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_mount.h"
  15 #include "xfs_alloc.h"
  16 #include "xfs_extent_busy.h"
  17 #include "xfs_trace.h"
  18 #include "xfs_trans.h"
  19 #include "xfs_log.h"
  20 #include "xfs_ag.h"
  21
  22 static void
  23 xfs_extent_busy_insert_list(
  24         struct xfs_perag        *pag,
  25         xfs_agblock_t           bno,
  26         xfs_extlen_t            len,
  27         unsigned int            flags,
  28         struct list_head        *busy_list)
  29 {
  30         struct xfs_extent_busy  *new;
  31         struct xfs_extent_busy  *busyp;
  32         struct rb_node          **rbp;
  33         struct rb_node          *parent = NULL;
  34
  35         new = kmem_zalloc(sizeof(struct xfs_extent_busy), 0);
  36         new->agno = pag->pag_agno;
  37         new->bno = bno;
  38         new->length = len;
  39         INIT_LIST_HEAD(&new->list);
  40         new->flags = flags;
  41
  42         /* trace before insert to be able to see failed inserts */
  43         trace_xfs_extent_busy(pag->pag_mount, pag->pag_agno, bno, len);
  44
  45         spin_lock(&pag->pagb_lock);
  46         rbp = &pag->pagb_tree.rb_node;
  47         while (*rbp) {
  48                 parent = *rbp;
  49                 busyp = rb_entry(parent, struct xfs_extent_busy, rb_node);
  50
  51                 if (new->bno < busyp->bno) {
  52                         rbp = &(*rbp)->rb_left;
  53                         ASSERT(new->bno + new->length <= busyp->bno);
  54                 } else if (new->bno > busyp->bno) {
  55                         rbp = &(*rbp)->rb_right;
  56                         ASSERT(bno >= busyp->bno + busyp->length);
  57                 } else {
  58                         ASSERT(0);
  59                 }
  60         }
  61
  62         rb_link_node(&new->rb_node, parent, rbp);
  63         rb_insert_color(&new->rb_node, &pag->pagb_tree);
  64
  65         list_add(&new->list, busy_list);
  66         spin_unlock(&pag->pagb_lock);
  67 }
  68
  69 void
  70 xfs_extent_busy_insert(
  71         struct xfs_trans        *tp,
  72         struct xfs_perag        *pag,
  73         xfs_agblock_t           bno,
  74         xfs_extlen_t            len,
  75         unsigned int            flags)
  76 {
  77         xfs_extent_busy_insert_list(pag, bno, len, flags, &tp->t_busy);
  78 }
  79
  80 void
  81 xfs_extent_busy_insert_discard(
  82         struct xfs_perag        *pag,
  83         xfs_agblock_t           bno,
  84         xfs_extlen_t            len,
  85         struct list_head        *busy_list)
  86 {
  87         xfs_extent_busy_insert_list(pag, bno, len, XFS_EXTENT_BUSY_DISCARDED,
  88                         busy_list);
  89 }
  90
  91 /*
  92  * Search for a busy extent within the range of the extent we are about to
  93  * allocate.  You need to be holding the busy extent tree lock when calling
  94  * xfs_extent_busy_search(). This function returns 0 for no overlapping busy
  95  * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
  96  * match. This is done so that a non-zero return indicates an overlap that
  97  * will require a synchronous transaction, but it can still be
  98  * used to distinguish between a partial or exact match.
