1 /* 2 * Copyright (c) 2000-2006 Silicon Graphics, Inc. 3 * All Rights Reserved. 4 * 5 * This program is free software; you can redistribute it and/or 6 * modify it under the terms of the GNU General Public License as 7 * published by the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it would be useful, 10 * but WITHOUT ANY WARRANTY; without even the implied warranty of 11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 * GNU General Public License for more details. 13 * 14 * You should have received a copy of the GNU General Public License 15 * along with this program; if not, write the Free Software Foundation, 16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA 17 */ 18 #include "xfs.h" 19 #include <linux/stddef.h> 20 #include <linux/errno.h> 21 #include <linux/gfp.h> 22 #include <linux/pagemap.h> 23 #include <linux/init.h> 24 #include <linux/vmalloc.h> 25 #include <linux/bio.h> 26 #include <linux/sysctl.h> 27 #include <linux/proc_fs.h> 28 #include <linux/workqueue.h> 29 #include <linux/percpu.h> 30 #include <linux/blkdev.h> 31 #include <linux/hash.h> 32 #include <linux/kthread.h> 33 #include <linux/migrate.h> 34 #include <linux/backing-dev.h> 35 #include <linux/freezer.h> 36 37 #include "xfs_log_format.h" 38 #include "xfs_trans_resv.h" 39 #include "xfs_sb.h" 40 #include "xfs_ag.h" 41 #include "xfs_mount.h" 42 #include "xfs_trace.h" 43 #include "xfs_log.h" 44 45 static kmem_zone_t *xfs_buf_zone; 46 47 static struct workqueue_struct *xfslogd_workqueue; 48 49 #ifdef XFS_BUF_LOCK_TRACKING 50 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid) 51 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1) 52 # define XB_GET_OWNER(bp) ((bp)->b_last_holder) 53 #else 54 # define XB_SET_OWNER(bp) do { } while (0) 55 # define XB_CLEAR_OWNER(bp) do { } while (0) 56 # define XB_GET_OWNER(bp) do { } while (0) 57 #endif 58 59 #define xb_to_gfp(flags) \ 60 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN) 61 62 63 static inline int 64 xfs_buf_is_vmapped( 65 struct xfs_buf *bp) 66 { 67 /* 68 * Return true if the buffer is vmapped. 69 * 70 * b_addr is null if the buffer is not mapped, but the code is clever 71 * enough to know it doesn't have to map a single page, so the check has 72 * to be both for b_addr and bp->b_page_count > 1. 73 */ 74 return bp->b_addr && bp->b_page_count > 1; 75 } 76 77 static inline int 78 xfs_buf_vmap_len( 79 struct xfs_buf *bp) 80 { 81 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset; 82 } 83 84 /* 85 * When we mark a buffer stale, we remove the buffer from the LRU and clear the 86 * b_lru_ref count so that the buffer is freed immediately when the buffer 87 * reference count falls to zero. If the buffer is already on the LRU, we need 88 * to remove the reference that LRU holds on the buffer. 89 * 90 * This prevents build-up of stale buffers on the LRU. 91 */ 92 void 93 xfs_buf_stale( 94 struct xfs_buf *bp) 95 { 96 ASSERT(xfs_buf_islocked(bp)); 97 98 bp->b_flags |= XBF_STALE; 99 100 /* 101 * Clear the delwri status so that a delwri queue walker will not 102 * flush this buffer to disk now that it is stale. The delwri queue has 103 * a reference to the buffer, so this is safe to do. 104 */ 105 bp->b_flags &= ~_XBF_DELWRI_Q; 106 107 spin_lock(&bp->b_lock); 108 atomic_set(&bp->b_lru_ref, 0); 109 if (!(bp->b_state & XFS_BSTATE_DISPOSE) && 110 (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru))) 111 atomic_dec(&bp->b_hold); 112 113 ASSERT(atomic_read(&bp->b_hold) >= 1); 114 spin_unlock(&bp->b_lock); 115 } 116 117 static int 118 xfs_buf_get_maps( 119 struct xfs_buf *bp, 120 int map_count) 121 { 122 ASSERT(bp->b_maps == NULL); 123 bp->b_map_count = map_count; 124 125 if (map_count == 1) { 126 bp->b_maps = &bp->__b_map; 127 return 0; 128 } 129 130 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map), 131 KM_NOFS); 132 if (!bp->b_maps) 133 return ENOMEM; 134 return 0; 135 } 136 137 /* 138 * Frees b_pages if it was allocated. 139 */ 140 static void 141 xfs_buf_free_maps( 142 struct xfs_buf *bp) 143 { 144 if (bp->b_maps != &bp->__b_map) { 145 kmem_free(bp->b_maps); 146 bp->b_maps = NULL; 147 } 148 } 149 150 struct xfs_buf * 151 _xfs_buf_alloc( 152 struct xfs_buftarg *target, 153 struct xfs_buf_map *map, 154 int nmaps, 155 xfs_buf_flags_t flags) 156 { 157 struct xfs_buf *bp; 158 int error; 159 int i; 160 161 bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS); 162 if (unlikely(!bp)) 163 return NULL; 164 165 /* 166 * We don't want certain flags to appear in b_flags unless they are 167 * specifically set by later operations on the buffer. 168 */ 169 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD); 170 171 atomic_set(&bp->b_hold, 1); 172 atomic_set(&bp->b_lru_ref, 1); 173 init_completion(&bp->b_iowait); 174 INIT_LIST_HEAD(&bp->b_lru); 175 INIT_LIST_HEAD(&bp->b_list); 176 RB_CLEAR_NODE(&bp->b_rbnode); 177 sema_init(&bp->b_sema, 0); /* held, no waiters */ 178 spin_lock_init(&bp->b_lock); 179 XB_SET_OWNER(bp); 180 bp->b_target = target; 181 bp->b_flags = flags; 182 183 /* 184 * Set length and io_length to the same value initially. 185 * I/O routines should use io_length, which will be the same in 186 * most cases but may be reset (e.g. XFS recovery). 187 */ 188 error = xfs_buf_get_maps(bp, nmaps); 189 if (error) { 190 kmem_zone_free(xfs_buf_zone, bp); 191 return NULL; 192 } 193 194 bp->b_bn = map[0].bm_bn; 195 bp->b_length = 0; 196 for (i = 0; i < nmaps; i++) { 197 bp->b_maps[i].bm_bn = map[i].bm_bn; 198 bp->b_maps[i].bm_len = map[i].bm_len; 199 bp->b_length += map[i].bm_len; 200 } 201 bp->b_io_length = bp->b_length; 202 203 atomic_set(&bp->b_pin_count, 0); 204 init_waitqueue_head(&bp->b_waiters); 205 206 XFS_STATS_INC(xb_create); 207 trace_xfs_buf_init(bp, _RET_IP_); 208 209 return bp; 210 } 211 212 /* 213 * Allocate a page array capable of holding a specified number 214 * of pages, and point the page buf at it. 215 */ 216 STATIC int 217 _xfs_buf_get_pages( 218 xfs_buf_t *bp, 219 int page_count) 220 { 221 /* Make sure that we have a page list */ 222 if (bp->b_pages == NULL) { 223 bp->b_page_count = page_count; 224 if (page_count <= XB_PAGES) { 225 bp->b_pages = bp->b_page_array; 226 } else { 227 bp->b_pages = kmem_alloc(sizeof(struct page *) * 228 page_count, KM_NOFS); 229 if (bp->b_pages == NULL) 230 return -ENOMEM; 231 } 232 memset(bp->b_pages, 0, sizeof(struct page *) * page_count); 233 } 234 return 0; 235 } 236 237 /* 238 * Frees b_pages if it was allocated. 