fs/xfs/xfs_extent_busy.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
   4  * Copyright (c) 2010 David Chinner.
   5  * Copyright (c) 2011 Christoph Hellwig.
   6  * All Rights Reserved.
   7  */
   8 #include "xfs.h"
   9 #include "xfs_fs.h"
  10 #include "xfs_format.h"
  11 #include "xfs_log_format.h"
  12 #include "xfs_shared.h"
  13 #include "xfs_trans_resv.h"
  14 #include "xfs_sb.h"
  15 #include "xfs_mount.h"
  16 #include "xfs_alloc.h"
  17 #include "xfs_extent_busy.h"
  18 #include "xfs_trace.h"
  19 #include "xfs_trans.h"
  20 #include "xfs_log.h"
  21
  22 void
  23 xfs_extent_busy_insert(
  24         struct xfs_trans        *tp,
  25         xfs_agnumber_t          agno,
  26         xfs_agblock_t           bno,
  27         xfs_extlen_t            len,
  28         unsigned int            flags)
  29 {
  30         struct xfs_extent_busy  *new;
  31         struct xfs_extent_busy  *busyp;
  32         struct xfs_perag        *pag;
  33         struct rb_node          **rbp;
  34         struct rb_node          *parent = NULL;
  35
  36         new = kmem_zalloc(sizeof(struct xfs_extent_busy), 0);
  37         new->agno = agno;
  38         new->bno = bno;
  39         new->length = len;
  40         INIT_LIST_HEAD(&new->list);
  41         new->flags = flags;
  42
  43         /* trace before insert to be able to see failed inserts */
  44         trace_xfs_extent_busy(tp->t_mountp, agno, bno, len);
  45
  46         pag = xfs_perag_get(tp->t_mountp, new->agno);
  47         spin_lock(&pag->pagb_lock);
  48         rbp = &pag->pagb_tree.rb_node;
  49         while (*rbp) {
  50                 parent = *rbp;
  51                 busyp = rb_entry(parent, struct xfs_extent_busy, rb_node);
  52
  53                 if (new->bno < busyp->bno) {
  54                         rbp = &(*rbp)->rb_left;
  55                         ASSERT(new->bno + new->length <= busyp->bno);
  56                 } else if (new->bno > busyp->bno) {
  57                         rbp = &(*rbp)->rb_right;
  58                         ASSERT(bno >= busyp->bno + busyp->length);
  59                 } else {
  60                         ASSERT(0);
  61                 }
  62         }
  63
  64         rb_link_node(&new->rb_node, parent, rbp);
  65         rb_insert_color(&new->rb_node, &pag->pagb_tree);
  66
  67         list_add(&new->list, &tp->t_busy);
  68         spin_unlock(&pag->pagb_lock);
  69         xfs_perag_put(pag);
  70 }
  71
  72 /*
  73  * Search for a busy extent within the range of the extent we are about to
  74  * allocate.  You need to be holding the busy extent tree lock when calling
  75  * xfs_extent_busy_search(). This function returns 0 for no overlapping busy
  76  * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
  77  * match. This is done so that a non-zero return indicates an overlap that
  78  * will require a synchronous transaction, but it can still be
  79  * used to distinguish between a partial or exact match.
