1 /* 2 * Copyright (c) 2000-2005 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 "xfs_fs.h" 20 #include "xfs_format.h" 21 #include "xfs_log_format.h" 22 #include "xfs_trans_resv.h" 23 #include "xfs_bit.h" 24 #include "xfs_sb.h" 25 #include "xfs_mount.h" 26 #include "xfs_trans.h" 27 #include "xfs_buf_item.h" 28 #include "xfs_trans_priv.h" 29 #include "xfs_error.h" 30 #include "xfs_trace.h" 31 #include "xfs_log.h" 32 #include "xfs_inode.h" 33 34 35 kmem_zone_t *xfs_buf_item_zone; 36 37 static inline struct xfs_buf_log_item *BUF_ITEM(struct xfs_log_item *lip) 38 { 39 return container_of(lip, struct xfs_buf_log_item, bli_item); 40 } 41 42 STATIC void xfs_buf_do_callbacks(struct xfs_buf *bp); 43 44 static inline int 45 xfs_buf_log_format_size( 46 struct xfs_buf_log_format *blfp) 47 { 48 return offsetof(struct xfs_buf_log_format, blf_data_map) + 49 (blfp->blf_map_size * sizeof(blfp->blf_data_map[0])); 50 } 51 52 /* 53 * This returns the number of log iovecs needed to log the 54 * given buf log item. 55 * 56 * It calculates this as 1 iovec for the buf log format structure 57 * and 1 for each stretch of non-contiguous chunks to be logged. 58 * Contiguous chunks are logged in a single iovec. 59 * 60 * If the XFS_BLI_STALE flag has been set, then log nothing. 61 */ 62 STATIC void 63 xfs_buf_item_size_segment( 64 struct xfs_buf_log_item *bip, 65 struct xfs_buf_log_format *blfp, 66 int *nvecs, 67 int *nbytes) 68 { 69 struct xfs_buf *bp = bip->bli_buf; 70 int next_bit; 71 int last_bit; 72 73 last_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size, 0); 74 if (last_bit == -1) 75 return; 76 77 /* 78 * initial count for a dirty buffer is 2 vectors - the format structure 79 * and the first dirty region. 80 */ 81 *nvecs += 2; 82 *nbytes += xfs_buf_log_format_size(blfp) + XFS_BLF_CHUNK; 83 84 while (last_bit != -1) { 85 /* 86 * This takes the bit number to start looking from and 87 * returns the next set bit from there. It returns -1 88 * if there are no more bits set or the start bit is 89 * beyond the end of the bitmap. 90 */ 91 next_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size, 92 last_bit + 1); 93 /* 94 * If we run out of bits, leave the loop, 95 * else if we find a new set of bits bump the number of vecs, 96 * else keep scanning the current set of bits. 97 */ 98 if (next_bit == -1) { 99 break; 100 } else if (next_bit != last_bit + 1) { 101 last_bit = next_bit; 102 (*nvecs)++; 103 } else if (xfs_buf_offset(bp, next_bit * XFS_BLF_CHUNK) != 104 (xfs_buf_offset(bp, last_bit * XFS_BLF_CHUNK) + 105 XFS_BLF_CHUNK)) { 106 last_bit = next_bit; 107 (*nvecs)++; 108 } else { 109 last_bit++; 110 } 111 *nbytes += XFS_BLF_CHUNK; 112 } 113 } 114 115 /* 116 * This returns the number of log iovecs needed to log the given buf log item. 117 * 118 * It calculates this as 1 iovec for the buf log format structure and 1 for each 119 * stretch of non-contiguous chunks to be logged. Contiguous chunks are logged 120 * in a single iovec. 121 * 122 * Discontiguous buffers need a format structure per region that that is being 123 * logged. This makes the changes in the buffer appear to log recovery as though 124 * they came from separate buffers, just like would occur if multiple buffers 125 * were used instead of a single discontiguous buffer. This enables 126 * discontiguous buffers to be in-memory constructs, completely transparent to 127 * what ends up on disk. 128 * 129 * If the XFS_BLI_STALE flag has been set, then log nothing but the buf log 130 * format structures. 131 */ 132 STATIC void 133 xfs_buf_item_size( 134 struct xfs_log_item *lip, 135 int *nvecs, 136 int *nbytes) 137 { 138 struct xfs_buf_log_item *bip = BUF_ITEM(lip); 139 int i; 140 141 ASSERT(atomic_read(&bip->bli_refcount) > 0); 142 if (bip->bli_flags & XFS_BLI_STALE) { 143 /* 144 * The buffer is stale, so all we need to log 145 * is the buf log format structure with the 146 * cancel flag in it. 147 */ 148 trace_xfs_buf_item_size_stale(bip); 149 ASSERT(bip->__bli_format.blf_flags & XFS_BLF_CANCEL); 150 *nvecs += bip->bli_format_count; 151 for (i = 0; i < bip->bli_format_count; i++) { 152 *nbytes += xfs_buf_log_format_size(&bip->bli_formats[i]); 153 } 154 return; 155 } 156 157 ASSERT(bip->bli_flags & XFS_BLI_LOGGED); 158 159 if (bip->bli_flags & XFS_BLI_ORDERED) { 160 /* 161 * The buffer has been logged just to order it. 162 * It is not being included in the transaction 163 * commit, so no vectors are used at all. 164 */ 165 trace_xfs_buf_item_size_ordered(bip); 166 *nvecs = XFS_LOG_VEC_ORDERED; 167 return; 168 } 169 170 /* 171 * the vector count is based on the number of buffer vectors we have 172 * dirty bits in. This will only be greater than one when we have a 173 * compound buffer with more than one segment dirty. Hence for compound 174 * buffers we need to track which segment the dirty bits correspond to, 175 * and when we move from one segment to the next increment the vector 176 * count for the extra buf log format structure that will need to be 177 * written. 