIC int __xfs_refcount_cow_alloc(struct xfs_btree_cur *rcur,
  51                 xfs_agblock_t agbno, xfs_extlen_t aglen,
  52                 struct xfs_defer_ops *dfops);
  53 STATIC int __xfs_refcount_cow_free(struct xfs_btree_cur *rcur,
  54                 xfs_agblock_t agbno, xfs_extlen_t aglen,
  55                 struct xfs_defer_ops *dfops);
  56
  57 /*
  58  * Look up the first record less than or equal to [bno, len] in the btree
  59  * given by cur.
  60  */
  61 int
  62 xfs_refcount_lookup_le(
  63         struct xfs_btree_cur    *cur,
  64         xfs_agblock_t           bno,
  65         int                     *stat)
  66 {
  67         trace_xfs_refcount_lookup(cur->bc_mp, cur->bc_private.a.agno, bno,
  68                         XFS_LOOKUP_LE);
  69         cur->bc_rec.rc.rc_startblock = bno;
  70         cur->bc_rec.rc.rc_blockcount = 0;
  71         return xfs_btree_lookup(cur, XFS_LOOKUP_LE, stat);
  72 }
  73
  74 /*
  75  * Look up the first record greater than or equal to [bno, len] in the btree
  76  * given by cur.
  77  */
  78 int
  79 xfs_refcount_lookup_ge(
  80         struct xfs_btree_cur    *cur,
  81         xfs_agblock_t           bno,
  82         int                     *stat)
  83 {
  84         trace_xfs_refcount_lookup(cur->bc_mp, cur->bc_private.a.agno, bno,
  85                         XFS_LOOKUP_GE);
  86         cur->bc_rec.rc.rc_startblock = bno;
  87         cur->bc_rec.rc.rc_blockcount = 0;
  88         return xfs_btree_lookup(cur, XFS_LOOKUP_GE, stat);
  89 }
  90
  91 /*
  92  * Look up the first record equal to [bno, len] in the btree
  93  * given by cur.
  94  */
  95 int
  96 xfs_refcount_lookup_eq(
  97         struct xfs_btree_cur    *cur,
  98         xfs_agblock_t           bno,
  99         int                     *stat)
 100 {
 101         trace_xfs_refcount_lookup(cur->bc_mp, cur->bc_private.a.agno, bno,
 102                         XFS_LOOKUP_LE);
 103         cur->bc_rec.rc.rc_startblock = bno;
 104         cur->bc_rec.rc.rc_blockcount = 0;
 105         return xfs_btree_lookup(cur, XFS_LOOKUP_EQ, stat);
 106 }
 107
 108 /* Convert on-disk record to in-core format. */
 109 void
 110 xfs_refcount_btrec_to_irec(
 111         union xfs_btree_rec             *rec,
 112         struct xfs_refcount_irec        *irec)
 113 {
 114         irec->rc_startblock = be32_to_cpu(rec->refc.rc_startblock);
 115         irec->rc_blockcount = be32_to_cpu(rec->refc.rc_blockcount);
 116         irec->rc_refcount = be32_to_cpu(rec->refc.rc_refcount);
 117 }
 118
 119 /*
 120  * Get the data from the pointed-to record.
 121  */
 122 int
 123 xfs_refcount_get_rec(
 124         struct xfs_btree_cur            *cur,
 125         struct xfs_refcount_irec        *irec,
 126         int                             *stat)
 127 {
 128         struct xfs_mount                *mp = cur->bc_mp;
 129         xfs_agnumber_t                  agno = cur->bc_private.a.agno;
 130         union xfs_btree_rec             *rec;
 131         int                             error;
 132         xfs_agblock_t                   realstart;
 133
 134         error = xfs_btree_get_rec(cur, &rec, stat);
 135         if (error || !*stat)
 136                 return error;
 137
 138         xfs_refcount_btrec_to_irec(rec, irec);
 139
 140         agno = cur->bc_private.a.agno;
 141         if (irec->rc_blockcount == 0 || irec->rc_blockcount > MAXREFCEXTLEN)
 142                 goto out_bad_rec;
 143
 144         /* handle special COW-staging state */
 145         realstart = irec->rc_startblock;
 146         if (realstart & XFS_REFC_COW_START) {
 147                 if (irec->rc_refcount != 1)
 148                         goto out_bad_rec;
 149                 realstart &= ~XFS_REFC_COW_START;
 150         } else if (irec->rc_refcount < 2) {
 151                 goto out_bad_rec;
 152         }
 153
 154         /* check for valid extent range, including overflow */
 155         if (!xfs_verify_agbno(mp, agno, realstart))
 156                 goto out_bad_rec;
 157         if (realstart > realstart + irec->rc_blockcount)
 158                 goto out_bad_rec;
 159         if (!xfs_verify_agbno(mp, agno, realstart + irec->rc_blockcount - 1))
 160                 goto out_bad_rec;
 161
 162         if (irec->rc_refcount == 0 || irec->rc_refcount > MAXREFCOUNT)
 163                 goto out_bad_rec;
 164
 165         trace_xfs_refcount_get(cur->bc_mp, cur->bc_private.a.agno, irec);
 166         return 0;
 167
 168 out_bad_rec:
 169         xfs_warn(mp,
 170                 "Refcount BTree record corruption in AG %d detected!", agno);
 171         xfs_warn(mp,
 172                 "Start block 0x%x, block count 0x%x, references 0x%x",
 173                 irec->rc_startblock, irec->rc_blockcount, irec->rc_refcount);
 174         return -EFSCORRUPTED;
 175 }
 176
 177 /*
 178  * Update the record referred to by cur to the value given
 179  * by [bno, len, refcount].
 180  * This either works (return 0) or gets an EFSCORRUPTED error.
 181  */
 182 STATIC int
 183 xfs_refcount_update(
 184         struct xfs_btree_cur            *cur,
 185         struct xfs_refcount_irec        *irec)
 186 {
 187         union xfs_btree_rec     rec;
 188         int                     error;
 189
 190         trace_xfs_refcount_update(cur->bc_mp, cur->bc_private.a.agno, irec);
 191         rec.refc.rc_startblock = cpu_to_be32(irec->rc_startblock);
 192         rec.refc.rc_blockcount = cpu_to_be32(irec->rc_blockcount);
 193         rec.refc.rc_refcount = cpu_to_be32(irec->rc_refcount);
 194         error = xfs_btree_update(cur, &rec);
 195         if (error)
 196                 trace_xfs_refcount_update_error(cur->bc_mp,
 197                                 cur->bc_private.a.agno, error, _RET_IP_);
 198         return error;
 199 }
 200
 201 /*
 202  * Insert the record referred to by cur to the value given
 203  * by [bno, len, refcount].
 204  * This either works (return 0) or gets an EFSCORRUPTED error.
 205  */
 206 int
 207 xfs_refcount_insert(
 208         struct xfs_btree_cur            *cur,
 209         struct xfs_refcount_irec        *irec,
 210         int                             *i)
 211 {
 212         int                             error;
 213
 214         trace_xfs_refcount_insert(cur->bc_mp, cur->bc_private.a.agno, irec);
 215         cur->bc_rec.rc.rc_startblock = irec->rc_startblock;
 216         cur->bc_rec.rc.rc_blockcount = irec->rc_blockcount;
 217         cur->bc_rec.rc.rc_refcount = irec->rc_refcount;
 218         error = xfs_btree_insert(cur, i);
 219         if (error)
 220                 goto out_error;
 221         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, *i == 1, out_error);
 222
 223 out_error:
 224         if (error)
 225                 trace_xfs_refcount_insert_error(cur->bc_mp,
 226                                 cur->bc_private.a.agno, error, _RET_IP_);
 227         return error;
 228 }
 229
 230 /*
 231  * Remove the record referred to by cur, then set the pointer to the spot
 232  * where the record could be re-inserted, in case we want to increment or
 233  * decrement the cursor.
 234  * This either works (return 0) or gets an EFSCORRUPTED error.
