Root/
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/module.h> |
31 | |
32 | static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh); |
33 | static void __jbd2_journal_unfile_buffer(struct journal_head *jh); |
34 | |
35 | /* |
36 | * jbd2_get_transaction: obtain a new transaction_t object. |
37 | * |
38 | * Simply allocate and initialise a new transaction. Create it in |
39 | * RUNNING state and add it to the current journal (which should not |
40 | * have an existing running transaction: we only make a new transaction |
41 | * once we have started to commit the old one). |
42 | * |
43 | * Preconditions: |
44 | * The journal MUST be locked. We don't perform atomic mallocs on the |
45 | * new transaction and we can't block without protecting against other |
46 | * processes trying to touch the journal while it is in transition. |
47 | * |
48 | */ |
49 | |
50 | static transaction_t * |
51 | jbd2_get_transaction(journal_t *journal, transaction_t *transaction) |
52 | { |
53 | transaction->t_journal = journal; |
54 | transaction->t_state = T_RUNNING; |
55 | transaction->t_start_time = ktime_get(); |
56 | transaction->t_tid = journal->j_transaction_sequence++; |
57 | transaction->t_expires = jiffies + journal->j_commit_interval; |
58 | spin_lock_init(&transaction->t_handle_lock); |
59 | atomic_set(&transaction->t_updates, 0); |
60 | atomic_set(&transaction->t_outstanding_credits, 0); |
61 | atomic_set(&transaction->t_handle_count, 0); |
62 | INIT_LIST_HEAD(&transaction->t_inode_list); |
63 | INIT_LIST_HEAD(&transaction->t_private_list); |
64 | |
65 | /* Set up the commit timer for the new transaction. */ |
66 | journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires); |
67 | add_timer(&journal->j_commit_timer); |
68 | |
69 | J_ASSERT(journal->j_running_transaction == NULL); |
70 | journal->j_running_transaction = transaction; |
71 | transaction->t_max_wait = 0; |
72 | transaction->t_start = jiffies; |
73 | |
74 | return transaction; |
75 | } |
76 | |
77 | /* |
78 | * Handle management. |
79 | * |
80 | * A handle_t is an object which represents a single atomic update to a |
81 | * filesystem, and which tracks all of the modifications which form part |
82 | * of that one update. |
83 | */ |
84 | |
85 | /* |
86 | * Update transaction's maximum wait time, if debugging is enabled. |
87 | * |
88 | * In order for t_max_wait to be reliable, it must be protected by a |
89 | * lock. But doing so will mean that start_this_handle() can not be |
90 | * run in parallel on SMP systems, which limits our scalability. So |
91 | * unless debugging is enabled, we no longer update t_max_wait, which |
92 | * means that maximum wait time reported by the jbd2_run_stats |
93 | * tracepoint will always be zero. |
94 | */ |
95 | static inline void update_t_max_wait(transaction_t *transaction, |
96 | unsigned long ts) |
97 | { |
98 | #ifdef CONFIG_JBD2_DEBUG |
99 | if (jbd2_journal_enable_debug && |
100 | time_after(transaction->t_start, ts)) { |
101 | ts = jbd2_time_diff(ts, transaction->t_start); |
102 | spin_lock(&transaction->t_handle_lock); |
103 | if (ts > transaction->t_max_wait) |
104 | transaction->t_max_wait = ts; |
105 | spin_unlock(&transaction->t_handle_lock); |
106 | } |
107 | #endif |
108 | } |
109 | |
110 | /* |
111 | * start_this_handle: Given a handle, deal with any locking or stalling |
112 | * needed to make sure that there is enough journal space for the handle |
113 | * to begin. Attach the handle to a transaction and set up the |
114 | * transaction's buffer credits. |
115 | */ |
116 | |
117 | static int start_this_handle(journal_t *journal, handle_t *handle, |
118 | int gfp_mask) |
119 | { |
120 | transaction_t *transaction, *new_transaction = NULL; |
121 | tid_t tid; |
122 | int needed, need_to_start; |
123 | int nblocks = handle->h_buffer_credits; |
124 | unsigned long ts = jiffies; |
125 | |
126 | if (nblocks > journal->j_max_transaction_buffers) { |
127 | printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n", |
128 | current->comm, nblocks, |
129 | journal->j_max_transaction_buffers); |
130 | return -ENOSPC; |
131 | } |
132 | |
133 | alloc_transaction: |
134 | if (!journal->j_running_transaction) { |
135 | new_transaction = kzalloc(sizeof(*new_transaction), gfp_mask); |
136 | if (!new_transaction) { |
137 | /* |
138 | * If __GFP_FS is not present, then we may be |
139 | * being called from inside the fs writeback |
140 | * layer, so we MUST NOT fail. Since |
141 | * __GFP_NOFAIL is going away, we will arrange |
142 | * to retry the allocation ourselves. |
143 | */ |
144 | if ((gfp_mask & __GFP_FS) == 0) { |
145 | congestion_wait(BLK_RW_ASYNC, HZ/50); |
146 | goto alloc_transaction; |
147 | } |
148 | return -ENOMEM; |
149 | } |
150 | } |
151 | |
152 | jbd_debug(3, "New handle %p going live.\n", handle); |
153 | |
154 | /* |
155 | * We need to hold j_state_lock until t_updates has been incremented, |
156 | * for proper journal barrier handling |
157 | */ |
158 | repeat: |
159 | read_lock(&journal->j_state_lock); |
160 | BUG_ON(journal->j_flags & JBD2_UNMOUNT); |
161 | if (is_journal_aborted(journal) || |
162 | (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) { |
163 | read_unlock(&journal->j_state_lock); |
164 | kfree(new_transaction); |
165 | return -EROFS; |
166 | } |
167 | |
168 | /* Wait on the journal's transaction barrier if necessary */ |
169 | if (journal->j_barrier_count) { |
170 | read_unlock(&journal->j_state_lock); |
171 | wait_event(journal->j_wait_transaction_locked, |
172 | journal->j_barrier_count == 0); |
173 | goto repeat; |
174 | } |
175 | |
176 | if (!journal->j_running_transaction) { |
177 | read_unlock(&journal->j_state_lock); |
178 | if (!new_transaction) |
179 | goto alloc_transaction; |
180 | write_lock(&journal->j_state_lock); |
181 | if (!journal->j_running_transaction) { |
182 | jbd2_get_transaction(journal, new_transaction); |
183 | new_transaction = NULL; |
184 | } |
185 | write_unlock(&journal->j_state_lock); |
186 | goto repeat; |
187 | } |
188 | |
189 | transaction = journal->j_running_transaction; |
190 | |
191 | /* |
192 | * If the current transaction is locked down for commit, wait for the |
193 | * lock to be released. |
194 | */ |
195 | if (transaction->t_state == T_LOCKED) { |
196 | DEFINE_WAIT(wait); |
197 | |
198 | prepare_to_wait(&journal->j_wait_transaction_locked, |
199 | &wait, TASK_UNINTERRUPTIBLE); |
200 | read_unlock(&journal->j_state_lock); |
201 | schedule(); |
202 | finish_wait(&journal->j_wait_transaction_locked, &wait); |
203 | goto repeat; |
204 | } |
205 | |
206 | /* |
207 | * If there is not enough space left in the log to write all potential |
208 | * buffers requested by this operation, we need to stall pending a log |
209 | * checkpoint to free some more log space. |
210 | */ |
211 | needed = atomic_add_return(nblocks, |
212 | &transaction->t_outstanding_credits); |
213 | |
214 | if (needed > journal->j_max_transaction_buffers) { |
215 | /* |
216 | * If the current transaction is already too large, then start |
217 | * to commit it: we can then go back and attach this handle to |
218 | * a new transaction. |
219 | */ |
220 | DEFINE_WAIT(wait); |
221 | |
222 | jbd_debug(2, "Handle %p starting new commit...\n", handle); |
223 | atomic_sub(nblocks, &transaction->t_outstanding_credits); |
224 | prepare_to_wait(&journal->j_wait_transaction_locked, &wait, |
225 | TASK_UNINTERRUPTIBLE); |
226 | tid = transaction->t_tid; |
227 | need_to_start = !tid_geq(journal->j_commit_request, tid); |
228 | read_unlock(&journal->j_state_lock); |
229 | if (need_to_start) |
230 | jbd2_log_start_commit(journal, tid); |
231 | schedule(); |
232 | finish_wait(&journal->j_wait_transaction_locked, &wait); |
233 | goto repeat; |
234 | } |
235 | |
236 | /* |
237 | * The commit code assumes that it can get enough log space |
238 | * without forcing a checkpoint. This is *critical* for |
239 | * correctness: a checkpoint of a buffer which is also |
240 | * associated with a committing transaction creates a deadlock, |
241 | * so commit simply cannot force through checkpoints. |
242 | * |
243 | * We must therefore ensure the necessary space in the journal |
244 | * *before* starting to dirty potentially checkpointed buffers |
245 | * in the new transaction. |
246 | * |
247 | * The worst part is, any transaction currently committing can |
248 | * reduce the free space arbitrarily. Be careful to account for |
249 | * those buffers when checkpointing. |
250 | */ |
251 | |
252 | /* |
253 | * @@@ AKPM: This seems rather over-defensive. We're giving commit |
254 | * a _lot_ of headroom: 1/4 of the journal plus the size of |
255 | * the committing transaction. Really, we only need to give it |
256 | * committing_transaction->t_outstanding_credits plus "enough" for |
257 | * the log control blocks. |
258 | * Also, this test is inconsistent with the matching one in |
259 | * jbd2_journal_extend(). |
260 | */ |
261 | if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) { |
262 | jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle); |
263 | atomic_sub(nblocks, &transaction->t_outstanding_credits); |
264 | read_unlock(&journal->j_state_lock); |
265 | write_lock(&journal->j_state_lock); |
266 | if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) |
267 | __jbd2_log_wait_for_space(journal); |
268 | write_unlock(&journal->j_state_lock); |
269 | goto repeat; |
270 | } |
271 | |
272 | /* OK, account for the buffers that this operation expects to |
273 | * use and add the handle to the running transaction. |
274 | */ |
275 | update_t_max_wait(transaction, ts); |
276 | handle->h_transaction = transaction; |
277 | atomic_inc(&transaction->t_updates); |
278 | atomic_inc(&transaction->t_handle_count); |
279 | jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n", |
280 | handle, nblocks, |
281 | atomic_read(&transaction->t_outstanding_credits), |
282 | __jbd2_log_space_left(journal)); |
283 | read_unlock(&journal->j_state_lock); |
284 | |
285 | lock_map_acquire(&handle->h_lockdep_map); |
286 | kfree(new_transaction); |
287 | return 0; |
288 | } |
289 | |
290 | static struct lock_class_key jbd2_handle_key; |
291 | |
292 | /* Allocate a new handle. This should probably be in a slab... */ |
293 | static handle_t *new_handle(int nblocks) |
