fs/xfs/xfs_extent_busy.c

   1 /*
   2  * Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
   3  * Copyright (c) 2010 David Chinner.
   4  * Copyright (c) 2011 Christoph Hellwig.
   5  * All Rights Reserved.
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License as
   9  * published by the Free Software Foundation.
  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_format.h"
  23 #include "xfs_log_format.h"
  24 #include "xfs_shared.h"
  25 #include "xfs_trans_resv.h"
  26 #include "xfs_sb.h"
  27 #include "xfs_ag.h"
  28 #include "xfs_mount.h"
  29 #include "xfs_bmap_btree.h"
  30 #include "xfs_alloc.h"
  31 #include "xfs_inode.h"
  32 #include "xfs_extent_busy.h"
  33 #include "xfs_trace.h"
  34 #include "xfs_trans.h"
  35 #include "xfs_log.h"
  36
  37 void
  38 xfs_extent_busy_insert(
  39         struct xfs_trans        *tp,
  40         xfs_agnumber_t          agno,
  41         xfs_agblock_t           bno,
  42         xfs_extlen_t            len,
  43         unsigned int            flags)
  44 {
  45         struct xfs_extent_busy  *new;
  46         struct xfs_extent_busy  *busyp;
  47         struct xfs_perag        *pag;
  48         struct rb_node          **rbp;
  49         struct rb_node          *parent = NULL;
  50
  51         new = kmem_zalloc(sizeof(struct xfs_extent_busy), KM_MAYFAIL);
  52         if (!new) {
  53                 /*
  54                  * No Memory!  Since it is now not possible to track the free
  55                  * block, make this a synchronous transaction to insure that
  56                  * the block is not reused before this transaction commits.
  57                  */
  58                 trace_xfs_extent_busy_enomem(tp->t_mountp, agno, bno, len);
  59                 xfs_trans_set_sync(tp);
  60                 return;
  61         }
  62
  63         new->agno = agno;
  64         new->bno = bno;
  65         new->length = len;
  66         INIT_LIST_HEAD(&new->list);
  67         new->flags = flags;
  68
  69         /* trace before insert to be able to see failed inserts */
  70         trace_xfs_extent_busy(tp->t_mountp, agno, bno, len);
  71
  72         pag = xfs_perag_get(tp->t_mountp, new->agno);
  73         spin_lock(&pag->pagb_lock);
  74         rbp = &pag->pagb_tree.rb_node;
  75         while (*rbp) {
  76                 parent = *rbp;
  77                 busyp = rb_entry(parent, struct xfs_extent_busy, rb_node);
  78
  79                 if (new->bno < busyp->bno) {
  80                         rbp = &(*rbp)->rb_left;
  81                         ASSERT(new->bno + new->length <= busyp->bno);
  82                 } else if (new->bno > busyp->bno) {
  83                         rbp = &(*rbp)->rb_right;
  84                         ASSERT(bno >= busyp->bno + busyp->length);
  85                 } else {
  86                         ASSERT(0);
  87                 }
  88         }
  89
  90         rb_link_node(&new->rb_node, parent, rbp);
  91         rb_insert_color(&new->rb_node, &pag->pagb_tree);
  92
  93         list_add(&new->list, &tp->t_busy);
  94         spin_unlock(&pag->pagb_lock);
  95         xfs_perag_put(pag);
  96 }
  97
  98 /*
  99  * Search for a busy extent within the range of the extent we are about to
 100  * allocate.  You need to be holding the busy extent tree lock when calling
 101  * xfs_extent_busy_search(). This function returns 0 for no overlapping busy
 102  * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
 103  * match. This is done so that a non-zero return indicates an overlap that
 104  * will require a synchronous transaction, but it can still be
 105  * used to distinguish between a partial or exact match.
