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