| 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 | #include <linux/list_sort.h> |
| 37 | |
| 38 | #include "xfs_sb.h" |
| 39 | #include "xfs_inum.h" |
| 40 | #include "xfs_log.h" |
| 41 | #include "xfs_ag.h" |
| 42 | #include "xfs_mount.h" |
| 43 | #include "xfs_trace.h" |
| 44 | |
| 45 | static kmem_zone_t *xfs_buf_zone; |
| 46 | STATIC int xfsbufd(void *); |
| 47 | STATIC int xfsbufd_wakeup(struct shrinker *, int, gfp_t); |
| 48 | STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int); |
| 49 | static struct shrinker xfs_buf_shake = { |
| 50 | .shrink = xfsbufd_wakeup, |
| 51 | .seeks = DEFAULT_SEEKS, |
| 52 | }; |
| 53 | |
| 54 | static struct workqueue_struct *xfslogd_workqueue; |
| 55 | struct workqueue_struct *xfsdatad_workqueue; |
| 56 | struct workqueue_struct *xfsconvertd_workqueue; |
| 57 | |
| 58 | #ifdef XFS_BUF_LOCK_TRACKING |
| 59 | # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid) |
| 60 | # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1) |
| 61 | # define XB_GET_OWNER(bp) ((bp)->b_last_holder) |
| 62 | #else |
| 63 | # define XB_SET_OWNER(bp) do { } while (0) |
| 64 | # define XB_CLEAR_OWNER(bp) do { } while (0) |
| 65 | # define XB_GET_OWNER(bp) do { } while (0) |
| 66 | #endif |
| 67 | |
| 68 | #define xb_to_gfp(flags) \ |
| 69 | ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \ |
| 70 | ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN) |
| 71 | |
| 72 | #define xb_to_km(flags) \ |
| 73 | (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP) |
| 74 | |
| 75 | #define xfs_buf_allocate(flags) \ |
| 76 | kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags)) |
| 77 | #define xfs_buf_deallocate(bp) \ |
| 78 | kmem_zone_free(xfs_buf_zone, (bp)); |
| 79 | |
| 80 | static inline int |
| 81 | xfs_buf_is_vmapped( |
| 82 | struct xfs_buf *bp) |
| 83 | { |
| 84 | /* |
| 85 | * Return true if the buffer is vmapped. |
| 86 | * |
| 87 | * The XBF_MAPPED flag is set if the buffer should be mapped, but the |
| 88 | * code is clever enough to know it doesn't have to map a single page, |
| 89 | * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1. |
| 90 | */ |
| 91 | return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1; |
| 92 | } |
| 93 | |
| 94 | static inline int |
| 95 | xfs_buf_vmap_len( |
| 96 | struct xfs_buf *bp) |
| 97 | { |
| 98 | return (bp->b_page_count * PAGE_SIZE) - bp->b_offset; |
| 99 | } |
| 100 | |
| 101 | /* |
| 102 | * Page Region interfaces. |
| 103 | * |
| 104 | * For pages in filesystems where the blocksize is smaller than the |
| 105 | * pagesize, we use the page->private field (long) to hold a bitmap |
| 106 | * of uptodate regions within the page. |
| 107 | * |
| 108 | * Each such region is "bytes per page / bits per long" bytes long. |
| 109 | * |
| 110 | * NBPPR == number-of-bytes-per-page-region |
| 111 | * BTOPR == bytes-to-page-region (rounded up) |
| 112 | * BTOPRT == bytes-to-page-region-truncated (rounded down) |
| 113 | */ |
| 114 | #if (BITS_PER_LONG == 32) |
| 115 | #define PRSHIFT (PAGE_CACHE_SHIFT - 5) /* (32 == 1<<5) */ |
| 116 | #elif (BITS_PER_LONG == 64) |
| 117 | #define PRSHIFT (PAGE_CACHE_SHIFT - 6) /* (64 == 1<<6) */ |
| 118 | #else |
| 119 | #error BITS_PER_LONG must be 32 or 64 |
| 120 | #endif |
| 121 | #define NBPPR (PAGE_CACHE_SIZE/BITS_PER_LONG) |
| 122 | #define BTOPR(b) (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT) |
| 123 | #define BTOPRT(b) (((unsigned int)(b) >> PRSHIFT)) |
| 124 | |
| 125 | STATIC unsigned long |
| 126 | page_region_mask( |
| 127 | size_t offset, |
| 128 | size_t length) |
| 129 | { |
| 130 | unsigned long mask; |
| 131 | int first, final; |
| 132 | |
| 133 | first = BTOPR(offset); |
| 134 | final = BTOPRT(offset + length - 1); |
| 135 | first = min(first, final); |
| 136 | |
| 137 | mask = ~0UL; |
| 138 | mask <<= BITS_PER_LONG - (final - first); |
| 139 | mask >>= BITS_PER_LONG - (final); |
| 140 | |
| 141 | ASSERT(offset + length <= PAGE_CACHE_SIZE); |
| 142 | ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0); |
| 143 | |
| 144 | return mask; |
| 145 | } |
| 146 | |
| 147 | STATIC void |
| 148 | set_page_region( |
| 149 | struct page *page, |
| 150 | size_t offset, |
| 151 | size_t length) |
| 152 | { |
| 153 | set_page_private(page, |
| 154 | page_private(page) | page_region_mask(offset, length)); |
| 155 | if (page_private(page) == ~0UL) |
| 156 | SetPageUptodate(page); |
| 157 | } |
| 158 | |
| 159 | STATIC int |
| 160 | test_page_region( |
| 161 | struct page *page, |
| 162 | size_t offset, |
| 163 | size_t length) |
| 164 | { |
| 165 | unsigned long mask = page_region_mask(offset, length); |
| 166 | |
| 167 | return (mask && (page_private(page) & mask) == mask); |
| 168 | } |
| 169 | |
| 170 | /* |
| 171 | * Internal xfs_buf_t object manipulation |
| 172 | */ |
| 173 | |
| 174 | STATIC void |
| 175 | _xfs_buf_initialize( |
| 176 | xfs_buf_t *bp, |
| 177 | xfs_buftarg_t *target, |
| 178 | xfs_off_t range_base, |
| 179 | size_t range_length, |
| 180 | xfs_buf_flags_t flags) |
| 181 | { |
| 182 | /* |
| 183 | * We don't want certain flags to appear in b_flags. |
| 184 | */ |
| 185 | flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD); |
| 186 | |
| 187 | memset(bp, 0, sizeof(xfs_buf_t)); |
| 188 | atomic_set(&bp->b_hold, 1); |
| 189 | init_completion(&bp->b_iowait); |
| 190 | INIT_LIST_HEAD(&bp->b_list); |
| 191 | INIT_LIST_HEAD(&bp->b_hash_list); |
| 192 | init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */ |
| 193 | XB_SET_OWNER(bp); |
| 194 | bp->b_target = target; |
| 195 | bp->b_file_offset = range_base; |
| 196 | /* |
| 197 | * Set buffer_length and count_desired to the same value initially. |
| 198 | * I/O routines should use count_desired, which will be the same in |
| 199 | * most cases but may be reset (e.g. XFS recovery). |
| 200 | */ |
| 201 | bp->b_buffer_length = bp->b_count_desired = range_length; |
| 202 | bp->b_flags = flags; |
| 203 | bp->b_bn = XFS_BUF_DADDR_NULL; |
| 204 | atomic_set(&bp->b_pin_count, 0); |
| 205 | init_waitqueue_head(&bp->b_waiters); |
| 206 | |
| 207 | XFS_STATS_INC(xb_create); |
| 208 | |
| 209 | trace_xfs_buf_init(bp, _RET_IP_); |
| 210 | } |
| 211 | |
| 212 | /* |
| 213 | * Allocate a page array capable of holding a specified number |
| 214 | * of pages, and point the page buf at it. |
| 215 | */ |
| 216 | STATIC int |
| 217 | _xfs_buf_get_pages( |
| 218 | xfs_buf_t *bp, |
| 219 | int page_count, |
| 220 | xfs_buf_flags_t flags) |
| 221 | { |
| 222 | /* Make sure that we have a page list */ |
| 223 | if (bp->b_pages == NULL) { |
| 224 | bp->b_offset = xfs_buf_poff(bp->b_file_offset); |
| 225 | bp->b_page_count = page_count; |
| 226 | if (page_count <= XB_PAGES) { |
| 227 | bp->b_pages = bp->b_page_array; |
| 228 | } else { |
| 229 | bp->b_pages = kmem_alloc(sizeof(struct page *) * |
| 230 | page_count, xb_to_km(flags)); |
| 231 | if (bp->b_pages == NULL) |
| 232 | return -ENOMEM; |
| 233 | } |
| 234 | memset(bp->b_pages, 0, sizeof(struct page *) * page_count); |
| 235 | } |
| 236 | return 0; |
| 237 | } |
| 238 | |
| 239 | /* |
| 240 | * Frees b_pages if it was allocated. |
| 241 | */ |
| 242 | STATIC void |
| 243 | _xfs_buf_free_pages( |
| 244 | xfs_buf_t *bp) |
| 245 | { |
| 246 | if (bp->b_pages != bp->b_page_array) { |
| 247 | kmem_free(bp->b_pages); |
| 248 | bp->b_pages = NULL; |
| 249 | } |
| 250 | } |
| 251 | |
| 252 | /* |
| 253 | * Releases the specified buffer. |
| 254 | * |
| 255 | * The modification state of any associated pages is left unchanged. |
| 256 | * The buffer most not be on any hash - use xfs_buf_rele instead for |
| 257 | * hashed and refcounted buffers |
| 258 | */ |
| 259 | void |
| 260 | xfs_buf_free( |
| 261 | xfs_buf_t *bp) |
| 262 | { |
| 263 | trace_xfs_buf_free(bp, _RET_IP_); |
| 264 | |
| 265 | ASSERT(list_empty(&bp->b_hash_list)); |
| 266 | |
| 267 | if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) { |
| 268 | uint i; |