 106  */
 107 int
 108 xfs_extent_busy_search(
 109         struct xfs_mount        *mp,
 110         xfs_agnumber_t          agno,
 111         xfs_agblock_t           bno,
 112         xfs_extlen_t            len)
 113 {
 114         struct xfs_perag        *pag;
 115         struct rb_node          *rbp;
 116         struct xfs_extent_busy  *busyp;
 117         int                     match = 0;
 118
 119         pag = xfs_perag_get(mp, agno);
 120         spin_lock(&pag->pagb_lock);
 121
 122         rbp = pag->pagb_tree.rb_node;
 123
 124         /* find closest start bno overlap */
 125         while (rbp) {
 126                 busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node);
 127                 if (bno < busyp->bno) {
 128                         /* may overlap, but exact start block is lower */
 129                         if (bno + len > busyp->bno)
 130                                 match = -1;
 131                         rbp = rbp->rb_left;
 132                 } else if (bno > busyp->bno) {
 133                         /* may overlap, but exact start block is higher */
 134                         if (bno < busyp->bno + busyp->length)
 135                                 match = -1;
 136                         rbp = rbp->rb_right;
 137                 } else {
 138                         /* bno matches busyp, length determines exact match */
 139                         match = (busyp->length == len) ? 1 : -1;
 140                         break;
 141                 }
 142         }
 143         spin_unlock(&pag->pagb_lock);
 144         xfs_perag_put(pag);
 145         return match;
 146 }
 147
 148 /*
 149  * The found free extent [fbno, fend] overlaps part or all of the given busy
 150  * extent.  If the overlap covers the beginning, the end, or all of the busy
 151  * extent, the overlapping portion can be made unbusy and used for the
 152  * allocation.  We can't split a busy extent because we can't modify a
 153  * transaction/CIL context busy list, but we can update an entry's block
 154  * number or length.
 155  *
 156  * Returns true if the extent can safely be reused, or false if the search
 157  * needs to be restarted.
 158  */
 159 STATIC bool
 160 xfs_extent_busy_update_extent(
 161         struct xfs_mount        *mp,
 162         struct xfs_perag        *pag,
 163         struct xfs_extent_busy  *busyp,
 164         xfs_agblock_t           fbno,
 165         xfs_extlen_t            flen,
 166         bool                    userdata) __releases(&pag->pagb_lock)
 167                                           __acquires(&pag->pagb_lock)
 168 {
 169         xfs_agblock_t           fend = fbno + flen;
 170         xfs_agblock_t           bbno = busyp->bno;
 171         xfs_agblock_t           bend = bbno + busyp->length;
 172
 173         /*
 174          * This extent is currently being discarded.  Give the thread
 175          * performing the discard a chance to mark the extent unbusy
 176          * and retry.
 177          */
 178         if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
 179                 spin_unlock(&pag->pagb_lock);
 180                 delay(1);
 181                 spin_lock(&pag->pagb_lock);
 182                 return false;
 183         }
 184
 185         /*
 186          * If there is a busy extent overlapping a user allocation, we have
 187          * no choice but to force the log and retry the search.
 188          *
 189          * Fortunately this does not happen during normal operation, but
 190          * only if the filesystem is very low on space and has to dip into
 191          * the AGFL for normal allocations.
 192          */
 193         if (userdata)
 194                 goto out_force_log;
 195
 196         if (bbno < fbno && bend > fend) {
 197                 /*
 198                  * Case 1:
 199                  *    bbno           bend
 200                  *    +BBBBBBBBBBBBBBBBB+
 201                  *        +---------+
 202                  *        fbno   fend
 203                  */
 204
 205                 /*
 206                  * We would have to split the busy extent to be able to track
 207                  * it correct, which we cannot do because we would have to
 208                  * modify the list of busy extents attached to the transaction
 209                  * or CIL context, which is immutable.
 210                  *
 211                  * Force out the log to clear the busy extent and retry the
 212                  * search.
 213                  */
 214                 goto out_force_log;
 215         } else if (bbno >= fbno && bend <= fend) {
 216                 /*
 217                  * Case 2:
 218                  *    bbno           bend
 219                  *    +BBBBBBBBBBBBBBBBB+
 220                  *    +-----------------+
 221                  *    fbno           fend
 222                  *
 223                  * Case 3:
 224                  *    bbno           bend
 225                  *    +BBBBBBBBBBBBBBBBB+
 226                  *    +--------------------------+
 227                  *    fbno                    fend
 228                  *
 229                  * Case 4:
 230                  *             bbno           bend
 231                  *             +BBBBBBBBBBBBBBBBB+
 232                  *    +--------------------------+
 233                  *    fbno                    fend
 234                  *
 235                  * Case 5:
 236                  *             bbno           bend
 237                  *             +BBBBBBBBBBBBBBBBB+
 238                  *    +-----------------------------------+
 239                  *    fbno                             fend
 240                  *
 241                  */
 242
 243                 /*
 244                  * The busy extent is fully covered by the extent we are
 245                  * allocating, and can simply be removed from the rbtree.