  80  */
  81 int
  82 xfs_extent_busy_search(
  83         struct xfs_mount        *mp,
  84         xfs_agnumber_t          agno,
  85         xfs_agblock_t           bno,
  86         xfs_extlen_t            len)
  87 {
  88         struct xfs_perag        *pag;
  89         struct rb_node          *rbp;
  90         struct xfs_extent_busy  *busyp;
  91         int                     match = 0;
  92
  93         pag = xfs_perag_get(mp, agno);
  94         spin_lock(&pag->pagb_lock);
  95
  96         rbp = pag->pagb_tree.rb_node;
  97
  98         /* find closest start bno overlap */
  99         while (rbp) {
 100                 busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node);
 101                 if (bno < busyp->bno) {
 102                         /* may overlap, but exact start block is lower */
 103                         if (bno + len > busyp->bno)
 104                                 match = -1;
 105                         rbp = rbp->rb_left;
 106                 } else if (bno > busyp->bno) {
 107                         /* may overlap, but exact start block is higher */
 108                         if (bno < busyp->bno + busyp->length)
 109                                 match = -1;
 110                         rbp = rbp->rb_right;
 111                 } else {
 112                         /* bno matches busyp, length determines exact match */
 113                         match = (busyp->length == len) ? 1 : -1;
 114                         break;
 115                 }
 116         }
 117         spin_unlock(&pag->pagb_lock);
 118         xfs_perag_put(pag);
 119         return match;
 120 }
 121
 122 /*
 123  * The found free extent [fbno, fend] overlaps part or all of the given busy
 124  * extent.  If the overlap covers the beginning, the end, or all of the busy
 125  * extent, the overlapping portion can be made unbusy and used for the
 126  * allocation.  We can't split a busy extent because we can't modify a
 127  * transaction/CIL context busy list, but we can update an entry's block
 128  * number or length.
 129  *
 130  * Returns true if the extent can safely be reused, or false if the search
 131  * needs to be restarted.
 132  */
 133 STATIC bool
 134 xfs_extent_busy_update_extent(
 135         struct xfs_mount        *mp,
 136         struct xfs_perag        *pag,
 137         struct xfs_extent_busy  *busyp,
 138         xfs_agblock_t           fbno,
 139         xfs_extlen_t            flen,
 140         bool                    userdata) __releases(&pag->pagb_lock)
 141                                           __acquires(&pag->pagb_lock)
 142 {
 143         xfs_agblock_t           fend = fbno + flen;
 144         xfs_agblock_t           bbno = busyp->bno;
 145         xfs_agblock_t           bend = bbno + busyp->length;
 146
 147         /*
 148          * This extent is currently being discarded.  Give the thread
 149          * performing the discard a chance to mark the extent unbusy
 150          * and retry.
 151          */
 152         if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
 153                 spin_unlock(&pag->pagb_lock);
 154                 delay(1);
 155                 spin_lock(&pag->pagb_lock);
 156                 return false;
 157         }
 158
 159         /*
 160          * If there is a busy extent overlapping a user allocation, we have
 161          * no choice but to force the log and retry the search.
 162          *
 163          * Fortunately this does not happen during normal operation, but
 164          * only if the filesystem is very low on space and has to dip into
 165          * the AGFL for normal allocations.
 166          */
 167         if (userdata)
 168                 goto out_force_log;
 169
 170         if (bbno < fbno && bend > fend) {
 171                 /*
 172                  * Case 1:
 173                  *    bbno           bend
 174                  *    +BBBBBBBBBBBBBBBBB+
 175                  *        +---------+
 176                  *        fbno   fend
 177                  */
 178
 179                 /*
 180                  * We would have to split the busy extent to be able to track
 181                  * it correct, which we cannot do because we would have to
 182                  * modify the list of busy extents attached to the transaction
 183                  * or CIL context, which is immutable.
 184                  *
 185                  * Force out the log to clear the busy extent and retry the
 186                  * search.
 187                  */
 188                 goto out_force_log;
 189         } else if (bbno >= fbno && bend <= fend) {
 190                 /*
 191                  * Case 2:
 192                  *    bbno           bend
 193                  *    +BBBBBBBBBBBBBBBBB+
 194                  *    +-----------------+
 195                  *    fbno           fend
 196                  *
 197                  * Case 3:
 198                  *    bbno           bend
 199                  *    +BBBBBBBBBBBBBBBBB+
 200                  *    +--------------------------+
 201                  *    fbno                    fend
 202                  *
 203                  * Case 4:
 204                  *             bbno           bend
 205                  *             +BBBBBBBBBBBBBBBBB+
 206                  *    +--------------------------+
 207                  *    fbno                    fend
 208                  *
 209                  * Case 5:
 210                  *             bbno           bend
 211                  *             +BBBBBBBBBBBBBBBBB+
 212                  *    +-----------------------------------+
 213                  *    fbno                             fend
 214                  *
 215                  */
 216
 217                 /*
 218                  * The busy extent is fully covered by the extent we are
 219                  * allocating, and can simply be removed from the rbtree.