 235  */
 236 STATIC int
 237 xfs_refcount_delete(
 238         struct xfs_btree_cur    *cur,
 239         int                     *i)
 240 {
 241         struct xfs_refcount_irec        irec;
 242         int                     found_rec;
 243         int                     error;
 244
 245         error = xfs_refcount_get_rec(cur, &irec, &found_rec);
 246         if (error)
 247                 goto out_error;
 248         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 249         trace_xfs_refcount_delete(cur->bc_mp, cur->bc_private.a.agno, &irec);
 250         error = xfs_btree_delete(cur, i);
 251         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, *i == 1, out_error);
 252         if (error)
 253                 goto out_error;
 254         error = xfs_refcount_lookup_ge(cur, irec.rc_startblock, &found_rec);
 255 out_error:
 256         if (error)
 257                 trace_xfs_refcount_delete_error(cur->bc_mp,
 258                                 cur->bc_private.a.agno, error, _RET_IP_);
 259         return error;
 260 }
 261
 262 /*
 263  * Adjusting the Reference Count
 264  *
 265  * As stated elsewhere, the reference count btree (refcbt) stores
 266  * >1 reference counts for extents of physical blocks.  In this
 267  * operation, we're either raising or lowering the reference count of
 268  * some subrange stored in the tree:
 269  *
 270  *      <------ adjustment range ------>
 271  * ----+   +---+-----+ +--+--------+---------
 272  *  2  |   | 3 |  4  | |17|   55   |   10
 273  * ----+   +---+-----+ +--+--------+---------
 274  * X axis is physical blocks number;
 275  * reference counts are the numbers inside the rectangles
 276  *
 277  * The first thing we need to do is to ensure that there are no
 278  * refcount extents crossing either boundary of the range to be
 279  * adjusted.  For any extent that does cross a boundary, split it into
 280  * two extents so that we can increment the refcount of one of the
 281  * pieces later:
 282  *
 283  *      <------ adjustment range ------>
 284  * ----+   +---+-----+ +--+--------+----+----
 285  *  2  |   | 3 |  2  | |17|   55   | 10 | 10
 286  * ----+   +---+-----+ +--+--------+----+----
 287  *
 288  * For this next step, let's assume that all the physical blocks in
 289  * the adjustment range are mapped to a file and are therefore in use
 290  * at least once.  Therefore, we can infer that any gap in the
 291  * refcount tree within the adjustment range represents a physical
 292  * extent with refcount == 1:
 293  *
 294  *      <------ adjustment range ------>
 295  * ----+---+---+-----+-+--+--------+----+----
 296  *  2  |"1"| 3 |  2  |1|17|   55   | 10 | 10
 297  * ----+---+---+-----+-+--+--------+----+----
 298  *      ^
 299  *
 300  * For each extent that falls within the interval range, figure out
 301  * which extent is to the left or the right of that extent.  Now we
 302  * have a left, current, and right extent.  If the new reference count
 303  * of the center extent enables us to merge left, center, and right
 304  * into one record covering all three, do so.  If the center extent is
 305  * at the left end of the range, abuts the left extent, and its new
 306  * reference count matches the left extent's record, then merge them.
 307  * If the center extent is at the right end of the range, abuts the
 308  * right extent, and the reference counts match, merge those.  In the
 309  * example, we can left merge (assuming an increment operation):
 310  *
 311  *      <------ adjustment range ------>
 312  * --------+---+-----+-+--+--------+----+----
 313  *    2    | 3 |  2  |1|17|   55   | 10 | 10
 314  * --------+---+-----+-+--+--------+----+----
 315  *          ^
 316  *
 317  * For all other extents within the range, adjust the reference count
 318  * or delete it if the refcount falls below 2.  If we were
 319  * incrementing, the end result looks like this:
 320  *
 321  *      <------ adjustment range ------>
 322  * --------+---+-----+-+--+--------+----+----
 323  *    2    | 4 |  3  |2|18|   56   | 11 | 10
 324  * --------+---+-----+-+--+--------+----+----
 325  *
 326  * The result of a decrement operation looks as such:
 327  *
 328  *      <------ adjustment range ------>
 329  * ----+   +---+       +--+--------+----+----
 330  *  2  |   | 2 |       |16|   54   |  9 | 10
 331  * ----+   +---+       +--+--------+----+----
 332  *      DDDD    111111DD
 333  *
 334  * The blocks marked "D" are freed; the blocks marked "1" are only
 335  * referenced once and therefore the record is removed from the
 336  * refcount btree.
 337  */
 338
 339 /* Next block after this extent. */
 340 static inline xfs_agblock_t
 341 xfs_refc_next(
 342         struct xfs_refcount_irec        *rc)
 343 {
 344         return rc->rc_startblock + rc->rc_blockcount;
 345 }
 346
 347 /*
 348  * Split a refcount extent that crosses agbno.
 349  */
 350 STATIC int
 351 xfs_refcount_split_extent(
 352         struct xfs_btree_cur            *cur,
 353         xfs_agblock_t                   agbno,
 354         bool                            *shape_changed)
 355 {
 356         struct xfs_refcount_irec        rcext, tmp;
 357         int                             found_rec;
 358         int                             error;
 359
 360         *shape_changed = false;
 361         error = xfs_refcount_lookup_le(cur, agbno, &found_rec);
 362         if (error)
 363                 goto out_error;
 364         if (!found_rec)
 365                 return 0;
 366
 367         error = xfs_refcount_get_rec(cur, &rcext, &found_rec);
 368         if (error)
 369                 goto out_error;
 370         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 371         if (rcext.rc_startblock == agbno || xfs_refc_next(&rcext) <= agbno)
 372                 return 0;
 373
 374         *shape_changed = true;
 375         trace_xfs_refcount_split_extent(cur->bc_mp, cur->bc_private.a.agno,
 376                         &rcext, agbno);
 377
 378         /* Establish the right extent. */
 379         tmp = rcext;
 380         tmp.rc_startblock = agbno;
 381         tmp.rc_blockcount -= (agbno - rcext.rc_startblock);
 382         error = xfs_refcount_update(cur, &tmp);
 383         if (error)
 384                 goto out_error;
 385
 386         /* Insert the left extent. */
 387         tmp = rcext;
 388         tmp.rc_blockcount = agbno - rcext.rc_startblock;
 389         error = xfs_refcount_insert(cur, &tmp, &found_rec);
 390         if (error)
 391                 goto out_error;
 392         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 393         return error;
 394
 395 out_error:
 396         trace_xfs_refcount_split_extent_error(cur->bc_mp,
 397                         cur->bc_private.a.agno, error, _RET_IP_);
 398         return error;
 399 }
 400
 401 /*
 402  * Merge the left, center, and right extents.
 403  */
 404 STATIC int
 405 xfs_refcount_merge_center_extents(
 406         struct xfs_btree_cur            *cur,
 407         struct xfs_refcount_irec        *left,
 408         struct xfs_refcount_irec        *center,
 409         struct xfs_refcount_irec        *right,
 410         unsigned long long              extlen,
 411         xfs_extlen_t                    *aglen)
 412 {
 413         int                             error;
 414         int                             found_rec;
 415
 416         trace_xfs_refcount_merge_center_extents(cur->bc_mp,
 417                         cur->bc_private.a.agno, left, center, right);
 418
 419         /*
 420          * Make sure the center and right extents are not in the btree.
 421          * If the center extent was synthesized, the first delete call
 422          * removes the right extent and we skip the second deletion.
 423          * If center and right were in the btree, then the first delete
 424          * call removes the center and the second one removes the right
 425          * extent.
 426          */
 427         error = xfs_refcount_lookup_ge(cur, center->rc_startblock,
 428                         &found_rec);
 429         if (error)
 430                 goto out_error;
 431         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 432
 433         error = xfs_refcount_delete(cur, &found_rec);
 434         if (error)
 435                 goto out_error;
 436         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 437
 438         if (center->rc_refcount > 1) {
 439                 error = xfs_refcount_delete(cur, &found_rec);
 440                 if (error)
 441                         goto out_error;
 442                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 443                                 out_error);
 444         }
 445
 446         /* Enlarge the left extent. */
 447         error = xfs_refcount_lookup_le(cur, left->rc_startblock,
 448                         &found_rec);
 449         if (error)
 450                 goto out_error;
 451         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 452
 453         left->rc_blockcount = extlen;
 454         error = xfs_refcount_update(cur, left);
 455         if (error)
 456                 goto out_error;
 457
 458         *aglen = 0;
 459         return error;
 460
 461 out_error:
 462         trace_xfs_refcount_merge_center_extents_error(cur->bc_mp,
 463                         cur->bc_private.a.agno, error, _RET_IP_);
 464         return error;
 465 }
 466
 467 /*
 468  * Merge with the left extent.
