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 /* always process discard lists in fifo order */ 66 list_add_tail(&new->list, busy_list); 67 spin_unlock(&pag->pagb_lock); 68 } 69 70 void 71 xfs_extent_busy_insert( 72 struct xfs_trans *tp, 73 struct xfs_perag *pag, 74 xfs_agblock_t bno, 75 xfs_extlen_t len, 76 unsigned int flags) 77 { 78 xfs_extent_busy_insert_list(pag, bno, len, flags, &tp->t_busy); 79 } 80 81 void 82 xfs_extent_busy_insert_discard( 83 struct xfs_perag *pag, 84 xfs_agblock_t bno, 85 xfs_extlen_t len, 86 struct list_head *busy_list) 87 { 88 xfs_extent_busy_insert_list(pag, bno, len, XFS_EXTENT_BUSY_DISCARDED, 89 busy_list); 90 } 91 92 /* 93 * Search for a busy extent within the range of the extent we are about to 94 * allocate. You need to be holding the busy extent tree lock when calling 95 * xfs_extent_busy_search(). This function returns 0 for no overlapping busy 96 * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact 97 * match. This is done so that a non-zero return indicates an overlap that 98 * will require a synchronous transaction, but it can still be 99 * used to distinguish between a partial or exact match. 100 */ 101 int 102 xfs_extent_busy_search( 103 struct xfs_mount *mp, 104 struct xfs_perag *pag, 105 xfs_agblock_t bno, 106 xfs_extlen_t len) 107 { 108 struct rb_node *rbp; 109 struct xfs_extent_busy *busyp; 110 int match = 0; 111 112 /* find closest start bno overlap */ 113 spin_lock(&pag->pagb_lock); 114 rbp = pag->pagb_tree.rb_node; 115 while (rbp) { 116 busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node); 117 if (bno < busyp->bno) { 118 /* may overlap, but exact start block is lower */ 119 if (bno + len > busyp->bno) 120 match = -1; 121 rbp = rbp->rb_left; 122 } else if (bno > busyp->bno) { 123 /* may overlap, but exact start block is higher */ 124 if (bno < busyp->bno + busyp->length) 125 match = -1; 126 rbp = rbp->rb_right; 127 } else { 128 /* bno matches busyp, length determines exact match */ 129 match = (busyp->length == len) ? 1 : -1; 130 break; 131 } 132 } 133 spin_unlock(&pag->pagb_lock); 134 return match; 135 } 136 137 /* 138 * The found free extent [fbno, fend] overlaps part or all of the given busy 139 * extent. If the overlap covers the beginning, the end, or all of the busy 140 * extent, the overlapping portion can be made unbusy and used for the 141 * allocation. We can't split a busy extent because we can't modify a 142 * transaction/CIL context busy list, but we can update an entry's block 143 * number or length. 144 * 145 * Returns true if the extent can safely be reused, or false if the search 146 * needs to be restarted. 147 */ 148 STATIC bool 149 xfs_extent_busy_update_extent( 150 struct xfs_mount *mp, 151 struct xfs_perag *pag, 152 struct xfs_extent_busy *busyp, 153 xfs_agblock_t fbno, 154 xfs_extlen_t flen, 155 bool userdata) __releases(&pag->pagb_lock) 156 __acquires(&pag->pagb_lock) 157 { 158 xfs_agblock_t fend = fbno + flen; 159 xfs_agblock_t bbno = busyp->bno; 160 xfs_agblock_t bend = bbno + busyp->length; 161 162 /* 163 * This extent is currently being discarded. Give the thread 164 * performing the discard a chance to mark the extent unbusy 165 * and retry. 