| 269 | |
| 270 | if (xfs_buf_is_vmapped(bp)) |
| 271 | vm_unmap_ram(bp->b_addr - bp->b_offset, |
| 272 | bp->b_page_count); |
| 273 | |
| 274 | for (i = 0; i < bp->b_page_count; i++) { |
| 275 | struct page *page = bp->b_pages[i]; |
| 276 | |
| 277 | if (bp->b_flags & _XBF_PAGE_CACHE) |
| 278 | ASSERT(!PagePrivate(page)); |
| 279 | page_cache_release(page); |
| 280 | } |
| 281 | } |
| 282 | _xfs_buf_free_pages(bp); |
| 283 | xfs_buf_deallocate(bp); |
| 284 | } |
| 285 | |
| 286 | /* |
| 287 | * Finds all pages for buffer in question and builds it's page list. |
| 288 | */ |
| 289 | STATIC int |
| 290 | _xfs_buf_lookup_pages( |
| 291 | xfs_buf_t *bp, |
| 292 | uint flags) |
| 293 | { |
| 294 | struct address_space *mapping = bp->b_target->bt_mapping; |
| 295 | size_t blocksize = bp->b_target->bt_bsize; |
| 296 | size_t size = bp->b_count_desired; |
| 297 | size_t nbytes, offset; |
| 298 | gfp_t gfp_mask = xb_to_gfp(flags); |
| 299 | unsigned short page_count, i; |
| 300 | pgoff_t first; |
| 301 | xfs_off_t end; |
| 302 | int error; |
| 303 | |
| 304 | end = bp->b_file_offset + bp->b_buffer_length; |
| 305 | page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset); |
| 306 | |
| 307 | error = _xfs_buf_get_pages(bp, page_count, flags); |
| 308 | if (unlikely(error)) |
| 309 | return error; |
| 310 | bp->b_flags |= _XBF_PAGE_CACHE; |
| 311 | |
| 312 | offset = bp->b_offset; |
| 313 | first = bp->b_file_offset >> PAGE_CACHE_SHIFT; |
| 314 | |
| 315 | for (i = 0; i < bp->b_page_count; i++) { |
| 316 | struct page *page; |
| 317 | uint retries = 0; |
| 318 | |
| 319 | retry: |
| 320 | page = find_or_create_page(mapping, first + i, gfp_mask); |
| 321 | if (unlikely(page == NULL)) { |
| 322 | if (flags & XBF_READ_AHEAD) { |
| 323 | bp->b_page_count = i; |
| 324 | for (i = 0; i < bp->b_page_count; i++) |
| 325 | unlock_page(bp->b_pages[i]); |
| 326 | return -ENOMEM; |
| 327 | } |
| 328 | |
| 329 | /* |
| 330 | * This could deadlock. |
| 331 | * |
| 332 | * But until all the XFS lowlevel code is revamped to |
| 333 | * handle buffer allocation failures we can't do much. |
| 334 | */ |
| 335 | if (!(++retries % 100)) |
| 336 | printk(KERN_ERR |
| 337 | "XFS: possible memory allocation " |
| 338 | "deadlock in %s (mode:0x%x)\n", |
| 339 | __func__, gfp_mask); |
| 340 | |
| 341 | XFS_STATS_INC(xb_page_retries); |
| 342 | xfsbufd_wakeup(NULL, 0, gfp_mask); |
| 343 | congestion_wait(BLK_RW_ASYNC, HZ/50); |
| 344 | goto retry; |
| 345 | } |
| 346 | |
| 347 | XFS_STATS_INC(xb_page_found); |
| 348 | |
| 349 | nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset); |
| 350 | size -= nbytes; |
| 351 | |
| 352 | ASSERT(!PagePrivate(page)); |
| 353 | if (!PageUptodate(page)) { |
| 354 | page_count--; |
| 355 | if (blocksize >= PAGE_CACHE_SIZE) { |
| 356 | if (flags & XBF_READ) |
| 357 | bp->b_flags |= _XBF_PAGE_LOCKED; |
| 358 | } else if (!PagePrivate(page)) { |
| 359 | if (test_page_region(page, offset, nbytes)) |
| 360 | page_count++; |
| 361 | } |
| 362 | } |
| 363 | |
| 364 | bp->b_pages[i] = page; |
| 365 | offset = 0; |
| 366 | } |
| 367 | |
| 368 | if (!(bp->b_flags & _XBF_PAGE_LOCKED)) { |
| 369 | for (i = 0; i < bp->b_page_count; i++) |
| 370 | unlock_page(bp->b_pages[i]); |
| 371 | } |
| 372 | |
| 373 | if (page_count == bp->b_page_count) |
| 374 | bp->b_flags |= XBF_DONE; |
| 375 | |
| 376 | return error; |
| 377 | } |
| 378 | |
| 379 | /* |
| 380 | * Map buffer into kernel address-space if nessecary. |
| 381 | */ |
| 382 | STATIC int |
| 383 | _xfs_buf_map_pages( |
| 384 | xfs_buf_t *bp, |
| 385 | uint flags) |
| 386 | { |
| 387 | /* A single page buffer is always mappable */ |
| 388 | if (bp->b_page_count == 1) { |
| 389 | bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset; |
| 390 | bp->b_flags |= XBF_MAPPED; |
| 391 | } else if (flags & XBF_MAPPED) { |
| 392 | bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count, |
| 393 | -1, PAGE_KERNEL); |
| 394 | if (unlikely(bp->b_addr == NULL)) |
| 395 | return -ENOMEM; |
| 396 | bp->b_addr += bp->b_offset; |
| 397 | bp->b_flags |= XBF_MAPPED; |
| 398 | } |
| 399 | |
| 400 | return 0; |
| 401 | } |
| 402 | |
| 403 | /* |
| 404 | * Finding and Reading Buffers |
| 405 | */ |
| 406 | |
| 407 | /* |
| 408 | * Look up, and creates if absent, a lockable buffer for |
| 409 | * a given range of an inode. The buffer is returned |
| 410 | * locked. If other overlapping buffers exist, they are |
| 411 | * released before the new buffer is created and locked, |
| 412 | * which may imply that this call will block until those buffers |
| 413 | * are unlocked. No I/O is implied by this call. |
| 414 | */ |
| 415 | xfs_buf_t * |
| 416 | _xfs_buf_find( |
| 417 | xfs_buftarg_t *btp, /* block device target */ |
| 418 | xfs_off_t ioff, /* starting offset of range */ |
| 419 | size_t isize, /* length of range */ |
| 420 | xfs_buf_flags_t flags, |
| 421 | xfs_buf_t *new_bp) |
| 422 | { |
| 423 | xfs_off_t range_base; |
| 424 | size_t range_length; |
| 425 | xfs_bufhash_t *hash; |
| 426 | xfs_buf_t *bp, *n; |
| 427 | |
| 428 | range_base = (ioff << BBSHIFT); |
| 429 | range_length = (isize << BBSHIFT); |
| 430 | |
| 431 | /* Check for IOs smaller than the sector size / not sector aligned */ |
| 432 | ASSERT(!(range_length < (1 << btp->bt_sshift))); |
| 433 | ASSERT(!(range_base & (xfs_off_t)btp->bt_smask)); |
| 434 | |
| 435 | hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)]; |
| 436 | |
| 437 | spin_lock(&hash->bh_lock); |
| 438 | |
| 439 | list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) { |
| 440 | ASSERT(btp == bp->b_target); |
| 441 | if (bp->b_file_offset == range_base && |
| 442 | bp->b_buffer_length == range_length) { |
| 443 | /* |
| 444 | * If we look at something, bring it to the |
| 445 | * front of the list for next time. |
| 446 | */ |
| 447 | atomic_inc(&bp->b_hold); |
| 448 | list_move(&bp->b_hash_list, &hash->bh_list); |
| 449 | goto found; |
| 450 | } |
| 451 | } |
| 452 | |
| 453 | /* No match found */ |
| 454 | if (new_bp) { |
| 455 | _xfs_buf_initialize(new_bp, btp, range_base, |
| 456 | range_length, flags); |
| 457 | new_bp->b_hash = hash; |
| 458 | list_add(&new_bp->b_hash_list, &hash->bh_list); |
| 459 | } else { |
| 460 | XFS_STATS_INC(xb_miss_locked); |
| 461 | } |
| 462 | |
| 463 | spin_unlock(&hash->bh_lock); |
| 464 | return new_bp; |
| 465 | |
| 466 | found: |
| 467 | spin_unlock(&hash->bh_lock); |
| 468 | |
| 469 | /* Attempt to get the semaphore without sleeping, |
| 470 | * if this does not work then we need to drop the |
| 471 | * spinlock and do a hard attempt on the semaphore. |
| 472 | */ |
| 473 | if (down_trylock(&bp->b_sema)) { |
| 474 | if (!(flags & XBF_TRYLOCK)) { |
| 475 | /* wait for buffer ownership */ |
| 476 | xfs_buf_lock(bp); |
| 477 | XFS_STATS_INC(xb_get_locked_waited); |
| 478 | } else { |
| 479 | /* We asked for a trylock and failed, no need |
| 480 | * to look at file offset and length here, we |
| 481 | * know that this buffer at least overlaps our |
| 482 | * buffer and is locked, therefore our buffer |
| 483 | * either does not exist, or is this buffer. |
| 484 | */ |
| 485 | xfs_buf_rele(bp); |
| 486 | XFS_STATS_INC(xb_busy_locked); |
| 487 | return NULL; |
| 488 | } |
| 489 | } else { |
| 490 | /* trylock worked */ |
| 491 | XB_SET_OWNER(bp); |
| 492 | } |
| 493 | |
| 494 | if (bp->b_flags & XBF_STALE) { |
| 495 | ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0); |
| 496 | bp->b_flags &= XBF_MAPPED; |
| 497 | } |
| 498 | |
| 499 | trace_xfs_buf_find(bp, flags, _RET_IP_); |
| 500 | XFS_STATS_INC(xb_get_locked); |
| 501 | return bp; |
| 502 | } |
| 503 | |
| 504 | /* |
| 505 | * Assembles a buffer covering the specified range. |
| 506 | * Storage in memory for all portions of the buffer will be allocated, |
| 507 | * although backing storage may not be. |
| 508 | */ |
| 509 | xfs_buf_t * |
| 510 | xfs_buf_get( |
| 511 | xfs_buftarg_t *target,/* target for buffer */ |