 469  */
 470 STATIC int
 471 xfs_refcount_merge_left_extent(
 472         struct xfs_btree_cur            *cur,
 473         struct xfs_refcount_irec        *left,
 474         struct xfs_refcount_irec        *cleft,
 475         xfs_agblock_t                   *agbno,
 476         xfs_extlen_t                    *aglen)
 477 {
 478         int                             error;
 479         int                             found_rec;
 480
 481         trace_xfs_refcount_merge_left_extent(cur->bc_mp,
 482                         cur->bc_private.a.agno, left, cleft);
 483
 484         /* If the extent at agbno (cleft) wasn't synthesized, remove it. */
 485         if (cleft->rc_refcount > 1) {
 486                 error = xfs_refcount_lookup_le(cur, cleft->rc_startblock,
 487                                 &found_rec);
 488                 if (error)
 489                         goto out_error;
 490                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 491                                 out_error);
 492
 493                 error = xfs_refcount_delete(cur, &found_rec);
 494                 if (error)
 495                         goto out_error;
 496                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 497                                 out_error);
 498         }
 499
 500         /* Enlarge the left extent. */
 501         error = xfs_refcount_lookup_le(cur, left->rc_startblock,
 502                         &found_rec);
 503         if (error)
 504                 goto out_error;
 505         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 506
 507         left->rc_blockcount += cleft->rc_blockcount;
 508         error = xfs_refcount_update(cur, left);
 509         if (error)
 510                 goto out_error;
 511
 512         *agbno += cleft->rc_blockcount;
 513         *aglen -= cleft->rc_blockcount;
 514         return error;
 515
 516 out_error:
 517         trace_xfs_refcount_merge_left_extent_error(cur->bc_mp,
 518                         cur->bc_private.a.agno, error, _RET_IP_);
 519         return error;
 520 }
 521
 522 /*
 523  * Merge with the right extent.
 524  */
 525 STATIC int
 526 xfs_refcount_merge_right_extent(
 527         struct xfs_btree_cur            *cur,
 528         struct xfs_refcount_irec        *right,
 529         struct xfs_refcount_irec        *cright,
 530         xfs_extlen_t                    *aglen)
 531 {
 532         int                             error;
 533         int                             found_rec;
 534
 535         trace_xfs_refcount_merge_right_extent(cur->bc_mp,
 536                         cur->bc_private.a.agno, cright, right);
 537
 538         /*
 539          * If the extent ending at agbno+aglen (cright) wasn't synthesized,
 540          * remove it.
 541          */
 542         if (cright->rc_refcount > 1) {
 543                 error = xfs_refcount_lookup_le(cur, cright->rc_startblock,
 544                         &found_rec);
 545                 if (error)
 546                         goto out_error;
 547                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 548                                 out_error);
 549
 550                 error = xfs_refcount_delete(cur, &found_rec);
 551                 if (error)
 552                         goto out_error;
 553                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 554                                 out_error);
 555         }
 556
 557         /* Enlarge the right extent. */
 558         error = xfs_refcount_lookup_le(cur, right->rc_startblock,
 559                         &found_rec);
 560         if (error)
 561                 goto out_error;
 562         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 563
 564         right->rc_startblock -= cright->rc_blockcount;
 565         right->rc_blockcount += cright->rc_blockcount;
 566         error = xfs_refcount_update(cur, right);
 567         if (error)
 568                 goto out_error;
 569
 570         *aglen -= cright->rc_blockcount;
 571         return error;
 572
 573 out_error:
 574         trace_xfs_refcount_merge_right_extent_error(cur->bc_mp,
 575                         cur->bc_private.a.agno, error, _RET_IP_);
 576         return error;
 577 }
 578
 579 #define XFS_FIND_RCEXT_SHARED   1
 580 #define XFS_FIND_RCEXT_COW      2
 581 /*
 582  * Find the left extent and the one after it (cleft).  This function assumes
 583  * that we've already split any extent crossing agbno.
 584  */
 585 STATIC int
 586 xfs_refcount_find_left_extents(
 587         struct xfs_btree_cur            *cur,
 588         struct xfs_refcount_irec        *left,
 589         struct xfs_refcount_irec        *cleft,
 590         xfs_agblock_t                   agbno,
 591         xfs_extlen_t                    aglen,
 592         int                             flags)
 593 {
 594         struct xfs_refcount_irec        tmp;
 595         int                             error;
 596         int                             found_rec;
 597
 598         left->rc_startblock = cleft->rc_startblock = NULLAGBLOCK;
 599         error = xfs_refcount_lookup_le(cur, agbno - 1, &found_rec);
 600         if (error)
 601                 goto out_error;
 602         if (!found_rec)
 603                 return 0;
 604
 605         error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
 606         if (error)
 607                 goto out_error;
 608         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 609
 610         if (xfs_refc_next(&tmp) != agbno)
 611                 return 0;
 612         if ((flags & XFS_FIND_RCEXT_SHARED) && tmp.rc_refcount < 2)
 613                 return 0;
 614         if ((flags & XFS_FIND_RCEXT_COW) && tmp.rc_refcount > 1)
 615                 return 0;
 616         /* We have a left extent; retrieve (or invent) the next right one */
 617         *left = tmp;
 618
 619         error = xfs_btree_increment(cur, 0, &found_rec);
 620         if (error)
 621                 goto out_error;
 622         if (found_rec) {
 623                 error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
 624                 if (error)
 625                         goto out_error;
 626                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 627                                 out_error);
 628
 629                 /* if tmp starts at the end of our range, just use that */
 630                 if (tmp.rc_startblock == agbno)
 631                         *cleft = tmp;
 632                 else {
 633                         /*
 634                          * There's a gap in the refcntbt at the start of the
 635                          * range we're interested in (refcount == 1) so
 636                          * synthesize the implied extent and pass it back.
 637                          * We assume here that the agbno/aglen range was
 638                          * passed in from a data fork extent mapping and
 639                          * therefore is allocated to exactly one owner.
 640                          */
 641                         cleft->rc_startblock = agbno;
 642                         cleft->rc_blockcount = min(aglen,
 643                                         tmp.rc_startblock - agbno);
 644                         cleft->rc_refcount = 1;
 645                 }
 646         } else {
 647                 /*
 648                  * No extents, so pretend that there's one covering the whole
 649                  * range.
 650                  */
 651                 cleft->rc_startblock = agbno;
 652                 cleft->rc_blockcount = aglen;
 653                 cleft->rc_refcount = 1;
 654         }
 655         trace_xfs_refcount_find_left_extent(cur->bc_mp, cur->bc_private.a.agno,
 656                         left, cleft, agbno);
 657         return error;
 658
 659 out_error:
 660         trace_xfs_refcount_find_left_extent_error(cur->bc_mp,
 661                         cur->bc_private.a.agno, error, _RET_IP_);
 662         return error;
 663 }
 664
 665 /*
 666  * Find the right extent and the one before it (cright).  This function
 667  * assumes that we've already split any extents crossing agbno + aglen.
 668  */
 669 STATIC int
 670 xfs_refcount_find_right_extents(
 671         struct xfs_btree_cur            *cur,
 672         struct xfs_refcount_irec        *right,
 673         struct xfs_refcount_irec        *cright,
 674         xfs_agblock_t                   agbno,
 675         xfs_extlen_t                    aglen,
 676         int                             flags)
 677 {
 678         struct xfs_refcount_irec        tmp;
 679         int                             error;
 680         int                             found_rec;
 681
 682         right->rc_startblock = cright->rc_startblock = NULLAGBLOCK;
 683         error = xfs_refcount_lookup_ge(cur, agbno + aglen, &found_rec);
 684         if (error)
 685                 goto out_error;
 686         if (!found_rec)
 687                 return 0;
 688
 689         error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
 690         if (error)
 691                 goto out_error;
 692         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1, out_error);
 693
 694         if (tmp.rc_startblock != agbno + aglen)
 695                 return 0;
 696         if ((flags & XFS_FIND_RCEXT_SHARED) && tmp.rc_refcount < 2)
 697                 return 0;
 698         if ((flags & XFS_FIND_RCEXT_COW) && tmp.rc_refcount > 1)
 699                 return 0;
 700         /* We have a right extent; retrieve (or invent) the next left one */
 701         *right = tmp;
 702
 703         error = xfs_btree_decrement(cur, 0, &found_rec);
 704         if (error)
 705                 goto out_error;
 706         if (found_rec) {
 707                 error = xfs_refcount_get_rec(cur, &tmp, &found_rec);
 708                 if (error)
 709                         goto out_error;
 710                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, found_rec == 1,
 711                                 out_error);
 712
 713                 /* if tmp ends at the end of our range, just use that */
 714                 if (xfs_refc_next(&tmp) == agbno + aglen)
 715                         *cright = tmp;
 716                 else {
 717                         /*
 718                          * There's a gap in the refcntbt at the end of the
 719                          * range we're interested in (refcount == 1) so
 720                          * create the implied extent and pass it back.