294 | { |
295 | handle_t *handle = jbd2_alloc_handle(GFP_NOFS); |
296 | if (!handle) |
297 | return NULL; |
298 | memset(handle, 0, sizeof(*handle)); |
299 | handle->h_buffer_credits = nblocks; |
300 | handle->h_ref = 1; |
301 | |
302 | lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle", |
303 | &jbd2_handle_key, 0); |
304 | |
305 | return handle; |
306 | } |
307 | |
308 | /** |
309 | * handle_t *jbd2_journal_start() - Obtain a new handle. |
310 | * @journal: Journal to start transaction on. |
311 | * @nblocks: number of block buffer we might modify |
312 | * |
313 | * We make sure that the transaction can guarantee at least nblocks of |
314 | * modified buffers in the log. We block until the log can guarantee |
315 | * that much space. |
316 | * |
317 | * This function is visible to journal users (like ext3fs), so is not |
318 | * called with the journal already locked. |
319 | * |
320 | * Return a pointer to a newly allocated handle, or an ERR_PTR() value |
321 | * on failure. |
322 | */ |
323 | handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int gfp_mask) |
324 | { |
325 | handle_t *handle = journal_current_handle(); |
326 | int err; |
327 | |
328 | if (!journal) |
329 | return ERR_PTR(-EROFS); |
330 | |
331 | if (handle) { |
332 | J_ASSERT(handle->h_transaction->t_journal == journal); |
333 | handle->h_ref++; |
334 | return handle; |
335 | } |
336 | |
337 | handle = new_handle(nblocks); |
338 | if (!handle) |
339 | return ERR_PTR(-ENOMEM); |
340 | |
341 | current->journal_info = handle; |
342 | |
343 | err = start_this_handle(journal, handle, gfp_mask); |
344 | if (err < 0) { |
345 | jbd2_free_handle(handle); |
346 | current->journal_info = NULL; |
347 | handle = ERR_PTR(err); |
348 | } |
349 | return handle; |
350 | } |
351 | EXPORT_SYMBOL(jbd2__journal_start); |
352 | |
353 | |
354 | handle_t *jbd2_journal_start(journal_t *journal, int nblocks) |
355 | { |
356 | return jbd2__journal_start(journal, nblocks, GFP_NOFS); |
357 | } |
358 | EXPORT_SYMBOL(jbd2_journal_start); |
359 | |
360 | |
361 | /** |
362 | * int jbd2_journal_extend() - extend buffer credits. |
363 | * @handle: handle to 'extend' |
364 | * @nblocks: nr blocks to try to extend by. |
365 | * |
366 | * Some transactions, such as large extends and truncates, can be done |
367 | * atomically all at once or in several stages. The operation requests |
368 | * a credit for a number of buffer modications in advance, but can |
369 | * extend its credit if it needs more. |
370 | * |
371 | * jbd2_journal_extend tries to give the running handle more buffer credits. |
372 | * It does not guarantee that allocation - this is a best-effort only. |
373 | * The calling process MUST be able to deal cleanly with a failure to |
374 | * extend here. |
375 | * |
376 | * Return 0 on success, non-zero on failure. |
377 | * |
378 | * return code < 0 implies an error |
379 | * return code > 0 implies normal transaction-full status. |
380 | */ |
381 | int jbd2_journal_extend(handle_t *handle, int nblocks) |
382 | { |
383 | transaction_t *transaction = handle->h_transaction; |
384 | journal_t *journal = transaction->t_journal; |
385 | int result; |
386 | int wanted; |
387 | |
388 | result = -EIO; |
389 | if (is_handle_aborted(handle)) |
390 | goto out; |
391 | |
392 | result = 1; |
393 | |
394 | read_lock(&journal->j_state_lock); |
395 | |
396 | /* Don't extend a locked-down transaction! */ |
397 | if (handle->h_transaction->t_state != T_RUNNING) { |
398 | jbd_debug(3, "denied handle %p %d blocks: " |
399 | "transaction not running\n", handle, nblocks); |
400 | goto error_out; |
401 | } |
402 | |
403 | spin_lock(&transaction->t_handle_lock); |
404 | wanted = atomic_read(&transaction->t_outstanding_credits) + nblocks; |
405 | |
406 | if (wanted > journal->j_max_transaction_buffers) { |
407 | jbd_debug(3, "denied handle %p %d blocks: " |
408 | "transaction too large\n", handle, nblocks); |
409 | goto unlock; |
410 | } |
411 | |
412 | if (wanted > __jbd2_log_space_left(journal)) { |
413 | jbd_debug(3, "denied handle %p %d blocks: " |
414 | "insufficient log space\n", handle, nblocks); |
415 | goto unlock; |
416 | } |
417 | |
418 | handle->h_buffer_credits += nblocks; |
419 | atomic_add(nblocks, &transaction->t_outstanding_credits); |
420 | result = 0; |
421 | |
422 | jbd_debug(3, "extended handle %p by %d\n", handle, nblocks); |
423 | unlock: |
424 | spin_unlock(&transaction->t_handle_lock); |
425 | error_out: |
426 | read_unlock(&journal->j_state_lock); |
427 | out: |
428 | return result; |
429 | } |
430 | |
431 | |
432 | /** |
433 | * int jbd2_journal_restart() - restart a handle . |
434 | * @handle: handle to restart |
435 | * @nblocks: nr credits requested |
436 | * |
437 | * Restart a handle for a multi-transaction filesystem |
438 | * operation. |
439 | * |
440 | * If the jbd2_journal_extend() call above fails to grant new buffer credits |
441 | * to a running handle, a call to jbd2_journal_restart will commit the |
442 | * handle's transaction so far and reattach the handle to a new |
443 | * transaction capabable of guaranteeing the requested number of |
444 | * credits. |
445 | */ |
446 | int jbd2__journal_restart(handle_t *handle, int nblocks, int gfp_mask) |
447 | { |
448 | transaction_t *transaction = handle->h_transaction; |
449 | journal_t *journal = transaction->t_journal; |
450 | tid_t tid; |
451 | int need_to_start, ret; |
452 | |
453 | /* If we've had an abort of any type, don't even think about |
454 | * actually doing the restart! */ |
455 | if (is_handle_aborted(handle)) |
456 | return 0; |
457 | |
458 | /* |
459 | * First unlink the handle from its current transaction, and start the |
460 | * commit on that. |
461 | */ |
462 | J_ASSERT(atomic_read(&transaction->t_updates) > 0); |
463 | J_ASSERT(journal_current_handle() == handle); |
464 | |
465 | read_lock(&journal->j_state_lock); |
466 | spin_lock(&transaction->t_handle_lock); |
467 | atomic_sub(handle->h_buffer_credits, |
468 | &transaction->t_outstanding_credits); |
469 | if (atomic_dec_and_test(&transaction->t_updates)) |
470 | wake_up(&journal->j_wait_updates); |
471 | spin_unlock(&transaction->t_handle_lock); |
472 | |
473 | jbd_debug(2, "restarting handle %p\n", handle); |
474 | tid = transaction->t_tid; |
475 | need_to_start = !tid_geq(journal->j_commit_request, tid); |
476 | read_unlock(&journal->j_state_lock); |
477 | if (need_to_start) |
478 | jbd2_log_start_commit(journal, tid); |
479 | |
480 | lock_map_release(&handle->h_lockdep_map); |
481 | handle->h_buffer_credits = nblocks; |
482 | ret = start_this_handle(journal, handle, gfp_mask); |
483 | return ret; |
484 | } |
485 | EXPORT_SYMBOL(jbd2__journal_restart); |
486 | |
487 | |
488 | int jbd2_journal_restart(handle_t *handle, int nblocks) |
489 | { |
490 | return jbd2__journal_restart(handle, nblocks, GFP_NOFS); |
491 | } |
492 | EXPORT_SYMBOL(jbd2_journal_restart); |
493 | |
494 | /** |
495 | * void jbd2_journal_lock_updates () - establish a transaction barrier. |
496 | * @journal: Journal to establish a barrier on. |
497 | * |
498 | * This locks out any further updates from being started, and blocks |
499 | * until all existing updates have completed, returning only once the |
500 | * journal is in a quiescent state with no updates running. |
501 | * |
502 | * The journal lock should not be held on entry. |
503 | */ |
504 | void jbd2_journal_lock_updates(journal_t *journal) |
505 | { |
506 | DEFINE_WAIT(wait); |
507 | |
508 | write_lock(&journal->j_state_lock); |
509 | ++journal->j_barrier_count; |
510 | |
511 | /* Wait until there are no running updates */ |
512 | while (1) { |
513 | transaction_t *transaction = journal->j_running_transaction; |
514 | |
515 | if (!transaction) |
516 | break; |
517 | |
518 | spin_lock(&transaction->t_handle_lock); |
519 | if (!atomic_read(&transaction->t_updates)) { |
520 | spin_unlock(&transaction->t_handle_lock); |
521 | break; |
522 | } |
523 | prepare_to_wait(&journal->j_wait_updates, &wait, |
524 | TASK_UNINTERRUPTIBLE); |
525 | spin_unlock(&transaction->t_handle_lock); |
526 | write_unlock(&journal->j_state_lock); |
527 | schedule(); |
528 | finish_wait(&journal->j_wait_updates, &wait); |
529 | write_lock(&journal->j_state_lock); |
530 | } |
531 | write_unlock(&journal->j_state_lock); |
532 | |
533 | /* |
534 | * We have now established a barrier against other normal updates, but |
535 | * we also need to barrier against other jbd2_journal_lock_updates() calls |
536 | * to make sure that we serialise special journal-locked operations |
537 | * too. |
538 | */ |
539 | mutex_lock(&journal->j_barrier); |
540 | } |
541 | |
542 | /** |
543 | * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier |
544 | * @journal: Journal to release the barrier on. |
545 | * |
546 | * Release a transaction barrier obtained with jbd2_journal_lock_updates(). |
547 | * |
548 | * Should be called without the journal lock held. |
549 | */ |
550 | void jbd2_journal_unlock_updates (journal_t *journal) |
551 | { |
552 | J_ASSERT(journal->j_barrier_count != 0); |
553 | |
554 | mutex_unlock(&journal->j_barrier); |
555 | write_lock(&journal->j_state_lock); |
556 | --journal->j_barrier_count; |
557 | write_unlock(&journal->j_state_lock); |
558 | wake_up(&journal->j_wait_transaction_locked); |
559 | } |
560 | |
561 | static void warn_dirty_buffer(struct buffer_head *bh) |
562 | { |
563 | char b[BDEVNAME_SIZE]; |
564 | |
565 | printk(KERN_WARNING |
566 | "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). " |
567 | "There's a risk of filesystem corruption in case of system " |
568 | "crash.\n", |
569 | bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr); |
570 | } |
571 | |
572 | /* |
573 | * If the buffer is already part of the current transaction, then there |
574 | * is nothing we need to do. If it is already part of a prior |
575 | * transaction which we are still committing to disk, then we need to |
576 | * make sure that we do not overwrite the old copy: we do copy-out to |
577 | * preserve the copy going to disk. We also account the buffer against |
578 | * the handle's metadata buffer credits (unless the buffer is already |
579 | * part of the transaction, that is). |
580 | * |
581 | */ |
582 | static int |
583 | do_get_write_access(handle_t *handle, struct journal_head *jh, |
