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 37 #include "xfs_log_format.h" 38 #include "xfs_trans_resv.h" 39 #include "xfs_sb.h" 40 #include "xfs_ag.h" 41 #include "xfs_mount.h" 42 #include "xfs_trace.h" 43 #include "xfs_log.h" 44 45 static kmem_zone_t *xfs_buf_zone; 46 47 static struct workqueue_struct *xfslogd_workqueue; 48 49 #ifdef XFS_BUF_LOCK_TRACKING 50 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid) 51 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1) 52 # define XB_GET_OWNER(bp) ((bp)->b_last_holder) 53 #else 54 # define XB_SET_OWNER(bp) do { } while (0) 55 # define XB_CLEAR_OWNER(bp) do { } while (0) 56 # define XB_GET_OWNER(bp) do { } while (0) 57 #endif 58 59 #define xb_to_gfp(flags) \ 60 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN) 61 62 63 static inline int 64 xfs_buf_is_vmapped( 65 struct xfs_buf *bp) 66 { 67 /* 68 * Return true if the buffer is vmapped. 69 * 70 * b_addr is null if the buffer is not mapped, but the code is clever 71 * enough to know it doesn't have to map a single page, so the check has 72 * to be both for b_addr and bp->b_page_count > 1. 73 */ 74 return bp->b_addr && bp->b_page_count > 1; 75 } 76 77 static inline int 78 xfs_buf_vmap_len( 79 struct xfs_buf *bp) 80 { 81 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset; 82 } 83 84 /* 85 * When we mark a buffer stale, we remove the buffer from the LRU and clear the 86 * b_lru_ref count so that the buffer is freed immediately when the buffer 87 * reference count falls to zero. If the buffer is already on the LRU, we need 88 * to remove the reference that LRU holds on the buffer. 89 * 90 * This prevents build-up of stale buffers on the LRU. 91 */ 92 void 93 xfs_buf_stale( 94 struct xfs_buf *bp) 95 { 96 ASSERT(xfs_buf_islocked(bp)); 97 98 bp->b_flags |= XBF_STALE; 99 100 /* 101 * Clear the delwri status so that a delwri queue walker will not 102 * flush this buffer to disk now that it is stale. The delwri queue has 103 * a reference to the buffer, so this is safe to do. 104 */ 105 bp->b_flags &= ~_XBF_DELWRI_Q; 106 107 spin_lock(&bp->b_lock); 108 atomic_set(&bp->b_lru_ref, 0); 109 if (!(bp->b_state & XFS_BSTATE_DISPOSE) && 110 (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru))) 111 atomic_dec(&bp->b_hold); 112 113 ASSERT(atomic_read(&bp->b_hold) >= 1); 114 spin_unlock(&bp->b_lock); 115 } 116 117 static int 118 xfs_buf_get_maps( 119 struct xfs_buf *bp, 120 int map_count) 121 { 122 ASSERT(bp->b_maps == NULL); 123 bp->b_map_count = map_count; 124 125 if (map_count == 1) { 126 bp->b_maps = &bp->__b_map; 127 return 0; 128 } 129 130 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map), 131 KM_NOFS); 132 if (!bp->b_maps) 133 return ENOMEM; 134 return 0; 135 } 136 137 /* 138 * Frees b_pages if it was allocated. 139 */ 140 static void 141 xfs_buf_free_maps( 142 struct xfs_buf *bp) 143 { 144 if (bp->b_maps != &bp->__b_map) { 145 kmem_free(bp->b_maps); 146 bp->b_maps = NULL; 147 } 148 } 149 150 struct xfs_buf * 151 _xfs_buf_alloc( 152 struct xfs_buftarg *target, 153 struct xfs_buf_map *map, 154 int nmaps, 155 xfs_buf_flags_t flags) 156 { 157 struct xfs_buf *bp; 158 int error; 159 int i; 160 161 bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS); 162 if (unlikely(!bp)) 163 return NULL; 164 165 /* 166 * We don't want certain flags to appear in b_flags unless they are 167 * specifically set by later operations on the buffer. 168 */ 169 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD); 170 171 atomic_set(&bp->b_hold, 1); 172 atomic_set(&bp->b_lru_ref, 1); 173 init_completion(&bp->b_iowait); 174 INIT_LIST_HEAD(&bp->b_lru); 175 INIT_LIST_HEAD(&bp->b_list); 176 RB_CLEAR_NODE(&bp->b_rbnode); 177 sema_init(&bp->b_sema, 0); /* held, no waiters */ 178 spin_lock_init(&bp->b_lock); 179 XB_SET_OWNER(bp); 180 bp->b_target = target; 181 bp->b_flags = flags; 182 183 /* 184 * Set length and io_length to the same value initially. 185 * I/O routines should use io_length, which will be the same in 186 * most cases but may be reset (e.g. XFS recovery). 187 */ 188 error = xfs_buf_get_maps(bp, nmaps); 189 if (error) { 190 kmem_zone_free(xfs_buf_zone, bp); 191 return NULL; 192 } 193 194 bp->b_bn = map[0].bm_bn; 195 bp->b_length = 0; 196 for (i = 0; i < nmaps; i++) { 197 bp->b_maps[i].bm_bn = map[i].bm_bn; 198 bp->b_maps[i].bm_len = map[i].bm_len; 199 bp->b_length += map[i].bm_len; 200 } 201 bp->b_io_length = bp->b_length; 202 203 atomic_set(&bp->b_pin_count, 0); 204 init_waitqueue_head(&bp->b_waiters); 205 206 XFS_STATS_INC(xb_create); 207 trace_xfs_buf_init(bp, _RET_IP_); 208 209 return bp; 210 } 211 212 /* 213 * Allocate a page array capable of holding a specified number 214 * of pages, and point the page buf at it. 215 */ 216 STATIC int 217 _xfs_buf_get_pages( 218 xfs_buf_t *bp, 219 int page_count, 220 xfs_buf_flags_t flags) 221 { 222 /* Make sure that we have a page list */ 223 if (bp->b_pages == NULL) { 224 bp->b_page_count = page_count; 225 if (page_count <= XB_PAGES) { 226 bp->b_pages = bp->b_page_array; 227 } else { 228 bp->b_pages = kmem_alloc(sizeof(struct page *) * 229 page_count, KM_NOFS); 230 if (bp->b_pages == NULL) 231 return -ENOMEM; 232 } 233 memset(bp->b_pages, 0, sizeof(struct page *) * page_count); 234 } 235 return 0; 236 } 237 238 /* 239 * Frees b_pages if it was allocated. 240 */ 241 STATIC void 242 _xfs_buf_free_pages( 243 xfs_buf_t *bp) 244 { 245 if (bp->b_pages != bp->b_page_array) { 246 kmem_free(bp->b_pages); 247 bp->b_pages = NULL; 248 } 249 } 250 251 /* 252 * Releases the specified buffer. 253 * 254 * The modification state of any associated pages is left unchanged. 