| 512 | xfs_off_t ioff, /* starting offset of range */ |
| 513 | size_t isize, /* length of range */ |
| 514 | xfs_buf_flags_t flags) |
| 515 | { |
| 516 | xfs_buf_t *bp, *new_bp; |
| 517 | int error = 0, i; |
| 518 | |
| 519 | new_bp = xfs_buf_allocate(flags); |
| 520 | if (unlikely(!new_bp)) |
| 521 | return NULL; |
| 522 | |
| 523 | bp = _xfs_buf_find(target, ioff, isize, flags, new_bp); |
| 524 | if (bp == new_bp) { |
| 525 | error = _xfs_buf_lookup_pages(bp, flags); |
| 526 | if (error) |
| 527 | goto no_buffer; |
| 528 | } else { |
| 529 | xfs_buf_deallocate(new_bp); |
| 530 | if (unlikely(bp == NULL)) |
| 531 | return NULL; |
| 532 | } |
| 533 | |
| 534 | for (i = 0; i < bp->b_page_count; i++) |
| 535 | mark_page_accessed(bp->b_pages[i]); |
| 536 | |
| 537 | if (!(bp->b_flags & XBF_MAPPED)) { |
| 538 | error = _xfs_buf_map_pages(bp, flags); |
| 539 | if (unlikely(error)) { |
| 540 | printk(KERN_WARNING "%s: failed to map pages\n", |
| 541 | __func__); |
| 542 | goto no_buffer; |
| 543 | } |
| 544 | } |
| 545 | |
| 546 | XFS_STATS_INC(xb_get); |
| 547 | |
| 548 | /* |
| 549 | * Always fill in the block number now, the mapped cases can do |
| 550 | * their own overlay of this later. |
| 551 | */ |
| 552 | bp->b_bn = ioff; |
| 553 | bp->b_count_desired = bp->b_buffer_length; |
| 554 | |
| 555 | trace_xfs_buf_get(bp, flags, _RET_IP_); |
| 556 | return bp; |
| 557 | |
| 558 | no_buffer: |
| 559 | if (flags & (XBF_LOCK | XBF_TRYLOCK)) |
| 560 | xfs_buf_unlock(bp); |
| 561 | xfs_buf_rele(bp); |
| 562 | return NULL; |
| 563 | } |
| 564 | |
| 565 | STATIC int |
| 566 | _xfs_buf_read( |
| 567 | xfs_buf_t *bp, |
| 568 | xfs_buf_flags_t flags) |
| 569 | { |
| 570 | int status; |
| 571 | |
| 572 | ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE))); |
| 573 | ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL); |
| 574 | |
| 575 | bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \ |
| 576 | XBF_READ_AHEAD | _XBF_RUN_QUEUES); |
| 577 | bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \ |
| 578 | XBF_READ_AHEAD | _XBF_RUN_QUEUES); |
| 579 | |
| 580 | status = xfs_buf_iorequest(bp); |
| 581 | if (!status && !(flags & XBF_ASYNC)) |
| 582 | status = xfs_buf_iowait(bp); |
| 583 | return status; |
| 584 | } |
| 585 | |
| 586 | xfs_buf_t * |
| 587 | xfs_buf_read( |
| 588 | xfs_buftarg_t *target, |
| 589 | xfs_off_t ioff, |
| 590 | size_t isize, |
| 591 | xfs_buf_flags_t flags) |
| 592 | { |
| 593 | xfs_buf_t *bp; |
| 594 | |
| 595 | flags |= XBF_READ; |
| 596 | |
| 597 | bp = xfs_buf_get(target, ioff, isize, flags); |
| 598 | if (bp) { |
| 599 | trace_xfs_buf_read(bp, flags, _RET_IP_); |
| 600 | |
| 601 | if (!XFS_BUF_ISDONE(bp)) { |
| 602 | XFS_STATS_INC(xb_get_read); |
| 603 | _xfs_buf_read(bp, flags); |
| 604 | } else if (flags & XBF_ASYNC) { |
| 605 | /* |
| 606 | * Read ahead call which is already satisfied, |
| 607 | * drop the buffer |
| 608 | */ |
| 609 | goto no_buffer; |
| 610 | } else { |
| 611 | /* We do not want read in the flags */ |
| 612 | bp->b_flags &= ~XBF_READ; |
| 613 | } |
| 614 | } |
| 615 | |
| 616 | return bp; |
| 617 | |
| 618 | no_buffer: |
| 619 | if (flags & (XBF_LOCK | XBF_TRYLOCK)) |
| 620 | xfs_buf_unlock(bp); |
| 621 | xfs_buf_rele(bp); |
| 622 | return NULL; |
| 623 | } |
| 624 | |
| 625 | /* |
| 626 | * If we are not low on memory then do the readahead in a deadlock |
| 627 | * safe manner. |
| 628 | */ |
| 629 | void |
| 630 | xfs_buf_readahead( |
| 631 | xfs_buftarg_t *target, |
| 632 | xfs_off_t ioff, |
| 633 | size_t isize, |
| 634 | xfs_buf_flags_t flags) |
| 635 | { |
| 636 | struct backing_dev_info *bdi; |
| 637 | |
| 638 | bdi = target->bt_mapping->backing_dev_info; |
| 639 | if (bdi_read_congested(bdi)) |
| 640 | return; |
| 641 | |
| 642 | flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD); |
| 643 | xfs_buf_read(target, ioff, isize, flags); |
| 644 | } |
| 645 | |
| 646 | xfs_buf_t * |
| 647 | xfs_buf_get_empty( |
| 648 | size_t len, |
| 649 | xfs_buftarg_t *target) |
| 650 | { |
| 651 | xfs_buf_t *bp; |
| 652 | |
| 653 | bp = xfs_buf_allocate(0); |
| 654 | if (bp) |
| 655 | _xfs_buf_initialize(bp, target, 0, len, 0); |
| 656 | return bp; |
| 657 | } |
| 658 | |
| 659 | static inline struct page * |
| 660 | mem_to_page( |
| 661 | void *addr) |
| 662 | { |
| 663 | if ((!is_vmalloc_addr(addr))) { |
| 664 | return virt_to_page(addr); |
| 665 | } else { |
| 666 | return vmalloc_to_page(addr); |
| 667 | } |
| 668 | } |
| 669 | |
| 670 | int |
| 671 | xfs_buf_associate_memory( |
| 672 | xfs_buf_t *bp, |
| 673 | void *mem, |
| 674 | size_t len) |
| 675 | { |
| 676 | int rval; |
| 677 | int i = 0; |
| 678 | unsigned long pageaddr; |
| 679 | unsigned long offset; |
| 680 | size_t buflen; |
| 681 | int page_count; |
| 682 | |
| 683 | pageaddr = (unsigned long)mem & PAGE_CACHE_MASK; |
| 684 | offset = (unsigned long)mem - pageaddr; |
| 685 | buflen = PAGE_CACHE_ALIGN(len + offset); |
| 686 | page_count = buflen >> PAGE_CACHE_SHIFT; |
| 687 | |
| 688 | /* Free any previous set of page pointers */ |
| 689 | if (bp->b_pages) |
| 690 | _xfs_buf_free_pages(bp); |
| 691 | |
| 692 | bp->b_pages = NULL; |
| 693 | bp->b_addr = mem; |
| 694 | |
| 695 | rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK); |
| 696 | if (rval) |
| 697 | return rval; |
| 698 | |
| 699 | bp->b_offset = offset; |
| 700 | |
| 701 | for (i = 0; i < bp->b_page_count; i++) { |
| 702 | bp->b_pages[i] = mem_to_page((void *)pageaddr); |
| 703 | pageaddr += PAGE_CACHE_SIZE; |
| 704 | } |
| 705 | |
| 706 | bp->b_count_desired = len; |
| 707 | bp->b_buffer_length = buflen; |
| 708 | bp->b_flags |= XBF_MAPPED; |
| 709 | bp->b_flags &= ~_XBF_PAGE_LOCKED; |
| 710 | |
| 711 | return 0; |
| 712 | } |
| 713 | |
| 714 | xfs_buf_t * |
| 715 | xfs_buf_get_noaddr( |
| 716 | size_t len, |
| 717 | xfs_buftarg_t *target) |
| 718 | { |
| 719 | unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT; |
| 720 | int error, i; |
| 721 | xfs_buf_t *bp; |
| 722 | |
| 723 | bp = xfs_buf_allocate(0); |
| 724 | if (unlikely(bp == NULL)) |
| 725 | goto fail; |
| 726 | _xfs_buf_initialize(bp, target, 0, len, 0); |
| 727 | |
| 728 | error = _xfs_buf_get_pages(bp, page_count, 0); |
| 729 | if (error) |
| 730 | goto fail_free_buf; |
| 731 | |
| 732 | for (i = 0; i < page_count; i++) { |
| 733 | bp->b_pages[i] = alloc_page(GFP_KERNEL); |
| 734 | if (!bp->b_pages[i]) |
| 735 | goto fail_free_mem; |
| 736 | } |
| 737 | bp->b_flags |= _XBF_PAGES; |
| 738 | |
| 739 | error = _xfs_buf_map_pages(bp, XBF_MAPPED); |
| 740 | if (unlikely(error)) { |
| 741 | printk(KERN_WARNING "%s: failed to map pages\n", |
| 742 | __func__); |
| 743 | goto fail_free_mem; |
| 744 | } |
| 745 | |
| 746 | xfs_buf_unlock(bp); |
| 747 | |
| 748 | trace_xfs_buf_get_noaddr(bp, _RET_IP_); |
| 749 | return bp; |
| 750 | |
| 751 | fail_free_mem: |
| 752 | while (--i >= 0) |
| 753 | __free_page(bp->b_pages[i]); |
| 754 | _xfs_buf_free_pages(bp); |
| 755 | fail_free_buf: |
| 756 | xfs_buf_deallocate(bp); |
| 757 | fail: |
| 758 | return NULL; |
| 759 | } |
| 760 | |
| 761 | /* |
| 762 | * Increment reference count on buffer, to hold the buffer concurrently |
| 763 | * with another thread which may release (free) the buffer asynchronously. |
| 764 | * Must hold the buffer already to call this function. |
| 765 | */ |
| 766 | void |
| 767 | xfs_buf_hold( |
| 768 | xfs_buf_t *bp) |
| 769 | { |
| 770 | trace_xfs_buf_hold(bp, _RET_IP_); |
| 771 | atomic_inc(&bp->b_hold); |
| 772 | } |
| 773 | |
| 774 | /* |
| 775 | * Releases a hold on the specified buffer. If the |
| 776 | * the hold count is 1, calls xfs_buf_free. |
| 777 | */ |
| 778 | void |
| 779 | xfs_buf_rele( |
| 780 | xfs_buf_t *bp) |
| 781 | { |
| 782 | xfs_bufhash_t *hash = bp->b_hash; |
| 783 | |
| 784 | trace_xfs_buf_rele(bp, _RET_IP_); |
| 785 | |
| 786 | if (unlikely(!hash)) { |
| 787 | ASSERT(!bp->b_relse); |
