1 /* 2 * Copyright (c) 2000-2006 Silicon Graphics, Inc. 3 * All Rights Reserved. 4 * 5 * This program is free software; you can redistribute it and/or 6 * modify it under the terms of the GNU General Public License as 7 * published by the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it would be useful, 10 * but WITHOUT ANY WARRANTY; without even the implied warranty of 11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 * GNU General Public License for more details. 13 * 14 * You should have received a copy of the GNU General Public License 15 * along with this program; if not, write the Free Software Foundation, 16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA 17 */ 18 #include "xfs.h" 19 #include <linux/stddef.h> 20 #include <linux/errno.h> 21 #include <linux/gfp.h> 22 #include <linux/pagemap.h> 23 #include <linux/init.h> 24 #include <linux/vmalloc.h> 25 #include <linux/bio.h> 26 #include <linux/sysctl.h> 27 #include <linux/proc_fs.h> 28 #include <linux/workqueue.h> 29 #include <linux/percpu.h> 30 #include <linux/blkdev.h> 31 #include <linux/hash.h> 32 #include <linux/kthread.h> 33 #include <linux/migrate.h> 34 #include <linux/backing-dev.h> 35 #include <linux/freezer.h> 36 #include <linux/sched/mm.h> 37 38 #include "xfs_format.h" 39 #include "xfs_log_format.h" 40 #include "xfs_trans_resv.h" 41 #include "xfs_sb.h" 42 #include "xfs_mount.h" 43 #include "xfs_trace.h" 44 #include "xfs_log.h" 45 46 static kmem_zone_t *xfs_buf_zone; 47 48 #ifdef XFS_BUF_LOCK_TRACKING 49 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid) 50 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1) 51 # define XB_GET_OWNER(bp) ((bp)->b_last_holder) 52 #else 53 # define XB_SET_OWNER(bp) do { } while (0) 54 # define XB_CLEAR_OWNER(bp) do { } while (0) 55 # define XB_GET_OWNER(bp) do { } while (0) 56 #endif 57 58 #define xb_to_gfp(flags) \ 59 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN) 60 61 62 static inline int 63 xfs_buf_is_vmapped( 64 struct xfs_buf *bp) 65 { 66 /* 67 * Return true if the buffer is vmapped. 68 * 69 * b_addr is null if the buffer is not mapped, but the code is clever 70 * enough to know it doesn't have to map a single page, so the check has 71 * to be both for b_addr and bp->b_page_count > 1. 72 */ 73 return bp->b_addr && bp->b_page_count > 1; 74 } 75 76 static inline int 77 xfs_buf_vmap_len( 78 struct xfs_buf *bp) 79 { 80 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset; 81 } 82 83 /* 84 * Bump the I/O in flight count on the buftarg if we haven't yet done so for 85 * this buffer. The count is incremented once per buffer (per hold cycle) 86 * because the corresponding decrement is deferred to buffer release. Buffers 87 * can undergo I/O multiple times in a hold-release cycle and per buffer I/O 88 * tracking adds unnecessary overhead. This is used for sychronization purposes 89 * with unmount (see xfs_wait_buftarg()), so all we really need is a count of 90 * in-flight buffers. 91 * 92 * Buffers that are never released (e.g., superblock, iclog buffers) must set 93 * the XBF_NO_IOACCT flag before I/O submission. Otherwise, the buftarg count 94 * never reaches zero and unmount hangs indefinitely. 95 */ 96 static inline void 97 xfs_buf_ioacct_inc( 98 struct xfs_buf *bp) 99 { 100 if (bp->b_flags & (XBF_NO_IOACCT|_XBF_IN_FLIGHT)) 101 return; 102 103 ASSERT(bp->b_flags & XBF_ASYNC); 104 bp->b_flags |= _XBF_IN_FLIGHT; 105 percpu_counter_inc(&bp->b_target->bt_io_count); 106 } 107 108 /* 109 * Clear the in-flight state on a buffer about to be released to the LRU or 110 * freed and unaccount from the buftarg. 111 */ 112 static inline void 113 xfs_buf_ioacct_dec( 114 struct xfs_buf *bp) 115 { 116 if (!(bp->b_flags & _XBF_IN_FLIGHT)) 117 return; 118 119 bp->b_flags &= ~_XBF_IN_FLIGHT; 120 percpu_counter_dec(&bp->b_target->bt_io_count); 121 } 122 123 /* 124 * When we mark a buffer stale, we remove the buffer from the LRU and clear the 125 * b_lru_ref count so that the buffer is freed immediately when the buffer 126 * reference count falls to zero. If the buffer is already on the LRU, we need 127 * to remove the reference that LRU holds on the buffer. 128 * 129 * This prevents build-up of stale buffers on the LRU. 130 */ 131 void 132 xfs_buf_stale( 133 struct xfs_buf *bp) 134 { 135 ASSERT(xfs_buf_islocked(bp)); 136 137 bp->b_flags |= XBF_STALE; 138 139 /* 140 * Clear the delwri status so that a delwri queue walker will not 141 * flush this buffer to disk now that it is stale. The delwri queue has 142 * a reference to the buffer, so this is safe to do. 143 */ 144 bp->b_flags &= ~_XBF_DELWRI_Q; 145 146 /* 147 * Once the buffer is marked stale and unlocked, a subsequent lookup 148 * could reset b_flags. There is no guarantee that the buffer is 149 * unaccounted (released to LRU) before that occurs. Drop in-flight 150 * status now to preserve accounting consistency. 151 */ 152 xfs_buf_ioacct_dec(bp); 153 154 spin_lock(&bp->b_lock); 155 atomic_set(&bp->b_lru_ref, 0); 156 if (!(bp->b_state & XFS_BSTATE_DISPOSE) && 157 (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru))) 158 atomic_dec(&bp->b_hold); 159 160 ASSERT(atomic_read(&bp->b_hold) >= 1); 161 spin_unlock(&bp->b_lock); 162 } 163 164 static int 165 xfs_buf_get_maps( 166 struct xfs_buf *bp, 167 int map_count) 168 { 169 ASSERT(bp->b_maps == NULL); 170 bp->b_map_count = map_count; 171 172 if (map_count == 1) { 173 bp->b_maps = &bp->__b_map; 174 return 0; 175 } 176 177 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map), 178 KM_NOFS); 179 if (!bp->b_maps) 180 return -ENOMEM; 181 return 0; 182 } 183 184 /* 185 * Frees b_pages if it was allocated. 186 */ 187 static void 188 xfs_buf_free_maps( 189 struct xfs_buf *bp) 190 { 191 if (bp->b_maps != &bp->__b_map) { 192 kmem_free(bp->b_maps); 193 bp->b_maps = NULL; 194 } 195 } 196 197 struct xfs_buf * 198 _xfs_buf_alloc( 199 struct xfs_buftarg *target, 200 struct xfs_buf_map *map, 201 int nmaps, 202 xfs_buf_flags_t flags) 203 { 204 struct xfs_buf *bp; 205 int error; 206 int i; 207 208 bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS); 209 if (unlikely(!bp)) 210 return NULL; 211 212 /* 213 * We don't want certain flags to appear in b_flags unless they are 214 * specifically set by later operations on the buffer. 215 */ 216 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD); 217 218 atomic_set(&bp->b_hold, 1); 219 atomic_set(&bp->b_lru_ref, 1); 220 init_completion(&bp->b_iowait); 221 INIT_LIST_HEAD(&bp->b_lru); 222 INIT_LIST_HEAD(&bp->b_list); 223 sema_init(&bp->b_sema, 0); /* held, no waiters */ 224 spin_lock_init(&bp->b_lock); 225 XB_SET_OWNER(bp); 226 bp->b_target = target; 227 bp->b_flags = flags; 228 229 /* 230 * Set length and io_length to the same value initially. 231 * I/O routines should use io_length, which will be the same in 232 * most cases but may be reset (e.g. XFS recovery). 