166 */ 167 if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) { 168 spin_unlock(&pag->pagb_lock); 169 delay(1); 170 spin_lock(&pag->pagb_lock); 171 return false; 172 } 173 174 /* 175 * If there is a busy extent overlapping a user allocation, we have 176 * no choice but to force the log and retry the search. 177 * 178 * Fortunately this does not happen during normal operation, but 179 * only if the filesystem is very low on space and has to dip into 180 * the AGFL for normal allocations. 181 */ 182 if (userdata) 183 goto out_force_log; 184 185 if (bbno < fbno && bend > fend) { 186 /* 187 * Case 1: 188 * bbno bend 189 * +BBBBBBBBBBBBBBBBB+ 190 * +---------+ 191 * fbno fend 192 */ 193 194 /* 195 * We would have to split the busy extent to be able to track 196 * it correct, which we cannot do because we would have to 197 * modify the list of busy extents attached to the transaction 198 * or CIL context, which is immutable. 199 * 200 * Force out the log to clear the busy extent and retry the 201 * search. 202 */ 203 goto out_force_log; 204 } else if (bbno >= fbno && bend <= fend) { 205 /* 206 * Case 2: 207 * bbno bend 208 * +BBBBBBBBBBBBBBBBB+ 209 * +-----------------+ 210 * fbno fend 211 * 212 * Case 3: 213 * bbno bend 214 * +BBBBBBBBBBBBBBBBB+ 215 * +--------------------------+ 216 * fbno fend 217 * 218 * Case 4: 219 * bbno bend 220 * +BBBBBBBBBBBBBBBBB+ 221 * +--------------------------+ 222 * fbno fend 223 * 224 * Case 5: 225 * bbno bend 226 * +BBBBBBBBBBBBBBBBB+ 227 * +-----------------------------------+ 228 * fbno fend 229 * 230 */ 231 232 /* 233 * The busy extent is fully covered by the extent we are 234 * allocating, and can simply be removed from the rbtree. 235 * However we cannot remove it from the immutable list 236 * tracking busy extents in the transaction or CIL context, 237 * so set the length to zero to mark it invalid. 238 * 239 * We also need to restart the busy extent search from the 240 * tree root, because erasing the node can rearrange the 241 * tree topology. 242 */ 243 rb_erase(&busyp->rb_node, &pag->pagb_tree); 244 busyp->length = 0; 245 return false; 246 } else if (fend < bend) { 247 /* 248 * Case 6: 249 * bbno bend 250 * +BBBBBBBBBBBBBBBBB+ 251 * +---------+ 252 * fbno fend 253 * 254 * Case 7: 255 * bbno bend 256 * +BBBBBBBBBBBBBBBBB+ 257 * +------------------+ 258 * fbno fend 259 * 260 */ 261 busyp->bno = fend; 262 busyp->length = bend - fend; 263 } else if (bbno < fbno) { 264 /* 265 * Case 8: 266 * bbno bend 267 * +BBBBBBBBBBBBBBBBB+ 268 * +-------------+ 269 * fbno fend 270 * 271 * Case 9: 272 * bbno bend 273 * +BBBBBBBBBBBBBBBBB+ 274 * +----------------------+ 275 * fbno fend 276 */ 277 busyp->length = fbno - busyp->bno; 278 } else { 279 ASSERT(0); 280 } 281 282 trace_xfs_extent_busy_reuse(mp, pag->pag_agno, fbno, flen); 283 return true; 284 285 out_force_log: 286 spin_unlock(&pag->pagb_lock); 287 xfs_log_force(mp, XFS_LOG_SYNC); 288 trace_xfs_extent_busy_force(mp, pag->pag_agno, fbno, flen); 289 spin_lock(&pag->pagb_lock); 290 return false; 291 } 292 293 294 /* 295 * For a given extent [fbno, flen], make sure we can reuse it safely. 