| 788 | if (atomic_dec_and_test(&bp->b_hold)) |
| 789 | xfs_buf_free(bp); |
| 790 | return; |
| 791 | } |
| 792 | |
| 793 | ASSERT(atomic_read(&bp->b_hold) > 0); |
| 794 | if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) { |
| 795 | if (bp->b_relse) { |
| 796 | atomic_inc(&bp->b_hold); |
| 797 | spin_unlock(&hash->bh_lock); |
| 798 | (*(bp->b_relse)) (bp); |
| 799 | } else if (bp->b_flags & XBF_FS_MANAGED) { |
| 800 | spin_unlock(&hash->bh_lock); |
| 801 | } else { |
| 802 | ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q))); |
| 803 | list_del_init(&bp->b_hash_list); |
| 804 | spin_unlock(&hash->bh_lock); |
| 805 | xfs_buf_free(bp); |
| 806 | } |
| 807 | } |
| 808 | } |
| 809 | |
| 810 | |
| 811 | /* |
| 812 | * Mutual exclusion on buffers. Locking model: |
| 813 | * |
| 814 | * Buffers associated with inodes for which buffer locking |
| 815 | * is not enabled are not protected by semaphores, and are |
| 816 | * assumed to be exclusively owned by the caller. There is a |
| 817 | * spinlock in the buffer, used by the caller when concurrent |
| 818 | * access is possible. |
| 819 | */ |
| 820 | |
| 821 | /* |
| 822 | * Locks a buffer object, if it is not already locked. |
| 823 | * Note that this in no way locks the underlying pages, so it is only |
| 824 | * useful for synchronizing concurrent use of buffer objects, not for |
| 825 | * synchronizing independent access to the underlying pages. |
| 826 | */ |
| 827 | int |
| 828 | xfs_buf_cond_lock( |
| 829 | xfs_buf_t *bp) |
| 830 | { |
| 831 | int locked; |
| 832 | |
| 833 | locked = down_trylock(&bp->b_sema) == 0; |
| 834 | if (locked) |
| 835 | XB_SET_OWNER(bp); |
| 836 | |
| 837 | trace_xfs_buf_cond_lock(bp, _RET_IP_); |
| 838 | return locked ? 0 : -EBUSY; |
| 839 | } |
| 840 | |
| 841 | int |
| 842 | xfs_buf_lock_value( |
| 843 | xfs_buf_t *bp) |
| 844 | { |
| 845 | return bp->b_sema.count; |
| 846 | } |
| 847 | |
| 848 | /* |
| 849 | * Locks a buffer object. |
| 850 | * Note that this in no way locks the underlying pages, so it is only |
| 851 | * useful for synchronizing concurrent use of buffer objects, not for |
| 852 | * synchronizing independent access to the underlying pages. |
| 853 | * |
| 854 | * If we come across a stale, pinned, locked buffer, we know that we |
| 855 | * are being asked to lock a buffer that has been reallocated. Because |
| 856 | * it is pinned, we know that the log has not been pushed to disk and |
| 857 | * hence it will still be locked. Rather than sleeping until someone |
| 858 | * else pushes the log, push it ourselves before trying to get the lock. |
| 859 | */ |
| 860 | void |
| 861 | xfs_buf_lock( |
| 862 | xfs_buf_t *bp) |
| 863 | { |
| 864 | trace_xfs_buf_lock(bp, _RET_IP_); |
| 865 | |
| 866 | if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE)) |
| 867 | xfs_log_force(bp->b_mount, 0); |
| 868 | if (atomic_read(&bp->b_io_remaining)) |
| 869 | blk_run_address_space(bp->b_target->bt_mapping); |
| 870 | down(&bp->b_sema); |
| 871 | XB_SET_OWNER(bp); |
| 872 | |
| 873 | trace_xfs_buf_lock_done(bp, _RET_IP_); |
| 874 | } |
| 875 | |
| 876 | /* |
| 877 | * Releases the lock on the buffer object. |
| 878 | * If the buffer is marked delwri but is not queued, do so before we |
| 879 | * unlock the buffer as we need to set flags correctly. We also need to |
| 880 | * take a reference for the delwri queue because the unlocker is going to |
| 881 | * drop their's and they don't know we just queued it. |
| 882 | */ |
| 883 | void |
| 884 | xfs_buf_unlock( |
| 885 | xfs_buf_t *bp) |
| 886 | { |
| 887 | if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) { |
| 888 | atomic_inc(&bp->b_hold); |
| 889 | bp->b_flags |= XBF_ASYNC; |
| 890 | xfs_buf_delwri_queue(bp, 0); |
| 891 | } |
| 892 | |
| 893 | XB_CLEAR_OWNER(bp); |
| 894 | up(&bp->b_sema); |
| 895 | |
| 896 | trace_xfs_buf_unlock(bp, _RET_IP_); |
| 897 | } |
| 898 | |
| 899 | |
| 900 | /* |
| 901 | * Pinning Buffer Storage in Memory |
| 902 | * Ensure that no attempt to force a buffer to disk will succeed. |
| 903 | */ |
| 904 | void |
| 905 | xfs_buf_pin( |
| 906 | xfs_buf_t *bp) |
| 907 | { |
| 908 | trace_xfs_buf_pin(bp, _RET_IP_); |
| 909 | atomic_inc(&bp->b_pin_count); |
| 910 | } |
| 911 | |
| 912 | void |
| 913 | xfs_buf_unpin( |
| 914 | xfs_buf_t *bp) |
| 915 | { |
| 916 | trace_xfs_buf_unpin(bp, _RET_IP_); |
| 917 | |
| 918 | if (atomic_dec_and_test(&bp->b_pin_count)) |
| 919 | wake_up_all(&bp->b_waiters); |
| 920 | } |
| 921 | |
| 922 | int |
| 923 | xfs_buf_ispin( |
| 924 | xfs_buf_t *bp) |
| 925 | { |
| 926 | return atomic_read(&bp->b_pin_count); |
| 927 | } |
| 928 | |
| 929 | STATIC void |
| 930 | xfs_buf_wait_unpin( |
| 931 | xfs_buf_t *bp) |
| 932 | { |
| 933 | DECLARE_WAITQUEUE (wait, current); |
| 934 | |
| 935 | if (atomic_read(&bp->b_pin_count) == 0) |
| 936 | return; |
| 937 | |
| 938 | add_wait_queue(&bp->b_waiters, &wait); |
| 939 | for (;;) { |
| 940 | set_current_state(TASK_UNINTERRUPTIBLE); |
| 941 | if (atomic_read(&bp->b_pin_count) == 0) |
| 942 | break; |
| 943 | if (atomic_read(&bp->b_io_remaining)) |
| 944 | blk_run_address_space(bp->b_target->bt_mapping); |
| 945 | schedule(); |
| 946 | } |
| 947 | remove_wait_queue(&bp->b_waiters, &wait); |
| 948 | set_current_state(TASK_RUNNING); |
| 949 | } |
| 950 | |
| 951 | /* |
| 952 | * Buffer Utility Routines |
| 953 | */ |
| 954 | |
| 955 | STATIC void |
| 956 | xfs_buf_iodone_work( |
| 957 | struct work_struct *work) |
| 958 | { |
| 959 | xfs_buf_t *bp = |
| 960 | container_of(work, xfs_buf_t, b_iodone_work); |
| 961 | |
| 962 | /* |
| 963 | * We can get an EOPNOTSUPP to ordered writes. Here we clear the |
| 964 | * ordered flag and reissue them. Because we can't tell the higher |
| 965 | * layers directly that they should not issue ordered I/O anymore, they |
| 966 | * need to check if the _XFS_BARRIER_FAILED flag was set during I/O completion. |
| 967 | */ |
| 968 | if ((bp->b_error == EOPNOTSUPP) && |
| 969 | (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) { |
| 970 | trace_xfs_buf_ordered_retry(bp, _RET_IP_); |
| 971 | bp->b_flags &= ~XBF_ORDERED; |
| 972 | bp->b_flags |= _XFS_BARRIER_FAILED; |
| 973 | xfs_buf_iorequest(bp); |
| 974 | } else if (bp->b_iodone) |
| 975 | (*(bp->b_iodone))(bp); |
| 976 | else if (bp->b_flags & XBF_ASYNC) |
| 977 | xfs_buf_relse(bp); |
| 978 | } |
| 979 | |
| 980 | void |
| 981 | xfs_buf_ioend( |
| 982 | xfs_buf_t *bp, |
| 983 | int schedule) |
| 984 | { |
| 985 | trace_xfs_buf_iodone(bp, _RET_IP_); |
| 986 | |
| 987 | bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD); |
| 988 | if (bp->b_error == 0) |
| 989 | bp->b_flags |= XBF_DONE; |
| 990 | |
| 991 | if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) { |
| 992 | if (schedule) { |
| 993 | INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work); |
| 994 | queue_work(xfslogd_workqueue, &bp->b_iodone_work); |
| 995 | } else { |
| 996 | xfs_buf_iodone_work(&bp->b_iodone_work); |
| 997 | } |
| 998 | } else { |
| 999 | complete(&bp->b_iowait); |
| 1000 | } |
| 1001 | } |
| 1002 | |
| 1003 | void |
| 1004 | xfs_buf_ioerror( |
| 1005 | xfs_buf_t *bp, |
| 1006 | int error) |
| 1007 | { |
| 1008 | ASSERT(error >= 0 && error <= 0xffff); |
| 1009 | bp->b_error = (unsigned short)error; |
| 1010 | trace_xfs_buf_ioerror(bp, error, _RET_IP_); |
| 1011 | } |
| 1012 | |
| 1013 | int |
| 1014 | xfs_bwrite( |
| 1015 | struct xfs_mount *mp, |
| 1016 | struct xfs_buf *bp) |
| 1017 | { |
| 1018 | int error; |
| 1019 | |
| 1020 | bp->b_strat = xfs_bdstrat_cb; |
| 1021 | bp->b_mount = mp; |
| 1022 | bp->b_flags |= XBF_WRITE; |
| 1023 | bp->b_flags &= ~(XBF_ASYNC | XBF_READ); |
| 1024 | |
| 1025 | xfs_buf_delwri_dequeue(bp); |
| 1026 | xfs_buf_iostrategy(bp); |
| 1027 | |
| 1028 | error = xfs_buf_iowait(bp); |
| 1029 | if (error) |
| 1030 | xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR); |
| 1031 | xfs_buf_relse(bp); |
| 1032 | return error; |
| 1033 | } |
