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