 721                          * We assume here that the agbno/aglen range was
 722                          * passed in from a data fork extent mapping and
 723                          * therefore is allocated to exactly one owner.
 724                          */
 725                         cright->rc_startblock = max(agbno, xfs_refc_next(&tmp));
 726                         cright->rc_blockcount = right->rc_startblock -
 727                                         cright->rc_startblock;
 728                         cright->rc_refcount = 1;
 729                 }
 730         } else {
 731                 /*
 732                  * No extents, so pretend that there's one covering the whole
 733                  * range.
 734                  */
 735                 cright->rc_startblock = agbno;
 736                 cright->rc_blockcount = aglen;
 737                 cright->rc_refcount = 1;
 738         }
 739         trace_xfs_refcount_find_right_extent(cur->bc_mp, cur->bc_private.a.agno,
 740                         cright, right, agbno + aglen);
 741         return error;
 742
 743 out_error:
 744         trace_xfs_refcount_find_right_extent_error(cur->bc_mp,
 745                         cur->bc_private.a.agno, error, _RET_IP_);
 746         return error;
 747 }
 748
 749 /* Is this extent valid? */
 750 static inline bool
 751 xfs_refc_valid(
 752         struct xfs_refcount_irec        *rc)
 753 {
 754         return rc->rc_startblock != NULLAGBLOCK;
 755 }
 756
 757 /*
 758  * Try to merge with any extents on the boundaries of the adjustment range.
 759  */
 760 STATIC int
 761 xfs_refcount_merge_extents(
 762         struct xfs_btree_cur    *cur,
 763         xfs_agblock_t           *agbno,
 764         xfs_extlen_t            *aglen,
 765         enum xfs_refc_adjust_op adjust,
 766         int                     flags,
 767         bool                    *shape_changed)
 768 {
 769         struct xfs_refcount_irec        left = {0}, cleft = {0};
 770         struct xfs_refcount_irec        cright = {0}, right = {0};
 771         int                             error;
 772         unsigned long long              ulen;
 773         bool                            cequal;
 774
 775         *shape_changed = false;
 776         /*
 777          * Find the extent just below agbno [left], just above agbno [cleft],
 778          * just below (agbno + aglen) [cright], and just above (agbno + aglen)
 779          * [right].
 780          */
 781         error = xfs_refcount_find_left_extents(cur, &left, &cleft, *agbno,
 782                         *aglen, flags);
 783         if (error)
 784                 return error;
 785         error = xfs_refcount_find_right_extents(cur, &right, &cright, *agbno,
 786                         *aglen, flags);
 787         if (error)
 788                 return error;
 789
 790         /* No left or right extent to merge; exit. */
 791         if (!xfs_refc_valid(&left) && !xfs_refc_valid(&right))
 792                 return 0;
 793
 794         cequal = (cleft.rc_startblock == cright.rc_startblock) &&
 795                  (cleft.rc_blockcount == cright.rc_blockcount);
 796
 797         /* Try to merge left, cleft, and right.  cleft must == cright. */
 798         ulen = (unsigned long long)left.rc_blockcount + cleft.rc_blockcount +
 799                         right.rc_blockcount;
 800         if (xfs_refc_valid(&left) && xfs_refc_valid(&right) &&
 801             xfs_refc_valid(&cleft) && xfs_refc_valid(&cright) && cequal &&
 802             left.rc_refcount == cleft.rc_refcount + adjust &&
 803             right.rc_refcount == cleft.rc_refcount + adjust &&
 804             ulen < MAXREFCEXTLEN) {
 805                 *shape_changed = true;
 806                 return xfs_refcount_merge_center_extents(cur, &left, &cleft,
 807                                 &right, ulen, aglen);
 808         }
 809
 810         /* Try to merge left and cleft. */
 811         ulen = (unsigned long long)left.rc_blockcount + cleft.rc_blockcount;
 812         if (xfs_refc_valid(&left) && xfs_refc_valid(&cleft) &&
 813             left.rc_refcount == cleft.rc_refcount + adjust &&
 814             ulen < MAXREFCEXTLEN) {
 815                 *shape_changed = true;
 816                 error = xfs_refcount_merge_left_extent(cur, &left, &cleft,
 817                                 agbno, aglen);
 818                 if (error)
 819                         return error;
 820
 821                 /*
 822                  * If we just merged left + cleft and cleft == cright,
 823                  * we no longer have a cright to merge with right.  We're done.
 824                  */
 825                 if (cequal)
 826                         return 0;
 827         }
 828
 829         /* Try to merge cright and right. */
 830         ulen = (unsigned long long)right.rc_blockcount + cright.rc_blockcount;
 831         if (xfs_refc_valid(&right) && xfs_refc_valid(&cright) &&
 832             right.rc_refcount == cright.rc_refcount + adjust &&
 833             ulen < MAXREFCEXTLEN) {
 834                 *shape_changed = true;
 835                 return xfs_refcount_merge_right_extent(cur, &right, &cright,
 836                                 aglen);
 837         }
 838
 839         return error;
 840 }
 841
 842 /*
 843  * XXX: This is a pretty hand-wavy estimate.  The penalty for guessing
 844  * true incorrectly is a shutdown FS; the penalty for guessing false
 845  * incorrectly is more transaction rolls than might be necessary.
 846  * Be conservative here.
 847  */
 848 static bool
 849 xfs_refcount_still_have_space(
 850         struct xfs_btree_cur            *cur)
 851 {
 852         unsigned long                   overhead;
 853
 854         overhead = cur->bc_private.a.priv.refc.shape_changes *
 855                         xfs_allocfree_log_count(cur->bc_mp, 1);
 856         overhead *= cur->bc_mp->m_sb.sb_blocksize;
 857
 858         /*
 859          * Only allow 2 refcount extent updates per transaction if the
 860          * refcount continue update "error" has been injected.
 861          */
 862         if (cur->bc_private.a.priv.refc.nr_ops > 2 &&
 863             XFS_TEST_ERROR(false, cur->bc_mp,
 864                         XFS_ERRTAG_REFCOUNT_CONTINUE_UPDATE))
 865                 return false;
 866
 867         if (cur->bc_private.a.priv.refc.nr_ops == 0)
 868                 return true;
 869         else if (overhead > cur->bc_tp->t_log_res)
 870                 return false;
 871         return  cur->bc_tp->t_log_res - overhead >
 872                 cur->bc_private.a.priv.refc.nr_ops * XFS_REFCOUNT_ITEM_OVERHEAD;
 873 }
 874
 875 /*
 876  * Adjust the refcounts of middle extents.  At this point we should have
 877  * split extents that crossed the adjustment range; merged with adjacent
 878  * extents; and updated agbno/aglen to reflect the merges.  Therefore,
 879  * all we have to do is update the extents inside [agbno, agbno + aglen].
 880  */
 881 STATIC int
 882 xfs_refcount_adjust_extents(
 883         struct xfs_btree_cur    *cur,
 884         xfs_agblock_t           *agbno,
 885         xfs_extlen_t            *aglen,
 886         enum xfs_refc_adjust_op adj,
 887         struct xfs_defer_ops    *dfops,
 888         struct xfs_owner_info   *oinfo)
 889 {
 890         struct xfs_refcount_irec        ext, tmp;
 891         int                             error;
 892         int                             found_rec, found_tmp;
 893         xfs_fsblock_t                   fsbno;
 894
 895         /* Merging did all the work already. */
 896         if (*aglen == 0)
 897                 return 0;
 898
 899         error = xfs_refcount_lookup_ge(cur, *agbno, &found_rec);
 900         if (error)
 901                 goto out_error;
 902
 903         while (*aglen > 0 && xfs_refcount_still_have_space(cur)) {
 904                 error = xfs_refcount_get_rec(cur, &ext, &found_rec);
 905                 if (error)
 906                         goto out_error;
 907                 if (!found_rec) {
 908                         ext.rc_startblock = cur->bc_mp->m_sb.sb_agblocks;
 909                         ext.rc_blockcount = 0;
 910                         ext.rc_refcount = 0;
 911                 }
 912
 913                 /*
 914                  * Deal with a hole in the refcount tree; if a file maps to
 915                  * these blocks and there's no refcountbt record, pretend that
 916                  * there is one with refcount == 1.