| 1034 | |
| 1035 | void |
| 1036 | xfs_bdwrite( |
| 1037 | void *mp, |
| 1038 | struct xfs_buf *bp) |
| 1039 | { |
| 1040 | trace_xfs_buf_bdwrite(bp, _RET_IP_); |
| 1041 | |
| 1042 | bp->b_strat = xfs_bdstrat_cb; |
| 1043 | bp->b_mount = mp; |
| 1044 | |
| 1045 | bp->b_flags &= ~XBF_READ; |
| 1046 | bp->b_flags |= (XBF_DELWRI | XBF_ASYNC); |
| 1047 | |
| 1048 | xfs_buf_delwri_queue(bp, 1); |
| 1049 | } |
| 1050 | |
| 1051 | /* |
| 1052 | * Called when we want to stop a buffer from getting written or read. |
| 1053 | * We attach the EIO error, muck with its flags, and call biodone |
| 1054 | * so that the proper iodone callbacks get called. |
| 1055 | */ |
| 1056 | STATIC int |
| 1057 | xfs_bioerror( |
| 1058 | xfs_buf_t *bp) |
| 1059 | { |
| 1060 | #ifdef XFSERRORDEBUG |
| 1061 | ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone); |
| 1062 | #endif |
| 1063 | |
| 1064 | /* |
| 1065 | * No need to wait until the buffer is unpinned, we aren't flushing it. |
| 1066 | */ |
| 1067 | XFS_BUF_ERROR(bp, EIO); |
| 1068 | |
| 1069 | /* |
| 1070 | * We're calling biodone, so delete XBF_DONE flag. |
| 1071 | */ |
| 1072 | XFS_BUF_UNREAD(bp); |
| 1073 | XFS_BUF_UNDELAYWRITE(bp); |
| 1074 | XFS_BUF_UNDONE(bp); |
| 1075 | XFS_BUF_STALE(bp); |
| 1076 | |
| 1077 | XFS_BUF_CLR_BDSTRAT_FUNC(bp); |
| 1078 | xfs_biodone(bp); |
| 1079 | |
| 1080 | return EIO; |
| 1081 | } |
| 1082 | |
| 1083 | /* |
| 1084 | * Same as xfs_bioerror, except that we are releasing the buffer |
| 1085 | * here ourselves, and avoiding the biodone call. |
| 1086 | * This is meant for userdata errors; metadata bufs come with |
| 1087 | * iodone functions attached, so that we can track down errors. |
| 1088 | */ |
| 1089 | STATIC int |
| 1090 | xfs_bioerror_relse( |
| 1091 | struct xfs_buf *bp) |
| 1092 | { |
| 1093 | int64_t fl = XFS_BUF_BFLAGS(bp); |
| 1094 | /* |
| 1095 | * No need to wait until the buffer is unpinned. |
| 1096 | * We aren't flushing it. |
| 1097 | * |
| 1098 | * chunkhold expects B_DONE to be set, whether |
| 1099 | * we actually finish the I/O or not. We don't want to |
| 1100 | * change that interface. |
| 1101 | */ |
| 1102 | XFS_BUF_UNREAD(bp); |
| 1103 | XFS_BUF_UNDELAYWRITE(bp); |
| 1104 | XFS_BUF_DONE(bp); |
| 1105 | XFS_BUF_STALE(bp); |
| 1106 | XFS_BUF_CLR_IODONE_FUNC(bp); |
| 1107 | XFS_BUF_CLR_BDSTRAT_FUNC(bp); |
| 1108 | if (!(fl & XBF_ASYNC)) { |
| 1109 | /* |
| 1110 | * Mark b_error and B_ERROR _both_. |
| 1111 | * Lot's of chunkcache code assumes that. |
| 1112 | * There's no reason to mark error for |
| 1113 | * ASYNC buffers. |
| 1114 | */ |
| 1115 | XFS_BUF_ERROR(bp, EIO); |
| 1116 | XFS_BUF_FINISH_IOWAIT(bp); |
| 1117 | } else { |
| 1118 | xfs_buf_relse(bp); |
| 1119 | } |
| 1120 | |
| 1121 | return EIO; |
| 1122 | } |
| 1123 | |
| 1124 | |
| 1125 | /* |
| 1126 | * All xfs metadata buffers except log state machine buffers |
| 1127 | * get this attached as their b_bdstrat callback function. |
| 1128 | * This is so that we can catch a buffer |
| 1129 | * after prematurely unpinning it to forcibly shutdown the filesystem. |
| 1130 | */ |
| 1131 | int |
| 1132 | xfs_bdstrat_cb( |
| 1133 | struct xfs_buf *bp) |
| 1134 | { |
| 1135 | if (XFS_FORCED_SHUTDOWN(bp->b_mount)) { |
| 1136 | trace_xfs_bdstrat_shut(bp, _RET_IP_); |
| 1137 | /* |
| 1138 | * Metadata write that didn't get logged but |
| 1139 | * written delayed anyway. These aren't associated |
| 1140 | * with a transaction, and can be ignored. |
| 1141 | */ |
| 1142 | if (!bp->b_iodone && !XFS_BUF_ISREAD(bp)) |
| 1143 | return xfs_bioerror_relse(bp); |
| 1144 | else |
| 1145 | return xfs_bioerror(bp); |
| 1146 | } |
| 1147 | |
| 1148 | xfs_buf_iorequest(bp); |
| 1149 | return 0; |
| 1150 | } |
| 1151 | |
| 1152 | /* |
| 1153 | * Wrapper around bdstrat so that we can stop data from going to disk in case |
| 1154 | * we are shutting down the filesystem. Typically user data goes thru this |
| 1155 | * path; one of the exceptions is the superblock. |
| 1156 | */ |
| 1157 | void |
| 1158 | xfsbdstrat( |
| 1159 | struct xfs_mount *mp, |
| 1160 | struct xfs_buf *bp) |
| 1161 | { |
| 1162 | if (XFS_FORCED_SHUTDOWN(mp)) { |
| 1163 | trace_xfs_bdstrat_shut(bp, _RET_IP_); |
| 1164 | xfs_bioerror_relse(bp); |
| 1165 | return; |
| 1166 | } |
| 1167 | |
| 1168 | xfs_buf_iorequest(bp); |
| 1169 | } |
| 1170 | |
| 1171 | STATIC void |
| 1172 | _xfs_buf_ioend( |
| 1173 | xfs_buf_t *bp, |
| 1174 | int schedule) |
| 1175 | { |
| 1176 | if (atomic_dec_and_test(&bp->b_io_remaining) == 1) { |
| 1177 | bp->b_flags &= ~_XBF_PAGE_LOCKED; |
| 1178 | xfs_buf_ioend(bp, schedule); |
| 1179 | } |
| 1180 | } |
| 1181 | |
| 1182 | STATIC void |
| 1183 | xfs_buf_bio_end_io( |
| 1184 | struct bio *bio, |
| 1185 | int error) |
| 1186 | { |
| 1187 | xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private; |
| 1188 | unsigned int blocksize = bp->b_target->bt_bsize; |
| 1189 | struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1; |
| 1190 | |
| 1191 | xfs_buf_ioerror(bp, -error); |
| 1192 | |
| 1193 | if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ)) |
| 1194 | invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp)); |
| 1195 | |
| 1196 | do { |
| 1197 | struct page *page = bvec->bv_page; |
| 1198 | |
| 1199 | ASSERT(!PagePrivate(page)); |
| 1200 | if (unlikely(bp->b_error)) { |
| 1201 | if (bp->b_flags & XBF_READ) |
| 1202 | ClearPageUptodate(page); |
| 1203 | } else if (blocksize >= PAGE_CACHE_SIZE) { |
| 1204 | SetPageUptodate(page); |
| 1205 | } else if (!PagePrivate(page) && |
| 1206 | (bp->b_flags & _XBF_PAGE_CACHE)) { |
| 1207 | set_page_region(page, bvec->bv_offset, bvec->bv_len); |
| 1208 | } |
| 1209 | |
| 1210 | if (--bvec >= bio->bi_io_vec) |
| 1211 | prefetchw(&bvec->bv_page->flags); |
| 1212 | |
| 1213 | if (bp->b_flags & _XBF_PAGE_LOCKED) |
| 1214 | unlock_page(page); |
| 1215 | } while (bvec >= bio->bi_io_vec); |
| 1216 | |
| 1217 | _xfs_buf_ioend(bp, 1); |
| 1218 | bio_put(bio); |
| 1219 | } |
| 1220 | |
| 1221 | STATIC void |
| 1222 | _xfs_buf_ioapply( |
| 1223 | xfs_buf_t *bp) |
| 1224 | { |
| 1225 | int rw, map_i, total_nr_pages, nr_pages; |
| 1226 | struct bio *bio; |
| 1227 | int offset = bp->b_offset; |
| 1228 | int size = bp->b_count_desired; |
| 1229 | sector_t sector = bp->b_bn; |
| 1230 | unsigned int blocksize = bp->b_target->bt_bsize; |
| 1231 | |
| 1232 | total_nr_pages = bp->b_page_count; |
| 1233 | map_i = 0; |
| 1234 | |
| 1235 | if (bp->b_flags & XBF_ORDERED) { |
| 1236 | ASSERT(!(bp->b_flags & XBF_READ)); |
| 1237 | rw = WRITE_BARRIER; |
| 1238 | } else if (bp->b_flags & XBF_LOG_BUFFER) { |
| 1239 | ASSERT(!(bp->b_flags & XBF_READ_AHEAD)); |
| 1240 | bp->b_flags &= ~_XBF_RUN_QUEUES; |
| 1241 | rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC; |
| 1242 | } else if (bp->b_flags & _XBF_RUN_QUEUES) { |
| 1243 | ASSERT(!(bp->b_flags & XBF_READ_AHEAD)); |
| 1244 | bp->b_flags &= ~_XBF_RUN_QUEUES; |
| 1245 | rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META; |
| 1246 | } else { |
| 1247 | rw = (bp->b_flags & XBF_WRITE) ? WRITE : |
| 1248 | (bp->b_flags & XBF_READ_AHEAD) ? READA : READ; |
| 1249 | } |
| 1250 | |
| 1251 | /* Special code path for reading a sub page size buffer in -- |
| 1252 | * we populate up the whole page, and hence the other metadata |
| 1253 | * in the same page. This optimization is only valid when the |
| 1254 | * filesystem block size is not smaller than the page size. |
| 1255 | */ |
| 1256 | if ((bp->b_buffer_length < PAGE_CACHE_SIZE) && |
| 1257 | ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) == |
| 1258 | (XBF_READ|_XBF_PAGE_LOCKED)) && |
| 1259 | (blocksize >= PAGE_CACHE_SIZE)) { |
| 1260 | bio = bio_alloc(GFP_NOIO, 1); |
| 1261 | |
| 1262 | bio->bi_bdev = bp->b_target->bt_bdev; |
| 1263 | bio->bi_sector = sector - (offset >> BBSHIFT); |
| 1264 | bio->bi_end_io = xfs_buf_bio_end_io; |
| 1265 | bio->bi_private = bp; |
| 1266 | |
