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_sb.h" 38 #include "xfs_log.h" 39 #include "xfs_ag.h" 40 #include "xfs_mount.h" 41 #include "xfs_trace.h" 42 43 static kmem_zone_t *xfs_buf_zone; 44 45 static struct workqueue_struct *xfslogd_workqueue; 46 47 #ifdef XFS_BUF_LOCK_TRACKING 48 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid) 49 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1) 50 # define XB_GET_OWNER(bp) ((bp)->b_last_holder) 51 #else 52 # define XB_SET_OWNER(bp) do { } while (0) 53 # define XB_CLEAR_OWNER(bp) do { } while (0) 54 # define XB_GET_OWNER(bp) do { } while (0) 55 #endif 56 57 #define xb_to_gfp(flags) \ 58 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN) 59 60 61 static inline int 62 xfs_buf_is_vmapped( 63 struct xfs_buf *bp) 64 { 65 /* 66 * Return true if the buffer is vmapped. 67 * 68 * b_addr is null if the buffer is not mapped, but the code is clever 69 * enough to know it doesn't have to map a single page, so the check has 70 * to be both for b_addr and bp->b_page_count > 1. 71 */ 72 return bp->b_addr && bp->b_page_count > 1; 73 } 74 75 static inline int 76 xfs_buf_vmap_len( 77 struct xfs_buf *bp) 78 { 79 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset; 80 } 81 82 /* 83 * xfs_buf_lru_add - add a buffer to the LRU. 84 * 85 * The LRU takes a new reference to the buffer so that it will only be freed 86 * once the shrinker takes the buffer off the LRU. 87 */ 88 STATIC void 89 xfs_buf_lru_add( 90 struct xfs_buf *bp) 91 { 92 struct xfs_buftarg *btp = bp->b_target; 93 94 spin_lock(&btp->bt_lru_lock); 95 if (list_empty(&bp->b_lru)) { 96 atomic_inc(&bp->b_hold); 97 list_add_tail(&bp->b_lru, &btp->bt_lru); 98 btp->bt_lru_nr++; 99 bp->b_lru_flags &= ~_XBF_LRU_DISPOSE; 100 } 101 spin_unlock(&btp->bt_lru_lock); 102 } 103 104 /* 105 * xfs_buf_lru_del - remove a buffer from the LRU 106 * 107 * The unlocked check is safe here because it only occurs when there are not 108 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there 109 * to optimise the shrinker removing the buffer from the LRU and calling 110 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the 111 * bt_lru_lock. 112 */ 113 STATIC void 114 xfs_buf_lru_del( 115 struct xfs_buf *bp) 116 { 117 struct xfs_buftarg *btp = bp->b_target; 118 119 if (list_empty(&bp->b_lru)) 120 return; 121 122 spin_lock(&btp->bt_lru_lock); 123 if (!list_empty(&bp->b_lru)) { 124 list_del_init(&bp->b_lru); 125 btp->bt_lru_nr--; 126 } 127 spin_unlock(&btp->bt_lru_lock); 128 } 129 130 /* 131 * When we mark a buffer stale, we remove the buffer from the LRU and clear the 132 * b_lru_ref count so that the buffer is freed immediately when the buffer 133 * reference count falls to zero. If the buffer is already on the LRU, we need 134 * to remove the reference that LRU holds on the buffer. 135 * 136 * This prevents build-up of stale buffers on the LRU. 137 */ 138 void 139 xfs_buf_stale( 140 struct xfs_buf *bp) 141 { 142 ASSERT(xfs_buf_islocked(bp)); 143 144 bp->b_flags |= XBF_STALE; 145 146 /* 147 * Clear the delwri status so that a delwri queue walker will not 148 * flush this buffer to disk now that it is stale. The delwri queue has 149 * a reference to the buffer, so this is safe to do. 150 */ 151 bp->b_flags &= ~_XBF_DELWRI_Q; 152 153 atomic_set(&(bp)->b_lru_ref, 0); 154 if (!list_empty(&bp->b_lru)) { 155 struct xfs_buftarg *btp = bp->b_target; 156 157 spin_lock(&btp->bt_lru_lock); 158 if (!list_empty(&bp->b_lru) && 159 !(bp->b_lru_flags & _XBF_LRU_DISPOSE)) { 160 list_del_init(&bp->b_lru); 161 btp->bt_lru_nr--; 162 atomic_dec(&bp->b_hold); 163 } 164 spin_unlock(&btp->bt_lru_lock); 165 } 166 ASSERT(atomic_read(&bp->b_hold) >= 1); 167 } 168 169 static int 170 xfs_buf_get_maps( 171 struct xfs_buf *bp, 172 int map_count) 173 { 174 ASSERT(bp->b_maps == NULL); 175 bp->b_map_count = map_count; 176 177 if (map_count == 1) { 178 bp->b_maps = &bp->__b_map; 179 return 0; 180 } 181 182 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map), 183 KM_NOFS); 184 if (!bp->b_maps) 185 return ENOMEM; 186 return 0; 187 } 188 189 /* 190 * Frees b_pages if it was allocated. 191 */ 192 static void 193 xfs_buf_free_maps( 194 struct xfs_buf *bp) 195 { 196 if (bp->b_maps != &bp->__b_map) { 197 kmem_free(bp->b_maps); 198 bp->b_maps = NULL; 199 } 200 } 201 202 struct xfs_buf * 203 _xfs_buf_alloc( 204 struct xfs_buftarg *target, 205 struct xfs_buf_map *map, 206 int nmaps, 207 xfs_buf_flags_t flags) 208 { 209 struct xfs_buf *bp; 210 int error; 211 int i; 212 213 bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS); 214 if (unlikely(!bp)) 215 return NULL; 216 217 /* 218 * We don't want certain flags to appear in b_flags unless they are 219 * specifically set by later operations on the buffer. 220 */ 221 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD); 222 223 atomic_set(&bp->b_hold, 1); 224 atomic_set(&bp->b_lru_ref, 1); 225 init_completion(&bp->b_iowait); 226 INIT_LIST_HEAD(&bp->b_lru); 227 INIT_LIST_HEAD(&bp->b_list); 228 RB_CLEAR_NODE(&bp->b_rbnode); 229 sema_init(&bp->b_sema, 0); /* held, no waiters */ 230 XB_SET_OWNER(bp); 231 bp->b_target = target; 232 bp->b_flags = flags; 233 234 /* 235 * Set length and io_length to the same value initially. 236 * I/O routines should use io_length, which will be the same in 237 * most cases but may be reset (e.g. XFS recovery). 