584 | int force_copy) |
585 | { |
586 | struct buffer_head *bh; |
587 | transaction_t *transaction; |
588 | journal_t *journal; |
589 | int error; |
590 | char *frozen_buffer = NULL; |
591 | int need_copy = 0; |
592 | |
593 | if (is_handle_aborted(handle)) |
594 | return -EROFS; |
595 | |
596 | transaction = handle->h_transaction; |
597 | journal = transaction->t_journal; |
598 | |
599 | jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy); |
600 | |
601 | JBUFFER_TRACE(jh, "entry"); |
602 | repeat: |
603 | bh = jh2bh(jh); |
604 | |
605 | /* @@@ Need to check for errors here at some point. */ |
606 | |
607 | lock_buffer(bh); |
608 | jbd_lock_bh_state(bh); |
609 | |
610 | /* We now hold the buffer lock so it is safe to query the buffer |
611 | * state. Is the buffer dirty? |
612 | * |
613 | * If so, there are two possibilities. The buffer may be |
614 | * non-journaled, and undergoing a quite legitimate writeback. |
615 | * Otherwise, it is journaled, and we don't expect dirty buffers |
616 | * in that state (the buffers should be marked JBD_Dirty |
617 | * instead.) So either the IO is being done under our own |
618 | * control and this is a bug, or it's a third party IO such as |
619 | * dump(8) (which may leave the buffer scheduled for read --- |
620 | * ie. locked but not dirty) or tune2fs (which may actually have |
621 | * the buffer dirtied, ugh.) */ |
622 | |
623 | if (buffer_dirty(bh)) { |
624 | /* |
625 | * First question: is this buffer already part of the current |
626 | * transaction or the existing committing transaction? |
627 | */ |
628 | if (jh->b_transaction) { |
629 | J_ASSERT_JH(jh, |
630 | jh->b_transaction == transaction || |
631 | jh->b_transaction == |
632 | journal->j_committing_transaction); |
633 | if (jh->b_next_transaction) |
634 | J_ASSERT_JH(jh, jh->b_next_transaction == |
635 | transaction); |
636 | warn_dirty_buffer(bh); |
637 | } |
638 | /* |
639 | * In any case we need to clean the dirty flag and we must |
640 | * do it under the buffer lock to be sure we don't race |
641 | * with running write-out. |
642 | */ |
643 | JBUFFER_TRACE(jh, "Journalling dirty buffer"); |
644 | clear_buffer_dirty(bh); |
645 | set_buffer_jbddirty(bh); |
646 | } |
647 | |
648 | unlock_buffer(bh); |
649 | |
650 | error = -EROFS; |
651 | if (is_handle_aborted(handle)) { |
652 | jbd_unlock_bh_state(bh); |
653 | goto out; |
654 | } |
655 | error = 0; |
656 | |
657 | /* |
658 | * The buffer is already part of this transaction if b_transaction or |
659 | * b_next_transaction points to it |
660 | */ |
661 | if (jh->b_transaction == transaction || |
662 | jh->b_next_transaction == transaction) |
663 | goto done; |
664 | |
665 | /* |
666 | * this is the first time this transaction is touching this buffer, |
667 | * reset the modified flag |
668 | */ |
669 | jh->b_modified = 0; |
670 | |
671 | /* |
672 | * If there is already a copy-out version of this buffer, then we don't |
673 | * need to make another one |
674 | */ |
675 | if (jh->b_frozen_data) { |
676 | JBUFFER_TRACE(jh, "has frozen data"); |
677 | J_ASSERT_JH(jh, jh->b_next_transaction == NULL); |
678 | jh->b_next_transaction = transaction; |
679 | goto done; |
680 | } |
681 | |
682 | /* Is there data here we need to preserve? */ |
683 | |
684 | if (jh->b_transaction && jh->b_transaction != transaction) { |
685 | JBUFFER_TRACE(jh, "owned by older transaction"); |
686 | J_ASSERT_JH(jh, jh->b_next_transaction == NULL); |
687 | J_ASSERT_JH(jh, jh->b_transaction == |
688 | journal->j_committing_transaction); |
689 | |
690 | /* There is one case we have to be very careful about. |
691 | * If the committing transaction is currently writing |
692 | * this buffer out to disk and has NOT made a copy-out, |
693 | * then we cannot modify the buffer contents at all |
694 | * right now. The essence of copy-out is that it is the |
695 | * extra copy, not the primary copy, which gets |
696 | * journaled. If the primary copy is already going to |
697 | * disk then we cannot do copy-out here. */ |
698 | |
699 | if (jh->b_jlist == BJ_Shadow) { |
700 | DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow); |
701 | wait_queue_head_t *wqh; |
702 | |
703 | wqh = bit_waitqueue(&bh->b_state, BH_Unshadow); |
704 | |
705 | JBUFFER_TRACE(jh, "on shadow: sleep"); |
706 | jbd_unlock_bh_state(bh); |
707 | /* commit wakes up all shadow buffers after IO */ |
708 | for ( ; ; ) { |
709 | prepare_to_wait(wqh, &wait.wait, |
710 | TASK_UNINTERRUPTIBLE); |
711 | if (jh->b_jlist != BJ_Shadow) |
712 | break; |
713 | schedule(); |
714 | } |
715 | finish_wait(wqh, &wait.wait); |
716 | goto repeat; |
717 | } |
718 | |
719 | /* Only do the copy if the currently-owning transaction |
720 | * still needs it. If it is on the Forget list, the |
721 | * committing transaction is past that stage. The |
722 | * buffer had better remain locked during the kmalloc, |
723 | * but that should be true --- we hold the journal lock |
724 | * still and the buffer is already on the BUF_JOURNAL |
725 | * list so won't be flushed. |
726 | * |
727 | * Subtle point, though: if this is a get_undo_access, |
728 | * then we will be relying on the frozen_data to contain |
729 | * the new value of the committed_data record after the |
730 | * transaction, so we HAVE to force the frozen_data copy |
731 | * in that case. */ |
732 | |
733 | if (jh->b_jlist != BJ_Forget || force_copy) { |
734 | JBUFFER_TRACE(jh, "generate frozen data"); |
735 | if (!frozen_buffer) { |
736 | JBUFFER_TRACE(jh, "allocate memory for buffer"); |
737 | jbd_unlock_bh_state(bh); |
738 | frozen_buffer = |
739 | jbd2_alloc(jh2bh(jh)->b_size, |
740 | GFP_NOFS); |
741 | if (!frozen_buffer) { |
742 | printk(KERN_EMERG |
743 | "%s: OOM for frozen_buffer\n", |
744 | __func__); |
745 | JBUFFER_TRACE(jh, "oom!"); |
746 | error = -ENOMEM; |
747 | jbd_lock_bh_state(bh); |
748 | goto done; |
749 | } |
750 | goto repeat; |
751 | } |
752 | jh->b_frozen_data = frozen_buffer; |
753 | frozen_buffer = NULL; |
754 | need_copy = 1; |
755 | } |
756 | jh->b_next_transaction = transaction; |
757 | } |
758 | |
759 | |
760 | /* |
761 | * Finally, if the buffer is not journaled right now, we need to make |
762 | * sure it doesn't get written to disk before the caller actually |
763 | * commits the new data |
764 | */ |
765 | if (!jh->b_transaction) { |
766 | JBUFFER_TRACE(jh, "no transaction"); |
767 | J_ASSERT_JH(jh, !jh->b_next_transaction); |
768 | JBUFFER_TRACE(jh, "file as BJ_Reserved"); |
769 | spin_lock(&journal->j_list_lock); |
770 | __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved); |
771 | spin_unlock(&journal->j_list_lock); |
772 | } |
773 | |
774 | done: |
775 | if (need_copy) { |
776 | struct page *page; |
777 | int offset; |
778 | char *source; |
779 | |
780 | J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)), |
781 | "Possible IO failure.\n"); |
782 | page = jh2bh(jh)->b_page; |
783 | offset = offset_in_page(jh2bh(jh)->b_data); |
784 | source = kmap_atomic(page, KM_USER0); |
785 | /* Fire data frozen trigger just before we copy the data */ |
786 | jbd2_buffer_frozen_trigger(jh, source + offset, |
787 | jh->b_triggers); |
788 | memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size); |
789 | kunmap_atomic(source, KM_USER0); |
790 | |
791 | /* |
792 | * Now that the frozen data is saved off, we need to store |
793 | * any matching triggers. |
794 | */ |
795 | jh->b_frozen_triggers = jh->b_triggers; |
796 | } |
797 | jbd_unlock_bh_state(bh); |
798 | |
799 | /* |
800 | * If we are about to journal a buffer, then any revoke pending on it is |
801 | * no longer valid |
802 | */ |
803 | jbd2_journal_cancel_revoke(handle, jh); |
804 | |
805 | out: |
806 | if (unlikely(frozen_buffer)) /* It's usually NULL */ |
807 | jbd2_free(frozen_buffer, bh->b_size); |
808 | |
809 | JBUFFER_TRACE(jh, "exit"); |
810 | return error; |
811 | } |
812 | |
813 | /** |
814 | * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update. |
815 | * @handle: transaction to add buffer modifications to |
816 | * @bh: bh to be used for metadata writes |
817 | * |
818 | * Returns an error code or 0 on success. |
819 | * |
820 | * In full data journalling mode the buffer may be of type BJ_AsyncData, |
821 | * because we're write()ing a buffer which is also part of a shared mapping. |
822 | */ |
823 | |
824 | int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh) |
825 | { |
826 | struct journal_head *jh = jbd2_journal_add_journal_head(bh); |
827 | int rc; |
828 | |
829 | /* We do not want to get caught playing with fields which the |
830 | * log thread also manipulates. Make sure that the buffer |
831 | * completes any outstanding IO before proceeding. */ |
832 | rc = do_get_write_access(handle, jh, 0); |
833 | jbd2_journal_put_journal_head(jh); |
834 | return rc; |
835 | } |
836 | |
837 | |
838 | /* |
839 | * When the user wants to journal a newly created buffer_head |
840 | * (ie. getblk() returned a new buffer and we are going to populate it |
841 | * manually rather than reading off disk), then we need to keep the |
842 | * buffer_head locked until it has been completely filled with new |
843 | * data. In this case, we should be able to make the assertion that |
844 | * the bh is not already part of an existing transaction. |
845 | * |
846 | * The buffer should already be locked by the caller by this point. |
847 | * There is no lock ranking violation: it was a newly created, |
848 | * unlocked buffer beforehand. */ |
849 | |
850 | /** |
851 | * int jbd2_journal_get_create_access () - notify intent to use newly created bh |
852 | * @handle: transaction to new buffer to |
853 | * @bh: new buffer. |
854 | * |
855 | * Call this if you create a new bh. |
856 | */ |
857 | int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh) |
858 | { |
859 | transaction_t *transaction = handle->h_transaction; |
860 | journal_t *journal = transaction->t_journal; |
861 | struct journal_head *jh = jbd2_journal_add_journal_head(bh); |
862 | int err; |
863 | |
864 | jbd_debug(5, "journal_head %p\n", jh); |
865 | err = -EROFS; |
866 | if (is_handle_aborted(handle)) |
867 | goto out; |
868 | err = 0; |
869 | |
870 | JBUFFER_TRACE(jh, "entry"); |
871 | /* |
872 | * The buffer may already belong to this transaction due to pre-zeroing |
