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 return NULL; 517 } 518 519 /* get tree root */ 520 pag = xfs_perag_get(btp->bt_mount, 521 xfs_daddr_to_agno(btp->bt_mount, blkno)); 522 523 /* walk tree */ 524 spin_lock(&pag->pag_buf_lock); 525 rbp = &pag->pag_buf_tree.rb_node; 526 parent = NULL; 527 bp = NULL; 528 while (*rbp) { 529 parent = *rbp; 530 bp = rb_entry(parent, struct xfs_buf, b_rbnode); 531 532 if (blkno < bp->b_bn) 533 rbp = &(*rbp)->rb_left; 534 else if (blkno > bp->b_bn) 535 rbp = &(*rbp)->rb_right; 536 else { 537 /* 538 * found a block number match. If the range doesn't 539 * match, the only way this is allowed is if the buffer 540 * in the cache is stale and the transaction that made 541 * it stale has not yet committed. i.e. we are 542 * reallocating a busy extent. Skip this buffer and 543 * continue searching to the right for an exact match. 544 */ 545 if (bp->b_length != numblks) { 546 ASSERT(bp->b_flags & XBF_STALE); 547 rbp = &(*rbp)->rb_right; 548 continue; 549 } 550 atomic_inc(&bp->b_hold); 551 goto found; 552 } 553 } 554 555 /* No match found */ 556 if (new_bp) { 557 rb_link_node(&new_bp->b_rbnode, parent, rbp); 558 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree); 559 /* the buffer keeps the perag reference until it is freed */ 560 new_bp->b_pag = pag; 561 spin_unlock(&pag->pag_buf_lock); 562 } else { 563 XFS_STATS_INC(xb_miss_locked); 564 spin_unlock(&pag->pag_buf_lock); 565 xfs_perag_put(pag); 566 } 567 return new_bp; 568 569 found: 570 spin_unlock(&pag->pag_buf_lock); 571 xfs_perag_put(pag); 572 573 if (!xfs_buf_trylock(bp)) { 574 if (flags & XBF_TRYLOCK) { 575 xfs_buf_rele(bp); 576 XFS_STATS_INC(xb_busy_locked); 577 return NULL; 578 } 579 xfs_buf_lock(bp); 580 XFS_STATS_INC(xb_get_locked_waited); 581 } 582 583 /* 584 * if the buffer is stale, clear all the external state associated with 585 * it. We need to keep flags such as how we allocated the buffer memory 586 * intact here. 587 */ 588 if (bp->b_flags & XBF_STALE) { 589 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0); 590 ASSERT(bp->b_iodone == NULL); 591 bp->b_flags &= _XBF_KMEM | _XBF_PAGES; 592 bp->b_ops = NULL; 593 } 594 595 trace_xfs_buf_find(bp, flags, _RET_IP_); 596 XFS_STATS_INC(xb_get_locked); 597 return bp; 598 } 599 600 /* 601 * Assembles a buffer covering the specified range. The code is optimised for 602 * cache hits, as metadata intensive workloads will see 3 orders of magnitude 603 * more hits than misses. 604 */ 605 struct xfs_buf * 606 xfs_buf_get_map( 607 struct xfs_buftarg *target, 608 struct xfs_buf_map *map, 609 int nmaps, 610 xfs_buf_flags_t flags) 611 { 612 struct xfs_buf *bp; 613 struct xfs_buf *new_bp; 614 int error = 0; 615 616 bp = _xfs_buf_find(target, map, nmaps, flags, NULL); 617 if (likely(bp)) 618 goto found; 619 620 new_bp = _xfs_buf_alloc(target, map, nmaps, flags); 621 if (unlikely(!new_bp)) 622 return NULL; 623 624 error = xfs_buf_allocate_memory(new_bp, flags); 625 if (error) { 626 xfs_buf_free(new_bp); 627 return NULL; 628 } 629 630 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp); 631 if (!bp) { 632 xfs_buf_free(new_bp); 633 return NULL; 634 } 635 636 if (bp != new_bp) 637 xfs_buf_free(new_bp); 638 639 found: 640 if (!bp->b_addr) { 641 error = _xfs_buf_map_pages(bp, flags); 642 if (unlikely(error)) { 643 xfs_warn(target->bt_mount, 644 "%s: failed to map pages\n", __func__); 645 xfs_buf_relse(bp); 646 return NULL; 647 } 648 } 649 650 XFS_STATS_INC(xb_get); 651 trace_xfs_buf_get(bp, flags, _RET_IP_); 652 return bp; 653 } 654 655 STATIC int 656 _xfs_buf_read( 657 xfs_buf_t *bp, 658 xfs_buf_flags_t flags) 659 { 660 ASSERT(!(flags & XBF_WRITE)); 661 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL); 662 663 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD); 664 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD); 665 666 xfs_buf_iorequest(bp); 667 if (flags & XBF_ASYNC) 668 return 0; 669 return xfs_buf_iowait(bp); 670 } 671 672 xfs_buf_t * 673 xfs_buf_read_map( 674 struct xfs_buftarg *target, 675 struct xfs_buf_map *map, 676 int nmaps, 677 xfs_buf_flags_t flags, 678 const struct xfs_buf_ops *ops) 679 { 680 struct xfs_buf *bp; 681 682 flags |= XBF_READ; 683 684 bp = xfs_buf_get_map(target, map, nmaps, flags); 685 if (bp) { 686 trace_xfs_buf_read(bp, flags, _RET_IP_); 687 688 if (!XFS_BUF_ISDONE(bp)) { 689 XFS_STATS_INC(xb_get_read); 690 bp->b_ops = ops; 691 _xfs_buf_read(bp, flags); 692 } else if (flags & XBF_ASYNC) { 693 /* 694 * Read ahead call which is already satisfied, 695 * drop the buffer 696 */ 697 xfs_buf_relse(bp); 698 return NULL; 699 } else { 700 /* We do not want read in the flags */ 701 bp->b_flags &= ~XBF_READ; 702 } 703 } 704 705 return bp; 706 } 707 708 /* 709 * If we are not low on memory then do the readahead in a deadlock 710 * safe manner. 