239 */ 240 STATIC void 241 _xfs_buf_free_pages( 242 xfs_buf_t *bp) 243 { 244 if (bp->b_pages != bp->b_page_array) { 245 kmem_free(bp->b_pages); 246 bp->b_pages = NULL; 247 } 248 } 249 250 /* 251 * Releases the specified buffer. 252 * 253 * The modification state of any associated pages is left unchanged. 254 * The buffer must not be on any hash - use xfs_buf_rele instead for 255 * hashed and refcounted buffers 256 */ 257 void 258 xfs_buf_free( 259 xfs_buf_t *bp) 260 { 261 trace_xfs_buf_free(bp, _RET_IP_); 262 263 ASSERT(list_empty(&bp->b_lru)); 264 265 if (bp->b_flags & _XBF_PAGES) { 266 uint i; 267 268 if (xfs_buf_is_vmapped(bp)) 269 vm_unmap_ram(bp->b_addr - bp->b_offset, 270 bp->b_page_count); 271 272 for (i = 0; i < bp->b_page_count; i++) { 273 struct page *page = bp->b_pages[i]; 274 275 __free_page(page); 276 } 277 } else if (bp->b_flags & _XBF_KMEM) 278 kmem_free(bp->b_addr); 279 _xfs_buf_free_pages(bp); 280 xfs_buf_free_maps(bp); 281 kmem_zone_free(xfs_buf_zone, bp); 282 } 283 284 /* 285 * Allocates all the pages for buffer in question and builds it's page list. 286 */ 287 STATIC int 288 xfs_buf_allocate_memory( 289 xfs_buf_t *bp, 290 uint flags) 291 { 292 size_t size; 293 size_t nbytes, offset; 294 gfp_t gfp_mask = xb_to_gfp(flags); 295 unsigned short page_count, i; 296 xfs_off_t start, end; 297 int error; 298 299 /* 300 * for buffers that are contained within a single page, just allocate 301 * the memory from the heap - there's no need for the complexity of 302 * page arrays to keep allocation down to order 0. 303 */ 304 size = BBTOB(bp->b_length); 305 if (size < PAGE_SIZE) { 306 bp->b_addr = kmem_alloc(size, KM_NOFS); 307 if (!bp->b_addr) { 308 /* low memory - use alloc_page loop instead */ 309 goto use_alloc_page; 310 } 311 312 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) != 313 ((unsigned long)bp->b_addr & PAGE_MASK)) { 314 /* b_addr spans two pages - use alloc_page instead */ 315 kmem_free(bp->b_addr); 316 bp->b_addr = NULL; 317 goto use_alloc_page; 318 } 319 bp->b_offset = offset_in_page(bp->b_addr); 320 bp->b_pages = bp->b_page_array; 321 bp->b_pages[0] = virt_to_page(bp->b_addr); 322 bp->b_page_count = 1; 323 bp->b_flags |= _XBF_KMEM; 324 return 0; 325 } 326 327 use_alloc_page: 328 start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT; 329 end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1) 330 >> PAGE_SHIFT; 331 page_count = end - start; 332 error = _xfs_buf_get_pages(bp, page_count); 333 if (unlikely(error)) 334 return error; 335 336 offset = bp->b_offset; 337 bp->b_flags |= _XBF_PAGES; 338 339 for (i = 0; i < bp->b_page_count; i++) { 340 struct page *page; 341 uint retries = 0; 342 retry: 343 page = alloc_page(gfp_mask); 344 if (unlikely(page == NULL)) { 345 if (flags & XBF_READ_AHEAD) { 346 bp->b_page_count = i; 347 error = ENOMEM; 348 goto out_free_pages; 349 } 350 351 /* 352 * This could deadlock. 353 * 354 * But until all the XFS lowlevel code is revamped to 355 * handle buffer allocation failures we can't do much. 356 */ 357 if (!(++retries % 100)) 358 xfs_err(NULL, 359 "possible memory allocation deadlock in %s (mode:0x%x)", 360 __func__, gfp_mask); 361 362 XFS_STATS_INC(xb_page_retries); 363 congestion_wait(BLK_RW_ASYNC, HZ/50); 364 goto retry; 365 } 366 367 XFS_STATS_INC(xb_page_found); 368 369 nbytes = min_t(size_t, size, PAGE_SIZE - offset); 370 size -= nbytes; 371 bp->b_pages[i] = page; 372 offset = 0; 373 } 374 return 0; 375 376 out_free_pages: 377 for (i = 0; i < bp->b_page_count; i++) 378 __free_page(bp->b_pages[i]); 379 return error; 380 } 381 382 /* 383 * Map buffer into kernel address-space if necessary. 384 */ 385 STATIC int 386 _xfs_buf_map_pages( 387 xfs_buf_t *bp, 388 uint flags) 389 { 390 ASSERT(bp->b_flags & _XBF_PAGES); 391 if (bp->b_page_count == 1) { 392 /* A single page buffer is always mappable */ 393 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset; 394 } else if (flags & XBF_UNMAPPED) { 395 bp->b_addr = NULL; 396 } else { 397 int retried = 0; 398 unsigned noio_flag; 399 400 /* 401 * vm_map_ram() will allocate auxillary structures (e.g. 402 * pagetables) with GFP_KERNEL, yet we are likely to be under 403 * GFP_NOFS context here. Hence we need to tell memory reclaim 404 * that we are in such a context via PF_MEMALLOC_NOIO to prevent 405 * memory reclaim re-entering the filesystem here and 406 * potentially deadlocking. 407 */ 408 noio_flag = memalloc_noio_save(); 409 do { 410 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count, 411 -1, PAGE_KERNEL); 412 if (bp->b_addr) 413 break; 414 vm_unmap_aliases(); 415 } while (retried++ <= 1); 416 memalloc_noio_restore(noio_flag); 417 418 if (!bp->b_addr) 419 return -ENOMEM; 420 bp->b_addr += bp->b_offset; 421 } 422 423 return 0; 424 } 425 426 /* 427 * Finding and Reading Buffers 428 */ 429 430 /* 431 * Look up, and creates if absent, a lockable buffer for 432 * a given range of an inode. The buffer is returned 433 * locked. No I/O is implied by this call. 434 */ 435 xfs_buf_t * 436 _xfs_buf_find( 437 struct xfs_buftarg *btp, 438 struct xfs_buf_map *map, 439 int nmaps, 440 xfs_buf_flags_t flags, 441 xfs_buf_t *new_bp) 442 { 443 size_t numbytes; 444 struct xfs_perag *pag; 445 struct rb_node **rbp; 446 struct rb_node *parent; 447 xfs_buf_t *bp; 448 xfs_daddr_t blkno = map[0].bm_bn; 449 xfs_daddr_t eofs; 450 int numblks = 0; 451 int i; 452 453 for (i = 0; i < nmaps; i++) 454 numblks += map[i].bm_len; 455 numbytes = BBTOB(numblks); 456 457 /* Check for IOs smaller than the sector size / not sector aligned */ 458 ASSERT(!(numbytes < btp->bt_meta_sectorsize)); 459 ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_meta_sectormask)); 460 461 /* 462 * Corrupted block numbers can get through to here, unfortunately, so we 463 * have to check that the buffer falls within the filesystem bounds. 