178 */ 179 for (i = 0; i < bip->bli_format_count; i++) { 180 xfs_buf_item_size_segment(bip, &bip->bli_formats[i], 181 nvecs, nbytes); 182 } 183 trace_xfs_buf_item_size(bip); 184 } 185 186 static inline void 187 xfs_buf_item_copy_iovec( 188 struct xfs_log_vec *lv, 189 struct xfs_log_iovec **vecp, 190 struct xfs_buf *bp, 191 uint offset, 192 int first_bit, 193 uint nbits) 194 { 195 offset += first_bit * XFS_BLF_CHUNK; 196 xlog_copy_iovec(lv, vecp, XLOG_REG_TYPE_BCHUNK, 197 xfs_buf_offset(bp, offset), 198 nbits * XFS_BLF_CHUNK); 199 } 200 201 static inline bool 202 xfs_buf_item_straddle( 203 struct xfs_buf *bp, 204 uint offset, 205 int next_bit, 206 int last_bit) 207 { 208 return xfs_buf_offset(bp, offset + (next_bit << XFS_BLF_SHIFT)) != 209 (xfs_buf_offset(bp, offset + (last_bit << XFS_BLF_SHIFT)) + 210 XFS_BLF_CHUNK); 211 } 212 213 static void 214 xfs_buf_item_format_segment( 215 struct xfs_buf_log_item *bip, 216 struct xfs_log_vec *lv, 217 struct xfs_log_iovec **vecp, 218 uint offset, 219 struct xfs_buf_log_format *blfp) 220 { 221 struct xfs_buf *bp = bip->bli_buf; 222 uint base_size; 223 int first_bit; 224 int last_bit; 225 int next_bit; 226 uint nbits; 227 228 /* copy the flags across from the base format item */ 229 blfp->blf_flags = bip->__bli_format.blf_flags; 230 231 /* 232 * Base size is the actual size of the ondisk structure - it reflects 233 * the actual size of the dirty bitmap rather than the size of the in 234 * memory structure. 235 */ 236 base_size = xfs_buf_log_format_size(blfp); 237 238 first_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size, 0); 239 if (!(bip->bli_flags & XFS_BLI_STALE) && first_bit == -1) { 240 /* 241 * If the map is not be dirty in the transaction, mark 242 * the size as zero and do not advance the vector pointer. 243 */ 244 return; 245 } 246 247 blfp = xlog_copy_iovec(lv, vecp, XLOG_REG_TYPE_BFORMAT, blfp, base_size); 248 blfp->blf_size = 1; 249 250 if (bip->bli_flags & XFS_BLI_STALE) { 251 /* 252 * The buffer is stale, so all we need to log 253 * is the buf log format structure with the 254 * cancel flag in it. 255 */ 256 trace_xfs_buf_item_format_stale(bip); 257 ASSERT(blfp->blf_flags & XFS_BLF_CANCEL); 258 return; 259 } 260 261 262 /* 263 * Fill in an iovec for each set of contiguous chunks. 264 */ 265 last_bit = first_bit; 266 nbits = 1; 267 for (;;) { 268 /* 269 * This takes the bit number to start looking from and 270 * returns the next set bit from there. It returns -1 271 * if there are no more bits set or the start bit is 272 * beyond the end of the bitmap. 273 */ 274 next_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size, 275 (uint)last_bit + 1); 276 /* 277 * If we run out of bits fill in the last iovec and get out of 278 * the loop. Else if we start a new set of bits then fill in 279 * the iovec for the series we were looking at and start 280 * counting the bits in the new one. Else we're still in the 281 * same set of bits so just keep counting and scanning. 282 */ 283 if (next_bit == -1) { 284 xfs_buf_item_copy_iovec(lv, vecp, bp, offset, 285 first_bit, nbits); 286 blfp->blf_size++; 287 break; 288 } else if (next_bit != last_bit + 1 || 289 xfs_buf_item_straddle(bp, offset, next_bit, last_bit)) { 290 xfs_buf_item_copy_iovec(lv, vecp, bp, offset, 291 first_bit, nbits); 292 blfp->blf_size++; 293 first_bit = next_bit; 294 last_bit = next_bit; 295 nbits = 1; 296 } else { 297 last_bit++; 298 nbits++; 299 } 300 } 301 } 302 303 /* 304 * This is called to fill in the vector of log iovecs for the 305 * given log buf item. It fills the first entry with a buf log 306 * format structure, and the rest point to contiguous chunks 307 * within the buffer. 308 */ 309 STATIC void 310 xfs_buf_item_format( 311 struct xfs_log_item *lip, 312 struct xfs_log_vec *lv) 313 { 314 struct xfs_buf_log_item *bip = BUF_ITEM(lip); 315 struct xfs_buf *bp = bip->bli_buf; 316 struct xfs_log_iovec *vecp = NULL; 317 uint offset = 0; 318 int i; 319 320 ASSERT(atomic_read(&bip->bli_refcount) > 0); 321 ASSERT((bip->bli_flags & XFS_BLI_LOGGED) || 322 (bip->bli_flags & XFS_BLI_STALE)); 323 ASSERT((bip->bli_flags & XFS_BLI_STALE) || 324 (xfs_blft_from_flags(&bip->__bli_format) > XFS_BLFT_UNKNOWN_BUF 325 && xfs_blft_from_flags(&bip->__bli_format) < XFS_BLFT_MAX_BUF)); 326 ASSERT(!(bip->bli_flags & XFS_BLI_ORDERED) || 327 (bip->bli_flags & XFS_BLI_STALE)); 328 329 330 /* 331 * If it is an inode buffer, transfer the in-memory state to the 332 * format flags and clear the in-memory state. 333 * 334 * For buffer based inode allocation, we do not transfer 335 * this state if the inode buffer allocation has not yet been committed 336 * to the log as setting the XFS_BLI_INODE_BUF flag will prevent 337 * correct replay of the inode allocation. 