711 */ 712 void 713 xfs_buf_readahead_map( 714 struct xfs_buftarg *target, 715 struct xfs_buf_map *map, 716 int nmaps, 717 const struct xfs_buf_ops *ops) 718 { 719 if (bdi_read_congested(target->bt_bdi)) 720 return; 721 722 xfs_buf_read_map(target, map, nmaps, 723 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops); 724 } 725 726 /* 727 * Read an uncached buffer from disk. Allocates and returns a locked 728 * buffer containing the disk contents or nothing. 729 */ 730 struct xfs_buf * 731 xfs_buf_read_uncached( 732 struct xfs_buftarg *target, 733 xfs_daddr_t daddr, 734 size_t numblks, 735 int flags, 736 const struct xfs_buf_ops *ops) 737 { 738 struct xfs_buf *bp; 739 740 bp = xfs_buf_get_uncached(target, numblks, flags); 741 if (!bp) 742 return NULL; 743 744 /* set up the buffer for a read IO */ 745 ASSERT(bp->b_map_count == 1); 746 bp->b_bn = daddr; 747 bp->b_maps[0].bm_bn = daddr; 748 bp->b_flags |= XBF_READ; 749 bp->b_ops = ops; 750 751 xfsbdstrat(target->bt_mount, bp); 752 xfs_buf_iowait(bp); 753 return bp; 754 } 755 756 /* 757 * Return a buffer allocated as an empty buffer and associated to external 758 * memory via xfs_buf_associate_memory() back to it's empty state. 759 */ 760 void 761 xfs_buf_set_empty( 762 struct xfs_buf *bp, 763 size_t numblks) 764 { 765 if (bp->b_pages) 766 _xfs_buf_free_pages(bp); 767 768 bp->b_pages = NULL; 769 bp->b_page_count = 0; 770 bp->b_addr = NULL; 771 bp->b_length = numblks; 772 bp->b_io_length = numblks; 773 774 ASSERT(bp->b_map_count == 1); 775 bp->b_bn = XFS_BUF_DADDR_NULL; 776 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL; 777 bp->b_maps[0].bm_len = bp->b_length; 778 } 779 780 static inline struct page * 781 mem_to_page( 782 void *addr) 783 { 784 if ((!is_vmalloc_addr(addr))) { 785 return virt_to_page(addr); 786 } else { 787 return vmalloc_to_page(addr); 788 } 789 } 790 791 int 792 xfs_buf_associate_memory( 793 xfs_buf_t *bp, 794 void *mem, 795 size_t len) 796 { 797 int rval; 798 int i = 0; 799 unsigned long pageaddr; 800 unsigned long offset; 801 size_t buflen; 802 int page_count; 803 804 pageaddr = (unsigned long)mem & PAGE_MASK; 805 offset = (unsigned long)mem - pageaddr; 806 buflen = PAGE_ALIGN(len + offset); 807 page_count = buflen >> PAGE_SHIFT; 808 809 /* Free any previous set of page pointers */ 810 if (bp->b_pages) 811 _xfs_buf_free_pages(bp); 812 813 bp->b_pages = NULL; 814 bp->b_addr = mem; 815 816 rval = _xfs_buf_get_pages(bp, page_count, 0); 817 if (rval) 818 return rval; 819 820 bp->b_offset = offset; 821 822 for (i = 0; i < bp->b_page_count; i++) { 823 bp->b_pages[i] = mem_to_page((void *)pageaddr); 824 pageaddr += PAGE_SIZE; 825 } 826 827 bp->b_io_length = BTOBB(len); 828 bp->b_length = BTOBB(buflen); 829 830 return 0; 831 } 832 833 xfs_buf_t * 834 xfs_buf_get_uncached( 835 struct xfs_buftarg *target, 836 size_t numblks, 837 int flags) 838 { 839 unsigned long page_count; 840 int error, i; 841 struct xfs_buf *bp; 842 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks); 843 844 bp = _xfs_buf_alloc(target, &map, 1, 0); 845 if (unlikely(bp == NULL)) 846 goto fail; 847 848 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT; 849 error = _xfs_buf_get_pages(bp, page_count, 0); 850 if (error) 851 goto fail_free_buf; 852 853 for (i = 0; i < page_count; i++) { 854 bp->b_pages[i] = alloc_page(xb_to_gfp(flags)); 855 if (!bp->b_pages[i]) 856 goto fail_free_mem; 857 } 858 bp->b_flags |= _XBF_PAGES; 859 860 error = _xfs_buf_map_pages(bp, 0); 861 if (unlikely(error)) { 862 xfs_warn(target->bt_mount, 863 "%s: failed to map pages\n", __func__); 864 goto fail_free_mem; 865 } 866 867 trace_xfs_buf_get_uncached(bp, _RET_IP_); 868 return bp; 869 870 fail_free_mem: 871 while (--i >= 0) 872 __free_page(bp->b_pages[i]); 873 _xfs_buf_free_pages(bp); 874 fail_free_buf: 875 xfs_buf_free_maps(bp); 876 kmem_zone_free(xfs_buf_zone, bp); 877 fail: 878 return NULL; 879 } 880 881 /* 882 * Increment reference count on buffer, to hold the buffer concurrently 883 * with another thread which may release (free) the buffer asynchronously. 