233 */ 234 error = xfs_buf_get_maps(bp, nmaps); 235 if (error) { 236 kmem_zone_free(xfs_buf_zone, bp); 237 return NULL; 238 } 239 240 bp->b_bn = map[0].bm_bn; 241 bp->b_length = 0; 242 for (i = 0; i < nmaps; i++) { 243 bp->b_maps[i].bm_bn = map[i].bm_bn; 244 bp->b_maps[i].bm_len = map[i].bm_len; 245 bp->b_length += map[i].bm_len; 246 } 247 bp->b_io_length = bp->b_length; 248 249 atomic_set(&bp->b_pin_count, 0); 250 init_waitqueue_head(&bp->b_waiters); 251 252 XFS_STATS_INC(target->bt_mount, xb_create); 253 trace_xfs_buf_init(bp, _RET_IP_); 254 255 return bp; 256 } 257 258 /* 259 * Allocate a page array capable of holding a specified number 260 * of pages, and point the page buf at it. 261 */ 262 STATIC int 263 _xfs_buf_get_pages( 264 xfs_buf_t *bp, 265 int page_count) 266 { 267 /* Make sure that we have a page list */ 268 if (bp->b_pages == NULL) { 269 bp->b_page_count = page_count; 270 if (page_count <= XB_PAGES) { 271 bp->b_pages = bp->b_page_array; 272 } else { 273 bp->b_pages = kmem_alloc(sizeof(struct page *) * 274 page_count, KM_NOFS); 275 if (bp->b_pages == NULL) 276 return -ENOMEM; 277 } 278 memset(bp->b_pages, 0, sizeof(struct page *) * page_count); 279 } 280 return 0; 281 } 282 283 /* 284 * Frees b_pages if it was allocated. 285 */ 286 STATIC void 287 _xfs_buf_free_pages( 288 xfs_buf_t *bp) 289 { 290 if (bp->b_pages != bp->b_page_array) { 291 kmem_free(bp->b_pages); 292 bp->b_pages = NULL; 293 } 294 } 295 296 /* 297 * Releases the specified buffer. 298 * 299 * The modification state of any associated pages is left unchanged. 300 * The buffer must not be on any hash - use xfs_buf_rele instead for 301 * hashed and refcounted buffers 302 */ 303 void 304 xfs_buf_free( 305 xfs_buf_t *bp) 306 { 307 trace_xfs_buf_free(bp, _RET_IP_); 308 309 ASSERT(list_empty(&bp->b_lru)); 310 311 if (bp->b_flags & _XBF_PAGES) { 312 uint i; 313 314 if (xfs_buf_is_vmapped(bp)) 315 vm_unmap_ram(bp->b_addr - bp->b_offset, 316 bp->b_page_count); 317 318 for (i = 0; i < bp->b_page_count; i++) { 319 struct page *page = bp->b_pages[i]; 320 321 __free_page(page); 322 } 323 } else if (bp->b_flags & _XBF_KMEM) 324 kmem_free(bp->b_addr); 325 _xfs_buf_free_pages(bp); 326 xfs_buf_free_maps(bp); 327 kmem_zone_free(xfs_buf_zone, bp); 328 } 329 330 /* 331 * Allocates all the pages for buffer in question and builds it's page list. 332 */ 333 STATIC int 334 xfs_buf_allocate_memory( 335 xfs_buf_t *bp, 336 uint flags) 337 { 338 size_t size; 339 size_t nbytes, offset; 340 gfp_t gfp_mask = xb_to_gfp(flags); 341 unsigned short page_count, i; 342 xfs_off_t start, end; 343 int error; 344 345 /* 346 * for buffers that are contained within a single page, just allocate 347 * the memory from the heap - there's no need for the complexity of 348 * page arrays to keep allocation down to order 0. 349 */ 350 size = BBTOB(bp->b_length); 351 if (size < PAGE_SIZE) { 352 bp->b_addr = kmem_alloc(size, KM_NOFS); 353 if (!bp->b_addr) { 354 /* low memory - use alloc_page loop instead */ 355 goto use_alloc_page; 356 } 357 358 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) != 359 ((unsigned long)bp->b_addr & PAGE_MASK)) { 360 /* b_addr spans two pages - use alloc_page instead */ 361 kmem_free(bp->b_addr); 362 bp->b_addr = NULL; 363 goto use_alloc_page; 364 } 365 bp->b_offset = offset_in_page(bp->b_addr); 366 bp->b_pages = bp->b_page_array; 367 bp->b_pages[0] = virt_to_page(bp->b_addr); 368 bp->b_page_count = 1; 369 bp->b_flags |= _XBF_KMEM; 370 return 0; 371 } 372 373 use_alloc_page: 374 start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT; 375 end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1) 376 >> PAGE_SHIFT; 377 page_count = end - start; 378 error = _xfs_buf_get_pages(bp, page_count); 379 if (unlikely(error)) 380 return error; 381 382 offset = bp->b_offset; 383 bp->b_flags |= _XBF_PAGES; 384 385 for (i = 0; i < bp->b_page_count; i++) { 386 struct page *page; 387 uint retries = 0; 388 retry: 389 page = alloc_page(gfp_mask); 390 if (unlikely(page == NULL)) { 391 if (flags & XBF_READ_AHEAD) { 392 bp->b_page_count = i; 393 error = -ENOMEM; 394 goto out_free_pages; 395 } 396 397 /* 398 * This could deadlock. 399 * 400 * But until all the XFS lowlevel code is revamped to 401 * handle buffer allocation failures we can't do much. 402 */ 403 if (!(++retries % 100)) 404 xfs_err(NULL, 405 "%s(%u) possible memory allocation deadlock in %s (mode:0x%x)", 406 current->comm, current->pid, 407 __func__, gfp_mask); 408 409 XFS_STATS_INC(bp->b_target->bt_mount, xb_page_retries); 410 congestion_wait(BLK_RW_ASYNC, HZ/50); 411 goto retry; 412 } 413 414 XFS_STATS_INC(bp->b_target->bt_mount, xb_page_found); 415 416 nbytes = min_t(size_t, size, PAGE_SIZE - offset); 417 size -= nbytes; 418 bp->b_pages[i] = page; 419 offset = 0; 420 } 421 return 0; 422 423 out_free_pages: 424 for (i = 0; i < bp->b_page_count; i++) 425 __free_page(bp->b_pages[i]); 426 bp->b_flags &= ~_XBF_PAGES; 427 return error; 428 } 429 430 /* 431 * Map buffer into kernel address-space if necessary. 432 */ 433 STATIC int 434 _xfs_buf_map_pages( 435 xfs_buf_t *bp, 436 uint flags) 437 { 438 ASSERT(bp->b_flags & _XBF_PAGES); 439 if (bp->b_page_count == 1) { 440 /* A single page buffer is always mappable */ 441 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset; 442 } else if (flags & XBF_UNMAPPED) { 443 bp->b_addr = NULL; 444 } else { 445 int retried = 0; 446 unsigned nofs_flag; 447 448 /* 449 * vm_map_ram() will allocate auxillary structures (e.g. 450 * pagetables) with GFP_KERNEL, yet we are likely to be under 451 * GFP_NOFS context here. Hence we need to tell memory reclaim 452 * that we are in such a context via PF_MEMALLOC_NOFS to prevent 453 * memory reclaim re-entering the filesystem here and 454 * potentially deadlocking. 455 */ 456 nofs_flag = memalloc_nofs_save(); 457 do { 458 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count, 459 -1, PAGE_KERNEL); 460 if (bp->b_addr) 461 break; 462 vm_unmap_aliases(); 463 } while (retried++ <= 1); 464 memalloc_nofs_restore(nofs_flag); 465 466 if (!bp->b_addr) 467 return -ENOMEM; 468 bp->b_addr += bp->b_offset; 469 } 470 471 return 0; 472 } 473 474 /* 475 * Finding and Reading Buffers 476 */ 477 static int 478 _xfs_buf_obj_cmp( 479 struct rhashtable_compare_arg *arg, 480 const void *obj) 481 { 482 const struct xfs_buf_map *map = arg->key; 483 const struct xfs_buf *bp = obj; 484 485 /* 486 * The key hashing in the lookup path depends on the key being the 487 * first element of the compare_arg, make sure to assert this. 488 */ 489 BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0); 490 491 if (bp->b_bn != map->bm_bn) 492 return 1; 493 494 if (unlikely(bp->b_length != map->bm_len)) { 495 /* 496 * found a block number match. If the range doesn't 497 * match, the only way this is allowed is if the buffer 498 * in the cache is stale and the transaction that made 499 * it stale has not yet committed. i.e. we are 500 * reallocating a busy extent. Skip this buffer and 501 * continue searching for an exact match. 