338 * 339 * For icreate item based inode allocation, the buffers aren't written 340 * to the journal during allocation, and hence we should always tag the 341 * buffer as an inode buffer so that the correct unlinked list replay 342 * occurs during recovery. 343 */ 344 if (bip->bli_flags & XFS_BLI_INODE_BUF) { 345 if (xfs_sb_version_hascrc(&lip->li_mountp->m_sb) || 346 !((bip->bli_flags & XFS_BLI_INODE_ALLOC_BUF) && 347 xfs_log_item_in_current_chkpt(lip))) 348 bip->__bli_format.blf_flags |= XFS_BLF_INODE_BUF; 349 bip->bli_flags &= ~XFS_BLI_INODE_BUF; 350 } 351 352 for (i = 0; i < bip->bli_format_count; i++) { 353 xfs_buf_item_format_segment(bip, lv, &vecp, offset, 354 &bip->bli_formats[i]); 355 offset += BBTOB(bp->b_maps[i].bm_len); 356 } 357 358 /* 359 * Check to make sure everything is consistent. 360 */ 361 trace_xfs_buf_item_format(bip); 362 } 363 364 /* 365 * This is called to pin the buffer associated with the buf log item in memory 366 * so it cannot be written out. 367 * 368 * We also always take a reference to the buffer log item here so that the bli 369 * is held while the item is pinned in memory. This means that we can 370 * unconditionally drop the reference count a transaction holds when the 371 * transaction is completed. 372 */ 373 STATIC void 374 xfs_buf_item_pin( 375 struct xfs_log_item *lip) 376 { 377 struct xfs_buf_log_item *bip = BUF_ITEM(lip); 378 379 ASSERT(atomic_read(&bip->bli_refcount) > 0); 380 ASSERT((bip->bli_flags & XFS_BLI_LOGGED) || 381 (bip->bli_flags & XFS_BLI_ORDERED) || 382 (bip->bli_flags & XFS_BLI_STALE)); 383 384 trace_xfs_buf_item_pin(bip); 385 386 atomic_inc(&bip->bli_refcount); 387 atomic_inc(&bip->bli_buf->b_pin_count); 388 } 389 390 /* 391 * This is called to unpin the buffer associated with the buf log 392 * item which was previously pinned with a call to xfs_buf_item_pin(). 393 * 394 * Also drop the reference to the buf item for the current transaction. 395 * If the XFS_BLI_STALE flag is set and we are the last reference, 396 * then free up the buf log item and unlock the buffer. 397 * 398 * If the remove flag is set we are called from uncommit in the 399 * forced-shutdown path. If that is true and the reference count on 400 * the log item is going to drop to zero we need to free the item's 401 * descriptor in the transaction. 402 */ 403 STATIC void 404 xfs_buf_item_unpin( 405 struct xfs_log_item *lip, 406 int remove) 407 { 408 struct xfs_buf_log_item *bip = BUF_ITEM(lip); 409 xfs_buf_t *bp = bip->bli_buf; 410 struct xfs_ail *ailp = lip->li_ailp; 411 int stale = bip->bli_flags & XFS_BLI_STALE; 412 int freed; 413 414 ASSERT(bp->b_log_item == bip); 415 ASSERT(atomic_read(&bip->bli_refcount) > 0); 416 417 trace_xfs_buf_item_unpin(bip); 418 419 freed = atomic_dec_and_test(&bip->bli_refcount); 420 421 if (atomic_dec_and_test(&bp->b_pin_count)) 422 wake_up_all(&bp->b_waiters); 423 424 if (freed && stale) { 425 ASSERT(bip->bli_flags & XFS_BLI_STALE); 426 ASSERT(xfs_buf_islocked(bp)); 427 ASSERT(bp->b_flags & XBF_STALE); 428 ASSERT(bip->__bli_format.blf_flags & XFS_BLF_CANCEL); 429 430 trace_xfs_buf_item_unpin_stale(bip); 431 432 if (remove) { 433 /* 434 * If we are in a transaction context, we have to 435 * remove the log item from the transaction as we are 436 * about to release our reference to the buffer. If we 437 * don't, the unlock that occurs later in 438 * xfs_trans_uncommit() will try to reference the 439 * buffer which we no longer have a hold on. 440 */ 441 if (!list_empty(&lip->li_trans)) 442 xfs_trans_del_item(lip); 443 444 /* 445 * Since the transaction no longer refers to the buffer, 446 * the buffer should no longer refer to the transaction. 447 */ 448 bp->b_transp = NULL; 449 } 450 451 /* 452 * If we get called here because of an IO error, we may 453 * or may not have the item on the AIL. xfs_trans_ail_delete() 454 * will take care of that situation. 455 * xfs_trans_ail_delete() drops the AIL lock. 456 */ 457 if (bip->bli_flags & XFS_BLI_STALE_INODE) { 458 xfs_buf_do_callbacks(bp); 459 bp->b_log_item = NULL; 460 list_del_init(&bp->b_li_list); 461 bp->b_iodone = NULL; 462 } else { 463 spin_lock(&ailp->ail_lock); 464 xfs_trans_ail_delete(ailp, lip, SHUTDOWN_LOG_IO_ERROR); 465 xfs_buf_item_relse(bp); 466 ASSERT(bp->b_log_item == NULL); 467 } 468 xfs_buf_relse(bp); 469 } else if (freed && remove) { 470 /* 471 * There are currently two references to the buffer - the active 472 * LRU reference and the buf log item. What we are about to do 473 * here - simulate a failed IO completion - requires 3 474 * references. 475 * 476 * The LRU reference is removed by the xfs_buf_stale() call. The 477 * buf item reference is removed by the xfs_buf_iodone() 478 * callback that is run by xfs_buf_do_callbacks() during ioend 479 * processing (via the bp->b_iodone callback), and then finally 480 * the ioend processing will drop the IO reference if the buffer 481 * is marked XBF_ASYNC. 