884 * Must hold the buffer already to call this function. 885 */ 886 void 887 xfs_buf_hold( 888 xfs_buf_t *bp) 889 { 890 trace_xfs_buf_hold(bp, _RET_IP_); 891 atomic_inc(&bp->b_hold); 892 } 893 894 /* 895 * Releases a hold on the specified buffer. If the 896 * the hold count is 1, calls xfs_buf_free. 897 */ 898 void 899 xfs_buf_rele( 900 xfs_buf_t *bp) 901 { 902 struct xfs_perag *pag = bp->b_pag; 903 904 trace_xfs_buf_rele(bp, _RET_IP_); 905 906 if (!pag) { 907 ASSERT(list_empty(&bp->b_lru)); 908 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode)); 909 if (atomic_dec_and_test(&bp->b_hold)) 910 xfs_buf_free(bp); 911 return; 912 } 913 914 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode)); 915 916 ASSERT(atomic_read(&bp->b_hold) > 0); 917 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) { 918 if (!(bp->b_flags & XBF_STALE) && 919 atomic_read(&bp->b_lru_ref)) { 920 xfs_buf_lru_add(bp); 921 spin_unlock(&pag->pag_buf_lock); 922 } else { 923 xfs_buf_lru_del(bp); 924 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); 925 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree); 926 spin_unlock(&pag->pag_buf_lock); 927 xfs_perag_put(pag); 928 xfs_buf_free(bp); 929 } 930 } 931 } 932 933 934 /* 935 * Lock a buffer object, if it is not already locked. 936 * 937 * If we come across a stale, pinned, locked buffer, we know that we are 938 * being asked to lock a buffer that has been reallocated. Because it is 939 * pinned, we know that the log has not been pushed to disk and hence it 940 * will still be locked. Rather than continuing to have trylock attempts 941 * fail until someone else pushes the log, push it ourselves before 942 * returning. This means that the xfsaild will not get stuck trying 943 * to push on stale inode buffers. 944 */ 945 int 946 xfs_buf_trylock( 947 struct xfs_buf *bp) 948 { 949 int locked; 950 951 locked = down_trylock(&bp->b_sema) == 0; 952 if (locked) 953 XB_SET_OWNER(bp); 954 955 trace_xfs_buf_trylock(bp, _RET_IP_); 956 return locked; 957 } 958 959 /* 960 * Lock a buffer object. 961 * 962 * If we come across a stale, pinned, locked buffer, we know that we 963 * are being asked to lock a buffer that has been reallocated. Because 964 * it is pinned, we know that the log has not been pushed to disk and 965 * hence it will still be locked. Rather than sleeping until someone 966 * else pushes the log, push it ourselves before trying to get the lock. 967 */ 968 void 969 xfs_buf_lock( 970 struct xfs_buf *bp) 971 { 972 trace_xfs_buf_lock(bp, _RET_IP_); 973 974 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE)) 975 xfs_log_force(bp->b_target->bt_mount, 0); 976 down(&bp->b_sema); 977 XB_SET_OWNER(bp); 978 979 trace_xfs_buf_lock_done(bp, _RET_IP_); 980 } 981 982 void 983 xfs_buf_unlock( 984 struct xfs_buf *bp) 985 { 986 XB_CLEAR_OWNER(bp); 987 up(&bp->b_sema); 988 989 trace_xfs_buf_unlock(bp, _RET_IP_); 990 } 991 992 STATIC void 993 xfs_buf_wait_unpin( 994 xfs_buf_t *bp) 995 { 996 DECLARE_WAITQUEUE (wait, current); 997 998 if (atomic_read(&bp->b_pin_count) == 0) 999 return; 1000 1001 add_wait_queue(&bp->b_waiters, &wait); 1002 for (;;) { 1003 set_current_state(TASK_UNINTERRUPTIBLE); 1004 if (atomic_read(&bp->b_pin_count) == 0) 1005 break; 1006 io_schedule(); 1007 } 1008 remove_wait_queue(&bp->b_waiters, &wait); 1009 set_current_state(TASK_RUNNING); 1010 } 1011 1012 /* 1013 * Buffer Utility Routines 1014 */ 1015 1016 STATIC void 1017 xfs_buf_iodone_work( 1018 struct work_struct *work) 1019 { 1020 struct xfs_buf *bp = 1021 container_of(work, xfs_buf_t, b_iodone_work); 1022 bool read = !!(bp->b_flags & XBF_READ); 1023 1024 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); 1025 if (read && bp->b_ops) 1026 bp->b_ops->verify_read(bp); 1027 1028 if (bp->b_iodone) 1029 (*(bp->b_iodone))(bp); 1030 else if (bp->b_flags & XBF_ASYNC) 1031 xfs_buf_relse(bp); 1032 else { 1033 ASSERT(read && bp->b_ops); 1034 complete(&bp->b_iowait); 1035 } 1036 } 1037 1038 void 1039 xfs_buf_ioend( 1040 struct xfs_buf *bp, 1041 int schedule) 1042 { 1043 bool read = !!