1 /* 2 * linux/fs/jbd2/transaction.c 3 * 4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998 5 * 6 * Copyright 1998 Red Hat corp --- All Rights Reserved 7 * 8 * This file is part of the Linux kernel and is made available under 9 * the terms of the GNU General Public License, version 2, or at your 10 * option, any later version, incorporated herein by reference. 11 * 12 * Generic filesystem transaction handling code; part of the ext2fs 13 * journaling system. 14 * 15 * This file manages transactions (compound commits managed by the 16 * journaling code) and handles (individual atomic operations by the 17 * filesystem). 18 */ 19 20 #include <linux/time.h> 21 #include <linux/fs.h> 22 #include <linux/jbd2.h> 23 #include <linux/errno.h> 24 #include <linux/slab.h> 25 #include <linux/timer.h> 26 #include <linux/mm.h> 27 #include <linux/highmem.h> 28 #include <linux/hrtimer.h> 29 #include <linux/backing-dev.h> 30 #include <linux/bug.h> 31 #include <linux/module.h> 32 33 #include <trace/events/jbd2.h> 34 35 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh); 36 static void __jbd2_journal_unfile_buffer(struct journal_head *jh); 37 38 static struct kmem_cache *transaction_cache; 39 int __init jbd2_journal_init_transaction_cache(void) 40 { 41 J_ASSERT(!transaction_cache); 42 transaction_cache = kmem_cache_create("jbd2_transaction_s", 43 sizeof(transaction_t), 44 0, 45 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY, 46 NULL); 47 if (transaction_cache) 48 return 0; 49 return -ENOMEM; 50 } 51 52 void jbd2_journal_destroy_transaction_cache(void) 53 { 54 if (transaction_cache) { 55 kmem_cache_destroy(transaction_cache); 56 transaction_cache = NULL; 57 } 58 } 59 60 void jbd2_journal_free_transaction(transaction_t *transaction) 61 { 62 if (unlikely(ZERO_OR_NULL_PTR(transaction))) 63 return; 64 kmem_cache_free(transaction_cache, transaction); 65 } 66 67 /* 68 * jbd2_get_transaction: obtain a new transaction_t object. 69 * 70 * Simply allocate and initialise a new transaction. Create it in 71 * RUNNING state and add it to the current journal (which should not 72 * have an existing running transaction: we only make a new transaction 73 * once we have started to commit the old one). 74 * 75 * Preconditions: 76 * The journal MUST be locked. We don't perform atomic mallocs on the 77 * new transaction and we can't block without protecting against other 78 * processes trying to touch the journal while it is in transition. 79 * 80 */ 81 82 static transaction_t * 83 jbd2_get_transaction(journal_t *journal, transaction_t *transaction) 84 { 85 transaction->t_journal = journal; 86 transaction->t_state = T_RUNNING; 87 transaction->t_start_time = ktime_get(); 88 transaction->t_tid = journal->j_transaction_sequence++; 89 transaction->t_expires = jiffies + journal->j_commit_interval; 90 spin_lock_init(&transaction->t_handle_lock); 91 atomic_set(&transaction->t_updates, 0); 92 atomic_set(&transaction->t_outstanding_credits, 93 atomic_read(&journal->j_reserved_credits)); 94 atomic_set(&transaction->t_handle_count, 0); 95 INIT_LIST_HEAD(&transaction->t_inode_list); 96 INIT_LIST_HEAD(&transaction->t_private_list); 97 98 /* Set up the commit timer for the new transaction. */ 99 journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires); 100 add_timer(&journal->j_commit_timer); 101 102 J_ASSERT(journal->j_running_transaction == NULL); 103 journal->j_running_transaction = transaction; 104 transaction->t_max_wait = 0; 105 transaction->t_start = jiffies; 106 transaction->t_requested = 0; 107 108 return transaction; 109 } 110 111 /* 112 * Handle management. 113 * 114 * A handle_t is an object which represents a single atomic update to a 115 * filesystem, and which tracks all of the modifications which form part 116 * of that one update. 117 */ 118 119 /* 120 * Update transaction's maximum wait time, if debugging is enabled. 121 * 122 * In order for t_max_wait to be reliable, it must be protected by a 123 * lock. But doing so will mean that start_this_handle() can not be 124 * run in parallel on SMP systems, which limits our scalability. So 125 * unless debugging is enabled, we no longer update t_max_wait, which 126 * means that maximum wait time reported by the jbd2_run_stats 127 * tracepoint will always be zero. 128 */ 129 static inline void update_t_max_wait(transaction_t *transaction, 130 unsigned long ts) 131 { 132 #ifdef CONFIG_JBD2_DEBUG 133 if (jbd2_journal_enable_debug && 134 time_after(transaction->t_start, ts)) { 135 ts = jbd2_time_diff(ts, transaction->t_start); 136 spin_lock(&transaction->t_handle_lock); 137 if (ts > transaction->t_max_wait) 138 transaction->t_max_wait = ts; 139 spin_unlock(&transaction->t_handle_lock); 140 } 141 #endif 142 } 143 144 /* 145 * Wait until running transaction passes T_LOCKED state. Also starts the commit 146 * if needed. The function expects running transaction to exist and releases 147 * j_state_lock. 148 */ 149 static void wait_transaction_locked(journal_t *journal) 150 __releases(journal->j_state_lock) 151 { 152 DEFINE_WAIT(wait); 153 int need_to_start; 154 tid_t tid = journal->j_running_transaction->t_tid; 155 156 prepare_to_wait(&journal->j_wait_transaction_locked, &wait, 157 TASK_UNINTERRUPTIBLE); 158 need_to_start = !tid_geq(journal->j_commit_request, tid); 159 read_unlock(&journal->j_state_lock); 160 if (need_to_start) 161 jbd2_log_start_commit(journal, tid); 162 schedule(); 163 finish_wait(&journal->j_wait_transaction_locked, &wait); 164 } 165 166 static void sub_reserved_credits(journal_t *journal, int blocks) 167 { 168 atomic_sub(blocks, &journal->j_reserved_credits); 169 wake_up(&journal->j_wait_reserved); 170 } 171 172 /* 173 * Wait until we can add credits for handle to the running transaction. Called 174 * with j_state_lock held for reading. Returns 0 if handle joined the running 175 * transaction. Returns 1 if we had to wait, j_state_lock is dropped, and 176 * caller must retry. 177 */ 178 static int add_transaction_credits(journal_t *journal, int blocks, 179 int rsv_blocks) 180 { 181 transaction_t *t = journal->j_running_transaction; 182 int needed; 183 int total = blocks + rsv_blocks; 184 185 /* 186 * If the current transaction is locked down for commit, wait 187 * for the lock to be released. 188 */ 189 if (t->t_state == T_LOCKED) { 190 wait_transaction_locked(journal); 191 return 1; 192 } 193 194 /* 195 * If there is not enough space left in the log to write all 196 * potential buffers requested by this operation, we need to 197 * stall pending a log checkpoint to free some more log space. 198 */ 199 needed = atomic_add_return(total, &t->t_outstanding_credits); 200 if (needed > journal->j_max_transaction_buffers) { 201 /* 202 * If the current transaction is already too large, 203 * then start to commit it: we can then go back and 204 * attach this handle to a new transaction. 205 */ 206 atomic_sub(total, &t->t_outstanding_credits); 207 208 /* 209 * Is the number of reserved credits in the current transaction too 210 * big to fit this handle? Wait until reserved credits are freed. 211 */ 212 if (atomic_read(&journal->j_reserved_credits) + total > 213 journal->j_max_transaction_buffers) { 214 read_unlock(&journal->j_state_lock); 215 wait_event(journal->j_wait_reserved, 216 atomic_read(&journal->j_reserved_credits) + total <= 217 journal->j_max_transaction_buffers); 218 return 1; 219 } 220 221 wait_transaction_locked(journal); 222 return 1; 223 } 224 225 /* 226 * The commit code assumes that it can get enough log space 227 * without forcing a checkpoint. This is *critical* for 228 * correctness: a checkpoint of a buffer which is also 229 * associated with a committing transaction creates a deadlock, 230 * so commit simply cannot force through checkpoints. 231 * 232 * We must therefore ensure the necessary space in the journal 233 * *before* starting to dirty potentially checkpointed buffers 234 * in the new transaction. 235 */ 236 if (jbd2_log_space_left(journal) < jbd2_space_needed(journal)) { 237 atomic_sub(total, &t->t_outstanding_credits); 238 read_unlock(&journal->j_state_lock); 239 write_lock(&journal->j_state_lock); 240 if (jbd2_log_space_left(journal) < jbd2_space_needed(journal)) 241 __jbd2_log_wait_for_space(journal); 242 write_unlock(&journal->j_state_lock); 243 return 1; 244 } 245 246 /* No reservation? We are done... */ 247 if (!rsv_blocks) 248 return 0; 249 250 needed = atomic_add_return(rsv_blocks, &journal->j_reserved_credits); 251 /* We allow at most half of a transaction to be reserved */ 252 if (needed > journal->j_max_transaction_buffers / 2) { 253 sub_reserved_credits(journal, rsv_blocks); 254 atomic_sub(total, &t->t_outstanding_credits); 255 read_unlock(&journal->j_state_lock); 256 wait_event(journal->j_wait_reserved, 257 atomic_read(&journal->j_reserved_credits) + rsv_blocks 258 <= journal->j_max_transaction_buffers / 2); 259 return 1; 260 } 261 return 0; 262 } 263 264 /* 265 * start_this_handle: Given a handle, deal with any locking or stalling 266 * needed to make sure that there is enough journal space for the handle 267 * to begin. Attach the handle to a transaction and set up the 268 * transaction's buffer credits. 269 */ 270 271 static int start_this_handle(journal_t *journal, handle_t *handle, 272 gfp_t gfp_mask) 273 { 274 transaction_t *transaction, *new_transaction = NULL; 275 int blocks = handle->h_buffer_credits; 276 int rsv_blocks = 0; 277 unsigned long ts = jiffies; 278 279 if (handle->h_rsv_handle) 280 rsv_blocks = handle->h_rsv_handle->h_buffer_credits; 281 282 /* 283 * Limit the number of reserved credits to 1/2 of maximum transaction 284 * size and limit the number of total credits to not exceed maximum 285 * transaction size per operation. 286 */ 287 if ((rsv_blocks > journal->j_max_transaction_buffers / 2) || 288 (rsv_blocks + blocks > journal->j_max_transaction_buffers)) { 289 printk(KERN_ERR "JBD2: %s wants too many credits " 290 "credits:%d rsv_credits:%d max:%d\n", 291 current->comm, blocks, rsv_blocks, 292 journal->j_max_transaction_buffers); 293 WARN_ON(1); 294 return -ENOSPC; 295 } 296 297 alloc_transaction: 298 if (!journal->j_running_transaction) { 299 /* 300 * If __GFP_FS is not present, then we may be being called from 301 * inside the fs writeback layer, so we MUST NOT fail. 302 */ 303 if ((gfp_mask & __GFP_FS) == 0) 304 gfp_mask |= __GFP_NOFAIL; 305 new_transaction = kmem_cache_zalloc(transaction_cache, 306 gfp_mask); 307 if (!new_transaction) 308 return -ENOMEM; 309 } 310 311 jbd_debug(3, "New handle %p going live.\n", handle); 312 313 /* 314 * We need to hold j_state_lock until t_updates has been incremented, 315 * for proper journal barrier handling 316 */ 317 repeat: 318 read_lock(&journal->j_state_lock); 319 BUG_ON(journal->j_flags & JBD2_UNMOUNT); 320 if (is_journal_aborted(journal) || 321 (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) { 322 read_unlock(&journal->j_state_lock); 323 jbd2_journal_free_transaction(new_transaction); 324 return -EROFS; 325 } 326 327 /* 328 * Wait on the journal's transaction barrier if necessary. Specifically 329 * we allow reserved handles to proceed because otherwise commit could 330 * deadlock on page writeback not being able to complete. 