482 * 483 * Hence we need to take an additional reference here so that IO 484 * completion processing doesn't free the buffer prematurely. 485 */ 486 xfs_buf_lock(bp); 487 xfs_buf_hold(bp); 488 bp->b_flags |= XBF_ASYNC; 489 xfs_buf_ioerror(bp, -EIO); 490 bp->b_flags &= ~XBF_DONE; 491 xfs_buf_stale(bp); 492 xfs_buf_ioend(bp); 493 } 494 } 495 496 /* 497 * Buffer IO error rate limiting. Limit it to no more than 10 messages per 30 498 * seconds so as to not spam logs too much on repeated detection of the same 499 * buffer being bad.. 500 */ 501 502 static DEFINE_RATELIMIT_STATE(xfs_buf_write_fail_rl_state, 30 * HZ, 10); 503 504 STATIC uint 505 xfs_buf_item_push( 506 struct xfs_log_item *lip, 507 struct list_head *buffer_list) 508 { 509 struct xfs_buf_log_item *bip = BUF_ITEM(lip); 510 struct xfs_buf *bp = bip->bli_buf; 511 uint rval = XFS_ITEM_SUCCESS; 512 513 if (xfs_buf_ispinned(bp)) 514 return XFS_ITEM_PINNED; 515 if (!xfs_buf_trylock(bp)) { 516 /* 517 * If we have just raced with a buffer being pinned and it has 518 * been marked stale, we could end up stalling until someone else 519 * issues a log force to unpin the stale buffer. Check for the 520 * race condition here so xfsaild recognizes the buffer is pinned 521 * and queues a log force to move it along. 522 */ 523 if (xfs_buf_ispinned(bp)) 524 return XFS_ITEM_PINNED; 525 return XFS_ITEM_LOCKED; 526 } 527 528 ASSERT(!(bip->bli_flags & XFS_BLI_STALE)); 529 530 trace_xfs_buf_item_push(bip); 531 532 /* has a previous flush failed due to IO errors? */ 533 if ((bp->b_flags & XBF_WRITE_FAIL) && 534 ___ratelimit(&xfs_buf_write_fail_rl_state, "XFS: Failing async write")) { 535 xfs_warn(bp->b_target->bt_mount, 536 "Failing async write on buffer block 0x%llx. Retrying async write.", 537 (long long)bp->b_bn); 538 } 539 540 if (!xfs_buf_delwri_queue(bp, buffer_list)) 541 rval = XFS_ITEM_FLUSHING; 542 xfs_buf_unlock(bp); 543 return rval; 544 } 545 546 /* 547 * Release the buffer associated with the buf log item. If there is no dirty 548 * logged data associated with the buffer recorded in the buf log item, then 549 * free the buf log item and remove the reference to it in the buffer. 550 * 551 * This call ignores the recursion count. It is only called when the buffer 552 * should REALLY be unlocked, regardless of the recursion count. 553 * 554 * We unconditionally drop the transaction's reference to the log item. If the 555 * item was logged, then another reference was taken when it was pinned, so we 556 * can safely drop the transaction reference now. This also allows us to avoid 557 * potential races with the unpin code freeing the bli by not referencing the 558 * bli after we've dropped the reference count. 559 * 560 * If the XFS_BLI_HOLD flag is set in the buf log item, then free the log item 561 * if necessary but do not unlock the buffer. This is for support of 562 * xfs_trans_bhold(). Make sure the XFS_BLI_HOLD field is cleared if we don't 563 * free the item. 564 */ 565 STATIC void 566 xfs_buf_item_unlock( 567 struct xfs_log_item *lip) 568 { 569 struct xfs_buf_log_item *bip = BUF_ITEM(lip); 570 struct xfs_buf *bp = bip->bli_buf; 571 bool aborted; 572 bool hold = !!(bip->bli_flags & XFS_BLI_HOLD); 573 bool dirty = !!(bip->bli_flags & XFS_BLI_DIRTY); 574 #if defined(DEBUG) || defined(XFS_WARN) 575 bool ordered = !!(bip->bli_flags & XFS_BLI_ORDERED); 576 #endif 577 578 aborted = test_bit(XFS_LI_ABORTED, &lip->li_flags); 579 580 /* Clear the buffer's association with this transaction. */ 581 bp->b_transp = NULL; 582 583 /* 584 * The per-transaction state has been copied above so clear it from the 585 * bli. 586 */ 587 bip->bli_flags &= ~(XFS_BLI_LOGGED | XFS_BLI_HOLD | XFS_BLI_ORDERED); 588 589 /* 590 * If the buf item is marked stale, then don't do anything. We'll 591 * unlock the buffer and free the buf item when the buffer is unpinned 592 * for the last time. 593 */ 594 if (bip->bli_flags & XFS_BLI_STALE) { 595 trace_xfs_buf_item_unlock_stale(bip); 596 ASSERT(bip->__bli_format.blf_flags & XFS_BLF_CANCEL); 597 if (!aborted) { 598 atomic_dec(&bip->bli_refcount); 599 return; 600 } 601 } 602 603 trace_xfs_buf_item_unlock(bip); 604 605 /* 606 * If the buf item isn't tracking any data, free it, otherwise drop the 607 * reference we hold to it. If we are aborting the transaction, this may 608 * be the only reference to the buf item, so we free it anyway 609 * regardless of whether it is dirty or not. A dirty abort implies a 610 * shutdown, anyway. 611 * 612 * The bli dirty state should match whether the blf has logged segments 613 * except for ordered buffers, where only the bli should be dirty. 