255 * The buffer must not be on any hash - use xfs_buf_rele instead for 256 * hashed and refcounted buffers 257 */ 258 void 259 xfs_buf_free( 260 xfs_buf_t *bp) 261 { 262 trace_xfs_buf_free(bp, _RET_IP_); 263 264 ASSERT(list_empty(&bp->b_lru)); 265 266 if (bp->b_flags & _XBF_PAGES) { 267 uint i; 268 269 if (xfs_buf_is_vmapped(bp)) 270 vm_unmap_ram(bp->b_addr - bp->b_offset, 271 bp->b_page_count); 272 273 for (i = 0; i < bp->b_page_count; i++) { 274 struct page *page = bp->b_pages[i]; 275 276 __free_page(page); 277 } 278 } else if (bp->b_flags & _XBF_KMEM) 279 kmem_free(bp->b_addr); 280 _xfs_buf_free_pages(bp); 281 xfs_buf_free_maps(bp); 282 kmem_zone_free(xfs_buf_zone, bp); 283 } 284 285 /* 286 * Allocates all the pages for buffer in question and builds it's page list. 287 */ 288 STATIC int 289 xfs_buf_allocate_memory( 290 xfs_buf_t *bp, 291 uint flags) 292 { 293 size_t size; 294 size_t nbytes, offset; 295 gfp_t gfp_mask = xb_to_gfp(flags); 296 unsigned short page_count, i; 297 xfs_off_t start, end; 298 int error; 299 300 /* 301 * for buffers that are contained within a single page, just allocate 302 * the memory from the heap - there's no need for the complexity of 303 * page arrays to keep allocation down to order 0. 304 */ 305 size = BBTOB(bp->b_length); 306 if (size < PAGE_SIZE) { 307 bp->b_addr = kmem_alloc(size, KM_NOFS); 308 if (!bp->b_addr) { 309 /* low memory - use alloc_page loop instead */ 310 goto use_alloc_page; 311 } 312 313 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) != 314 ((unsigned long)bp->b_addr & PAGE_MASK)) { 315 /* b_addr spans two pages - use alloc_page instead */ 316 kmem_free(bp->b_addr); 317 bp->b_addr = NULL; 318 goto use_alloc_page; 319 } 320 bp->b_offset = offset_in_page(bp->b_addr); 321 bp->b_pages = bp->b_page_array; 322 bp->b_pages[0] = virt_to_page(bp->b_addr); 323 bp->b_page_count = 1; 324 bp->b_flags |= _XBF_KMEM; 325 return 0; 326 } 327 328 use_alloc_page: 329 start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT; 330 end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1) 331 >> PAGE_SHIFT; 332 page_count = end - start; 333 error = _xfs_buf_get_pages(bp, page_count, flags); 334 if (unlikely(error)) 335 return error; 336 337 offset = bp->b_offset; 338 bp->b_flags |= _XBF_PAGES; 339 340 for (i = 0; i < bp->b_page_count; i++) { 341 struct page *page; 342 uint retries = 0; 343 retry: 344 page = alloc_page(gfp_mask); 345 if (unlikely(page == NULL)) { 346 if (flags & XBF_READ_AHEAD) { 347 bp->b_page_count = i; 348 error = ENOMEM; 349 goto out_free_pages; 350 } 351 352 /* 353 * This could deadlock. 354 * 355 * But until all the XFS lowlevel code is revamped to 356 * handle buffer allocation failures we can't do much. 357 */ 358 if (!(++retries % 100)) 359 xfs_err(NULL, 360 "possible memory allocation deadlock in %s (mode:0x%x)", 361 __func__, gfp_mask); 362 363 XFS_STATS_INC(xb_page_retries); 364 congestion_wait(BLK_RW_ASYNC, HZ/50); 365 goto retry; 366 } 367 368 XFS_STATS_INC(xb_page_found); 369 370 nbytes = min_t(size_t, size, PAGE_SIZE - offset); 371 size -= nbytes; 372 bp->b_pages[i] = page; 373 offset = 0; 374 } 375 return 0; 376 377 out_free_pages: 378 for (i = 0; i < bp->b_page_count; i++) 379 __free_page(bp->b_pages[i]); 380 return error; 381 } 382 383 /* 384 * Map buffer into kernel address-space if necessary. 385 */ 386 STATIC int 387 _xfs_buf_map_pages( 388 xfs_buf_t *bp, 389 uint flags) 390 { 391 ASSERT(bp->b_flags & _XBF_PAGES); 392 if (bp->b_page_count == 1) { 393 /* A single page buffer is always mappable */ 394 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset; 395 } else if (flags & XBF_UNMAPPED) { 396 bp->b_addr = NULL; 397 } else { 398 int retried = 0; 399 400 do { 401 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count, 402 -1, PAGE_KERNEL); 403 if (bp->b_addr) 404 break; 405 vm_unmap_aliases(); 406 } while (retried++ <= 1); 407 408 if (!bp->b_addr) 409 return -ENOMEM; 410 bp->b_addr += bp->b_offset; 411 } 412 413 return 0; 414 } 415 416 /* 417 * Finding and Reading Buffers 418 */ 419 420 /* 421 * Look up, and creates if absent, a lockable buffer for 422 * a given range of an inode. The buffer is returned 423 * locked. No I/O is implied by this call. 424 */ 425 xfs_buf_t * 426 _xfs_buf_find( 427 struct xfs_buftarg *btp, 428 struct xfs_buf_map *map, 429 int nmaps, 430 xfs_buf_flags_t flags, 431 xfs_buf_t *new_bp) 432 { 433 size_t numbytes; 434 struct xfs_perag *pag; 435 struct rb_node **rbp; 436 struct rb_node *parent; 437 xfs_buf_t *bp; 438 xfs_daddr_t blkno = map[0].bm_bn; 439 xfs_daddr_t eofs; 440 int numblks = 0; 441 int i; 442 443 for (i = 0; i < nmaps; i++) 444 numblks += map[i].bm_len; 445 numbytes = BBTOB(numblks); 446 447 /* Check for IOs smaller than the sector size / not sector aligned */ 448 ASSERT(!(numbytes < btp->bt_meta_sectorsize)); 449 ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_meta_sectormask)); 450 451 /* 452 * Corrupted block numbers can get through to here, unfortunately, so we 453 * have to check that the buffer falls within the filesystem bounds. 454 */ 455 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks); 456 if (blkno >= eofs) { 457 /* 458 * XXX (dgc): we should really be returning EFSCORRUPTED here, 459 * but none of the higher level infrastructure supports 460 * returning a specific error on buffer lookup failures. 461 */ 462 xfs_alert(btp->bt_mount, 463 "%s: Block out of range: block 0x%llx, EOFS 0x%llx ", 464 __func__, blkno, eofs); 465 WARN_ON(1); 466 return NULL; 467 } 468 469 /* get tree root */ 470 pag = xfs_perag_get(btp->bt_mount, 471 xfs_daddr_to_agno(btp->bt_mount, blkno)); 472 473 /* walk tree */ 474 spin_lock(&pag->pag_buf_lock); 475 rbp = &pag->pag_buf_tree.rb_node; 476 parent = NULL; 477 bp = NULL; 478 while (*rbp) { 479 parent = *rbp; 480 bp = rb_entry(parent, struct xfs_buf, b_rbnode); 481 482 if (blkno < bp->b_bn) 483 rbp = &(*rbp)->rb_left; 484 else if (blkno > bp->b_bn) 485 rbp = &(*rbp)->rb_right; 486 else { 487 /* 488 * found a block number match. If the range doesn't 489 * match, the only way this is allowed is if the buffer 490 * in the cache is stale and the transaction that made 491 * it stale has not yet committed. i.e. we are 492 * reallocating a busy extent. Skip this buffer and 493 * continue searching to the right for an exact match. 