873 | * in the filesystem's new_block code. It may also be on the previous, |
874 | * committing transaction's lists, but it HAS to be in Forget state in |
875 | * that case: the transaction must have deleted the buffer for it to be |
876 | * reused here. |
877 | */ |
878 | jbd_lock_bh_state(bh); |
879 | spin_lock(&journal->j_list_lock); |
880 | J_ASSERT_JH(jh, (jh->b_transaction == transaction || |
881 | jh->b_transaction == NULL || |
882 | (jh->b_transaction == journal->j_committing_transaction && |
883 | jh->b_jlist == BJ_Forget))); |
884 | |
885 | J_ASSERT_JH(jh, jh->b_next_transaction == NULL); |
886 | J_ASSERT_JH(jh, buffer_locked(jh2bh(jh))); |
887 | |
888 | if (jh->b_transaction == NULL) { |
889 | /* |
890 | * Previous jbd2_journal_forget() could have left the buffer |
891 | * with jbddirty bit set because it was being committed. When |
892 | * the commit finished, we've filed the buffer for |
893 | * checkpointing and marked it dirty. Now we are reallocating |
894 | * the buffer so the transaction freeing it must have |
895 | * committed and so it's safe to clear the dirty bit. |
896 | */ |
897 | clear_buffer_dirty(jh2bh(jh)); |
898 | /* first access by this transaction */ |
899 | jh->b_modified = 0; |
900 | |
901 | JBUFFER_TRACE(jh, "file as BJ_Reserved"); |
902 | __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved); |
903 | } else if (jh->b_transaction == journal->j_committing_transaction) { |
904 | /* first access by this transaction */ |
905 | jh->b_modified = 0; |
906 | |
907 | JBUFFER_TRACE(jh, "set next transaction"); |
908 | jh->b_next_transaction = transaction; |
909 | } |
910 | spin_unlock(&journal->j_list_lock); |
911 | jbd_unlock_bh_state(bh); |
912 | |
913 | /* |
914 | * akpm: I added this. ext3_alloc_branch can pick up new indirect |
915 | * blocks which contain freed but then revoked metadata. We need |
916 | * to cancel the revoke in case we end up freeing it yet again |
917 | * and the reallocating as data - this would cause a second revoke, |
918 | * which hits an assertion error. |
919 | */ |
920 | JBUFFER_TRACE(jh, "cancelling revoke"); |
921 | jbd2_journal_cancel_revoke(handle, jh); |
922 | out: |
923 | jbd2_journal_put_journal_head(jh); |
924 | return err; |
925 | } |
926 | |
927 | /** |
928 | * int jbd2_journal_get_undo_access() - Notify intent to modify metadata with |
929 | * non-rewindable consequences |
930 | * @handle: transaction |
931 | * @bh: buffer to undo |
932 | * |
933 | * Sometimes there is a need to distinguish between metadata which has |
934 | * been committed to disk and that which has not. The ext3fs code uses |
935 | * this for freeing and allocating space, we have to make sure that we |
936 | * do not reuse freed space until the deallocation has been committed, |
937 | * since if we overwrote that space we would make the delete |
938 | * un-rewindable in case of a crash. |
939 | * |
940 | * To deal with that, jbd2_journal_get_undo_access requests write access to a |
941 | * buffer for parts of non-rewindable operations such as delete |
942 | * operations on the bitmaps. The journaling code must keep a copy of |
943 | * the buffer's contents prior to the undo_access call until such time |
944 | * as we know that the buffer has definitely been committed to disk. |
945 | * |
946 | * We never need to know which transaction the committed data is part |
947 | * of, buffers touched here are guaranteed to be dirtied later and so |
948 | * will be committed to a new transaction in due course, at which point |
949 | * we can discard the old committed data pointer. |
950 | * |
951 | * Returns error number or 0 on success. |
952 | */ |
953 | int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh) |
954 | { |
955 | int err; |
956 | struct journal_head *jh = jbd2_journal_add_journal_head(bh); |
957 | char *committed_data = NULL; |
958 | |
959 | JBUFFER_TRACE(jh, "entry"); |
960 | |
961 | /* |
962 | * Do this first --- it can drop the journal lock, so we want to |
963 | * make sure that obtaining the committed_data is done |
964 | * atomically wrt. completion of any outstanding commits. |
965 | */ |
966 | err = do_get_write_access(handle, jh, 1); |
967 | if (err) |
968 | goto out; |
969 | |
970 | repeat: |
971 | if (!jh->b_committed_data) { |
972 | committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS); |
973 | if (!committed_data) { |
974 | printk(KERN_EMERG "%s: No memory for committed data\n", |
975 | __func__); |
976 | err = -ENOMEM; |
977 | goto out; |
978 | } |
979 | } |
980 | |
981 | jbd_lock_bh_state(bh); |
982 | if (!jh->b_committed_data) { |
983 | /* Copy out the current buffer contents into the |
984 | * preserved, committed copy. */ |
985 | JBUFFER_TRACE(jh, "generate b_committed data"); |
986 | if (!committed_data) { |
987 | jbd_unlock_bh_state(bh); |
988 | goto repeat; |
989 | } |
990 | |
991 | jh->b_committed_data = committed_data; |
992 | committed_data = NULL; |
993 | memcpy(jh->b_committed_data, bh->b_data, bh->b_size); |
994 | } |
995 | jbd_unlock_bh_state(bh); |
996 | out: |
997 | jbd2_journal_put_journal_head(jh); |
998 | if (unlikely(committed_data)) |
999 | jbd2_free(committed_data, bh->b_size); |
1000 | return err; |
1001 | } |
1002 | |
1003 | /** |
1004 | * void jbd2_journal_set_triggers() - Add triggers for commit writeout |
1005 | * @bh: buffer to trigger on |
1006 | * @type: struct jbd2_buffer_trigger_type containing the trigger(s). |
1007 | * |
1008 | * Set any triggers on this journal_head. This is always safe, because |
1009 | * triggers for a committing buffer will be saved off, and triggers for |
1010 | * a running transaction will match the buffer in that transaction. |
1011 | * |
1012 | * Call with NULL to clear the triggers. |
1013 | */ |
1014 | void jbd2_journal_set_triggers(struct buffer_head *bh, |
1015 | struct jbd2_buffer_trigger_type *type) |
1016 | { |
1017 | struct journal_head *jh = bh2jh(bh); |
1018 | |
1019 | jh->b_triggers = type; |
1020 | } |
1021 | |
1022 | void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data, |
1023 | struct jbd2_buffer_trigger_type *triggers) |
1024 | { |
1025 | struct buffer_head *bh = jh2bh(jh); |
1026 | |
1027 | if (!triggers || !triggers->t_frozen) |
1028 | return; |
1029 | |
1030 | triggers->t_frozen(triggers, bh, mapped_data, bh->b_size); |
1031 | } |
1032 | |
1033 | void jbd2_buffer_abort_trigger(struct journal_head *jh, |
1034 | struct jbd2_buffer_trigger_type *triggers) |
1035 | { |
1036 | if (!triggers || !triggers->t_abort) |
1037 | return; |
1038 | |
1039 | triggers->t_abort(triggers, jh2bh(jh)); |
1040 | } |
1041 | |
1042 | |
1043 | |
1044 | /** |
1045 | * int jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata |
1046 | * @handle: transaction to add buffer to. |
1047 | * @bh: buffer to mark |
1048 | * |
1049 | * mark dirty metadata which needs to be journaled as part of the current |
1050 | * transaction. |
1051 | * |
1052 | * The buffer is placed on the transaction's metadata list and is marked |
1053 | * as belonging to the transaction. |
1054 | * |
1055 | * Returns error number or 0 on success. |
1056 | * |
1057 | * Special care needs to be taken if the buffer already belongs to the |
1058 | * current committing transaction (in which case we should have frozen |
1059 | * data present for that commit). In that case, we don't relink the |
1060 | * buffer: that only gets done when the old transaction finally |
1061 | * completes its commit. |
1062 | */ |
1063 | int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh) |
1064 | { |
1065 | transaction_t *transaction = handle->h_transaction; |
1066 | journal_t *journal = transaction->t_journal; |
1067 | struct journal_head *jh = bh2jh(bh); |
1068 | |
1069 | jbd_debug(5, "journal_head %p\n", jh); |
1070 | JBUFFER_TRACE(jh, "entry"); |
1071 | if (is_handle_aborted(handle)) |
1072 | goto out; |
1073 | |
1074 | jbd_lock_bh_state(bh); |
1075 | |
1076 | if (jh->b_modified == 0) { |
1077 | /* |
1078 | * This buffer's got modified and becoming part |
1079 | * of the transaction. This needs to be done |
1080 | * once a transaction -bzzz |
1081 | */ |
1082 | jh->b_modified = 1; |
1083 | J_ASSERT_JH(jh, handle->h_buffer_credits > 0); |
1084 | handle->h_buffer_credits--; |
1085 | } |
1086 | |
1087 | /* |
1088 | * fastpath, to avoid expensive locking. If this buffer is already |
1089 | * on the running transaction's metadata list there is nothing to do. |
1090 | * Nobody can take it off again because there is a handle open. |
1091 | * I _think_ we're OK here with SMP barriers - a mistaken decision will |
1092 | * result in this test being false, so we go in and take the locks. |
1093 | */ |
1094 | if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) { |
1095 | JBUFFER_TRACE(jh, "fastpath"); |
1096 | J_ASSERT_JH(jh, jh->b_transaction == |
1097 | journal->j_running_transaction); |
1098 | goto out_unlock_bh; |
1099 | } |
1100 | |
1101 | set_buffer_jbddirty(bh); |
1102 | |
1103 | /* |
1104 | * Metadata already on the current transaction list doesn't |
1105 | * need to be filed. Metadata on another transaction's list must |
1106 | * be committing, and will be refiled once the commit completes: |
1107 | * leave it alone for now. |
1108 | */ |
1109 | if (jh->b_transaction != transaction) { |
1110 | JBUFFER_TRACE(jh, "already on other transaction"); |
1111 | J_ASSERT_JH(jh, jh->b_transaction == |
1112 | journal->j_committing_transaction); |
1113 | J_ASSERT_JH(jh, jh->b_next_transaction == transaction); |
1114 | /* And this case is illegal: we can't reuse another |
1115 | * transaction's data buffer, ever. */ |
1116 | goto out_unlock_bh; |
1117 | } |
1118 | |
1119 | /* That test should have eliminated the following case: */ |
1120 | J_ASSERT_JH(jh, jh->b_frozen_data == NULL); |
1121 | |
1122 | JBUFFER_TRACE(jh, "file as BJ_Metadata"); |
1123 | spin_lock(&journal->j_list_lock); |
1124 | __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata); |
1125 | spin_unlock(&journal->j_list_lock); |
1126 | out_unlock_bh: |
1127 | jbd_unlock_bh_state(bh); |
1128 | out: |
1129 | JBUFFER_TRACE(jh, "exit"); |
1130 | return 0; |
1131 | } |
1132 | |
1133 | /* |
1134 | * jbd2_journal_release_buffer: undo a get_write_access without any buffer |
1135 | * updates, if the update decided in the end that it didn't need access. |