331 */ 332 if (!handle->h_reserved && journal->j_barrier_count) { 333 read_unlock(&journal->j_state_lock); 334 wait_event(journal->j_wait_transaction_locked, 335 journal->j_barrier_count == 0); 336 goto repeat; 337 } 338 339 if (!journal->j_running_transaction) { 340 read_unlock(&journal->j_state_lock); 341 if (!new_transaction) 342 goto alloc_transaction; 343 write_lock(&journal->j_state_lock); 344 if (!journal->j_running_transaction && 345 (handle->h_reserved || !journal->j_barrier_count)) { 346 jbd2_get_transaction(journal, new_transaction); 347 new_transaction = NULL; 348 } 349 write_unlock(&journal->j_state_lock); 350 goto repeat; 351 } 352 353 transaction = journal->j_running_transaction; 354 355 if (!handle->h_reserved) { 356 /* We may have dropped j_state_lock - restart in that case */ 357 if (add_transaction_credits(journal, blocks, rsv_blocks)) 358 goto repeat; 359 } else { 360 /* 361 * We have handle reserved so we are allowed to join T_LOCKED 362 * transaction and we don't have to check for transaction size 363 * and journal space. 364 */ 365 sub_reserved_credits(journal, blocks); 366 handle->h_reserved = 0; 367 } 368 369 /* OK, account for the buffers that this operation expects to 370 * use and add the handle to the running transaction. 371 */ 372 update_t_max_wait(transaction, ts); 373 handle->h_transaction = transaction; 374 handle->h_requested_credits = blocks; 375 handle->h_start_jiffies = jiffies; 376 atomic_inc(&transaction->t_updates); 377 atomic_inc(&transaction->t_handle_count); 378 jbd_debug(4, "Handle %p given %d credits (total %d, free %lu)\n", 379 handle, blocks, 380 atomic_read(&transaction->t_outstanding_credits), 381 jbd2_log_space_left(journal)); 382 read_unlock(&journal->j_state_lock); 383 current->journal_info = handle; 384 385 lock_map_acquire(&handle->h_lockdep_map); 386 jbd2_journal_free_transaction(new_transaction); 387 return 0; 388 } 389 390 static struct lock_class_key jbd2_handle_key; 391 392 /* Allocate a new handle. This should probably be in a slab... */ 393 static handle_t *new_handle(int nblocks) 394 { 395 handle_t *handle = jbd2_alloc_handle(GFP_NOFS); 396 if (!handle) 397 return NULL; 398 handle->h_buffer_credits = nblocks; 399 handle->h_ref = 1; 400 401 lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle", 402 &jbd2_handle_key, 0); 403 404 return handle; 405 } 406 407 /** 408 * handle_t *jbd2_journal_start() - Obtain a new handle. 409 * @journal: Journal to start transaction on. 410 * @nblocks: number of block buffer we might modify 411 * 412 * We make sure that the transaction can guarantee at least nblocks of 413 * modified buffers in the log. We block until the log can guarantee 414 * that much space. Additionally, if rsv_blocks > 0, we also create another 415 * handle with rsv_blocks reserved blocks in the journal. This handle is 416 * is stored in h_rsv_handle. It is not attached to any particular transaction 417 * and thus doesn't block transaction commit. If the caller uses this reserved 418 * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop() 419 * on the parent handle will dispose the reserved one. Reserved handle has to 420 * be converted to a normal handle using jbd2_journal_start_reserved() before 421 * it can be used. 422 * 423 * Return a pointer to a newly allocated handle, or an ERR_PTR() value 424 * on failure. 425 */ 426 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int rsv_blocks, 427 gfp_t gfp_mask, unsigned int type, 428 unsigned int line_no) 429 { 430 handle_t *handle = journal_current_handle(); 431 int err; 432 433 if (!journal) 434 return ERR_PTR(-EROFS); 435 436 if (handle) { 437 J_ASSERT(handle->h_transaction->t_journal == journal); 438 handle->h_ref++; 439 return handle; 440 } 441 442 handle = new_handle(nblocks); 443 if (!handle) 444 return ERR_PTR(-ENOMEM); 445 if (rsv_blocks) { 446 handle_t *rsv_handle; 447 448 rsv_handle = new_handle(rsv_blocks); 449 if (!rsv_handle) { 450 jbd2_free_handle(handle); 451 return ERR_PTR(-ENOMEM); 452 } 453 rsv_handle->h_reserved = 1; 454 rsv_handle->h_journal = journal; 455 handle->h_rsv_handle = rsv_handle; 456 } 457 458 err = start_this_handle(journal, handle, gfp_mask); 459 if (err < 0) { 460 if (handle->h_rsv_handle) 461 jbd2_free_handle(handle->h_rsv_handle); 462 jbd2_free_handle(handle); 463 return ERR_PTR(err); 464 } 465 handle->h_type = type; 466 handle->h_line_no = line_no; 467 trace_jbd2_handle_start(journal->j_fs_dev->bd_dev, 468 handle->h_transaction->t_tid, type, 469 line_no, nblocks); 470 return handle; 471 } 472 EXPORT_SYMBOL(jbd2__journal_start); 473 474 475 handle_t *jbd2_journal_start(journal_t *journal, int nblocks) 476 { 477 return jbd2__journal_start(journal, nblocks, 0, GFP_NOFS, 0, 0); 478 } 479 EXPORT_SYMBOL(jbd2_journal_start); 480 481 void jbd2_journal_free_reserved(handle_t *handle) 482 { 483 journal_t *journal = handle->h_journal; 484 485 WARN_ON(!handle->h_reserved); 486 sub_reserved_credits(journal, handle->h_buffer_credits); 487 jbd2_free_handle(handle); 488 } 489 EXPORT_SYMBOL(jbd2_journal_free_reserved); 490 491 /** 492 * int jbd2_journal_start_reserved(handle_t *handle) - start reserved handle 493 * @handle: handle to start 494 * 495 * Start handle that has been previously reserved with jbd2_journal_reserve(). 496 * This attaches @handle to the running transaction (or creates one if there's 497 * not transaction running). Unlike jbd2_journal_start() this function cannot 498 * block on journal commit, checkpointing, or similar stuff. It can block on 499 * memory allocation or frozen journal though. 500 * 501 * Return 0 on success, non-zero on error - handle is freed in that case. 502 */ 503 int jbd2_journal_start_reserved(handle_t *handle, unsigned int type, 504 unsigned int line_no) 505 { 506 journal_t *journal = handle->h_journal; 507 int ret = -EIO; 508 509 if (WARN_ON(!handle->h_reserved)) { 510 /* Someone passed in normal handle? Just stop it. */ 511 jbd2_journal_stop(handle); 512 return ret; 513 } 514 /* 515 * Usefulness of mixing of reserved and unreserved handles is 516 * questionable. So far nobody seems to need it so just error out. 517 */ 518 if (WARN_ON(current->journal_info)) { 519 jbd2_journal_free_reserved(handle); 520 return ret; 521 } 522 523 handle->h_journal = NULL; 524 /* 525 * GFP_NOFS is here because callers are likely from writeback or 526 * similarly constrained call sites 527 */ 528 ret = start_this_handle(journal, handle, GFP_NOFS); 529 if (ret < 0) { 530 jbd2_journal_free_reserved(handle); 531 return ret; 532 } 533 handle->h_type = type; 534 handle->h_line_no = line_no; 535 return 0; 536 } 537 EXPORT_SYMBOL(jbd2_journal_start_reserved); 538 539 /** 540 * int jbd2_journal_extend() - extend buffer credits. 541 * @handle: handle to 'extend' 542 * @nblocks: nr blocks to try to extend by. 543 * 544 * Some transactions, such as large extends and truncates, can be done 545 * atomically all at once or in several stages. The operation requests 546 * a credit for a number of buffer modications in advance, but can 547 * extend its credit if it needs more. 548 * 549 * jbd2_journal_extend tries to give the running handle more buffer credits. 550 * It does not guarantee that allocation - this is a best-effort only. 551 * The calling process MUST be able to deal cleanly with a failure to 552 * extend here. 553 * 554 * Return 0 on success, non-zero on failure. 555 * 556 * return code < 0 implies an error 557 * return code > 0 implies normal transaction-full status. 558 */ 559 int jbd2_journal_extend(handle_t *handle, int nblocks) 560 { 561 transaction_t *transaction = handle->h_transaction; 562 journal_t *journal; 563 int result; 564 int wanted; 565 566 if (is_handle_aborted(handle)) 567 return -EROFS; 568 journal = transaction->t_journal; 569 570 result = 1; 571 572 read_lock(&journal->j_state_lock); 573 574 /* Don't extend a locked-down transaction! */ 575 if (transaction->t_state != T_RUNNING) { 576 jbd_debug(3, "denied handle %p %d blocks: " 577 "transaction not running\n", handle, nblocks); 578 goto error_out; 579 } 580 581 spin_lock(&transaction->t_handle_lock); 582 wanted = atomic_add_return(nblocks, 583 &transaction->t_outstanding_credits); 584 585 if (wanted > journal->j_max_transaction_buffers) { 586 jbd_debug(3, "denied handle %p %d blocks: " 587 "transaction too large\n", handle, nblocks); 588 atomic_sub(nblocks, &transaction->t_outstanding_credits); 589 goto unlock; 590 } 591 592 if (wanted + (wanted >> JBD2_CONTROL_BLOCKS_SHIFT) > 593 jbd2_log_space_left(journal)) { 594 jbd_debug(3, "denied handle %p %d blocks: " 595 "insufficient log space\n", handle, nblocks); 596 atomic_sub(nblocks, &transaction->t_outstanding_credits); 597 goto unlock; 598 } 599 600 trace_jbd2_handle_extend(journal->j_fs_dev->bd_dev, 601 transaction->t_tid, 602 handle->h_type, handle->h_line_no, 603 handle->h_buffer_credits, 604 nblocks); 605 606 handle->h_buffer_credits += nblocks; 607 handle->h_requested_credits += nblocks; 608 result = 0; 609 610 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks); 611 unlock: 612 spin_unlock(&transaction->t_handle_lock); 613 error_out: 614 read_unlock(&journal->j_state_lock); 615 return result; 616 } 617 618 619 /** 620 * int jbd2_journal_restart() - restart a handle . 621 * @handle: handle to restart 622 * @nblocks: nr credits requested 623 * 624 * Restart a handle for a multi-transaction filesystem 625 * operation. 626 * 627 * If the jbd2_journal_extend() call above fails to grant new buffer credits 628 * to a running handle, a call to jbd2_journal_restart will commit the 629 * handle's transaction so far and reattach the handle to a new 630 * transaction capabable of guaranteeing the requested number of 631 * credits. We preserve reserved handle if there's any attached to the 632 * passed in handle. 633 */ 634 int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask) 635 { 636 transaction_t *transaction = handle->h_transaction; 637 journal_t *journal; 638 tid_t tid; 639 int need_to_start, ret; 640 641 /* If we've had an abort of any type, don't even think about 642 * actually doing the restart! */ 643 if (is_handle_aborted(handle)) 644 return 0; 645 journal = transaction->t_journal; 646 647 /* 648 * First unlink the handle from its current transaction, and start the 649 * commit on that. 650 */ 651 J_ASSERT(atomic_read(&transaction->t_updates) > 0); 652 J_ASSERT(journal_current_handle() == handle); 653 654 read_lock(&journal->j_state_lock); 655 spin_lock(&transaction->t_handle_lock); 656 atomic_sub(handle->h_buffer_credits, 657 &transaction->t_outstanding_credits); 658 if (handle->h_rsv_handle) { 659 sub_reserved_credits(journal, 660 handle->h_rsv_handle->h_buffer_credits); 661 } 662 if (atomic_dec_and_test(&transaction->t_updates)) 663 wake_up(&journal->j_wait_updates); 664 tid = transaction->t_tid; 665 spin_unlock(&transaction->t_handle_lock); 666 handle->h_transaction = NULL; 667 current->journal_info = NULL; 668 669 jbd_debug(2, "restarting handle %p\n", handle); 670 need_to_start = !tid_geq(journal->j_commit_request, tid); 671 read_unlock(&journal->j_state_lock); 672 if (need_to_start) 673 jbd2_log_start_commit(journal, tid); 674 675 lock_map_release(&handle->h_lockdep_map); 676 handle->h_buffer_credits = nblocks; 677 ret = start_this_handle(journal, handle, gfp_mask); 678 return ret; 679 } 680 EXPORT_SYMBOL(jbd2__journal_restart); 681 682 683 int jbd2_journal_restart(handle_t *handle, int nblocks) 684 { 685 return jbd2__journal_restart(handle, nblocks, GFP_NOFS); 686 } 687 EXPORT_SYMBOL(jbd2_journal_restart); 688 689 /** 690 * void jbd2_journal_lock_updates () - establish a transaction barrier. 