614 */ 615 ASSERT((!ordered && dirty == xfs_buf_item_dirty_format(bip)) || 616 (ordered && dirty && !xfs_buf_item_dirty_format(bip))); 617 618 /* 619 * Clean buffers, by definition, cannot be in the AIL. However, aborted 620 * buffers may be in the AIL regardless of dirty state. An aborted 621 * transaction that invalidates a buffer already in the AIL may have 622 * marked it stale and cleared the dirty state, for example. 623 * 624 * Therefore if we are aborting a buffer and we've just taken the last 625 * reference away, we have to check if it is in the AIL before freeing 626 * it. We need to free it in this case, because an aborted transaction 627 * has already shut the filesystem down and this is the last chance we 628 * will have to do so. 629 */ 630 if (atomic_dec_and_test(&bip->bli_refcount)) { 631 if (aborted) { 632 ASSERT(XFS_FORCED_SHUTDOWN(lip->li_mountp)); 633 xfs_trans_ail_remove(lip, SHUTDOWN_LOG_IO_ERROR); 634 xfs_buf_item_relse(bp); 635 } else if (!dirty) 636 xfs_buf_item_relse(bp); 637 } 638 639 if (!hold) 640 xfs_buf_relse(bp); 641 } 642 643 /* 644 * This is called to find out where the oldest active copy of the 645 * buf log item in the on disk log resides now that the last log 646 * write of it completed at the given lsn. 647 * We always re-log all the dirty data in a buffer, so usually the 648 * latest copy in the on disk log is the only one that matters. For 649 * those cases we simply return the given lsn. 650 * 651 * The one exception to this is for buffers full of newly allocated 652 * inodes. These buffers are only relogged with the XFS_BLI_INODE_BUF 653 * flag set, indicating that only the di_next_unlinked fields from the 654 * inodes in the buffers will be replayed during recovery. If the 655 * original newly allocated inode images have not yet been flushed 656 * when the buffer is so relogged, then we need to make sure that we 657 * keep the old images in the 'active' portion of the log. We do this 658 * by returning the original lsn of that transaction here rather than 659 * the current one. 660 */ 661 STATIC xfs_lsn_t 662 xfs_buf_item_committed( 663 struct xfs_log_item *lip, 664 xfs_lsn_t lsn) 665 { 666 struct xfs_buf_log_item *bip = BUF_ITEM(lip); 667 668 trace_xfs_buf_item_committed(bip); 669 670 if ((bip->bli_flags & XFS_BLI_INODE_ALLOC_BUF) && lip->li_lsn != 0) 671 return lip->li_lsn; 672 return lsn; 673 } 674 675 STATIC void 676 xfs_buf_item_committing( 677 struct xfs_log_item *lip, 678 xfs_lsn_t commit_lsn) 679 { 680 } 681 682 /* 683 * This is the ops vector shared by all buf log items. 684 */ 685 static const struct xfs_item_ops xfs_buf_item_ops = { 686 .iop_size = xfs_buf_item_size, 687 .iop_format = xfs_buf_item_format, 688 .iop_pin = xfs_buf_item_pin, 689 .iop_unpin = xfs_buf_item_unpin, 690 .iop_unlock = xfs_buf_item_unlock, 691 .iop_committed = xfs_buf_item_committed, 692 .iop_push = xfs_buf_item_push, 693 .iop_committing = xfs_buf_item_committing 694 }; 695 696 STATIC int 697 xfs_buf_item_get_format( 698 struct xfs_buf_log_item *bip, 699 int count) 700 { 701 ASSERT(bip->bli_formats == NULL); 702 bip->bli_format_count = count; 703 704 if (count == 1) { 705 bip->bli_formats = &bip->__bli_format; 706 return 0; 707 } 708 709 bip->bli_formats = kmem_zalloc(count * sizeof(struct xfs_buf_log_format), 710 KM_SLEEP); 711 if (!bip->bli_formats) 712 return -ENOMEM; 713 return 0; 714 } 715 716 STATIC void 717 xfs_buf_item_free_format( 718 struct xfs_buf_log_item *bip) 719 { 720 if (bip->bli_formats != &bip->__bli_format) { 721 kmem_free(bip->bli_formats); 722 bip->bli_formats = NULL; 723 } 724 } 725 726 /* 727 * Allocate a new buf log item to go with the given buffer. 728 * Set the buffer's b_log_item field to point to the new 729 * buf log item. 730 */ 731 int 732 xfs_buf_item_init( 733 struct xfs_buf *bp, 734 struct xfs_mount *mp) 735 { 736 struct xfs_buf_log_item *bip = bp->b_log_item; 737 int chunks; 738 int map_size; 739 int error; 740 int i; 741 742 /* 743 * Check to see if there is already a buf log item for 744 * this buffer. If we do already have one, there is 745 * nothing to do here so return. 746 */ 747 ASSERT(bp->b_target->bt_mount == mp); 748 if (bip) { 749 ASSERT(bip->bli_item.li_type == XFS_LI_BUF); 750 ASSERT(!bp->b_transp); 751 ASSERT(bip->bli_buf == bp); 752 return 0; 753 } 754 755 bip = kmem_zone_zalloc(xfs_buf_item_zone, KM_SLEEP); 756 xfs_log_item_init(mp, &bip->bli_item, XFS_LI_BUF, &xfs_buf_item_ops); 757 bip->bli_buf = bp; 758 759 /* 760 * chunks is the number of XFS_BLF_CHUNK size pieces the buffer 761 * can be divided into. Make sure not to truncate any pieces. 762 * map_size is the size of the bitmap needed to describe the 763 * chunks of the buffer. 764 * 765 * Discontiguous buffer support follows the layout of the underlying 766 * buffer. This makes the implementation as simple as possible. 