  99  */
 100 int
 101 xfs_extent_busy_search(
 102         struct xfs_mount        *mp,
 103         struct xfs_perag        *pag,
 104         xfs_agblock_t           bno,
 105         xfs_extlen_t            len)
 106 {
 107         struct rb_node          *rbp;
 108         struct xfs_extent_busy  *busyp;
 109         int                     match = 0;
 110
 111         /* find closest start bno overlap */
 112         spin_lock(&pag->pagb_lock);
 113         rbp = pag->pagb_tree.rb_node;
 114         while (rbp) {
 115                 busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node);
 116                 if (bno < busyp->bno) {
 117                         /* may overlap, but exact start block is lower */
 118                         if (bno + len > busyp->bno)
 119                                 match = -1;
 120                         rbp = rbp->rb_left;
 121                 } else if (bno > busyp->bno) {
 122                         /* may overlap, but exact start block is higher */
 123                         if (bno < busyp->bno + busyp->length)
 124                                 match = -1;
 125                         rbp = rbp->rb_right;
 126                 } else {
 127                         /* bno matches busyp, length determines exact match */
 128                         match = (busyp->length == len) ? 1 : -1;
 129                         break;
 130                 }
 131         }
 132         spin_unlock(&pag->pagb_lock);
 133         return match;
 134 }
 135
 136 /*
 137  * The found free extent [fbno, fend] overlaps part or all of the given busy
 138  * extent.  If the overlap covers the beginning, the end, or all of the busy
 139  * extent, the overlapping portion can be made unbusy and used for the
 140  * allocation.  We can't split a busy extent because we can't modify a
 141  * transaction/CIL context busy list, but we can update an entry's block
 142  * number or length.
 143  *
 144  * Returns true if the extent can safely be reused, or false if the search
 145  * needs to be restarted.
 146  */
 147 STATIC bool
 148 xfs_extent_busy_update_extent(
 149         struct xfs_mount        *mp,
 150         struct xfs_perag        *pag,
 151         struct xfs_extent_busy  *busyp,
 152         xfs_agblock_t           fbno,
 153         xfs_extlen_t            flen,
 154         bool                    userdata) __releases(&pag->pagb_lock)
 155                                           __acquires(&pag->pagb_lock)
 156 {
 157         xfs_agblock_t           fend = fbno + flen;
 158         xfs_agblock_t           bbno = busyp->bno;
 159         xfs_agblock_t           bend = bbno + busyp->length;
 160
 161         /*
 162          * This extent is currently being discarded.  Give the thread
 163          * performing the discard a chance to mark the extent unbusy
 164          * and retry.
 165          */
 166         if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
 167                 spin_unlock(&pag->pagb_lock);
 168                 delay(1);
 169                 spin_lock(&pag->pagb_lock);
 170                 return false;
 171         }
 172
 173         /*
 174          * If there is a busy extent overlapping a user allocation, we have
 175          * no choice but to force the log and retry the search.
 176          *
 177          * Fortunately this does not happen during normal operation, but
 178          * only if the filesystem is very low on space and has to dip into
 179          * the AGFL for normal allocations.
 180          */
 181         if (userdata)
 182                 goto out_force_log;
 183
 184         if (bbno < fbno && bend > fend) {
 185                 /*
 186                  * Case 1:
 187                  *    bbno           bend
 188                  *    +BBBBBBBBBBBBBBBBB+
 189                  *        +---------+
 190                  *        fbno   fend
 191                  */
 192
 193                 /*
 194                  * We would have to split the busy extent to be able to track
 195                  * it correct, which we cannot do because we would have to
 196                  * modify the list of busy extents attached to the transaction
 197                  * or CIL context, which is immutable.
 198                  *
 199                  * Force out the log to clear the busy extent and retry the
 200                  * search.
 201                  */
 202                 goto out_force_log;
 203         } else if (bbno >= fbno && bend <= fend) {
 204                 /*
 205                  * Case 2:
 206                  *    bbno           bend
 207                  *    +BBBBBBBBBBBBBBBBB+
 208                  *    +-----------------+
 209                  *    fbno           fend
 210                  *
 211                  * Case 3:
 212                  *    bbno           bend
 213                  *    +BBBBBBBBBBBBBBBBB+
 214                  *    +--------------------------+
 215                  *    fbno                    fend
 216                  *
 217                  * Case 4:
 218                  *             bbno           bend
 219                  *             +BBBBBBBBBBBBBBBBB+
 220                  *    +--------------------------+
 221                  *    fbno                    fend
 222                  *
 223                  * Case 5:
 224                  *             bbno           bend
 225                  *             +BBBBBBBBBBBBBBBBB+
 226                  *    +-----------------------------------+
 227                  *    fbno                             fend
 228                  *
 229                  */
 230
 231                 /*
 232                  * The busy extent is fully covered by the extent we are
 233                  * allocating, and can simply be removed from the rbtree.