502 */ 503 ASSERT(bp->b_flags & XBF_STALE); 504 return 1; 505 } 506 return 0; 507 } 508 509 static const struct rhashtable_params xfs_buf_hash_params = { 510 .min_size = 32, /* empty AGs have minimal footprint */ 511 .nelem_hint = 16, 512 .key_len = sizeof(xfs_daddr_t), 513 .key_offset = offsetof(struct xfs_buf, b_bn), 514 .head_offset = offsetof(struct xfs_buf, b_rhash_head), 515 .automatic_shrinking = true, 516 .obj_cmpfn = _xfs_buf_obj_cmp, 517 }; 518 519 int 520 xfs_buf_hash_init( 521 struct xfs_perag *pag) 522 { 523 spin_lock_init(&pag->pag_buf_lock); 524 return rhashtable_init(&pag->pag_buf_hash, &xfs_buf_hash_params); 525 } 526 527 void 528 xfs_buf_hash_destroy( 529 struct xfs_perag *pag) 530 { 531 rhashtable_destroy(&pag->pag_buf_hash); 532 } 533 534 /* 535 * Look up, and creates if absent, a lockable buffer for 536 * a given range of an inode. The buffer is returned 537 * locked. No I/O is implied by this call. 538 */ 539 xfs_buf_t * 540 _xfs_buf_find( 541 struct xfs_buftarg *btp, 542 struct xfs_buf_map *map, 543 int nmaps, 544 xfs_buf_flags_t flags, 545 xfs_buf_t *new_bp) 546 { 547 struct xfs_perag *pag; 548 xfs_buf_t *bp; 549 struct xfs_buf_map cmap = { .bm_bn = map[0].bm_bn }; 550 xfs_daddr_t eofs; 551 int i; 552 553 for (i = 0; i < nmaps; i++) 554 cmap.bm_len += map[i].bm_len; 555 556 /* Check for IOs smaller than the sector size / not sector aligned */ 557 ASSERT(!(BBTOB(cmap.bm_len) < btp->bt_meta_sectorsize)); 558 ASSERT(!(BBTOB(cmap.bm_bn) & (xfs_off_t)btp->bt_meta_sectormask)); 559 560 /* 561 * Corrupted block numbers can get through to here, unfortunately, so we 562 * have to check that the buffer falls within the filesystem bounds. 563 */ 564 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks); 565 if (cmap.bm_bn < 0 || cmap.bm_bn >= eofs) { 566 /* 567 * XXX (dgc): we should really be returning -EFSCORRUPTED here, 568 * but none of the higher level infrastructure supports 569 * returning a specific error on buffer lookup failures. 570 */ 571 xfs_alert(btp->bt_mount, 572 "%s: Block out of range: block 0x%llx, EOFS 0x%llx ", 573 __func__, cmap.bm_bn, eofs); 574 WARN_ON(1); 575 return NULL; 576 } 577 578 pag = xfs_perag_get(btp->bt_mount, 579 xfs_daddr_to_agno(btp->bt_mount, cmap.bm_bn)); 580 581 spin_lock(&pag->pag_buf_lock); 582 bp = rhashtable_lookup_fast(&pag->pag_buf_hash, &cmap, 583 xfs_buf_hash_params); 584 if (bp) { 585 atomic_inc(&bp->b_hold); 586 goto found; 587 } 588 589 /* No match found */ 590 if (new_bp) { 591 /* the buffer keeps the perag reference until it is freed */ 592 new_bp->b_pag = pag; 593 rhashtable_insert_fast(&pag->pag_buf_hash, 594 &new_bp->b_rhash_head, 595 xfs_buf_hash_params); 596 spin_unlock(&pag->pag_buf_lock); 597 } else { 598 XFS_STATS_INC(btp->bt_mount, xb_miss_locked); 599 spin_unlock(&pag->pag_buf_lock); 600 xfs_perag_put(pag); 601 } 602 return new_bp; 603 604 found: 605 spin_unlock(&pag->pag_buf_lock); 606 xfs_perag_put(pag); 607 608 if (!xfs_buf_trylock(bp)) { 609 if (flags & XBF_TRYLOCK) { 610 xfs_buf_rele(bp); 611 XFS_STATS_INC(btp->bt_mount, xb_busy_locked); 612 return NULL; 613 } 614 xfs_buf_lock(bp); 615 XFS_STATS_INC(btp->bt_mount, xb_get_locked_waited); 616 } 617 618 /* 619 * if the buffer is stale, clear all the external state associated with 620 * it. We need to keep flags such as how we allocated the buffer memory 621 * intact here. 622 */ 623 if (bp->b_flags & XBF_STALE) { 624 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0); 625 ASSERT(bp->b_iodone == NULL); 626 bp->b_flags &= _XBF_KMEM | _XBF_PAGES; 627 bp->b_ops = NULL; 628 } 629 630 trace_xfs_buf_find(bp, flags, _RET_IP_); 631 XFS_STATS_INC(btp->bt_mount, xb_get_locked); 632 return bp; 633 } 634 635 /* 636 * Assembles a buffer covering the specified range. The code is optimised for 637 * cache hits, as metadata intensive workloads will see 3 orders of magnitude 638 * more hits than misses. 639 */ 640 struct xfs_buf * 641 xfs_buf_get_map( 642 struct xfs_buftarg *target, 643 struct xfs_buf_map *map, 644 int nmaps, 645 xfs_buf_flags_t flags) 646 { 647 struct xfs_buf *bp; 648 struct xfs_buf *new_bp; 649 int error = 0; 650 651 bp = _xfs_buf_find(target, map, nmaps, flags, NULL); 652 if (likely(bp)) 653 goto found; 654 655 new_bp = _xfs_buf_alloc(target, map, nmaps, flags); 656 if (unlikely(!new_bp)) 657 return NULL; 658 659 error = xfs_buf_allocate_memory(new_bp, flags); 660 if (error) { 661 xfs_buf_free(new_bp); 662 return NULL; 663 } 664 665 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp); 666 if (!bp) { 667 xfs_buf_free(new_bp); 668 return NULL; 669 } 670 671 if (bp != new_bp) 672 xfs_buf_free(new_bp); 673 674 found: 675 if (!bp->b_addr) { 676 error = _xfs_buf_map_pages(bp, flags); 677 if (unlikely(error)) { 678 xfs_warn(target->bt_mount, 679 "%s: failed to map pagesn", __func__); 680 xfs_buf_relse(bp); 681 return NULL; 682 } 683 } 684 685 /* 686 * Clear b_error if this is a lookup from a caller that doesn't expect 687 * valid data to be found in the buffer. 688 */ 689 if (!(flags & XBF_READ)) 690 xfs_buf_ioerror(bp, 0); 691 692 XFS_STATS_INC(target->bt_mount, xb_get); 693 trace_xfs_buf_get(bp, flags, _RET_IP_); 694 return bp; 695 } 696 697 STATIC int 698 _xfs_buf_read( 699 xfs_buf_t *bp, 700 xfs_buf_flags_t flags) 701 { 702 ASSERT(!(flags & XBF_WRITE)); 703 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL); 704 705 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD); 706 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD); 707 708 if (flags & XBF_ASYNC) { 709 xfs_buf_submit(bp); 710 return 0; 711 } 712 return xfs_buf_submit_wait(bp); 713 } 714 715 xfs_buf_t * 716 xfs_buf_read_map( 717 struct xfs_buftarg *target, 718 struct xfs_buf_map *map, 719 int nmaps, 720 xfs_buf_flags_t flags, 721 const struct xfs_buf_ops *ops) 722 { 723 struct xfs_buf *bp; 724 725 flags |= XBF_READ; 726 727 bp = xfs_buf_get_map(target, map, nmaps, flags); 728 if (bp) { 729 trace_xfs_buf_read(bp, flags, _RET_IP_); 730 731 if (!(bp->b_flags & XBF_DONE)) { 732 XFS_STATS_INC(target->bt_mount, xb_get_read); 733 bp->b_ops = ops; 734 _xfs_buf_read(bp, flags); 735 } else if (flags & XBF_ASYNC) { 736 /* 737 * Read ahead call which is already satisfied, 738 * drop the buffer 739 */ 740 xfs_buf_relse(bp); 741 return NULL; 742 } else { 743 /* We do not want read in the flags */ 744 bp->b_flags &= ~XBF_READ; 745 } 746 } 747 748 return bp; 749 } 750 751 /* 752 * If we are not low on memory then do the readahead in a deadlock 753 * safe manner. 754 */ 755 void 756 xfs_buf_readahead_map( 757 struct xfs_buftarg *target, 758 struct xfs_buf_map *map, 759 int nmaps, 760 const struct xfs_buf_ops *ops) 761 { 762 if (bdi_read_congested(target->bt_bdev->bd_bdi)) 763 return; 764 765 xfs_buf_read_map(target, map, nmaps, 766 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops); 767 } 768 769 /* 770 * Read an uncached buffer from disk. Allocates and returns a locked 771 * buffer containing the disk contents or nothing. 