767 */ 768 error = xfs_buf_item_get_format(bip, bp->b_map_count); 769 ASSERT(error == 0); 770 if (error) { /* to stop gcc throwing set-but-unused warnings */ 771 kmem_zone_free(xfs_buf_item_zone, bip); 772 return error; 773 } 774 775 776 for (i = 0; i < bip->bli_format_count; i++) { 777 chunks = DIV_ROUND_UP(BBTOB(bp->b_maps[i].bm_len), 778 XFS_BLF_CHUNK); 779 map_size = DIV_ROUND_UP(chunks, NBWORD); 780 781 bip->bli_formats[i].blf_type = XFS_LI_BUF; 782 bip->bli_formats[i].blf_blkno = bp->b_maps[i].bm_bn; 783 bip->bli_formats[i].blf_len = bp->b_maps[i].bm_len; 784 bip->bli_formats[i].blf_map_size = map_size; 785 } 786 787 bp->b_log_item = bip; 788 xfs_buf_hold(bp); 789 return 0; 790 } 791 792 793 /* 794 * Mark bytes first through last inclusive as dirty in the buf 795 * item's bitmap. 796 */ 797 static void 798 xfs_buf_item_log_segment( 799 uint first, 800 uint last, 801 uint *map) 802 { 803 uint first_bit; 804 uint last_bit; 805 uint bits_to_set; 806 uint bits_set; 807 uint word_num; 808 uint *wordp; 809 uint bit; 810 uint end_bit; 811 uint mask; 812 813 /* 814 * Convert byte offsets to bit numbers. 815 */ 816 first_bit = first >> XFS_BLF_SHIFT; 817 last_bit = last >> XFS_BLF_SHIFT; 818 819 /* 820 * Calculate the total number of bits to be set. 821 */ 822 bits_to_set = last_bit - first_bit + 1; 823 824 /* 825 * Get a pointer to the first word in the bitmap 826 * to set a bit in. 827 */ 828 word_num = first_bit >> BIT_TO_WORD_SHIFT; 829 wordp = &map[word_num]; 830 831 /* 832 * Calculate the starting bit in the first word. 833 */ 834 bit = first_bit & (uint)(NBWORD - 1); 835 836 /* 837 * First set any bits in the first word of our range. 838 * If it starts at bit 0 of the word, it will be 839 * set below rather than here. That is what the variable 840 * bit tells us. The variable bits_set tracks the number 841 * of bits that have been set so far. End_bit is the number 842 * of the last bit to be set in this word plus one. 843 */ 844 if (bit) { 845 end_bit = MIN(bit + bits_to_set, (uint)NBWORD); 846 mask = ((1U << (end_bit - bit)) - 1) << bit; 847 *wordp |= mask; 848 wordp++; 849 bits_set = end_bit - bit; 850 } else { 851 bits_set = 0; 852 } 853 854 /* 855 * Now set bits a whole word at a time that are between 856 * first_bit and last_bit. 857 */ 858 while ((bits_to_set - bits_set) >= NBWORD) { 859 *wordp |= 0xffffffff; 860 bits_set += NBWORD; 861 wordp++; 862 } 863 864 /* 865 * Finally, set any bits left to be set in one last partial word. 866 */ 867 end_bit = bits_to_set - bits_set; 868 if (end_bit) { 869 mask = (1U << end_bit) - 1; 870 *wordp |= mask; 871 } 872 } 873 874 /* 875 * Mark bytes first through last inclusive as dirty in the buf 876 * item's bitmap. 877 */ 878 void 879 xfs_buf_item_log( 880 struct xfs_buf_log_item *bip, 881 uint first, 882 uint last) 883 { 884 int i; 885 uint start; 886 uint end; 887 struct xfs_buf *bp = bip->bli_buf; 888 889 /* 890 * walk each buffer segment and mark them dirty appropriately. 891 */ 892 start = 0; 893 for (i = 0; i < bip->bli_format_count; i++) { 894 if (start > last) 895 break; 896 end = start + BBTOB(bp->b_maps[i].bm_len) - 1; 897 898 /* skip to the map that includes the first byte to log */ 899 if (first > end) { 900 start += BBTOB(bp->b_maps[i].bm_len); 901 continue; 902 } 903 904 /* 905 * Trim the range to this segment and mark it in the bitmap. 906 * Note that we must convert buffer offsets to segment relative 907 * offsets (e.g., the first byte of each segment is byte 0 of 908 * that segment). 909 */ 910 if (first < start) 911 first = start; 912 if (end > last) 913 end = last; 914 xfs_buf_item_log_segment(first - start, end - start, 915 &bip->bli_formats[i].blf_data_map[0]); 916 917 start += BBTOB(bp->b_maps[i].bm_len); 918 } 919 } 920 921 922 /* 923 * Return true if the buffer has any ranges logged/dirtied by a transaction, 924 * false otherwise. 925 */ 926 bool 927 xfs_buf_item_dirty_format( 928 struct xfs_buf_log_item *bip) 929 { 930 int i; 931 932 for (i = 0; i < bip->bli_format_count; i++) { 933 if (!xfs_bitmap_empty(bip->bli_formats[i].blf_data_map, 934 bip->bli_formats[i].blf_map_size)) 935 return true; 936 } 937 938 return false; 939 } 940 941 STATIC void 942 xfs_buf_item_free( 943 struct xfs_buf_log_item *bip) 944 { 945 xfs_buf_item_free_format(bip); 946 kmem_free(bip->bli_item.li_lv_shadow); 947 kmem_zone_free(xfs_buf_item_zone, bip); 948 } 949 950 /* 951 * This is called when the buf log item is no longer needed. It should 952 * free the buf log item associated with the given buffer and clear 953 * the buffer's pointer to the buf log item. If there are no more 954 * items in the list, clear the b_iodone field of the buffer (see 955 * xfs_buf_attach_iodone() below). 956 */ 957 void 958 xfs_buf_item_relse( 959 xfs_buf_t *bp) 960 { 961 struct xfs_buf_log_item *bip = bp->b_log_item; 962 963 trace_xfs_buf_item_relse(bp, _RET_IP_); 964 ASSERT(!