296 */ 297 void 298 xfs_extent_busy_reuse( 299 struct xfs_mount *mp, 300 struct xfs_perag *pag, 301 xfs_agblock_t fbno, 302 xfs_extlen_t flen, 303 bool userdata) 304 { 305 struct rb_node *rbp; 306 307 ASSERT(flen > 0); 308 spin_lock(&pag->pagb_lock); 309 restart: 310 rbp = pag->pagb_tree.rb_node; 311 while (rbp) { 312 struct xfs_extent_busy *busyp = 313 rb_entry(rbp, struct xfs_extent_busy, rb_node); 314 xfs_agblock_t bbno = busyp->bno; 315 xfs_agblock_t bend = bbno + busyp->length; 316 317 if (fbno + flen <= bbno) { 318 rbp = rbp->rb_left; 319 continue; 320 } else if (fbno >= bend) { 321 rbp = rbp->rb_right; 322 continue; 323 } 324 325 if (!xfs_extent_busy_update_extent(mp, pag, busyp, fbno, flen, 326 userdata)) 327 goto restart; 328 } 329 spin_unlock(&pag->pagb_lock); 330 } 331 332 /* 333 * For a given extent [fbno, flen], search the busy extent list to find a 334 * subset of the extent that is not busy. If *rlen is smaller than 335 * args->minlen no suitable extent could be found, and the higher level 336 * code needs to force out the log and retry the allocation. 337 * 338 * Return the current busy generation for the AG if the extent is busy. This 339 * value can be used to wait for at least one of the currently busy extents 340 * to be cleared. Note that the busy list is not guaranteed to be empty after 341 * the gen is woken. The state of a specific extent must always be confirmed 342 * with another call to xfs_extent_busy_trim() before it can be used. 343 */ 344 bool 345 xfs_extent_busy_trim( 346 struct xfs_alloc_arg *args, 347 xfs_agblock_t *bno, 348 xfs_extlen_t *len, 349 unsigned *busy_gen) 350 { 351 xfs_agblock_t fbno; 352 xfs_extlen_t flen; 353 struct rb_node *rbp; 354 bool ret = false; 355 356 ASSERT(*len > 0); 357 358 spin_lock(&args->pag->pagb_lock); 359 fbno = *bno; 360 flen = *len; 361 rbp = args->pag->pagb_tree.rb_node; 362 while (rbp && flen >= args->minlen) { 363 struct xfs_extent_busy *busyp = 364 rb_entry(rbp, struct xfs_extent_busy, rb_node); 365 xfs_agblock_t fend = fbno + flen; 366 xfs_agblock_t bbno = busyp->bno; 367 xfs_agblock_t bend = bbno + busyp->length; 368 369 if (fend <= bbno) { 370 rbp = rbp->rb_left; 371 continue; 372 } else if (fbno >= bend) { 373 rbp = rbp->rb_right; 374 continue; 375 } 376 377 if (bbno <= fbno) { 378 /* start overlap */ 379 380 /* 381 * Case 1: 382 * bbno bend 383 * +BBBBBBBBBBBBBBBBB+ 384 * +---------+ 385 * fbno fend 386 * 387 * Case 2: 388 * bbno bend 389 * +BBBBBBBBBBBBBBBBB+ 390 * +-------------+ 391 * fbno fend 392 * 393 * Case 3: 394 * bbno bend 395 * +BBBBBBBBBBBBBBBBB+ 396 * +-------------+ 397 * fbno fend 398 * 399 * Case 4: 400 * bbno bend 401 * +BBBBBBBBBBBBBBBBB+ 402 * +-----------------+ 403 * fbno fend 404 * 405 * No unbusy region in extent, return failure. 406 */ 407 if (fend <= bend) 408 goto fail; 409 410 /* 411 * Case 5: 412 * bbno bend 413 * +BBBBBBBBBBBBBBBBB+ 414 * +----------------------+ 415 * fbno fend 416 * 417 * Case 6: 418 * bbno bend 419 * +BBBBBBBBBBBBBBBBB+ 420 * +--------------------------+ 421 * fbno fend 422 * 423 * Needs to be trimmed to: 424 * +-------+ 425 * fbno fend 426 */ 427 fbno = bend; 428 } else if (bend >= fend) { 429 /* end overlap */ 430 431 /* 432 * Case 7: 433 * bbno bend 434 * +BBBBBBBBBBBBBBBBB+ 435 * +------------------+ 436 * fbno fend 437 * 438 * Case 