 917                  */
 918                 if (ext.rc_startblock != *agbno) {
 919                         tmp.rc_startblock = *agbno;
 920                         tmp.rc_blockcount = min(*aglen,
 921                                         ext.rc_startblock - *agbno);
 922                         tmp.rc_refcount = 1 + adj;
 923                         trace_xfs_refcount_modify_extent(cur->bc_mp,
 924                                         cur->bc_private.a.agno, &tmp);
 925
 926                         /*
 927                          * Either cover the hole (increment) or
 928                          * delete the range (decrement).
 929                          */
 930                         if (tmp.rc_refcount) {
 931                                 error = xfs_refcount_insert(cur, &tmp,
 932                                                 &found_tmp);
 933                                 if (error)
 934                                         goto out_error;
 935                                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
 936                                                 found_tmp == 1, out_error);
 937                                 cur->bc_private.a.priv.refc.nr_ops++;
 938                         } else {
 939                                 fsbno = XFS_AGB_TO_FSB(cur->bc_mp,
 940                                                 cur->bc_private.a.agno,
 941                                                 tmp.rc_startblock);
 942                                 xfs_bmap_add_free(cur->bc_mp, dfops, fsbno,
 943                                                 tmp.rc_blockcount, oinfo);
 944                         }
 945
 946                         (*agbno) += tmp.rc_blockcount;
 947                         (*aglen) -= tmp.rc_blockcount;
 948
 949                         error = xfs_refcount_lookup_ge(cur, *agbno,
 950                                         &found_rec);
 951                         if (error)
 952                                 goto out_error;
 953                 }
 954
 955                 /* Stop if there's nothing left to modify */
 956                 if (*aglen == 0 || !xfs_refcount_still_have_space(cur))
 957                         break;
 958
 959                 /*
 960                  * Adjust the reference count and either update the tree
 961                  * (incr) or free the blocks (decr).
 962                  */
 963                 if (ext.rc_refcount == MAXREFCOUNT)
 964                         goto skip;
 965                 ext.rc_refcount += adj;
 966                 trace_xfs_refcount_modify_extent(cur->bc_mp,
 967                                 cur->bc_private.a.agno, &ext);
 968                 if (ext.rc_refcount > 1) {
 969                         error = xfs_refcount_update(cur, &ext);
 970                         if (error)
 971                                 goto out_error;
 972                         cur->bc_private.a.priv.refc.nr_ops++;
 973                 } else if (ext.rc_refcount == 1) {
 974                         error = xfs_refcount_delete(cur, &found_rec);
 975                         if (error)
 976                                 goto out_error;
 977                         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
 978                                         found_rec == 1, out_error);
 979                         cur->bc_private.a.priv.refc.nr_ops++;
 980                         goto advloop;
 981                 } else {
 982                         fsbno = XFS_AGB_TO_FSB(cur->bc_mp,
 983                                         cur->bc_private.a.agno,
 984                                         ext.rc_startblock);
 985                         xfs_bmap_add_free(cur->bc_mp, dfops, fsbno,
 986                                         ext.rc_blockcount, oinfo);
 987                 }
 988
 989 skip:
 990                 error = xfs_btree_increment(cur, 0, &found_rec);
 991                 if (error)
 992                         goto out_error;
 993
 994 advloop:
 995                 (*agbno) += ext.rc_blockcount;
 996                 (*aglen) -= ext.rc_blockcount;
 997         }
 998
 999         return error;
1000 out_error:
1001         trace_xfs_refcount_modify_extent_error(cur->bc_mp,
1002                         cur->bc_private.a.agno, error, _RET_IP_);
1003         return error;
1004 }
1005
1006 /* Adjust the reference count of a range of AG blocks. */
1007 STATIC int
1008 xfs_refcount_adjust(
1009         struct xfs_btree_cur    *cur,
1010         xfs_agblock_t           agbno,
1011         xfs_extlen_t            aglen,
1012         xfs_agblock_t           *new_agbno,
1013         xfs_extlen_t            *new_aglen,
1014         enum xfs_refc_adjust_op adj,
1015         struct xfs_defer_ops    *dfops,
1016         struct xfs_owner_info   *oinfo)
1017 {
1018         bool                    shape_changed;
1019         int                     shape_changes = 0;
1020         int                     error;
1021
1022         *new_agbno = agbno;
1023         *new_aglen = aglen;
1024         if (adj == XFS_REFCOUNT_ADJUST_INCREASE)
1025                 trace_xfs_refcount_increase(cur->bc_mp, cur->bc_private.a.agno,
1026                                 agbno, aglen);
1027         else
1028                 trace_xfs_refcount_decrease(cur->bc_mp, cur->bc_private.a.agno,
1029                                 agbno, aglen);
1030
1031         /*
1032          * Ensure that no rcextents cross the boundary of the adjustment range.
1033          */
1034         error = xfs_refcount_split_extent(cur, agbno, &shape_changed);
1035         if (error)
1036                 goto out_error;
1037         if (shape_changed)
1038                 shape_changes++;
1039
1040         error = xfs_refcount_split_extent(cur, agbno + aglen, &shape_changed);
1041         if (error)
1042                 goto out_error;
1043         if (shape_changed)
1044                 shape_changes++;
1045
1046         /*
1047          * Try to merge with the left or right extents of the range.
1048          */
1049         error = xfs_refcount_merge_extents(cur, new_agbno, new_aglen, adj,
1050                         XFS_FIND_RCEXT_SHARED, &shape_changed);
1051         if (error)
1052                 goto out_error;
1053         if (shape_changed)
1054                 shape_changes++;
1055         if (shape_changes)
1056                 cur->bc_private.a.priv.refc.shape_changes++;
1057
1058         /* Now that we've taken care of the ends, adjust the middle extents */
1059         error = xfs_refcount_adjust_extents(cur, new_agbno, new_aglen,
1060                         adj, dfops, oinfo);
1061         if (error)
1062                 goto out_error;
1063
1064         return 0;
1065
1066 out_error:
1067         trace_xfs_refcount_adjust_error(cur->bc_mp, cur->bc_private.a.agno,
1068                         error, _RET_IP_);
1069         return error;
1070 }
1071
1072 /* Clean up after calling xfs_refcount_finish_one. */
1073 void
1074 xfs_refcount_finish_one_cleanup(
1075         struct xfs_trans        *tp,
1076         struct xfs_btree_cur    *rcur,
1077         int                     error)
1078 {
1079         struct xfs_buf          *agbp;
1080
1081         if (rcur == NULL)
1082                 return;
1083         agbp = rcur->bc_private.a.agbp;
1084         xfs_btree_del_cursor(rcur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
1085         if (error)
1086                 xfs_trans_brelse(tp, agbp);
1087 }
1088
1089 /*
1090  * Process one of the deferred refcount operations.  We pass back the
1091  * btree cursor to maintain our lock on the btree between calls.
1092  * This saves time and eliminates a buffer deadlock between the
1093  * superblock and the AGF because we'll always grab them in the same
1094  * order.
1095  */
1096 int
1097 xfs_refcount_finish_one(
1098         struct xfs_trans                *tp,
1099         struct xfs_defer_ops            *dfops,
1100         enum xfs_refcount_intent_type   type,
1101         xfs_fsblock_t                   startblock,
1102         xfs_extlen_t                    blockcount,
1103         xfs_fsblock_t                   *new_fsb,
1104         xfs_extlen_t                    *new_len,
1105         struct xfs_btree_cur            **pcur)
1106 {
1107         struct xfs_mount                *mp = tp->t_mountp;
1108         struct xfs_btree_cur            *rcur;
1109         struct xfs_buf                  *agbp = NULL;
1110         int                             error = 0;
1111         xfs_agnumber_t                  agno;
1112         xfs_agblock_t                   bno;
1113         xfs_agblock_t                   new_agbno;
1114         unsigned long                   nr_ops = 0;
1115         int                             shape_changes = 0;
1116
1117         agno = XFS_FSB_TO_AGNO(mp, startblock);
1118         ASSERT(agno != NULLAGNUMBER);
1119         bno = XFS_FSB_TO_AGBNO(mp, startblock);
1120
1121         trace_xfs_refcount_deferred(mp, XFS_FSB_TO_AGNO(mp, startblock),
1122                         type, XFS_FSB_TO_AGBNO(mp, startblock),
1123                         blockcount);
1124
1125         if (XFS_TEST_ERROR(false, mp,
1126                         XFS_ERRTAG_REFCOUNT_FINISH_ONE))
1127                 return -EIO;
1128
1129         /*
1130          * If we haven't gotten a cursor or the cursor AG doesn't match
1131          * the startblock, get one now.