 220                  * However we cannot remove it from the immutable list
 221                  * tracking busy extents in the transaction or CIL context,
 222                  * so set the length to zero to mark it invalid.
 223                  *
 224                  * We also need to restart the busy extent search from the
 225                  * tree root, because erasing the node can rearrange the
 226                  * tree topology.
 227                  */
 228                 rb_erase(&busyp->rb_node, &pag->pagb_tree);
 229                 busyp->length = 0;
 230                 return false;
 231         } else if (fend < bend) {
 232                 /*
 233                  * Case 6:
 234                  *              bbno           bend
 235                  *             +BBBBBBBBBBBBBBBBB+
 236                  *             +---------+
 237                  *             fbno   fend
 238                  *
 239                  * Case 7:
 240                  *             bbno           bend
 241                  *             +BBBBBBBBBBBBBBBBB+
 242                  *    +------------------+
 243                  *    fbno            fend
 244                  *
 245                  */
 246                 busyp->bno = fend;
 247         } else if (bbno < fbno) {
 248                 /*
 249                  * Case 8:
 250                  *    bbno           bend
 251                  *    +BBBBBBBBBBBBBBBBB+
 252                  *        +-------------+
 253                  *        fbno       fend
 254                  *
 255                  * Case 9:
 256                  *    bbno           bend
 257                  *    +BBBBBBBBBBBBBBBBB+
 258                  *        +----------------------+
 259                  *        fbno                fend
 260                  */
 261                 busyp->length = fbno - busyp->bno;
 262         } else {
 263                 ASSERT(0);
 264         }
 265
 266         trace_xfs_extent_busy_reuse(mp, pag->pag_agno, fbno, flen);
 267         return true;
 268
 269 out_force_log:
 270         spin_unlock(&pag->pagb_lock);
 271         xfs_log_force(mp, XFS_LOG_SYNC);
 272         trace_xfs_extent_busy_force(mp, pag->pag_agno, fbno, flen);
 273         spin_lock(&pag->pagb_lock);
 274         return false;
 275 }
 276
 277
 278 /*
 279  * For a given extent [fbno, flen], make sure we can reuse it safely.
 280  */
 281 void
 282 xfs_extent_busy_reuse(
 283         struct xfs_mount        *mp,
 284         xfs_agnumber_t          agno,
 285         xfs_agblock_t           fbno,
 286         xfs_extlen_t            flen,
 287         bool                    userdata)
 288 {
 289         struct xfs_perag        *pag;
 290         struct rb_node          *rbp;
 291
 292         ASSERT(flen > 0);
 293
 294         pag = xfs_perag_get(mp, agno);
 295         spin_lock(&pag->pagb_lock);
 296 restart:
 297         rbp = pag->pagb_tree.rb_node;
 298         while (rbp) {
 299                 struct xfs_extent_busy *busyp =
 300                         rb_entry(rbp, struct xfs_extent_busy, rb_node);
 301                 xfs_agblock_t   bbno = busyp->bno;
 302                 xfs_agblock_t   bend = bbno + busyp->length;
 303
 304                 if (fbno + flen <= bbno) {
 305                         rbp = rbp->rb_left;
 306                         continue;
 307                 } else if (fbno >= bend) {
 308                         rbp = rbp->rb_right;
 309                         continue;
 310                 }
 311
 312                 if (!xfs_extent_busy_update_extent(mp, pag, busyp, fbno, flen,
 313                                                   userdata))
 314                         goto restart;
 315         }
 316         spin_unlock(&pag->pagb_lock);
 317         xfs_perag_put(pag);
 318 }
 319
 320 /*
 321  * For a given extent [fbno, flen], search the busy extent list to find a
 322  * subset of the extent that is not busy.  If *rlen is smaller than
 323  * args->minlen no suitable extent could be found, and the higher level
 324  * code needs to force out the log and retry the allocation.