 234                  * However we cannot remove it from the immutable list
 235                  * tracking busy extents in the transaction or CIL context,
 236                  * so set the length to zero to mark it invalid.
 237                  *
 238                  * We also need to restart the busy extent search from the
 239                  * tree root, because erasing the node can rearrange the
 240                  * tree topology.
 241                  */
 242                 rb_erase(&busyp->rb_node, &pag->pagb_tree);
 243                 busyp->length = 0;
 244                 return false;
 245         } else if (fend < bend) {
 246                 /*
 247                  * Case 6:
 248                  *              bbno           bend
 249                  *             +BBBBBBBBBBBBBBBBB+
 250                  *             +---------+
 251                  *             fbno   fend
 252                  *
 253                  * Case 7:
 254                  *             bbno           bend
 255                  *             +BBBBBBBBBBBBBBBBB+
 256                  *    +------------------+
 257                  *    fbno            fend
 258                  *
 259                  */
 260                 busyp->bno = fend;
 261                 busyp->length = bend - fend;
 262         } else if (bbno < fbno) {
 263                 /*
 264                  * Case 8:
 265                  *    bbno           bend
 266                  *    +BBBBBBBBBBBBBBBBB+
 267                  *        +-------------+
 268                  *        fbno       fend
 269                  *
 270                  * Case 9:
 271                  *    bbno           bend
 272                  *    +BBBBBBBBBBBBBBBBB+
 273                  *        +----------------------+
 274                  *        fbno                fend
 275                  */
 276                 busyp->length = fbno - busyp->bno;
 277         } else {
 278                 ASSERT(0);
 279         }
 280
 281         trace_xfs_extent_busy_reuse(mp, pag->pag_agno, fbno, flen);
 282         return true;
 283
 284 out_force_log:
 285         spin_unlock(&pag->pagb_lock);
 286         xfs_log_force(mp, XFS_LOG_SYNC);
 287         trace_xfs_extent_busy_force(mp, pag->pag_agno, fbno, flen);
 288         spin_lock(&pag->pagb_lock);
 289         return false;
 290 }
 291
 292
 293 /*
 294  * For a given extent [fbno, flen], make sure we can reuse it safely.
 295  */
 296 void
 297 xfs_extent_busy_reuse(
 298         struct xfs_mount        *mp,
 299         struct xfs_perag        *pag,
 300         xfs_agblock_t           fbno,
 301         xfs_extlen_t            flen,
 302         bool                    userdata)
 303 {
 304         struct rb_node          *rbp;
 305
 306         ASSERT(flen > 0);
 307         spin_lock(&pag->pagb_lock);
 308 restart:
 309         rbp = pag->pagb_tree.rb_node;
 310         while (rbp) {
 311                 struct xfs_extent_busy *busyp =
 312                         rb_entry(rbp, struct xfs_extent_busy, rb_node);
 313                 xfs_agblock_t   bbno = busyp->bno;
 314                 xfs_agblock_t   bend = bbno + busyp->length;
 315
 316                 if (fbno + flen <= bbno) {
 317                         rbp = rbp->rb_left;
 318                         continue;
 319                 } else if (fbno >= bend) {
 320                         rbp = rbp->rb_right;
 321                         continue;
 322                 }
 323
 324                 if (!xfs_extent_busy_update_extent(mp, pag, busyp, fbno, flen,
 325                                                   userdata))
 326                         goto restart;
 327         }
 328         spin_unlock(&pag->pagb_lock);
 329 }
 330
 331 /*
 332  * For a given extent [fbno, flen], search the busy extent list to find a
 333  * subset of the extent that is not busy.  If *rlen is smaller than
 334  * args->minlen no suitable extent could be found, and the higher level
 335  * code needs to force out the log and retry the allocation.