772 */ 773 int 774 xfs_buf_read_uncached( 775 struct xfs_buftarg *target, 776 xfs_daddr_t daddr, 777 size_t numblks, 778 int flags, 779 struct xfs_buf **bpp, 780 const struct xfs_buf_ops *ops) 781 { 782 struct xfs_buf *bp; 783 784 *bpp = NULL; 785 786 bp = xfs_buf_get_uncached(target, numblks, flags); 787 if (!bp) 788 return -ENOMEM; 789 790 /* set up the buffer for a read IO */ 791 ASSERT(bp->b_map_count == 1); 792 bp->b_bn = XFS_BUF_DADDR_NULL; /* always null for uncached buffers */ 793 bp->b_maps[0].bm_bn = daddr; 794 bp->b_flags |= XBF_READ; 795 bp->b_ops = ops; 796 797 xfs_buf_submit_wait(bp); 798 if (bp->b_error) { 799 int error = bp->b_error; 800 xfs_buf_relse(bp); 801 return error; 802 } 803 804 *bpp = bp; 805 return 0; 806 } 807 808 /* 809 * Return a buffer allocated as an empty buffer and associated to external 810 * memory via xfs_buf_associate_memory() back to it's empty state. 811 */ 812 void 813 xfs_buf_set_empty( 814 struct xfs_buf *bp, 815 size_t numblks) 816 { 817 if (bp->b_pages) 818 _xfs_buf_free_pages(bp); 819 820 bp->b_pages = NULL; 821 bp->b_page_count = 0; 822 bp->b_addr = NULL; 823 bp->b_length = numblks; 824 bp->b_io_length = numblks; 825 826 ASSERT(bp->b_map_count == 1); 827 bp->b_bn = XFS_BUF_DADDR_NULL; 828 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL; 829 bp->b_maps[0].bm_len = bp->b_length; 830 } 831 832 static inline struct page * 833 mem_to_page( 834 void *addr) 835 { 836 if ((!is_vmalloc_addr(addr))) { 837 return virt_to_page(addr); 838 } else { 839 return vmalloc_to_page(addr); 840 } 841 } 842 843 int 844 xfs_buf_associate_memory( 845 xfs_buf_t *bp, 846 void *mem, 847 size_t len) 848 { 849 int rval; 850 int i = 0; 851 unsigned long pageaddr; 852 unsigned long offset; 853 size_t buflen; 854 int page_count; 855 856 pageaddr = (unsigned long)mem & PAGE_MASK; 857 offset = (unsigned long)mem - pageaddr; 858 buflen = PAGE_ALIGN(len + offset); 859 page_count = buflen >> PAGE_SHIFT; 860 861 /* Free any previous set of page pointers */ 862 if (bp->b_pages) 863 _xfs_buf_free_pages(bp); 864 865 bp->b_pages = NULL; 866 bp->b_addr = mem; 867 868 rval = _xfs_buf_get_pages(bp, page_count); 869 if (rval) 870 return rval; 871 872 bp->b_offset = offset; 873 874 for (i = 0; i < bp->b_page_count; i++) { 875 bp->b_pages[i] = mem_to_page((void *)pageaddr); 876 pageaddr += PAGE_SIZE; 877 } 878 879 bp->b_io_length = BTOBB(len); 880 bp->b_length = BTOBB(buflen); 881 882 return 0; 883 } 884 885 xfs_buf_t * 886 xfs_buf_get_uncached( 887 struct xfs_buftarg *target, 888 size_t numblks, 889 int flags) 890 { 891 unsigned long page_count; 892 int error, i; 893 struct xfs_buf *bp; 894 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks); 895 896 /* flags might contain irrelevant bits, pass only what we care about */ 897 bp = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT); 898 if (unlikely(bp == NULL)) 899 goto fail; 900 901 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT; 902 error = _xfs_buf_get_pages(bp, page_count); 903 if (error) 904 goto fail_free_buf; 905 906 for (i = 0; i < page_count; i++) { 907 bp->b_pages[i] = alloc_page(xb_to_gfp(flags)); 908 if (!bp->b_pages[i]) 909 goto fail_free_mem; 910 } 911 bp->b_flags |= _XBF_PAGES; 912 913 error = _xfs_buf_map_pages(bp, 0); 914 if (unlikely(error)) { 915 xfs_warn(target->bt_mount, 916 "%s: failed to map pages", __func__); 917 goto fail_free_mem; 918 } 919 920 trace_xfs_buf_get_uncached(bp, _RET_IP_); 921 return bp; 922 923 fail_free_mem: 924 while (--i >= 0) 925 __free_page(bp->b_pages[i]); 926 _xfs_buf_free_pages(bp); 927 fail_free_buf: 928 xfs_buf_free_maps(bp); 929 kmem_zone_free(xfs_buf_zone, bp); 930 fail: 931 return NULL; 932 } 933 934 /* 935 * Increment reference count on buffer, to hold the buffer concurrently 936 * with another thread which may release (free) the buffer asynchronously. 937 * Must hold the buffer already to call this function. 938 */ 939 void 940 xfs_buf_hold( 941 xfs_buf_t *bp) 942 { 943 trace_xfs_buf_hold(bp, _RET_IP_); 944 atomic_inc(&bp->b_hold); 945 } 946 947 /* 948 * Release a hold on the specified buffer. If the hold count is 1, the buffer is 949 * placed on LRU or freed (depending on b_lru_ref). 950 */ 951 void 952 xfs_buf_rele( 953 xfs_buf_t *bp) 954 { 955 struct xfs_perag *pag = bp->b_pag; 956 bool release; 957 bool freebuf = false; 958 959 trace_xfs_buf_rele(bp, _RET_IP_); 960 961 if (!pag) { 962 ASSERT(list_empty(&bp->b_lru)); 963 if (atomic_dec_and_test(&bp->b_hold)) { 964 xfs_buf_ioacct_dec(bp); 965 xfs_buf_free(bp); 966 } 967 return; 968 } 969 970 ASSERT(atomic_read(&bp->b_hold) > 0); 971 972 release = atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock); 973 spin_lock(&bp->b_lock); 974 if (!release) { 975 /* 976 * Drop the in-flight state if the buffer is already on the LRU 977 * and it holds the only reference. This is racy because we 978 * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT 979 * ensures the decrement occurs only once per-buf. 980 */ 981 if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru)) 982 xfs_buf_ioacct_dec(bp); 983 goto out_unlock; 984 } 985 986 /* the last reference has been dropped ... */ 987 xfs_buf_ioacct_dec(bp); 988 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) { 989 /* 990 * If the buffer is added to the LRU take a new reference to the 991 * buffer for the LRU and clear the (now stale) dispose list 992 * state flag 993 */ 994 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) { 995 bp->b_state &= ~XFS_BSTATE_DISPOSE; 996 atomic_inc(&bp->b_hold); 997 } 998 spin_unlock(&pag->pag_buf_lock); 999 } else { 1000 /* 1001 * most of the time buffers will already be removed from the 1002 * LRU, so optimise that case by checking for the 1003 * XFS_BSTATE_DISPOSE flag indicating the last list the buffer 1004 * was on was the disposal list 1005 */ 1006 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) { 1007 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru); 1008 } else { 1009 ASSERT(list_empty(&bp->b_lru)); 1010 } 1011 1012 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); 1013 rhashtable_remove_fast(&pag->pag_buf_hash, &bp->b_rhash_head, 1014 xfs_buf_hash_params); 1015 spin_unlock(&pag->pag_buf_lock); 1016 xfs_perag_put(pag); 1017 freebuf = true; 1018 } 1019 1020 out_unlock: 1021 spin_unlock(&bp->b_lock); 1022 1023 if (freebuf) 1024 xfs_buf_free(bp); 1025 } 1026 1027 1028 /* 1029 * Lock a buffer object, if it is not already locked. 1030 * 1031 * If we come across a stale, pinned, locked buffer, we know that we are 1032 * being asked to lock a buffer that has been reallocated. Because it is 1033 * pinned, we know that the log has not been pushed to disk and hence it 1034 * will still be locked. Rather than continuing to have trylock attempts 1035 * fail until someone else pushes the log, push it ourselves before 1036 * returning. This means that the xfsaild will not get stuck trying 1037 * to push on stale inode buffers. 1038 */ 1039 int 1040 xfs_buf_trylock( 1041 struct xfs_buf *bp) 1042 { 1043 int locked; 1044 1045 locked = down_trylock(&bp->b_sema) == 0; 1046 if (locked) { 1047 XB_SET_OWNER(bp); 1048 trace_xfs_buf_trylock(bp, _RET_IP_); 1049 } else { 1050 trace_xfs_buf_trylock_fail(bp, _RET_IP_); 1051 } 1052 return locked; 1053 } 1054 1055 /* 1056 * Lock a buffer object. 