 325  *
 326  * Return the current busy generation for the AG if the extent is busy. This
 327  * value can be used to wait for at least one of the currently busy extents
 328  * to be cleared. Note that the busy list is not guaranteed to be empty after
 329  * the gen is woken. The state of a specific extent must always be confirmed
 330  * with another call to xfs_extent_busy_trim() before it can be used.
 331  */
 332 bool
 333 xfs_extent_busy_trim(
 334         struct xfs_alloc_arg    *args,
 335         xfs_agblock_t           *bno,
 336         xfs_extlen_t            *len,
 337         unsigned                *busy_gen)
 338 {
 339         xfs_agblock_t           fbno;
 340         xfs_extlen_t            flen;
 341         struct rb_node          *rbp;
 342         bool                    ret = false;
 343
 344         ASSERT(*len > 0);
 345
 346         spin_lock(&args->pag->pagb_lock);
 347 restart:
 348         fbno = *bno;
 349         flen = *len;
 350         rbp = args->pag->pagb_tree.rb_node;
 351         while (rbp && flen >= args->minlen) {
 352                 struct xfs_extent_busy *busyp =
 353                         rb_entry(rbp, struct xfs_extent_busy, rb_node);
 354                 xfs_agblock_t   fend = fbno + flen;
 355                 xfs_agblock_t   bbno = busyp->bno;
 356                 xfs_agblock_t   bend = bbno + busyp->length;
 357
 358                 if (fend <= bbno) {
 359                         rbp = rbp->rb_left;
 360                         continue;
 361                 } else if (fbno >= bend) {
 362                         rbp = rbp->rb_right;
 363                         continue;
 364                 }
 365
 366                 /*
 367                  * If this is a metadata allocation, try to reuse the busy
 368                  * extent instead of trimming the allocation.
 369                  */
 370                 if (!xfs_alloc_is_userdata(args->datatype) &&
 371                     !(busyp->flags & XFS_EXTENT_BUSY_DISCARDED)) {
 372                         if (!xfs_extent_busy_update_extent(args->mp, args->pag,
 373                                                           busyp, fbno, flen,
 374                                                           false))
 375                                 goto restart;
 376                         continue;
 377                 }
 378
 379                 if (bbno <= fbno) {
 380                         /* start overlap */
 381
 382                         /*
 383                          * Case 1:
 384                          *    bbno           bend
 385                          *    +BBBBBBBBBBBBBBBBB+
 386                          *        +---------+
 387                          *        fbno   fend
 388                          *
 389                          * Case 2:
 390                          *    bbno           bend
 391                          *    +BBBBBBBBBBBBBBBBB+
 392                          *    +-------------+
 393                          *    fbno       fend
 394                          *
 395                          * Case 3:
 396                          *    bbno           bend
 397                          *    +BBBBBBBBBBBBBBBBB+
 398                          *        +-------------+
 399                          *        fbno       fend
 400                          *
 401                          * Case 4:
 402                          *    bbno           bend
 403                          *    +BBBBBBBBBBBBBBBBB+
 404                          *    +-----------------+
 405                          *    fbno           fend
 406                          *
 407                          * No unbusy region in extent, return failure.