(bip->bli_item.li_flags & XFS_LI_IN_AIL)); 965 966 bp->b_log_item = NULL; 967 if (list_empty(&bp->b_li_list)) 968 bp->b_iodone = NULL; 969 970 xfs_buf_rele(bp); 971 xfs_buf_item_free(bip); 972 } 973 974 975 /* 976 * Add the given log item with its callback to the list of callbacks 977 * to be called when the buffer's I/O completes. If it is not set 978 * already, set the buffer's b_iodone() routine to be 979 * xfs_buf_iodone_callbacks() and link the log item into the list of 980 * items rooted at b_li_list. 981 */ 982 void 983 xfs_buf_attach_iodone( 984 xfs_buf_t *bp, 985 void (*cb)(xfs_buf_t *, xfs_log_item_t *), 986 xfs_log_item_t *lip) 987 { 988 ASSERT(xfs_buf_islocked(bp)); 989 990 lip->li_cb = cb; 991 list_add_tail(&lip->li_bio_list, &bp->b_li_list); 992 993 ASSERT(bp->b_iodone == NULL || 994 bp->b_iodone == xfs_buf_iodone_callbacks); 995 bp->b_iodone = xfs_buf_iodone_callbacks; 996 } 997 998 /* 999 * We can have many callbacks on a buffer. Running the callbacks individually 1000 * can cause a lot of contention on the AIL lock, so we allow for a single 1001 * callback to be able to scan the remaining items in bp->b_li_list for other 1002 * items of the same type and callback to be processed in the first call. 1003 * 1004 * As a result, the loop walking the callback list below will also modify the 1005 * list. it removes the first item from the list and then runs the callback. 1006 * The loop then restarts from the new first item int the list. This allows the 1007 * callback to scan and modify the list attached to the buffer and we don't 1008 * have to care about maintaining a next item pointer. 1009 */ 1010 STATIC void 1011 xfs_buf_do_callbacks( 1012 struct xfs_buf *bp) 1013 { 1014 struct xfs_buf_log_item *blip = bp->b_log_item; 1015 struct xfs_log_item *lip; 1016 1017 /* If there is a buf_log_item attached, run its callback */ 1018 if (blip) { 1019 lip = &blip->bli_item; 1020 lip->li_cb(bp, lip); 1021 } 1022 1023 while (!list_empty(&bp->b_li_list)) { 1024 lip = list_first_entry(&bp->b_li_list, struct xfs_log_item, 1025 li_bio_list); 1026 1027 /* 1028 * Remove the item from the list, so we don't have any 1029 * confusion if the item is added to another buf. 1030 * Don't touch the log item after calling its 1031 * callback, because it could have freed itself. 1032 */ 1033 list_del_init(&lip->li_bio_list); 1034 lip->li_cb(bp, lip); 1035 } 1036 } 1037 1038 /* 1039 * Invoke the error state callback for each log item affected by the failed I/O. 1040 * 1041 * If a metadata buffer write fails with a non-permanent error, the buffer is 1042 * eventually resubmitted and so the completion callbacks are not run. The error 1043 * state may need to be propagated to the log items attached to the buffer, 1044 * however, so the next AIL push of the item knows hot to handle it correctly. 1045 */ 1046 STATIC void 1047 xfs_buf_do_callbacks_fail( 1048 struct xfs_buf *bp) 1049 { 1050 struct xfs_log_item *lip; 1051 struct xfs_ail *ailp; 1052 1053 /* 1054 * Buffer log item errors are handled directly by xfs_buf_item_push() 1055 * and xfs_buf_iodone_callback_error, and they have no IO error 1056 * callbacks. Check only for items in b_li_list. 1057 */ 1058 if (list_empty(&bp->b_li_list)) 1059 return; 1060 1061 lip = list_first_entry(&bp->b_li_list, struct xfs_log_item, 1062 li_bio_list); 1063 ailp = lip->li_ailp; 1064 spin_lock(&ailp->ail_lock); 1065 list_for_each_entry(lip, &bp->b_li_list, li_bio_list) { 1066 if (lip->li_ops->iop_error) 1067 lip->li_ops->iop_error(lip, bp); 1068 } 1069 spin_unlock(&ailp->ail_lock); 1070 } 1071 1072 static bool 1073 xfs_buf_iodone_callback_error( 1074 struct xfs_buf *bp) 1075 { 1076 struct xfs_buf_log_item *bip = bp->b_log_item; 1077 struct xfs_log_item *lip; 1078 struct xfs_mount *mp; 1079 static ulong lasttime; 1080 static xfs_buftarg_t *lasttarg; 1081 struct xfs_error_cfg *cfg; 1082 1083 /* 1084 * The failed buffer might not have a buf_log_item attached or the 1085 * log_item list might be empty. Get the mp from the available 1086 * xfs_log_item 1087 */ 1088 lip = list_first_entry_or_null(&bp->b_li_list, struct xfs_log_item, 1089 li_bio_list); 1090 mp = lip ? lip->li_mountp : bip->bli_item.li_mountp; 1091 1092 /* 1093 * If we've already decided to shutdown the filesystem because of 1094 * I/O errors, there's no point in giving this a retry. 1095 */ 1096 if (XFS_FORCED_SHUTDOWN(mp)) 1097 goto out_stale; 1098 1099 if (bp->b_target != lasttarg || 1100 time_after(jiffies, (lasttime + 5*HZ))) { 1101 lasttime = jiffies; 1102 xfs_buf_ioerror_alert(bp, __func__); 1103 } 1104 lasttarg = bp->b_target; 1105 1106 /* synchronous writes will have callers process the error */ 1107 if (!