8: 439 * bbno bend 440 * +BBBBBBBBBBBBBBBBB+ 441 * +--------------------------+ 442 * fbno fend 443 * 444 * Needs to be trimmed to: 445 * +-------+ 446 * fbno fend 447 */ 448 fend = bbno; 449 } else { 450 /* middle overlap */ 451 452 /* 453 * Case 9: 454 * bbno bend 455 * +BBBBBBBBBBBBBBBBB+ 456 * +-----------------------------------+ 457 * fbno fend 458 * 459 * Can be trimmed to: 460 * +-------+ OR +-------+ 461 * fbno fend fbno fend 462 * 463 * Backward allocation leads to significant 464 * fragmentation of directories, which degrades 465 * directory performance, therefore we always want to 466 * choose the option that produces forward allocation 467 * patterns. 468 * Preferring the lower bno extent will make the next 469 * request use "fend" as the start of the next 470 * allocation; if the segment is no longer busy at 471 * that point, we'll get a contiguous allocation, but 472 * even if it is still busy, we will get a forward 473 * allocation. 474 * We try to avoid choosing the segment at "bend", 475 * because that can lead to the next allocation 476 * taking the segment at "fbno", which would be a 477 * backward allocation. We only use the segment at 478 * "fbno" if it is much larger than the current 479 * requested size, because in that case there's a 480 * good chance subsequent allocations will be 481 * contiguous. 482 */ 483 if (bbno - fbno >= args->maxlen) { 484 /* left candidate fits perfect */ 485 fend = bbno; 486 } else if (fend - bend >= args->maxlen * 4) { 487 /* right candidate has enough free space */ 488 fbno = bend; 489 } else if (bbno - fbno >= args->minlen) { 490 /* left candidate fits minimum requirement */ 491 fend = bbno; 492 } else { 493 goto fail; 494 } 495 } 496 497 flen = fend - fbno; 498 } 499 out: 500 501 if (fbno != *bno || flen != *len) { 502 trace_xfs_extent_busy_trim(args->mp, args->agno, *bno, *len, 503 fbno, flen); 504 *bno = fbno; 505 *len = flen; 506 *busy_gen = args->pag->pagb_gen; 507 ret = true; 508 } 509 spin_unlock(&args->pag->pagb_lock); 510 return ret; 511 fail: 512 /* 513 * Return a zero extent length as failure indications. All callers 514 * re-check if the trimmed extent satisfies the minlen requirement. 515 */ 516 flen = 0; 517 goto out; 518 } 519 520 STATIC void 521 xfs_extent_busy_clear_one( 522 struct xfs_mount *mp, 523 struct xfs_perag *pag, 524 struct xfs_extent_busy *busyp) 525 { 526 if (busyp->length) { 527 trace_xfs_extent_busy_clear(mp, busyp->agno, busyp->bno, 528 busyp->length); 529 rb_erase(&busyp->rb_node, &pag->pagb_tree); 530 } 531 532 list_del_init(&busyp->list); 533 kmem_free(busyp); 534 } 535 536 static void 537 xfs_extent_busy_put_pag( 538 struct xfs_perag *pag, 539 bool wakeup) 540 __releases(pag->pagb_lock) 541 { 542 if (wakeup) { 543 pag->pagb_gen++; 544 wake_up_all(&pag->pagb_wait); 545 } 546 547 spin_unlock(&pag->pagb_lock); 548 xfs_perag_put(pag); 549 } 550 551 /* 552 * Remove all extents on the passed in list from the busy extents tree. 553 * If do_discard is set skip extents that need to be discarded, and mark 554 * these as undergoing a discard operation instead. 