| 1267 | bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0); |
| 1268 | size = 0; |
| 1269 | |
| 1270 | atomic_inc(&bp->b_io_remaining); |
| 1271 | |
| 1272 | goto submit_io; |
| 1273 | } |
| 1274 | |
| 1275 | next_chunk: |
| 1276 | atomic_inc(&bp->b_io_remaining); |
| 1277 | nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT); |
| 1278 | if (nr_pages > total_nr_pages) |
| 1279 | nr_pages = total_nr_pages; |
| 1280 | |
| 1281 | bio = bio_alloc(GFP_NOIO, nr_pages); |
| 1282 | bio->bi_bdev = bp->b_target->bt_bdev; |
| 1283 | bio->bi_sector = sector; |
| 1284 | bio->bi_end_io = xfs_buf_bio_end_io; |
| 1285 | bio->bi_private = bp; |
| 1286 | |
| 1287 | for (; size && nr_pages; nr_pages--, map_i++) { |
| 1288 | int rbytes, nbytes = PAGE_CACHE_SIZE - offset; |
| 1289 | |
| 1290 | if (nbytes > size) |
| 1291 | nbytes = size; |
| 1292 | |
| 1293 | rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset); |
| 1294 | if (rbytes < nbytes) |
| 1295 | break; |
| 1296 | |
| 1297 | offset = 0; |
| 1298 | sector += nbytes >> BBSHIFT; |
| 1299 | size -= nbytes; |
| 1300 | total_nr_pages--; |
| 1301 | } |
| 1302 | |
| 1303 | submit_io: |
| 1304 | if (likely(bio->bi_size)) { |
| 1305 | if (xfs_buf_is_vmapped(bp)) { |
| 1306 | flush_kernel_vmap_range(bp->b_addr, |
| 1307 | xfs_buf_vmap_len(bp)); |
| 1308 | } |
| 1309 | submit_bio(rw, bio); |
| 1310 | if (size) |
| 1311 | goto next_chunk; |
| 1312 | } else { |
| 1313 | bio_put(bio); |
| 1314 | xfs_buf_ioerror(bp, EIO); |
| 1315 | } |
| 1316 | } |
| 1317 | |
| 1318 | int |
| 1319 | xfs_buf_iorequest( |
| 1320 | xfs_buf_t *bp) |
| 1321 | { |
| 1322 | trace_xfs_buf_iorequest(bp, _RET_IP_); |
| 1323 | |
| 1324 | if (bp->b_flags & XBF_DELWRI) { |
| 1325 | xfs_buf_delwri_queue(bp, 1); |
| 1326 | return 0; |
| 1327 | } |
| 1328 | |
| 1329 | if (bp->b_flags & XBF_WRITE) { |
| 1330 | xfs_buf_wait_unpin(bp); |
| 1331 | } |
| 1332 | |
| 1333 | xfs_buf_hold(bp); |
| 1334 | |
| 1335 | /* Set the count to 1 initially, this will stop an I/O |
| 1336 | * completion callout which happens before we have started |
| 1337 | * all the I/O from calling xfs_buf_ioend too early. |
| 1338 | */ |
| 1339 | atomic_set(&bp->b_io_remaining, 1); |
| 1340 | _xfs_buf_ioapply(bp); |
| 1341 | _xfs_buf_ioend(bp, 0); |
| 1342 | |
| 1343 | xfs_buf_rele(bp); |
| 1344 | return 0; |
| 1345 | } |
| 1346 | |
| 1347 | /* |
| 1348 | * Waits for I/O to complete on the buffer supplied. |
| 1349 | * It returns immediately if no I/O is pending. |
| 1350 | * It returns the I/O error code, if any, or 0 if there was no error. |
| 1351 | */ |
| 1352 | int |
| 1353 | xfs_buf_iowait( |
| 1354 | xfs_buf_t *bp) |
| 1355 | { |
| 1356 | trace_xfs_buf_iowait(bp, _RET_IP_); |
| 1357 | |
| 1358 | if (atomic_read(&bp->b_io_remaining)) |
| 1359 | blk_run_address_space(bp->b_target->bt_mapping); |
| 1360 | wait_for_completion(&bp->b_iowait); |
| 1361 | |
| 1362 | trace_xfs_buf_iowait_done(bp, _RET_IP_); |
| 1363 | return bp->b_error; |
| 1364 | } |
| 1365 | |
| 1366 | xfs_caddr_t |
| 1367 | xfs_buf_offset( |
| 1368 | xfs_buf_t *bp, |
| 1369 | size_t offset) |
| 1370 | { |
| 1371 | struct page *page; |
| 1372 | |
| 1373 | if (bp->b_flags & XBF_MAPPED) |
| 1374 | return XFS_BUF_PTR(bp) + offset; |
| 1375 | |
| 1376 | offset += bp->b_offset; |
| 1377 | page = bp->b_pages[offset >> PAGE_CACHE_SHIFT]; |
| 1378 | return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1)); |
| 1379 | } |
| 1380 | |
| 1381 | /* |
| 1382 | * Move data into or out of a buffer. |
| 1383 | */ |
| 1384 | void |
| 1385 | xfs_buf_iomove( |
| 1386 | xfs_buf_t *bp, /* buffer to process */ |
| 1387 | size_t boff, /* starting buffer offset */ |
| 1388 | size_t bsize, /* length to copy */ |
| 1389 | void *data, /* data address */ |
| 1390 | xfs_buf_rw_t mode) /* read/write/zero flag */ |
| 1391 | { |
| 1392 | size_t bend, cpoff, csize; |
| 1393 | struct page *page; |
| 1394 | |
| 1395 | bend = boff + bsize; |
| 1396 | while (boff < bend) { |
| 1397 | page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)]; |
| 1398 | cpoff = xfs_buf_poff(boff + bp->b_offset); |
| 1399 | csize = min_t(size_t, |
| 1400 | PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff); |
| 1401 | |
| 1402 | ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE)); |
| 1403 | |
| 1404 | switch (mode) { |
| 1405 | case XBRW_ZERO: |
| 1406 | memset(page_address(page) + cpoff, 0, csize); |
| 1407 | break; |
| 1408 | case XBRW_READ: |
| 1409 | memcpy(data, page_address(page) + cpoff, csize); |
| 1410 | break; |
| 1411 | case XBRW_WRITE: |
| 1412 | memcpy(page_address(page) + cpoff, data, csize); |
| 1413 | } |
| 1414 | |
| 1415 | boff += csize; |
| 1416 | data += csize; |
| 1417 | } |
| 1418 | } |
| 1419 | |
| 1420 | /* |
| 1421 | * Handling of buffer targets (buftargs). |
| 1422 | */ |
| 1423 | |
| 1424 | /* |
| 1425 | * Wait for any bufs with callbacks that have been submitted but |
| 1426 | * have not yet returned... walk the hash list for the target. |
| 1427 | */ |
| 1428 | void |
| 1429 | xfs_wait_buftarg( |
| 1430 | xfs_buftarg_t *btp) |
| 1431 | { |
| 1432 | xfs_buf_t *bp, *n; |
| 1433 | xfs_bufhash_t *hash; |
| 1434 | uint i; |
| 1435 | |
| 1436 | for (i = 0; i < (1 << btp->bt_hashshift); i++) { |
| 1437 | hash = &btp->bt_hash[i]; |
| 1438 | again: |
| 1439 | spin_lock(&hash->bh_lock); |
| 1440 | list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) { |
| 1441 | ASSERT(btp == bp->b_target); |
| 1442 | if (!(bp->b_flags & XBF_FS_MANAGED)) { |
| 1443 | spin_unlock(&hash->bh_lock); |
| 1444 | /* |
| 1445 | * Catch superblock reference count leaks |
| 1446 | * immediately |
| 1447 | */ |
| 1448 | BUG_ON(bp->b_bn == 0); |
| 1449 | delay(100); |
| 1450 | goto again; |
| 1451 | } |
| 1452 | } |
| 1453 | spin_unlock(&hash->bh_lock); |
| 1454 | } |
| 1455 | } |
| 1456 | |
| 1457 | /* |
| 1458 | * Allocate buffer hash table for a given target. |
| 1459 | * For devices containing metadata (i.e. not the log/realtime devices) |
| 1460 | * we need to allocate a much larger hash table. |
| 1461 | */ |
| 1462 | STATIC void |
| 1463 | xfs_alloc_bufhash( |
| 1464 | xfs_buftarg_t *btp, |
| 1465 | int external) |
| 1466 | { |
| 1467 | unsigned int i; |
| 1468 | |
| 1469 | btp->bt_hashshift = external ? 3 : 8; /* 8 or 256 buckets */ |
| 1470 | btp->bt_hashmask = (1 << btp->bt_hashshift) - 1; |
| 1471 | btp->bt_hash = kmem_zalloc_large((1 << btp->bt_hashshift) * |
| 1472 | sizeof(xfs_bufhash_t)); |
| 1473 | for (i = 0; i < (1 << btp->bt_hashshift); i++) { |
| 1474 | spin_lock_init(&btp->bt_hash[i].bh_lock); |
| 1475 | INIT_LIST_HEAD(&btp->bt_hash[i].bh_list); |
| 1476 | } |
| 1477 | } |
| 1478 | |
| 1479 | STATIC void |
| 1480 | xfs_free_bufhash( |
| 1481 | xfs_buftarg_t *btp) |
| 1482 | { |
| 1483 | kmem_free_large(btp->bt_hash); |
| 1484 | btp->bt_hash = NULL; |
| 1485 | } |
| 1486 | |
| 1487 | /* |
| 1488 | * buftarg list for delwrite queue processing |
| 1489 | */ |
| 1490 | static LIST_HEAD(xfs_buftarg_list); |
| 1491 | static DEFINE_SPINLOCK(xfs_buftarg_lock); |
| 1492 | |
| 1493 | STATIC void |
| 1494 | xfs_register_buftarg( |
| 1495 | xfs_buftarg_t *btp) |
| 1496 | { |
| 1497 | spin_lock(&xfs_buftarg_lock); |
| 1498 | list_add(&btp->bt_list, &xfs_buftarg_list); |
| 1499 | spin_unlock(&xfs_buftarg_lock); |
| 1500 | } |
| 1501 | |
| 1502 | STATIC void |
| 1503 | xfs_unregister_buftarg( |
| 1504 | xfs_buftarg_t *btp) |
| 1505 | { |
| 1506 | spin_lock(&xfs_buftarg_lock); |
| 1507 | list_del(&btp->bt_list); |
| 1508 | spin_unlock(&xfs_buftarg_lock); |
| 1509 | } |
| 1510 | |
| 1511 | void |
| 1512 | xfs_free_buftarg( |
| 1513 | struct xfs_mount *mp, |
| 1514 | struct xfs_buftarg *btp) |
| 1515 | { |
| 1516 | xfs_flush_buftarg(btp, 1); |
| 1517 | if (mp->m_flags & XFS_MOUNT_BARRIER) |
| 1518 | xfs_blkdev_issue_flush(btp); |
| 1519 | xfs_free_bufhash(btp); |
| 1520 | iput(btp->bt_mapping->host); |
| 1521 | |
| 1522 | /* Unregister the buftarg first so that we don't get a |