1136 | * |
1137 | */ |
1138 | void |
1139 | jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh) |
1140 | { |
1141 | BUFFER_TRACE(bh, "entry"); |
1142 | } |
1143 | |
1144 | /** |
1145 | * void jbd2_journal_forget() - bforget() for potentially-journaled buffers. |
1146 | * @handle: transaction handle |
1147 | * @bh: bh to 'forget' |
1148 | * |
1149 | * We can only do the bforget if there are no commits pending against the |
1150 | * buffer. If the buffer is dirty in the current running transaction we |
1151 | * can safely unlink it. |
1152 | * |
1153 | * bh may not be a journalled buffer at all - it may be a non-JBD |
1154 | * buffer which came off the hashtable. Check for this. |
1155 | * |
1156 | * Decrements bh->b_count by one. |
1157 | * |
1158 | * Allow this call even if the handle has aborted --- it may be part of |
1159 | * the caller's cleanup after an abort. |
1160 | */ |
1161 | int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh) |
1162 | { |
1163 | transaction_t *transaction = handle->h_transaction; |
1164 | journal_t *journal = transaction->t_journal; |
1165 | struct journal_head *jh; |
1166 | int drop_reserve = 0; |
1167 | int err = 0; |
1168 | int was_modified = 0; |
1169 | |
1170 | BUFFER_TRACE(bh, "entry"); |
1171 | |
1172 | jbd_lock_bh_state(bh); |
1173 | spin_lock(&journal->j_list_lock); |
1174 | |
1175 | if (!buffer_jbd(bh)) |
1176 | goto not_jbd; |
1177 | jh = bh2jh(bh); |
1178 | |
1179 | /* Critical error: attempting to delete a bitmap buffer, maybe? |
1180 | * Don't do any jbd operations, and return an error. */ |
1181 | if (!J_EXPECT_JH(jh, !jh->b_committed_data, |
1182 | "inconsistent data on disk")) { |
1183 | err = -EIO; |
1184 | goto not_jbd; |
1185 | } |
1186 | |
1187 | /* keep track of wether or not this transaction modified us */ |
1188 | was_modified = jh->b_modified; |
1189 | |
1190 | /* |
1191 | * The buffer's going from the transaction, we must drop |
1192 | * all references -bzzz |
1193 | */ |
1194 | jh->b_modified = 0; |
1195 | |
1196 | if (jh->b_transaction == handle->h_transaction) { |
1197 | J_ASSERT_JH(jh, !jh->b_frozen_data); |
1198 | |
1199 | /* If we are forgetting a buffer which is already part |
1200 | * of this transaction, then we can just drop it from |
1201 | * the transaction immediately. */ |
1202 | clear_buffer_dirty(bh); |
1203 | clear_buffer_jbddirty(bh); |
1204 | |
1205 | JBUFFER_TRACE(jh, "belongs to current transaction: unfile"); |
1206 | |
1207 | /* |
1208 | * we only want to drop a reference if this transaction |
1209 | * modified the buffer |
1210 | */ |
1211 | if (was_modified) |
1212 | drop_reserve = 1; |
1213 | |
1214 | /* |
1215 | * We are no longer going to journal this buffer. |
1216 | * However, the commit of this transaction is still |
1217 | * important to the buffer: the delete that we are now |
1218 | * processing might obsolete an old log entry, so by |
1219 | * committing, we can satisfy the buffer's checkpoint. |
1220 | * |
1221 | * So, if we have a checkpoint on the buffer, we should |
1222 | * now refile the buffer on our BJ_Forget list so that |
1223 | * we know to remove the checkpoint after we commit. |
1224 | */ |
1225 | |
1226 | if (jh->b_cp_transaction) { |
1227 | __jbd2_journal_temp_unlink_buffer(jh); |
1228 | __jbd2_journal_file_buffer(jh, transaction, BJ_Forget); |
1229 | } else { |
1230 | __jbd2_journal_unfile_buffer(jh); |
1231 | if (!buffer_jbd(bh)) { |
1232 | spin_unlock(&journal->j_list_lock); |
1233 | jbd_unlock_bh_state(bh); |
1234 | __bforget(bh); |
1235 | goto drop; |
1236 | } |
1237 | } |
1238 | } else if (jh->b_transaction) { |
1239 | J_ASSERT_JH(jh, (jh->b_transaction == |
1240 | journal->j_committing_transaction)); |
1241 | /* However, if the buffer is still owned by a prior |
1242 | * (committing) transaction, we can't drop it yet... */ |
1243 | JBUFFER_TRACE(jh, "belongs to older transaction"); |
1244 | /* ... but we CAN drop it from the new transaction if we |
1245 | * have also modified it since the original commit. */ |
1246 | |
1247 | if (jh->b_next_transaction) { |
1248 | J_ASSERT(jh->b_next_transaction == transaction); |
1249 | jh->b_next_transaction = NULL; |
1250 | |
1251 | /* |
1252 | * only drop a reference if this transaction modified |
1253 | * the buffer |
1254 | */ |
1255 | if (was_modified) |
1256 | drop_reserve = 1; |
1257 | } |
1258 | } |
1259 | |
1260 | not_jbd: |
1261 | spin_unlock(&journal->j_list_lock); |
1262 | jbd_unlock_bh_state(bh); |
1263 | __brelse(bh); |
1264 | drop: |
1265 | if (drop_reserve) { |
1266 | /* no need to reserve log space for this block -bzzz */ |
1267 | handle->h_buffer_credits++; |
1268 | } |
1269 | return err; |
1270 | } |
1271 | |
1272 | /** |
1273 | * int jbd2_journal_stop() - complete a transaction |
1274 | * @handle: tranaction to complete. |
1275 | * |
1276 | * All done for a particular handle. |
1277 | * |
1278 | * There is not much action needed here. We just return any remaining |
1279 | * buffer credits to the transaction and remove the handle. The only |
1280 | * complication is that we need to start a commit operation if the |
1281 | * filesystem is marked for synchronous update. |
1282 | * |
1283 | * jbd2_journal_stop itself will not usually return an error, but it may |
1284 | * do so in unusual circumstances. In particular, expect it to |
1285 | * return -EIO if a jbd2_journal_abort has been executed since the |
1286 | * transaction began. |
1287 | */ |
1288 | int jbd2_journal_stop(handle_t *handle) |
1289 | { |
1290 | transaction_t *transaction = handle->h_transaction; |
1291 | journal_t *journal = transaction->t_journal; |
1292 | int err, wait_for_commit = 0; |
1293 | tid_t tid; |
1294 | pid_t pid; |
1295 | |
1296 | J_ASSERT(journal_current_handle() == handle); |
1297 | |
1298 | if (is_handle_aborted(handle)) |
1299 | err = -EIO; |
1300 | else { |
1301 | J_ASSERT(atomic_read(&transaction->t_updates) > 0); |
1302 | err = 0; |
1303 | } |
1304 | |
1305 | if (--handle->h_ref > 0) { |
1306 | jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1, |
1307 | handle->h_ref); |
1308 | return err; |
1309 | } |
1310 | |
1311 | jbd_debug(4, "Handle %p going down\n", handle); |
1312 | |
1313 | /* |
1314 | * Implement synchronous transaction batching. If the handle |
1315 | * was synchronous, don't force a commit immediately. Let's |
1316 | * yield and let another thread piggyback onto this |
1317 | * transaction. Keep doing that while new threads continue to |
1318 | * arrive. It doesn't cost much - we're about to run a commit |
1319 | * and sleep on IO anyway. Speeds up many-threaded, many-dir |
1320 | * operations by 30x or more... |
1321 | * |
1322 | * We try and optimize the sleep time against what the |
1323 | * underlying disk can do, instead of having a static sleep |
1324 | * time. This is useful for the case where our storage is so |
1325 | * fast that it is more optimal to go ahead and force a flush |
1326 | * and wait for the transaction to be committed than it is to |
1327 | * wait for an arbitrary amount of time for new writers to |
1328 | * join the transaction. We achieve this by measuring how |
1329 | * long it takes to commit a transaction, and compare it with |
1330 | * how long this transaction has been running, and if run time |
1331 | * < commit time then we sleep for the delta and commit. This |
1332 | * greatly helps super fast disks that would see slowdowns as |
1333 | * more threads started doing fsyncs. |
1334 | * |
1335 | * But don't do this if this process was the most recent one |
1336 | * to perform a synchronous write. We do this to detect the |
1337 | * case where a single process is doing a stream of sync |
1338 | * writes. No point in waiting for joiners in that case. |
1339 | */ |
1340 | pid = current->pid; |
1341 | if (handle->h_sync && journal->j_last_sync_writer != pid) { |
1342 | u64 commit_time, trans_time; |
1343 | |
1344 | journal->j_last_sync_writer = pid; |
1345 | |
1346 | read_lock(&journal->j_state_lock); |
1347 | commit_time = journal->j_average_commit_time; |
1348 | read_unlock(&journal->j_state_lock); |
1349 | |
1350 | trans_time = ktime_to_ns(ktime_sub(ktime_get(), |
1351 | transaction->t_start_time)); |
1352 | |
1353 | commit_time = max_t(u64, commit_time, |
1354 | 1000*journal->j_min_batch_time); |
1355 | commit_time = min_t(u64, commit_time, |
1356 | 1000*journal->j_max_batch_time); |
1357 | |
1358 | if (trans_time < commit_time) { |
1359 | ktime_t expires = ktime_add_ns(ktime_get(), |
1360 | commit_time); |
1361 | set_current_state(TASK_UNINTERRUPTIBLE); |
1362 | schedule_hrtimeout(&expires, HRTIMER_MODE_ABS); |
1363 | } |
1364 | } |
1365 | |
1366 | if (handle->h_sync) |
1367 | transaction->t_synchronous_commit = 1; |
1368 | current->journal_info = NULL; |
1369 | atomic_sub(handle->h_buffer_credits, |
1370 | &transaction->t_outstanding_credits); |
1371 | |
1372 | /* |
1373 | * If the handle is marked SYNC, we need to set another commit |
1374 | * going! We also want to force a commit if the current |
1375 | * transaction is occupying too much of the log, or if the |
1376 | * transaction is too old now. |
1377 | */ |
1378 | if (handle->h_sync || |
1379 | (atomic_read(&transaction->t_outstanding_credits) > |
1380 | journal->j_max_transaction_buffers) || |
1381 | time_after_eq(jiffies, transaction->t_expires)) { |
1382 | /* Do this even for aborted journals: an abort still |
1383 | * completes the commit thread, it just doesn't write |
1384 | * anything to disk. */ |
1385 | |
1386 | jbd_debug(2, "transaction too old, requesting commit for " |
1387 | "handle %p\n", handle); |
1388 | /* This is non-blocking */ |
1389 | jbd2_log_start_commit(journal, transaction->t_tid); |
1390 | |
1391 | /* |
1392 | * Special case: JBD2_SYNC synchronous updates require us |
1393 | * to wait for the commit to complete. |
1394 | */ |
1395 | if (handle->h_sync && !(current->flags & PF_MEMALLOC)) |
1396 | wait_for_commit = 1; |
1397 | } |
1398 | |
1399 | /* |
1400 | * Once we drop t_updates, if it goes to zero the transaction |
1401 | * could start committing on us and eventually disappear. So |
1402 | * once we do this, we must not dereference transaction |
1403 | * pointer again. |
1404 | */ |
1405 | tid = transaction->t_tid; |
1406 | if (atomic_dec_and_test(&transaction->t_updates)) { |