464 */ 465 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks); 466 if (blkno >= eofs) { 467 /* 468 * XXX (dgc): we should really be returning EFSCORRUPTED here, 469 * but none of the higher level infrastructure supports 470 * returning a specific error on buffer lookup failures. 471 */ 472 xfs_alert(btp->bt_mount, 473 "%s: Block out of range: block 0x%llx, EOFS 0x%llx ", 474 __func__, blkno, eofs); 475 WARN_ON(1); 476 return NULL; 477 } 478 479 /* get tree root */ 480 pag = xfs_perag_get(btp->bt_mount, 481 xfs_daddr_to_agno(btp->bt_mount, blkno)); 482 483 /* walk tree */ 484 spin_lock(&pag->pag_buf_lock); 485 rbp = &pag->pag_buf_tree.rb_node; 486 parent = NULL; 487 bp = NULL; 488 while (*rbp) { 489 parent = *rbp; 490 bp = rb_entry(parent, struct xfs_buf, b_rbnode); 491 492 if (blkno < bp->b_bn) 493 rbp = &(*rbp)->rb_left; 494 else if (blkno > bp->b_bn) 495 rbp = &(*rbp)->rb_right; 496 else { 497 /* 498 * found a block number match. If the range doesn't 499 * match, the only way this is allowed is if the buffer 500 * in the cache is stale and the transaction that made 501 * it stale has not yet committed. i.e. we are 502 * reallocating a busy extent. Skip this buffer and 503 * continue searching to the right for an exact match. 504 */ 505 if (bp->b_length != numblks) { 506 ASSERT(bp->b_flags & XBF_STALE); 507 rbp = &(*rbp)->rb_right; 508 continue; 509 } 510 atomic_inc(&bp->b_hold); 511 goto found; 512 } 513 } 514 515 /* No match found */ 516 if (new_bp) { 517 rb_link_node(&new_bp->b_rbnode, parent, rbp); 518 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree); 519 /* the buffer keeps the perag reference until it is freed */ 520 new_bp->b_pag = pag; 521 spin_unlock(&pag->pag_buf_lock); 522 } else { 523 XFS_STATS_INC(xb_miss_locked); 524 spin_unlock(&pag->pag_buf_lock); 525 xfs_perag_put(pag); 526 } 527 return new_bp; 528 529 found: 530 spin_unlock(&pag->pag_buf_lock); 531 xfs_perag_put(pag); 532 533 if (!xfs_buf_trylock(bp)) { 534 if (flags & XBF_TRYLOCK) { 535 xfs_buf_rele(bp); 536 XFS_STATS_INC(xb_busy_locked); 537 return NULL; 538 } 539 xfs_buf_lock(bp); 540 XFS_STATS_INC(xb_get_locked_waited); 541 } 542 543 /* 544 * if the buffer is stale, clear all the external state associated with 545 * it. We need to keep flags such as how we allocated the buffer memory 546 * intact here. 547 */ 548 if (bp->b_flags & XBF_STALE) { 549 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0); 550 ASSERT(bp->b_iodone == NULL); 551 bp->b_flags &= _XBF_KMEM | _XBF_PAGES; 552 bp->b_ops = NULL; 553 } 554 555 trace_xfs_buf_find(bp, flags, _RET_IP_); 556 XFS_STATS_INC(xb_get_locked); 557 return bp; 558 } 559 560 /* 561 * Assembles a buffer covering the specified range. The code is optimised for 562 * cache hits, as metadata intensive workloads will see 3 orders of magnitude 563 * more hits than misses. 564 */ 565 struct xfs_buf * 566 xfs_buf_get_map( 567 struct xfs_buftarg *target, 568 struct xfs_buf_map *map, 569 int nmaps, 570 xfs_buf_flags_t flags) 571 { 572 struct xfs_buf *bp; 573 struct xfs_buf *new_bp; 574 int error = 0; 575 576 bp = _xfs_buf_find(target, map, nmaps, flags, NULL); 577 if (likely(bp)) 578 goto found; 579 580 new_bp = _xfs_buf_alloc(target, map, nmaps, flags); 581 if (unlikely(!new_bp)) 582 return NULL; 583 584 error = xfs_buf_allocate_memory(new_bp, flags); 585 if (error) { 586 xfs_buf_free(new_bp); 587 return NULL; 588 } 589 590 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp); 591 if (!bp) { 592 xfs_buf_free(new_bp); 593 return NULL; 594 } 595 596 if (bp != new_bp) 597 xfs_buf_free(new_bp); 598 599 found: 600 if (!bp->b_addr) { 601 error = _xfs_buf_map_pages(bp, flags); 602 if (unlikely(error)) { 603 xfs_warn(target->bt_mount, 604 "%s: failed to map pagesn", __func__); 605 xfs_buf_relse(bp); 606 return NULL; 607 } 608 } 609 610 XFS_STATS_INC(xb_get); 611 trace_xfs_buf_get(bp, flags, _RET_IP_); 612 return bp; 613 } 614 615 STATIC int 616 _xfs_buf_read( 617 xfs_buf_t *bp, 618 xfs_buf_flags_t flags) 619 { 620 ASSERT(!(flags & XBF_WRITE)); 621 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL); 622 623 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD); 624 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD); 625 626 xfs_buf_iorequest(bp); 627 if (flags & XBF_ASYNC) 628 return 0; 629 return xfs_buf_iowait(bp); 630 } 631 632 xfs_buf_t * 633 xfs_buf_read_map( 634 struct xfs_buftarg *target, 635 struct xfs_buf_map *map, 636 int nmaps, 637 xfs_buf_flags_t flags, 638 const struct xfs_buf_ops *ops) 639 { 640 struct xfs_buf *bp; 641 642 flags |= XBF_READ; 643 644 bp = xfs_buf_get_map(target, map, nmaps, flags); 645 if (bp) { 646 trace_xfs_buf_read(bp, flags, _RET_IP_); 647 648 if (!XFS_BUF_ISDONE(bp)) { 649 XFS_STATS_INC(xb_get_read); 650 bp->b_ops = ops; 651 _xfs_buf_read(bp, flags); 652 } else if (flags & XBF_ASYNC) { 653 /* 654 * Read ahead call which is already satisfied, 655 * drop the buffer 656 */ 657 xfs_buf_relse(bp); 658 return NULL; 659 } else { 660 /* We do not want read in the flags */ 661 bp->b_flags &= ~XBF_READ; 662 } 663 } 664 665 return bp; 666 } 667 668 /* 669 * If we are not low on memory then do the readahead in a deadlock 670 * safe manner. 671 */ 672 void 673 xfs_buf_readahead_map( 674 struct xfs_buftarg *target, 675 struct xfs_buf_map *map, 676 int nmaps, 677 const struct xfs_buf_ops *ops) 678 { 679 if (bdi_read_congested(target->bt_bdi)) 680 return; 681 682 xfs_buf_read_map(target, map, nmaps, 683 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops); 684 } 685 686 /* 687 * Read an uncached buffer from disk. Allocates and returns a locked 688 * buffer containing the disk contents or nothing. 689 */ 690 struct xfs_buf * 691 xfs_buf_read_uncached( 692 struct xfs_buftarg *target, 693 xfs_daddr_t daddr, 694 size_t numblks, 695 int flags, 696 const struct xfs_buf_ops *ops) 697 { 698 struct xfs_buf *bp; 699 700 bp = xfs_buf_get_uncached(target, numblks, flags); 701 if (!bp) 702 return NULL; 703 704 /* set up the buffer for a read IO */ 705 ASSERT(bp->b_map_count == 1); 706 bp->b_bn = daddr; 707 bp->b_maps[0].bm_bn = daddr; 708 bp->b_flags |= XBF_READ; 709 bp->b_ops = ops; 710 711 if (XFS_FORCED_SHUTDOWN(target->bt_mount)) { 712 xfs_buf_relse(bp); 713 return NULL; 714 } 715 xfs_buf_iorequest(bp); 716 xfs_buf_iowait(bp); 717 return bp; 718 } 719 720 /* 721 * Return a buffer allocated as an empty buffer and associated to external 722 * memory via xfs_buf_associate_memory() back to it's empty state. 