238 */ 239 error = xfs_buf_get_maps(bp, nmaps); 240 if (error) { 241 kmem_zone_free(xfs_buf_zone, bp); 242 return NULL; 243 } 244 245 bp->b_bn = map[0].bm_bn; 246 bp->b_length = 0; 247 for (i = 0; i < nmaps; i++) { 248 bp->b_maps[i].bm_bn = map[i].bm_bn; 249 bp->b_maps[i].bm_len = map[i].bm_len; 250 bp->b_length += map[i].bm_len; 251 } 252 bp->b_io_length = bp->b_length; 253 254 atomic_set(&bp->b_pin_count, 0); 255 init_waitqueue_head(&bp->b_waiters); 256 257 XFS_STATS_INC(xb_create); 258 trace_xfs_buf_init(bp, _RET_IP_); 259 260 return bp; 261 } 262 263 /* 264 * Allocate a page array capable of holding a specified number 265 * of pages, and point the page buf at it. 266 */ 267 STATIC int 268 _xfs_buf_get_pages( 269 xfs_buf_t *bp, 270 int page_count, 271 xfs_buf_flags_t flags) 272 { 273 /* Make sure that we have a page list */ 274 if (bp->b_pages == NULL) { 275 bp->b_page_count = page_count; 276 if (page_count <= XB_PAGES) { 277 bp->b_pages = bp->b_page_array; 278 } else { 279 bp->b_pages = kmem_alloc(sizeof(struct page *) * 280 page_count, KM_NOFS); 281 if (bp->b_pages == NULL) 282 return -ENOMEM; 283 } 284 memset(bp->b_pages, 0, sizeof(struct page *) * page_count); 285 } 286 return 0; 287 } 288 289 /* 290 * Frees b_pages if it was allocated. 291 */ 292 STATIC void 293 _xfs_buf_free_pages( 294 xfs_buf_t *bp) 295 { 296 if (bp->b_pages != bp->b_page_array) { 297 kmem_free(bp->b_pages); 298 bp->b_pages = NULL; 299 } 300 } 301 302 /* 303 * Releases the specified buffer. 304 * 305 * The modification state of any associated pages is left unchanged. 306 * The buffer most not be on any hash - use xfs_buf_rele instead for 307 * hashed and refcounted buffers 308 */ 309 void 310 xfs_buf_free( 311 xfs_buf_t *bp) 312 { 313 trace_xfs_buf_free(bp, _RET_IP_); 314 315 ASSERT(list_empty(&bp->b_lru)); 316 317 if (bp->b_flags & _XBF_PAGES) { 318 uint i; 319 320 if (xfs_buf_is_vmapped(bp)) 321 vm_unmap_ram(bp->b_addr - bp->b_offset, 322 bp->b_page_count); 323 324 for (i = 0; i < bp->b_page_count; i++) { 325 struct page *page = bp->b_pages[i]; 326 327 __free_page(page); 328 } 329 } else if (bp->b_flags & _XBF_KMEM) 330 kmem_free(bp->b_addr); 331 _xfs_buf_free_pages(bp); 332 xfs_buf_free_maps(bp); 333 kmem_zone_free(xfs_buf_zone, bp); 334 } 335 336 /* 337 * Allocates all the pages for buffer in question and builds it's page list. 338 */ 339 STATIC int 340 xfs_buf_allocate_memory( 341 xfs_buf_t *bp, 342 uint flags) 343 { 344 size_t size; 345 size_t nbytes, offset; 346 gfp_t gfp_mask = xb_to_gfp(flags); 347 unsigned short page_count, i; 348 xfs_off_t start, end; 349 int error; 350 351 /* 352 * for buffers that are contained within a single page, just allocate 353 * the memory from the heap - there's no need for the complexity of 354 * page arrays to keep allocation down to order 0. 355 */ 356 size = BBTOB(bp->b_length); 357 if (size < PAGE_SIZE) { 358 bp->b_addr = kmem_alloc(size, KM_NOFS); 359 if (!bp->b_addr) { 360 /* low memory - use alloc_page loop instead */ 361 goto use_alloc_page; 362 } 363 364 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) != 365 ((unsigned long)bp->b_addr & PAGE_MASK)) { 366 /* b_addr spans two pages - use alloc_page instead */ 367 kmem_free(bp->b_addr); 368 bp->b_addr = NULL; 369 goto use_alloc_page; 370 } 371 bp->b_offset = offset_in_page(bp->b_addr); 372 bp->b_pages = bp->b_page_array; 373 bp->b_pages[0] = virt_to_page(bp->b_addr); 374 bp->b_page_count = 1; 375 bp->b_flags |= _XBF_KMEM; 376 return 0; 377 } 378 379 use_alloc_page: 380 start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT; 381 end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1) 382 >> PAGE_SHIFT; 383 page_count = end - start; 384 error = _xfs_buf_get_pages(bp, page_count, flags); 385 if (unlikely(error)) 386 return error; 387 388 offset = bp->b_offset; 389 bp->b_flags |= _XBF_PAGES; 390 391 for (i = 0; i < bp->b_page_count; i++) { 392 struct page *page; 393 uint retries = 0; 394 retry: 395 page = alloc_page(gfp_mask); 396 if (unlikely(page == NULL)) { 397 if (flags & XBF_READ_AHEAD) { 398 bp->b_page_count = i; 399 error = ENOMEM; 400 goto out_free_pages; 401 } 402 403 /* 404 * This could deadlock. 405 * 406 * But until all the XFS lowlevel code is revamped to 407 * handle buffer allocation failures we can't do much. 408 */ 409 if (!(++retries % 100)) 410 xfs_err(NULL, 411 "possible memory allocation deadlock in %s (mode:0x%x)", 412 __func__, gfp_mask); 413 414 XFS_STATS_INC(xb_page_retries); 415 congestion_wait(BLK_RW_ASYNC, HZ/50); 416 goto retry; 417 } 418 419 XFS_STATS_INC(xb_page_found); 420 421 nbytes = min_t(size_t, size, PAGE_SIZE - offset); 422 size -= nbytes; 423 bp->b_pages[i] = page; 424 offset = 0; 425 } 426 return 0; 427 428 out_free_pages: 429 for (i = 0; i < bp->b_page_count; i++) 430 __free_page(bp->b_pages[i]); 431 return error; 432 } 433 434 /* 435 * Map buffer into kernel address-space if necessary. 436 */ 437 STATIC int 438 _xfs_buf_map_pages( 439 xfs_buf_t *bp, 440 uint flags) 441 { 442 ASSERT(bp->b_flags & _XBF_PAGES); 443 if (bp->b_page_count == 1) { 444 /* A single page buffer is always mappable */ 445 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset; 446 } else if (flags & XBF_UNMAPPED) { 447 bp->b_addr = NULL; 448 } else { 449 int retried = 0; 450 451 do { 452 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count, 453 -1, PAGE_KERNEL); 454 if (bp->b_addr) 455 break; 456 vm_unmap_aliases(); 457 } while (retried++ <= 1); 458 459 if (!bp->b_addr) 460 return -ENOMEM; 461 bp->b_addr += bp->b_offset; 462 } 463 464 return 0; 465 } 466 467 /* 468 * Finding and Reading Buffers 469 */ 470 471 /* 472 * Look up, and creates if absent, a lockable buffer for 473 * a given range of an inode. The buffer is returned 474 * locked. No I/O is implied by this call. 475 */ 476 xfs_buf_t * 477 _xfs_buf_find( 478 struct xfs_buftarg *btp, 479 struct xfs_buf_map *map, 480 int nmaps, 481 xfs_buf_flags_t flags, 482 xfs_buf_t *new_bp) 483 { 484 size_t numbytes; 485 struct xfs_perag *pag; 486 struct rb_node **rbp; 487 struct rb_node *parent; 488 xfs_buf_t *bp; 489 xfs_daddr_t blkno = map[0].bm_bn; 490 xfs_daddr_t eofs; 491 int numblks = 0; 492 int i; 493 494 for (i = 0; i < nmaps; i++) 495 numblks += map[i].bm_len; 496 numbytes = BBTOB(numblks); 497 498 /* Check for IOs smaller than the sector size / not sector aligned */ 499 ASSERT(!