691 * @journal: Journal to establish a barrier on. 692 * 693 * This locks out any further updates from being started, and blocks 694 * until all existing updates have completed, returning only once the 695 * journal is in a quiescent state with no updates running. 696 * 697 * The journal lock should not be held on entry. 698 */ 699 void jbd2_journal_lock_updates(journal_t *journal) 700 { 701 DEFINE_WAIT(wait); 702 703 write_lock(&journal->j_state_lock); 704 ++journal->j_barrier_count; 705 706 /* Wait until there are no reserved handles */ 707 if (atomic_read(&journal->j_reserved_credits)) { 708 write_unlock(&journal->j_state_lock); 709 wait_event(journal->j_wait_reserved, 710 atomic_read(&journal->j_reserved_credits) == 0); 711 write_lock(&journal->j_state_lock); 712 } 713 714 /* Wait until there are no running updates */ 715 while (1) { 716 transaction_t *transaction = journal->j_running_transaction; 717 718 if (!transaction) 719 break; 720 721 spin_lock(&transaction->t_handle_lock); 722 prepare_to_wait(&journal->j_wait_updates, &wait, 723 TASK_UNINTERRUPTIBLE); 724 if (!atomic_read(&transaction->t_updates)) { 725 spin_unlock(&transaction->t_handle_lock); 726 finish_wait(&journal->j_wait_updates, &wait); 727 break; 728 } 729 spin_unlock(&transaction->t_handle_lock); 730 write_unlock(&journal->j_state_lock); 731 schedule(); 732 finish_wait(&journal->j_wait_updates, &wait); 733 write_lock(&journal->j_state_lock); 734 } 735 write_unlock(&journal->j_state_lock); 736 737 /* 738 * We have now established a barrier against other normal updates, but 739 * we also need to barrier against other jbd2_journal_lock_updates() calls 740 * to make sure that we serialise special journal-locked operations 741 * too. 742 */ 743 mutex_lock(&journal->j_barrier); 744 } 745 746 /** 747 * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier 748 * @journal: Journal to release the barrier on. 749 * 750 * Release a transaction barrier obtained with jbd2_journal_lock_updates(). 751 * 752 * Should be called without the journal lock held. 753 */ 754 void jbd2_journal_unlock_updates (journal_t *journal) 755 { 756 J_ASSERT(journal->j_barrier_count != 0); 757 758 mutex_unlock(&journal->j_barrier); 759 write_lock(&journal->j_state_lock); 760 --journal->j_barrier_count; 761 write_unlock(&journal->j_state_lock); 762 wake_up(&journal->j_wait_transaction_locked); 763 } 764 765 static void warn_dirty_buffer(struct buffer_head *bh) 766 { 767 char b[BDEVNAME_SIZE]; 768 769 printk(KERN_WARNING 770 "JBD2: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). " 771 "There's a risk of filesystem corruption in case of system " 772 "crash.\n", 773 bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr); 774 } 775 776 /* Call t_frozen trigger and copy buffer data into jh->b_frozen_data. */ 777 static void jbd2_freeze_jh_data(struct journal_head *jh) 778 { 779 struct page *page; 780 int offset; 781 char *source; 782 struct buffer_head *bh = jh2bh(jh); 783 784 J_EXPECT_JH(jh, buffer_uptodate(bh), "Possible IO failure.\n"); 785 page = bh->b_page; 786 offset = offset_in_page(bh->b_data); 787 source = kmap_atomic(page); 788 /* Fire data frozen trigger just before we copy the data */ 789 jbd2_buffer_frozen_trigger(jh, source + offset, jh->b_triggers); 790 memcpy(jh->b_frozen_data, source + offset, bh->b_size); 791 kunmap_atomic(source); 792 793 /* 794 * Now that the frozen data is saved off, we need to store any matching 795 * triggers. 796 */ 797 jh->b_frozen_triggers = jh->b_triggers; 798 } 799 800 /* 801 * If the buffer is already part of the current transaction, then there 802 * is nothing we need to do. If it is already part of a prior 803 * transaction which we are still committing to disk, then we need to 804 * make sure that we do not overwrite the old copy: we do copy-out to 805 * preserve the copy going to disk. We also account the buffer against 806 * the handle's metadata buffer credits (unless the buffer is already 807 * part of the transaction, that is). 808 * 809 */ 810 static int 811 do_get_write_access(handle_t *handle, struct journal_head *jh, 812 int force_copy) 813 { 814 struct buffer_head *bh; 815 transaction_t *transaction = handle->h_transaction; 816 journal_t *journal; 817 int error; 818 char *frozen_buffer = NULL; 819 unsigned long start_lock, time_lock; 820 821 if (is_handle_aborted(handle)) 822 return -EROFS; 823 journal = transaction->t_journal; 824 825 jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy); 826 827 JBUFFER_TRACE(jh, "entry"); 828 repeat: 829 bh = jh2bh(jh); 830 831 /* @@@ Need to check for errors here at some point. */ 832 833 start_lock = jiffies; 834 lock_buffer(bh); 835 jbd_lock_bh_state(bh); 836 837 /* If it takes too long to lock the buffer, trace it */ 838 time_lock = jbd2_time_diff(start_lock, jiffies); 839 if (time_lock > HZ/10) 840 trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev, 841 jiffies_to_msecs(time_lock)); 842 843 /* We now hold the buffer lock so it is safe to query the buffer 844 * state. Is the buffer dirty? 845 * 846 * If so, there are two possibilities. The buffer may be 847 * non-journaled, and undergoing a quite legitimate writeback. 848 * Otherwise, it is journaled, and we don't expect dirty buffers 849 * in that state (the buffers should be marked JBD_Dirty 850 * instead.) So either the IO is being done under our own 851 * control and this is a bug, or it's a third party IO such as 852 * dump(8) (which may leave the buffer scheduled for read --- 853 * ie. locked but not dirty) or tune2fs (which may actually have 854 * the buffer dirtied, ugh.) */ 855 856 if (buffer_dirty(bh)) { 857 /* 858 * First question: is this buffer already part of the current 859 * transaction or the existing committing transaction? 860 */ 861 if (jh->b_transaction) { 862 J_ASSERT_JH(jh, 863 jh->b_transaction == transaction || 864 jh->b_transaction == 865 journal->j_committing_transaction); 866 if (jh->b_next_transaction) 867 J_ASSERT_JH(jh, jh->b_next_transaction == 868 transaction); 869 warn_dirty_buffer(bh); 870 } 871 /* 872 * In any case we need to clean the dirty flag and we must 873 * do it under the buffer lock to be sure we don't race 874 * with running write-out. 875 */ 876 JBUFFER_TRACE(jh, "Journalling dirty buffer"); 877 clear_buffer_dirty(bh); 878 set_buffer_jbddirty(bh); 879 } 880 881 unlock_buffer(bh); 882 883 error = -EROFS; 884 if (is_handle_aborted(handle)) { 885 jbd_unlock_bh_state(bh); 886 goto out; 887 } 888 error = 0; 889 890 /* 891 * The buffer is already part of this transaction if b_transaction or 892 * b_next_transaction points to it 893 */ 894 if (jh->b_transaction == transaction || 895 jh->b_next_transaction == transaction) 896 goto done; 897 898 /* 899 * this is the first time this transaction is touching this buffer, 900 * reset the modified flag 901 */ 902 jh->b_modified = 0; 903 904 /* 905 * If the buffer is not journaled right now, we need to make sure it 906 * doesn't get written to disk before the caller actually commits the 907 * new data 908 */ 909 if (!jh->b_transaction) { 910 JBUFFER_TRACE(jh, "no transaction"); 911 J_ASSERT_JH(jh, !jh->b_next_transaction); 912 JBUFFER_TRACE(jh, "file as BJ_Reserved"); 913 /* 914 * Make sure all stores to jh (b_modified, b_frozen_data) are 915 * visible before attaching it to the running transaction. 916 * Paired with barrier in jbd2_write_access_granted() 917 */ 918 smp_wmb(); 919 spin_lock(&journal->j_list_lock); 920 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved); 921 spin_unlock(&journal->j_list_lock); 922 goto done; 923 } 924 /* 925 * If there is already a copy-out version of this buffer, then we don't 926 * need to make another one 927 */ 928 if (jh->b_frozen_data) { 929 JBUFFER_TRACE(jh, "has frozen data"); 930 J_ASSERT_JH(jh, jh->b_next_transaction == NULL); 931 goto attach_next; 932 } 933 934 JBUFFER_TRACE(jh, "owned by older transaction"); 935 J_ASSERT_JH(jh, jh->b_next_transaction == NULL); 936 J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction); 937 938 /* 939 * There is one case we have to be very careful about. If the 940 * committing transaction is currently writing this buffer out to disk 941 * and has NOT made a copy-out, then we cannot modify the buffer 942 * contents at all right now. The essence of copy-out is that it is 943 * the extra copy, not the primary copy, which gets journaled. If the 944 * primary copy is already going to disk then we cannot do copy-out 945 * here. 946 */ 947 if (buffer_shadow(bh)) { 948 JBUFFER_TRACE(jh, "on shadow: sleep"); 949 jbd_unlock_bh_state(bh); 950 wait_on_bit_io(&bh->b_state, BH_Shadow, TASK_UNINTERRUPTIBLE); 951 goto repeat; 952 } 953 954 /* 955 * Only do the copy if the currently-owning transaction still needs it. 956 * If buffer isn't on BJ_Metadata list, the committing transaction is 957 * past that stage (here we use the fact that BH_Shadow is set under 958 * bh_state lock together with refiling to BJ_Shadow list and at this 959 * point we know the buffer doesn't have BH_Shadow set). 