723 */ 724 void 725 xfs_buf_set_empty( 726 struct xfs_buf *bp, 727 size_t numblks) 728 { 729 if (bp->b_pages) 730 _xfs_buf_free_pages(bp); 731 732 bp->b_pages = NULL; 733 bp->b_page_count = 0; 734 bp->b_addr = NULL; 735 bp->b_length = numblks; 736 bp->b_io_length = numblks; 737 738 ASSERT(bp->b_map_count == 1); 739 bp->b_bn = XFS_BUF_DADDR_NULL; 740 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL; 741 bp->b_maps[0].bm_len = bp->b_length; 742 } 743 744 static inline struct page * 745 mem_to_page( 746 void *addr) 747 { 748 if ((!is_vmalloc_addr(addr))) { 749 return virt_to_page(addr); 750 } else { 751 return vmalloc_to_page(addr); 752 } 753 } 754 755 int 756 xfs_buf_associate_memory( 757 xfs_buf_t *bp, 758 void *mem, 759 size_t len) 760 { 761 int rval; 762 int i = 0; 763 unsigned long pageaddr; 764 unsigned long offset; 765 size_t buflen; 766 int page_count; 767 768 pageaddr = (unsigned long)mem & PAGE_MASK; 769 offset = (unsigned long)mem - pageaddr; 770 buflen = PAGE_ALIGN(len + offset); 771 page_count = buflen >> PAGE_SHIFT; 772 773 /* Free any previous set of page pointers */ 774 if (bp->b_pages) 775 _xfs_buf_free_pages(bp); 776 777 bp->b_pages = NULL; 778 bp->b_addr = mem; 779 780 rval = _xfs_buf_get_pages(bp, page_count); 781 if (rval) 782 return rval; 783 784 bp->b_offset = offset; 785 786 for (i = 0; i < bp->b_page_count; i++) { 787 bp->b_pages[i] = mem_to_page((void *)pageaddr); 788 pageaddr += PAGE_SIZE; 789 } 790 791 bp->b_io_length = BTOBB(len); 792 bp->b_length = BTOBB(buflen); 793 794 return 0; 795 } 796 797 xfs_buf_t * 798 xfs_buf_get_uncached( 799 struct xfs_buftarg *target, 800 size_t numblks, 801 int flags) 802 { 803 unsigned long page_count; 804 int error, i; 805 struct xfs_buf *bp; 806 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks); 807 808 bp = _xfs_buf_alloc(target, &map, 1, 0); 809 if (unlikely(bp == NULL)) 810 goto fail; 811 812 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT; 813 error = _xfs_buf_get_pages(bp, page_count); 814 if (error) 815 goto fail_free_buf; 816 817 for (i = 0; i < page_count; i++) { 818 bp->b_pages[i] = alloc_page(xb_to_gfp(flags)); 819 if (!bp->b_pages[i]) 820 goto fail_free_mem; 821 } 822 bp->b_flags |= _XBF_PAGES; 823 824 error = _xfs_buf_map_pages(bp, 0); 825 if (unlikely(error)) { 826 xfs_warn(target->bt_mount, 827 "%s: failed to map pages", __func__); 828 goto fail_free_mem; 829 } 830 831 trace_xfs_buf_get_uncached(bp, _RET_IP_); 832 return bp; 833 834 fail_free_mem: 835 while (--i >= 0) 836 __free_page(bp->b_pages[i]); 837 _xfs_buf_free_pages(bp); 838 fail_free_buf: 839 xfs_buf_free_maps(bp); 840 kmem_zone_free(xfs_buf_zone, bp); 841 fail: 842 return NULL; 843 } 844 845 /* 846 * Increment reference count on buffer, to hold the buffer concurrently 847 * with another thread which may release (free) the buffer asynchronously. 848 * Must hold the buffer already to call this function. 849 */ 850 void 851 xfs_buf_hold( 852 xfs_buf_t *bp) 853 { 854 trace_xfs_buf_hold(bp, _RET_IP_); 855 atomic_inc(&bp->b_hold); 856 } 857 858 /* 859 * Releases a hold on the specified buffer. If the 860 * the hold count is 1, calls xfs_buf_free. 861 */ 862 void 863 xfs_buf_rele( 864 xfs_buf_t *bp) 865 { 866 struct xfs_perag *pag = bp->b_pag; 867 868 trace_xfs_buf_rele(bp, _RET_IP_); 869 870 if (!pag) { 871 ASSERT(list_empty(&bp->b_lru)); 872 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode)); 873 if (atomic_dec_and_test(&bp->b_hold)) 874 xfs_buf_free(bp); 875 return; 876 } 877 878 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode)); 879 880 ASSERT(atomic_read(&bp->b_hold) > 0); 881 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) { 882 spin_lock(&bp->b_lock); 883 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) { 884 /* 885 * If the buffer is added to the LRU take a new 886 * reference to the buffer for the LRU and clear the 887 * (now stale) dispose list state flag 888 */ 889 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) { 890 bp->b_state &= ~XFS_BSTATE_DISPOSE; 891 atomic_inc(&bp->b_hold); 892 } 893 spin_unlock(&bp->b_lock); 894 spin_unlock(&pag->pag_buf_lock); 895 } else { 896 /* 897 * most of the time buffers will already be removed from 898 * the LRU, so optimise that case by checking for the 899 * XFS_BSTATE_DISPOSE flag indicating the last list the 900 * buffer was on was the disposal list 901 */ 902 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) { 903 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru); 904 } else { 905 ASSERT(list_empty(&bp->b_lru)); 906 } 907 spin_unlock(&bp->b_lock); 908 909 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); 910 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree); 911 spin_unlock(&pag->pag_buf_lock); 912 xfs_perag_put(pag); 913 xfs_buf_free(bp); 914 } 915 } 916 } 917 918 919 /* 920 * Lock a buffer object, if it is not already locked. 921 * 922 * If we come across a stale, pinned, locked buffer, we know that we are 923 * being asked to lock a buffer that has been reallocated. Because it is 924 * pinned, we know that the log has not been pushed to disk and hence it 925 * will still be locked. Rather than continuing to have trylock attempts 926 * fail until someone else pushes the log, push it ourselves before 927 * returning. This means that the xfsaild will not get stuck trying 928 * to push on stale inode buffers. 929 */ 930 int 931 xfs_buf_trylock( 932 struct xfs_buf *bp) 933 { 934 int locked; 935 936 locked = down_trylock(&bp->b_sema) == 0; 937 if (locked) 938 XB_SET_OWNER(bp); 939 940 trace_xfs_buf_trylock(bp, _RET_IP_); 941 return locked; 942 } 943 944 /* 945 * Lock a buffer object. 