1132          */
1133         rcur = *pcur;
1134         if (rcur != NULL && rcur->bc_private.a.agno != agno) {
1135                 nr_ops = rcur->bc_private.a.priv.refc.nr_ops;
1136                 shape_changes = rcur->bc_private.a.priv.refc.shape_changes;
1137                 xfs_refcount_finish_one_cleanup(tp, rcur, 0);
1138                 rcur = NULL;
1139                 *pcur = NULL;
1140         }
1141         if (rcur == NULL) {
1142                 error = xfs_alloc_read_agf(tp->t_mountp, tp, agno,
1143                                 XFS_ALLOC_FLAG_FREEING, &agbp);
1144                 if (error)
1145                         return error;
1146                 if (!agbp)
1147                         return -EFSCORRUPTED;
1148
1149                 rcur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno, dfops);
1150                 if (!rcur) {
1151                         error = -ENOMEM;
1152                         goto out_cur;
1153                 }
1154                 rcur->bc_private.a.priv.refc.nr_ops = nr_ops;
1155                 rcur->bc_private.a.priv.refc.shape_changes = shape_changes;
1156         }
1157         *pcur = rcur;
1158
1159         switch (type) {
1160         case XFS_REFCOUNT_INCREASE:
1161                 error = xfs_refcount_adjust(rcur, bno, blockcount, &new_agbno,
1162                         new_len, XFS_REFCOUNT_ADJUST_INCREASE, dfops, NULL);
1163                 *new_fsb = XFS_AGB_TO_FSB(mp, agno, new_agbno);
1164                 break;
1165         case XFS_REFCOUNT_DECREASE:
1166                 error = xfs_refcount_adjust(rcur, bno, blockcount, &new_agbno,
1167                         new_len, XFS_REFCOUNT_ADJUST_DECREASE, dfops, NULL);
1168                 *new_fsb = XFS_AGB_TO_FSB(mp, agno, new_agbno);
1169                 break;
1170         case XFS_REFCOUNT_ALLOC_COW:
1171                 *new_fsb = startblock + blockcount;
1172                 *new_len = 0;
1173                 error = __xfs_refcount_cow_alloc(rcur, bno, blockcount, dfops);
1174                 break;
1175         case XFS_REFCOUNT_FREE_COW:
1176                 *new_fsb = startblock + blockcount;
1177                 *new_len = 0;
1178                 error = __xfs_refcount_cow_free(rcur, bno, blockcount, dfops);
1179                 break;
1180         default:
1181                 ASSERT(0);
1182                 error = -EFSCORRUPTED;
1183         }
1184         if (!error && *new_len > 0)
1185                 trace_xfs_refcount_finish_one_leftover(mp, agno, type,
1186                                 bno, blockcount, new_agbno, *new_len);
1187         return error;
1188
1189 out_cur:
1190         xfs_trans_brelse(tp, agbp);
1191
1192         return error;
1193 }
1194
1195 /*
1196  * Record a refcount intent for later processing.
1197  */
1198 static int
1199 __xfs_refcount_add(
1200         struct xfs_mount                *mp,
1201         struct xfs_defer_ops            *dfops,
1202         enum xfs_refcount_intent_type   type,
1203         xfs_fsblock_t                   startblock,
1204         xfs_extlen_t                    blockcount)
1205 {
1206         struct xfs_refcount_intent      *ri;
1207
1208         trace_xfs_refcount_defer(mp, XFS_FSB_TO_AGNO(mp, startblock),
1209                         type, XFS_FSB_TO_AGBNO(mp, startblock),
1210                         blockcount);
1211
1212         ri = kmem_alloc(sizeof(struct xfs_refcount_intent),
1213                         KM_SLEEP | KM_NOFS);
1214         INIT_LIST_HEAD(&ri->ri_list);
1215         ri->ri_type = type;
1216         ri->ri_startblock = startblock;
1217         ri->ri_blockcount = blockcount;
1218
1219         xfs_defer_add(dfops, XFS_DEFER_OPS_TYPE_REFCOUNT, &ri->ri_list);
1220         return 0;
1221 }
1222
1223 /*
1224  * Increase the reference count of the blocks backing a file's extent.
1225  */
1226 int
1227 xfs_refcount_increase_extent(
1228         struct xfs_mount                *mp,
1229         struct xfs_defer_ops            *dfops,
1230         struct xfs_bmbt_irec            *PREV)
1231 {
1232         if (!xfs_sb_version_hasreflink(&mp->m_sb))
1233                 return 0;
1234
1235         return __xfs_refcount_add(mp, dfops, XFS_REFCOUNT_INCREASE,
1236                         PREV->br_startblock, PREV->br_blockcount);
1237 }
1238
1239 /*
1240  * Decrease the reference count of the blocks backing a file's extent.
1241  */
1242 int
1243 xfs_refcount_decrease_extent(
1244         struct xfs_mount                *mp,
1245         struct xfs_defer_ops            *dfops,
1246         struct xfs_bmbt_irec            *PREV)
1247 {
1248         if (!xfs_sb_version_hasreflink(&mp->m_sb))
1249                 return 0;
1250
1251         return __xfs_refcount_add(mp, dfops, XFS_REFCOUNT_DECREASE,
1252                         PREV->br_startblock, PREV->br_blockcount);
1253 }
1254
1255 /*
1256  * Given an AG extent, find the lowest-numbered run of shared blocks
1257  * within that range and return the range in fbno/flen.  If
1258  * find_end_of_shared is set, return the longest contiguous extent of
1259  * shared blocks; if not, just return the first extent we find.  If no
1260  * shared blocks are found, fbno and flen will be set to NULLAGBLOCK
1261  * and 0, respectively.
1262  */
1263 int
1264 xfs_refcount_find_shared(
1265         struct xfs_btree_cur            *cur,
1266         xfs_agblock_t                   agbno,
1267         xfs_extlen_t                    aglen,
1268         xfs_agblock_t                   *fbno,
1269         xfs_extlen_t                    *flen,
1270         bool                            find_end_of_shared)
1271 {
1272         struct xfs_refcount_irec        tmp;
1273         int                             i;
1274         int                             have;
1275         int                             error;
1276
1277         trace_xfs_refcount_find_shared(cur->bc_mp, cur->bc_private.a.agno,
1278                         agbno, aglen);
1279
1280         /* By default, skip the whole range */
1281         *fbno = NULLAGBLOCK;
1282         *flen = 0;
1283
1284         /* Try to find a refcount extent that crosses the start */
1285         error = xfs_refcount_lookup_le(cur, agbno, &have);
1286         if (error)
1287                 goto out_error;
1288         if (!have) {
1289                 /* No left extent, look at the next one */
1290                 error = xfs_btree_increment(cur, 0, &have);
1291                 if (error)
1292                         goto out_error;
1293                 if (!have)
1294                         goto done;
1295         }
1296         error = xfs_refcount_get_rec(cur, &tmp, &i);
1297         if (error)
1298                 goto out_error;
1299         XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, i == 1, out_error);
1300
1301         /* If the extent ends before the start, look at the next one */
1302         if (tmp.rc_startblock + tmp.rc_blockcount <= agbno) {
1303                 error = xfs_btree_increment(cur, 0, &have);
1304                 if (error)
1305                         goto out_error;
1306                 if (!have)
1307                         goto done;
1308                 error = xfs_refcount_get_rec(cur, &tmp, &i);
1309                 if (error)
1310                         goto out_error;
1311                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, i == 1, out_error);
1312         }
1313
1314         /* If the extent starts after the range we want, bail out */
1315         if (tmp.rc_startblock >= agbno + aglen)
1316                 goto done;
1317
1318         /* We found the start of a shared extent! */
1319         if (tmp.rc_startblock < agbno) {
1320                 tmp.rc_blockcount -= (agbno - tmp.rc_startblock);
1321                 tmp.rc_startblock = agbno;
1322         }
1323
1324         *fbno = tmp.rc_startblock;
1325         *flen = min(tmp.rc_blockcount, agbno + aglen - *fbno);
1326         if (!find_end_of_shared)
1327                 goto done;
1328
1329         /* Otherwise, find the end of this shared extent */
1330         while (*fbno + *flen < agbno + aglen) {
1331                 error = xfs_btree_increment(cur, 0, &have);
1332                 if (error)
1333                         goto out_error;
1334                 if (!have)
1335                         break;
1336                 error = xfs_refcount_get_rec(cur, &tmp, &i);
1337                 if (error)
1338                         goto out_error;
1339                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp, i == 1, out_error);
1340                 if (tmp.rc_startblock >= agbno + aglen ||
1341                     tmp.rc_startblock != *fbno + *flen)
1342                         break;
1343                 *flen = min(*flen + tmp.rc_blockcount, agbno + aglen - *fbno);
1344         }
1345
1346 done:
1347         trace_xfs_refcount_find_shared_result(cur->bc_mp,
1348                         cur->bc_private.a.agno, *fbno, *flen);
1349
1350 out_error:
1351         if (error)
1352                 trace_xfs_refcount_find_shared_error(cur->bc_mp,
1353                                 cur->bc_private.a.agno, error, _RET_IP_);
1354         return error;
1355 }
1356
1357 /*
1358  * Recovering CoW Blocks After a Crash
1359  *
1360  * Due to the way that the copy on write mechanism works, there's a window of
1361  * opportunity in which we can lose track of allocated blocks during a crash.