960 * 961 * Subtle point, though: if this is a get_undo_access, then we will be 962 * relying on the frozen_data to contain the new value of the 963 * committed_data record after the transaction, so we HAVE to force the 964 * frozen_data copy in that case. 965 */ 966 if (jh->b_jlist == BJ_Metadata || force_copy) { 967 JBUFFER_TRACE(jh, "generate frozen data"); 968 if (!frozen_buffer) { 969 JBUFFER_TRACE(jh, "allocate memory for buffer"); 970 jbd_unlock_bh_state(bh); 971 frozen_buffer = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS); 972 if (!frozen_buffer) { 973 printk(KERN_ERR "%s: OOM for frozen_buffer\n", 974 __func__); 975 JBUFFER_TRACE(jh, "oom!"); 976 error = -ENOMEM; 977 goto out; 978 } 979 goto repeat; 980 } 981 jh->b_frozen_data = frozen_buffer; 982 frozen_buffer = NULL; 983 jbd2_freeze_jh_data(jh); 984 } 985 attach_next: 986 /* 987 * Make sure all stores to jh (b_modified, b_frozen_data) are visible 988 * before attaching it to the running transaction. Paired with barrier 989 * in jbd2_write_access_granted() 990 */ 991 smp_wmb(); 992 jh->b_next_transaction = transaction; 993 994 done: 995 jbd_unlock_bh_state(bh); 996 997 /* 998 * If we are about to journal a buffer, then any revoke pending on it is 999 * no longer valid 1000 */ 1001 jbd2_journal_cancel_revoke(handle, jh); 1002 1003 out: 1004 if (unlikely(frozen_buffer)) /* It's usually NULL */ 1005 jbd2_free(frozen_buffer, bh->b_size); 1006 1007 JBUFFER_TRACE(jh, "exit"); 1008 return error; 1009 } 1010 1011 /* Fast check whether buffer is already attached to the required transaction */ 1012 static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh, 1013 bool undo) 1014 { 1015 struct journal_head *jh; 1016 bool ret = false; 1017 1018 /* Dirty buffers require special handling... */ 1019 if (buffer_dirty(bh)) 1020 return false; 1021 1022 /* 1023 * RCU protects us from dereferencing freed pages. So the checks we do 1024 * are guaranteed not to oops. However the jh slab object can get freed 1025 * & reallocated while we work with it. So we have to be careful. When 1026 * we see jh attached to the running transaction, we know it must stay 1027 * so until the transaction is committed. Thus jh won't be freed and 1028 * will be attached to the same bh while we run. However it can 1029 * happen jh gets freed, reallocated, and attached to the transaction 1030 * just after we get pointer to it from bh. So we have to be careful 1031 * and recheck jh still belongs to our bh before we return success. 1032 */ 1033 rcu_read_lock(); 1034 if (!buffer_jbd(bh)) 1035 goto out; 1036 /* This should be bh2jh() but that doesn't work with inline functions */ 1037 jh = READ_ONCE(bh->b_private); 1038 if (!jh) 1039 goto out; 1040 /* For undo access buffer must have data copied */ 1041 if (undo && !jh->b_committed_data) 1042 goto out; 1043 if (jh->b_transaction != handle->h_transaction && 1044 jh->b_next_transaction != handle->h_transaction) 1045 goto out; 1046 /* 1047 * There are two reasons for the barrier here: 1048 * 1) Make sure to fetch b_bh after we did previous checks so that we 1049 * detect when jh went through free, realloc, attach to transaction 1050 * while we were checking. Paired with implicit barrier in that path. 1051 * 2) So that access to bh done after jbd2_write_access_granted() 1052 * doesn't get reordered and see inconsistent state of concurrent 1053 * do_get_write_access(). 1054 */ 1055 smp_mb(); 1056 if (unlikely(jh->b_bh != bh)) 1057 goto out; 1058 ret = true; 1059 out: 1060 rcu_read_unlock(); 1061 return ret; 1062 } 1063 1064 /** 1065 * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update. 1066 * @handle: transaction to add buffer modifications to 1067 * @bh: bh to be used for metadata writes 1068 * 1069 * Returns an error code or 0 on success. 1070 * 1071 * In full data journalling mode the buffer may be of type BJ_AsyncData, 1072 * because we're write()ing a buffer which is also part of a shared mapping. 1073 */ 1074 1075 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh) 1076 { 1077 struct journal_head *jh; 1078 int rc; 1079 1080 if (jbd2_write_access_granted(handle, bh, false)) 1081 return 0; 1082 1083 jh = jbd2_journal_add_journal_head(bh); 1084 /* We do not want to get caught playing with fields which the 1085 * log thread also manipulates. Make sure that the buffer 1086 * completes any outstanding IO before proceeding. */ 1087 rc = do_get_write_access(handle, jh, 0); 1088 jbd2_journal_put_journal_head(jh); 1089 return rc; 1090 } 1091 1092 1093 /* 1094 * When the user wants to journal a newly created buffer_head 1095 * (ie. getblk() returned a new buffer and we are going to populate it 1096 * manually rather than reading off disk), then we need to keep the 1097 * buffer_head locked until it has been completely filled with new 1098 * data. In this case, we should be able to make the assertion that 1099 * the bh is not already part of an existing transaction. 1100 * 1101 * The buffer should already be locked by the caller by this point. 1102 * There is no lock ranking violation: it was a newly created, 1103 * unlocked buffer beforehand. */ 1104 1105 /** 1106 * int jbd2_journal_get_create_access () - notify intent to use newly created bh 1107 * @handle: transaction to new buffer to 1108 * @bh: new buffer. 1109 * 1110 * Call this if you create a new bh. 1111 */ 1112 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh) 1113 { 1114 transaction_t *transaction = handle->h_transaction; 1115 journal_t *journal; 1116 struct journal_head *jh = jbd2_journal_add_journal_head(bh); 1117 int err; 1118 1119 jbd_debug(5, "journal_head %p\n", jh); 1120 err = -EROFS; 1121 if (is_handle_aborted(handle)) 1122 goto out; 1123 journal = transaction->t_journal; 1124 err = 0; 1125 1126 JBUFFER_TRACE(jh, "entry"); 1127 /* 1128 * The buffer may already belong to this transaction due to pre-zeroing 1129 * in the filesystem's new_block code. It may also be on the previous, 1130 * committing transaction's lists, but it HAS to be in Forget state in 1131 * that case: the transaction must have deleted the buffer for it to be 1132 * reused here. 1133 */ 1134 jbd_lock_bh_state(bh); 1135 J_ASSERT_JH(jh, (jh->b_transaction == transaction || 1136 jh->b_transaction == NULL || 1137 (jh->b_transaction == journal->j_committing_transaction && 1138 jh->b_jlist == BJ_Forget))); 1139 1140 J_ASSERT_JH(jh, jh->b_next_transaction == NULL); 1141 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh))); 1142 1143 if (jh->b_transaction == NULL) { 1144 /* 1145 * Previous jbd2_journal_forget() could have left the buffer 1146 * with jbddirty bit set because it was being committed. When 1147 * the commit finished, we've filed the buffer for 1148 * checkpointing and marked it dirty. Now we are reallocating 1149 * the buffer so the transaction freeing it must have 1150 * committed and so it's safe to clear the dirty bit. 1151 */ 1152 clear_buffer_dirty(jh2bh(jh)); 1153 /* first access by this transaction */ 1154 jh->b_modified = 0; 1155 1156 JBUFFER_TRACE(jh, "file as BJ_Reserved"); 1157 spin_lock(&journal->j_list_lock); 1158 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved); 1159 } else if (jh->b_transaction == journal->j_committing_transaction) { 1160 /* first access by this transaction */ 1161 jh->b_modified = 0; 1162 1163 JBUFFER_TRACE(jh, "set next transaction"); 1164 spin_lock(&journal->j_list_lock); 1165 jh->b_next_transaction = transaction; 1166 } 1167 spin_unlock(&journal->j_list_lock); 1168 jbd_unlock_bh_state(bh); 1169 1170 /* 1171 * akpm: I added this. ext3_alloc_branch can pick up new indirect 1172 * blocks which contain freed but then revoked metadata. We need 1173 * to cancel the revoke in case we end up freeing it yet again 1174 * and the reallocating as data - this would cause a second revoke, 1175 * which hits an assertion error. 1176 */ 1177 JBUFFER_TRACE(jh, "cancelling revoke"); 1178 jbd2_journal_cancel_revoke(handle, jh); 1179 out: 1180 jbd2_journal_put_journal_head(jh); 1181 return err; 1182 } 1183 1184 /** 1185 * int jbd2_journal_get_undo_access() - Notify intent to modify metadata with 1186 * non-rewindable consequences 1187 * @handle: transaction 1188 * @bh: buffer to undo 1189 * 1190 * Sometimes there is a need to distinguish between metadata which has 1191 * been committed to disk and that which has not. The ext3fs code uses 1192 * this for freeing and allocating space, we have to make sure that we 1193 * do not reuse freed space until the deallocation has been committed, 1194 * since if we overwrote that space we would make the delete 1195 * un-rewindable in case of a crash. 1196 * 1197 * To deal with that, jbd2_journal_get_undo_access requests write access to a 1198 * buffer for parts of non-rewindable operations such as delete 1199 * operations on the bitmaps. The journaling code must keep a copy of 1200 * the buffer's contents prior to the undo_access call until such time 1201 * as we know that the buffer has definitely been committed to disk. 1202 * 1203 * We never need to know which transaction the committed data is part 1204 * of, buffers touched here are guaranteed to be dirtied later and so 1205 * will be committed to a new transaction in due course, at which point 1206 * we can discard the old committed data pointer. 1207 * 1208 * Returns error number or 0 on success. 1209 */ 1210 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh) 1211 { 1212 int err; 1213 struct journal_head *jh; 1214 char *committed_data = NULL; 1215 1216 JBUFFER_TRACE(jh, "entry"); 1217 if (jbd2_write_access_granted(handle, bh, true)) 1218 return 0; 1219 1220 jh = jbd2_journal_add_journal_head(bh); 1221 /* 1222 * Do this first --- it can drop the journal lock, so we want to 1223 * make sure that obtaining the committed_data is done 1224 * atomically wrt. completion of any outstanding commits. 1225 */ 1226 err = do_get_write_access(handle, jh, 1); 1227 if (err) 1228 goto out; 1229 1230 repeat: 1231 if (!jh->b_committed_data) { 1232 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS); 1233 if (!committed_data) { 1234 printk(KERN_ERR "%s: No memory for committed data\n", 1235 __func__); 1236 err = -ENOMEM; 1237 goto out; 1238 } 1239 } 1240 1241 jbd_lock_bh_state(bh); 1242 if (!jh->b_committed_data) { 1243 /* Copy out the current buffer contents into the 1244 * preserved, committed copy. */ 1245 JBUFFER_TRACE(jh, "generate b_committed data"); 1246 if (!committed_data) { 1247 jbd_unlock_bh_state(bh); 1248 goto repeat; 1249 } 1250 1251 jh->b_committed_data = committed_data; 1252 committed_data = NULL; 1253 memcpy(jh->b_committed_data, bh->b_data, bh->b_size); 1254 } 1255 jbd_unlock_bh_state(bh); 1256 out: 1257 jbd2_journal_put_journal_head(jh); 1258 if (unlikely(committed_data)) 1259 jbd2_free(committed_data, bh->b_size); 1260 return err; 1261 } 1262 1263 /** 1264 * void jbd2_journal_set_triggers() - Add triggers for commit writeout 1265 * @bh: buffer to trigger on 1266 * @type: struct jbd2_buffer_trigger_type containing the trigger(s). 1267 * 1268 * Set any triggers on this journal_head. This is always safe, because 1269 * triggers for a committing buffer will be saved off, and triggers for 1270 * a running transaction will match the buffer in that transaction. 1271 * 1272 * Call with NULL to clear the triggers. 1273 */ 1274 void jbd2_journal_set_triggers(struct buffer_head *bh, 1275 struct jbd2_buffer_trigger_type *type) 1276 { 1277 struct journal_head *jh = jbd2_journal_grab_journal_head(bh); 1278 1279 if (WARN_ON(!jh)) 1280 return; 1281 jh->b_triggers = type; 1282 jbd2_journal_put_journal_head(jh); 1283 } 1284 1285 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data, 1286 struct jbd2_buffer_trigger_type *triggers) 1287 { 1288 struct buffer_head *bh = jh2bh(jh); 1289 1290 if (!triggers || !triggers->t_frozen) 1291 return; 1292 1293 triggers->t_frozen(triggers, bh, mapped_data, bh->b_size); 1294 } 1295 1296 void jbd2_buffer_abort_trigger(struct journal_head *jh, 1297 struct jbd2_buffer_trigger_type *triggers) 1298 { 1299 if (!triggers || !triggers->t_abort) 1300 return; 1301 1302 triggers->t_abort(triggers, jh2bh(jh)); 1303 } 1304 1305 /** 1306 * int jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata 1307 * @handle: transaction to add buffer to. 1308 * @bh: buffer to mark 1309 * 1310 * mark dirty metadata which needs to be journaled as part of the current 1311 * transaction. 