1057 * 1058 * If we come across a stale, pinned, locked buffer, we know that we 1059 * are being asked to lock a buffer that has been reallocated. Because 1060 * it is pinned, we know that the log has not been pushed to disk and 1061 * hence it will still be locked. Rather than sleeping until someone 1062 * else pushes the log, push it ourselves before trying to get the lock. 1063 */ 1064 void 1065 xfs_buf_lock( 1066 struct xfs_buf *bp) 1067 { 1068 trace_xfs_buf_lock(bp, _RET_IP_); 1069 1070 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE)) 1071 xfs_log_force(bp->b_target->bt_mount, 0); 1072 down(&bp->b_sema); 1073 XB_SET_OWNER(bp); 1074 1075 trace_xfs_buf_lock_done(bp, _RET_IP_); 1076 } 1077 1078 void 1079 xfs_buf_unlock( 1080 struct xfs_buf *bp) 1081 { 1082 ASSERT(xfs_buf_islocked(bp)); 1083 1084 XB_CLEAR_OWNER(bp); 1085 up(&bp->b_sema); 1086 1087 trace_xfs_buf_unlock(bp, _RET_IP_); 1088 } 1089 1090 STATIC void 1091 xfs_buf_wait_unpin( 1092 xfs_buf_t *bp) 1093 { 1094 DECLARE_WAITQUEUE (wait, current); 1095 1096 if (atomic_read(&bp->b_pin_count) == 0) 1097 return; 1098 1099 add_wait_queue(&bp->b_waiters, &wait); 1100 for (;;) { 1101 set_current_state(TASK_UNINTERRUPTIBLE); 1102 if (atomic_read(&bp->b_pin_count) == 0) 1103 break; 1104 io_schedule(); 1105 } 1106 remove_wait_queue(&bp->b_waiters, &wait); 1107 set_current_state(TASK_RUNNING); 1108 } 1109 1110 /* 1111 * Buffer Utility Routines 1112 */ 1113 1114 void 1115 xfs_buf_ioend( 1116 struct xfs_buf *bp) 1117 { 1118 bool read = bp->b_flags & XBF_READ; 1119 1120 trace_xfs_buf_iodone(bp, _RET_IP_); 1121 1122 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); 1123 1124 /* 1125 * Pull in IO completion errors now. We are guaranteed to be running 1126 * single threaded, so we don't need the lock to read b_io_error. 1127 */ 1128 if (!bp->b_error && bp->b_io_error) 1129 xfs_buf_ioerror(bp, bp->b_io_error); 1130 1131 /* Only validate buffers that were read without errors */ 1132 if (read && !bp->b_error && bp->b_ops) { 1133 ASSERT(!bp->b_iodone); 1134 bp->b_ops->verify_read(bp); 1135 } 1136 1137 if (!bp->b_error) 1138 bp->b_flags |= XBF_DONE; 1139 1140 if (bp->b_iodone) 1141 (*(bp->b_iodone))(bp); 1142 else if (bp->b_flags & XBF_ASYNC) 1143 xfs_buf_relse(bp); 1144 else 1145 complete(&bp->b_iowait); 1146 } 1147 1148 static void 1149 xfs_buf_ioend_work( 1150 struct work_struct *work) 1151 { 1152 struct xfs_buf *bp = 1153 container_of(work, xfs_buf_t, b_ioend_work); 1154 1155 xfs_buf_ioend(bp); 1156 } 1157 1158 static void 1159 xfs_buf_ioend_async( 1160 struct xfs_buf *bp) 1161 { 1162 INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work); 1163 queue_work(bp->b_ioend_wq, &bp->b_ioend_work); 1164 } 1165 1166 void 1167 xfs_buf_ioerror( 1168 xfs_buf_t *bp, 1169 int error) 1170 { 1171 ASSERT(error <= 0 && error >= -1000); 1172 bp->b_error = error; 1173 trace_xfs_buf_ioerror(bp, error, _RET_IP_); 1174 } 1175 1176 void 1177 xfs_buf_ioerror_alert( 1178 struct xfs_buf *bp, 1179 const char *func) 1180 { 1181 xfs_alert(bp->b_target->bt_mount, 1182 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d", 1183 (__uint64_t)XFS_BUF_ADDR(bp), func, -bp->b_error, bp->b_length); 1184 } 1185 1186 int 1187 xfs_bwrite( 1188 struct xfs_buf *bp) 1189 { 1190 int error; 1191 1192 ASSERT(xfs_buf_islocked(bp)); 1193 1194 bp->b_flags |= XBF_WRITE; 1195 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q | 1196 XBF_WRITE_FAIL | XBF_DONE); 1197 1198 error = xfs_buf_submit_wait(bp); 1199 if (error) { 1200 xfs_force_shutdown(bp->b_target->bt_mount, 1201 SHUTDOWN_META_IO_ERROR); 1202 } 1203 return error; 1204 } 1205 1206 static void 1207 xfs_buf_bio_end_io( 1208 struct bio *bio) 1209 { 1210 struct xfs_buf *bp = (struct xfs_buf *)bio->bi_private; 1211 1212 /* 1213 * don't overwrite existing errors - otherwise we can lose errors on 1214 * buffers that require multiple bios to complete. 1215 */ 1216 if (bio->bi_error) 1217 cmpxchg(&bp->b_io_error, 0, bio->bi_error); 1218 1219 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ)) 1220 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp)); 1221 1222 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) 1223 xfs_buf_ioend_async(bp); 1224 bio_put(bio); 1225 } 1226 1227 static void 1228 xfs_buf_ioapply_map( 1229 struct xfs_buf *bp, 1230 int map, 1231 int *buf_offset, 1232 int *count, 1233 int op, 1234 int op_flags) 1235 { 1236 int page_index; 1237 int total_nr_pages = bp->b_page_count; 1238 int nr_pages; 1239 struct bio *bio; 1240 sector_t sector = bp->b_maps[map].bm_bn; 1241 int size; 1242 int offset; 1243 1244 total_nr_pages = bp->b_page_count; 1245 1246 /* skip the pages in the buffer before the start offset */ 1247 page_index = 0; 1248 offset = *buf_offset; 1249 while (offset >= PAGE_SIZE) { 1250 page_index++; 1251 offset -= PAGE_SIZE; 1252 } 1253 1254 /* 1255 * Limit the IO size to the length of the current vector, and update the 1256 * remaining IO count for the next time around. 1257 */ 1258 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count); 1259 *count -= size; 1260 *buf_offset += size; 1261 1262 next_chunk: 1263 atomic_inc(&bp->b_io_remaining); 1264 nr_pages = min(total_nr_pages, BIO_MAX_PAGES); 1265 1266 bio = bio_alloc(GFP_NOIO, nr_pages); 1267 bio->bi_bdev = bp->b_target->bt_bdev; 1268 bio->bi_iter.bi_sector = sector; 1269 bio->bi_end_io = xfs_buf_bio_end_io; 1270 bio->bi_private = bp; 1271 bio_set_op_attrs(bio, op, op_flags); 1272 1273 for (; size && nr_pages; nr_pages--, page_index++) { 1274 int rbytes, nbytes = PAGE_SIZE - offset; 1275 1276 if (nbytes > size) 1277 nbytes = size; 1278 1279 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes, 1280 offset); 1281 if (rbytes < nbytes) 1282 break; 1283 1284 offset = 0; 1285 sector += BTOBB(nbytes); 1286 size -= nbytes; 1287 total_nr_pages--; 1288 } 1289 1290 if (likely(bio->bi_iter.bi_size)) { 1291 if (xfs_buf_is_vmapped(bp)) { 1292 flush_kernel_vmap_range(bp->b_addr, 1293 xfs_buf_vmap_len(bp)); 1294 } 1295 submit_bio(bio); 1296 if (size) 1297 goto next_chunk; 1298 } else { 1299 /* 1300 * This is guaranteed not to be the last io reference count 1301 * because the caller (xfs_buf_submit) holds a count itself. 1302 */ 1303 atomic_dec(&bp->b_io_remaining); 1304 xfs_buf_ioerror(bp, -EIO); 1305 bio_put(bio); 1306 } 1307 1308 } 1309 1310 STATIC void 1311 _xfs_buf_ioapply( 1312 struct xfs_buf *bp) 1313 { 1314 struct blk_plug plug; 1315 int op; 1316 int op_flags = 0; 1317 int offset; 1318 int size; 1319 int i; 1320 1321 /* 1322 * Make sure we capture only current IO errors rather than stale errors 1323 * left over from previous use of the buffer (e.g. failed readahead). 