946 * 947 * If we come across a stale, pinned, locked buffer, we know that we 948 * are being asked to lock a buffer that has been reallocated. Because 949 * it is pinned, we know that the log has not been pushed to disk and 950 * hence it will still be locked. Rather than sleeping until someone 951 * else pushes the log, push it ourselves before trying to get the lock. 952 */ 953 void 954 xfs_buf_lock( 955 struct xfs_buf *bp) 956 { 957 trace_xfs_buf_lock(bp, _RET_IP_); 958 959 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE)) 960 xfs_log_force(bp->b_target->bt_mount, 0); 961 down(&bp->b_sema); 962 XB_SET_OWNER(bp); 963 964 trace_xfs_buf_lock_done(bp, _RET_IP_); 965 } 966 967 void 968 xfs_buf_unlock( 969 struct xfs_buf *bp) 970 { 971 XB_CLEAR_OWNER(bp); 972 up(&bp->b_sema); 973 974 trace_xfs_buf_unlock(bp, _RET_IP_); 975 } 976 977 STATIC void 978 xfs_buf_wait_unpin( 979 xfs_buf_t *bp) 980 { 981 DECLARE_WAITQUEUE (wait, current); 982 983 if (atomic_read(&bp->b_pin_count) == 0) 984 return; 985 986 add_wait_queue(&bp->b_waiters, &wait); 987 for (;;) { 988 set_current_state(TASK_UNINTERRUPTIBLE); 989 if (atomic_read(&bp->b_pin_count) == 0) 990 break; 991 io_schedule(); 992 } 993 remove_wait_queue(&bp->b_waiters, &wait); 994 set_current_state(TASK_RUNNING); 995 } 996 997 /* 998 * Buffer Utility Routines 999 */ 1000 1001 STATIC void 1002 xfs_buf_iodone_work( 1003 struct work_struct *work) 1004 { 1005 struct xfs_buf *bp = 1006 container_of(work, xfs_buf_t, b_iodone_work); 1007 bool read = !!(bp->b_flags & XBF_READ); 1008 1009 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); 1010 1011 /* only validate buffers that were read without errors */ 1012 if (read && bp->b_ops && !bp->b_error && (bp->b_flags & XBF_DONE)) 1013 bp->b_ops->verify_read(bp); 1014 1015 if (bp->b_iodone) 1016 (*(bp->b_iodone))(bp); 1017 else if (bp->b_flags & XBF_ASYNC) 1018 xfs_buf_relse(bp); 1019 else { 1020 ASSERT(read && bp->b_ops); 1021 complete(&bp->b_iowait); 1022 } 1023 } 1024 1025 void 1026 xfs_buf_ioend( 1027 struct xfs_buf *bp, 1028 int schedule) 1029 { 1030 bool read = !!(bp->b_flags & XBF_READ); 1031 1032 trace_xfs_buf_iodone(bp, _RET_IP_); 1033 1034 if (bp->b_error == 0) 1035 bp->b_flags |= XBF_DONE; 1036 1037 if (bp->b_iodone || (read && bp->b_ops) || (bp->b_flags & XBF_ASYNC)) { 1038 if (schedule) { 1039 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work); 1040 queue_work(xfslogd_workqueue, &bp->b_iodone_work); 1041 } else { 1042 xfs_buf_iodone_work(&bp->b_iodone_work); 1043 } 1044 } else { 1045 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); 1046 complete(&bp->b_iowait); 1047 } 1048 } 1049 1050 void 1051 xfs_buf_ioerror( 1052 xfs_buf_t *bp, 1053 int error) 1054 { 1055 ASSERT(error >= 0 && error <= 0xffff); 1056 bp->b_error = (unsigned short)error; 1057 trace_xfs_buf_ioerror(bp, error, _RET_IP_); 1058 } 1059 1060 void 1061 xfs_buf_ioerror_alert( 1062 struct xfs_buf *bp, 1063 const char *func) 1064 { 1065 xfs_alert(bp->b_target->bt_mount, 1066 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d", 1067 (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length); 1068 } 1069 1070 /* 1071 * Called when we want to stop a buffer from getting written or read. 1072 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend 1073 * so that the proper iodone callbacks get called. 1074 */ 1075 STATIC int 1076 xfs_bioerror( 1077 xfs_buf_t *bp) 1078 { 1079 #ifdef XFSERRORDEBUG 1080 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone); 1081 #endif 1082 1083 /* 1084 * No need to wait until the buffer is unpinned, we aren't flushing it. 1085 */ 1086 xfs_buf_ioerror(bp, EIO); 1087 1088 /* 1089 * We're calling xfs_buf_ioend, so delete XBF_DONE flag. 1090 */ 1091 XFS_BUF_UNREAD(bp); 1092 XFS_BUF_UNDONE(bp); 1093 xfs_buf_stale(bp); 1094 1095 xfs_buf_ioend(bp, 0); 1096 1097 return EIO; 1098 } 1099 1100 /* 1101 * Same as xfs_bioerror, except that we are releasing the buffer 1102 * here ourselves, and avoiding the xfs_buf_ioend call. 1103 * This is meant for userdata errors; metadata bufs come with 1104 * iodone functions attached, so that we can track down errors. 1105 */ 1106 int 1107 xfs_bioerror_relse( 1108 struct xfs_buf *bp) 1109 { 1110 int64_t fl = bp->b_flags; 1111 /* 1112 * No need to wait until the buffer is unpinned. 1113 * We aren't flushing it. 1114 * 1115 * chunkhold expects B_DONE to be set, whether 1116 * we actually finish the I/O or not. We don't want to 1117 * change that interface. 1118 */ 1119 XFS_BUF_UNREAD(bp); 1120 XFS_BUF_DONE(bp); 1121 xfs_buf_stale(bp); 1122 bp->b_iodone = NULL; 1123 if (!(fl & XBF_ASYNC)) { 1124 /* 1125 * Mark b_error and B_ERROR _both_. 1126 * Lot's of chunkcache code assumes that. 1127 * There's no reason to mark error for 1128 * ASYNC buffers. 1129 */ 1130 xfs_buf_ioerror(bp, EIO); 1131 complete(&bp->b_iowait); 1132 } else { 1133 xfs_buf_relse(bp); 1134 } 1135 1136 return EIO; 1137 } 1138 1139 STATIC int 1140 xfs_bdstrat_cb( 1141 struct xfs_buf *bp) 1142 { 1143 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) { 1144 trace_xfs_bdstrat_shut(bp, _RET_IP_); 1145 /* 1146 * Metadata write that didn't get logged but 1147 * written delayed anyway. These aren't associated 1148 * with a transaction, and can be ignored. 1149 */ 1150 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp)) 1151 return xfs_bioerror_relse(bp); 1152 else 1153 return xfs_bioerror(bp); 1154 } 1155 1156 xfs_buf_iorequest(bp); 1157 return 0; 1158 } 1159 1160 int 1161 xfs_bwrite( 1162 struct xfs_buf *bp) 1163 { 1164 int error; 1165 1166 ASSERT(xfs_buf_islocked(bp)); 1167 1168 bp->b_flags |= XBF_WRITE; 1169 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q | XBF_WRITE_FAIL); 1170 1171 xfs_bdstrat_cb(bp); 1172 1173 error = xfs_buf_iowait(bp); 1174 if (error) { 1175 xfs_force_shutdown(bp->b_target->bt_mount, 1176 SHUTDOWN_META_IO_ERROR); 1177 } 1178 return error; 1179 } 1180 1181 STATIC void 1182 _xfs_buf_ioend( 1183 xfs_buf_t *bp, 1184 int schedule) 1185 { 1186 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) 1187 xfs_buf_ioend(bp, schedule); 1188 } 1189 1190 STATIC void 1191 xfs_buf_bio_end_io( 1192 struct bio *bio, 1193 int error) 1194 { 1195 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private; 1196 1197 /* 1198 * don't overwrite existing errors - otherwise we can lose errors on 1199 * buffers that require multiple bios to complete. 