1312 * 1313 * The buffer must have previously had jbd2_journal_get_write_access() 1314 * called so that it has a valid journal_head attached to the buffer 1315 * head. 1316 * 1317 * The buffer is placed on the transaction's metadata list and is marked 1318 * as belonging to the transaction. 1319 * 1320 * Returns error number or 0 on success. 1321 * 1322 * Special care needs to be taken if the buffer already belongs to the 1323 * current committing transaction (in which case we should have frozen 1324 * data present for that commit). In that case, we don't relink the 1325 * buffer: that only gets done when the old transaction finally 1326 * completes its commit. 1327 */ 1328 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh) 1329 { 1330 transaction_t *transaction = handle->h_transaction; 1331 journal_t *journal; 1332 struct journal_head *jh; 1333 int ret = 0; 1334 1335 if (is_handle_aborted(handle)) 1336 return -EROFS; 1337 if (!buffer_jbd(bh)) { 1338 ret = -EUCLEAN; 1339 goto out; 1340 } 1341 /* 1342 * We don't grab jh reference here since the buffer must be part 1343 * of the running transaction. 1344 */ 1345 jh = bh2jh(bh); 1346 /* 1347 * This and the following assertions are unreliable since we may see jh 1348 * in inconsistent state unless we grab bh_state lock. But this is 1349 * crucial to catch bugs so let's do a reliable check until the 1350 * lockless handling is fully proven. 1351 */ 1352 if (jh->b_transaction != transaction && 1353 jh->b_next_transaction != transaction) { 1354 jbd_lock_bh_state(bh); 1355 J_ASSERT_JH(jh, jh->b_transaction == transaction || 1356 jh->b_next_transaction == transaction); 1357 jbd_unlock_bh_state(bh); 1358 } 1359 if (jh->b_modified == 1) { 1360 /* If it's in our transaction it must be in BJ_Metadata list. */ 1361 if (jh->b_transaction == transaction && 1362 jh->b_jlist != BJ_Metadata) { 1363 jbd_lock_bh_state(bh); 1364 J_ASSERT_JH(jh, jh->b_transaction != transaction || 1365 jh->b_jlist == BJ_Metadata); 1366 jbd_unlock_bh_state(bh); 1367 } 1368 goto out; 1369 } 1370 1371 journal = transaction->t_journal; 1372 jbd_debug(5, "journal_head %p\n", jh); 1373 JBUFFER_TRACE(jh, "entry"); 1374 1375 jbd_lock_bh_state(bh); 1376 1377 if (jh->b_modified == 0) { 1378 /* 1379 * This buffer's got modified and becoming part 1380 * of the transaction. This needs to be done 1381 * once a transaction -bzzz 1382 */ 1383 jh->b_modified = 1; 1384 if (handle->h_buffer_credits <= 0) { 1385 ret = -ENOSPC; 1386 goto out_unlock_bh; 1387 } 1388 handle->h_buffer_credits--; 1389 } 1390 1391 /* 1392 * fastpath, to avoid expensive locking. If this buffer is already 1393 * on the running transaction's metadata list there is nothing to do. 1394 * Nobody can take it off again because there is a handle open. 1395 * I _think_ we're OK here with SMP barriers - a mistaken decision will 1396 * result in this test being false, so we go in and take the locks. 1397 */ 1398 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) { 1399 JBUFFER_TRACE(jh, "fastpath"); 1400 if (unlikely(jh->b_transaction != 1401 journal->j_running_transaction)) { 1402 printk(KERN_ERR "JBD2: %s: " 1403 "jh->b_transaction (%llu, %p, %u) != " 1404 "journal->j_running_transaction (%p, %u)\n", 1405 journal->j_devname, 1406 (unsigned long long) bh->b_blocknr, 1407 jh->b_transaction, 1408 jh->b_transaction ? jh->b_transaction->t_tid : 0, 1409 journal->j_running_transaction, 1410 journal->j_running_transaction ? 1411 journal->j_running_transaction->t_tid : 0); 1412 ret = -EINVAL; 1413 } 1414 goto out_unlock_bh; 1415 } 1416 1417 set_buffer_jbddirty(bh); 1418 1419 /* 1420 * Metadata already on the current transaction list doesn't 1421 * need to be filed. Metadata on another transaction's list must 1422 * be committing, and will be refiled once the commit completes: 1423 * leave it alone for now. 1424 */ 1425 if (jh->b_transaction != transaction) { 1426 JBUFFER_TRACE(jh, "already on other transaction"); 1427 if (unlikely(((jh->b_transaction != 1428 journal->j_committing_transaction)) || 1429 (jh->b_next_transaction != transaction))) { 1430 printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: " 1431 "bad jh for block %llu: " 1432 "transaction (%p, %u), " 1433 "jh->b_transaction (%p, %u), " 1434 "jh->b_next_transaction (%p, %u), jlist %u\n", 1435 journal->j_devname, 1436 (unsigned long long) bh->b_blocknr, 1437 transaction, transaction->t_tid, 1438 jh->b_transaction, 1439 jh->b_transaction ? 1440 jh->b_transaction->t_tid : 0, 1441 jh->b_next_transaction, 1442 jh->b_next_transaction ? 1443 jh->b_next_transaction->t_tid : 0, 1444 jh->b_jlist); 1445 WARN_ON(1); 1446 ret = -EINVAL; 1447 } 1448 /* And this case is illegal: we can't reuse another 1449 * transaction's data buffer, ever. */ 1450 goto out_unlock_bh; 1451 } 1452 1453 /* That test should have eliminated the following case: */ 1454 J_ASSERT_JH(jh, jh->b_frozen_data == NULL); 1455 1456 JBUFFER_TRACE(jh, "file as BJ_Metadata"); 1457 spin_lock(&journal->j_list_lock); 1458 __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata); 1459 spin_unlock(&journal->j_list_lock); 1460 out_unlock_bh: 1461 jbd_unlock_bh_state(bh); 1462 out: 1463 JBUFFER_TRACE(jh, "exit"); 1464 return ret; 1465 } 1466 1467 /** 1468 * void jbd2_journal_forget() - bforget() for potentially-journaled buffers. 1469 * @handle: transaction handle 1470 * @bh: bh to 'forget' 1471 * 1472 * We can only do the bforget if there are no commits pending against the 1473 * buffer. If the buffer is dirty in the current running transaction we 1474 * can safely unlink it. 1475 * 1476 * bh may not be a journalled buffer at all - it may be a non-JBD 1477 * buffer which came off the hashtable. Check for this. 1478 * 1479 * Decrements bh->b_count by one. 1480 * 1481 * Allow this call even if the handle has aborted --- it may be part of 1482 * the caller's cleanup after an abort. 1483 */ 1484 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh) 1485 { 1486 transaction_t *transaction = handle->h_transaction; 1487 journal_t *journal; 1488 struct journal_head *jh; 1489 int drop_reserve = 0; 1490 int err = 0; 1491 int was_modified = 0; 1492 1493 if (is_handle_aborted(handle)) 1494 return -EROFS; 1495 journal = transaction->t_journal; 1496 1497 BUFFER_TRACE(bh, "entry"); 1498 1499 jbd_lock_bh_state(bh); 1500 1501 if (!buffer_jbd(bh)) 1502 goto not_jbd; 1503 jh = bh2jh(bh); 1504 1505 /* Critical error: attempting to delete a bitmap buffer, maybe? 1506 * Don't do any jbd operations, and return an error. */ 1507 if (!J_EXPECT_JH(jh, !jh->b_committed_data, 1508 "inconsistent data on disk")) { 1509 err = -EIO; 1510 goto not_jbd; 1511 } 1512 1513 /* keep track of whether or not this transaction modified us */ 1514 was_modified = jh->b_modified; 1515 1516 /* 1517 * The buffer's going from the transaction, we must drop 1518 * all references -bzzz 1519 */ 1520 jh->b_modified = 0; 1521 1522 if (jh->b_transaction == transaction) { 1523 J_ASSERT_JH(jh, !jh->b_frozen_data); 1524 1525 /* If we are forgetting a buffer which is already part 1526 * of this transaction, then we can just drop it from 1527 * the transaction immediately. */ 1528 clear_buffer_dirty(bh); 1529 clear_buffer_jbddirty(bh); 1530 1531 JBUFFER_TRACE(jh, "belongs to current transaction: unfile"); 1532 1533 /* 1534 * we only want to drop a reference if this transaction 1535 * modified the buffer 1536 */ 1537 if (was_modified) 1538 drop_reserve = 1; 1539 1540 /* 1541 * We are no longer going to journal this buffer. 1542 * However, the commit of this transaction is still 1543 * important to the buffer: the delete that we are now 1544 * processing might obsolete an old log entry, so by 1545 * committing, we can satisfy the buffer's checkpoint. 1546 * 1547 * So, if we have a checkpoint on the buffer, we should 1548 * now refile the buffer on our BJ_Forget list so that 1549 * we know to remove the checkpoint after we commit. 1550 */ 1551 1552 spin_lock(&journal->j_list_lock); 1553 if (jh->b_cp_transaction) { 1554 __jbd2_journal_temp_unlink_buffer(jh); 1555 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget); 1556 } else { 1557 __jbd2_journal_unfile_buffer(jh); 1558 if (!buffer_jbd(bh)) { 1559 spin_unlock(&journal->j_list_lock); 1560 jbd_unlock_bh_state(bh); 1561 __bforget(bh); 1562 goto drop; 1563 } 1564 } 1565 spin_unlock(&journal->j_list_lock); 1566 } else if (jh->b_transaction) { 1567 J_ASSERT_JH(jh, (jh->b_transaction == 1568 journal->j_committing_transaction)); 1569 /* However, if the buffer is still owned by a prior 1570 * (committing) transaction, we can't drop it yet... */ 1571 JBUFFER_TRACE(jh, "belongs to older transaction"); 1572 /* ... but we CAN drop it from the new transaction if we 1573 * have also modified it since the original commit. */ 1574 1575 if (jh->b_next_transaction) { 1576 J_ASSERT(jh->b_next_transaction == transaction); 1577 spin_lock(&journal->j_list_lock); 1578 jh->b_next_transaction = NULL; 1579 spin_unlock(&journal->j_list_lock); 1580 1581 /* 1582 * only drop a reference if this transaction modified 1583 * the buffer 1584 */ 1585 if (was_modified) 1586 drop_reserve = 1; 1587 } 1588 } 1589 1590 not_jbd: 1591 jbd_unlock_bh_state(bh); 1592 __brelse(bh); 1593 drop: 1594 if (drop_reserve) { 1595 /* no need to reserve log space for this block -bzzz */ 1596 handle->h_buffer_credits++; 1597 } 1598 return err; 1599 } 1600 1601 /** 1602 * int jbd2_journal_stop() - complete a transaction 1603 * @handle: tranaction to complete. 1604 * 1605 * All done for a particular handle. 1606 * 1607 * There is not much action needed here. We just return any remaining 1608 * buffer credits to the transaction and remove the handle. The only 1609 * complication is that we need to start a commit operation if the 1610 * filesystem is marked for synchronous update. 1611 * 1612 * jbd2_journal_stop itself will not usually return an error, but it may 1613 * do so in unusual circumstances. In particular, expect it to 1614 * return -EIO if a jbd2_journal_abort has been executed since the 1615 * transaction began. 