(bp->b_flags & XBF_READ); 1044 1045 trace_xfs_buf_iodone(bp, _RET_IP_); 1046 1047 if (bp->b_error == 0) 1048 bp->b_flags |= XBF_DONE; 1049 1050 if (bp->b_iodone || (read && bp->b_ops) || (bp->b_flags & XBF_ASYNC)) { 1051 if (schedule) { 1052 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work); 1053 queue_work(xfslogd_workqueue, &bp->b_iodone_work); 1054 } else { 1055 xfs_buf_iodone_work(&bp->b_iodone_work); 1056 } 1057 } else { 1058 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); 1059 complete(&bp->b_iowait); 1060 } 1061 } 1062 1063 void 1064 xfs_buf_ioerror( 1065 xfs_buf_t *bp, 1066 int error) 1067 { 1068 ASSERT(error >= 0 && error <= 0xffff); 1069 bp->b_error = (unsigned short)error; 1070 trace_xfs_buf_ioerror(bp, error, _RET_IP_); 1071 } 1072 1073 void 1074 xfs_buf_ioerror_alert( 1075 struct xfs_buf *bp, 1076 const char *func) 1077 { 1078 xfs_alert(bp->b_target->bt_mount, 1079 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d", 1080 (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length); 1081 } 1082 1083 /* 1084 * Called when we want to stop a buffer from getting written or read. 1085 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend 1086 * so that the proper iodone callbacks get called. 1087 */ 1088 STATIC int 1089 xfs_bioerror( 1090 xfs_buf_t *bp) 1091 { 1092 #ifdef XFSERRORDEBUG 1093 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone); 1094 #endif 1095 1096 /* 1097 * No need to wait until the buffer is unpinned, we aren't flushing it. 1098 */ 1099 xfs_buf_ioerror(bp, EIO); 1100 1101 /* 1102 * We're calling xfs_buf_ioend, so delete XBF_DONE flag. 1103 */ 1104 XFS_BUF_UNREAD(bp); 1105 XFS_BUF_UNDONE(bp); 1106 xfs_buf_stale(bp); 1107 1108 xfs_buf_ioend(bp, 0); 1109 1110 return EIO; 1111 } 1112 1113 /* 1114 * Same as xfs_bioerror, except that we are releasing the buffer 1115 * here ourselves, and avoiding the xfs_buf_ioend call. 1116 * This is meant for userdata errors; metadata bufs come with 1117 * iodone functions attached, so that we can track down errors. 1118 */ 1119 STATIC int 1120 xfs_bioerror_relse( 1121 struct xfs_buf *bp) 1122 { 1123 int64_t fl = bp->b_flags; 1124 /* 1125 * No need to wait until the buffer is unpinned. 1126 * We aren't flushing it. 1127 * 1128 * chunkhold expects B_DONE to be set, whether 1129 * we actually finish the I/O or not. We don't want to 1130 * change that interface. 1131 */ 1132 XFS_BUF_UNREAD(bp); 1133 XFS_BUF_DONE(bp); 1134 xfs_buf_stale(bp); 1135 bp->b_iodone = NULL; 1136 if (!(fl & XBF_ASYNC)) { 1137 /* 1138 * Mark b_error and B_ERROR _both_. 1139 * Lot's of chunkcache code assumes that. 1140 * There's no reason to mark error for 1141 * ASYNC buffers. 1142 */ 1143 xfs_buf_ioerror(bp, EIO); 1144 complete(&bp->b_iowait); 1145 } else { 1146 xfs_buf_relse(bp); 1147 } 1148 1149 return EIO; 1150 } 1151 1152 STATIC int 1153 xfs_bdstrat_cb( 1154 struct xfs_buf *bp) 1155 { 1156 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) { 1157 trace_xfs_bdstrat_shut(bp, _RET_IP_); 1158 /* 1159 * Metadata write that didn't get logged but 1160 * written delayed anyway. These aren't associated 1161 * with a transaction, and can be ignored. 1162 */ 1163 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp)) 1164 return xfs_bioerror_relse(bp); 1165 else 1166 return xfs_bioerror(bp); 1167 } 1168 1169 xfs_buf_iorequest(bp); 1170 return 0; 1171 } 1172 1173 int 1174 xfs_bwrite( 1175 struct xfs_buf *bp) 1176 { 1177 int error; 1178 1179 ASSERT(xfs_buf_islocked(bp)); 1180 1181 bp->b_flags |= XBF_WRITE; 1182 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q); 1183 1184 xfs_bdstrat_cb(bp); 1185 1186 error = xfs_buf_iowait(bp); 1187 if (error) { 1188 xfs_force_shutdown(bp->b_target->bt_mount, 1189 SHUTDOWN_META_IO_ERROR); 1190 } 1191 return error; 1192 } 1193 1194 /* 1195 * Wrapper around bdstrat so that we can stop data from going to disk in case 1196 * we are shutting down the filesystem. Typically user data goes thru this 1197 * path; one of the exceptions is the superblock. 