(numbytes < (1 << btp->bt_sshift))); 500 ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_smask)); 501 502 /* 503 * Corrupted block numbers can get through to here, unfortunately, so we 504 * have to check that the buffer falls within the filesystem bounds. 505 */ 506 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks); 507 if (blkno >= eofs) { 508 /* 509 * XXX (dgc): we should really be returning EFSCORRUPTED here, 510 * but none of the higher level infrastructure supports 511 * returning a specific error on buffer lookup failures. 512 */ 513 xfs_alert(btp->bt_mount, 514 "%s: Block out of range: block 0x%llx, EOFS 0x%llx ", 515 __func__, blkno, eofs); 516 WARN_ON(1); 517 return NULL; 518 } 519 520 /* get tree root */ 521 pag = xfs_perag_get(btp->bt_mount, 522 xfs_daddr_to_agno(btp->bt_mount, blkno)); 523 524 /* walk tree */ 525 spin_lock(&pag->pag_buf_lock); 526 rbp = &pag->pag_buf_tree.rb_node; 527 parent = NULL; 528 bp = NULL; 529 while (*rbp) { 530 parent = *rbp; 531 bp = rb_entry(parent, struct xfs_buf, b_rbnode); 532 533 if (blkno < bp->b_bn) 534 rbp = &(*rbp)->rb_left; 535 else if (blkno > bp->b_bn) 536 rbp = &(*rbp)->rb_right; 537 else { 538 /* 539 * found a block number match. If the range doesn't 540 * match, the only way this is allowed is if the buffer 541 * in the cache is stale and the transaction that made 542 * it stale has not yet committed. i.e. we are 543 * reallocating a busy extent. Skip this buffer and 544 * continue searching to the right for an exact match. 545 */ 546 if (bp->b_length != numblks) { 547 ASSERT(bp->b_flags & XBF_STALE); 548 rbp = &(*rbp)->rb_right; 549 continue; 550 } 551 atomic_inc(&bp->b_hold); 552 goto found; 553 } 554 } 555 556 /* No match found */ 557 if (new_bp) { 558 rb_link_node(&new_bp->b_rbnode, parent, rbp); 559 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree); 560 /* the buffer keeps the perag reference until it is freed */ 561 new_bp->b_pag = pag; 562 spin_unlock(&pag->pag_buf_lock); 563 } else { 564 XFS_STATS_INC(xb_miss_locked); 565 spin_unlock(&pag->pag_buf_lock); 566 xfs_perag_put(pag); 567 } 568 return new_bp; 569 570 found: 571 spin_unlock(&pag->pag_buf_lock); 572 xfs_perag_put(pag); 573 574 if (!xfs_buf_trylock(bp)) { 575 if (flags & XBF_TRYLOCK) { 576 xfs_buf_rele(bp); 577 XFS_STATS_INC(xb_busy_locked); 578 return NULL; 579 } 580 xfs_buf_lock(bp); 581 XFS_STATS_INC(xb_get_locked_waited); 582 } 583 584 /* 585 * if the buffer is stale, clear all the external state associated with 586 * it. We need to keep flags such as how we allocated the buffer memory 587 * intact here. 588 */ 589 if (bp->b_flags & XBF_STALE) { 590 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0); 591 ASSERT(bp->b_iodone == NULL); 592 bp->b_flags &= _XBF_KMEM | _XBF_PAGES; 593 bp->b_ops = NULL; 594 } 595 596 trace_xfs_buf_find(bp, flags, _RET_IP_); 597 XFS_STATS_INC(xb_get_locked); 598 return bp; 599 } 600 601 /* 602 * Assembles a buffer covering the specified range. The code is optimised for 603 * cache hits, as metadata intensive workloads will see 3 orders of magnitude 604 * more hits than misses. 605 */ 606 struct xfs_buf * 607 xfs_buf_get_map( 608 struct xfs_buftarg *target, 609 struct xfs_buf_map *map, 610 int nmaps, 611 xfs_buf_flags_t flags) 612 { 613 struct xfs_buf *bp; 614 struct xfs_buf *new_bp; 615 int error = 0; 616 617 bp = _xfs_buf_find(target, map, nmaps, flags, NULL); 618 if (likely(bp)) 619 goto found; 620 621 new_bp = _xfs_buf_alloc(target, map, nmaps, flags); 622 if (unlikely(!new_bp)) 623 return NULL; 624 625 error = xfs_buf_allocate_memory(new_bp, flags); 626 if (error) { 627 xfs_buf_free(new_bp); 628 return NULL; 629 } 630 631 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp); 632 if (!bp) { 633 xfs_buf_free(new_bp); 634 return NULL; 635 } 636 637 if (bp != new_bp) 638 xfs_buf_free(new_bp); 639 640 found: 641 if (!bp->b_addr) { 642 error = _xfs_buf_map_pages(bp, flags); 643 if (unlikely(error)) { 644 xfs_warn(target->bt_mount, 645 "%s: failed to map pages\n", __func__); 646 xfs_buf_relse(bp); 647 return NULL; 648 } 649 } 650 651 XFS_STATS_INC(xb_get); 652 trace_xfs_buf_get(bp, flags, _RET_IP_); 653 return bp; 654 } 655 656 STATIC int 657 _xfs_buf_read( 658 xfs_buf_t *bp, 659 xfs_buf_flags_t flags) 660 { 661 ASSERT(!(flags & XBF_WRITE)); 662 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL); 663 664 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD); 665 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD); 666 667 xfs_buf_iorequest(bp); 668 if (flags & XBF_ASYNC) 669 return 0; 670 return xfs_buf_iowait(bp); 671 } 672 673 xfs_buf_t * 674 xfs_buf_read_map( 675 struct xfs_buftarg *target, 676 struct xfs_buf_map *map, 677 int nmaps, 678 xfs_buf_flags_t flags, 679 const struct xfs_buf_ops *ops) 680 { 681 struct xfs_buf *bp; 682 683 flags |= XBF_READ; 684 685 bp = xfs_buf_get_map(target, map, nmaps, flags); 686 if (bp) { 687 trace_xfs_buf_read(bp, flags, _RET_IP_); 688 689 if (!XFS_BUF_ISDONE(bp)) { 690 XFS_STATS_INC(xb_get_read); 691 bp->b_ops = ops; 692 _xfs_buf_read(bp, flags); 693 } else if (flags & XBF_ASYNC) { 694 /* 695 * Read ahead call which is already satisfied, 696 * drop the buffer 697 */ 698 xfs_buf_relse(bp); 699 return NULL; 700 } else { 701 /* We do not want read in the flags */ 702 bp->b_flags &= ~XBF_READ; 703 } 704 } 705 706 return bp; 707 } 708 709 /* 710 * If we are not low on memory then do the readahead in a deadlock 711 * safe manner. 712 */ 713 void 714 xfs_buf_readahead_map( 715 struct xfs_buftarg *target, 716 struct xfs_buf_map *map, 717 int nmaps, 718 const struct xfs_buf_ops *ops) 719 { 720 if (bdi_read_congested(target->bt_bdi)) 721 return; 722 723 xfs_buf_read_map(target, map, nmaps, 724 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops); 725 } 726 727 /* 728 * Read an uncached buffer from disk. Allocates and returns a locked 729 * buffer containing the disk contents or nothing. 