| 1523 | * wakeup finding a non-existent task |
| 1524 | */ |
| 1525 | xfs_unregister_buftarg(btp); |
| 1526 | kthread_stop(btp->bt_task); |
| 1527 | |
| 1528 | kmem_free(btp); |
| 1529 | } |
| 1530 | |
| 1531 | STATIC int |
| 1532 | xfs_setsize_buftarg_flags( |
| 1533 | xfs_buftarg_t *btp, |
| 1534 | unsigned int blocksize, |
| 1535 | unsigned int sectorsize, |
| 1536 | int verbose) |
| 1537 | { |
| 1538 | btp->bt_bsize = blocksize; |
| 1539 | btp->bt_sshift = ffs(sectorsize) - 1; |
| 1540 | btp->bt_smask = sectorsize - 1; |
| 1541 | |
| 1542 | if (set_blocksize(btp->bt_bdev, sectorsize)) { |
| 1543 | printk(KERN_WARNING |
| 1544 | "XFS: Cannot set_blocksize to %u on device %s\n", |
| 1545 | sectorsize, XFS_BUFTARG_NAME(btp)); |
| 1546 | return EINVAL; |
| 1547 | } |
| 1548 | |
| 1549 | if (verbose && |
| 1550 | (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) { |
| 1551 | printk(KERN_WARNING |
| 1552 | "XFS: %u byte sectors in use on device %s. " |
| 1553 | "This is suboptimal; %u or greater is ideal.\n", |
| 1554 | sectorsize, XFS_BUFTARG_NAME(btp), |
| 1555 | (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG); |
| 1556 | } |
| 1557 | |
| 1558 | return 0; |
| 1559 | } |
| 1560 | |
| 1561 | /* |
| 1562 | * When allocating the initial buffer target we have not yet |
| 1563 | * read in the superblock, so don't know what sized sectors |
| 1564 | * are being used is at this early stage. Play safe. |
| 1565 | */ |
| 1566 | STATIC int |
| 1567 | xfs_setsize_buftarg_early( |
| 1568 | xfs_buftarg_t *btp, |
| 1569 | struct block_device *bdev) |
| 1570 | { |
| 1571 | return xfs_setsize_buftarg_flags(btp, |
| 1572 | PAGE_CACHE_SIZE, bdev_logical_block_size(bdev), 0); |
| 1573 | } |
| 1574 | |
| 1575 | int |
| 1576 | xfs_setsize_buftarg( |
| 1577 | xfs_buftarg_t *btp, |
| 1578 | unsigned int blocksize, |
| 1579 | unsigned int sectorsize) |
| 1580 | { |
| 1581 | return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1); |
| 1582 | } |
| 1583 | |
| 1584 | STATIC int |
| 1585 | xfs_mapping_buftarg( |
| 1586 | xfs_buftarg_t *btp, |
| 1587 | struct block_device *bdev) |
| 1588 | { |
| 1589 | struct backing_dev_info *bdi; |
| 1590 | struct inode *inode; |
| 1591 | struct address_space *mapping; |
| 1592 | static const struct address_space_operations mapping_aops = { |
| 1593 | .sync_page = block_sync_page, |
| 1594 | .migratepage = fail_migrate_page, |
| 1595 | }; |
| 1596 | |
| 1597 | inode = new_inode(bdev->bd_inode->i_sb); |
| 1598 | if (!inode) { |
| 1599 | printk(KERN_WARNING |
| 1600 | "XFS: Cannot allocate mapping inode for device %s\n", |
| 1601 | XFS_BUFTARG_NAME(btp)); |
| 1602 | return ENOMEM; |
| 1603 | } |
| 1604 | inode->i_mode = S_IFBLK; |
| 1605 | inode->i_bdev = bdev; |
| 1606 | inode->i_rdev = bdev->bd_dev; |
| 1607 | bdi = blk_get_backing_dev_info(bdev); |
| 1608 | if (!bdi) |
| 1609 | bdi = &default_backing_dev_info; |
| 1610 | mapping = &inode->i_data; |
| 1611 | mapping->a_ops = &mapping_aops; |
| 1612 | mapping->backing_dev_info = bdi; |
| 1613 | mapping_set_gfp_mask(mapping, GFP_NOFS); |
| 1614 | btp->bt_mapping = mapping; |
| 1615 | return 0; |
| 1616 | } |
| 1617 | |
| 1618 | STATIC int |
| 1619 | xfs_alloc_delwrite_queue( |
| 1620 | xfs_buftarg_t *btp, |
| 1621 | const char *fsname) |
| 1622 | { |
| 1623 | int error = 0; |
| 1624 | |
| 1625 | INIT_LIST_HEAD(&btp->bt_list); |
| 1626 | INIT_LIST_HEAD(&btp->bt_delwrite_queue); |
| 1627 | spin_lock_init(&btp->bt_delwrite_lock); |
| 1628 | btp->bt_flags = 0; |
| 1629 | btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname); |
| 1630 | if (IS_ERR(btp->bt_task)) { |
| 1631 | error = PTR_ERR(btp->bt_task); |
| 1632 | goto out_error; |
| 1633 | } |
| 1634 | xfs_register_buftarg(btp); |
| 1635 | out_error: |
| 1636 | return error; |
| 1637 | } |
| 1638 | |
| 1639 | xfs_buftarg_t * |
| 1640 | xfs_alloc_buftarg( |
| 1641 | struct block_device *bdev, |
| 1642 | int external, |
| 1643 | const char *fsname) |
| 1644 | { |
| 1645 | xfs_buftarg_t *btp; |
| 1646 | |
| 1647 | btp = kmem_zalloc(sizeof(*btp), KM_SLEEP); |
| 1648 | |
| 1649 | btp->bt_dev = bdev->bd_dev; |
| 1650 | btp->bt_bdev = bdev; |
| 1651 | if (xfs_setsize_buftarg_early(btp, bdev)) |
| 1652 | goto error; |
| 1653 | if (xfs_mapping_buftarg(btp, bdev)) |
| 1654 | goto error; |
| 1655 | if (xfs_alloc_delwrite_queue(btp, fsname)) |
| 1656 | goto error; |
| 1657 | xfs_alloc_bufhash(btp, external); |
| 1658 | return btp; |
| 1659 | |
| 1660 | error: |
| 1661 | kmem_free(btp); |
| 1662 | return NULL; |
| 1663 | } |
| 1664 | |
| 1665 | |
| 1666 | /* |
| 1667 | * Delayed write buffer handling |
| 1668 | */ |
| 1669 | STATIC void |
| 1670 | xfs_buf_delwri_queue( |
| 1671 | xfs_buf_t *bp, |
| 1672 | int unlock) |
| 1673 | { |
| 1674 | struct list_head *dwq = &bp->b_target->bt_delwrite_queue; |
| 1675 | spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock; |
| 1676 | |
| 1677 | trace_xfs_buf_delwri_queue(bp, _RET_IP_); |
| 1678 | |
| 1679 | ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC)); |
| 1680 | |
| 1681 | spin_lock(dwlk); |
| 1682 | /* If already in the queue, dequeue and place at tail */ |
| 1683 | if (!list_empty(&bp->b_list)) { |
| 1684 | ASSERT(bp->b_flags & _XBF_DELWRI_Q); |
| 1685 | if (unlock) |
| 1686 | atomic_dec(&bp->b_hold); |
| 1687 | list_del(&bp->b_list); |
| 1688 | } |
| 1689 | |
| 1690 | if (list_empty(dwq)) { |
| 1691 | /* start xfsbufd as it is about to have something to do */ |
| 1692 | wake_up_process(bp->b_target->bt_task); |
| 1693 | } |
| 1694 | |
| 1695 | bp->b_flags |= _XBF_DELWRI_Q; |
| 1696 | list_add_tail(&bp->b_list, dwq); |
| 1697 | bp->b_queuetime = jiffies; |
| 1698 | spin_unlock(dwlk); |
| 1699 | |
| 1700 | if (unlock) |
| 1701 | xfs_buf_unlock(bp); |
| 1702 | } |
| 1703 | |
| 1704 | void |
| 1705 | xfs_buf_delwri_dequeue( |
| 1706 | xfs_buf_t *bp) |
| 1707 | { |
| 1708 | spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock; |
| 1709 | int dequeued = 0; |
| 1710 | |
| 1711 | spin_lock(dwlk); |
| 1712 | if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) { |
| 1713 | ASSERT(bp->b_flags & _XBF_DELWRI_Q); |
| 1714 | list_del_init(&bp->b_list); |
| 1715 | dequeued = 1; |
| 1716 | } |
| 1717 | bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q); |
| 1718 | spin_unlock(dwlk); |
| 1719 | |
| 1720 | if (dequeued) |
| 1721 | xfs_buf_rele(bp); |
| 1722 | |
| 1723 | trace_xfs_buf_delwri_dequeue(bp, _RET_IP_); |
| 1724 | } |
| 1725 | |
| 1726 | /* |
| 1727 | * If a delwri buffer needs to be pushed before it has aged out, then promote |
| 1728 | * it to the head of the delwri queue so that it will be flushed on the next |
| 1729 | * xfsbufd run. We do this by resetting the queuetime of the buffer to be older |
| 1730 | * than the age currently needed to flush the buffer. Hence the next time the |
| 1731 | * xfsbufd sees it is guaranteed to be considered old enough to flush. |
| 1732 | */ |
| 1733 | void |
| 1734 | xfs_buf_delwri_promote( |
| 1735 | struct xfs_buf *bp) |
| 1736 | { |
| 1737 | struct xfs_buftarg *btp = bp->b_target; |
| 1738 | long age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1; |
| 1739 | |
| 1740 | ASSERT(bp->b_flags & XBF_DELWRI); |
| 1741 | ASSERT(bp->b_flags & _XBF_DELWRI_Q); |
| 1742 | |
| 1743 | /* |
| 1744 | * Check the buffer age before locking the delayed write queue as we |
| 1745 | * don't need to promote buffers that are already past the flush age. |
| 1746 | */ |
| 1747 | if (bp->b_queuetime < jiffies - age) |
| 1748 | return; |
| 1749 | bp->b_queuetime = jiffies - age; |
| 1750 | spin_lock(&btp->bt_delwrite_lock); |
| 1751 | list_move(&bp->b_list, &btp->bt_delwrite_queue); |
| 1752 | spin_unlock(&btp->bt_delwrite_lock); |
| 1753 | } |
| 1754 | |
| 1755 | STATIC void |
| 1756 | xfs_buf_runall_queues( |
| 1757 | struct workqueue_struct *queue) |
| 1758 | { |
| 1759 | flush_workqueue(queue); |
| 1760 | } |
| 1761 | |
| 1762 | STATIC int |