494 */ 495 if (bp->b_length != numblks) { 496 ASSERT(bp->b_flags & XBF_STALE); 497 rbp = &(*rbp)->rb_right; 498 continue; 499 } 500 atomic_inc(&bp->b_hold); 501 goto found; 502 } 503 } 504 505 /* No match found */ 506 if (new_bp) { 507 rb_link_node(&new_bp->b_rbnode, parent, rbp); 508 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree); 509 /* the buffer keeps the perag reference until it is freed */ 510 new_bp->b_pag = pag; 511 spin_unlock(&pag->pag_buf_lock); 512 } else { 513 XFS_STATS_INC(xb_miss_locked); 514 spin_unlock(&pag->pag_buf_lock); 515 xfs_perag_put(pag); 516 } 517 return new_bp; 518 519 found: 520 spin_unlock(&pag->pag_buf_lock); 521 xfs_perag_put(pag); 522 523 if (!xfs_buf_trylock(bp)) { 524 if (flags & XBF_TRYLOCK) { 525 xfs_buf_rele(bp); 526 XFS_STATS_INC(xb_busy_locked); 527 return NULL; 528 } 529 xfs_buf_lock(bp); 530 XFS_STATS_INC(xb_get_locked_waited); 531 } 532 533 /* 534 * if the buffer is stale, clear all the external state associated with 535 * it. We need to keep flags such as how we allocated the buffer memory 536 * intact here. 537 */ 538 if (bp->b_flags & XBF_STALE) { 539 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0); 540 ASSERT(bp->b_iodone == NULL); 541 bp->b_flags &= _XBF_KMEM | _XBF_PAGES; 542 bp->b_ops = NULL; 543 } 544 545 trace_xfs_buf_find(bp, flags, _RET_IP_); 546 XFS_STATS_INC(xb_get_locked); 547 return bp; 548 } 549 550 /* 551 * Assembles a buffer covering the specified range. The code is optimised for 552 * cache hits, as metadata intensive workloads will see 3 orders of magnitude 553 * more hits than misses. 554 */ 555 struct xfs_buf * 556 xfs_buf_get_map( 557 struct xfs_buftarg *target, 558 struct xfs_buf_map *map, 559 int nmaps, 560 xfs_buf_flags_t flags) 561 { 562 struct xfs_buf *bp; 563 struct xfs_buf *new_bp; 564 int error = 0; 565 566 bp = _xfs_buf_find(target, map, nmaps, flags, NULL); 567 if (likely(bp)) 568 goto found; 569 570 new_bp = _xfs_buf_alloc(target, map, nmaps, flags); 571 if (unlikely(!new_bp)) 572 return NULL; 573 574 error = xfs_buf_allocate_memory(new_bp, flags); 575 if (error) { 576 xfs_buf_free(new_bp); 577 return NULL; 578 } 579 580 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp); 581 if (!bp) { 582 xfs_buf_free(new_bp); 583 return NULL; 584 } 585 586 if (bp != new_bp) 587 xfs_buf_free(new_bp); 588 589 found: 590 if (!bp->b_addr) { 591 error = _xfs_buf_map_pages(bp, flags); 592 if (unlikely(error)) { 593 xfs_warn(target->bt_mount, 594 "%s: failed to map pagesn", __func__); 595 xfs_buf_relse(bp); 596 return NULL; 597 } 598 } 599 600 XFS_STATS_INC(xb_get); 601 trace_xfs_buf_get(bp, flags, _RET_IP_); 602 return bp; 603 } 604 605 STATIC int 606 _xfs_buf_read( 607 xfs_buf_t *bp, 608 xfs_buf_flags_t flags) 609 { 610 ASSERT(!(flags & XBF_WRITE)); 611 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL); 612 613 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD); 614 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD); 615 616 xfs_buf_iorequest(bp); 617 if (flags & XBF_ASYNC) 618 return 0; 619 return xfs_buf_iowait(bp); 620 } 621 622 xfs_buf_t * 623 xfs_buf_read_map( 624 struct xfs_buftarg *target, 625 struct xfs_buf_map *map, 626 int nmaps, 627 xfs_buf_flags_t flags, 628 const struct xfs_buf_ops *ops) 629 { 630 struct xfs_buf *bp; 631 632 flags |= XBF_READ; 633 634 bp = xfs_buf_get_map(target, map, nmaps, flags); 635 if (bp) { 636 trace_xfs_buf_read(bp, flags, _RET_IP_); 637 638 if (!XFS_BUF_ISDONE(bp)) { 639 XFS_STATS_INC(xb_get_read); 640 bp->b_ops = ops; 641 _xfs_buf_read(bp, flags); 642 } else if (flags & XBF_ASYNC) { 643 /* 644 * Read ahead call which is already satisfied, 645 * drop the buffer 646 */ 647 xfs_buf_relse(bp); 648 return NULL; 649 } else { 650 /* We do not want read in the flags */ 651 bp->b_flags &= ~XBF_READ; 652 } 653 } 654 655 return bp; 656 } 657 658 /* 659 * If we are not low on memory then do the readahead in a deadlock 660 * safe manner. 661 */ 662 void 663 xfs_buf_readahead_map( 664 struct xfs_buftarg *target, 665 struct xfs_buf_map *map, 666 int nmaps, 667 const struct xfs_buf_ops *ops) 668 { 669 if (bdi_read_congested(target->bt_bdi)) 670 return; 671 672 xfs_buf_read_map(target, map, nmaps, 673 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops); 674 } 675 676 /* 677 * Read an uncached buffer from disk. Allocates and returns a locked 678 * buffer containing the disk contents or nothing. 679 */ 680 struct xfs_buf * 681 xfs_buf_read_uncached( 682 struct xfs_buftarg *target, 683 xfs_daddr_t daddr, 684 size_t numblks, 685 int flags, 686 const struct xfs_buf_ops *ops) 687 { 688 struct xfs_buf *bp; 689 690 bp = xfs_buf_get_uncached(target, numblks, flags); 691 if (!bp) 692 return NULL; 693 694 /* set up the buffer for a read IO */ 695 ASSERT(bp->b_map_count == 1); 696 bp->b_bn = daddr; 697 bp->b_maps[0].bm_bn = daddr; 698 bp->b_flags |= XBF_READ; 699 bp->b_ops = ops; 700 701 if (XFS_FORCED_SHUTDOWN(target->bt_mount)) { 702 xfs_buf_relse(bp); 703 return NULL; 704 } 705 xfs_buf_iorequest(bp); 706 xfs_buf_iowait(bp); 707 return bp; 708 } 709 710 /* 711 * Return a buffer allocated as an empty buffer and associated to external 712 * memory via xfs_buf_associate_memory() back to it's empty state. 