1324 */ 1325 bp->b_error = 0; 1326 1327 /* 1328 * Initialize the I/O completion workqueue if we haven't yet or the 1329 * submitter has not opted to specify a custom one. 1330 */ 1331 if (!bp->b_ioend_wq) 1332 bp->b_ioend_wq = bp->b_target->bt_mount->m_buf_workqueue; 1333 1334 if (bp->b_flags & XBF_WRITE) { 1335 op = REQ_OP_WRITE; 1336 if (bp->b_flags & XBF_SYNCIO) 1337 op_flags = REQ_SYNC; 1338 if (bp->b_flags & XBF_FUA) 1339 op_flags |= REQ_FUA; 1340 if (bp->b_flags & XBF_FLUSH) 1341 op_flags |= REQ_PREFLUSH; 1342 1343 /* 1344 * Run the write verifier callback function if it exists. If 1345 * this function fails it will mark the buffer with an error and 1346 * the IO should not be dispatched. 1347 */ 1348 if (bp->b_ops) { 1349 bp->b_ops->verify_write(bp); 1350 if (bp->b_error) { 1351 xfs_force_shutdown(bp->b_target->bt_mount, 1352 SHUTDOWN_CORRUPT_INCORE); 1353 return; 1354 } 1355 } else if (bp->b_bn != XFS_BUF_DADDR_NULL) { 1356 struct xfs_mount *mp = bp->b_target->bt_mount; 1357 1358 /* 1359 * non-crc filesystems don't attach verifiers during 1360 * log recovery, so don't warn for such filesystems. 1361 */ 1362 if (xfs_sb_version_hascrc(&mp->m_sb)) { 1363 xfs_warn(mp, 1364 "%s: no ops on block 0x%llx/0x%x", 1365 __func__, bp->b_bn, bp->b_length); 1366 xfs_hex_dump(bp->b_addr, 64); 1367 dump_stack(); 1368 } 1369 } 1370 } else if (bp->b_flags & XBF_READ_AHEAD) { 1371 op = REQ_OP_READ; 1372 op_flags = REQ_RAHEAD; 1373 } else { 1374 op = REQ_OP_READ; 1375 } 1376 1377 /* we only use the buffer cache for meta-data */ 1378 op_flags |= REQ_META; 1379 1380 /* 1381 * Walk all the vectors issuing IO on them. Set up the initial offset 1382 * into the buffer and the desired IO size before we start - 1383 * _xfs_buf_ioapply_vec() will modify them appropriately for each 1384 * subsequent call. 1385 */ 1386 offset = bp->b_offset; 1387 size = BBTOB(bp->b_io_length); 1388 blk_start_plug(&plug); 1389 for (i = 0; i < bp->b_map_count; i++) { 1390 xfs_buf_ioapply_map(bp, i, &offset, &size, op, op_flags); 1391 if (bp->b_error) 1392 break; 1393 if (size <= 0) 1394 break; /* all done */ 1395 } 1396 blk_finish_plug(&plug); 1397 } 1398 1399 /* 1400 * Asynchronous IO submission path. This transfers the buffer lock ownership and 1401 * the current reference to the IO. It is not safe to reference the buffer after 1402 * a call to this function unless the caller holds an additional reference 1403 * itself. 1404 */ 1405 void 1406 xfs_buf_submit( 1407 struct xfs_buf *bp) 1408 { 1409 trace_xfs_buf_submit(bp, _RET_IP_); 1410 1411 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); 1412 ASSERT(bp->b_flags & XBF_ASYNC); 1413 1414 /* on shutdown we stale and complete the buffer immediately */ 1415 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) { 1416 xfs_buf_ioerror(bp, -EIO); 1417 bp->b_flags &= ~XBF_DONE; 1418 xfs_buf_stale(bp); 1419 xfs_buf_ioend(bp); 1420 return; 1421 } 1422 1423 if (bp->b_flags & XBF_WRITE) 1424 xfs_buf_wait_unpin(bp); 1425 1426 /* clear the internal error state to avoid spurious errors */ 1427 bp->b_io_error = 0; 1428 1429 /* 1430 * The caller's reference is released during I/O completion. 1431 * This occurs some time after the last b_io_remaining reference is 1432 * released, so after we drop our Io reference we have to have some 1433 * other reference to ensure the buffer doesn't go away from underneath 1434 * us. Take a direct reference to ensure we have safe access to the 1435 * buffer until we are finished with it. 1436 */ 1437 xfs_buf_hold(bp); 1438 1439 /* 1440 * Set the count to 1 initially, this will stop an I/O completion 1441 * callout which happens before we have started all the I/O from calling 1442 * xfs_buf_ioend too early. 1443 */ 1444 atomic_set(&bp->b_io_remaining, 1); 1445 xfs_buf_ioacct_inc(bp); 1446 _xfs_buf_ioapply(bp); 1447 1448 /* 1449 * If _xfs_buf_ioapply failed, we can get back here with only the IO 1450 * reference we took above. If we drop it to zero, run completion so 1451 * that we don't return to the caller with completion still pending. 1452 */ 1453 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) { 1454 if (bp->b_error) 1455 xfs_buf_ioend(bp); 1456 else 1457 xfs_buf_ioend_async(bp); 1458 } 1459 1460 xfs_buf_rele(bp); 1461 /* Note: it is not safe to reference bp now we've dropped our ref */ 1462 } 1463 1464 /* 1465 * Synchronous buffer IO submission path, read or write. 1466 */ 1467 int 1468 xfs_buf_submit_wait( 1469 struct xfs_buf *bp) 1470 { 1471 int error; 1472 1473 trace_xfs_buf_submit_wait(bp, _RET_IP_); 1474 1475 ASSERT(!(bp->b_flags & (_XBF_DELWRI_Q | XBF_ASYNC))); 1476 1477 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) { 1478 xfs_buf_ioerror(bp, -EIO); 1479 xfs_buf_stale(bp); 1480 bp->b_flags &= ~XBF_DONE; 1481 return -EIO; 1482 } 1483 1484 if (bp->b_flags & XBF_WRITE) 1485 xfs_buf_wait_unpin(bp); 1486 1487 /* clear the internal error state to avoid spurious errors */ 1488 bp->b_io_error = 0; 1489 1490 /* 1491 * For synchronous IO, the IO does not inherit the submitters reference 1492 * count, nor the buffer lock. Hence we cannot release the reference we 1493 * are about to take until we've waited for all IO completion to occur, 1494 * including any xfs_buf_ioend_async() work that may be pending. 1495 */ 1496 xfs_buf_hold(bp); 1497 1498 /* 1499 * Set the count to 1 initially, this will stop an I/O completion 1500 * callout which happens before we have started all the I/O from calling 1501 * xfs_buf_ioend too early. 1502 */ 1503 atomic_set(&bp->b_io_remaining, 1); 1504 _xfs_buf_ioapply(bp); 1505 1506 /* 1507 * make sure we run completion synchronously if it raced with us and is 1508 * already complete. 1509 */ 1510 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) 1511 xfs_buf_ioend(bp); 1512 1513 /* wait for completion before gathering the error from the buffer */ 1514 trace_xfs_buf_iowait(bp, _RET_IP_); 1515 wait_for_completion(&bp->b_iowait); 1516 trace_xfs_buf_iowait_done(bp, _RET_IP_); 1517 error = bp->b_error; 1518 1519 /* 1520 * all done now, we can release the hold that keeps the buffer 1521 * referenced for the entire IO. 1522 */ 1523 xfs_buf_rele(bp); 1524 return error; 1525 } 1526 1527 void * 1528 xfs_buf_offset( 1529 struct xfs_buf *bp, 1530 size_t offset) 1531 { 1532 struct page *page; 1533 1534 if (bp->b_addr) 1535 return bp->b_addr + offset; 1536 1537 offset += bp->b_offset; 1538 page = bp->b_pages[offset >> PAGE_SHIFT]; 1539 return page_address(page) + (offset & (PAGE_SIZE-1)); 1540 } 1541 1542 /* 1543 * Move data into or out of a buffer. 1544 */ 1545 void 1546 xfs_buf_iomove( 1547 xfs_buf_t *bp, /* buffer to process */ 1548 size_t boff, /* starting buffer offset */ 1549 size_t bsize, /* length to copy */ 1550 void *data, /* data address */ 1551 xfs_buf_rw_t mode) /* read/write/zero flag */ 1552 { 1553 size_t bend; 1554 1555 bend = boff + bsize; 1556 while (boff < bend) { 1557 struct page *page; 1558 int page_index, page_offset, csize; 1559 1560 page_index = (boff + bp->b_offset) >> PAGE_SHIFT; 1561 page_offset = (boff + bp->b_offset) & ~PAGE_MASK; 1562 page = bp->b_pages[page_index]; 1563 csize = min_t(size_t, PAGE_SIZE - page_offset, 1564 BBTOB(bp->b_io_length) - boff); 1565 1566 ASSERT((csize + page_offset) <= PAGE_SIZE); 1567 1568 switch (mode) { 1569 case XBRW_ZERO: 1570 memset(page_address(page) + page_offset, 0, csize); 1571 break; 1572 case XBRW_READ: 1573 memcpy(data, page_address(page) + page_offset, csize); 1574 break; 1575 case XBRW_WRITE: 1576 memcpy(page_address(page) + page_offset, data, csize); 1577 } 1578 1579 boff += csize; 1580 data += csize; 1581 } 1582 } 1583 1584 /* 1585 * Handling of buffer targets (buftargs). 