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