 246                  * However we cannot remove it from the immutable list
 247                  * tracking busy extents in the transaction or CIL context,
 248                  * so set the length to zero to mark it invalid.
 249                  *
 250                  * We also need to restart the busy extent search from the
 251                  * tree root, because erasing the node can rearrange the
 252                  * tree topology.
 253                  */
 254                 rb_erase(&busyp->rb_node, &pag->pagb_tree);
 255                 busyp->length = 0;
 256                 return false;
 257         } else if (fend < bend) {
 258                 /*
 259                  * Case 6:
 260                  *              bbno           bend
 261                  *             +BBBBBBBBBBBBBBBBB+
 262                  *             +---------+
 263                  *             fbno   fend
 264                  *
 265                  * Case 7:
 266                  *             bbno           bend
 267                  *             +BBBBBBBBBBBBBBBBB+
 268                  *    +------------------+
 269                  *    fbno            fend
 270                  *
 271                  */
 272                 busyp->bno = fend;
 273         } else if (bbno < fbno) {
 274                 /*
 275                  * Case 8:
 276                  *    bbno           bend
 277                  *    +BBBBBBBBBBBBBBBBB+
 278                  *        +-------------+
 279                  *        fbno       fend
 280                  *
 281                  * Case 9:
 282                  *    bbno           bend
 283                  *    +BBBBBBBBBBBBBBBBB+
 284                  *        +----------------------+
 285                  *        fbno                fend
 286                  */
 287                 busyp->length = fbno - busyp->bno;
 288         } else {
 289                 ASSERT(0);
 290         }
 291
 292         trace_xfs_extent_busy_reuse(mp, pag->pag_agno, fbno, flen);
 293         return true;
 294
 295 out_force_log:
 296         spin_unlock(&pag->pagb_lock);
 297         xfs_log_force(mp, XFS_LOG_SYNC);
 298         trace_xfs_extent_busy_force(mp, pag->pag_agno, fbno, flen);
 299         spin_lock(&pag->pagb_lock);
 300         return false;
 301 }
 302
 303
 304 /*
 305  * For a given extent [fbno, flen], make sure we can reuse it safely.
 306  */
 307 void
 308 xfs_extent_busy_reuse(
 309         struct xfs_mount        *mp,
 310         xfs_agnumber_t          agno,
 311         xfs_agblock_t           fbno,
 312         xfs_extlen_t            flen,
 313         bool                    userdata)
 314 {
 315         struct xfs_perag        *pag;
 316         struct rb_node          *rbp;
 317
 318         ASSERT(flen > 0);
 319
 320         pag = xfs_perag_get(mp, agno);
 321         spin_lock(&pag->pagb_lock);
 322 restart:
 323         rbp = pag->pagb_tree.rb_node;
 324         while (rbp) {
 325                 struct xfs_extent_busy *busyp =
 326                         rb_entry(rbp, struct xfs_extent_busy, rb_node);
 327                 xfs_agblock_t   bbno = busyp->bno;
 328                 xfs_agblock_t   bend = bbno + busyp->length;
 329
 330                 if (fbno + flen <= bbno) {
 331                         rbp = rbp->rb_left;
 332                         continue;
 333                 } else if (fbno >= bend) {
 334                         rbp = rbp->rb_right;
 335                         continue;
 336                 }
 337
 338                 if (!xfs_extent_busy_update_extent(mp, pag, busyp, fbno, flen,
 339                                                   userdata))
 340                         goto restart;
 341         }
 342         spin_unlock(&pag->pagb_lock);
 343         xfs_perag_put(pag);
 344 }
 345
 346 /*
 347  * For a given extent [fbno, flen], search the busy extent list to find a
 348  * subset of the extent that is not busy.  If *rlen is smaller than
 349  * args->minlen no suitable extent could be found, and the higher level
 350  * code needs to force out the log and retry the allocation.
