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