730 */ 731 struct xfs_buf * 732 xfs_buf_read_uncached( 733 struct xfs_buftarg *target, 734 xfs_daddr_t daddr, 735 size_t numblks, 736 int flags, 737 const struct xfs_buf_ops *ops) 738 { 739 struct xfs_buf *bp; 740 741 bp = xfs_buf_get_uncached(target, numblks, flags); 742 if (!bp) 743 return NULL; 744 745 /* set up the buffer for a read IO */ 746 ASSERT(bp->b_map_count == 1); 747 bp->b_bn = daddr; 748 bp->b_maps[0].bm_bn = daddr; 749 bp->b_flags |= XBF_READ; 750 bp->b_ops = ops; 751 752 xfsbdstrat(target->bt_mount, bp); 753 xfs_buf_iowait(bp); 754 return bp; 755 } 756 757 /* 758 * Return a buffer allocated as an empty buffer and associated to external 759 * memory via xfs_buf_associate_memory() back to it's empty state. 760 */ 761 void 762 xfs_buf_set_empty( 763 struct xfs_buf *bp, 764 size_t numblks) 765 { 766 if (bp->b_pages) 767 _xfs_buf_free_pages(bp); 768 769 bp->b_pages = NULL; 770 bp->b_page_count = 0; 771 bp->b_addr = NULL; 772 bp->b_length = numblks; 773 bp->b_io_length = numblks; 774 775 ASSERT(bp->b_map_count == 1); 776 bp->b_bn = XFS_BUF_DADDR_NULL; 777 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL; 778 bp->b_maps[0].bm_len = bp->b_length; 779 } 780 781 static inline struct page * 782 mem_to_page( 783 void *addr) 784 { 785 if ((!is_vmalloc_addr(addr))) { 786 return virt_to_page(addr); 787 } else { 788 return vmalloc_to_page(addr); 789 } 790 } 791 792 int 793 xfs_buf_associate_memory( 794 xfs_buf_t *bp, 795 void *mem, 796 size_t len) 797 { 798 int rval; 799 int i = 0; 800 unsigned long pageaddr; 801 unsigned long offset; 802 size_t buflen; 803 int page_count; 804 805 pageaddr = (unsigned long)mem & PAGE_MASK; 806 offset = (unsigned long)mem - pageaddr; 807 buflen = PAGE_ALIGN(len + offset); 808 page_count = buflen >> PAGE_SHIFT; 809 810 /* Free any previous set of page pointers */ 811 if (bp->b_pages) 812 _xfs_buf_free_pages(bp); 813 814 bp->b_pages = NULL; 815 bp->b_addr = mem; 816 817 rval = _xfs_buf_get_pages(bp, page_count, 0); 818 if (rval) 819 return rval; 820 821 bp->b_offset = offset; 822 823 for (i = 0; i < bp->b_page_count; i++) { 824 bp->b_pages[i] = mem_to_page((void *)pageaddr); 825 pageaddr += PAGE_SIZE; 826 } 827 828 bp->b_io_length = BTOBB(len); 829 bp->b_length = BTOBB(buflen); 830 831 return 0; 832 } 833 834 xfs_buf_t * 835 xfs_buf_get_uncached( 836 struct xfs_buftarg *target, 837 size_t numblks, 838 int flags) 839 { 840 unsigned long page_count; 841 int error, i; 842 struct xfs_buf *bp; 843 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks); 844 845 bp = _xfs_buf_alloc(target, &map, 1, 0); 846 if (unlikely(bp == NULL)) 847 goto fail; 848 849 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT; 850 error = _xfs_buf_get_pages(bp, page_count, 0); 851 if (error) 852 goto fail_free_buf; 853 854 for (i = 0; i < page_count; i++) { 855 bp->b_pages[i] = alloc_page(xb_to_gfp(flags)); 856 if (!bp->b_pages[i]) 857 goto fail_free_mem; 858 } 859 bp->b_flags |= _XBF_PAGES; 860 861 error = _xfs_buf_map_pages(bp, 0); 862 if (unlikely(error)) { 863 xfs_warn(target->bt_mount, 864 "%s: failed to map pages\n", __func__); 865 goto fail_free_mem; 866 } 867 868 trace_xfs_buf_get_uncached(bp, _RET_IP_); 869 return bp; 870 871 fail_free_mem: 872 while (--i >= 0) 873 __free_page(bp->b_pages[i]); 874 _xfs_buf_free_pages(bp); 875 fail_free_buf: 876 xfs_buf_free_maps(bp); 877 kmem_zone_free(xfs_buf_zone, bp); 878 fail: 879 return NULL; 880 } 881 882 /* 883 * Increment reference count on buffer, to hold the buffer concurrently 884 * with another thread which may release (free) the buffer asynchronously. 885 * Must hold the buffer already to call this function. 886 */ 887 void 888 xfs_buf_hold( 889 xfs_buf_t *bp) 890 { 891 trace_xfs_buf_hold(bp, _RET_IP_); 892 atomic_inc(&bp->b_hold); 893 } 894 895 /* 896 * Releases a hold on the specified buffer. If the 897 * the hold count is 1, calls xfs_buf_free. 898 */ 899 void 900 xfs_buf_rele( 901 xfs_buf_t *bp) 902 { 903 struct xfs_perag *pag = bp->b_pag; 904 905 trace_xfs_buf_rele(bp, _RET_IP_); 906 907 if (!pag) { 908 ASSERT(list_empty(&bp->b_lru)); 909 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode)); 910 if (atomic_dec_and_test(&bp->b_hold)) 911 xfs_buf_free(bp); 912 return; 913 } 914 915 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode)); 916 917 ASSERT(atomic_read(&bp->b_hold) > 0); 918 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) { 919 if (!(bp->b_flags & XBF_STALE) && 920 atomic_read(&bp->b_lru_ref)) { 921 xfs_buf_lru_add(bp); 922 spin_unlock(&pag->pag_buf_lock); 923 } else { 924 xfs_buf_lru_del(bp); 925 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); 926 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree); 927 spin_unlock(&pag->pag_buf_lock); 928 xfs_perag_put(pag); 929 xfs_buf_free(bp); 930 } 931 } 932 } 933 934 935 /* 936 * Lock a buffer object, if it is not already locked. 937 * 938 * If we come across a stale, pinned, locked buffer, we know that we are 939 * being asked to lock a buffer that has been reallocated. Because it is 940 * pinned, we know that the log has not been pushed to disk and hence it 941 * will still be locked. Rather than continuing to have trylock attempts 942 * fail until someone else pushes the log, push it ourselves before 943 * returning. This means that the xfsaild will not get stuck trying 944 * to push on stale inode buffers. 945 */ 946 int 947 xfs_buf_trylock( 948 struct xfs_buf *bp) 949 { 950 int locked; 951 952 locked = down_trylock(&bp->b_sema) == 0; 953 if (locked) 954 XB_SET_OWNER(bp); 955 956 trace_xfs_buf_trylock(bp, _RET_IP_); 957 return locked; 958 } 959 960 /* 961 * Lock a buffer object. 