 408                          */
 409                         if (fend <= bend)
 410                                 goto fail;
 411
 412                         /*
 413                          * Case 5:
 414                          *    bbno           bend
 415                          *    +BBBBBBBBBBBBBBBBB+
 416                          *        +----------------------+
 417                          *        fbno                fend
 418                          *
 419                          * Case 6:
 420                          *    bbno           bend
 421                          *    +BBBBBBBBBBBBBBBBB+
 422                          *    +--------------------------+
 423                          *    fbno                    fend
 424                          *
 425                          * Needs to be trimmed to:
 426                          *                       +-------+
 427                          *                       fbno fend
 428                          */
 429                         fbno = bend;
 430                 } else if (bend >= fend) {
 431                         /* end overlap */
 432
 433                         /*
 434                          * Case 7:
 435                          *             bbno           bend
 436                          *             +BBBBBBBBBBBBBBBBB+
 437                          *    +------------------+
 438                          *    fbno            fend
 439                          *
 440                          * Case 8:
 441                          *             bbno           bend
 442                          *             +BBBBBBBBBBBBBBBBB+
 443                          *    +--------------------------+
 444                          *    fbno                    fend
 445                          *
 446                          * Needs to be trimmed to:
 447                          *    +-------+
 448                          *    fbno fend
 449                          */
 450                         fend = bbno;
 451                 } else {
 452                         /* middle overlap */
 453
 454                         /*
 455                          * Case 9:
 456                          *             bbno           bend
 457                          *             +BBBBBBBBBBBBBBBBB+
 458                          *    +-----------------------------------+
 459                          *    fbno                             fend
 460                          *
 461                          * Can be trimmed to:
 462                          *    +-------+        OR         +-------+
 463                          *    fbno fend                   fbno fend
 464                          *
 465                          * Backward allocation leads to significant
 466                          * fragmentation of directories, which degrades
 467                          * directory performance, therefore we always want to
 468                          * choose the option that produces forward allocation
 469                          * patterns.
 470                          * Preferring the lower bno extent will make the next
 471                          * request use "fend" as the start of the next
 472                          * allocation;  if the segment is no longer busy at
 473                          * that point, we'll get a contiguous allocation, but
 474                          * even if it is still busy, we will get a forward
 475                          * allocation.
 476                          * We try to avoid choosing the segment at "bend",
 477                          * because that can lead to the next allocation
 478                          * taking the segment at "fbno", which would be a
 479                          * backward allocation.  We only use the segment at
 480                          * "fbno" if it is much larger than the current
 481                          * requested size, because in that case there's a
 482                          * good chance subsequent allocations will be
 483                          * contiguous.
 484                          */
 485                         if (bbno - fbno >= args->maxlen) {
 486                                 /* left candidate fits perfect */
 487                                 fend = bbno;
 488                         } else if (fend - bend >= args->maxlen * 4) {
 489                                 /* right candidate has enough free space */
 490                                 fbno = bend;
 491                         } else if (bbno - fbno >= args->minlen) {
 492                                 /* left candidate fits minimum requirement */
 493                                 fend = bbno;
 494                         } else {
 495                                 goto fail;
 496                         }
 497                 }
 498
 499                 flen = fend - fbno;
 500         }
 501 out:
 502
 503         if (fbno != *bno || flen != *len) {
 504                 trace_xfs_extent_busy_trim(args->mp, args->agno, *bno, *len,
 505                                           fbno, flen);
 506                 *bno = fbno;
 507                 *len = flen;
 508                 *busy_gen = args->pag->pagb_gen;
 509                 ret = true;
 510         }
 511         spin_unlock(&args->pag->pagb_lock);
 512         return ret;
 513 fail:
 514         /*
 515          * Return a zero extent length as failure indications.  All callers
 516          * re-check if the trimmed extent satisfies the minlen requirement.
 517          */
 518         flen = 0;
 519         goto out;
 520 }
 521
 522 STATIC void
 523 xfs_extent_busy_clear_one(
 524         struct xfs_mount        *mp,
 525         struct xfs_perag        *pag,
 526         struct xfs_extent_busy  *busyp)
 527 {
 528         if (busyp->length) {
 529                 trace_xfs_extent_busy_clear(mp, busyp->agno, busyp->bno,
 530                                                 busyp->length);
 531                 rb_erase(&busyp->rb_node, &pag->pagb_tree);
 532         }
 533
 534         list_del_init(&busyp->list);
 535         kmem_free(busyp);
 536 }
 537
 538 static void
 539 xfs_extent_busy_put_pag(
 540         struct xfs_perag        *pag,
 541         bool                    wakeup)
 542                 __releases(pag->pagb_lock)
 543 {
 544         if (wakeup) {
 545                 pag->pagb_gen++;
 546                 wake_up_all(&pag->pagb_wait);
 547         }
 548
 549         spin_unlock(&pag->pagb_lock);
 550         xfs_perag_put(pag);
 551 }
 552
 553 /*
 554  * Remove all extents on the passed in list from the busy extents tree.