1362  * Because CoW uses delayed allocation in the in-core CoW fork, writeback
1363  * causes blocks to be allocated and stored in the CoW fork.  The blocks are
1364  * no longer in the free space btree but are not otherwise recorded anywhere
1365  * until the write completes and the blocks are mapped into the file.  A crash
1366  * in between allocation and remapping results in the replacement blocks being
1367  * lost.  This situation is exacerbated by the CoW extent size hint because
1368  * allocations can hang around for long time.
1369  *
1370  * However, there is a place where we can record these allocations before they
1371  * become mappings -- the reference count btree.  The btree does not record
1372  * extents with refcount == 1, so we can record allocations with a refcount of
1373  * 1.  Blocks being used for CoW writeout cannot be shared, so there should be
1374  * no conflict with shared block records.  These mappings should be created
1375  * when we allocate blocks to the CoW fork and deleted when they're removed
1376  * from the CoW fork.
1377  *
1378  * Minor nit: records for in-progress CoW allocations and records for shared
1379  * extents must never be merged, to preserve the property that (except for CoW
1380  * allocations) there are no refcount btree entries with refcount == 1.  The
1381  * only time this could potentially happen is when unsharing a block that's
1382  * adjacent to CoW allocations, so we must be careful to avoid this.
1383  *
1384  * At mount time we recover lost CoW allocations by searching the refcount
1385  * btree for these refcount == 1 mappings.  These represent CoW allocations
1386  * that were in progress at the time the filesystem went down, so we can free
1387  * them to get the space back.
1388  *
1389  * This mechanism is superior to creating EFIs for unmapped CoW extents for
1390  * several reasons -- first, EFIs pin the tail of the log and would have to be
1391  * periodically relogged to avoid filling up the log.  Second, CoW completions
1392  * will have to file an EFD and create new EFIs for whatever remains in the
1393  * CoW fork; this partially takes care of (1) but extent-size reservations
1394  * will have to periodically relog even if there's no writeout in progress.
1395  * This can happen if the CoW extent size hint is set, which you really want.
1396  * Third, EFIs cannot currently be automatically relogged into newer
1397  * transactions to advance the log tail.  Fourth, stuffing the log full of
1398  * EFIs places an upper bound on the number of CoW allocations that can be
1399  * held filesystem-wide at any given time.  Recording them in the refcount
1400  * btree doesn't require us to maintain any state in memory and doesn't pin
1401  * the log.
1402  */
1403 /*
1404  * Adjust the refcounts of CoW allocations.  These allocations are "magic"
1405  * in that they're not referenced anywhere else in the filesystem, so we
1406  * stash them in the refcount btree with a refcount of 1 until either file
1407  * remapping (or CoW cancellation) happens.
1408  */
1409 STATIC int
1410 xfs_refcount_adjust_cow_extents(
1411         struct xfs_btree_cur    *cur,
1412         xfs_agblock_t           agbno,
1413         xfs_extlen_t            aglen,
1414         enum xfs_refc_adjust_op adj)
1415 {
1416         struct xfs_refcount_irec        ext, tmp;
1417         int                             error;
1418         int                             found_rec, found_tmp;
1419
1420         if (aglen == 0)
1421                 return 0;
1422
1423         /* Find any overlapping refcount records */
1424         error = xfs_refcount_lookup_ge(cur, agbno, &found_rec);
1425         if (error)
1426                 goto out_error;
1427         error = xfs_refcount_get_rec(cur, &ext, &found_rec);
1428         if (error)
1429                 goto out_error;
1430         if (!found_rec) {
1431                 ext.rc_startblock = cur->bc_mp->m_sb.sb_agblocks +
1432                                 XFS_REFC_COW_START;
1433                 ext.rc_blockcount = 0;
1434                 ext.rc_refcount = 0;
1435         }
1436
1437         switch (adj) {
1438         case XFS_REFCOUNT_ADJUST_COW_ALLOC:
1439                 /* Adding a CoW reservation, there should be nothing here. */
1440                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1441                                 ext.rc_startblock >= agbno + aglen, out_error);
1442
1443                 tmp.rc_startblock = agbno;
1444                 tmp.rc_blockcount = aglen;
1445                 tmp.rc_refcount = 1;
1446                 trace_xfs_refcount_modify_extent(cur->bc_mp,
1447                                 cur->bc_private.a.agno, &tmp);
1448
1449                 error = xfs_refcount_insert(cur, &tmp,
1450                                 &found_tmp);
1451                 if (error)
1452                         goto out_error;
1453                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1454                                 found_tmp == 1, out_error);
1455                 break;
1456         case XFS_REFCOUNT_ADJUST_COW_FREE:
1457                 /* Removing a CoW reservation, there should be one extent. */
1458                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1459                         ext.rc_startblock == agbno, out_error);
1460                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1461                         ext.rc_blockcount == aglen, out_error);
1462                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1463                         ext.rc_refcount == 1, out_error);
1464
1465                 ext.rc_refcount = 0;
1466                 trace_xfs_refcount_modify_extent(cur->bc_mp,
1467                                 cur->bc_private.a.agno, &ext);
1468                 error = xfs_refcount_delete(cur, &found_rec);
1469                 if (error)
1470                         goto out_error;
1471                 XFS_WANT_CORRUPTED_GOTO(cur->bc_mp,
1472                                 found_rec == 1, out_error);
1473                 break;
1474         default:
1475                 ASSERT(0);
1476         }
1477
1478         return error;
1479 out_error:
1480         trace_xfs_refcount_modify_extent_error(cur->bc_mp,
1481                         cur->bc_private.a.agno, error, _RET_IP_);
1482         return error;
1483 }
1484
1485 /*
1486  * Add or remove refcount btree entries for CoW reservations.
1487  */
1488 STATIC int
1489 xfs_refcount_adjust_cow(
1490         struct xfs_btree_cur    *cur,
1491         xfs_agblock_t           agbno,
1492         xfs_extlen_t            aglen,
1493         enum xfs_refc_adjust_op adj)
1494 {
1495         bool                    shape_changed;
1496         int                     error;
1497
1498         agbno += XFS_REFC_COW_START;
1499
1500         /*
1501          * Ensure that no rcextents cross the boundary of the adjustment range.
1502          */
1503         error = xfs_refcount_split_extent(cur, agbno, &shape_changed);
1504         if (error)
1505                 goto out_error;
1506
1507         error = xfs_refcount_split_extent(cur, agbno + aglen, &shape_changed);
1508         if (error)
1509                 goto out_error;
1510
1511         /*
1512          * Try to merge with the left or right extents of the range.
1513          */
1514         error = xfs_refcount_merge_extents(cur, &agbno, &aglen, adj,
1515                         XFS_FIND_RCEXT_COW, &shape_changed);
1516         if (error)
1517                 goto out_error;
1518
1519         /* Now that we've taken care of the ends, adjust the middle extents */
1520         error = xfs_refcount_adjust_cow_extents(cur, agbno, aglen, adj);
1521         if (error)
1522                 goto out_error;
1523
1524         return 0;
1525
1526 out_error:
1527         trace_xfs_refcount_adjust_cow_error(cur->bc_mp, cur->bc_private.a.agno,
1528                         error, _RET_IP_);
1529         return error;
1530 }
1531
1532 /*
1533  * Record a CoW allocation in the refcount btree.
1534  */
1535 STATIC int
1536 __xfs_refcount_cow_alloc(
1537         struct xfs_btree_cur    *rcur,
1538         xfs_agblock_t           agbno,
1539         xfs_extlen_t            aglen,
1540         struct xfs_defer_ops    *dfops)
1541 {
1542         trace_xfs_refcount_cow_increase(rcur->bc_mp, rcur->bc_private.a.agno,
1543                         agbno, aglen);
1544
1545         /* Add refcount btree reservation */
1546         return xfs_refcount_adjust_cow(rcur, agbno, aglen,
1547                         XFS_REFCOUNT_ADJUST_COW_ALLOC);
1548 }
1549
1550 /*
1551  * Remove a CoW allocation from the refcount btree.