962 * 963 * If we come across a stale, pinned, locked buffer, we know that we 964 * are being asked to lock a buffer that has been reallocated. Because 965 * it is pinned, we know that the log has not been pushed to disk and 966 * hence it will still be locked. Rather than sleeping until someone 967 * else pushes the log, push it ourselves before trying to get the lock. 968 */ 969 void 970 xfs_buf_lock( 971 struct xfs_buf *bp) 972 { 973 trace_xfs_buf_lock(bp, _RET_IP_); 974 975 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE)) 976 xfs_log_force(bp->b_target->bt_mount, 0); 977 down(&bp->b_sema); 978 XB_SET_OWNER(bp); 979 980 trace_xfs_buf_lock_done(bp, _RET_IP_); 981 } 982 983 void 984 xfs_buf_unlock( 985 struct xfs_buf *bp) 986 { 987 XB_CLEAR_OWNER(bp); 988 up(&bp->b_sema); 989 990 trace_xfs_buf_unlock(bp, _RET_IP_); 991 } 992 993 STATIC void 994 xfs_buf_wait_unpin( 995 xfs_buf_t *bp) 996 { 997 DECLARE_WAITQUEUE (wait, current); 998 999 if (atomic_read(&bp->b_pin_count) == 0) 1000 return; 1001 1002 add_wait_queue(&bp->b_waiters, &wait); 1003 for (;;) { 1004 set_current_state(TASK_UNINTERRUPTIBLE); 1005 if (atomic_read(&bp->b_pin_count) == 0) 1006 break; 1007 io_schedule(); 1008 } 1009 remove_wait_queue(&bp->b_waiters, &wait); 1010 set_current_state(TASK_RUNNING); 1011 } 1012 1013 /* 1014 * Buffer Utility Routines 1015 */ 1016 1017 STATIC void 1018 xfs_buf_iodone_work( 1019 struct work_struct *work) 1020 { 1021 struct xfs_buf *bp = 1022 container_of(work, xfs_buf_t, b_iodone_work); 1023 bool read = !!(bp->b_flags & XBF_READ); 1024 1025 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); 1026 1027 /* only validate buffers that were read without errors */ 1028 if (read && bp->b_ops && !bp->b_error && (bp->b_flags & XBF_DONE)) 1029 bp->b_ops->verify_read(bp); 1030 1031 if (bp->b_iodone) 1032 (*(bp->b_iodone))(bp); 1033 else if (bp->b_flags & XBF_ASYNC) 1034 xfs_buf_relse(bp); 1035 else { 1036 ASSERT(read && bp->b_ops); 1037 complete(&bp->b_iowait); 1038 } 1039 } 1040 1041 void 1042 xfs_buf_ioend( 1043 struct xfs_buf *bp, 1044 int schedule) 1045 { 1046 bool read = !!(bp->b_flags & XBF_READ); 1047 1048 trace_xfs_buf_iodone(bp, _RET_IP_); 1049 1050 if (bp->b_error == 0) 1051 bp->b_flags |= XBF_DONE; 1052 1053 if (bp->b_iodone || (read && bp->b_ops) || (bp->b_flags & XBF_ASYNC)) { 1054 if (schedule) { 1055 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work); 1056 queue_work(xfslogd_workqueue, &bp->b_iodone_work); 1057 } else { 1058 xfs_buf_iodone_work(&bp->b_iodone_work); 1059 } 1060 } else { 1061 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); 1062 complete(&bp->b_iowait); 1063 } 1064 } 1065 1066 void 1067 xfs_buf_ioerror( 1068 xfs_buf_t *bp, 1069 int error) 1070 { 1071 ASSERT(error >= 0 && error <= 0xffff); 1072 bp->b_error = (unsigned short)error; 1073 trace_xfs_buf_ioerror(bp, error, _RET_IP_); 1074 } 1075 1076 void 1077 xfs_buf_ioerror_alert( 1078 struct xfs_buf *bp, 1079 const char *func) 1080 { 1081 xfs_alert(bp->b_target->bt_mount, 1082 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d", 1083 (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length); 1084 } 1085 1086 /* 1087 * Called when we want to stop a buffer from getting written or read. 1088 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend 1089 * so that the proper iodone callbacks get called. 1090 */ 1091 STATIC int 1092 xfs_bioerror( 1093 xfs_buf_t *bp) 1094 { 1095 #ifdef XFSERRORDEBUG 1096 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone); 1097 #endif 1098 1099 /* 1100 * No need to wait until the buffer is unpinned, we aren't flushing it. 1101 */ 1102 xfs_buf_ioerror(bp, EIO); 1103 1104 /* 1105 * We're calling xfs_buf_ioend, so delete XBF_DONE flag. 1106 */ 1107 XFS_BUF_UNREAD(bp); 1108 XFS_BUF_UNDONE(bp); 1109 xfs_buf_stale(bp); 1110 1111 xfs_buf_ioend(bp, 0); 1112 1113 return EIO; 1114 } 1115 1116 /* 1117 * Same as xfs_bioerror, except that we are releasing the buffer 1118 * here ourselves, and avoiding the xfs_buf_ioend call. 1119 * This is meant for userdata errors; metadata bufs come with 1120 * iodone functions attached, so that we can track down errors. 1121 */ 1122 STATIC int 1123 xfs_bioerror_relse( 1124 struct xfs_buf *bp) 1125 { 1126 int64_t fl = bp->b_flags; 1127 /* 1128 * No need to wait until the buffer is unpinned. 1129 * We aren't flushing it. 1130 * 1131 * chunkhold expects B_DONE to be set, whether 1132 * we actually finish the I/O or not. We don't want to 1133 * change that interface. 1134 */ 1135 XFS_BUF_UNREAD(bp); 1136 XFS_BUF_DONE(bp); 1137 xfs_buf_stale(bp); 1138 bp->b_iodone = NULL; 1139 if (!(fl & XBF_ASYNC)) { 1140 /* 1141 * Mark b_error and B_ERROR _both_. 1142 * Lot's of chunkcache code assumes that. 1143 * There's no reason to mark error for 1144 * ASYNC buffers. 1145 */ 1146 xfs_buf_ioerror(bp, EIO); 1147 complete(&bp->b_iowait); 1148 } else { 1149 xfs_buf_relse(bp); 1150 } 1151 1152 return EIO; 1153 } 1154 1155 STATIC int 1156 xfs_bdstrat_cb( 1157 struct xfs_buf *bp) 1158 { 1159 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) { 1160 trace_xfs_bdstrat_shut(bp, _RET_IP_); 1161 /* 1162 * Metadata write that didn't get logged but 1163 * written delayed anyway. These aren't associated 1164 * with a transaction, and can be ignored. 1165 */ 1166 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp)) 1167 return xfs_bioerror_relse(bp); 1168 else 1169 return xfs_bioerror(bp); 1170 } 1171 1172 xfs_buf_iorequest(bp); 1173 return 0; 1174 } 1175 1176 int 1177 xfs_bwrite( 1178 struct xfs_buf *bp) 1179 { 1180 int error; 1181 1182 ASSERT(xfs_buf_islocked(bp)); 1183 1184 bp->b_flags |= XBF_WRITE; 1185 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q); 1186 1187 xfs_bdstrat_cb(bp); 1188 1189 error = xfs_buf_iowait(bp); 1190 if (error) { 1191 xfs_force_shutdown(bp->b_target->bt_mount, 1192 SHUTDOWN_META_IO_ERROR); 1193 } 1194 return error; 1195 } 1196 1197 /* 1198 * Wrapper around bdstrat so that we can stop data from going to disk in case 1199 * we are shutting down the filesystem. Typically user data goes thru this 1200 * path; one of the exceptions is the superblock. 