1198 */ 1199 void 1200 xfsbdstrat( 1201 struct xfs_mount *mp, 1202 struct xfs_buf *bp) 1203 { 1204 if (XFS_FORCED_SHUTDOWN(mp)) { 1205 trace_xfs_bdstrat_shut(bp, _RET_IP_); 1206 xfs_bioerror_relse(bp); 1207 return; 1208 } 1209 1210 xfs_buf_iorequest(bp); 1211 } 1212 1213 STATIC void 1214 _xfs_buf_ioend( 1215 xfs_buf_t *bp, 1216 int schedule) 1217 { 1218 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) 1219 xfs_buf_ioend(bp, schedule); 1220 } 1221 1222 STATIC void 1223 xfs_buf_bio_end_io( 1224 struct bio *bio, 1225 int error) 1226 { 1227 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private; 1228 1229 /* 1230 * don't overwrite existing errors - otherwise we can lose errors on 1231 * buffers that require multiple bios to complete. 1232 */ 1233 if (!bp->b_error) 1234 xfs_buf_ioerror(bp, -error); 1235 1236 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ)) 1237 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp)); 1238 1239 _xfs_buf_ioend(bp, 1); 1240 bio_put(bio); 1241 } 1242 1243 static void 1244 xfs_buf_ioapply_map( 1245 struct xfs_buf *bp, 1246 int map, 1247 int *buf_offset, 1248 int *count, 1249 int rw) 1250 { 1251 int page_index; 1252 int total_nr_pages = bp->b_page_count; 1253 int nr_pages; 1254 struct bio *bio; 1255 sector_t sector = bp->b_maps[map].bm_bn; 1256 int size; 1257 int offset; 1258 1259 total_nr_pages = bp->b_page_count; 1260 1261 /* skip the pages in the buffer before the start offset */ 1262 page_index = 0; 1263 offset = *buf_offset; 1264 while (offset >= PAGE_SIZE) { 1265 page_index++; 1266 offset -= PAGE_SIZE; 1267 } 1268 1269 /* 1270 * Limit the IO size to the length of the current vector, and update the 1271 * remaining IO count for the next time around. 1272 */ 1273 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count); 1274 *count -= size; 1275 *buf_offset += size; 1276 1277 next_chunk: 1278 atomic_inc(&bp->b_io_remaining); 1279 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT); 1280 if (nr_pages > total_nr_pages) 1281 nr_pages = total_nr_pages; 1282 1283 bio = bio_alloc(GFP_NOIO, nr_pages); 1284 bio->bi_bdev = bp->b_target->bt_bdev; 1285 bio->bi_sector = sector; 1286 bio->bi_end_io = xfs_buf_bio_end_io; 1287 bio->bi_private = bp; 1288 1289 1290 for (; size && nr_pages; nr_pages--, page_index++) { 1291 int rbytes, nbytes = PAGE_SIZE - offset; 1292 1293 if (nbytes > size) 1294 nbytes = size; 1295 1296 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes, 1297 offset); 1298 if (rbytes < nbytes) 1299 break; 1300 1301 offset = 0; 1302 sector += BTOBB(nbytes); 1303 size -= nbytes; 1304 total_nr_pages--; 1305 } 1306 1307 if (likely(bio->bi_size)) { 1308 if (xfs_buf_is_vmapped(bp)) { 1309 flush_kernel_vmap_range(bp->b_addr, 1310 xfs_buf_vmap_len(bp)); 1311 } 1312 submit_bio(rw, bio); 1313 if (size) 1314 goto next_chunk; 1315 } else { 1316 /* 1317 * This is guaranteed not to be the last io reference count 1318 * because the caller (xfs_buf_iorequest) holds a count itself. 1319 */ 1320 atomic_dec(&bp->b_io_remaining); 1321 xfs_buf_ioerror(bp, EIO); 1322 bio_put(bio); 1323 } 1324 1325 } 1326 1327 STATIC void 1328 _xfs_buf_ioapply( 1329 struct xfs_buf *bp) 1330 { 1331 struct blk_plug plug; 1332 int rw; 1333 int offset; 1334 int size; 1335 int i; 1336 1337 if (bp->b_flags & XBF_WRITE) { 1338 if (bp->b_flags & XBF_SYNCIO) 1339 rw = WRITE_SYNC; 1340 else 1341 rw = WRITE; 1342 if (bp->b_flags & XBF_FUA) 1343 rw |= REQ_FUA; 1344 if (bp->b_flags & XBF_FLUSH) 1345 rw |= REQ_FLUSH; 1346 1347 /* 1348 * Run the write verifier callback function if it exists. If 1349 * this function fails it will mark the buffer with an error and 1350 * the IO should not be dispatched. 1351 */ 1352 if (bp->b_ops) { 1353 bp->b_ops->verify_write(bp); 1354 if (bp->b_error) { 1355 xfs_force_shutdown(bp->b_target->bt_mount, 1356 SHUTDOWN_CORRUPT_INCORE); 1357 return; 1358 } 1359 } 1360 } else if (bp->b_flags & XBF_READ_AHEAD) { 1361 rw = READA; 1362 } else { 1363 rw = READ; 1364 } 1365 1366 /* we only use the buffer cache for meta-data */ 1367 rw |= REQ_META; 1368 1369 /* 1370 * Walk all the vectors issuing IO on them. Set up the initial offset 1371 * into the buffer and the desired IO size before we start - 1372 * _xfs_buf_ioapply_vec() will modify them appropriately for each 1373 * subsequent call. 1374 */ 1375 offset = bp->b_offset; 1376 size = BBTOB(bp->b_io_length); 1377 blk_start_plug(&plug); 1378 for (i = 0; i < bp->b_map_count; i++) { 1379 xfs_buf_ioapply_map(bp, i, &offset, &size, rw); 1380 if (bp->b_error) 1381 break; 1382 if (size <= 0) 1383 break; /* all done */ 1384 } 1385 blk_finish_plug(&plug); 1386 } 1387 1388 void 1389 xfs_buf_iorequest( 1390 xfs_buf_t *bp) 1391 { 1392 trace_xfs_buf_iorequest(bp, _RET_IP_); 1393 1394 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); 1395 1396 if (bp->b_flags & XBF_WRITE) 1397 xfs_buf_wait_unpin(bp); 1398 xfs_buf_hold(bp); 1399 1400 /* Set the count to 1 initially, this will stop an I/O 1401 * completion callout which happens before we have started 1402 * all the I/O from calling xfs_buf_ioend too early. 