1407 | wake_up(&journal->j_wait_updates); |
1408 | if (journal->j_barrier_count) |
1409 | wake_up(&journal->j_wait_transaction_locked); |
1410 | } |
1411 | |
1412 | if (wait_for_commit) |
1413 | err = jbd2_log_wait_commit(journal, tid); |
1414 | |
1415 | lock_map_release(&handle->h_lockdep_map); |
1416 | |
1417 | jbd2_free_handle(handle); |
1418 | return err; |
1419 | } |
1420 | |
1421 | /** |
1422 | * int jbd2_journal_force_commit() - force any uncommitted transactions |
1423 | * @journal: journal to force |
1424 | * |
1425 | * For synchronous operations: force any uncommitted transactions |
1426 | * to disk. May seem kludgy, but it reuses all the handle batching |
1427 | * code in a very simple manner. |
1428 | */ |
1429 | int jbd2_journal_force_commit(journal_t *journal) |
1430 | { |
1431 | handle_t *handle; |
1432 | int ret; |
1433 | |
1434 | handle = jbd2_journal_start(journal, 1); |
1435 | if (IS_ERR(handle)) { |
1436 | ret = PTR_ERR(handle); |
1437 | } else { |
1438 | handle->h_sync = 1; |
1439 | ret = jbd2_journal_stop(handle); |
1440 | } |
1441 | return ret; |
1442 | } |
1443 | |
1444 | /* |
1445 | * |
1446 | * List management code snippets: various functions for manipulating the |
1447 | * transaction buffer lists. |
1448 | * |
1449 | */ |
1450 | |
1451 | /* |
1452 | * Append a buffer to a transaction list, given the transaction's list head |
1453 | * pointer. |
1454 | * |
1455 | * j_list_lock is held. |
1456 | * |
1457 | * jbd_lock_bh_state(jh2bh(jh)) is held. |
1458 | */ |
1459 | |
1460 | static inline void |
1461 | __blist_add_buffer(struct journal_head **list, struct journal_head *jh) |
1462 | { |
1463 | if (!*list) { |
1464 | jh->b_tnext = jh->b_tprev = jh; |
1465 | *list = jh; |
1466 | } else { |
1467 | /* Insert at the tail of the list to preserve order */ |
1468 | struct journal_head *first = *list, *last = first->b_tprev; |
1469 | jh->b_tprev = last; |
1470 | jh->b_tnext = first; |
1471 | last->b_tnext = first->b_tprev = jh; |
1472 | } |
1473 | } |
1474 | |
1475 | /* |
1476 | * Remove a buffer from a transaction list, given the transaction's list |
1477 | * head pointer. |
1478 | * |
1479 | * Called with j_list_lock held, and the journal may not be locked. |
1480 | * |
1481 | * jbd_lock_bh_state(jh2bh(jh)) is held. |
1482 | */ |
1483 | |
1484 | static inline void |
1485 | __blist_del_buffer(struct journal_head **list, struct journal_head *jh) |
1486 | { |
1487 | if (*list == jh) { |
1488 | *list = jh->b_tnext; |
1489 | if (*list == jh) |
1490 | *list = NULL; |
1491 | } |
1492 | jh->b_tprev->b_tnext = jh->b_tnext; |
1493 | jh->b_tnext->b_tprev = jh->b_tprev; |
1494 | } |
1495 | |
1496 | /* |
1497 | * Remove a buffer from the appropriate transaction list. |
1498 | * |
1499 | * Note that this function can *change* the value of |
1500 | * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list, |
1501 | * t_log_list or t_reserved_list. If the caller is holding onto a copy of one |
1502 | * of these pointers, it could go bad. Generally the caller needs to re-read |
1503 | * the pointer from the transaction_t. |
1504 | * |
1505 | * Called under j_list_lock. The journal may not be locked. |
1506 | */ |
1507 | void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh) |
1508 | { |
1509 | struct journal_head **list = NULL; |
1510 | transaction_t *transaction; |
1511 | struct buffer_head *bh = jh2bh(jh); |
1512 | |
1513 | J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh)); |
1514 | transaction = jh->b_transaction; |
1515 | if (transaction) |
1516 | assert_spin_locked(&transaction->t_journal->j_list_lock); |
1517 | |
1518 | J_ASSERT_JH(jh, jh->b_jlist < BJ_Types); |
1519 | if (jh->b_jlist != BJ_None) |
1520 | J_ASSERT_JH(jh, transaction != NULL); |
1521 | |
1522 | switch (jh->b_jlist) { |
1523 | case BJ_None: |
1524 | return; |
1525 | case BJ_Metadata: |
1526 | transaction->t_nr_buffers--; |
1527 | J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0); |
1528 | list = &transaction->t_buffers; |
1529 | break; |
1530 | case BJ_Forget: |
1531 | list = &transaction->t_forget; |
1532 | break; |
1533 | case BJ_IO: |
1534 | list = &transaction->t_iobuf_list; |
1535 | break; |
1536 | case BJ_Shadow: |
1537 | list = &transaction->t_shadow_list; |
1538 | break; |
1539 | case BJ_LogCtl: |
1540 | list = &transaction->t_log_list; |
1541 | break; |
1542 | case BJ_Reserved: |
1543 | list = &transaction->t_reserved_list; |
1544 | break; |
1545 | } |
1546 | |
1547 | __blist_del_buffer(list, jh); |
1548 | jh->b_jlist = BJ_None; |
1549 | if (test_clear_buffer_jbddirty(bh)) |
1550 | mark_buffer_dirty(bh); /* Expose it to the VM */ |
1551 | } |
1552 | |
1553 | /* |
1554 | * Remove buffer from all transactions. |
1555 | * |
1556 | * Called with bh_state lock and j_list_lock |
1557 | * |
1558 | * jh and bh may be already freed when this function returns. |
1559 | */ |
1560 | static void __jbd2_journal_unfile_buffer(struct journal_head *jh) |
1561 | { |
1562 | __jbd2_journal_temp_unlink_buffer(jh); |
1563 | jh->b_transaction = NULL; |
1564 | jbd2_journal_put_journal_head(jh); |
1565 | } |
1566 | |
1567 | void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh) |
1568 | { |
1569 | struct buffer_head *bh = jh2bh(jh); |
1570 | |
1571 | /* Get reference so that buffer cannot be freed before we unlock it */ |
1572 | get_bh(bh); |
1573 | jbd_lock_bh_state(bh); |
1574 | spin_lock(&journal->j_list_lock); |
1575 | __jbd2_journal_unfile_buffer(jh); |
1576 | spin_unlock(&journal->j_list_lock); |
1577 | jbd_unlock_bh_state(bh); |
1578 | __brelse(bh); |
1579 | } |
1580 | |
1581 | /* |
1582 | * Called from jbd2_journal_try_to_free_buffers(). |
1583 | * |
1584 | * Called under jbd_lock_bh_state(bh) |
1585 | */ |
1586 | static void |
1587 | __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh) |
1588 | { |
1589 | struct journal_head *jh; |
1590 | |
1591 | jh = bh2jh(bh); |
1592 | |
1593 | if (buffer_locked(bh) || buffer_dirty(bh)) |
1594 | goto out; |
1595 | |
1596 | if (jh->b_next_transaction != NULL) |
1597 | goto out; |
1598 | |
1599 | spin_lock(&journal->j_list_lock); |
1600 | if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) { |
1601 | /* written-back checkpointed metadata buffer */ |
1602 | if (jh->b_jlist == BJ_None) { |
1603 | JBUFFER_TRACE(jh, "remove from checkpoint list"); |
1604 | __jbd2_journal_remove_checkpoint(jh); |
1605 | } |
1606 | } |
1607 | spin_unlock(&journal->j_list_lock); |
1608 | out: |
1609 | return; |
1610 | } |
1611 | |
1612 | /** |
1613 | * int jbd2_journal_try_to_free_buffers() - try to free page buffers. |
1614 | * @journal: journal for operation |
1615 | * @page: to try and free |
1616 | * @gfp_mask: we use the mask to detect how hard should we try to release |
1617 | * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to |
1618 | * release the buffers. |
1619 | * |
1620 | * |
1621 | * For all the buffers on this page, |
1622 | * if they are fully written out ordered data, move them onto BUF_CLEAN |
1623 | * so try_to_free_buffers() can reap them. |
1624 | * |
1625 | * This function returns non-zero if we wish try_to_free_buffers() |
1626 | * to be called. We do this if the page is releasable by try_to_free_buffers(). |
1627 | * We also do it if the page has locked or dirty buffers and the caller wants |
1628 | * us to perform sync or async writeout. |
1629 | * |
1630 | * This complicates JBD locking somewhat. We aren't protected by the |
1631 | * BKL here. We wish to remove the buffer from its committing or |
1632 | * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer. |
1633 | * |
1634 | * This may *change* the value of transaction_t->t_datalist, so anyone |
1635 | * who looks at t_datalist needs to lock against this function. |
1636 | * |
1637 | * Even worse, someone may be doing a jbd2_journal_dirty_data on this |
1638 | * buffer. So we need to lock against that. jbd2_journal_dirty_data() |
1639 | * will come out of the lock with the buffer dirty, which makes it |
1640 | * ineligible for release here. |
1641 | * |
1642 | * Who else is affected by this? hmm... Really the only contender |
1643 | * is do_get_write_access() - it could be looking at the buffer while |
1644 | * journal_try_to_free_buffer() is changing its state. But that |
1645 | * cannot happen because we never reallocate freed data as metadata |
1646 | * while the data is part of a transaction. Yes? |
1647 | * |
1648 | * Return 0 on failure, 1 on success |
1649 | */ |
1650 | int jbd2_journal_try_to_free_buffers(journal_t *journal, |
1651 | struct page *page, gfp_t gfp_mask) |
1652 | { |
1653 | struct buffer_head *head; |
1654 | struct buffer_head *bh; |
1655 | int ret = 0; |
1656 | |
1657 | J_ASSERT(PageLocked(page)); |
1658 | |
1659 | head = page_buffers(page); |
1660 | bh = head; |
1661 | do { |
1662 | struct journal_head *jh; |
1663 | |
1664 | /* |
1665 | * We take our own ref against the journal_head here to avoid |
1666 | * having to add tons of locking around each instance of |
1667 | * jbd2_journal_put_journal_head(). |
1668 | */ |
1669 | jh = jbd2_journal_grab_journal_head(bh); |
1670 | if (!jh) |
1671 | continue; |
1672 | |
1673 | jbd_lock_bh_state(bh); |
1674 | __journal_try_to_free_buffer(journal, bh); |
1675 | jbd2_journal_put_journal_head(jh); |
1676 | jbd_unlock_bh_state(bh); |
1677 | if (buffer_jbd(bh)) |
1678 | goto busy; |
1679 | } while ((bh = bh->b_this_page) != head); |
1680 | |
1681 | ret = try_to_free_buffers(page); |
1682 | |
1683 | busy: |
1684 | return ret; |
1685 | } |
1686 | |
1687 | /* |
1688 | * This buffer is no longer needed. If it is on an older transaction's |
1689 | * checkpoint list we need to record it on this transaction's forget list |
1690 | * to pin this buffer (and hence its checkpointing transaction) down until |
1691 | * this transaction commits. If the buffer isn't on a checkpoint list, we |
1692 | * release it. |
1693 | * Returns non-zero if JBD no longer has an interest in the buffer. |
1694 | * |
1695 | * Called under j_list_lock. |
1696 | * |
1697 | * Called under jbd_lock_bh_state(bh). |
1698 | */ |
1699 | static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction) |
1700 | { |
1701 | int may_free = 1; |
1702 | struct buffer_head *bh = jh2bh(jh); |
1703 | |
1704 | if (jh->b_cp_transaction) { |
1705 | JBUFFER_TRACE(jh, "on running+cp transaction"); |