1200 */ 1201 if (!bp->b_error) 1202 xfs_buf_ioerror(bp, -error); 1203 1204 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ)) 1205 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp)); 1206 1207 _xfs_buf_ioend(bp, 1); 1208 bio_put(bio); 1209 } 1210 1211 static void 1212 xfs_buf_ioapply_map( 1213 struct xfs_buf *bp, 1214 int map, 1215 int *buf_offset, 1216 int *count, 1217 int rw) 1218 { 1219 int page_index; 1220 int total_nr_pages = bp->b_page_count; 1221 int nr_pages; 1222 struct bio *bio; 1223 sector_t sector = bp->b_maps[map].bm_bn; 1224 int size; 1225 int offset; 1226 1227 total_nr_pages = bp->b_page_count; 1228 1229 /* skip the pages in the buffer before the start offset */ 1230 page_index = 0; 1231 offset = *buf_offset; 1232 while (offset >= PAGE_SIZE) { 1233 page_index++; 1234 offset -= PAGE_SIZE; 1235 } 1236 1237 /* 1238 * Limit the IO size to the length of the current vector, and update the 1239 * remaining IO count for the next time around. 1240 */ 1241 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count); 1242 *count -= size; 1243 *buf_offset += size; 1244 1245 next_chunk: 1246 atomic_inc(&bp->b_io_remaining); 1247 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT); 1248 if (nr_pages > total_nr_pages) 1249 nr_pages = total_nr_pages; 1250 1251 bio = bio_alloc(GFP_NOIO, nr_pages); 1252 bio->bi_bdev = bp->b_target->bt_bdev; 1253 bio->bi_iter.bi_sector = sector; 1254 bio->bi_end_io = xfs_buf_bio_end_io; 1255 bio->bi_private = bp; 1256 1257 1258 for (; size && nr_pages; nr_pages--, page_index++) { 1259 int rbytes, nbytes = PAGE_SIZE - offset; 1260 1261 if (nbytes > size) 1262 nbytes = size; 1263 1264 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes, 1265 offset); 1266 if (rbytes < nbytes) 1267 break; 1268 1269 offset = 0; 1270 sector += BTOBB(nbytes); 1271 size -= nbytes; 1272 total_nr_pages--; 1273 } 1274 1275 if (likely(bio->bi_iter.bi_size)) { 1276 if (xfs_buf_is_vmapped(bp)) { 1277 flush_kernel_vmap_range(bp->b_addr, 1278 xfs_buf_vmap_len(bp)); 1279 } 1280 submit_bio(rw, bio); 1281 if (size) 1282 goto next_chunk; 1283 } else { 1284 /* 1285 * This is guaranteed not to be the last io reference count 1286 * because the caller (xfs_buf_iorequest) holds a count itself. 1287 */ 1288 atomic_dec(&bp->b_io_remaining); 1289 xfs_buf_ioerror(bp, EIO); 1290 bio_put(bio); 1291 } 1292 1293 } 1294 1295 STATIC void 1296 _xfs_buf_ioapply( 1297 struct xfs_buf *bp) 1298 { 1299 struct blk_plug plug; 1300 int rw; 1301 int offset; 1302 int size; 1303 int i; 1304 1305 /* 1306 * Make sure we capture only current IO errors rather than stale errors 1307 * left over from previous use of the buffer (e.g. failed readahead). 1308 */ 1309 bp->b_error = 0; 1310 1311 if (bp->b_flags & XBF_WRITE) { 1312 if (bp->b_flags & XBF_SYNCIO) 1313 rw = WRITE_SYNC; 1314 else 1315 rw = WRITE; 1316 if (bp->b_flags & XBF_FUA) 1317 rw |= REQ_FUA; 1318 if (bp->b_flags & XBF_FLUSH) 1319 rw |= REQ_FLUSH; 1320 1321 /* 1322 * Run the write verifier callback function if it exists. If 1323 * this function fails it will mark the buffer with an error and 1324 * the IO should not be dispatched. 1325 */ 1326 if (bp->b_ops) { 1327 bp->b_ops->verify_write(bp); 1328 if (bp->b_error) { 1329 xfs_force_shutdown(bp->b_target->bt_mount, 1330 SHUTDOWN_CORRUPT_INCORE); 1331 return; 1332 } 1333 } 1334 } else if (bp->b_flags & XBF_READ_AHEAD) { 1335 rw = READA; 1336 } else { 1337 rw = READ; 1338 } 1339 1340 /* we only use the buffer cache for meta-data */ 1341 rw |= REQ_META; 1342 1343 /* 1344 * Walk all the vectors issuing IO on them. Set up the initial offset 1345 * into the buffer and the desired IO size before we start - 1346 * _xfs_buf_ioapply_vec() will modify them appropriately for each 1347 * subsequent call. 1348 */ 1349 offset = bp->b_offset; 1350 size = BBTOB(bp->b_io_length); 1351 blk_start_plug(&plug); 1352 for (i = 0; i < bp->b_map_count; i++) { 1353 xfs_buf_ioapply_map(bp, i, &offset, &size, rw); 1354 if (bp->b_error) 1355 break; 1356 if (size <= 0) 1357 break; /* all done */ 1358 } 1359 blk_finish_plug(&plug); 1360 } 1361 1362 void 1363 xfs_buf_iorequest( 1364 xfs_buf_t *bp) 1365 { 1366 trace_xfs_buf_iorequest(bp, _RET_IP_); 1367 1368 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q)); 1369 1370 if (bp->b_flags & XBF_WRITE) 1371 xfs_buf_wait_unpin(bp); 1372 xfs_buf_hold(bp); 1373 1374 /* 1375 * Set the count to 1 initially, this will stop an I/O 1376 * completion callout which happens before we have started 1377 * all the I/O from calling xfs_buf_ioend too early. 1378 */ 1379 atomic_set(&bp->b_io_remaining, 1); 1380 _xfs_buf_ioapply(bp); 1381 /* 1382 * If _xfs_buf_ioapply failed, we'll get back here with 1383 * only the reference we took above. _xfs_buf_ioend will 1384 * drop it to zero, so we'd better not queue it for later, 1385 * or we'll free it before it's done. 1386 */ 1387 _xfs_buf_ioend(bp, bp->b_error ? 0 : 1); 1388 1389 xfs_buf_rele(bp); 1390 } 1391 1392 /* 1393 * Waits for I/O to complete on the buffer supplied. It returns immediately if 1394 * no I/O is pending or there is already a pending error on the buffer, in which 1395 * case nothing will ever complete. It returns the I/O error code, if any, or 1396 * 0 if there was no error. 1397 */ 1398 int 1399 xfs_buf_iowait( 1400 xfs_buf_t *bp) 1401 { 1402 trace_xfs_buf_iowait(bp, _RET_IP_); 1403 1404 if (!bp->b_error) 1405 wait_for_completion(&bp->b_iowait); 1406 1407 trace_xfs_buf_iowait_done(bp, _RET_IP_); 1408 return bp->b_error; 1409 } 1410 1411 xfs_caddr_t 1412 xfs_buf_offset( 1413 xfs_buf_t *bp, 1414 size_t offset) 1415 { 1416 struct page *page; 1417 1418 if (bp->b_addr) 1419 return bp->b_addr + offset; 1420 1421 offset += bp->b_offset; 1422 page = bp->b_pages[offset >> PAGE_SHIFT]; 1423 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1)); 1424 } 1425 1426 /* 1427 * Move data into or out of a buffer. 