1616 */ 1617 int jbd2_journal_stop(handle_t *handle) 1618 { 1619 transaction_t *transaction = handle->h_transaction; 1620 journal_t *journal; 1621 int err = 0, wait_for_commit = 0; 1622 tid_t tid; 1623 pid_t pid; 1624 1625 if (!transaction) { 1626 /* 1627 * Handle is already detached from the transaction so 1628 * there is nothing to do other than decrease a refcount, 1629 * or free the handle if refcount drops to zero 1630 */ 1631 if (--handle->h_ref > 0) { 1632 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1, 1633 handle->h_ref); 1634 return err; 1635 } else { 1636 if (handle->h_rsv_handle) 1637 jbd2_free_handle(handle->h_rsv_handle); 1638 goto free_and_exit; 1639 } 1640 } 1641 journal = transaction->t_journal; 1642 1643 J_ASSERT(journal_current_handle() == handle); 1644 1645 if (is_handle_aborted(handle)) 1646 err = -EIO; 1647 else 1648 J_ASSERT(atomic_read(&transaction->t_updates) > 0); 1649 1650 if (--handle->h_ref > 0) { 1651 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1, 1652 handle->h_ref); 1653 return err; 1654 } 1655 1656 jbd_debug(4, "Handle %p going down\n", handle); 1657 trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev, 1658 transaction->t_tid, 1659 handle->h_type, handle->h_line_no, 1660 jiffies - handle->h_start_jiffies, 1661 handle->h_sync, handle->h_requested_credits, 1662 (handle->h_requested_credits - 1663 handle->h_buffer_credits)); 1664 1665 /* 1666 * Implement synchronous transaction batching. If the handle 1667 * was synchronous, don't force a commit immediately. Let's 1668 * yield and let another thread piggyback onto this 1669 * transaction. Keep doing that while new threads continue to 1670 * arrive. It doesn't cost much - we're about to run a commit 1671 * and sleep on IO anyway. Speeds up many-threaded, many-dir 1672 * operations by 30x or more... 1673 * 1674 * We try and optimize the sleep time against what the 1675 * underlying disk can do, instead of having a static sleep 1676 * time. This is useful for the case where our storage is so 1677 * fast that it is more optimal to go ahead and force a flush 1678 * and wait for the transaction to be committed than it is to 1679 * wait for an arbitrary amount of time for new writers to 1680 * join the transaction. We achieve this by measuring how 1681 * long it takes to commit a transaction, and compare it with 1682 * how long this transaction has been running, and if run time 1683 * < commit time then we sleep for the delta and commit. This 1684 * greatly helps super fast disks that would see slowdowns as 1685 * more threads started doing fsyncs. 1686 * 1687 * But don't do this if this process was the most recent one 1688 * to perform a synchronous write. We do this to detect the 1689 * case where a single process is doing a stream of sync 1690 * writes. No point in waiting for joiners in that case. 1691 * 1692 * Setting max_batch_time to 0 disables this completely. 1693 */ 1694 pid = current->pid; 1695 if (handle->h_sync && journal->j_last_sync_writer != pid && 1696 journal->j_max_batch_time) { 1697 u64 commit_time, trans_time; 1698 1699 journal->j_last_sync_writer = pid; 1700 1701 read_lock(&journal->j_state_lock); 1702 commit_time = journal->j_average_commit_time; 1703 read_unlock(&journal->j_state_lock); 1704 1705 trans_time = ktime_to_ns(ktime_sub(ktime_get(), 1706 transaction->t_start_time)); 1707 1708 commit_time = max_t(u64, commit_time, 1709 1000*journal->j_min_batch_time); 1710 commit_time = min_t(u64, commit_time, 1711 1000*journal->j_max_batch_time); 1712 1713 if (trans_time < commit_time) { 1714 ktime_t expires = ktime_add_ns(ktime_get(), 1715 commit_time); 1716 set_current_state(TASK_UNINTERRUPTIBLE); 1717 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS); 1718 } 1719 } 1720 1721 if (handle->h_sync) 1722 transaction->t_synchronous_commit = 1; 1723 current->journal_info = NULL; 1724 atomic_sub(handle->h_buffer_credits, 1725 &transaction->t_outstanding_credits); 1726 1727 /* 1728 * If the handle is marked SYNC, we need to set another commit 1729 * going! We also want to force a commit if the current 1730 * transaction is occupying too much of the log, or if the 1731 * transaction is too old now. 1732 */ 1733 if (handle->h_sync || 1734 (atomic_read(&transaction->t_outstanding_credits) > 1735 journal->j_max_transaction_buffers) || 1736 time_after_eq(jiffies, transaction->t_expires)) { 1737 /* Do this even for aborted journals: an abort still 1738 * completes the commit thread, it just doesn't write 1739 * anything to disk. */ 1740 1741 jbd_debug(2, "transaction too old, requesting commit for " 1742 "handle %p\n", handle); 1743 /* This is non-blocking */ 1744 jbd2_log_start_commit(journal, transaction->t_tid); 1745 1746 /* 1747 * Special case: JBD2_SYNC synchronous updates require us 1748 * to wait for the commit to complete. 1749 */ 1750 if (handle->h_sync && !(current->flags & PF_MEMALLOC)) 1751 wait_for_commit = 1; 1752 } 1753 1754 /* 1755 * Once we drop t_updates, if it goes to zero the transaction 1756 * could start committing on us and eventually disappear. So 1757 * once we do this, we must not dereference transaction 1758 * pointer again. 1759 */ 1760 tid = transaction->t_tid; 1761 if (atomic_dec_and_test(&transaction->t_updates)) { 1762 wake_up(&journal->j_wait_updates); 1763 if (journal->j_barrier_count) 1764 wake_up(&journal->j_wait_transaction_locked); 1765 } 1766 1767 if (wait_for_commit) 1768 err = jbd2_log_wait_commit(journal, tid); 1769 1770 lock_map_release(&handle->h_lockdep_map); 1771 1772 if (handle->h_rsv_handle) 1773 jbd2_journal_free_reserved(handle->h_rsv_handle); 1774 free_and_exit: 1775 jbd2_free_handle(handle); 1776 return err; 1777 } 1778 1779 /* 1780 * 1781 * List management code snippets: various functions for manipulating the 1782 * transaction buffer lists. 1783 * 1784 */ 1785 1786 /* 1787 * Append a buffer to a transaction list, given the transaction's list head 1788 * pointer. 1789 * 1790 * j_list_lock is held. 1791 * 1792 * jbd_lock_bh_state(jh2bh(jh)) is held. 1793 */ 1794 1795 static inline void 1796 __blist_add_buffer(struct journal_head **list, struct journal_head *jh) 1797 { 1798 if (!*list) { 1799 jh->b_tnext = jh->b_tprev = jh; 1800 *list = jh; 1801 } else { 1802 /* Insert at the tail of the list to preserve order */ 1803 struct journal_head *first = *list, *last = first->b_tprev; 1804 jh->b_tprev = last; 1805 jh->b_tnext = first; 1806 last->b_tnext = first->b_tprev = jh; 1807 } 1808 } 1809 1810 /* 1811 * Remove a buffer from a transaction list, given the transaction's list 1812 * head pointer. 1813 * 1814 * Called with j_list_lock held, and the journal may not be locked. 1815 * 1816 * jbd_lock_bh_state(jh2bh(jh)) is held. 1817 */ 1818 1819 static inline void 1820 __blist_del_buffer(struct journal_head **list, struct journal_head *jh) 1821 { 1822 if (*list == jh) { 1823 *list = jh->b_tnext; 1824 if (*list == jh) 1825 *list = NULL; 1826 } 1827 jh->b_tprev->b_tnext = jh->b_tnext; 1828 jh->b_tnext->b_tprev = jh->b_tprev; 1829 } 1830 1831 /* 1832 * Remove a buffer from the appropriate transaction list. 1833 * 1834 * Note that this function can *change* the value of 1835 * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or 1836 * t_reserved_list. If the caller is holding onto a copy of one of these 1837 * pointers, it could go bad. Generally the caller needs to re-read the 1838 * pointer from the transaction_t. 1839 * 1840 * Called under j_list_lock. 1841 */ 1842 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh) 1843 { 1844 struct journal_head **list = NULL; 1845 transaction_t *transaction; 1846 struct buffer_head *bh = jh2bh(jh); 1847 1848 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh)); 1849 transaction = jh->b_transaction; 1850 if (transaction) 1851 assert_spin_locked(&transaction->t_journal->j_list_lock); 1852 1853 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types); 1854 if (jh->b_jlist != BJ_None) 1855 J_ASSERT_JH(jh, transaction != NULL); 1856 1857 switch (jh->b_jlist) { 1858 case BJ_None: 1859 return; 1860 case BJ_Metadata: 1861 transaction->t_nr_buffers--; 1862 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0); 1863 list = &transaction->t_buffers; 1864 break; 1865 case BJ_Forget: 1866 list = &transaction->t_forget; 1867 break; 1868 case BJ_Shadow: 1869 list = &transaction->t_shadow_list; 1870 break; 1871 case BJ_Reserved: 1872 list = &transaction->t_reserved_list; 1873 break; 1874 } 1875 1876 __blist_del_buffer(list, jh); 1877 jh->b_jlist = BJ_None; 1878 if (test_clear_buffer_jbddirty(bh)) 1879 mark_buffer_dirty(bh); /* Expose it to the VM */ 1880 } 1881 1882 /* 1883 * Remove buffer from all transactions. 1884 * 1885 * Called with bh_state lock and j_list_lock 1886 * 1887 * jh and bh may be already freed when this function returns. 1888 */ 1889 static void __jbd2_journal_unfile_buffer(struct journal_head *jh) 1890 { 1891 __jbd2_journal_temp_unlink_buffer(jh); 1892 jh->b_transaction = NULL; 1893 jbd2_journal_put_journal_head(jh); 1894 } 1895 1896 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh) 1897 { 1898 struct buffer_head *bh = jh2bh(jh); 1899 1900 /* Get reference so that buffer cannot be freed before we unlock it */ 1901 get_bh(bh); 1902 jbd_lock_bh_state(bh); 1903 spin_lock(&journal->j_list_lock); 1904 __jbd2_journal_unfile_buffer(jh); 1905 spin_unlock(&journal->j_list_lock); 1906 jbd_unlock_bh_state(bh); 1907 __brelse(bh); 1908 } 1909 1910 /* 1911 * Called from jbd2_journal_try_to_free_buffers(). 1912 * 1913 * Called under jbd_lock_bh_state(bh) 1914 */ 1915 static void 1916 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh) 1917 { 1918 struct journal_head *jh; 1919 1920 jh = bh2jh(bh); 1921 1922 if (buffer_locked(bh) || buffer_dirty(bh)) 1923 goto out; 1924 1925 if (jh->b_next_transaction != NULL || jh->b_transaction != NULL) 1926 goto out; 1927 1928 spin_lock(&journal->j_list_lock); 1929 if (jh->b_cp_transaction != NULL) { 1930 /* written-back checkpointed metadata buffer */ 1931 JBUFFER_TRACE(jh, "remove from checkpoint list"); 1932 __jbd2_journal_remove_checkpoint(jh); 1933 } 1934 spin_unlock(&journal->j_list_lock); 1935 out: 1936 return; 1937 } 1938 1939 /** 1940 * int jbd2_journal_try_to_free_buffers() - try to free page buffers. 1941 * @journal: journal for operation 1942 * @page: to try and free 1943 * @gfp_mask: we use the mask to detect how hard should we try to release 1944 * buffers. If __GFP_DIRECT_RECLAIM and __GFP_FS is set, we wait for commit 1945 * code to release the buffers. 1946 * 1947 * 1948 * For all the buffers on this page, 1949 * if they are fully written out ordered data, move them onto BUF_CLEAN 1950 * so try_to_free_buffers() can reap them. 1951 * 1952 * This function returns non-zero if we wish try_to_free_buffers() 1953 * to be called. We do this if the page is releasable by try_to_free_buffers(). 1954 * We also do it if the page has locked or dirty buffers and the caller wants 1955 * us to perform sync or async writeout. 1956 * 1957 * This complicates JBD locking somewhat. We aren't protected by the 1958 * BKL here. We wish to remove the buffer from its committing or 1959 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer. 1960 * 1961 * This may *change* the value of transaction_t->t_datalist, so anyone 1962 * who looks at t_datalist needs to lock against this function. 