 336  *
 337  * Return the current busy generation for the AG if the extent is busy. This
 338  * value can be used to wait for at least one of the currently busy extents
 339  * to be cleared. Note that the busy list is not guaranteed to be empty after
 340  * the gen is woken. The state of a specific extent must always be confirmed
 341  * with another call to xfs_extent_busy_trim() before it can be used.
 342  */
 343 bool
 344 xfs_extent_busy_trim(
 345         struct xfs_alloc_arg    *args,
 346         xfs_agblock_t           *bno,
 347         xfs_extlen_t            *len,
 348         unsigned                *busy_gen)
 349 {
 350         xfs_agblock_t           fbno;
 351         xfs_extlen_t            flen;
 352         struct rb_node          *rbp;
 353         bool                    ret = false;
 354
 355         ASSERT(*len > 0);
 356
 357         spin_lock(&args->pag->pagb_lock);
 358         fbno = *bno;
 359         flen = *len;
 360         rbp = args->pag->pagb_tree.rb_node;
 361         while (rbp && flen >= args->minlen) {
 362                 struct xfs_extent_busy *busyp =
 363                         rb_entry(rbp, struct xfs_extent_busy, rb_node);
 364                 xfs_agblock_t   fend = fbno + flen;
 365                 xfs_agblock_t   bbno = busyp->bno;
 366                 xfs_agblock_t   bend = bbno + busyp->length;
 367
 368                 if (fend <= bbno) {
 369                         rbp = rbp->rb_left;
 370                         continue;
 371                 } else if (fbno >= bend) {
 372                         rbp = rbp->rb_right;
 373                         continue;
 374                 }
 375
 376                 if (bbno <= fbno) {
 377                         /* start overlap */
 378
 379                         /*
 380                          * Case 1:
 381                          *    bbno           bend
 382                          *    +BBBBBBBBBBBBBBBBB+
 383                          *        +---------+
 384                          *        fbno   fend
 385                          *
 386                          * Case 2:
 387                          *    bbno           bend
 388                          *    +BBBBBBBBBBBBBBBBB+
 389                          *    +-------------+
 390                          *    fbno       fend
 391                          *
 392                          * Case 3:
 393                          *    bbno           bend
 394                          *    +BBBBBBBBBBBBBBBBB+
 395                          *        +-------------+
 396                          *        fbno       fend
 397                          *
 398                          * Case 4:
 399                          *    bbno           bend
 400                          *    +BBBBBBBBBBBBBBBBB+
 401                          *    +-----------------+
 402                          *    fbno           fend
 403                          *
 404                          * No unbusy region in extent, return failure.
 405                          */
 406                         if (fend <= bend)
 407                                 goto fail;
 408
 409                         /*
 410                          * Case 5:
 411                          *    bbno           bend
 412                          *    +BBBBBBBBBBBBBBBBB+
 413                          *        +----------------------+
 414                          *        fbno                fend
 415                          *
 416                          * Case 6:
 417                          *    bbno           bend
 418                          *    +BBBBBBBBBBBBBBBBB+
 419                          *    +--------------------------+
 420                          *    fbno                    fend
 421                          *
 422                          * Needs to be trimmed to:
 423                          *                       +-------+
 424                          *                       fbno fend
 425                          */
 426                         fbno = bend;
 427                 } else if (bend >= fend) {
 428                         /* end overlap */
 429
 430                         /*
 431                          * Case 7:
 432                          *             bbno           bend
 433                          *             +BBBBBBBBBBBBBBBBB+
 434                          *    +------------------+
 435                          *    fbno            fend
 436                          *
 437                          * Case 8:
 438                          *             bbno           bend
 439                          *             +BBBBBBBBBBBBBBBBB+
 440                          *    +--------------------------+
 441                          *    fbno                    fend
 442                          *
 443                          * Needs to be trimmed to:
 444                          *    +-------+
 445                          *    fbno fend
 446                          */
 447                         fend = bbno;
 448                 } else {
 449                         /* middle overlap */
 450
 451                         /*
 452                          * Case 9:
 453                          *             bbno           bend
 454                          *             +BBBBBBBBBBBBBBBBB+
 455                          *    +-----------------------------------+
 456                          *    fbno                             fend
 457                          *
 458                          * Can be trimmed to:
 459                          *    +-------+        OR         +-------+
 460                          *    fbno fend                   fbno fend
 461                          *
 462                          * Backward allocation leads to significant
 463                          * fragmentation of directories, which degrades
 464                          * directory performance, therefore we always want to
 465                          * choose the option that produces forward allocation
 466                          * patterns.