1403 */ 1404 atomic_set(&bp->b_io_remaining, 1); 1405 _xfs_buf_ioapply(bp); 1406 _xfs_buf_ioend(bp, 1); 1407 1408 xfs_buf_rele(bp); 1409 } 1410 1411 /* 1412 * Waits for I/O to complete on the buffer supplied. It returns immediately if 1413 * no I/O is pending or there is already a pending error on the buffer. It 1414 * returns the I/O error code, if any, or 0 if there was no error. 1415 */ 1416 int 1417 xfs_buf_iowait( 1418 xfs_buf_t *bp) 1419 { 1420 trace_xfs_buf_iowait(bp, _RET_IP_); 1421 1422 if (!bp->b_error) 1423 wait_for_completion(&bp->b_iowait); 1424 1425 trace_xfs_buf_iowait_done(bp, _RET_IP_); 1426 return bp->b_error; 1427 } 1428 1429 xfs_caddr_t 1430 xfs_buf_offset( 1431 xfs_buf_t *bp, 1432 size_t offset) 1433 { 1434 struct page *page; 1435 1436 if (bp->b_addr) 1437 return bp->b_addr + offset; 1438 1439 offset += bp->b_offset; 1440 page = bp->b_pages[offset >> PAGE_SHIFT]; 1441 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1)); 1442 } 1443 1444 /* 1445 * Move data into or out of a buffer. 1446 */ 1447 void 1448 xfs_buf_iomove( 1449 xfs_buf_t *bp, /* buffer to process */ 1450 size_t boff, /* starting buffer offset */ 1451 size_t bsize, /* length to copy */ 1452 void *data, /* data address */ 1453 xfs_buf_rw_t mode) /* read/write/zero flag */ 1454 { 1455 size_t bend; 1456 1457 bend = boff + bsize; 1458 while (boff < bend) { 1459 struct page *page; 1460 int page_index, page_offset, csize; 1461 1462 page_index = (boff + bp->b_offset) >> PAGE_SHIFT; 1463 page_offset = (boff + bp->b_offset) & ~PAGE_MASK; 1464 page = bp->b_pages[page_index]; 1465 csize = min_t(size_t, PAGE_SIZE - page_offset, 1466 BBTOB(bp->b_io_length) - boff); 1467 1468 ASSERT((csize + page_offset) <= PAGE_SIZE); 1469 1470 switch (mode) { 1471 case XBRW_ZERO: 1472 memset(page_address(page) + page_offset, 0, csize); 1473 break; 1474 case XBRW_READ: 1475 memcpy(data, page_address(page) + page_offset, csize); 1476 break; 1477 case XBRW_WRITE: 1478 memcpy(page_address(page) + page_offset, data, csize); 1479 } 1480 1481 boff += csize; 1482 data += csize; 1483 } 1484 } 1485 1486 /* 1487 * Handling of buffer targets (buftargs). 1488 */ 1489 1490 /* 1491 * Wait for any bufs with callbacks that have been submitted but have not yet 1492 * returned. These buffers will have an elevated hold count, so wait on those 1493 * while freeing all the buffers only held by the LRU. 1494 */ 1495 void 1496 xfs_wait_buftarg( 1497 struct xfs_buftarg *btp) 1498 { 1499 struct xfs_buf *bp; 1500 1501 restart: 1502 spin_lock(&btp->bt_lru_lock); 1503 while (!list_empty(&btp->bt_lru)) { 1504 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru); 1505 if (atomic_read(&bp->b_hold) > 1) { 1506 trace_xfs_buf_wait_buftarg(bp, _RET_IP_); 1507 list_move_tail(&bp->b_lru, &btp->bt_lru); 1508 spin_unlock(&btp->bt_lru_lock); 1509 delay(100); 1510 goto restart; 1511 } 1512 /* 1513 * clear the LRU reference count so the buffer doesn't get 1514 * ignored in xfs_buf_rele(). 1515 */ 1516 atomic_set(&bp->b_lru_ref, 0); 1517 spin_unlock(&btp->bt_lru_lock); 1518 xfs_buf_rele(bp); 1519 spin_lock(&btp->bt_lru_lock); 1520 } 1521 spin_unlock(&btp->bt_lru_lock); 1522 } 1523 1524 int 1525 xfs_buftarg_shrink( 1526 struct shrinker *shrink, 1527 struct shrink_control *sc) 1528 { 1529 struct xfs_buftarg *btp = container_of(shrink, 1530 struct xfs_buftarg, bt_shrinker); 1531 struct xfs_buf *bp; 1532 int nr_to_scan = sc->nr_to_scan; 1533 LIST_HEAD(dispose); 1534 1535 if (!nr_to_scan) 1536 return btp->bt_lru_nr; 1537 1538 spin_lock(&btp->bt_lru_lock); 1539 while (!list_empty(&btp->bt_lru)) { 1540 if (nr_to_scan-- <= 0) 1541 break; 1542 1543 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru); 1544 1545 /* 1546 * Decrement the b_lru_ref count unless the value is already 1547 * zero. If the value is already zero, we need to reclaim the 1548 * buffer, otherwise it gets another trip through the LRU. 1549 */ 1550 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) { 1551 list_move_tail(&bp->b_lru, &btp->bt_lru); 1552 continue; 1553 } 1554 1555 /* 1556 * remove the buffer from the LRU now to avoid needing another 1557 * lock round trip inside xfs_buf_rele(). 