1963 * 1964 * Even worse, someone may be doing a jbd2_journal_dirty_data on this 1965 * buffer. So we need to lock against that. jbd2_journal_dirty_data() 1966 * will come out of the lock with the buffer dirty, which makes it 1967 * ineligible for release here. 1968 * 1969 * Who else is affected by this? hmm... Really the only contender 1970 * is do_get_write_access() - it could be looking at the buffer while 1971 * journal_try_to_free_buffer() is changing its state. But that 1972 * cannot happen because we never reallocate freed data as metadata 1973 * while the data is part of a transaction. Yes? 1974 * 1975 * Return 0 on failure, 1 on success 1976 */ 1977 int jbd2_journal_try_to_free_buffers(journal_t *journal, 1978 struct page *page, gfp_t gfp_mask) 1979 { 1980 struct buffer_head *head; 1981 struct buffer_head *bh; 1982 int ret = 0; 1983 1984 J_ASSERT(PageLocked(page)); 1985 1986 head = page_buffers(page); 1987 bh = head; 1988 do { 1989 struct journal_head *jh; 1990 1991 /* 1992 * We take our own ref against the journal_head here to avoid 1993 * having to add tons of locking around each instance of 1994 * jbd2_journal_put_journal_head(). 1995 */ 1996 jh = jbd2_journal_grab_journal_head(bh); 1997 if (!jh) 1998 continue; 1999 2000 jbd_lock_bh_state(bh); 2001 __journal_try_to_free_buffer(journal, bh); 2002 jbd2_journal_put_journal_head(jh); 2003 jbd_unlock_bh_state(bh); 2004 if (buffer_jbd(bh)) 2005 goto busy; 2006 } while ((bh = bh->b_this_page) != head); 2007 2008 ret = try_to_free_buffers(page); 2009 2010 busy: 2011 return ret; 2012 } 2013 2014 /* 2015 * This buffer is no longer needed. If it is on an older transaction's 2016 * checkpoint list we need to record it on this transaction's forget list 2017 * to pin this buffer (and hence its checkpointing transaction) down until 2018 * this transaction commits. If the buffer isn't on a checkpoint list, we 2019 * release it. 2020 * Returns non-zero if JBD no longer has an interest in the buffer. 2021 * 2022 * Called under j_list_lock. 2023 * 2024 * Called under jbd_lock_bh_state(bh). 2025 */ 2026 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction) 2027 { 2028 int may_free = 1; 2029 struct buffer_head *bh = jh2bh(jh); 2030 2031 if (jh->b_cp_transaction) { 2032 JBUFFER_TRACE(jh, "on running+cp transaction"); 2033 __jbd2_journal_temp_unlink_buffer(jh); 2034 /* 2035 * We don't want to write the buffer anymore, clear the 2036 * bit so that we don't confuse checks in 2037 * __journal_file_buffer 2038 */ 2039 clear_buffer_dirty(bh); 2040 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget); 2041 may_free = 0; 2042 } else { 2043 JBUFFER_TRACE(jh, "on running transaction"); 2044 __jbd2_journal_unfile_buffer(jh); 2045 } 2046 return may_free; 2047 } 2048 2049 /* 2050 * jbd2_journal_invalidatepage 2051 * 2052 * This code is tricky. It has a number of cases to deal with. 2053 * 2054 * There are two invariants which this code relies on: 2055 * 2056 * i_size must be updated on disk before we start calling invalidatepage on the 2057 * data. 2058 * 2059 * This is done in ext3 by defining an ext3_setattr method which 2060 * updates i_size before truncate gets going. By maintaining this 2061 * invariant, we can be sure that it is safe to throw away any buffers 2062 * attached to the current transaction: once the transaction commits, 2063 * we know that the data will not be needed. 2064 * 2065 * Note however that we can *not* throw away data belonging to the 2066 * previous, committing transaction! 2067 * 2068 * Any disk blocks which *are* part of the previous, committing 2069 * transaction (and which therefore cannot be discarded immediately) are 2070 * not going to be reused in the new running transaction 2071 * 2072 * The bitmap committed_data images guarantee this: any block which is 2073 * allocated in one transaction and removed in the next will be marked 2074 * as in-use in the committed_data bitmap, so cannot be reused until 2075 * the next transaction to delete the block commits. This means that 2076 * leaving committing buffers dirty is quite safe: the disk blocks 2077 * cannot be reallocated to a different file and so buffer aliasing is 2078 * not possible. 2079 * 2080 * 2081 * The above applies mainly to ordered data mode. In writeback mode we 2082 * don't make guarantees about the order in which data hits disk --- in 2083 * particular we don't guarantee that new dirty data is flushed before 2084 * transaction commit --- so it is always safe just to discard data 2085 * immediately in that mode. --sct 2086 */ 2087 2088 /* 2089 * The journal_unmap_buffer helper function returns zero if the buffer 2090 * concerned remains pinned as an anonymous buffer belonging to an older 2091 * transaction. 2092 * 2093 * We're outside-transaction here. Either or both of j_running_transaction 2094 * and j_committing_transaction may be NULL. 2095 */ 2096 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh, 2097 int partial_page) 2098 { 2099 transaction_t *transaction; 2100 struct journal_head *jh; 2101 int may_free = 1; 2102 2103 BUFFER_TRACE(bh, "entry"); 2104 2105 /* 2106 * It is safe to proceed here without the j_list_lock because the 2107 * buffers cannot be stolen by try_to_free_buffers as long as we are 2108 * holding the page lock. --sct 2109 */ 2110 2111 if (!buffer_jbd(bh)) 2112 goto zap_buffer_unlocked; 2113 2114 /* OK, we have data buffer in journaled mode */ 2115 write_lock(&journal->j_state_lock); 2116 jbd_lock_bh_state(bh); 2117 spin_lock(&journal->j_list_lock); 2118 2119 jh = jbd2_journal_grab_journal_head(bh); 2120 if (!jh) 2121 goto zap_buffer_no_jh; 2122 2123 /* 2124 * We cannot remove the buffer from checkpoint lists until the 2125 * transaction adding inode to orphan list (let's call it T) 2126 * is committed. Otherwise if the transaction changing the 2127 * buffer would be cleaned from the journal before T is 2128 * committed, a crash will cause that the correct contents of 2129 * the buffer will be lost. On the other hand we have to 2130 * clear the buffer dirty bit at latest at the moment when the 2131 * transaction marking the buffer as freed in the filesystem 2132 * structures is committed because from that moment on the 2133 * block can be reallocated and used by a different page. 2134 * Since the block hasn't been freed yet but the inode has 2135 * already been added to orphan list, it is safe for us to add 2136 * the buffer to BJ_Forget list of the newest transaction. 2137 * 2138 * Also we have to clear buffer_mapped flag of a truncated buffer 2139 * because the buffer_head may be attached to the page straddling 2140 * i_size (can happen only when blocksize < pagesize) and thus the 2141 * buffer_head can be reused when the file is extended again. So we end 2142 * up keeping around invalidated buffers attached to transactions' 2143 * BJ_Forget list just to stop checkpointing code from cleaning up 2144 * the transaction this buffer was modified in. 2145 */ 2146 transaction = jh->b_transaction; 2147 if (transaction == NULL) { 2148 /* First case: not on any transaction. If it 2149 * has no checkpoint link, then we can zap it: 2150 * it's a writeback-mode buffer so we don't care 2151 * if it hits disk safely. */ 2152 if (!jh->b_cp_transaction) { 2153 JBUFFER_TRACE(jh, "not on any transaction: zap"); 2154 goto zap_buffer; 2155 } 2156 2157 if (!buffer_dirty(bh)) { 2158 /* bdflush has written it. We can drop it now */ 2159 __jbd2_journal_remove_checkpoint(jh); 2160 goto zap_buffer; 2161 } 2162 2163 /* OK, it must be in the journal but still not 2164 * written fully to disk: it's metadata or 2165 * journaled data... */ 2166 2167 if (journal->j_running_transaction) { 2168 /* ... and once the current transaction has 2169 * committed, the buffer won't be needed any 2170 * longer. */ 2171 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget"); 2172 may_free = __dispose_buffer(jh, 2173 journal->j_running_transaction); 2174 goto zap_buffer; 2175 } else { 2176 /* There is no currently-running transaction. So the 2177 * orphan record which we wrote for this file must have 2178 * passed into commit. We must attach this buffer to 2179 * the committing transaction, if it exists. */ 2180 if (journal->j_committing_transaction) { 2181 JBUFFER_TRACE(jh, "give to committing trans"); 2182 may_free = __dispose_buffer(jh, 2183 journal->j_committing_transaction); 2184 goto zap_buffer; 2185 } else { 2186 /* The orphan record's transaction has 2187 * committed. We can cleanse this buffer */ 2188 clear_buffer_jbddirty(bh); 2189 __jbd2_journal_remove_checkpoint(jh); 2190 goto zap_buffer; 2191 } 2192 } 2193 } else if (transaction == journal->j_committing_transaction) { 2194 JBUFFER_TRACE(jh, "on committing transaction"); 2195 /* 2196 * The buffer is committing, we simply cannot touch 2197 * it. If the page is straddling i_size we have to wait 2198 * for commit and try again. 2199 */ 2200 if (partial_page) { 2201 jbd2_journal_put_journal_head(jh); 2202 spin_unlock(&journal->j_list_lock); 2203 jbd_unlock_bh_state(bh); 2204 write_unlock(&journal->j_state_lock); 2205 return -EBUSY; 2206 } 2207 /* 2208 * OK, buffer won't be reachable after truncate. We just set 2209 * j_next_transaction to the running transaction (if there is 2210 * one) and mark buffer as freed so that commit code knows it 2211 * should clear dirty bits when it is done with the buffer. 2212 */ 2213 set_buffer_freed(bh); 2214 if (journal->j_running_transaction && buffer_jbddirty(bh)) 2215 jh->b_next_transaction = journal->j_running_transaction; 2216 jbd2_journal_put_journal_head(jh); 2217 spin_unlock(&journal->j_list_lock); 2218 jbd_unlock_bh_state(bh); 2219 write_unlock(&journal->j_state_lock); 2220 return 0; 2221 } else { 2222 /* Good, the buffer belongs to the running transaction. 2223 * We are writing our own transaction's data, not any 2224 * previous one's, so it is safe to throw it away 2225 * (remember that we expect the filesystem to have set 2226 * i_size already for this truncate so recovery will not 2227 * expose the disk blocks we are discarding here.) */ 2228 J_ASSERT_JH(jh, transaction == journal->j_running_transaction); 2229 JBUFFER_TRACE(jh, "on running transaction"); 2230 may_free = __dispose_buffer(jh, transaction); 2231 } 2232 2233 zap_buffer: 2234 /* 2235 * This is tricky. Although the buffer is truncated, it may be reused 2236 * if blocksize < pagesize and it is attached to the page straddling 2237 * EOF. Since the buffer might have been added to BJ_Forget list of the 2238 * running transaction, journal_get_write_access() won't clear 2239 * b_modified and credit accounting gets confused. So clear b_modified 2240 * here. 2241 */ 2242 jh->b_modified = 0; 2243 jbd2_journal_put_journal_head(jh); 2244 zap_buffer_no_jh: 2245 spin_unlock(&journal->j_list_lock); 2246 jbd_unlock_bh_state(bh); 2247 write_unlock(&journal->j_state_lock); 2248 zap_buffer_unlocked: 2249 clear_buffer_dirty(bh); 2250 J_ASSERT_BH(bh, !buffer_jbddirty(bh)); 2251 clear_buffer_mapped(bh); 2252 clear_buffer_req(bh); 2253 clear_buffer_new(bh); 2254 clear_buffer_delay(bh); 2255 clear_buffer_unwritten(bh); 2256 bh->b_bdev = NULL; 2257 return may_free; 2258 } 2259 2260 /** 2261 * void jbd2_journal_invalidatepage() 2262 * @journal: journal to use for flush... 