1201 */ 1202 void 1203 xfsbdstrat( 1204 struct xfs_mount *mp, 1205 struct xfs_buf *bp) 1206 { 1207 if (XFS_FORCED_SHUTDOWN(mp)) { 1208 trace_xfs_bdstrat_shut(bp, _RET_IP_); 1209 xfs_bioerror_relse(bp); 1210 return; 1211 } 1212 1213 xfs_buf_iorequest(bp); 1214 } 1215 1216 STATIC void 1217 _xfs_buf_ioend( 1218 xfs_buf_t *bp, 1219 int schedule) 1220 { 1221 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) 1222 xfs_buf_ioend(bp, schedule); 1223 } 1224 1225 STATIC void 1226 xfs_buf_bio_end_io( 1227 struct bio *bio, 1228 int error) 1229 { 1230 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private; 1231 1232 /* 1233 * don't overwrite existing errors - otherwise we can lose errors on 1234 * buffers that require multiple bios to complete. 1235 */ 1236 if (!bp->b_error) 1237 xfs_buf_ioerror(bp, -error); 1238 1239 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ)) 1240 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp)); 1241 1242 _xfs_buf_ioend(bp, 1); 1243 bio_put(bio); 1244 } 1245 1246 static void 1247 xfs_buf_ioapply_map( 1248 struct xfs_buf *bp, 1249 int map, 1250 int *buf_offset, 1251 int *count, 1252 int rw) 1253 { 1254 int page_index; 1255 int total_nr_pages = bp->b_page_count; 1256 int nr_pages; 1257 struct bio *bio; 1258 sector_t sector = bp->b_maps[map].bm_bn; 1259 int size; 1260 int offset; 1261 1262 total_nr_pages = bp->b_page_count; 1263 1264 /* skip the pages in the buffer before the start offset */ 1265 page_index = 0; 1266 offset = *buf_offset; 1267 while (offset >= PAGE_SIZE) { 1268 page_index++; 1269 offset -= PAGE_SIZE; 1270 } 1271 1272 /* 1273 * Limit the IO size to the length of the current vector, and update the 1274 * remaining IO count for the next time around. 1275 */ 1276 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count); 1277 *count -= size; 1278 *buf_offset += size; 1279 1280 next_chunk: 1281 atomic_inc(&bp->b_io_remaining); 1282 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT); 1283 if (nr_pages > total_nr_pages) 1284 nr_pages = total_nr_pages; 1285 1286 bio = bio_alloc(GFP_NOIO, nr_pages); 1287 bio->bi_bdev = bp->b_target->bt_bdev; 1288 bio->bi_sector = sector; 1289 bio->bi_end_io = xfs_buf_bio_end_io; 1290 bio->bi_private = bp; 1291 1292 1293 for (; size && nr_pages; nr_pages--, page_index++) { 1294 int rbytes, nbytes = PAGE_SIZE - offset; 1295 1296 if (nbytes > size) 1297 nbytes = size; 1298 1299 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes, 1300 offset); 1301 if (rbytes < nbytes) 1302 break; 1303 1304 offset = 0; 1305 sector += BTOBB(nbytes); 1306 size -= nbytes; 1307 total_nr_pages--; 1308 } 1309 1310 if (likely(bio->bi_size)) { 1311 if (xfs_buf_is_vmapped(bp)) { 1312 flush_kernel_vmap_range(bp->b_addr, 1313 xfs_buf_vmap_len(bp)); 1314 } 1315 submit_bio(rw, bio); 1316 if (size) 1317 goto next_chunk; 1318 } else { 1319 /* 1320 * This is guaranteed not to be the last io reference count 1321 * because the caller (xfs_buf_iorequest) holds a count itself. 1322 */ 1323 atomic_dec(&bp->b_io_remaining); 1324 xfs_buf_ioerror(bp, EIO); 1325 bio_put(bio); 1326 } 1327 1328 } 1329 1330 STATIC void 1331 _xfs_buf_ioapply( 1332 struct xfs_buf *bp) 1333 { 1334 struct blk_plug plug; 1335 int rw; 1336 int offset; 1337 int size; 1338 int i; 1339 1340 /* 1341 * Make sure we capture only current IO errors rather than stale errors 1342 * left over from previous use of the buffer (e.g. failed readahead). 1343 */ 1344 bp->b_error = 0; 1345 1346 if (bp->b_flags & XBF_WRITE) { 1347 if (bp->b_flags & XBF_SYNCIO) 1348 rw = WRITE_SYNC; 1349 else 1350 rw = WRITE; 1351 if (bp->b_flags & XBF_FUA) 1352 rw |= REQ_FUA; 1353 if (bp->b_flags & XBF_FLUSH) 1354 rw |= REQ_FLUSH; 1355 1356 /* 1357 * Run the write verifier callback function if it exists. If 1358 * this function fails it will mark the buffer with an error and 1359 * the IO should not be dispatched. 1360 */ 1361 if (bp->b_ops) { 1362 bp->b_ops->verify_write(bp); 1363 if (bp->b_error) { 1364 xfs_force_shutdown(bp->b_target->bt_mount, 1365 SHUTDOWN_CORRUPT_INCORE); 1366 return; 1367 } 1368 } 1369 } else if (bp->b_flags & XBF_READ_AHEAD) { 1370 rw = READA; 1371 } else { 1372 rw = READ; 1373 } 1374 1375 /* we only use the buffer cache for meta-data */ 1376 rw |= REQ_META; 1377 1378 /* 1379 * Walk all the vectors issuing IO on them. Set up the initial offset 1380 * into the buffer and the desired IO size before we start - 1381 * _xfs_buf_ioapply_vec() will modify them appropriately for each 1382 * subsequent call. 1383 */ 1384 offset = bp->b_offset; 1385 size = BBTOB(bp->b_io_length); 1386 blk_start_plug(&plug); 1387 for (i = 0; i < bp->b_map_count; i++) { 1388 xfs_buf_ioapply_map(bp, i, &offset, &size, rw); 1389 if (bp->b_error) 1390 break; 1391 if (size <= 0) 1392 break; /* all done */ 1393 } 1394 blk_finish_plug(&plug); 1395 } 1396 1397 void 1398 xfs_buf_iorequest( 1399 xfs_buf_t *bp) 1400 { 1401 trace_xfs_buf_iorequest(bp, _RET_IP_); 1402 1403 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); 1404 1405 if (bp->b_flags & XBF_WRITE) 1406 xfs_buf_wait_unpin(bp); 1407 xfs_buf_hold(bp); 1408 1409 /* Set the count to 1 initially, this will stop an I/O 1410 * completion callout which happens before we have started 1411 * all the I/O from calling xfs_buf_ioend too early. 1412 */ 1413 atomic_set(&bp->b_io_remaining, 1); 1414 _xfs_buf_ioapply(bp); 1415 _xfs_buf_ioend(bp, 1); 1416 1417 xfs_buf_rele(bp); 1418 } 1419 1420 /* 1421 * Waits for I/O to complete on the buffer supplied. It returns immediately if 1422 * no I/O is pending or there is already a pending error on the buffer. It 1423 * returns the I/O error code, if any, or 0 if there was no error. 