1586 */ 1587 1588 /* 1589 * Wait for any bufs with callbacks that have been submitted but have not yet 1590 * returned. These buffers will have an elevated hold count, so wait on those 1591 * while freeing all the buffers only held by the LRU. 1592 */ 1593 static enum lru_status 1594 xfs_buftarg_wait_rele( 1595 struct list_head *item, 1596 struct list_lru_one *lru, 1597 spinlock_t *lru_lock, 1598 void *arg) 1599 1600 { 1601 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru); 1602 struct list_head *dispose = arg; 1603 1604 if (atomic_read(&bp->b_hold) > 1) { 1605 /* need to wait, so skip it this pass */ 1606 trace_xfs_buf_wait_buftarg(bp, _RET_IP_); 1607 return LRU_SKIP; 1608 } 1609 if (!spin_trylock(&bp->b_lock)) 1610 return LRU_SKIP; 1611 1612 /* 1613 * clear the LRU reference count so the buffer doesn't get 1614 * ignored in xfs_buf_rele(). 1615 */ 1616 atomic_set(&bp->b_lru_ref, 0); 1617 bp->b_state |= XFS_BSTATE_DISPOSE; 1618 list_lru_isolate_move(lru, item, dispose); 1619 spin_unlock(&bp->b_lock); 1620 return LRU_REMOVED; 1621 } 1622 1623 void 1624 xfs_wait_buftarg( 1625 struct xfs_buftarg *btp) 1626 { 1627 LIST_HEAD(dispose); 1628 int loop = 0; 1629 1630 /* 1631 * First wait on the buftarg I/O count for all in-flight buffers to be 1632 * released. This is critical as new buffers do not make the LRU until 1633 * they are released. 1634 * 1635 * Next, flush the buffer workqueue to ensure all completion processing 1636 * has finished. Just waiting on buffer locks is not sufficient for 1637 * async IO as the reference count held over IO is not released until 1638 * after the buffer lock is dropped. Hence we need to ensure here that 1639 * all reference counts have been dropped before we start walking the 1640 * LRU list. 1641 */ 1642 while (percpu_counter_sum(&btp->bt_io_count)) 1643 delay(100); 1644 flush_workqueue(btp->bt_mount->m_buf_workqueue); 1645 1646 /* loop until there is nothing left on the lru list. */ 1647 while (list_lru_count(&btp->bt_lru)) { 1648 list_lru_walk(&btp->bt_lru, xfs_buftarg_wait_rele, 1649 &dispose, LONG_MAX); 1650 1651 while (!list_empty(&dispose)) { 1652 struct xfs_buf *bp; 1653 bp = list_first_entry(&dispose, struct xfs_buf, b_lru); 1654 list_del_init(&bp->b_lru); 1655 if (bp->b_flags & XBF_WRITE_FAIL) { 1656 xfs_alert(btp->bt_mount, 1657 "Corruption Alert: Buffer at block 0x%llx had permanent write failures!", 1658 (long long)bp->b_bn); 1659 xfs_alert(btp->bt_mount, 1660 "Please run xfs_repair to determine the extent of the problem."); 1661 } 1662 xfs_buf_rele(bp); 1663 } 1664 if (loop++ != 0) 1665 delay(100); 1666 } 1667 } 1668 1669 static enum lru_status 1670 xfs_buftarg_isolate( 1671 struct list_head *item, 1672 struct list_lru_one *lru, 1673 spinlock_t *lru_lock, 1674 void *arg) 1675 { 1676 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru); 1677 struct list_head *dispose = arg; 1678 1679 /* 1680 * we are inverting the lru lock/bp->b_lock here, so use a trylock. 1681 * If we fail to get the lock, just skip it. 1682 */ 1683 if (!spin_trylock(&bp->b_lock)) 1684 return LRU_SKIP; 1685 /* 1686 * Decrement the b_lru_ref count unless the value is already 1687 * zero. If the value is already zero, we need to reclaim the 1688 * buffer, otherwise it gets another trip through the LRU. 1689 */ 1690 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) { 1691 spin_unlock(&bp->b_lock); 1692 return LRU_ROTATE; 1693 } 1694 1695 bp->b_state |= XFS_BSTATE_DISPOSE; 1696 list_lru_isolate_move(lru, item, dispose); 1697 spin_unlock(&bp->b_lock); 1698 return LRU_REMOVED; 1699 } 1700 1701 static unsigned long 1702 xfs_buftarg_shrink_scan( 1703 struct shrinker *shrink, 1704 struct shrink_control *sc) 1705 { 1706 struct xfs_buftarg *btp = container_of(shrink, 1707 struct xfs_buftarg, bt_shrinker); 1708 LIST_HEAD(dispose); 1709 unsigned long freed; 1710 1711 freed = list_lru_shrink_walk(&btp->bt_lru, sc, 1712 xfs_buftarg_isolate, &dispose); 1713 1714 while (!list_empty(&dispose)) { 1715 struct xfs_buf *bp; 1716 bp = list_first_entry(&dispose, struct xfs_buf, b_lru); 1717 list_del_init(&bp->b_lru); 1718 xfs_buf_rele(bp); 1719 } 1720 1721 return freed; 1722 } 1723 1724 static unsigned long 1725 xfs_buftarg_shrink_count( 1726 struct shrinker *shrink, 1727 struct shrink_control *sc) 1728 { 1729 struct xfs_buftarg *btp = container_of(shrink, 1730 struct xfs_buftarg, bt_shrinker); 1731 return list_lru_shrink_count(&btp->bt_lru, sc); 1732 } 1733 1734 void 1735 xfs_free_buftarg( 1736 struct xfs_mount *mp, 1737 struct xfs_buftarg *btp) 1738 { 1739 unregister_shrinker(&btp->bt_shrinker); 1740 ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0); 1741 percpu_counter_destroy(&btp->bt_io_count); 1742 list_lru_destroy(&btp->bt_lru); 1743 1744 xfs_blkdev_issue_flush(btp); 1745 1746 kmem_free(btp); 1747 } 1748 1749 int 1750 xfs_setsize_buftarg( 1751 xfs_buftarg_t *btp, 1752 unsigned int sectorsize) 1753 { 1754 /* Set up metadata sector size info */ 1755 btp->bt_meta_sectorsize = sectorsize; 1756 btp->bt_meta_sectormask = sectorsize - 1; 1757 1758 if (set_blocksize(btp->bt_bdev, sectorsize)) { 1759 xfs_warn(btp->bt_mount, 1760 "Cannot set_blocksize to %u on device %pg", 1761 sectorsize, btp->bt_bdev); 1762 return -EINVAL; 1763 } 1764 1765 /* Set up device logical sector size mask */ 1766 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev); 1767 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1; 1768 1769 return 0; 1770 } 1771 1772 /* 1773 * When allocating the initial buffer target we have not yet 1774 * read in the superblock, so don't know what sized sectors 1775 * are being used at this early stage. Play safe. 1776 */ 1777 STATIC int 1778 xfs_setsize_buftarg_early( 1779 xfs_buftarg_t *btp, 1780 struct block_device *bdev) 1781 { 1782 return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev)); 1783 } 1784 1785 xfs_buftarg_t * 1786 xfs_alloc_buftarg( 1787 struct xfs_mount *mp, 1788 struct block_device *bdev) 1789 { 1790 xfs_buftarg_t *btp; 1791 1792 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS); 1793 1794 btp->bt_mount = mp; 1795 btp->bt_dev = bdev->bd_dev; 1796 btp->bt_bdev = bdev; 1797 1798 if (xfs_setsize_buftarg_early(btp, bdev)) 1799 goto error; 1800 1801 if (list_lru_init(&btp->bt_lru)) 1802 goto error; 1803 1804 if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL)) 1805 goto error; 1806 1807 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count; 1808 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan; 1809 btp->bt_shrinker.seeks = DEFAULT_SEEKS; 1810 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE; 1811 register_shrinker(&btp->bt_shrinker); 1812 return btp; 1813 1814 error: 1815 kmem_free(btp); 1816 return NULL; 1817 } 1818 1819 /* 1820 * Cancel a delayed write list. 1821 * 1822 * Remove each buffer from the list, clear the delwri queue flag and drop the 1823 * associated buffer reference. 