1706 | __jbd2_journal_temp_unlink_buffer(jh); |
1707 | /* |
1708 | * We don't want to write the buffer anymore, clear the |
1709 | * bit so that we don't confuse checks in |
1710 | * __journal_file_buffer |
1711 | */ |
1712 | clear_buffer_dirty(bh); |
1713 | __jbd2_journal_file_buffer(jh, transaction, BJ_Forget); |
1714 | may_free = 0; |
1715 | } else { |
1716 | JBUFFER_TRACE(jh, "on running transaction"); |
1717 | __jbd2_journal_unfile_buffer(jh); |
1718 | } |
1719 | return may_free; |
1720 | } |
1721 | |
1722 | /* |
1723 | * jbd2_journal_invalidatepage |
1724 | * |
1725 | * This code is tricky. It has a number of cases to deal with. |
1726 | * |
1727 | * There are two invariants which this code relies on: |
1728 | * |
1729 | * i_size must be updated on disk before we start calling invalidatepage on the |
1730 | * data. |
1731 | * |
1732 | * This is done in ext3 by defining an ext3_setattr method which |
1733 | * updates i_size before truncate gets going. By maintaining this |
1734 | * invariant, we can be sure that it is safe to throw away any buffers |
1735 | * attached to the current transaction: once the transaction commits, |
1736 | * we know that the data will not be needed. |
1737 | * |
1738 | * Note however that we can *not* throw away data belonging to the |
1739 | * previous, committing transaction! |
1740 | * |
1741 | * Any disk blocks which *are* part of the previous, committing |
1742 | * transaction (and which therefore cannot be discarded immediately) are |
1743 | * not going to be reused in the new running transaction |
1744 | * |
1745 | * The bitmap committed_data images guarantee this: any block which is |
1746 | * allocated in one transaction and removed in the next will be marked |
1747 | * as in-use in the committed_data bitmap, so cannot be reused until |
1748 | * the next transaction to delete the block commits. This means that |
1749 | * leaving committing buffers dirty is quite safe: the disk blocks |
1750 | * cannot be reallocated to a different file and so buffer aliasing is |
1751 | * not possible. |
1752 | * |
1753 | * |
1754 | * The above applies mainly to ordered data mode. In writeback mode we |
1755 | * don't make guarantees about the order in which data hits disk --- in |
1756 | * particular we don't guarantee that new dirty data is flushed before |
1757 | * transaction commit --- so it is always safe just to discard data |
1758 | * immediately in that mode. --sct |
1759 | */ |
1760 | |
1761 | /* |
1762 | * The journal_unmap_buffer helper function returns zero if the buffer |
1763 | * concerned remains pinned as an anonymous buffer belonging to an older |
1764 | * transaction. |
1765 | * |
1766 | * We're outside-transaction here. Either or both of j_running_transaction |
1767 | * and j_committing_transaction may be NULL. |
1768 | */ |
1769 | static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh) |
1770 | { |
1771 | transaction_t *transaction; |
1772 | struct journal_head *jh; |
1773 | int may_free = 1; |
1774 | int ret; |
1775 | |
1776 | BUFFER_TRACE(bh, "entry"); |
1777 | |
1778 | /* |
1779 | * It is safe to proceed here without the j_list_lock because the |
1780 | * buffers cannot be stolen by try_to_free_buffers as long as we are |
1781 | * holding the page lock. --sct |
1782 | */ |
1783 | |
1784 | if (!buffer_jbd(bh)) |
1785 | goto zap_buffer_unlocked; |
1786 | |
1787 | /* OK, we have data buffer in journaled mode */ |
1788 | write_lock(&journal->j_state_lock); |
1789 | jbd_lock_bh_state(bh); |
1790 | spin_lock(&journal->j_list_lock); |
1791 | |
1792 | jh = jbd2_journal_grab_journal_head(bh); |
1793 | if (!jh) |
1794 | goto zap_buffer_no_jh; |
1795 | |
1796 | /* |
1797 | * We cannot remove the buffer from checkpoint lists until the |
1798 | * transaction adding inode to orphan list (let's call it T) |
1799 | * is committed. Otherwise if the transaction changing the |
1800 | * buffer would be cleaned from the journal before T is |
1801 | * committed, a crash will cause that the correct contents of |
1802 | * the buffer will be lost. On the other hand we have to |
1803 | * clear the buffer dirty bit at latest at the moment when the |
1804 | * transaction marking the buffer as freed in the filesystem |
1805 | * structures is committed because from that moment on the |
1806 | * buffer can be reallocated and used by a different page. |
1807 | * Since the block hasn't been freed yet but the inode has |
1808 | * already been added to orphan list, it is safe for us to add |
1809 | * the buffer to BJ_Forget list of the newest transaction. |
1810 | */ |
1811 | transaction = jh->b_transaction; |
1812 | if (transaction == NULL) { |
1813 | /* First case: not on any transaction. If it |
1814 | * has no checkpoint link, then we can zap it: |
1815 | * it's a writeback-mode buffer so we don't care |
1816 | * if it hits disk safely. */ |
1817 | if (!jh->b_cp_transaction) { |
1818 | JBUFFER_TRACE(jh, "not on any transaction: zap"); |
1819 | goto zap_buffer; |
1820 | } |
1821 | |
1822 | if (!buffer_dirty(bh)) { |
1823 | /* bdflush has written it. We can drop it now */ |
1824 | goto zap_buffer; |
1825 | } |
1826 | |
1827 | /* OK, it must be in the journal but still not |
1828 | * written fully to disk: it's metadata or |
1829 | * journaled data... */ |
1830 | |
1831 | if (journal->j_running_transaction) { |
1832 | /* ... and once the current transaction has |
1833 | * committed, the buffer won't be needed any |
1834 | * longer. */ |
1835 | JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget"); |
1836 | ret = __dispose_buffer(jh, |
1837 | journal->j_running_transaction); |
1838 | jbd2_journal_put_journal_head(jh); |
1839 | spin_unlock(&journal->j_list_lock); |
1840 | jbd_unlock_bh_state(bh); |
1841 | write_unlock(&journal->j_state_lock); |
1842 | return ret; |
1843 | } else { |
1844 | /* There is no currently-running transaction. So the |
1845 | * orphan record which we wrote for this file must have |
1846 | * passed into commit. We must attach this buffer to |
1847 | * the committing transaction, if it exists. */ |
1848 | if (journal->j_committing_transaction) { |
1849 | JBUFFER_TRACE(jh, "give to committing trans"); |
1850 | ret = __dispose_buffer(jh, |
1851 | journal->j_committing_transaction); |
1852 | jbd2_journal_put_journal_head(jh); |
1853 | spin_unlock(&journal->j_list_lock); |
1854 | jbd_unlock_bh_state(bh); |
1855 | write_unlock(&journal->j_state_lock); |
1856 | return ret; |
1857 | } else { |
1858 | /* The orphan record's transaction has |
1859 | * committed. We can cleanse this buffer */ |
1860 | clear_buffer_jbddirty(bh); |
1861 | goto zap_buffer; |
1862 | } |
1863 | } |
1864 | } else if (transaction == journal->j_committing_transaction) { |
1865 | JBUFFER_TRACE(jh, "on committing transaction"); |
1866 | /* |
1867 | * The buffer is committing, we simply cannot touch |
1868 | * it. So we just set j_next_transaction to the |
1869 | * running transaction (if there is one) and mark |
1870 | * buffer as freed so that commit code knows it should |
1871 | * clear dirty bits when it is done with the buffer. |
1872 | */ |
1873 | set_buffer_freed(bh); |
1874 | if (journal->j_running_transaction && buffer_jbddirty(bh)) |
1875 | jh->b_next_transaction = journal->j_running_transaction; |
1876 | jbd2_journal_put_journal_head(jh); |
1877 | spin_unlock(&journal->j_list_lock); |
1878 | jbd_unlock_bh_state(bh); |
1879 | write_unlock(&journal->j_state_lock); |
1880 | return 0; |
1881 | } else { |
1882 | /* Good, the buffer belongs to the running transaction. |
1883 | * We are writing our own transaction's data, not any |
1884 | * previous one's, so it is safe to throw it away |
1885 | * (remember that we expect the filesystem to have set |
1886 | * i_size already for this truncate so recovery will not |
1887 | * expose the disk blocks we are discarding here.) */ |
1888 | J_ASSERT_JH(jh, transaction == journal->j_running_transaction); |
1889 | JBUFFER_TRACE(jh, "on running transaction"); |
1890 | may_free = __dispose_buffer(jh, transaction); |
1891 | } |
1892 | |
1893 | zap_buffer: |
1894 | jbd2_journal_put_journal_head(jh); |
1895 | zap_buffer_no_jh: |
1896 | spin_unlock(&journal->j_list_lock); |
1897 | jbd_unlock_bh_state(bh); |
1898 | write_unlock(&journal->j_state_lock); |
1899 | zap_buffer_unlocked: |
1900 | clear_buffer_dirty(bh); |
1901 | J_ASSERT_BH(bh, !buffer_jbddirty(bh)); |
1902 | clear_buffer_mapped(bh); |
1903 | clear_buffer_req(bh); |
1904 | clear_buffer_new(bh); |
1905 | bh->b_bdev = NULL; |
1906 | return may_free; |
1907 | } |
1908 | |
1909 | /** |
1910 | * void jbd2_journal_invalidatepage() |
1911 | * @journal: journal to use for flush... |
1912 | * @page: page to flush |
1913 | * @offset: length of page to invalidate. |
1914 | * |
1915 | * Reap page buffers containing data after offset in page. |
1916 | * |
1917 | */ |
1918 | void jbd2_journal_invalidatepage(journal_t *journal, |
1919 | struct page *page, |
1920 | unsigned long offset) |
1921 | { |
1922 | struct buffer_head *head, *bh, *next; |
1923 | unsigned int curr_off = 0; |
1924 | int may_free = 1; |
1925 | |
1926 | if (!PageLocked(page)) |
1927 | BUG(); |
1928 | if (!page_has_buffers(page)) |
1929 | return; |
1930 | |
1931 | /* We will potentially be playing with lists other than just the |
1932 | * data lists (especially for journaled data mode), so be |
1933 | * cautious in our locking. */ |
1934 | |
1935 | head = bh = page_buffers(page); |
1936 | do { |
1937 | unsigned int next_off = curr_off + bh->b_size; |
1938 | next = bh->b_this_page; |
1939 | |
1940 | if (offset <= curr_off) { |
1941 | /* This block is wholly outside the truncation point */ |
1942 | lock_buffer(bh); |
1943 | may_free &= journal_unmap_buffer(journal, bh); |
1944 | unlock_buffer(bh); |
1945 | } |
1946 | curr_off = next_off; |
1947 | bh = next; |
1948 | |
1949 | } while (bh != head); |
1950 | |
1951 | if (!offset) { |
1952 | if (may_free && try_to_free_buffers(page)) |
1953 | J_ASSERT(!page_has_buffers(page)); |
1954 | } |
1955 | } |
1956 | |
1957 | /* |
1958 | * File a buffer on the given transaction list. |
1959 | */ |
1960 | void __jbd2_journal_file_buffer(struct journal_head *jh, |
1961 | transaction_t *transaction, int jlist) |
1962 | { |
1963 | struct journal_head **list = NULL; |