1558 */ 1559 list_move(&bp->b_lru, &dispose); 1560 btp->bt_lru_nr--; 1561 bp->b_lru_flags |= _XBF_LRU_DISPOSE; 1562 } 1563 spin_unlock(&btp->bt_lru_lock); 1564 1565 while (!list_empty(&dispose)) { 1566 bp = list_first_entry(&dispose, struct xfs_buf, b_lru); 1567 list_del_init(&bp->b_lru); 1568 xfs_buf_rele(bp); 1569 } 1570 1571 return btp->bt_lru_nr; 1572 } 1573 1574 void 1575 xfs_free_buftarg( 1576 struct xfs_mount *mp, 1577 struct xfs_buftarg *btp) 1578 { 1579 unregister_shrinker(&btp->bt_shrinker); 1580 1581 if (mp->m_flags & XFS_MOUNT_BARRIER) 1582 xfs_blkdev_issue_flush(btp); 1583 1584 kmem_free(btp); 1585 } 1586 1587 STATIC int 1588 xfs_setsize_buftarg_flags( 1589 xfs_buftarg_t *btp, 1590 unsigned int blocksize, 1591 unsigned int sectorsize, 1592 int verbose) 1593 { 1594 btp->bt_bsize = blocksize; 1595 btp->bt_sshift = ffs(sectorsize) - 1; 1596 btp->bt_smask = sectorsize - 1; 1597 1598 if (set_blocksize(btp->bt_bdev, sectorsize)) { 1599 char name[BDEVNAME_SIZE]; 1600 1601 bdevname(btp->bt_bdev, name); 1602 1603 xfs_warn(btp->bt_mount, 1604 "Cannot set_blocksize to %u on device %s\n", 1605 sectorsize, name); 1606 return EINVAL; 1607 } 1608 1609 return 0; 1610 } 1611 1612 /* 1613 * When allocating the initial buffer target we have not yet 1614 * read in the superblock, so don't know what sized sectors 1615 * are being used is at this early stage. Play safe. 1616 */ 1617 STATIC int 1618 xfs_setsize_buftarg_early( 1619 xfs_buftarg_t *btp, 1620 struct block_device *bdev) 1621 { 1622 return xfs_setsize_buftarg_flags(btp, 1623 PAGE_SIZE, bdev_logical_block_size(bdev), 0); 1624 } 1625 1626 int 1627 xfs_setsize_buftarg( 1628 xfs_buftarg_t *btp, 1629 unsigned int blocksize, 1630 unsigned int sectorsize) 1631 { 1632 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1); 1633 } 1634 1635 xfs_buftarg_t * 1636 xfs_alloc_buftarg( 1637 struct xfs_mount *mp, 1638 struct block_device *bdev, 1639 int external, 1640 const char *fsname) 1641 { 1642 xfs_buftarg_t *btp; 1643 1644 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP); 1645 1646 btp->bt_mount = mp; 1647 btp->bt_dev = bdev->bd_dev; 1648 btp->bt_bdev = bdev; 1649 btp->bt_bdi = blk_get_backing_dev_info(bdev); 1650 if (!btp->bt_bdi) 1651 goto error; 1652 1653 INIT_LIST_HEAD(&btp->bt_lru); 1654 spin_lock_init(&btp->bt_lru_lock); 1655 if (xfs_setsize_buftarg_early(btp, bdev)) 1656 goto error; 1657 btp->bt_shrinker.shrink = xfs_buftarg_shrink; 1658 btp->bt_shrinker.seeks = DEFAULT_SEEKS; 1659 register_shrinker(&btp->bt_shrinker); 1660 return btp; 1661 1662 error: 1663 kmem_free(btp); 1664 return NULL; 1665 } 1666 1667 /* 1668 * Add a buffer to the delayed write list. 1669 * 1670 * This queues a buffer for writeout if it hasn't already been. Note that 1671 * neither this routine nor the buffer list submission functions perform 1672 * any internal synchronization. It is expected that the lists are thread-local 1673 * to the callers. 1674 * 1675 * Returns true if we queued up the buffer, or false if it already had 1676 * been on the buffer list. 1677 */ 1678 bool 1679 xfs_buf_delwri_queue( 1680 struct xfs_buf *bp, 1681 struct list_head *list) 1682 { 1683 ASSERT(xfs_buf_islocked(bp)); 1684 ASSERT(!(bp->b_flags & XBF_READ)); 1685 1686 /* 1687 * If the buffer is already marked delwri it already is queued up 1688 * by someone else for imediate writeout. Just ignore it in that 1689 * case. 1690 */ 1691 if (bp->b_flags & _XBF_DELWRI_Q) { 1692 trace_xfs_buf_delwri_queued(bp, _RET_IP_); 1693 return false; 1694 } 1695 1696 trace_xfs_buf_delwri_queue(bp, _RET_IP_); 1697 1698 /* 1699 * If a buffer gets written out synchronously or marked stale while it 1700 * is on a delwri list we lazily remove it. To do this, the other party 1701 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone. 1702 * It remains referenced and on the list. In a rare corner case it 1703 * might get readded to a delwri list after the synchronous writeout, in 1704 * which case we