 555  * If do_discard is set skip extents that need to be discarded, and mark
 556  * these as undergoing a discard operation instead.
 557  */
 558 void
 559 xfs_extent_busy_clear(
 560         struct xfs_mount        *mp,
 561         struct list_head        *list,
 562         bool                    do_discard)
 563 {
 564         struct xfs_extent_busy  *busyp, *n;
 565         struct xfs_perag        *pag = NULL;
 566         xfs_agnumber_t          agno = NULLAGNUMBER;
 567         bool                    wakeup = false;
 568
 569         list_for_each_entry_safe(busyp, n, list, list) {
 570                 if (busyp->agno != agno) {
 571                         if (pag)
 572                                 xfs_extent_busy_put_pag(pag, wakeup);
 573                         agno = busyp->agno;
 574                         pag = xfs_perag_get(mp, agno);
 575                         spin_lock(&pag->pagb_lock);
 576                         wakeup = false;
 577                 }
 578
 579                 if (do_discard && busyp->length &&
 580                     !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD)) {
 581                         busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
 582                 } else {
 583                         xfs_extent_busy_clear_one(mp, pag, busyp);
 584                         wakeup = true;
 585                 }
 586         }
 587
 588         if (pag)
 589                 xfs_extent_busy_put_pag(pag, wakeup);
 590 }
 591
 592 /*
 593  * Flush out all busy extents for this AG.
 594  */
 595 void
 596 xfs_extent_busy_flush(
 597         struct xfs_mount        *mp,
 598         struct xfs_perag        *pag,
 599         unsigned                busy_gen)
 600 {
 601         DEFINE_WAIT             (wait);
 602         int                     error;
 603
 604         error = xfs_log_force(mp, XFS_LOG_SYNC);
 605         if (error)
 606                 return;
 607
 608         do {
 609                 prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
 610                 if  (busy_gen != READ_ONCE(pag->pagb_gen))
 611                         break;
 612                 schedule();
 613         } while (1);
 614
 615         finish_wait(&pag->pagb_wait, &wait);
 616 }
 617
 618 void
 619 xfs_extent_busy_wait_all(
 620         struct xfs_mount        *mp)
 621 {
 622         DEFINE_WAIT             (wait);
 623         xfs_agnumber_t          agno;
 624
 625         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
 626                 struct xfs_perag *pag = xfs_perag_get(mp, agno);
 627
 628                 do {
 629                         prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
 630                         if  (RB_EMPTY_ROOT(&pag->pagb_tree))
 631                                 break;
 632                         schedule();
 633                 } while (1);
 634                 finish_wait(&pag->pagb_wait, &wait);
 635
 636                 xfs_perag_put(pag);
 637         }
 638 }
 639
 640 /*
 641  * Callback for list_sort to sort busy extents by the AG they reside in.
 642  */
 643 int
 644 xfs_extent_busy_ag_cmp(
 645         void                    *priv,
 646         struct list_head        *l1,
 647         struct list_head        *l2)
 648 {
 649         struct xfs_extent_busy  *b1 =
 650                 container_of(l1, struct xfs_extent_busy, list);
 651         struct xfs_extent_busy  *b2 =
 652                 container_of(l2, struct xfs_extent_busy, list);
 653         s32 diff;
 654
 655         diff = b1->agno - b2->agno;
 656         if (!diff)
 657                 diff = b1->bno - b2->bno;
 658         return diff;
 659 }