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