1552  */
1553 STATIC int
1554 __xfs_refcount_cow_free(
1555         struct xfs_btree_cur    *rcur,
1556         xfs_agblock_t           agbno,
1557         xfs_extlen_t            aglen,
1558         struct xfs_defer_ops    *dfops)
1559 {
1560         trace_xfs_refcount_cow_decrease(rcur->bc_mp, rcur->bc_private.a.agno,
1561                         agbno, aglen);
1562
1563         /* Remove refcount btree reservation */
1564         return xfs_refcount_adjust_cow(rcur, agbno, aglen,
1565                         XFS_REFCOUNT_ADJUST_COW_FREE);
1566 }
1567
1568 /* Record a CoW staging extent in the refcount btree. */
1569 int
1570 xfs_refcount_alloc_cow_extent(
1571         struct xfs_mount                *mp,
1572         struct xfs_defer_ops            *dfops,
1573         xfs_fsblock_t                   fsb,
1574         xfs_extlen_t                    len)
1575 {
1576         int                             error;
1577
1578         if (!xfs_sb_version_hasreflink(&mp->m_sb))
1579                 return 0;
1580
1581         error = __xfs_refcount_add(mp, dfops, XFS_REFCOUNT_ALLOC_COW,
1582                         fsb, len);
1583         if (error)
1584                 return error;
1585
1586         /* Add rmap entry */
1587         return xfs_rmap_alloc_extent(mp, dfops, XFS_FSB_TO_AGNO(mp, fsb),
1588                         XFS_FSB_TO_AGBNO(mp, fsb), len, XFS_RMAP_OWN_COW);
1589 }
1590
1591 /* Forget a CoW staging event in the refcount btree. */
1592 int
1593 xfs_refcount_free_cow_extent(
1594         struct xfs_mount                *mp,
1595         struct xfs_defer_ops            *dfops,
1596         xfs_fsblock_t                   fsb,
1597         xfs_extlen_t                    len)
1598 {
1599         int                             error;
1600
1601         if (!xfs_sb_version_hasreflink(&mp->m_sb))
1602                 return 0;
1603
1604         /* Remove rmap entry */
1605         error = xfs_rmap_free_extent(mp, dfops, XFS_FSB_TO_AGNO(mp, fsb),
1606                         XFS_FSB_TO_AGBNO(mp, fsb), len, XFS_RMAP_OWN_COW);
1607         if (error)
1608                 return error;
1609
1610         return __xfs_refcount_add(mp, dfops, XFS_REFCOUNT_FREE_COW,
1611                         fsb, len);
1612 }
1613
1614 struct xfs_refcount_recovery {
1615         struct list_head                rr_list;
1616         struct xfs_refcount_irec        rr_rrec;
1617 };
1618
1619 /* Stuff an extent on the recovery list. */
1620 STATIC int
1621 xfs_refcount_recover_extent(
1622         struct xfs_btree_cur            *cur,
1623         union xfs_btree_rec             *rec,
1624         void                            *priv)
1625 {
1626         struct list_head                *debris = priv;
1627         struct xfs_refcount_recovery    *rr;
1628
1629         if (be32_to_cpu(rec->refc.rc_refcount) != 1)
1630                 return -EFSCORRUPTED;
1631
1632         rr = kmem_alloc(sizeof(struct xfs_refcount_recovery), KM_SLEEP);
1633         xfs_refcount_btrec_to_irec(rec, &rr->rr_rrec);
1634         list_add_tail(&rr->rr_list, debris);
1635
1636         return 0;
1637 }
1638
1639 /* Find and remove leftover CoW reservations. */
1640 int
1641 xfs_refcount_recover_cow_leftovers(
1642         struct xfs_mount                *mp,
1643         xfs_agnumber_t                  agno)
1644 {
1645         struct xfs_trans                *tp;
1646         struct xfs_btree_cur            *cur;
1647         struct xfs_buf                  *agbp;
1648         struct xfs_refcount_recovery    *rr, *n;
1649         struct list_head                debris;
1650         union xfs_btree_irec            low;
1651         union xfs_btree_irec            high;
1652         struct xfs_defer_ops            dfops;
1653         xfs_fsblock_t                   fsb;
1654         xfs_agblock_t                   agbno;
1655         int                             error;
1656
1657         if (mp->m_sb.sb_agblocks >= XFS_REFC_COW_START)
1658                 return -EOPNOTSUPP;
1659
1660         INIT_LIST_HEAD(&debris);
1661
1662         /*
1663          * In this first part, we use an empty transaction to gather up
1664          * all the leftover CoW extents so that we can subsequently
1665          * delete them.  The empty transaction is used to avoid
1666          * a buffer lock deadlock if there happens to be a loop in the
1667          * refcountbt because we're allowed to re-grab a buffer that is
1668          * already attached to our transaction.  When we're done
1669          * recording the CoW debris we cancel the (empty) transaction
1670          * and everything goes away cleanly.
1671          */
1672         error = xfs_trans_alloc_empty(mp, &tp);
1673         if (error)
1674                 return error;
1675
1676         error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
1677         if (error)
1678                 goto out_trans;
1679         if (!agbp) {
1680                 error = -ENOMEM;
1681                 goto out_trans;
1682         }
1683         cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno, NULL);
1684
1685         /* Find all the leftover CoW staging extents. */
1686         memset(&low, 0, sizeof(low));
1687         memset(&high, 0, sizeof(high));
1688         low.rc.rc_startblock = XFS_REFC_COW_START;
1689         high.rc.rc_startblock = -1U;
1690         error = xfs_btree_query_range(cur, &low, &high,
1691                         xfs_refcount_recover_extent, &debris);
1692         if (error)
1693                 goto out_cursor;
1694         xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1695         xfs_trans_brelse(tp, agbp);
1696         xfs_trans_cancel(tp);
1697
1698         /* Now iterate the list to free the leftovers */
1699         list_for_each_entry_safe(rr, n, &debris, rr_list) {
1700                 /* Set up transaction. */
1701                 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1702                 if (error)
1703                         goto out_free;
1704
1705                 trace_xfs_refcount_recover_extent(mp, agno, &rr->rr_rrec);
1706
1707                 /* Free the orphan record */
1708                 xfs_defer_init(&dfops, &fsb);
1709                 agbno = rr->rr_rrec.rc_startblock - XFS_REFC_COW_START;
1710                 fsb = XFS_AGB_TO_FSB(mp, agno, agbno);
1711                 error = xfs_refcount_free_cow_extent(mp, &dfops, fsb,
1712                                 rr->rr_rrec.rc_blockcount);
1713                 if (error)
1714                         goto out_defer;
1715
1716                 /* Free the block. */
1717                 xfs_bmap_add_free(mp, &dfops, fsb,
1718                                 rr->rr_rrec.rc_blockcount, NULL);
1719
1720                 error = xfs_defer_finish(&tp, &dfops);
1721                 if (error)
1722                         goto out_defer;
1723
1724                 error = xfs_trans_commit(tp);
1725                 if (error)
1726                         goto out_free;
1727
1728                 list_del(&rr->rr_list);
1729                 kmem_free(rr);
1730         }
1731
1732         return error;
1733 out_defer:
1734         xfs_defer_cancel(&dfops);
1735 out_trans:
1736         xfs_trans_cancel(tp);
1737 out_free:
1738         /* Free the leftover list */
1739         list_for_each_entry_safe(rr, n, &debris, rr_list) {
1740                 list_del(&rr->rr_list);
1741                 kmem_free(rr);
1742         }
1743         return error;
1744
1745 out_cursor:
1746         xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1747         xfs_trans_brelse(tp, agbp);
1748         goto out_trans;
1749 }
1750
1751 /* Is there a record covering a given extent? */
1752 int
1753 xfs_refcount_has_record(
1754         struct xfs_btree_cur    *cur,
1755         xfs_agblock_t           bno,
1756         xfs_extlen_t            len,
1757         bool                    *exists)
1758 {
1759         union xfs_btree_irec    low;
1760         union xfs_btree_irec    high;
1761
1762         memset(&low, 0, sizeof(low));
1763         low.rc.rc_startblock = bno;
1764         memset(&high, 0xFF, sizeof(high));
1765         high.rc.rc_startblock = bno + len - 1;
1766
1767         return xfs_btree_has_record(cur, &low, &high, exists);
1768 }