need just need to re-add the flag here. 1705 */ 1706 bp->b_flags |= _XBF_DELWRI_Q; 1707 if (list_empty(&bp->b_list)) { 1708 atomic_inc(&bp->b_hold); 1709 list_add_tail(&bp->b_list, list); 1710 } 1711 1712 return true; 1713 } 1714 1715 /* 1716 * Compare function is more complex than it needs to be because 1717 * the return value is only 32 bits and we are doing comparisons 1718 * on 64 bit values 1719 */ 1720 static int 1721 xfs_buf_cmp( 1722 void *priv, 1723 struct list_head *a, 1724 struct list_head *b) 1725 { 1726 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list); 1727 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list); 1728 xfs_daddr_t diff; 1729 1730 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn; 1731 if (diff < 0) 1732 return -1; 1733 if (diff > 0) 1734 return 1; 1735 return 0; 1736 } 1737 1738 static int 1739 __xfs_buf_delwri_submit( 1740 struct list_head *buffer_list, 1741 struct list_head *io_list, 1742 bool wait) 1743 { 1744 struct blk_plug plug; 1745 struct xfs_buf *bp, *n; 1746 int pinned = 0; 1747 1748 list_for_each_entry_safe(bp, n, buffer_list, b_list) { 1749 if (!wait) { 1750 if (xfs_buf_ispinned(bp)) { 1751 pinned++; 1752 continue; 1753 } 1754 if (!xfs_buf_trylock(bp)) 1755 continue; 1756 } else { 1757 xfs_buf_lock(bp); 1758 } 1759 1760 /* 1761 * Someone else might have written the buffer synchronously or 1762 * marked it stale in the meantime. In that case only the 1763 * _XBF_DELWRI_Q flag got cleared, and we have to drop the 1764 * reference and remove it from the list here. 1765 */ 1766 if (!(bp->b_flags & _XBF_DELWRI_Q)) { 1767 list_del_init(&bp->b_list); 1768 xfs_buf_relse(bp); 1769 continue; 1770 } 1771 1772 list_move_tail(&bp->b_list, io_list); 1773 trace_xfs_buf_delwri_split(bp, _RET_IP_); 1774 } 1775 1776 list_sort(NULL, io_list, xfs_buf_cmp); 1777 1778 blk_start_plug(&plug); 1779 list_for_each_entry_safe(bp, n, io_list, b_list) { 1780 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC); 1781 bp->b_flags |= XBF_WRITE; 1782 1783 if (!wait) { 1784 bp->b_flags |= XBF_ASYNC; 1785 list_del_init(&bp->b_list); 1786 } 1787 xfs_bdstrat_cb(bp); 1788 } 1789 blk_finish_plug(&plug); 1790 1791 return pinned; 1792 } 1793 1794 /* 1795 * Write out a buffer list asynchronously. 1796 * 1797 * This will take the @buffer_list, write all non-locked and non-pinned buffers 1798 * out and not wait for I/O completion on any of the buffers. This interface 1799 * is only safely useable for callers that can track I/O completion by higher 1800 * level means, e.g. AIL pushing as the @buffer_list is consumed in this 1801 * function. 1802 */ 1803 int 1804 xfs_buf_delwri_submit_nowait( 1805 struct list_head *buffer_list) 1806 { 1807 LIST_HEAD (io_list); 1808 return __xfs_buf_delwri_submit(buffer_list, &io_list, false); 1809 } 1810 1811 /* 1812 * Write out a buffer list synchronously. 1813 * 1814 * This will take the @buffer_list, write all buffers out and wait for I/O 1815 * completion on all of the buffers. @buffer_list is consumed by the function, 1816 * so callers must have some other way of tracking buffers if they require such 1817 * functionality. 1818 */ 1819 int 1820 xfs_buf_delwri_submit( 1821 struct list_head *buffer_list) 1822 { 1823 LIST_HEAD (io_list); 1824 int error = 0, error2; 1825 struct xfs_buf *bp; 1826 1827 __xfs_buf_delwri_submit(buffer_list, &io_list, true); 1828 1829 /* Wait for IO to complete. */ 1830 while (!list_empty(&io_list)) { 1831 bp = list_first_entry(&io_list, struct xfs_buf, b_list); 1832 1833 list_del_init(&bp->b_list); 1834 error2 = xfs_buf_iowait(bp); 1835 xfs_buf_relse(bp); 1836 if (!error) 1837 error = error2; 1838 } 1839 1840 return error; 1841 } 1842 1843 int __init 1844 xfs_buf_init(void) 1845 { 1846 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf", 1847 KM_ZONE_HWALIGN, NULL); 1848 if (!xfs_buf_zone) 1849 goto out; 1850 1851 xfslogd_workqueue = alloc_workqueue("xfslogd", 1852 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1); 1853 if (!xfslogd_workqueue) 1854 goto out_free_buf_zone; 1855 1856 return 0; 1857 1858 out_free_buf_zone: 1859 kmem_zone_destroy(xfs_buf_zone); 1860 out: 1861 return -ENOMEM; 1862 } 1863 1864 void 1865 xfs_buf_terminate(void) 1866 { 1867 destroy_workqueue(xfslogd_workqueue); 1868 kmem_zone_destroy(xfs_buf_zone); 1869 } 1870