 467                          * Preferring the lower bno extent will make the next
 468                          * request use "fend" as the start of the next
 469                          * allocation;  if the segment is no longer busy at
 470                          * that point, we'll get a contiguous allocation, but
 471                          * even if it is still busy, we will get a forward
 472                          * allocation.
 473                          * We try to avoid choosing the segment at "bend",
 474                          * because that can lead to the next allocation
 475                          * taking the segment at "fbno", which would be a
 476                          * backward allocation.  We only use the segment at
 477                          * "fbno" if it is much larger than the current
 478                          * requested size, because in that case there's a
 479                          * good chance subsequent allocations will be
 480                          * contiguous.
 481                          */
 482                         if (bbno - fbno >= args->maxlen) {
 483                                 /* left candidate fits perfect */
 484                                 fend = bbno;
 485                         } else if (fend - bend >= args->maxlen * 4) {
 486                                 /* right candidate has enough free space */
 487                                 fbno = bend;
 488                         } else if (bbno - fbno >= args->minlen) {
 489                                 /* left candidate fits minimum requirement */
 490                                 fend = bbno;
 491                         } else {
 492                                 goto fail;
 493                         }
 494                 }
 495
 496                 flen = fend - fbno;
 497         }
 498 out:
 499
 500         if (fbno != *bno || flen != *len) {
 501                 trace_xfs_extent_busy_trim(args->mp, args->agno, *bno, *len,
 502                                           fbno, flen);
 503                 *bno = fbno;
 504                 *len = flen;
 505                 *busy_gen = args->pag->pagb_gen;
 506                 ret = true;
 507         }
 508         spin_unlock(&args->pag->pagb_lock);
 509         return ret;
 510 fail:
 511         /*
 512          * Return a zero extent length as failure indications.  All callers
 513          * re-check if the trimmed extent satisfies the minlen requirement.
 514          */
 515         flen = 0;
 516         goto out;
 517 }
 518
 519 STATIC void
 520 xfs_extent_busy_clear_one(
 521         struct xfs_mount        *mp,
 522         struct xfs_perag        *pag,
 523         struct xfs_extent_busy  *busyp)
 524 {
 525         if (busyp->length) {
 526                 trace_xfs_extent_busy_clear(mp, busyp->agno, busyp->bno,
 527                                                 busyp->length);
 528                 rb_erase(&busyp->rb_node, &pag->pagb_tree);
 529         }
 530
 531         list_del_init(&busyp->list);
 532         kmem_free(busyp);
 533 }
 534
 535 static void
 536 xfs_extent_busy_put_pag(
 537         struct xfs_perag        *pag,
 538         bool                    wakeup)
 539                 __releases(pag->pagb_lock)
 540 {
 541         if (wakeup) {
 542                 pag->pagb_gen++;
 543                 wake_up_all(&pag->pagb_wait);
 544         }
 545
 546         spin_unlock(&pag->pagb_lock);
 547         xfs_perag_put(pag);
 548 }
 549
 550 /*
 551  * Remove all extents on the passed in list from the busy extents tree.
 552  * If do_discard is set skip extents that need to be discarded, and mark
 553  * these as undergoing a discard operation instead.