1964 | int was_dirty = 0; |
1965 | struct buffer_head *bh = jh2bh(jh); |
1966 | |
1967 | J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh)); |
1968 | assert_spin_locked(&transaction->t_journal->j_list_lock); |
1969 | |
1970 | J_ASSERT_JH(jh, jh->b_jlist < BJ_Types); |
1971 | J_ASSERT_JH(jh, jh->b_transaction == transaction || |
1972 | jh->b_transaction == NULL); |
1973 | |
1974 | if (jh->b_transaction && jh->b_jlist == jlist) |
1975 | return; |
1976 | |
1977 | if (jlist == BJ_Metadata || jlist == BJ_Reserved || |
1978 | jlist == BJ_Shadow || jlist == BJ_Forget) { |
1979 | /* |
1980 | * For metadata buffers, we track dirty bit in buffer_jbddirty |
1981 | * instead of buffer_dirty. We should not see a dirty bit set |
1982 | * here because we clear it in do_get_write_access but e.g. |
1983 | * tune2fs can modify the sb and set the dirty bit at any time |
1984 | * so we try to gracefully handle that. |
1985 | */ |
1986 | if (buffer_dirty(bh)) |
1987 | warn_dirty_buffer(bh); |
1988 | if (test_clear_buffer_dirty(bh) || |
1989 | test_clear_buffer_jbddirty(bh)) |
1990 | was_dirty = 1; |
1991 | } |
1992 | |
1993 | if (jh->b_transaction) |
1994 | __jbd2_journal_temp_unlink_buffer(jh); |
1995 | else |
1996 | jbd2_journal_grab_journal_head(bh); |
1997 | jh->b_transaction = transaction; |
1998 | |
1999 | switch (jlist) { |
2000 | case BJ_None: |
2001 | J_ASSERT_JH(jh, !jh->b_committed_data); |
2002 | J_ASSERT_JH(jh, !jh->b_frozen_data); |
2003 | return; |
2004 | case BJ_Metadata: |
2005 | transaction->t_nr_buffers++; |
2006 | list = &transaction->t_buffers; |
2007 | break; |
2008 | case BJ_Forget: |
2009 | list = &transaction->t_forget; |
2010 | break; |
2011 | case BJ_IO: |
2012 | list = &transaction->t_iobuf_list; |
2013 | break; |
2014 | case BJ_Shadow: |
2015 | list = &transaction->t_shadow_list; |
2016 | break; |
2017 | case BJ_LogCtl: |
2018 | list = &transaction->t_log_list; |
2019 | break; |
2020 | case BJ_Reserved: |
2021 | list = &transaction->t_reserved_list; |
2022 | break; |
2023 | } |
2024 | |
2025 | __blist_add_buffer(list, jh); |
2026 | jh->b_jlist = jlist; |
2027 | |
2028 | if (was_dirty) |
2029 | set_buffer_jbddirty(bh); |
2030 | } |
2031 | |
2032 | void jbd2_journal_file_buffer(struct journal_head *jh, |
2033 | transaction_t *transaction, int jlist) |
2034 | { |
2035 | jbd_lock_bh_state(jh2bh(jh)); |
2036 | spin_lock(&transaction->t_journal->j_list_lock); |
2037 | __jbd2_journal_file_buffer(jh, transaction, jlist); |
2038 | spin_unlock(&transaction->t_journal->j_list_lock); |
2039 | jbd_unlock_bh_state(jh2bh(jh)); |
2040 | } |
2041 | |
2042 | /* |
2043 | * Remove a buffer from its current buffer list in preparation for |
2044 | * dropping it from its current transaction entirely. If the buffer has |
2045 | * already started to be used by a subsequent transaction, refile the |
2046 | * buffer on that transaction's metadata list. |
2047 | * |
2048 | * Called under j_list_lock |
2049 | * Called under jbd_lock_bh_state(jh2bh(jh)) |
2050 | * |
2051 | * jh and bh may be already free when this function returns |
2052 | */ |
2053 | void __jbd2_journal_refile_buffer(struct journal_head *jh) |
2054 | { |
2055 | int was_dirty, jlist; |
2056 | struct buffer_head *bh = jh2bh(jh); |
2057 | |
2058 | J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh)); |
2059 | if (jh->b_transaction) |
2060 | assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock); |
2061 | |
2062 | /* If the buffer is now unused, just drop it. */ |
2063 | if (jh->b_next_transaction == NULL) { |
2064 | __jbd2_journal_unfile_buffer(jh); |
2065 | return; |
2066 | } |
2067 | |
2068 | /* |
2069 | * It has been modified by a later transaction: add it to the new |
2070 | * transaction's metadata list. |
2071 | */ |
2072 | |
2073 | was_dirty = test_clear_buffer_jbddirty(bh); |
2074 | __jbd2_journal_temp_unlink_buffer(jh); |
2075 | /* |
2076 | * We set b_transaction here because b_next_transaction will inherit |
2077 | * our jh reference and thus __jbd2_journal_file_buffer() must not |
2078 | * take a new one. |
2079 | */ |
2080 | jh->b_transaction = jh->b_next_transaction; |
2081 | jh->b_next_transaction = NULL; |
2082 | if (buffer_freed(bh)) |
2083 | jlist = BJ_Forget; |
2084 | else if (jh->b_modified) |
2085 | jlist = BJ_Metadata; |
2086 | else |
2087 | jlist = BJ_Reserved; |
2088 | __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist); |
2089 | J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING); |
2090 | |
2091 | if (was_dirty) |
2092 | set_buffer_jbddirty(bh); |
2093 | } |
2094 | |
2095 | /* |
2096 | * __jbd2_journal_refile_buffer() with necessary locking added. We take our |
2097 | * bh reference so that we can safely unlock bh. |
2098 | * |
2099 | * The jh and bh may be freed by this call. |
2100 | */ |
2101 | void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh) |
2102 | { |
2103 | struct buffer_head *bh = jh2bh(jh); |
2104 | |
2105 | /* Get reference so that buffer cannot be freed before we unlock it */ |
2106 | get_bh(bh); |
2107 | jbd_lock_bh_state(bh); |
2108 | spin_lock(&journal->j_list_lock); |
2109 | __jbd2_journal_refile_buffer(jh); |
2110 | jbd_unlock_bh_state(bh); |
2111 | spin_unlock(&journal->j_list_lock); |
2112 | __brelse(bh); |
2113 | } |
2114 | |
2115 | /* |
2116 | * File inode in the inode list of the handle's transaction |
2117 | */ |
2118 | int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode) |
2119 | { |
2120 | transaction_t *transaction = handle->h_transaction; |
2121 | journal_t *journal = transaction->t_journal; |
2122 | |
2123 | if (is_handle_aborted(handle)) |
2124 | return -EIO; |
2125 | |
2126 | jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino, |
2127 | transaction->t_tid); |
2128 | |
2129 | /* |
2130 | * First check whether inode isn't already on the transaction's |
2131 | * lists without taking the lock. Note that this check is safe |
2132 | * without the lock as we cannot race with somebody removing inode |
2133 | * from the transaction. The reason is that we remove inode from the |
2134 | * transaction only in journal_release_jbd_inode() and when we commit |
2135 | * the transaction. We are guarded from the first case by holding |
2136 | * a reference to the inode. We are safe against the second case |
2137 | * because if jinode->i_transaction == transaction, commit code |
2138 | * cannot touch the transaction because we hold reference to it, |
2139 | * and if jinode->i_next_transaction == transaction, commit code |
2140 | * will only file the inode where we want it. |
2141 | */ |
2142 | if (jinode->i_transaction == transaction || |
2143 | jinode->i_next_transaction == transaction) |
2144 | return 0; |
2145 | |
2146 | spin_lock(&journal->j_list_lock); |
2147 | |
2148 | if (jinode->i_transaction == transaction || |
2149 | jinode->i_next_transaction == transaction) |
2150 | goto done; |
2151 | |
2152 | /* |
2153 | * We only ever set this variable to 1 so the test is safe. Since |
2154 | * t_need_data_flush is likely to be set, we do the test to save some |
2155 | * cacheline bouncing |
2156 | */ |
2157 | if (!transaction->t_need_data_flush) |
2158 | transaction->t_need_data_flush = 1; |
2159 | /* On some different transaction's list - should be |
2160 | * the committing one */ |
2161 | if (jinode->i_transaction) { |
2162 | J_ASSERT(jinode->i_next_transaction == NULL); |
2163 | J_ASSERT(jinode->i_transaction == |
2164 | journal->j_committing_transaction); |
2165 | jinode->i_next_transaction = transaction; |
2166 | goto done; |
2167 | } |
2168 | /* Not on any transaction list... */ |
2169 | J_ASSERT(!jinode->i_next_transaction); |
2170 | jinode->i_transaction = transaction; |
2171 | list_add(&jinode->i_list, &transaction->t_inode_list); |
2172 | done: |
2173 | spin_unlock(&journal->j_list_lock); |
2174 | |
2175 | return 0; |
2176 | } |
2177 | |
2178 | /* |
2179 | * File truncate and transaction commit interact with each other in a |
2180 | * non-trivial way. If a transaction writing data block A is |
2181 | * committing, we cannot discard the data by truncate until we have |
2182 | * written them. Otherwise if we crashed after the transaction with |
2183 | * write has committed but before the transaction with truncate has |
2184 | * committed, we could see stale data in block A. This function is a |
2185 | * helper to solve this problem. It starts writeout of the truncated |
2186 | * part in case it is in the committing transaction. |
2187 | * |
2188 | * Filesystem code must call this function when inode is journaled in |
2189 | * ordered mode before truncation happens and after the inode has been |
2190 | * placed on orphan list with the new inode size. The second condition |
2191 | * avoids the race that someone writes new data and we start |
2192 | * committing the transaction after this function has been called but |
2193 | * before a transaction for truncate is started (and furthermore it |
2194 | * allows us to optimize the case where the addition to orphan list |
2195 | * happens in the same transaction as write --- we don't have to write |
2196 | * any data in such case). |
2197 | */ |
2198 | int jbd2_journal_begin_ordered_truncate(journal_t *journal, |
2199 | struct jbd2_inode *jinode, |
2200 | loff_t new_size) |
2201 | { |
2202 | transaction_t *inode_trans, *commit_trans; |
2203 | int ret = 0; |
2204 | |
2205 | /* This is a quick check to avoid locking if not necessary */ |
2206 | if (!jinode->i_transaction) |
2207 | goto out; |
2208 | /* Locks are here just to force reading of recent values, it is |
2209 | * enough that the transaction was not committing before we started |
2210 | * a transaction adding the inode to orphan list */ |
2211 | read_lock(&journal->j_state_lock); |
2212 | commit_trans = journal->j_committing_transaction; |
2213 | read_unlock(&journal->j_state_lock); |
2214 | spin_lock(&journal->j_list_lock); |
2215 | inode_trans = jinode->i_transaction; |
2216 | spin_unlock(&journal->j_list_lock); |
2217 | if (inode_trans == commit_trans) { |
2218 | ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping, |
2219 | new_size, LLONG_MAX); |
2220 | if (ret) |
2221 | jbd2_journal_abort(journal, ret); |
2222 | } |
2223 | out: |
2224 | return ret; |
2225 | } |
2226 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9