1424 */ 1425 int 1426 xfs_buf_iowait( 1427 xfs_buf_t *bp) 1428 { 1429 trace_xfs_buf_iowait(bp, _RET_IP_); 1430 1431 if (!bp->b_error) 1432 wait_for_completion(&bp->b_iowait); 1433 1434 trace_xfs_buf_iowait_done(bp, _RET_IP_); 1435 return bp->b_error; 1436 } 1437 1438 xfs_caddr_t 1439 xfs_buf_offset( 1440 xfs_buf_t *bp, 1441 size_t offset) 1442 { 1443 struct page *page; 1444 1445 if (bp->b_addr) 1446 return bp->b_addr + offset; 1447 1448 offset += bp->b_offset; 1449 page = bp->b_pages[offset >> PAGE_SHIFT]; 1450 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1)); 1451 } 1452 1453 /* 1454 * Move data into or out of a buffer. 1455 */ 1456 void 1457 xfs_buf_iomove( 1458 xfs_buf_t *bp, /* buffer to process */ 1459 size_t boff, /* starting buffer offset */ 1460 size_t bsize, /* length to copy */ 1461 void *data, /* data address */ 1462 xfs_buf_rw_t mode) /* read/write/zero flag */ 1463 { 1464 size_t bend; 1465 1466 bend = boff + bsize; 1467 while (boff < bend) { 1468 struct page *page; 1469 int page_index, page_offset, csize; 1470 1471 page_index = (boff + bp->b_offset) >> PAGE_SHIFT; 1472 page_offset = (boff + bp->b_offset) & ~PAGE_MASK; 1473 page = bp->b_pages[page_index]; 1474 csize = min_t(size_t, PAGE_SIZE - page_offset, 1475 BBTOB(bp->b_io_length) - boff); 1476 1477 ASSERT((csize + page_offset) <= PAGE_SIZE); 1478 1479 switch (mode) { 1480 case XBRW_ZERO: 1481 memset(page_address(page) + page_offset, 0, csize); 1482 break; 1483 case XBRW_READ: 1484 memcpy(data, page_address(page) + page_offset, csize); 1485 break; 1486 case XBRW_WRITE: 1487 memcpy(page_address(page) + page_offset, data, csize); 1488 } 1489 1490 boff += csize; 1491 data += csize; 1492 } 1493 } 1494 1495 /* 1496 * Handling of buffer targets (buftargs). 1497 */ 1498 1499 /* 1500 * Wait for any bufs with callbacks that have been submitted but have not yet 1501 * returned. These buffers will have an elevated hold count, so wait on those 1502 * while freeing all the buffers only held by the LRU. 1503 */ 1504 void 1505 xfs_wait_buftarg( 1506 struct xfs_buftarg *btp) 1507 { 1508 struct xfs_buf *bp; 1509 1510 restart: 1511 spin_lock(&btp->bt_lru_lock); 1512 while (!list_empty(&btp->bt_lru)) { 1513 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru); 1514 if (atomic_read(&bp->b_hold) > 1) { 1515 trace_xfs_buf_wait_buftarg(bp, _RET_IP_); 1516 list_move_tail(&bp->b_lru, &btp->bt_lru); 1517 spin_unlock(&btp->bt_lru_lock); 1518 delay(100); 1519 goto restart; 1520 } 1521 /* 1522 * clear the LRU reference count so the buffer doesn't get 1523 * ignored in xfs_buf_rele(). 1524 */ 1525 atomic_set(&bp->b_lru_ref, 0); 1526 spin_unlock(&btp->bt_lru_lock); 1527 xfs_buf_rele(bp); 1528 spin_lock(&btp->bt_lru_lock); 1529 } 1530 spin_unlock(&btp->bt_lru_lock); 1531 } 1532 1533 int 1534 xfs_buftarg_shrink( 1535 struct shrinker *shrink, 1536 struct shrink_control *sc) 1537 { 1538 struct xfs_buftarg *btp = container_of(shrink, 1539 struct xfs_buftarg, bt_shrinker); 1540 struct xfs_buf *bp; 1541 int nr_to_scan = sc->nr_to_scan; 1542 LIST_HEAD(dispose); 1543 1544 if (!nr_to_scan) 1545 return btp->bt_lru_nr; 1546 1547 spin_lock(&btp->bt_lru_lock); 1548 while (!list_empty(&btp->bt_lru)) { 1549 if (nr_to_scan-- <= 0) 1550 break; 1551 1552 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru); 1553 1554 /* 1555 * Decrement the b_lru_ref count unless the value is already 1556 * zero. If the value is already zero, we need to reclaim the 1557 * buffer, otherwise it gets another trip through the LRU. 1558 */ 1559 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) { 1560 list_move_tail(&bp->b_lru, &btp->bt_lru); 1561 continue; 1562 } 1563 1564 /* 1565 * remove the buffer from the LRU now to avoid needing another 1566 * lock round trip inside xfs_buf_rele(). 1567 */ 1568 list_move(&bp->b_lru, &dispose); 1569 btp->bt_lru_nr--; 1570 bp->b_lru_flags |= _XBF_LRU_DISPOSE; 1571 } 1572 spin_unlock(&btp->bt_lru_lock); 1573 1574 while (!list_empty(&dispose)) { 1575 bp = list_first_entry(&dispose, struct xfs_buf, b_lru); 1576 list_del_init(&bp->b_lru); 1577 xfs_buf_rele(bp); 1578 } 1579 1580 return btp->bt_lru_nr; 1581 } 1582 1583 void 1584 xfs_free_buftarg( 1585 struct xfs_mount *mp, 1586 struct xfs_buftarg *btp) 1587 { 1588 unregister_shrinker(&btp->bt_shrinker); 1589 1590 if (mp->m_flags & XFS_MOUNT_BARRIER) 1591 xfs_blkdev_issue_flush(btp); 1592 1593 kmem_free(btp); 1594 } 1595 1596 STATIC int 1597 xfs_setsize_buftarg_flags( 1598 xfs_buftarg_t *btp, 1599 unsigned int blocksize, 1600 unsigned int sectorsize, 1601 int verbose) 1602 { 1603 btp->bt_bsize = blocksize; 1604 btp->bt_sshift = ffs(sectorsize) - 1; 1605 btp->bt_smask = sectorsize - 1; 1606 1607 if (set_blocksize(btp->bt_bdev, sectorsize)) { 1608 char name[BDEVNAME_SIZE]; 1609 1610 bdevname(btp->bt_bdev, name); 1611 1612 xfs_warn(btp->bt_mount, 1613 "Cannot set_blocksize to %u on device %s\n", 1614 sectorsize, name); 1615 return EINVAL; 1616 } 1617 1618 return 0; 1619 } 1620 1621 /* 1622 * When allocating the initial buffer target we have not yet 1623 * read in the superblock, so don't know what sized sectors 1624 * are being used is at this early stage. Play safe. 1625 */ 1626 STATIC int 1627 xfs_setsize_buftarg_early( 1628 xfs_buftarg_t *btp, 1629 struct block_device *bdev) 1630 { 1631 return xfs_setsize_buftarg_flags(btp, 1632 PAGE_SIZE, bdev_logical_block_size(bdev), 0); 1633 } 1634 1635 int 1636 xfs_setsize_buftarg( 1637 xfs_buftarg_t *btp, 1638 unsigned int blocksize, 1639 unsigned int sectorsize) 1640 { 1641 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1); 1642 } 1643 1644 xfs_buftarg_t * 1645 xfs_alloc_buftarg( 1646 struct xfs_mount *mp, 1647 struct block_device *bdev, 1648 int external, 1649 const char *fsname) 1650 { 1651 xfs_buftarg_t *btp; 1652 1653 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS); 1654 1655 btp->bt_mount = mp; 1656 btp->bt_dev = bdev->bd_dev; 1657 btp->bt_bdev = bdev; 1658 btp->bt_bdi = blk_get_backing_dev_info(bdev); 1659 if (!btp->bt_bdi) 1660 goto error; 1661 1662 INIT_LIST_HEAD(&btp->bt_lru); 1663 spin_lock_init(&btp->bt_lru_lock); 1664 if (xfs_setsize_buftarg_early(btp, bdev)) 1665 goto error; 1666 btp->bt_shrinker.shrink = xfs_buftarg_shrink; 1667 btp->bt_shrinker.seeks = DEFAULT_SEEKS; 1668 register_shrinker(&btp->bt_shrinker); 1669 return btp; 1670 1671 error: 1672 kmem_free(btp); 1673 return NULL; 1674 } 1675 1676 /* 1677 * Add a buffer to the delayed write list. 1678 * 1679 * This queues a buffer for writeout if it hasn't already been. Note that 1680 * neither this routine nor the buffer list submission functions perform 1681 * any internal synchronization. It is expected that the lists are thread-local 1682 * to the callers. 