 554  */
 555 void
 556 xfs_extent_busy_clear(
 557         struct xfs_mount        *mp,
 558         struct list_head        *list,
 559         bool                    do_discard)
 560 {
 561         struct xfs_extent_busy  *busyp, *n;
 562         struct xfs_perag        *pag = NULL;
 563         xfs_agnumber_t          agno = NULLAGNUMBER;
 564         bool                    wakeup = false;
 565
 566         list_for_each_entry_safe(busyp, n, list, list) {
 567                 if (busyp->agno != agno) {
 568                         if (pag)
 569                                 xfs_extent_busy_put_pag(pag, wakeup);
 570                         agno = busyp->agno;
 571                         pag = xfs_perag_get(mp, agno);
 572                         spin_lock(&pag->pagb_lock);
 573                         wakeup = false;
 574                 }
 575
 576                 if (do_discard && busyp->length &&
 577                     !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD)) {
 578                         busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
 579                 } else {
 580                         xfs_extent_busy_clear_one(mp, pag, busyp);
 581                         wakeup = true;
 582                 }
 583         }
 584
 585         if (pag)
 586                 xfs_extent_busy_put_pag(pag, wakeup);
 587 }
 588
 589 /*
 590  * Flush out all busy extents for this AG.
 591  *
 592  * If the current transaction is holding busy extents, the caller may not want
 593  * to wait for committed busy extents to resolve. If we are being told just to
 594  * try a flush or progress has been made since we last skipped a busy extent,
 595  * return immediately to allow the caller to try again.
 596  *
 597  * If we are freeing extents, we might actually be holding the only free extents
 598  * in the transaction busy list and the log force won't resolve that situation.
 599  * In this case, we must return -EAGAIN to avoid a deadlock by informing the
 600  * caller it needs to commit the busy extents it holds before retrying the
 601  * extent free operation.
 602  */
 603 int
 604 xfs_extent_busy_flush(
 605         struct xfs_trans        *tp,
 606         struct xfs_perag        *pag,
 607         unsigned                busy_gen,
 608         uint32_t                alloc_flags)
 609 {
 610         DEFINE_WAIT             (wait);
 611         int                     error;
 612
 613         error = xfs_log_force(tp->t_mountp, XFS_LOG_SYNC);
 614         if (error)
 615                 return error;
 616
 617         /* Avoid deadlocks on uncommitted busy extents. */
 618         if (!list_empty(&tp->t_busy)) {
 619                 if (alloc_flags & XFS_ALLOC_FLAG_TRYFLUSH)
 620                         return 0;
 621
 622                 if (busy_gen != READ_ONCE(pag->pagb_gen))
 623                         return 0;
 624
 625                 if (alloc_flags & XFS_ALLOC_FLAG_FREEING)
 626                         return -EAGAIN;
 627         }
 628
 629         /* Wait for committed busy extents to resolve. */
 630         do {
 631                 prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
 632                 if  (busy_gen != READ_ONCE(pag->pagb_gen))
 633                         break;
 634                 schedule();
 635         } while (1);
 636
 637         finish_wait(&pag->pagb_wait, &wait);
 638         return 0;
 639 }
 640
 641 void
 642 xfs_extent_busy_wait_all(
 643         struct xfs_mount        *mp)
 644 {
 645         struct xfs_perag        *pag;
 646         DEFINE_WAIT             (wait);
 647         xfs_agnumber_t          agno;
 648
 649         for_each_perag(mp, agno, pag) {
 650                 do {
 651                         prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
 652                         if  (RB_EMPTY_ROOT(&pag->pagb_tree))
 653                                 break;
 654                         schedule();
 655                 } while (1);
 656                 finish_wait(&pag->pagb_wait, &wait);
 657         }
 658 }
 659
 660 /*
 661  * Callback for list_sort to sort busy extents by the AG they reside in.
 662  */
 663 int
 664 xfs_extent_busy_ag_cmp(
 665         void                    *priv,
 666         const struct list_head  *l1,
 667         const struct list_head  *l2)
 668 {
 669         struct xfs_extent_busy  *b1 =
 670                 container_of(l1, struct xfs_extent_busy, list);
 671         struct xfs_extent_busy  *b2 =
 672                 container_of(l2, struct xfs_extent_busy, list);
 673         s32 diff;
 674
 675         diff = b1->agno - b2->agno;
 676         if (!diff)
 677                 diff = b1->bno - b2->bno;
 678         return diff;
 679 }