1428 */ 1429 void 1430 xfs_buf_iomove( 1431 xfs_buf_t *bp, /* buffer to process */ 1432 size_t boff, /* starting buffer offset */ 1433 size_t bsize, /* length to copy */ 1434 void *data, /* data address */ 1435 xfs_buf_rw_t mode) /* read/write/zero flag */ 1436 { 1437 size_t bend; 1438 1439 bend = boff + bsize; 1440 while (boff < bend) { 1441 struct page *page; 1442 int page_index, page_offset, csize; 1443 1444 page_index = (boff + bp->b_offset) >> PAGE_SHIFT; 1445 page_offset = (boff + bp->b_offset) & ~PAGE_MASK; 1446 page = bp->b_pages[page_index]; 1447 csize = min_t(size_t, PAGE_SIZE - page_offset, 1448 BBTOB(bp->b_io_length) - boff); 1449 1450 ASSERT((csize + page_offset) <= PAGE_SIZE); 1451 1452 switch (mode) { 1453 case XBRW_ZERO: 1454 memset(page_address(page) + page_offset, 0, csize); 1455 break; 1456 case XBRW_READ: 1457 memcpy(data, page_address(page) + page_offset, csize); 1458 break; 1459 case XBRW_WRITE: 1460 memcpy(page_address(page) + page_offset, data, csize); 1461 } 1462 1463 boff += csize; 1464 data += csize; 1465 } 1466 } 1467 1468 /* 1469 * Handling of buffer targets (buftargs). 1470 */ 1471 1472 /* 1473 * Wait for any bufs with callbacks that have been submitted but have not yet 1474 * returned. These buffers will have an elevated hold count, so wait on those 1475 * while freeing all the buffers only held by the LRU. 1476 */ 1477 static enum lru_status 1478 xfs_buftarg_wait_rele( 1479 struct list_head *item, 1480 spinlock_t *lru_lock, 1481 void *arg) 1482 1483 { 1484 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru); 1485 struct list_head *dispose = arg; 1486 1487 if (atomic_read(&bp->b_hold) > 1) { 1488 /* need to wait, so skip it this pass */ 1489 trace_xfs_buf_wait_buftarg(bp, _RET_IP_); 1490 return LRU_SKIP; 1491 } 1492 if (!spin_trylock(&bp->b_lock)) 1493 return LRU_SKIP; 1494 1495 /* 1496 * clear the LRU reference count so the buffer doesn't get 1497 * ignored in xfs_buf_rele(). 1498 */ 1499 atomic_set(&bp->b_lru_ref, 0); 1500 bp->b_state |= XFS_BSTATE_DISPOSE; 1501 list_move(item, dispose); 1502 spin_unlock(&bp->b_lock); 1503 return LRU_REMOVED; 1504 } 1505 1506 void 1507 xfs_wait_buftarg( 1508 struct xfs_buftarg *btp) 1509 { 1510 LIST_HEAD(dispose); 1511 int loop = 0; 1512 1513 /* loop until there is nothing left on the lru list. */ 1514 while (list_lru_count(&btp->bt_lru)) { 1515 list_lru_walk(&btp->bt_lru, xfs_buftarg_wait_rele, 1516 &dispose, LONG_MAX); 1517 1518 while (!list_empty(&dispose)) { 1519 struct xfs_buf *bp; 1520 bp = list_first_entry(&dispose, struct xfs_buf, b_lru); 1521 list_del_init(&bp->b_lru); 1522 if (bp->b_flags & XBF_WRITE_FAIL) { 1523 xfs_alert(btp->bt_mount, 1524 "Corruption Alert: Buffer at block 0x%llx had permanent write failures!\n" 1525 "Please run xfs_repair to determine the extent of the problem.", 1526 (long long)bp->b_bn); 1527 } 1528 xfs_buf_rele(bp); 1529 } 1530 if (loop++ != 0) 1531 delay(100); 1532 } 1533 } 1534 1535 static enum lru_status 1536 xfs_buftarg_isolate( 1537 struct list_head *item, 1538 spinlock_t *lru_lock, 1539 void *arg) 1540 { 1541 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru); 1542 struct list_head *dispose = arg; 1543 1544 /* 1545 * we are inverting the lru lock/bp->b_lock here, so use a trylock. 1546 * If we fail to get the lock, just skip it. 1547 */ 1548 if (!spin_trylock(&bp->b_lock)) 1549 return LRU_SKIP; 1550 /* 1551 * Decrement the b_lru_ref count unless the value is already 1552 * zero. If the value is already zero, we need to reclaim the 1553 * buffer, otherwise it gets another trip through the LRU. 1554 */ 1555 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) { 1556 spin_unlock(&bp->b_lock); 1557 return LRU_ROTATE; 1558 } 1559 1560 bp->b_state |= XFS_BSTATE_DISPOSE; 1561 list_move(item, dispose); 1562 spin_unlock(&bp->b_lock); 1563 return LRU_REMOVED; 1564 } 1565 1566 static unsigned long 1567 xfs_buftarg_shrink_scan( 1568 struct shrinker *shrink, 1569 struct shrink_control *sc) 1570 { 1571 struct xfs_buftarg *btp = container_of(shrink, 1572 struct xfs_buftarg, bt_shrinker); 1573 LIST_HEAD(dispose); 1574 unsigned long freed; 1575 unsigned long nr_to_scan = sc->nr_to_scan; 1576 1577 freed = list_lru_walk_node(&btp->bt_lru, sc->nid, xfs_buftarg_isolate, 1578 &dispose, &nr_to_scan); 1579 1580 while (!list_empty(&dispose)) { 1581 struct xfs_buf *bp; 1582 bp = list_first_entry(&dispose, struct xfs_buf, b_lru); 1583 list_del_init(&bp->b_lru); 1584 xfs_buf_rele(bp); 1585 } 1586 1587 return freed; 1588 } 1589 1590 static unsigned long 1591 xfs_buftarg_shrink_count( 1592 struct shrinker *shrink, 1593 struct shrink_control *sc) 1594 { 1595 struct xfs_buftarg *btp = container_of(shrink, 1596 struct xfs_buftarg, bt_shrinker); 1597 return list_lru_count_node(&btp->bt_lru, sc->nid); 1598 } 1599 1600 void 1601 xfs_free_buftarg( 1602 struct xfs_mount *mp, 1603 struct xfs_buftarg *btp) 1604 { 1605 unregister_shrinker(&btp->bt_shrinker); 1606 list_lru_destroy(&btp->bt_lru); 1607 1608 if (mp->m_flags & XFS_MOUNT_BARRIER) 1609 xfs_blkdev_issue_flush(btp); 1610 1611 kmem_free(btp); 1612 } 1613 1614 int 1615 xfs_setsize_buftarg( 1616 xfs_buftarg_t *btp, 1617 unsigned int sectorsize) 1618 { 1619 /* Set up metadata sector size info */ 1620 btp->bt_meta_sectorsize = sectorsize; 1621 btp->bt_meta_sectormask = sectorsize - 1; 1622 1623 if (set_blocksize(btp->bt_bdev, sectorsize)) { 1624 char name[BDEVNAME_SIZE]; 1625 1626 bdevname(btp->bt_bdev, name); 1627 1628 xfs_warn(btp->bt_mount, 1629 "Cannot set_blocksize to %u on device %s", 1630 sectorsize, name); 1631 return EINVAL; 1632 } 1633 1634 /* Set up device logical sector size mask */ 1635 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev); 1636 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1; 1637 1638 return 0; 1639 } 1640 1641 /* 1642 * When allocating the initial buffer target we have not yet 1643 * read in the superblock, so don't know what sized sectors 1644 * are being used at this early stage. Play safe. 1645 */ 1646 STATIC int 1647 xfs_setsize_buftarg_early( 1648 xfs_buftarg_t *btp, 1649 struct block_device *bdev) 1650 { 1651 return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev)); 1652 } 1653 1654 xfs_buftarg_t * 1655 xfs_alloc_buftarg( 1656 struct xfs_mount *mp, 1657 struct block_device *bdev) 1658 { 1659 xfs_buftarg_t *btp; 1660 1661 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS); 1662 1663 btp->bt_mount = mp; 1664 btp->bt_dev = bdev->bd_dev; 1665 btp->bt_bdev = bdev; 1666 btp->bt_bdi = blk_get_backing_dev_info(bdev); 1667 if (!btp->bt_bdi) 1668 goto error; 1669 1670 if (xfs_setsize_buftarg_early(btp, bdev)) 1671 goto error; 1672 1673 if (list_lru_init(&btp->bt_lru)) 1674 goto error; 1675 1676 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count; 1677 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan; 1678 btp->bt_shrinker.seeks = DEFAULT_SEEKS; 1679 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE; 1680 register_shrinker(&btp->bt_shrinker); 1681 return btp; 1682 1683 error: 1684 kmem_free(btp); 1685 return NULL; 1686 } 1687 1688 /* 1689 * Add a buffer to the delayed write list. 1690 * 1691 * This queues a buffer for writeout if it hasn't already been. Note that 1692 * neither this routine nor the buffer list submission functions perform 1693 * any internal synchronization. It is expected that the lists are thread-local 1694 * to the callers. 