(bp->b_flags & XBF_ASYNC)) 1108 goto out_stale; 1109 1110 trace_xfs_buf_item_iodone_async(bp, _RET_IP_); 1111 ASSERT(bp->b_iodone != NULL); 1112 1113 cfg = xfs_error_get_cfg(mp, XFS_ERR_METADATA, bp->b_error); 1114 1115 /* 1116 * If the write was asynchronous then no one will be looking for the 1117 * error. If this is the first failure of this type, clear the error 1118 * state and write the buffer out again. This means we always retry an 1119 * async write failure at least once, but we also need to set the buffer 1120 * up to behave correctly now for repeated failures. 1121 */ 1122 if (!(bp->b_flags & (XBF_STALE | XBF_WRITE_FAIL)) || 1123 bp->b_last_error != bp->b_error) { 1124 bp->b_flags |= (XBF_WRITE | XBF_DONE | XBF_WRITE_FAIL); 1125 bp->b_last_error = bp->b_error; 1126 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER && 1127 !bp->b_first_retry_time) 1128 bp->b_first_retry_time = jiffies; 1129 1130 xfs_buf_ioerror(bp, 0); 1131 xfs_buf_submit(bp); 1132 return true; 1133 } 1134 1135 /* 1136 * Repeated failure on an async write. Take action according to the 1137 * error configuration we have been set up to use. 1138 */ 1139 1140 if (cfg->max_retries != XFS_ERR_RETRY_FOREVER && 1141 ++bp->b_retries > cfg->max_retries) 1142 goto permanent_error; 1143 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER && 1144 time_after(jiffies, cfg->retry_timeout + bp->b_first_retry_time)) 1145 goto permanent_error; 1146 1147 /* At unmount we may treat errors differently */ 1148 if ((mp->m_flags & XFS_MOUNT_UNMOUNTING) && mp->m_fail_unmount) 1149 goto permanent_error; 1150 1151 /* 1152 * Still a transient error, run IO completion failure callbacks and let 1153 * the higher layers retry the buffer. 1154 */ 1155 xfs_buf_do_callbacks_fail(bp); 1156 xfs_buf_ioerror(bp, 0); 1157 xfs_buf_relse(bp); 1158 return true; 1159 1160 /* 1161 * Permanent error - we need to trigger a shutdown if we haven't already 1162 * to indicate that inconsistency will result from this action. 1163 */ 1164 permanent_error: 1165 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR); 1166 out_stale: 1167 xfs_buf_stale(bp); 1168 bp->b_flags |= XBF_DONE; 1169 trace_xfs_buf_error_relse(bp, _RET_IP_); 1170 return false; 1171 } 1172 1173 /* 1174 * This is the iodone() function for buffers which have had callbacks attached 1175 * to them by xfs_buf_attach_iodone(). We need to iterate the items on the 1176 * callback list, mark the buffer as having no more callbacks and then push the 1177 * buffer through IO completion processing. 1178 */ 1179 void 1180 xfs_buf_iodone_callbacks( 1181 struct xfs_buf *bp) 1182 { 1183 /* 1184 * If there is an error, process it. Some errors require us 1185 * to run callbacks after failure processing is done so we 1186 * detect that and take appropriate action. 1187 */ 1188 if (bp->b_error && xfs_buf_iodone_callback_error(bp)) 1189 return; 1190 1191 /* 1192 * Successful IO or permanent error. Either way, we can clear the 1193 * retry state here in preparation for the next error that may occur. 1194 */ 1195 bp->b_last_error = 0; 1196 bp->b_retries = 0; 1197 bp->b_first_retry_time = 0; 1198 1199 xfs_buf_do_callbacks(bp); 1200 bp->b_log_item = NULL; 1201 list_del_init(&bp->b_li_list); 1202 bp->b_iodone = NULL; 1203 xfs_buf_ioend(bp); 1204 } 1205 1206 /* 1207 * This is the iodone() function for buffers which have been 1208 * logged. It is called when they are eventually flushed out. 1209 * It should remove the buf item from the AIL, and free the buf item. 1210 * It is called by xfs_buf_iodone_callbacks() above which will take 1211 * care of cleaning up the buffer itself. 1212 */ 1213 void 1214 xfs_buf_iodone( 1215 struct xfs_buf *bp, 1216 struct xfs_log_item *lip) 1217 { 1218 struct xfs_ail *ailp = lip->li_ailp; 1219 1220 ASSERT(BUF_ITEM(lip)->bli_buf == bp); 1221 1222 xfs_buf_rele(bp); 1223 1224 /* 1225 * If we are forcibly shutting down, this may well be 1226 * off the AIL already. That's because we simulate the 1227 * log-committed callbacks to unpin these buffers. Or we may never 1228 * have put this item on AIL because of the transaction was 1229 * aborted forcibly. xfs_trans_ail_delete() takes care of these. 1230 * 1231 * Either way, AIL is useless if we're forcing a shutdown. 1232 */ 1233 spin_lock(&ailp->ail_lock); 1234 xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE); 1235 xfs_buf_item_free(BUF_ITEM(lip)); 1236 } 1237 1238 /* 1239 * Requeue a failed buffer for writeback 1240 * 1241 * Return true if the buffer has been re-queued properly, false otherwise 1242 */ 1243 bool 1244 xfs_buf_resubmit_failed_buffers( 1245 struct xfs_buf *bp, 1246 struct list_head *buffer_list) 1247 { 1248 struct xfs_log_item *lip; 1249 1250 /* 1251 * Clear XFS_LI_FAILED flag from all items before resubmit 1252 * 1253 * XFS_LI_FAILED set/clear is protected by ail_lock, caller this 1254 * function already have it acquired 1255 */ 1256 list_for_each_entry(lip, &bp->b_li_list, li_bio_list) 1257 xfs_clear_li_failed(lip); 1258 1259 /* Add this buffer back to the delayed write list */ 1260 return xfs_buf_delwri_queue(bp, buffer_list); 1261 } 1262