1683 * 1684 * Returns true if we queued up the buffer, or false if it already had 1685 * been on the buffer list. 1686 */ 1687 bool 1688 xfs_buf_delwri_queue( 1689 struct xfs_buf *bp, 1690 struct list_head *list) 1691 { 1692 ASSERT(xfs_buf_islocked(bp)); 1693 ASSERT(!(bp->b_flags & XBF_READ)); 1694 1695 /* 1696 * If the buffer is already marked delwri it already is queued up 1697 * by someone else for imediate writeout. Just ignore it in that 1698 * case. 1699 */ 1700 if (bp->b_flags & _XBF_DELWRI_Q) { 1701 trace_xfs_buf_delwri_queued(bp, _RET_IP_); 1702 return false; 1703 } 1704 1705 trace_xfs_buf_delwri_queue(bp, _RET_IP_); 1706 1707 /* 1708 * If a buffer gets written out synchronously or marked stale while it 1709 * is on a delwri list we lazily remove it. To do this, the other party 1710 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone. 1711 * It remains referenced and on the list. In a rare corner case it 1712 * might get readded to a delwri list after the synchronous writeout, in 1713 * which case we need just need to re-add the flag here. 1714 */ 1715 bp->b_flags |= _XBF_DELWRI_Q; 1716 if (list_empty(&bp->b_list)) { 1717 atomic_inc(&bp->b_hold); 1718 list_add_tail(&bp->b_list, list); 1719 } 1720 1721 return true; 1722 } 1723 1724 /* 1725 * Compare function is more complex than it needs to be because 1726 * the return value is only 32 bits and we are doing comparisons 1727 * on 64 bit values 1728 */ 1729 static int 1730 xfs_buf_cmp( 1731 void *priv, 1732 struct list_head *a, 1733 struct list_head *b) 1734 { 1735 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list); 1736 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list); 1737 xfs_daddr_t diff; 1738 1739 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn; 1740 if (diff < 0) 1741 return -1; 1742 if (diff > 0) 1743 return 1; 1744 return 0; 1745 } 1746 1747 static int 1748 __xfs_buf_delwri_submit( 1749 struct list_head *buffer_list, 1750 struct list_head *io_list, 1751 bool wait) 1752 { 1753 struct blk_plug plug; 1754 struct xfs_buf *bp, *n; 1755 int pinned = 0; 1756 1757 list_for_each_entry_safe(bp, n, buffer_list, b_list) { 1758 if (!wait) { 1759 if (xfs_buf_ispinned(bp)) { 1760 pinned++; 1761 continue; 1762 } 1763 if (!xfs_buf_trylock(bp)) 1764 continue; 1765 } else { 1766 xfs_buf_lock(bp); 1767 } 1768 1769 /* 1770 * Someone else might have written the buffer synchronously or 1771 * marked it stale in the meantime. In that case only the 1772 * _XBF_DELWRI_Q flag got cleared, and we have to drop the 1773 * reference and remove it from the list here. 1774 */ 1775 if (!(bp->b_flags & _XBF_DELWRI_Q)) { 1776 list_del_init(&bp->b_list); 1777 xfs_buf_relse(bp); 1778 continue; 1779 } 1780 1781 list_move_tail(&bp->b_list, io_list); 1782 trace_xfs_buf_delwri_split(bp, _RET_IP_); 1783 } 1784 1785 list_sort(NULL, io_list, xfs_buf_cmp); 1786 1787 blk_start_plug(&plug); 1788 list_for_each_entry_safe(bp, n, io_list, b_list) { 1789 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC); 1790 bp->b_flags |= XBF_WRITE; 1791 1792 if (!wait) { 1793 bp->b_flags |= XBF_ASYNC; 1794 list_del_init(&bp->b_list); 1795 } 1796 xfs_bdstrat_cb(bp); 1797 } 1798 blk_finish_plug(&plug); 1799 1800 return pinned; 1801 } 1802 1803 /* 1804 * Write out a buffer list asynchronously. 1805 * 1806 * This will take the @buffer_list, write all non-locked and non-pinned buffers 1807 * out and not wait for I/O completion on any of the buffers. This interface 1808 * is only safely useable for callers that can track I/O completion by higher 1809 * level means, e.g. AIL pushing as the @buffer_list is consumed in this 1810 * function. 1811 */ 1812 int 1813 xfs_buf_delwri_submit_nowait( 1814 struct list_head *buffer_list) 1815 { 1816 LIST_HEAD (io_list); 1817 return __xfs_buf_delwri_submit(buffer_list, &io_list, false); 1818 } 1819 1820 /* 1821 * Write out a buffer list synchronously. 1822 * 1823 * This will take the @buffer_list, write all buffers out and wait for I/O 1824 * completion on all of the buffers. @buffer_list is consumed by the function, 1825 * so callers must have some other way of tracking buffers if they require such 1826 * functionality. 1827 */ 1828 int 1829 xfs_buf_delwri_submit( 1830 struct list_head *buffer_list) 1831 { 1832 LIST_HEAD (io_list); 1833 int error = 0, error2; 1834 struct xfs_buf *bp; 1835 1836 __xfs_buf_delwri_submit(buffer_list, &io_list, true); 1837 1838 /* Wait for IO to complete. */ 1839 while (!list_empty(&io_list)) { 1840 bp = list_first_entry(&io_list, struct xfs_buf, b_list); 1841 1842 list_del_init(&bp->b_list); 1843 error2 = xfs_buf_iowait(bp); 1844 xfs_buf_relse(bp); 1845 if (!error) 1846 error = error2; 1847 } 1848 1849 return error; 1850 } 1851 1852 int __init 1853 xfs_buf_init(void) 1854 { 1855 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf", 1856 KM_ZONE_HWALIGN, NULL); 1857 if (!xfs_buf_zone) 1858 goto out; 1859 1860 xfslogd_workqueue = alloc_workqueue("xfslogd", 1861 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1); 1862 if (!xfslogd_workqueue) 1863 goto out_free_buf_zone; 1864 1865 return 0; 1866 1867 out_free_buf_zone: 1868 kmem_zone_destroy(xfs_buf_zone); 1869 out: 1870 return -ENOMEM; 1871 } 1872 1873 void 1874 xfs_buf_terminate(void) 1875 { 1876 destroy_workqueue(xfslogd_workqueue); 1877 kmem_zone_destroy(xfs_buf_zone); 1878 } 1879