555 */ 556 void 557 xfs_extent_busy_clear( 558 struct xfs_mount *mp, 559 struct list_head *list, 560 bool do_discard) 561 { 562 struct xfs_extent_busy *busyp, *n; 563 struct xfs_perag *pag = NULL; 564 xfs_agnumber_t agno = NULLAGNUMBER; 565 bool wakeup = false; 566 567 list_for_each_entry_safe(busyp, n, list, list) { 568 if (busyp->agno != agno) { 569 if (pag) 570 xfs_extent_busy_put_pag(pag, wakeup); 571 agno = busyp->agno; 572 pag = xfs_perag_get(mp, agno); 573 spin_lock(&pag->pagb_lock); 574 wakeup = false; 575 } 576 577 if (do_discard && busyp->length && 578 !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD)) { 579 busyp->flags = XFS_EXTENT_BUSY_DISCARDED; 580 } else { 581 xfs_extent_busy_clear_one(mp, pag, busyp); 582 wakeup = true; 583 } 584 } 585 586 if (pag) 587 xfs_extent_busy_put_pag(pag, wakeup); 588 } 589 590 /* 591 * Flush out all busy extents for this AG. 592 * 593 * If the current transaction is holding busy extents, the caller may not want 594 * to wait for committed busy extents to resolve. If we are being told just to 595 * try a flush or progress has been made since we last skipped a busy extent, 596 * return immediately to allow the caller to try again. 597 * 598 * If we are freeing extents, we might actually be holding the only free extents 599 * in the transaction busy list and the log force won't resolve that situation. 600 * In this case, we must return -EAGAIN to avoid a deadlock by informing the 601 * caller it needs to commit the busy extents it holds before retrying the 602 * extent free operation. 603 */ 604 int 605 xfs_extent_busy_flush( 606 struct xfs_trans *tp, 607 struct xfs_perag *pag, 608 unsigned busy_gen, 609 uint32_t alloc_flags) 610 { 611 DEFINE_WAIT (wait); 612 int error; 613 614 error = xfs_log_force(tp->t_mountp, XFS_LOG_SYNC); 615 if (error) 616 return error; 617 618 /* Avoid deadlocks on uncommitted busy extents. */ 619 if (!list_empty(&tp->t_busy)) { 620 if (alloc_flags & XFS_ALLOC_FLAG_TRYFLUSH) 621 return 0; 622 623 if (busy_gen != READ_ONCE(pag->pagb_gen)) 624 return 0; 625 626 if (alloc_flags & XFS_ALLOC_FLAG_FREEING) 627 return -EAGAIN; 628 } 629 630 /* Wait for committed busy extents to resolve. */ 631 do { 632 prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE); 633 if (busy_gen != READ_ONCE(pag->pagb_gen)) 634 break; 635 schedule(); 636 } while (1); 637 638 finish_wait(&pag->pagb_wait, &wait); 639 return 0; 640 } 641 642 void 643 xfs_extent_busy_wait_all( 644 struct xfs_mount *mp) 645 { 646 struct xfs_perag *pag; 647 DEFINE_WAIT (wait); 648 xfs_agnumber_t agno; 649 650 for_each_perag(mp, agno, pag) { 651 do { 652 prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE); 653 if (RB_EMPTY_ROOT(&pag->pagb_tree)) 654 break; 655 schedule(); 656 } while (1); 657 finish_wait(&pag->pagb_wait, &wait); 658 } 659 } 660 661 /* 662 * Callback for list_sort to sort busy extents by the AG they reside in. 663 */ 664 int 665 xfs_extent_busy_ag_cmp( 666 void *priv, 667 const struct list_head *l1, 668 const struct list_head *l2) 669 { 670 struct xfs_extent_busy *b1 = 671 container_of(l1, struct xfs_extent_busy, list); 672 struct xfs_extent_busy *b2 = 673 container_of(l2, struct xfs_extent_busy, list); 674 s32 diff; 675 676 diff = b1->agno - b2->agno; 677 if (!diff) 678 diff = b1->bno - b2->bno; 679 return diff; 680 } 681