 351  */
 352 void
 353 xfs_extent_busy_trim(
 354         struct xfs_alloc_arg    *args,
 355         xfs_agblock_t           bno,
 356         xfs_extlen_t            len,
 357         xfs_agblock_t           *rbno,
 358         xfs_extlen_t            *rlen)
 359 {
 360         xfs_agblock_t           fbno;
 361         xfs_extlen_t            flen;
 362         struct rb_node          *rbp;
 363
 364         ASSERT(len > 0);
 365
 366         spin_lock(&args->pag->pagb_lock);
 367 restart:
 368         fbno = bno;
 369         flen = len;
 370         rbp = args->pag->pagb_tree.rb_node;
 371         while (rbp && flen >= args->minlen) {
 372                 struct xfs_extent_busy *busyp =
 373                         rb_entry(rbp, struct xfs_extent_busy, rb_node);
 374                 xfs_agblock_t   fend = fbno + flen;
 375                 xfs_agblock_t   bbno = busyp->bno;
 376                 xfs_agblock_t   bend = bbno + busyp->length;
 377
 378                 if (fend <= bbno) {
 379                         rbp = rbp->rb_left;
 380                         continue;
 381                 } else if (fbno >= bend) {
 382                         rbp = rbp->rb_right;
 383                         continue;
 384                 }
 385
 386                 /*
 387                  * If this is a metadata allocation, try to reuse the busy
 388                  * extent instead of trimming the allocation.
 389                  */
 390                 if (!args->userdata &&
 391                     !(busyp->flags & XFS_EXTENT_BUSY_DISCARDED)) {
 392                         if (!xfs_extent_busy_update_extent(args->mp, args->pag,
 393                                                           busyp, fbno, flen,
 394                                                           false))
 395                                 goto restart;
 396                         continue;
 397                 }
 398
 399                 if (bbno <= fbno) {
 400                         /* start overlap */
 401
 402                         /*
 403                          * Case 1:
 404                          *    bbno           bend
 405                          *    +BBBBBBBBBBBBBBBBB+
 406                          *        +---------+
 407                          *        fbno   fend
 408                          *
 409                          * Case 2:
 410                          *    bbno           bend
 411                          *    +BBBBBBBBBBBBBBBBB+
 412                          *    +-------------+
 413                          *    fbno       fend
 414                          *
 415                          * Case 3:
 416                          *    bbno           bend
 417                          *    +BBBBBBBBBBBBBBBBB+
 418                          *        +-------------+
 419                          *        fbno       fend
 420                          *
 421                          * Case 4:
 422                          *    bbno           bend
 423                          *    +BBBBBBBBBBBBBBBBB+
 424                          *    +-----------------+
 425                          *    fbno           fend
 426                          *
 427                          * No unbusy region in extent, return failure.
 428                          */
 429                         if (fend <= bend)
 430                                 goto fail;
 431
 432                         /*
 433                          * Case 5:
 434                          *    bbno           bend
 435                          *    +BBBBBBBBBBBBBBBBB+
 436                          *        +----------------------+
 437                          *        fbno                fend
 438                          *
 439                          * Case 6:
 440                          *    bbno           bend
 441                          *    +BBBBBBBBBBBBBBBBB+
 442                          *    +--------------------------+
 443                          *    fbno                    fend
 444                          *
 445                          * Needs to be trimmed to:
 446                          *                       +-------+
 447                          *                       fbno fend
 448                          */
 449                         fbno = bend;
 450                 } else if (bend >= fend) {
 451                         /* end overlap */
 452
 453                         /*
 454                          * Case 7:
 455                          *             bbno           bend
 456                          *             +BBBBBBBBBBBBBBBBB+
 457                          *    +------------------+
 458                          *    fbno            fend
 459                          *
 460                          * Case 8:
 461                          *             bbno           bend
 462                          *             +BBBBBBBBBBBBBBBBB+
 463                          *    +--------------------------+
 464                          *    fbno                    fend
 465                          *
 466                          * Needs to be trimmed to:
 467                          *    +-------+
 468                          *    fbno fend
 469                          */
 470                         fend = bbno;
 471                 } else {
 472                         /* middle overlap */
 473
 474                         /*
 475                          * Case 9:
 476                          *             bbno           bend
 477                          *             +BBBBBBBBBBBBBBBBB+
 478                          *    +-----------------------------------+
 479                          *    fbno                             fend
 480                          *
 481                          * Can be trimmed to:
 482                          *    +-------+        OR         +-------+
 483                          *    fbno fend                   fbno fend
 484                          *
 485                          * Backward allocation leads to significant
 486                          * fragmentation of directories, which degrades
 487                          * directory performance, therefore we always want to
 488                          * choose the option that produces forward allocation
 489                          * patterns.