2263 * @page: page to flush 2264 * @offset: start of the range to invalidate 2265 * @length: length of the range to invalidate 2266 * 2267 * Reap page buffers containing data after in the specified range in page. 2268 * Can return -EBUSY if buffers are part of the committing transaction and 2269 * the page is straddling i_size. Caller then has to wait for current commit 2270 * and try again. 2271 */ 2272 int jbd2_journal_invalidatepage(journal_t *journal, 2273 struct page *page, 2274 unsigned int offset, 2275 unsigned int length) 2276 { 2277 struct buffer_head *head, *bh, *next; 2278 unsigned int stop = offset + length; 2279 unsigned int curr_off = 0; 2280 int partial_page = (offset || length < PAGE_CACHE_SIZE); 2281 int may_free = 1; 2282 int ret = 0; 2283 2284 if (!PageLocked(page)) 2285 BUG(); 2286 if (!page_has_buffers(page)) 2287 return 0; 2288 2289 BUG_ON(stop > PAGE_CACHE_SIZE || stop < length); 2290 2291 /* We will potentially be playing with lists other than just the 2292 * data lists (especially for journaled data mode), so be 2293 * cautious in our locking. */ 2294 2295 head = bh = page_buffers(page); 2296 do { 2297 unsigned int next_off = curr_off + bh->b_size; 2298 next = bh->b_this_page; 2299 2300 if (next_off > stop) 2301 return 0; 2302 2303 if (offset <= curr_off) { 2304 /* This block is wholly outside the truncation point */ 2305 lock_buffer(bh); 2306 ret = journal_unmap_buffer(journal, bh, partial_page); 2307 unlock_buffer(bh); 2308 if (ret < 0) 2309 return ret; 2310 may_free &= ret; 2311 } 2312 curr_off = next_off; 2313 bh = next; 2314 2315 } while (bh != head); 2316 2317 if (!partial_page) { 2318 if (may_free && try_to_free_buffers(page)) 2319 J_ASSERT(!page_has_buffers(page)); 2320 } 2321 return 0; 2322 } 2323 2324 /* 2325 * File a buffer on the given transaction list. 2326 */ 2327 void __jbd2_journal_file_buffer(struct journal_head *jh, 2328 transaction_t *transaction, int jlist) 2329 { 2330 struct journal_head **list = NULL; 2331 int was_dirty = 0; 2332 struct buffer_head *bh = jh2bh(jh); 2333 2334 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh)); 2335 assert_spin_locked(&transaction->t_journal->j_list_lock); 2336 2337 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types); 2338 J_ASSERT_JH(jh, jh->b_transaction == transaction || 2339 jh->b_transaction == NULL); 2340 2341 if (jh->b_transaction && jh->b_jlist == jlist) 2342 return; 2343 2344 if (jlist == BJ_Metadata || jlist == BJ_Reserved || 2345 jlist == BJ_Shadow || jlist == BJ_Forget) { 2346 /* 2347 * For metadata buffers, we track dirty bit in buffer_jbddirty 2348 * instead of buffer_dirty. We should not see a dirty bit set 2349 * here because we clear it in do_get_write_access but e.g. 2350 * tune2fs can modify the sb and set the dirty bit at any time 2351 * so we try to gracefully handle that. 2352 */ 2353 if (buffer_dirty(bh)) 2354 warn_dirty_buffer(bh); 2355 if (test_clear_buffer_dirty(bh) || 2356 test_clear_buffer_jbddirty(bh)) 2357 was_dirty = 1; 2358 } 2359 2360 if (jh->b_transaction) 2361 __jbd2_journal_temp_unlink_buffer(jh); 2362 else 2363 jbd2_journal_grab_journal_head(bh); 2364 jh->b_transaction = transaction; 2365 2366 switch (jlist) { 2367 case BJ_None: 2368 J_ASSERT_JH(jh, !jh->b_committed_data); 2369 J_ASSERT_JH(jh, !jh->b_frozen_data); 2370 return; 2371 case BJ_Metadata: 2372 transaction->t_nr_buffers++; 2373 list = &transaction->t_buffers; 2374 break; 2375 case BJ_Forget: 2376 list = &transaction->t_forget; 2377 break; 2378 case BJ_Shadow: 2379 list = &transaction->t_shadow_list; 2380 break; 2381 case BJ_Reserved: 2382 list = &transaction->t_reserved_list; 2383 break; 2384 } 2385 2386 __blist_add_buffer(list, jh); 2387 jh->b_jlist = jlist; 2388 2389 if (was_dirty) 2390 set_buffer_jbddirty(bh); 2391 } 2392 2393 void jbd2_journal_file_buffer(struct journal_head *jh, 2394 transaction_t *transaction, int jlist) 2395 { 2396 jbd_lock_bh_state(jh2bh(jh)); 2397 spin_lock(&transaction->t_journal->j_list_lock); 2398 __jbd2_journal_file_buffer(jh, transaction, jlist); 2399 spin_unlock(&transaction->t_journal->j_list_lock); 2400 jbd_unlock_bh_state(jh2bh(jh)); 2401 } 2402 2403 /* 2404 * Remove a buffer from its current buffer list in preparation for 2405 * dropping it from its current transaction entirely. If the buffer has 2406 * already started to be used by a subsequent transaction, refile the 2407 * buffer on that transaction's metadata list. 2408 * 2409 * Called under j_list_lock 2410 * Called under jbd_lock_bh_state(jh2bh(jh)) 2411 * 2412 * jh and bh may be already free when this function returns 2413 */ 2414 void __jbd2_journal_refile_buffer(struct journal_head *jh) 2415 { 2416 int was_dirty, jlist; 2417 struct buffer_head *bh = jh2bh(jh); 2418 2419 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh)); 2420 if (jh->b_transaction) 2421 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock); 2422 2423 /* If the buffer is now unused, just drop it. */ 2424 if (jh->b_next_transaction == NULL) { 2425 __jbd2_journal_unfile_buffer(jh); 2426 return; 2427 } 2428 2429 /* 2430 * It has been modified by a later transaction: add it to the new 2431 * transaction's metadata list. 2432 */ 2433 2434 was_dirty = test_clear_buffer_jbddirty(bh); 2435 __jbd2_journal_temp_unlink_buffer(jh); 2436 /* 2437 * We set b_transaction here because b_next_transaction will inherit 2438 * our jh reference and thus __jbd2_journal_file_buffer() must not 2439 * take a new one. 2440 */ 2441 jh->b_transaction = jh->b_next_transaction; 2442 jh->b_next_transaction = NULL; 2443 if (buffer_freed(bh)) 2444 jlist = BJ_Forget; 2445 else if (jh->b_modified) 2446 jlist = BJ_Metadata; 2447 else 2448 jlist = BJ_Reserved; 2449 __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist); 2450 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING); 2451 2452 if (was_dirty) 2453 set_buffer_jbddirty(bh); 2454 } 2455 2456 /* 2457 * __jbd2_journal_refile_buffer() with necessary locking added. We take our 2458 * bh reference so that we can safely unlock bh. 2459 * 2460 * The jh and bh may be freed by this call. 2461 */ 2462 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh) 2463 { 2464 struct buffer_head *bh = jh2bh(jh); 2465 2466 /* Get reference so that buffer cannot be freed before we unlock it */ 2467 get_bh(bh); 2468 jbd_lock_bh_state(bh); 2469 spin_lock(&journal->j_list_lock); 2470 __jbd2_journal_refile_buffer(jh); 2471 jbd_unlock_bh_state(bh); 2472 spin_unlock(&journal->j_list_lock); 2473 __brelse(bh); 2474 } 2475 2476 /* 2477 * File inode in the inode list of the handle's transaction 2478 */ 2479 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode) 2480 { 2481 transaction_t *transaction = handle->h_transaction; 2482 journal_t *journal; 2483 2484 if (is_handle_aborted(handle)) 2485 return -EROFS; 2486 journal = transaction->t_journal; 2487 2488 jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino, 2489 transaction->t_tid); 2490 2491 /* 2492 * First check whether inode isn't already on the transaction's 2493 * lists without taking the lock. Note that this check is safe 2494 * without the lock as we cannot race with somebody removing inode 2495 * from the transaction. The reason is that we remove inode from the 2496 * transaction only in journal_release_jbd_inode() and when we commit 2497 * the transaction. We are guarded from the first case by holding 2498 * a reference to the inode. We are safe against the second case 2499 * because if jinode->i_transaction == transaction, commit code 2500 * cannot touch the transaction because we hold reference to it, 2501 * and if jinode->i_next_transaction == transaction, commit code 2502 * will only file the inode where we want it. 2503 */ 2504 if (jinode->i_transaction == transaction || 2505 jinode->i_next_transaction == transaction) 2506 return 0; 2507 2508 spin_lock(&journal->j_list_lock); 2509 2510 if (jinode->i_transaction == transaction || 2511 jinode->i_next_transaction == transaction) 2512 goto done; 2513 2514 /* 2515 * We only ever set this variable to 1 so the test is safe. Since 2516 * t_need_data_flush is likely to be set, we do the test to save some 2517 * cacheline bouncing 2518 */ 2519 if (!transaction->t_need_data_flush) 2520 transaction->t_need_data_flush = 1; 2521 /* On some different transaction's list - should be 2522 * the committing one */ 2523 if (jinode->i_transaction) { 2524 J_ASSERT(jinode->i_next_transaction == NULL); 2525 J_ASSERT(jinode->i_transaction == 2526 journal->j_committing_transaction); 2527 jinode->i_next_transaction = transaction; 2528 goto done; 2529 } 2530 /* Not on any transaction list... */ 2531 J_ASSERT(!jinode->i_next_transaction); 2532 jinode->i_transaction = transaction; 2533 list_add(&jinode->i_list, &transaction->t_inode_list); 2534 done: 2535 spin_unlock(&journal->j_list_lock); 2536 2537 return 0; 2538 } 2539 2540 /* 2541 * File truncate and transaction commit interact with each other in a 2542 * non-trivial way. If a transaction writing data block A is 2543 * committing, we cannot discard the data by truncate until we have 2544 * written them. Otherwise if we crashed after the transaction with 2545 * write has committed but before the transaction with truncate has 2546 * committed, we could see stale data in block A. This function is a 2547 * helper to solve this problem. It starts writeout of the truncated 2548 * part in case it is in the committing transaction. 2549 * 2550 * Filesystem code must call this function when inode is journaled in 2551 * ordered mode before truncation happens and after the inode has been 2552 * placed on orphan list with the new inode size. The second condition 2553 * avoids the race that someone writes new data and we start 2554 * committing the transaction after this function has been called but 2555 * before a transaction for truncate is started (and furthermore it 2556 * allows us to optimize the case where the addition to orphan list 2557 * happens in the same transaction as write --- we don't have to write 2558 * any data in such case). 2559 */ 2560 int jbd2_journal_begin_ordered_truncate(journal_t *journal, 2561 struct jbd2_inode *jinode, 2562 loff_t new_size) 2563 { 2564 transaction_t *inode_trans, *commit_trans; 2565 int ret = 0; 2566 2567 /* This is a quick check to avoid locking if not necessary */ 2568 if (!jinode->i_transaction) 2569 goto out; 2570 /* Locks are here just to force reading of recent values, it is 2571 * enough that the transaction was not committing before we started 2572 * a transaction adding the inode to orphan list */ 2573 read_lock(&journal->j_state_lock); 2574 commit_trans = journal->j_committing_transaction; 2575 read_unlock(&journal->j_state_lock); 2576 spin_lock(&journal->j_list_lock); 2577 inode_trans = jinode->i_transaction; 2578 spin_unlock(&journal->j_list_lock); 2579 if (inode_trans == commit_trans) { 2580 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping, 2581 new_size, LLONG_MAX); 2582 if (ret) 2583 jbd2_journal_abort(journal, ret); 2584 } 2585 out: 2586 return ret; 2587 } 2588