713 */ 714 void 715 xfs_buf_set_empty( 716 struct xfs_buf *bp, 717 size_t numblks) 718 { 719 if (bp->b_pages) 720 _xfs_buf_free_pages(bp); 721 722 bp->b_pages = NULL; 723 bp->b_page_count = 0; 724 bp->b_addr = NULL; 725 bp->b_length = numblks; 726 bp->b_io_length = numblks; 727 728 ASSERT(bp->b_map_count == 1); 729 bp->b_bn = XFS_BUF_DADDR_NULL; 730 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL; 731 bp->b_maps[0].bm_len = bp->b_length; 732 } 733 734 static inline struct page * 735 mem_to_page( 736 void *addr) 737 { 738 if ((!is_vmalloc_addr(addr))) { 739 return virt_to_page(addr); 740 } else { 741 return vmalloc_to_page(addr); 742 } 743 } 744 745 int 746 xfs_buf_associate_memory( 747 xfs_buf_t *bp, 748 void *mem, 749 size_t len) 750 { 751 int rval; 752 int i = 0; 753 unsigned long pageaddr; 754 unsigned long offset; 755 size_t buflen; 756 int page_count; 757 758 pageaddr = (unsigned long)mem & PAGE_MASK; 759 offset = (unsigned long)mem - pageaddr; 760 buflen = PAGE_ALIGN(len + offset); 761 page_count = buflen >> PAGE_SHIFT; 762 763 /* Free any previous set of page pointers */ 764 if (bp->b_pages) 765 _xfs_buf_free_pages(bp); 766 767 bp->b_pages = NULL; 768 bp->b_addr = mem; 769 770 rval = _xfs_buf_get_pages(bp, page_count, 0); 771 if (rval) 772 return rval; 773 774 bp->b_offset = offset; 775 776 for (i = 0; i < bp->b_page_count; i++) { 777 bp->b_pages[i] = mem_to_page((void *)pageaddr); 778 pageaddr += PAGE_SIZE; 779 } 780 781 bp->b_io_length = BTOBB(len); 782 bp->b_length = BTOBB(buflen); 783 784 return 0; 785 } 786 787 xfs_buf_t * 788 xfs_buf_get_uncached( 789 struct xfs_buftarg *target, 790 size_t numblks, 791 int flags) 792 { 793 unsigned long page_count; 794 int error, i; 795 struct xfs_buf *bp; 796 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks); 797 798 bp = _xfs_buf_alloc(target, &map, 1, 0); 799 if (unlikely(bp == NULL)) 800 goto fail; 801 802 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT; 803 error = _xfs_buf_get_pages(bp, page_count, 0); 804 if (error) 805 goto fail_free_buf; 806 807 for (i = 0; i < page_count; i++) { 808 bp->b_pages[i] = alloc_page(xb_to_gfp(flags)); 809 if (!bp->b_pages[i]) 810 goto fail_free_mem; 811 } 812 bp->b_flags |= _XBF_PAGES; 813 814 error = _xfs_buf_map_pages(bp, 0); 815 if (unlikely(error)) { 816 xfs_warn(target->bt_mount, 817 "%s: failed to map pages", __func__); 818 goto fail_free_mem; 819 } 820 821 trace_xfs_buf_get_uncached(bp, _RET_IP_); 822 return bp; 823 824 fail_free_mem: 825 while (--i >= 0) 826 __free_page(bp->b_pages[i]); 827 _xfs_buf_free_pages(bp); 828 fail_free_buf: 829 xfs_buf_free_maps(bp); 830 kmem_zone_free(xfs_buf_zone, bp); 831 fail: 832 return NULL; 833 } 834 835 /* 836 * Increment reference count on buffer, to hold the buffer concurrently 837 * with another thread which may release (free) the buffer asynchronously. 838 * Must hold the buffer already to call this function. 839 */ 840 void 841 xfs_buf_hold( 842 xfs_buf_t *bp) 843 { 844 trace_xfs_buf_hold(bp, _RET_IP_); 845 atomic_inc(&bp->b_hold); 846 } 847 848 /* 849 * Releases a hold on the specified buffer. If the 850 * the hold count is 1, calls xfs_buf_free. 851 */ 852 void 853 xfs_buf_rele( 854 xfs_buf_t *bp) 855 { 856 struct xfs_perag *pag = bp->b_pag; 857 858 trace_xfs_buf_rele(bp, _RET_IP_); 859 860 if (!pag) { 861 ASSERT(list_empty(&bp->b_lru)); 862 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode)); 863 if (atomic_dec_and_test(&bp->b_hold)) 864 xfs_buf_free(bp); 865 return; 866 } 867 868 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode)); 869 870 ASSERT(atomic_read(&bp->b_hold) > 0); 871 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) { 872 spin_lock(&bp->b_lock); 873 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) { 874 /* 875 * If the buffer is added to the LRU take a new 876 * reference to the buffer for the LRU and clear the 877 * (now stale) dispose list state flag 878 */ 879 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) { 880 bp->b_state &= ~XFS_BSTATE_DISPOSE; 881 atomic_inc(&bp->b_hold); 882 } 883 spin_unlock(&bp->b_lock); 884 spin_unlock(&pag->pag_buf_lock); 885 } else { 886 /* 887 * most of the time buffers will already be removed from 888 * the LRU, so optimise that case by checking for the 889 * XFS_BSTATE_DISPOSE flag indicating the last list the 890 * buffer was on was the disposal list 891 */ 892 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) { 893 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru); 894 } else { 895 ASSERT(list_empty(&bp->b_lru)); 896 } 897 spin_unlock(&bp->b_lock); 898 899 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); 900 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree); 901 spin_unlock(&pag->pag_buf_lock); 902 xfs_perag_put(pag); 903 xfs_buf_free(bp); 904 } 905 } 906 } 907 908 909 /* 910 * Lock a buffer object, if it is not already locked. 911 * 912 * If we come across a stale, pinned, locked buffer, we know that we are 913 * being asked to lock a buffer that has been reallocated. Because it is 914 * pinned, we know that the log has not been pushed to disk and hence it 915 * will still be locked. Rather than continuing to have trylock attempts 916 * fail until someone else pushes the log, push it ourselves before 917 * returning. This means that the xfsaild will not get stuck trying 918 * to push on stale inode buffers. 919 */ 920 int 921 xfs_buf_trylock( 922 struct xfs_buf *bp) 923 { 924 int locked; 925 926 locked = down_trylock(&bp->b_sema) == 0; 927 if (locked) 928 XB_SET_OWNER(bp); 929 930 trace_xfs_buf_trylock(bp, _RET_IP_); 931 return locked; 932 } 933 934 /* 935 * Lock a buffer object. 936 * 937 * If we come across a stale, pinned, locked buffer, we know that we 938 * are being asked to lock a buffer that has been reallocated. Because 939 * it is pinned, we know that the log has not been pushed to disk and 940 * hence it will still be locked. Rather than sleeping until someone 941 * else pushes the log, push it ourselves before trying to get the lock. 942 */ 943 void 944 xfs_buf_lock( 945 struct xfs_buf *bp) 946 { 947 trace_xfs_buf_lock(bp, _RET_IP_); 948 949 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE)) 950 xfs_log_force(bp->b_target->bt_mount, 0); 951 down(&bp->b_sema); 952 XB_SET_OWNER(bp); 953 954 trace_xfs_buf_lock_done(bp, _RET_IP_); 955 } 956 957 void 958 xfs_buf_unlock( 959 struct xfs_buf *bp) 960 { 961 XB_CLEAR_OWNER(bp); 962 up(&bp->b_sema); 963 964 trace_xfs_buf_unlock(bp, _RET_IP_); 965 } 966 967 STATIC void 968 xfs_buf_wait_unpin( 969 xfs_buf_t *bp) 970 { 971 DECLARE_WAITQUEUE (wait, current); 972 973 if (atomic_read(&bp->b_pin_count) == 0) 974 return; 975 976 add_wait_queue(&bp->b_waiters, &wait); 977 for (;;) { 978 set_current_state(TASK_UNINTERRUPTIBLE); 979 if (atomic_read(&bp->b_pin_count) == 0) 980 break; 981 io_schedule(); 982 } 983 remove_wait_queue(&bp->b_waiters, &wait); 984 set_current_state(TASK_RUNNING); 985 } 986 987 /* 988 * Buffer Utility Routines 989 */ 990 991 STATIC void 992 xfs_buf_iodone_work( 993 struct work_struct *work) 994 { 995 struct xfs_buf *bp = 996 container_of(work, xfs_buf_t, b_iodone_work); 997 bool read = !!