1824 */ 1825 void 1826 xfs_buf_delwri_cancel( 1827 struct list_head *list) 1828 { 1829 struct xfs_buf *bp; 1830 1831 while (!list_empty(list)) { 1832 bp = list_first_entry(list, struct xfs_buf, b_list); 1833 1834 xfs_buf_lock(bp); 1835 bp->b_flags &= ~_XBF_DELWRI_Q; 1836 list_del_init(&bp->b_list); 1837 xfs_buf_relse(bp); 1838 } 1839 } 1840 1841 /* 1842 * Add a buffer to the delayed write list. 1843 * 1844 * This queues a buffer for writeout if it hasn't already been. Note that 1845 * neither this routine nor the buffer list submission functions perform 1846 * any internal synchronization. It is expected that the lists are thread-local 1847 * to the callers. 1848 * 1849 * Returns true if we queued up the buffer, or false if it already had 1850 * been on the buffer list. 1851 */ 1852 bool 1853 xfs_buf_delwri_queue( 1854 struct xfs_buf *bp, 1855 struct list_head *list) 1856 { 1857 ASSERT(xfs_buf_islocked(bp)); 1858 ASSERT(!(bp->b_flags & XBF_READ)); 1859 1860 /* 1861 * If the buffer is already marked delwri it already is queued up 1862 * by someone else for imediate writeout. Just ignore it in that 1863 * case. 1864 */ 1865 if (bp->b_flags & _XBF_DELWRI_Q) { 1866 trace_xfs_buf_delwri_queued(bp, _RET_IP_); 1867 return false; 1868 } 1869 1870 trace_xfs_buf_delwri_queue(bp, _RET_IP_); 1871 1872 /* 1873 * If a buffer gets written out synchronously or marked stale while it 1874 * is on a delwri list we lazily remove it. To do this, the other party 1875 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone. 1876 * It remains referenced and on the list. In a rare corner case it 1877 * might get readded to a delwri list after the synchronous writeout, in 1878 * which case we need just need to re-add the flag here. 1879 */ 1880 bp->b_flags |= _XBF_DELWRI_Q; 1881 if (list_empty(&bp->b_list)) { 1882 atomic_inc(&bp->b_hold); 1883 list_add_tail(&bp->b_list, list); 1884 } 1885 1886 return true; 1887 } 1888 1889 /* 1890 * Compare function is more complex than it needs to be because 1891 * the return value is only 32 bits and we are doing comparisons 1892 * on 64 bit values 1893 */ 1894 static int 1895 xfs_buf_cmp( 1896 void *priv, 1897 struct list_head *a, 1898 struct list_head *b) 1899 { 1900 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list); 1901 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list); 1902 xfs_daddr_t diff; 1903 1904 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn; 1905 if (diff < 0) 1906 return -1; 1907 if (diff > 0) 1908 return 1; 1909 return 0; 1910 } 1911 1912 /* 1913 * submit buffers for write. 1914 * 1915 * When we have a large buffer list, we do not want to hold all the buffers 1916 * locked while we block on the request queue waiting for IO dispatch. To avoid 1917 * this problem, we lock and submit buffers in groups of 50, thereby minimising 1918 * the lock hold times for lists which may contain thousands of objects. 1919 * 1920 * To do this, we sort the buffer list before we walk the list to lock and 1921 * submit buffers, and we plug and unplug around each group of buffers we 1922 * submit. 1923 */ 1924 static int 1925 xfs_buf_delwri_submit_buffers( 1926 struct list_head *buffer_list, 1927 struct list_head *wait_list) 1928 { 1929 struct xfs_buf *bp, *n; 1930 LIST_HEAD (submit_list); 1931 int pinned = 0; 1932 struct blk_plug plug; 1933 1934 list_sort(NULL, buffer_list, xfs_buf_cmp); 1935 1936 blk_start_plug(&plug); 1937 list_for_each_entry_safe(bp, n, buffer_list, b_list) { 1938 if (!wait_list) { 1939 if (xfs_buf_ispinned(bp)) { 1940 pinned++; 1941 continue; 1942 } 1943 if (!xfs_buf_trylock(bp)) 1944 continue; 1945 } else { 1946 xfs_buf_lock(bp); 1947 } 1948 1949 /* 1950 * Someone else might have written the buffer synchronously or 1951 * marked it stale in the meantime. In that case only the 1952 * _XBF_DELWRI_Q flag got cleared, and we have to drop the 1953 * reference and remove it from the list here. 1954 */ 1955 if (!(bp->b_flags & _XBF_DELWRI_Q)) { 1956 list_del_init(&bp->b_list); 1957 xfs_buf_relse(bp); 1958 continue; 1959 } 1960 1961 trace_xfs_buf_delwri_split(bp, _RET_IP_); 1962 1963 /* 1964 * We do all IO submission async. This means if we need 1965 * to wait for IO completion we need to take an extra 1966 * reference so the buffer is still valid on the other 1967 * side. We need to move the buffer onto the io_list 1968 * at this point so the caller can still access it. 1969 */ 1970 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_WRITE_FAIL); 1971 bp->b_flags |= XBF_WRITE | XBF_ASYNC; 1972 if (wait_list) { 1973 xfs_buf_hold(bp); 1974 list_move_tail(&bp->b_list, wait_list); 1975 } else 1976 list_del_init(&bp->b_list); 1977 1978 xfs_buf_submit(bp); 1979 } 1980 blk_finish_plug(&plug); 1981 1982 return pinned; 1983 } 1984 1985 /* 1986 * Write out a buffer list asynchronously. 1987 * 1988 * This will take the @buffer_list, write all non-locked and non-pinned buffers 1989 * out and not wait for I/O completion on any of the buffers. This interface 1990 * is only safely useable for callers that can track I/O completion by higher 1991 * level means, e.g. AIL pushing as the @buffer_list is consumed in this 1992 * function. 1993 */ 1994 int 1995 xfs_buf_delwri_submit_nowait( 1996 struct list_head *buffer_list) 1997 { 1998 return xfs_buf_delwri_submit_buffers(buffer_list, NULL); 1999 } 2000 2001 /* 2002 * Write out a buffer list synchronously. 2003 * 2004 * This will take the @buffer_list, write all buffers out and wait for I/O 2005 * completion on all of the buffers. @buffer_list is consumed by the function, 2006 * so callers must have some other way of tracking buffers if they require such 2007 * functionality. 2008 */ 2009 int 2010 xfs_buf_delwri_submit( 2011 struct list_head *buffer_list) 2012 { 2013 LIST_HEAD (wait_list); 2014 int error = 0, error2; 2015 struct xfs_buf *bp; 2016 2017 xfs_buf_delwri_submit_buffers(buffer_list, &wait_list); 2018 2019 /* Wait for IO to complete. */ 2020 while (!list_empty(&wait_list)) { 2021 bp = list_first_entry(&wait_list, struct xfs_buf, b_list); 2022 2023 list_del_init(&bp->b_list); 2024 2025 /* locking the buffer will wait for async IO completion. */ 2026 xfs_buf_lock(bp); 2027 error2 = bp->b_error; 2028 xfs_buf_relse(bp); 2029 if (!error) 2030 error = error2; 2031 } 2032 2033 return error; 2034 } 2035 2036 int __init 2037 xfs_buf_init(void) 2038 { 2039 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf", 2040 KM_ZONE_HWALIGN, NULL); 2041 if (!xfs_buf_zone) 2042 goto out; 2043 2044 return 0; 2045 2046 out: 2047 return -ENOMEM; 2048 } 2049 2050 void 2051 xfs_buf_terminate(void) 2052 { 2053 kmem_zone_destroy(xfs_buf_zone); 2054 } 2055