1695 * 1696 * Returns true if we queued up the buffer, or false if it already had 1697 * been on the buffer list. 1698 */ 1699 bool 1700 xfs_buf_delwri_queue( 1701 struct xfs_buf *bp, 1702 struct list_head *list) 1703 { 1704 ASSERT(xfs_buf_islocked(bp)); 1705 ASSERT(!(bp->b_flags & XBF_READ)); 1706 1707 /* 1708 * If the buffer is already marked delwri it already is queued up 1709 * by someone else for imediate writeout. Just ignore it in that 1710 * case. 1711 */ 1712 if (bp->b_flags & _XBF_DELWRI_Q) { 1713 trace_xfs_buf_delwri_queued(bp, _RET_IP_); 1714 return false; 1715 } 1716 1717 trace_xfs_buf_delwri_queue(bp, _RET_IP_); 1718 1719 /* 1720 * If a buffer gets written out synchronously or marked stale while it 1721 * is on a delwri list we lazily remove it. To do this, the other party 1722 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone. 1723 * It remains referenced and on the list. In a rare corner case it 1724 * might get readded to a delwri list after the synchronous writeout, in 1725 * which case we need just need to re-add the flag here. 1726 */ 1727 bp->b_flags |= _XBF_DELWRI_Q; 1728 if (list_empty(&bp->b_list)) { 1729 atomic_inc(&bp->b_hold); 1730 list_add_tail(&bp->b_list, list); 1731 } 1732 1733 return true; 1734 } 1735 1736 /* 1737 * Compare function is more complex than it needs to be because 1738 * the return value is only 32 bits and we are doing comparisons 1739 * on 64 bit values 1740 */ 1741 static int 1742 xfs_buf_cmp( 1743 void *priv, 1744 struct list_head *a, 1745 struct list_head *b) 1746 { 1747 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list); 1748 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list); 1749 xfs_daddr_t diff; 1750 1751 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn; 1752 if (diff < 0) 1753 return -1; 1754 if (diff > 0) 1755 return 1; 1756 return 0; 1757 } 1758 1759 static int 1760 __xfs_buf_delwri_submit( 1761 struct list_head *buffer_list, 1762 struct list_head *io_list, 1763 bool wait) 1764 { 1765 struct blk_plug plug; 1766 struct xfs_buf *bp, *n; 1767 int pinned = 0; 1768 1769 list_for_each_entry_safe(bp, n, buffer_list, b_list) { 1770 if (!wait) { 1771 if (xfs_buf_ispinned(bp)) { 1772 pinned++; 1773 continue; 1774 } 1775 if (!xfs_buf_trylock(bp)) 1776 continue; 1777 } else { 1778 xfs_buf_lock(bp); 1779 } 1780 1781 /* 1782 * Someone else might have written the buffer synchronously or 1783 * marked it stale in the meantime. In that case only the 1784 * _XBF_DELWRI_Q flag got cleared, and we have to drop the 1785 * reference and remove it from the list here. 1786 */ 1787 if (!(bp->b_flags & _XBF_DELWRI_Q)) { 1788 list_del_init(&bp->b_list); 1789 xfs_buf_relse(bp); 1790 continue; 1791 } 1792 1793 list_move_tail(&bp->b_list, io_list); 1794 trace_xfs_buf_delwri_split(bp, _RET_IP_); 1795 } 1796 1797 list_sort(NULL, io_list, xfs_buf_cmp); 1798 1799 blk_start_plug(&plug); 1800 list_for_each_entry_safe(bp, n, io_list, b_list) { 1801 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC | XBF_WRITE_FAIL); 1802 bp->b_flags |= XBF_WRITE; 1803 1804 if (!wait) { 1805 bp->b_flags |= XBF_ASYNC; 1806 list_del_init(&bp->b_list); 1807 } 1808 xfs_bdstrat_cb(bp); 1809 } 1810 blk_finish_plug(&plug); 1811 1812 return pinned; 1813 } 1814 1815 /* 1816 * Write out a buffer list asynchronously. 1817 * 1818 * This will take the @buffer_list, write all non-locked and non-pinned buffers 1819 * out and not wait for I/O completion on any of the buffers. This interface 1820 * is only safely useable for callers that can track I/O completion by higher 1821 * level means, e.g. AIL pushing as the @buffer_list is consumed in this 1822 * function. 1823 */ 1824 int 1825 xfs_buf_delwri_submit_nowait( 1826 struct list_head *buffer_list) 1827 { 1828 LIST_HEAD (io_list); 1829 return __xfs_buf_delwri_submit(buffer_list, &io_list, false); 1830 } 1831 1832 /* 1833 * Write out a buffer list synchronously. 1834 * 1835 * This will take the @buffer_list, write all buffers out and wait for I/O 1836 * completion on all of the buffers. @buffer_list is consumed by the function, 1837 * so callers must have some other way of tracking buffers if they require such 1838 * functionality. 1839 */ 1840 int 1841 xfs_buf_delwri_submit( 1842 struct list_head *buffer_list) 1843 { 1844 LIST_HEAD (io_list); 1845 int error = 0, error2; 1846 struct xfs_buf *bp; 1847 1848 __xfs_buf_delwri_submit(buffer_list, &io_list, true); 1849 1850 /* Wait for IO to complete. */ 1851 while (!list_empty(&io_list)) { 1852 bp = list_first_entry(&io_list, struct xfs_buf, b_list); 1853 1854 list_del_init(&bp->b_list); 1855 error2 = xfs_buf_iowait(bp); 1856 xfs_buf_relse(bp); 1857 if (!error) 1858 error = error2; 1859 } 1860 1861 return error; 1862 } 1863 1864 int __init 1865 xfs_buf_init(void) 1866 { 1867 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf", 1868 KM_ZONE_HWALIGN, NULL); 1869 if (!xfs_buf_zone) 1870 goto out; 1871 1872 xfslogd_workqueue = alloc_workqueue("xfslogd", 1873 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1); 1874 if (!xfslogd_workqueue) 1875 goto out_free_buf_zone; 1876 1877 return 0; 1878 1879 out_free_buf_zone: 1880 kmem_zone_destroy(xfs_buf_zone); 1881 out: 1882 return -ENOMEM; 1883 } 1884 1885 void 1886 xfs_buf_terminate(void) 1887 { 1888 destroy_workqueue(xfslogd_workqueue); 1889 kmem_zone_destroy(xfs_buf_zone); 1890 } 1891