(bp->b_flags & XBF_READ); 998 999 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); 1000 1001 /* only validate buffers that were read without errors */ 1002 if (read && bp->b_ops && !bp->b_error && (bp->b_flags & XBF_DONE)) 1003 bp->b_ops->verify_read(bp); 1004 1005 if (bp->b_iodone) 1006 (*(bp->b_iodone))(bp); 1007 else if (bp->b_flags & XBF_ASYNC) 1008 xfs_buf_relse(bp); 1009 else { 1010 ASSERT(read && bp->b_ops); 1011 complete(&bp->b_iowait); 1012 } 1013 } 1014 1015 void 1016 xfs_buf_ioend( 1017 struct xfs_buf *bp, 1018 int schedule) 1019 { 1020 bool read = !!(bp->b_flags & XBF_READ); 1021 1022 trace_xfs_buf_iodone(bp, _RET_IP_); 1023 1024 if (bp->b_error == 0) 1025 bp->b_flags |= XBF_DONE; 1026 1027 if (bp->b_iodone || (read && bp->b_ops) || (bp->b_flags & XBF_ASYNC)) { 1028 if (schedule) { 1029 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work); 1030 queue_work(xfslogd_workqueue, &bp->b_iodone_work); 1031 } else { 1032 xfs_buf_iodone_work(&bp->b_iodone_work); 1033 } 1034 } else { 1035 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); 1036 complete(&bp->b_iowait); 1037 } 1038 } 1039 1040 void 1041 xfs_buf_ioerror( 1042 xfs_buf_t *bp, 1043 int error) 1044 { 1045 ASSERT(error >= 0 && error <= 0xffff); 1046 bp->b_error = (unsigned short)error; 1047 trace_xfs_buf_ioerror(bp, error, _RET_IP_); 1048 } 1049 1050 void 1051 xfs_buf_ioerror_alert( 1052 struct xfs_buf *bp, 1053 const char *func) 1054 { 1055 xfs_alert(bp->b_target->bt_mount, 1056 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d", 1057 (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length); 1058 } 1059 1060 /* 1061 * Called when we want to stop a buffer from getting written or read. 1062 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend 1063 * so that the proper iodone callbacks get called. 1064 */ 1065 STATIC int 1066 xfs_bioerror( 1067 xfs_buf_t *bp) 1068 { 1069 #ifdef XFSERRORDEBUG 1070 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone); 1071 #endif 1072 1073 /* 1074 * No need to wait until the buffer is unpinned, we aren't flushing it. 1075 */ 1076 xfs_buf_ioerror(bp, EIO); 1077 1078 /* 1079 * We're calling xfs_buf_ioend, so delete XBF_DONE flag. 1080 */ 1081 XFS_BUF_UNREAD(bp); 1082 XFS_BUF_UNDONE(bp); 1083 xfs_buf_stale(bp); 1084 1085 xfs_buf_ioend(bp, 0); 1086 1087 return EIO; 1088 } 1089 1090 /* 1091 * Same as xfs_bioerror, except that we are releasing the buffer 1092 * here ourselves, and avoiding the xfs_buf_ioend call. 1093 * This is meant for userdata errors; metadata bufs come with 1094 * iodone functions attached, so that we can track down errors. 1095 */ 1096 int 1097 xfs_bioerror_relse( 1098 struct xfs_buf *bp) 1099 { 1100 int64_t fl = bp->b_flags; 1101 /* 1102 * No need to wait until the buffer is unpinned. 1103 * We aren't flushing it. 1104 * 1105 * chunkhold expects B_DONE to be set, whether 1106 * we actually finish the I/O or not. We don't want to 1107 * change that interface. 1108 */ 1109 XFS_BUF_UNREAD(bp); 1110 XFS_BUF_DONE(bp); 1111 xfs_buf_stale(bp); 1112 bp->b_iodone = NULL; 1113 if (!(fl & XBF_ASYNC)) { 1114 /* 1115 * Mark b_error and B_ERROR _both_. 1116 * Lot's of chunkcache code assumes that. 1117 * There's no reason to mark error for 1118 * ASYNC buffers. 1119 */ 1120 xfs_buf_ioerror(bp, EIO); 1121 complete(&bp->b_iowait); 1122 } else { 1123 xfs_buf_relse(bp); 1124 } 1125 1126 return EIO; 1127 } 1128 1129 STATIC int 1130 xfs_bdstrat_cb( 1131 struct xfs_buf *bp) 1132 { 1133 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) { 1134 trace_xfs_bdstrat_shut(bp, _RET_IP_); 1135 /* 1136 * Metadata write that didn't get logged but 1137 * written delayed anyway. These aren't associated 1138 * with a transaction, and can be ignored. 1139 */ 1140 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp)) 1141 return xfs_bioerror_relse(bp); 1142 else 1143 return xfs_bioerror(bp); 1144 } 1145 1146 xfs_buf_iorequest(bp); 1147 return 0; 1148 } 1149 1150 int 1151 xfs_bwrite( 1152 struct xfs_buf *bp) 1153 { 1154 int error; 1155 1156 ASSERT(xfs_buf_islocked(bp)); 1157 1158 bp->b_flags |= XBF_WRITE; 1159 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q | XBF_WRITE_FAIL); 1160 1161 xfs_bdstrat_cb(bp); 1162 1163 error = xfs_buf_iowait(bp); 1164 if (error) { 1165 xfs_force_shutdown(bp->b_target->bt_mount, 1166 SHUTDOWN_META_IO_ERROR); 1167 } 1168 return error; 1169 } 1170 1171 STATIC void 1172 _xfs_buf_ioend( 1173 xfs_buf_t *bp, 1174 int schedule) 1175 { 1176 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) 1177 xfs_buf_ioend(bp, schedule); 1178 } 1179 1180 STATIC void 1181 xfs_buf_bio_end_io( 1182 struct bio *bio, 1183 int error) 1184 { 1185 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private; 1186 1187 /* 1188 * don't overwrite existing errors - otherwise we can lose errors on 1189 * buffers that require multiple bios to complete. 1190 */ 1191 if (!bp->b_error) 1192 xfs_buf_ioerror(bp, -error); 1193 1194 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ)) 1195 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp)); 1196 1197 _xfs_buf_ioend(bp, 1); 1198 bio_put(bio); 1199 } 1200 1201 static void 1202 xfs_buf_ioapply_map( 1203 struct xfs_buf *bp, 1204 int map, 1205 int *buf_offset, 1206 int *count, 1207 int rw) 1208 { 1209 int page_index; 1210 int total_nr_pages = bp->b_page_count; 1211 int nr_pages; 1212 struct bio *bio; 1213 sector_t sector = bp->b_maps[map].bm_bn; 1214 int size; 1215 int offset; 1216 1217 total_nr_pages = bp->b_page_count; 1218 1219 /* skip the pages in the buffer before the start offset */ 1220 page_index = 0; 1221 offset = *buf_offset; 1222 while (offset >= PAGE_SIZE) { 1223 page_index++; 1224 offset -= PAGE_SIZE; 1225 } 1226 1227 /* 1228 * Limit the IO size to the length of the current vector, and update the 1229 * remaining IO count for the next time around. 