 490                          * Preferring the lower bno extent will make the next
 491                          * request use "fend" as the start of the next
 492                          * allocation;  if the segment is no longer busy at
 493                          * that point, we'll get a contiguous allocation, but
 494                          * even if it is still busy, we will get a forward
 495                          * allocation.
 496                          * We try to avoid choosing the segment at "bend",
 497                          * because that can lead to the next allocation
 498                          * taking the segment at "fbno", which would be a
 499                          * backward allocation.  We only use the segment at
 500                          * "fbno" if it is much larger than the current
 501                          * requested size, because in that case there's a
 502                          * good chance subsequent allocations will be
 503                          * contiguous.
 504                          */
 505                         if (bbno - fbno >= args->maxlen) {
 506                                 /* left candidate fits perfect */
 507                                 fend = bbno;
 508                         } else if (fend - bend >= args->maxlen * 4) {
 509                                 /* right candidate has enough free space */
 510                                 fbno = bend;
 511                         } else if (bbno - fbno >= args->minlen) {
 512                                 /* left candidate fits minimum requirement */
 513                                 fend = bbno;
 514                         } else {
 515                                 goto fail;
 516                         }
 517                 }
 518
 519                 flen = fend - fbno;
 520         }
 521         spin_unlock(&args->pag->pagb_lock);
 522
 523         if (fbno != bno || flen != len) {
 524                 trace_xfs_extent_busy_trim(args->mp, args->agno, bno, len,
 525                                           fbno, flen);
 526         }
 527         *rbno = fbno;
 528         *rlen = flen;
 529         return;
 530 fail:
 531         /*
 532          * Return a zero extent length as failure indications.  All callers
 533          * re-check if the trimmed extent satisfies the minlen requirement.
 534          */
 535         spin_unlock(&args->pag->pagb_lock);
 536         trace_xfs_extent_busy_trim(args->mp, args->agno, bno, len, fbno, 0);
 537         *rbno = fbno;
 538         *rlen = 0;
 539 }
 540
 541 STATIC void
 542 xfs_extent_busy_clear_one(
 543         struct xfs_mount        *mp,
 544         struct xfs_perag        *pag,
 545         struct xfs_extent_busy  *busyp)
 546 {
 547         if (busyp->length) {
 548                 trace_xfs_extent_busy_clear(mp, busyp->agno, busyp->bno,
 549                                                 busyp->length);
 550                 rb_erase(&busyp->rb_node, &pag->pagb_tree);
 551         }
 552
 553         list_del_init(&busyp->list);
 554         kmem_free(busyp);
 555 }
 556
 557 /*
 558  * Remove all extents on the passed in list from the busy extents tree.
 559  * If do_discard is set skip extents that need to be discarded, and mark
 560  * these as undergoing a discard operation instead.
 561  */
 562 void
 563 xfs_extent_busy_clear(
 564         struct xfs_mount        *mp,
 565         struct list_head        *list,
 566         bool                    do_discard)
 567 {
 568         struct xfs_extent_busy  *busyp, *n;
 569         struct xfs_perag        *pag = NULL;
 570         xfs_agnumber_t          agno = NULLAGNUMBER;
 571
 572         list_for_each_entry_safe(busyp, n, list, list) {
 573                 if (busyp->agno != agno) {
 574                         if (pag) {
 575                                 spin_unlock(&pag->pagb_lock);
 576                                 xfs_perag_put(pag);
 577                         }
 578                         pag = xfs_perag_get(mp, busyp->agno);
 579                         spin_lock(&pag->pagb_lock);
 580                         agno = busyp->agno;
 581                 }
 582
 583                 if (do_discard && busyp->length &&
 584                     !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD))
 585                         busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
 586                 else
 587                         xfs_extent_busy_clear_one(mp, pag, busyp);
 588         }
 589
 590         if (pag) {
 591                 spin_unlock(&pag->pagb_lock);
 592                 xfs_perag_put(pag);
 593         }
 594 }
 595
 596 /*
 597  * Callback for list_sort to sort busy extents by the AG they reside in.
 598  */
 599 int
 600 xfs_extent_busy_ag_cmp(
 601         void                    *priv,
 602         struct list_head        *a,
 603         struct list_head        *b)
 604 {
 605         return container_of(a, struct xfs_extent_busy, list)->agno -
 606                 container_of(b, struct xfs_extent_busy, list)->agno;
 607 }