| 1763 | xfsbufd_wakeup( |
| 1764 | struct shrinker *shrink, |
| 1765 | int priority, |
| 1766 | gfp_t mask) |
| 1767 | { |
| 1768 | xfs_buftarg_t *btp; |
| 1769 | |
| 1770 | spin_lock(&xfs_buftarg_lock); |
| 1771 | list_for_each_entry(btp, &xfs_buftarg_list, bt_list) { |
| 1772 | if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags)) |
| 1773 | continue; |
| 1774 | if (list_empty(&btp->bt_delwrite_queue)) |
| 1775 | continue; |
| 1776 | set_bit(XBT_FORCE_FLUSH, &btp->bt_flags); |
| 1777 | wake_up_process(btp->bt_task); |
| 1778 | } |
| 1779 | spin_unlock(&xfs_buftarg_lock); |
| 1780 | return 0; |
| 1781 | } |
| 1782 | |
| 1783 | /* |
| 1784 | * Move as many buffers as specified to the supplied list |
| 1785 | * idicating if we skipped any buffers to prevent deadlocks. |
| 1786 | */ |
| 1787 | STATIC int |
| 1788 | xfs_buf_delwri_split( |
| 1789 | xfs_buftarg_t *target, |
| 1790 | struct list_head *list, |
| 1791 | unsigned long age) |
| 1792 | { |
| 1793 | xfs_buf_t *bp, *n; |
| 1794 | struct list_head *dwq = &target->bt_delwrite_queue; |
| 1795 | spinlock_t *dwlk = &target->bt_delwrite_lock; |
| 1796 | int skipped = 0; |
| 1797 | int force; |
| 1798 | |
| 1799 | force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags); |
| 1800 | INIT_LIST_HEAD(list); |
| 1801 | spin_lock(dwlk); |
| 1802 | list_for_each_entry_safe(bp, n, dwq, b_list) { |
| 1803 | trace_xfs_buf_delwri_split(bp, _RET_IP_); |
| 1804 | ASSERT(bp->b_flags & XBF_DELWRI); |
| 1805 | |
| 1806 | if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) { |
| 1807 | if (!force && |
| 1808 | time_before(jiffies, bp->b_queuetime + age)) { |
| 1809 | xfs_buf_unlock(bp); |
| 1810 | break; |
| 1811 | } |
| 1812 | |
| 1813 | bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q| |
| 1814 | _XBF_RUN_QUEUES); |
| 1815 | bp->b_flags |= XBF_WRITE; |
| 1816 | list_move_tail(&bp->b_list, list); |
| 1817 | } else |
| 1818 | skipped++; |
| 1819 | } |
| 1820 | spin_unlock(dwlk); |
| 1821 | |
| 1822 | return skipped; |
| 1823 | |
| 1824 | } |
| 1825 | |
| 1826 | /* |
| 1827 | * Compare function is more complex than it needs to be because |
| 1828 | * the return value is only 32 bits and we are doing comparisons |
| 1829 | * on 64 bit values |
| 1830 | */ |
| 1831 | static int |
| 1832 | xfs_buf_cmp( |
| 1833 | void *priv, |
| 1834 | struct list_head *a, |
| 1835 | struct list_head *b) |
| 1836 | { |
| 1837 | struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list); |
| 1838 | struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list); |
| 1839 | xfs_daddr_t diff; |
| 1840 | |
| 1841 | diff = ap->b_bn - bp->b_bn; |
| 1842 | if (diff < 0) |
| 1843 | return -1; |
| 1844 | if (diff > 0) |
| 1845 | return 1; |
| 1846 | return 0; |
| 1847 | } |
| 1848 | |
| 1849 | void |
| 1850 | xfs_buf_delwri_sort( |
| 1851 | xfs_buftarg_t *target, |
| 1852 | struct list_head *list) |
| 1853 | { |
| 1854 | list_sort(NULL, list, xfs_buf_cmp); |
| 1855 | } |
| 1856 | |
| 1857 | STATIC int |
| 1858 | xfsbufd( |
| 1859 | void *data) |
| 1860 | { |
| 1861 | xfs_buftarg_t *target = (xfs_buftarg_t *)data; |
| 1862 | |
| 1863 | current->flags |= PF_MEMALLOC; |
| 1864 | |
| 1865 | set_freezable(); |
| 1866 | |
| 1867 | do { |
| 1868 | long age = xfs_buf_age_centisecs * msecs_to_jiffies(10); |
| 1869 | long tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10); |
| 1870 | int count = 0; |
| 1871 | struct list_head tmp; |
| 1872 | |
| 1873 | if (unlikely(freezing(current))) { |
| 1874 | set_bit(XBT_FORCE_SLEEP, &target->bt_flags); |
| 1875 | refrigerator(); |
| 1876 | } else { |
| 1877 | clear_bit(XBT_FORCE_SLEEP, &target->bt_flags); |
| 1878 | } |
| 1879 | |
| 1880 | /* sleep for a long time if there is nothing to do. */ |
| 1881 | if (list_empty(&target->bt_delwrite_queue)) |
| 1882 | tout = MAX_SCHEDULE_TIMEOUT; |
| 1883 | schedule_timeout_interruptible(tout); |
| 1884 | |
| 1885 | xfs_buf_delwri_split(target, &tmp, age); |
| 1886 | list_sort(NULL, &tmp, xfs_buf_cmp); |
| 1887 | while (!list_empty(&tmp)) { |
| 1888 | struct xfs_buf *bp; |
| 1889 | bp = list_first_entry(&tmp, struct xfs_buf, b_list); |
| 1890 | list_del_init(&bp->b_list); |
| 1891 | xfs_buf_iostrategy(bp); |
| 1892 | count++; |
| 1893 | } |
| 1894 | if (count) |
| 1895 | blk_run_address_space(target->bt_mapping); |
| 1896 | |
| 1897 | } while (!kthread_should_stop()); |
| 1898 | |
| 1899 | return 0; |
| 1900 | } |
| 1901 | |
| 1902 | /* |
| 1903 | * Go through all incore buffers, and release buffers if they belong to |
| 1904 | * the given device. This is used in filesystem error handling to |
| 1905 | * preserve the consistency of its metadata. |
| 1906 | */ |
| 1907 | int |
| 1908 | xfs_flush_buftarg( |
| 1909 | xfs_buftarg_t *target, |
| 1910 | int wait) |
| 1911 | { |
| 1912 | xfs_buf_t *bp; |
| 1913 | int pincount = 0; |
| 1914 | LIST_HEAD(tmp_list); |
| 1915 | LIST_HEAD(wait_list); |
| 1916 | |
| 1917 | xfs_buf_runall_queues(xfsconvertd_workqueue); |
| 1918 | xfs_buf_runall_queues(xfsdatad_workqueue); |
| 1919 | xfs_buf_runall_queues(xfslogd_workqueue); |
| 1920 | |
| 1921 | set_bit(XBT_FORCE_FLUSH, &target->bt_flags); |
| 1922 | pincount = xfs_buf_delwri_split(target, &tmp_list, 0); |
| 1923 | |
| 1924 | /* |
| 1925 | * Dropped the delayed write list lock, now walk the temporary list. |
| 1926 | * All I/O is issued async and then if we need to wait for completion |
| 1927 | * we do that after issuing all the IO. |
| 1928 | */ |
| 1929 | list_sort(NULL, &tmp_list, xfs_buf_cmp); |
| 1930 | while (!list_empty(&tmp_list)) { |
| 1931 | bp = list_first_entry(&tmp_list, struct xfs_buf, b_list); |
| 1932 | ASSERT(target == bp->b_target); |
| 1933 | list_del_init(&bp->b_list); |
| 1934 | if (wait) { |
| 1935 | bp->b_flags &= ~XBF_ASYNC; |
| 1936 | list_add(&bp->b_list, &wait_list); |
| 1937 | } |
| 1938 | xfs_buf_iostrategy(bp); |
| 1939 | } |
| 1940 | |
| 1941 | if (wait) { |
| 1942 | /* Expedite and wait for IO to complete. */ |
| 1943 | blk_run_address_space(target->bt_mapping); |
| 1944 | while (!list_empty(&wait_list)) { |
| 1945 | bp = list_first_entry(&wait_list, struct xfs_buf, b_list); |
| 1946 | |
| 1947 | list_del_init(&bp->b_list); |
| 1948 | xfs_iowait(bp); |
| 1949 | xfs_buf_relse(bp); |
| 1950 | } |
| 1951 | } |
| 1952 | |
| 1953 | return pincount; |
| 1954 | } |
| 1955 | |
| 1956 | int __init |
| 1957 | xfs_buf_init(void) |
| 1958 | { |
| 1959 | xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf", |
| 1960 | KM_ZONE_HWALIGN, NULL); |
| 1961 | if (!xfs_buf_zone) |
| 1962 | goto out; |
| 1963 | |
| 1964 | xfslogd_workqueue = create_workqueue("xfslogd"); |
| 1965 | if (!xfslogd_workqueue) |
| 1966 | goto out_free_buf_zone; |
| 1967 | |
| 1968 | xfsdatad_workqueue = create_workqueue("xfsdatad"); |
| 1969 | if (!xfsdatad_workqueue) |
| 1970 | goto out_destroy_xfslogd_workqueue; |
| 1971 | |
| 1972 | xfsconvertd_workqueue = create_workqueue("xfsconvertd"); |
| 1973 | if (!xfsconvertd_workqueue) |
| 1974 | goto out_destroy_xfsdatad_workqueue; |
| 1975 | |
| 1976 | register_shrinker(&xfs_buf_shake); |
| 1977 | return 0; |
| 1978 | |
| 1979 | out_destroy_xfsdatad_workqueue: |
| 1980 | destroy_workqueue(xfsdatad_workqueue); |
| 1981 | out_destroy_xfslogd_workqueue: |
| 1982 | destroy_workqueue(xfslogd_workqueue); |
| 1983 | out_free_buf_zone: |
| 1984 | kmem_zone_destroy(xfs_buf_zone); |
| 1985 | out: |
| 1986 | return -ENOMEM; |
| 1987 | } |
| 1988 | |
| 1989 | void |
| 1990 | xfs_buf_terminate(void) |
| 1991 | { |
| 1992 | unregister_shrinker(&xfs_buf_shake); |
| 1993 | destroy_workqueue(xfsconvertd_workqueue); |
| 1994 | destroy_workqueue(xfsdatad_workqueue); |
| 1995 | destroy_workqueue(xfslogd_workqueue); |
| 1996 | kmem_zone_destroy(xfs_buf_zone); |
| 1997 | } |
| 1998 | |
| 1999 | #ifdef CONFIG_KDB_MODULES |
| 2000 | struct list_head * |
| 2001 | xfs_get_buftarg_list(void) |
| 2002 | { |
| 2003 | return &xfs_buftarg_list; |
| 2004 | } |
| 2005 | #endif |