1230 */ 1231 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count); 1232 *count -= size; 1233 *buf_offset += size; 1234 1235 next_chunk: 1236 atomic_inc(&bp->b_io_remaining); 1237 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT); 1238 if (nr_pages > total_nr_pages) 1239 nr_pages = total_nr_pages; 1240 1241 bio = bio_alloc(GFP_NOIO, nr_pages); 1242 bio->bi_bdev = bp->b_target->bt_bdev; 1243 bio->bi_iter.bi_sector = sector; 1244 bio->bi_end_io = xfs_buf_bio_end_io; 1245 bio->bi_private = bp; 1246 1247 1248 for (; size && nr_pages; nr_pages--, page_index++) { 1249 int rbytes, nbytes = PAGE_SIZE - offset; 1250 1251 if (nbytes > size) 1252 nbytes = size; 1253 1254 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes, 1255 offset); 1256 if (rbytes < nbytes) 1257 break; 1258 1259 offset = 0; 1260 sector += BTOBB(nbytes); 1261 size -= nbytes; 1262 total_nr_pages--; 1263 } 1264 1265 if (likely(bio->bi_iter.bi_size)) { 1266 if (xfs_buf_is_vmapped(bp)) { 1267 flush_kernel_vmap_range(bp->b_addr, 1268 xfs_buf_vmap_len(bp)); 1269 } 1270 submit_bio(rw, bio); 1271 if (size) 1272 goto next_chunk; 1273 } else { 1274 /* 1275 * This is guaranteed not to be the last io reference count 1276 * because the caller (xfs_buf_iorequest) holds a count itself. 1277 */ 1278 atomic_dec(&bp->b_io_remaining); 1279 xfs_buf_ioerror(bp, EIO); 1280 bio_put(bio); 1281 } 1282 1283 } 1284 1285 STATIC void 1286 _xfs_buf_ioapply( 1287 struct xfs_buf *bp) 1288 { 1289 struct blk_plug plug; 1290 int rw; 1291 int offset; 1292 int size; 1293 int i; 1294 1295 /* 1296 * Make sure we capture only current IO errors rather than stale errors 1297 * left over from previous use of the buffer (e.g. failed readahead). 1298 */ 1299 bp->b_error = 0; 1300 1301 if (bp->b_flags & XBF_WRITE) { 1302 if (bp->b_flags & XBF_SYNCIO) 1303 rw = WRITE_SYNC; 1304 else 1305 rw = WRITE; 1306 if (bp->b_flags & XBF_FUA) 1307 rw |= REQ_FUA; 1308 if (bp->b_flags & XBF_FLUSH) 1309 rw |= REQ_FLUSH; 1310 1311 /* 1312 * Run the write verifier callback function if it exists. If 1313 * this function fails it will mark the buffer with an error and 1314 * the IO should not be dispatched. 1315 */ 1316 if (bp->b_ops) { 1317 bp->b_ops->verify_write(bp); 1318 if (bp->b_error) { 1319 xfs_force_shutdown(bp->b_target->bt_mount, 1320 SHUTDOWN_CORRUPT_INCORE); 1321 return; 1322 } 1323 } 1324 } else if (bp->b_flags & XBF_READ_AHEAD) { 1325 rw = READA; 1326 } else { 1327 rw = READ; 1328 } 1329 1330 /* we only use the buffer cache for meta-data */ 1331 rw |= REQ_META; 1332 1333 /* 1334 * Walk all the vectors issuing IO on them. Set up the initial offset 1335 * into the buffer and the desired IO size before we start - 1336 * _xfs_buf_ioapply_vec() will modify them appropriately for each 1337 * subsequent call. 1338 */ 1339 offset = bp->b_offset; 1340 size = BBTOB(bp->b_io_length); 1341 blk_start_plug(&plug); 1342 for (i = 0; i < bp->b_map_count; i++) { 1343 xfs_buf_ioapply_map(bp, i, &offset, &size, rw); 1344 if (bp->b_error) 1345 break; 1346 if (size <= 0) 1347 break; /* all done */ 1348 } 1349 blk_finish_plug(&plug); 1350 } 1351 1352 void 1353 xfs_buf_iorequest( 1354 xfs_buf_t *bp) 1355 { 1356 trace_xfs_buf_iorequest(bp, _RET_IP_); 1357 1358 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); 1359 1360 if (bp->b_flags & XBF_WRITE) 1361 xfs_buf_wait_unpin(bp); 1362 xfs_buf_hold(bp); 1363 1364 /* Set the count to 1 initially, this will stop an I/O 1365 * completion callout which happens before we have started 1366 * all the I/O from calling xfs_buf_ioend too early. 1367 */ 1368 atomic_set(&bp->b_io_remaining, 1); 1369 _xfs_buf_ioapply(bp); 1370 _xfs_buf_ioend(bp, 1); 1371 1372 xfs_buf_rele(bp); 1373 } 1374 1375 /* 1376 * Waits for I/O to complete on the buffer supplied. It returns immediately if 1377 * no I/O is pending or there is already a pending error on the buffer. It 1378 * returns the I/O error code, if any, or 0 if there was no error. 1379 */ 1380 int 1381 xfs_buf_iowait( 1382 xfs_buf_t *bp) 1383 { 1384 trace_xfs_buf_iowait(bp, _RET_IP_); 1385 1386 if (!bp->b_error) 1387 wait_for_completion(&bp->b_iowait); 1388 1389 trace_xfs_buf_iowait_done(bp, _RET_IP_); 1390 return bp->b_error; 1391 } 1392 1393 xfs_caddr_t 1394 xfs_buf_offset( 1395 xfs_buf_t *bp, 1396 size_t offset) 1397 { 1398 struct page *page; 1399 1400 if (bp->b_addr) 1401 return bp->b_addr + offset; 1402 1403 offset += bp->b_offset; 1404 page = bp->b_pages[offset >> PAGE_SHIFT]; 1405 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1)); 1406 } 1407 1408 /* 1409 * Move data into or out of a buffer. 1410 */ 1411 void 1412 xfs_buf_iomove( 1413 xfs_buf_t *bp, /* buffer to process */ 1414 size_t boff, /* starting buffer offset */ 1415 size_t bsize, /* length to copy */ 1416 void *data, /* data address */ 1417 xfs_buf_rw_t mode) /* read/write/zero flag */ 1418 { 1419 size_t bend; 1420 1421 bend = boff + bsize; 1422 while (boff < bend) { 1423 struct page *page; 1424 int page_index, page_offset, csize; 1425 1426 page_index = (boff + bp->b_offset) >> PAGE_SHIFT; 1427 page_offset = (boff + bp->b_offset) & ~PAGE_MASK; 1428 page = bp->b_pages[page_index]; 1429 csize = min_t(size_t, PAGE_SIZE - page_offset, 1430 BBTOB(bp->b_io_length) - boff); 1431 1432 ASSERT((csize + page_offset) <= PAGE_SIZE); 1433 1434 switch (mode) { 1435 case XBRW_ZERO: 1436 memset(page_address(page) + page_offset, 0, csize); 1437 break; 1438 case XBRW_READ: 1439 memcpy(data, page_address(page) + page_offset, csize); 1440 break; 1441 case XBRW_WRITE: 1442 memcpy(page_address(page) + page_offset, data, csize); 1443 } 1444 1445 boff += csize; 1446 data += csize; 1447 } 1448 } 1449 1450 /* 1451 * Handling of buffer targets (buftargs). 1452 */ 1453 1454 /* 1455 * Wait for any bufs with callbacks that have been submitted but have not yet 1456 * returned. These buffers will have an elevated hold count, so wait on those 1457 * while freeing all the buffers only held by the LRU. 1458 */ 1459 static enum lru_status 1460 xfs_buftarg_wait_rele( 1461 struct list_head *item, 1462 spinlock_t *lru_lock, 1463 void *arg) 1464 1465 { 1466 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru); 1467 struct list_head *dispose = arg; 1468 1469 if (atomic_read(&bp->b_hold) > 1) { 1470 /* need to wait, so skip it this pass */ 1471 trace_xfs_buf_wait_buftarg(bp, _RET_IP_); 1472 return LRU_SKIP; 1473 } 1474 if (!spin_trylock(&bp->b_lock)) 1475 return LRU_SKIP; 1476 1477 /* 1478 * clear the LRU reference count so the buffer doesn't get 1479 * ignored in xfs_buf_rele(). 1480 */ 1481 atomic_set(&bp->b_lru_ref, 0); 1482 bp->b_state |= XFS_BSTATE_DISPOSE; 1483 list_move(item, dispose); 1484 spin_unlock(&bp->b_lock); 1485 return LRU_REMOVED; 1486 } 1487 1488 void 1489 xfs_wait_buftarg( 1490 struct xfs_buftarg *btp) 1491 { 1492 LIST_HEAD(dispose); 1493 int loop = 0; 1494 1495 /* loop until there is nothing left on the lru list. */ 1496 while (list_lru_count(&btp->bt_lru)) { 1497 list_lru_walk(&btp->bt_lru, xfs_buftarg_wait_rele, 1498 &dispose, LONG_MAX); 1499 1500 while (!list_empty(&dispose)) { 1501 struct xfs_buf *bp; 1502 bp = list_first_entry(&dispose, struct xfs_buf, b_lru); 1503 list_del_init(&bp->b_lru); 1504 if (bp->b_flags & XBF_WRITE_FAIL) { 1505 xfs_alert(btp->bt_mount, 1506 "Corruption Alert: Buffer at block 0x%llx had permanent write failures!\n" 1507 "Please run xfs_repair to determine the extent of the problem.", 1508 (long long)bp->b_bn); 1509 } 1510 xfs_buf_rele(bp); 1511 } 1512 if (loop++ != 0) 1513 delay(100); 1514 } 1515 } 1516 1517 static enum lru_status 1518 xfs_buftarg_isolate( 1519 struct list_head *item, 1520 spinlock_t *lru_lock, 1521 void *arg) 1522 { 1523 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru); 1524 struct list_head *dispose = arg; 1525 1526 /* 1527 * we are inverting the lru lock/bp->b_lock here, so use a trylock. 1528 * If we fail to get the lock, just skip it. 1529 */ 1530 if (!spin_trylock(&bp->b_lock)) 1531 return LRU_SKIP; 1532 /* 1533 * Decrement the b_lru_ref count unless the value is already 1534 * zero. If the value is already zero, we need to reclaim the 1535 * buffer, otherwise it gets another trip through the LRU. 1536 */ 1537 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) { 1538 spin_unlock(&bp->b_lock); 1539 return LRU_ROTATE; 1540 } 1541 1542 bp->b_state |= XFS_BSTATE_DISPOSE; 1543 list_move(item, dispose); 1544 spin_unlock(&bp->b_lock); 1545 return LRU_REMOVED; 1546 } 1547 1548 static unsigned long 1549 xfs_buftarg_shrink_scan( 1550 struct shrinker *shrink, 1551 struct shrink_control *sc) 1552 { 1553 struct xfs_buftarg *btp = container_of(shrink, 1554 struct xfs_buftarg, bt_shrinker); 1555 LIST_HEAD(dispose); 1556 unsigned long freed; 1557 unsigned long nr_to_scan = sc->nr_to_scan; 1558 1559 freed = list_lru_walk_node(&btp->bt_lru, sc->nid, xfs_buftarg_isolate, 1560 &dispose, &nr_to_scan); 1561 1562 while (!list_empty(&dispose)) { 1563 struct xfs_buf *bp; 1564 bp = list_first_entry(&dispose, struct xfs_buf, b_lru); 1565 list_del_init(&bp->b_lru); 1566 xfs_buf_rele(bp); 1567 } 1568 1569 return freed; 1570 } 1571 1572 static unsigned long 1573 xfs_buftarg_shrink_count( 1574 struct shrinker *shrink, 1575 struct shrink_control *sc) 1576 { 1577 struct xfs_buftarg *btp = container_of(shrink, 1578 struct xfs_buftarg, bt_shrinker); 1579 return list_lru_count_node(&btp->bt_lru, sc->nid); 1580 } 1581 1582 void 1583 xfs_free_buftarg( 1584 struct xfs_mount *mp, 1585 struct xfs_buftarg *btp) 1586 { 1587 unregister_shrinker(&btp->bt_shrinker); 1588 list_lru_destroy(&btp->bt_lru); 1589 1590 if (mp->m_flags & XFS_MOUNT_BARRIER) 1591 xfs_blkdev_issue_flush(btp); 1592 1593 kmem_free(btp); 1594 } 1595 1596 int 1597 xfs_setsize_buftarg( 1598 xfs_buftarg_t *btp, 1599 unsigned int blocksize, 1600 unsigned int sectorsize) 1601 { 1602 /* Set up metadata sector size info */ 1603 btp->bt_meta_sectorsize = sectorsize; 1604 btp->bt_meta_sectormask = sectorsize - 1; 1605 1606 if (set_blocksize(btp->bt_bdev, sectorsize)) { 1607 char name[BDEVNAME_SIZE]; 1608 1609 bdevname(btp->bt_bdev, name); 1610 1611 xfs_warn(btp->bt_mount, 1612 "Cannot set_blocksize to %u on device %s", 1613 sectorsize, name); 1614 return EINVAL; 1615 } 1616 1617 /* Set up device logical sector size mask */ 1618 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev); 1619 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1; 1620 1621 return 0; 1622 } 1623 1624 /* 1625 * When allocating the initial buffer target we have not yet 1626 * read in the superblock, so don't know what sized sectors 1627 * are being used at this early stage. Play safe. 1628 */ 1629 STATIC int 1630 xfs_setsize_buftarg_early( 1631 xfs_buftarg_t *btp, 1632 struct block_device *bdev) 1633 { 1634 return xfs_setsize_buftarg(btp, PAGE_SIZE, 1635 bdev_logical_block_size(bdev)); 1636 } 1637 1638 xfs_buftarg_t * 1639 xfs_alloc_buftarg( 1640 struct xfs_mount *mp, 1641 struct block_device *bdev, 1642 int external, 1643 const char *fsname) 1644 { 1645 xfs_buftarg_t *btp; 1646 1647 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS); 1648 1649 btp->bt_mount = mp; 1650 btp->bt_dev = bdev->bd_dev; 1651 btp->bt_bdev = bdev; 1652 btp->bt_bdi = blk_get_backing_dev_info(bdev); 1653 if (!btp->bt_bdi) 1654 goto error; 1655 1656 if (xfs_setsize_buftarg_early(btp, bdev)) 1657 goto error; 1658 1659 if (list_lru_init(&btp->bt_lru)) 1660 goto error; 1661 1662 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count; 1663 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan; 1664 btp->bt_shrinker.seeks = DEFAULT_SEEKS; 1665 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE; 1666 register_shrinker(&btp->bt_shrinker); 1667 return btp; 1668 1669 error: 1670 kmem_free(btp); 1671 return NULL; 1672 } 1673 1674 /* 1675 * Add a buffer to the delayed write list. 1676 * 1677 * This queues a buffer for writeout if it hasn't already been. Note that 1678 * neither this routine nor the buffer list submission functions perform 1679 * any internal synchronization. It is expected that the lists are thread-local 1680 * to the callers. 1681 * 1682 * Returns true if we queued up the buffer, or false if it already had 1683 * been on the buffer list. 1684 */ 1685 bool 1686 xfs_buf_delwri_queue( 1687 struct xfs_buf *bp, 1688 struct list_head *list) 1689 { 1690 ASSERT(xfs_buf_islocked(bp)); 1691 ASSERT(!(bp->b_flags & XBF_READ)); 1692 1693 /* 1694 * If the buffer is already marked delwri it already is queued up 1695 * by someone else for imediate writeout. Just ignore it in that 1696 * case. 1697 */ 1698 if (bp->b_flags & _XBF_DELWRI_Q) { 1699 trace_xfs_buf_delwri_queued(bp, _RET_IP_); 1700 return false; 1701 } 1702 1703 trace_xfs_buf_delwri_queue(bp, _RET_IP_); 1704 1705 /* 1706 * If a buffer gets written out synchronously or marked stale while it 1707 * is on a delwri list we lazily remove it. To do this, the other party 1708 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone. 1709 * It remains referenced and on the list. In a rare corner case it 1710 * might get readded to a delwri list after the synchronous writeout, in 1711 * which case we need just need to re-add the flag here. 1712 */ 1713 bp->b_flags |= _XBF_DELWRI_Q; 1714 if (list_empty(&bp->b_list)) { 1715 atomic_inc(&bp->b_hold); 1716 list_add_tail(&bp->b_list, list); 1717 } 1718 1719 return true; 1720 } 1721 1722 /* 1723 * Compare function is more complex than it needs to be because 1724 * the return value is only 32 bits and we are doing comparisons 1725 * on 64 bit values 1726 */ 1727 static int 1728 xfs_buf_cmp( 1729 void *priv, 1730 struct list_head *a, 1731 struct list_head *b) 1732 { 1733 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list); 1734 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list); 1735 xfs_daddr_t diff; 1736 1737 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn; 1738 if (diff < 0) 1739 return -1; 1740 if (diff > 0) 1741 return 1; 1742 return 0; 1743 } 1744 1745 static int 1746 __xfs_buf_delwri_submit( 1747 struct list_head *buffer_list, 1748 struct list_head *io_list, 1749 bool wait) 1750 { 1751 struct blk_plug plug; 1752 struct xfs_buf *bp, *n; 1753 int pinned = 0; 1754 1755 list_for_each_entry_safe(bp, n, buffer_list, b_list) { 1756 if (!wait) { 1757 if (xfs_buf_ispinned(bp)) { 1758 pinned++; 1759 continue; 1760 } 1761 if (!xfs_buf_trylock(bp)) 1762 continue; 1763 } else { 1764 xfs_buf_lock(bp); 1765 } 1766 1767 /* 1768 * Someone else might have written the buffer synchronously or 1769 * marked it stale in the meantime. In that case only the 1770 * _XBF_DELWRI_Q flag got cleared, and we have to drop the 1771 * reference and remove it from the list here. 1772 */ 1773 if (!(bp->b_flags & _XBF_DELWRI_Q)) { 1774 list_del_init(&bp->b_list); 1775 xfs_buf_relse(bp); 1776 continue; 1777 } 1778 1779 list_move_tail(&bp->b_list, io_list); 1780 trace_xfs_buf_delwri_split(bp, _RET_IP_); 1781 } 1782 1783 list_sort(NULL, io_list, xfs_buf_cmp); 1784 1785 blk_start_plug(&plug); 1786 list_for_each_entry_safe(bp, n, io_list, b_list) { 1787 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC | XBF_WRITE_FAIL); 1788 bp->b_flags |= XBF_WRITE; 1789 1790 if (!wait) { 1791 bp->b_flags |= XBF_ASYNC; 1792 list_del_init(&bp->b_list); 1793 } 1794 xfs_bdstrat_cb(bp); 1795 } 1796 blk_finish_plug(&plug); 1797 1798 return pinned; 1799 } 1800 1801 /* 1802 * Write out a buffer list asynchronously. 1803 * 1804 * This will take the @buffer_list, write all non-locked and non-pinned buffers 1805 * out and not wait for I/O completion on any of the buffers. This interface 1806 * is only safely useable for callers that can track I/O completion by higher 1807 * level means, e.g. AIL pushing as the @buffer_list is consumed in this 1808 * function. 1809 */ 1810 int 1811 xfs_buf_delwri_submit_nowait( 1812 struct list_head *buffer_list) 1813 { 1814 LIST_HEAD (io_list); 1815 return __xfs_buf_delwri_submit(buffer_list, &io_list, false); 1816 } 1817 1818 /* 1819 * Write out a buffer list synchronously. 1820 * 1821 * This will take the @buffer_list, write all buffers out and wait for I/O 1822 * completion on all of the buffers. @buffer_list is consumed by the function, 1823 * so callers must have some other way of tracking buffers if they require such 1824 * functionality. 1825 */ 1826 int 1827 xfs_buf_delwri_submit( 1828 struct list_head *buffer_list) 1829 { 1830 LIST_HEAD (io_list); 1831 int error = 0, error2; 1832 struct xfs_buf *bp; 1833 1834 __xfs_buf_delwri_submit(buffer_list, &io_list, true); 1835 1836 /* Wait for IO to complete. */ 1837 while (!list_empty(&io_list)) { 1838 bp = list_first_entry(&io_list, struct xfs_buf, b_list); 1839 1840 list_del_init(&bp->b_list); 1841 error2 = xfs_buf_iowait(bp); 1842 xfs_buf_relse(bp); 1843 if (!error) 1844 error = error2; 1845 } 1846 1847 return error; 1848 } 1849 1850 int __init 1851 xfs_buf_init(void) 1852 { 1853 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf", 1854 KM_ZONE_HWALIGN, NULL); 1855 if (!xfs_buf_zone) 1856 goto out; 1857 1858 xfslogd_workqueue = alloc_workqueue("xfslogd", 1859 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1); 1860 if (!xfslogd_workqueue) 1861 goto out_free_buf_zone; 1862 1863 return 0; 1864 1865 out_free_buf_zone: 1866 kmem_zone_destroy(xfs_buf_zone); 1867 out: 1868 return -ENOMEM; 1869 } 1870 1871 void 1872 xfs_buf_terminate(void) 1873 { 1874 destroy_workqueue(xfslogd_workqueue); 1875 kmem_zone_destroy(xfs_buf_zone); 1876 } 1877