blob: cceaf57e37781304df0b553ca2f8e9957fd65ecb [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/fs/transaction.c
Mingming Caoae6ddcc2006-09-27 01:49:27 -07003 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 * 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
Mingming Caoae6ddcc2006-09-27 01:49:27 -070013 * journaling system.
Linus Torvalds1da177e2005-04-16 15:20:36 -070014 *
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/jbd.h>
23#include <linux/errno.h>
24#include <linux/slab.h>
25#include <linux/timer.h>
26#include <linux/smp_lock.h>
27#include <linux/mm.h>
28#include <linux/highmem.h>
29
Adrian Bunkd394e122006-12-06 20:38:26 -080030static void __journal_temp_unlink_buffer(struct journal_head *jh);
31
Linus Torvalds1da177e2005-04-16 15:20:36 -070032/*
33 * get_transaction: obtain a new transaction_t object.
34 *
35 * Simply allocate and initialise a new transaction. Create it in
36 * RUNNING state and add it to the current journal (which should not
37 * have an existing running transaction: we only make a new transaction
38 * once we have started to commit the old one).
39 *
40 * Preconditions:
41 * The journal MUST be locked. We don't perform atomic mallocs on the
42 * new transaction and we can't block without protecting against other
43 * processes trying to touch the journal while it is in transition.
44 *
45 * Called under j_state_lock
46 */
47
48static transaction_t *
49get_transaction(journal_t *journal, transaction_t *transaction)
50{
51 transaction->t_journal = journal;
52 transaction->t_state = T_RUNNING;
53 transaction->t_tid = journal->j_transaction_sequence++;
54 transaction->t_expires = jiffies + journal->j_commit_interval;
55 spin_lock_init(&transaction->t_handle_lock);
56
57 /* Set up the commit timer for the new transaction. */
Arjan van de Ven44d306e2006-12-10 02:21:26 -080058 journal->j_commit_timer.expires = round_jiffies(transaction->t_expires);
Andrew Mortone3df1892006-03-25 03:06:53 -080059 add_timer(&journal->j_commit_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
61 J_ASSERT(journal->j_running_transaction == NULL);
62 journal->j_running_transaction = transaction;
63
64 return transaction;
65}
66
67/*
68 * Handle management.
69 *
70 * A handle_t is an object which represents a single atomic update to a
71 * filesystem, and which tracks all of the modifications which form part
72 * of that one update.
73 */
74
75/*
76 * start_this_handle: Given a handle, deal with any locking or stalling
77 * needed to make sure that there is enough journal space for the handle
78 * to begin. Attach the handle to a transaction and set up the
Mingming Caoae6ddcc2006-09-27 01:49:27 -070079 * transaction's buffer credits.
Linus Torvalds1da177e2005-04-16 15:20:36 -070080 */
81
82static int start_this_handle(journal_t *journal, handle_t *handle)
83{
84 transaction_t *transaction;
85 int needed;
86 int nblocks = handle->h_buffer_credits;
87 transaction_t *new_transaction = NULL;
88 int ret = 0;
89
90 if (nblocks > journal->j_max_transaction_buffers) {
91 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
92 current->comm, nblocks,
93 journal->j_max_transaction_buffers);
94 ret = -ENOSPC;
95 goto out;
96 }
97
98alloc_transaction:
99 if (!journal->j_running_transaction) {
100 new_transaction = jbd_kmalloc(sizeof(*new_transaction),
101 GFP_NOFS);
102 if (!new_transaction) {
103 ret = -ENOMEM;
104 goto out;
105 }
106 memset(new_transaction, 0, sizeof(*new_transaction));
107 }
108
109 jbd_debug(3, "New handle %p going live.\n", handle);
110
111repeat:
112
113 /*
114 * We need to hold j_state_lock until t_updates has been incremented,
115 * for proper journal barrier handling
116 */
117 spin_lock(&journal->j_state_lock);
118repeat_locked:
119 if (is_journal_aborted(journal) ||
120 (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
121 spin_unlock(&journal->j_state_lock);
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700122 ret = -EROFS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123 goto out;
124 }
125
126 /* Wait on the journal's transaction barrier if necessary */
127 if (journal->j_barrier_count) {
128 spin_unlock(&journal->j_state_lock);
129 wait_event(journal->j_wait_transaction_locked,
130 journal->j_barrier_count == 0);
131 goto repeat;
132 }
133
134 if (!journal->j_running_transaction) {
135 if (!new_transaction) {
136 spin_unlock(&journal->j_state_lock);
137 goto alloc_transaction;
138 }
139 get_transaction(journal, new_transaction);
140 new_transaction = NULL;
141 }
142
143 transaction = journal->j_running_transaction;
144
145 /*
146 * If the current transaction is locked down for commit, wait for the
147 * lock to be released.
148 */
149 if (transaction->t_state == T_LOCKED) {
150 DEFINE_WAIT(wait);
151
152 prepare_to_wait(&journal->j_wait_transaction_locked,
153 &wait, TASK_UNINTERRUPTIBLE);
154 spin_unlock(&journal->j_state_lock);
155 schedule();
156 finish_wait(&journal->j_wait_transaction_locked, &wait);
157 goto repeat;
158 }
159
160 /*
161 * If there is not enough space left in the log to write all potential
162 * buffers requested by this operation, we need to stall pending a log
163 * checkpoint to free some more log space.
164 */
165 spin_lock(&transaction->t_handle_lock);
166 needed = transaction->t_outstanding_credits + nblocks;
167
168 if (needed > journal->j_max_transaction_buffers) {
169 /*
170 * If the current transaction is already too large, then start
171 * to commit it: we can then go back and attach this handle to
172 * a new transaction.
173 */
174 DEFINE_WAIT(wait);
175
176 jbd_debug(2, "Handle %p starting new commit...\n", handle);
177 spin_unlock(&transaction->t_handle_lock);
178 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
179 TASK_UNINTERRUPTIBLE);
180 __log_start_commit(journal, transaction->t_tid);
181 spin_unlock(&journal->j_state_lock);
182 schedule();
183 finish_wait(&journal->j_wait_transaction_locked, &wait);
184 goto repeat;
185 }
186
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700187 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 * The commit code assumes that it can get enough log space
189 * without forcing a checkpoint. This is *critical* for
190 * correctness: a checkpoint of a buffer which is also
191 * associated with a committing transaction creates a deadlock,
192 * so commit simply cannot force through checkpoints.
193 *
194 * We must therefore ensure the necessary space in the journal
195 * *before* starting to dirty potentially checkpointed buffers
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700196 * in the new transaction.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197 *
198 * The worst part is, any transaction currently committing can
199 * reduce the free space arbitrarily. Be careful to account for
200 * those buffers when checkpointing.
201 */
202
203 /*
204 * @@@ AKPM: This seems rather over-defensive. We're giving commit
205 * a _lot_ of headroom: 1/4 of the journal plus the size of
206 * the committing transaction. Really, we only need to give it
207 * committing_transaction->t_outstanding_credits plus "enough" for
208 * the log control blocks.
209 * Also, this test is inconsitent with the matching one in
210 * journal_extend().
211 */
212 if (__log_space_left(journal) < jbd_space_needed(journal)) {
213 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
214 spin_unlock(&transaction->t_handle_lock);
215 __log_wait_for_space(journal);
216 goto repeat_locked;
217 }
218
219 /* OK, account for the buffers that this operation expects to
220 * use and add the handle to the running transaction. */
221
222 handle->h_transaction = transaction;
223 transaction->t_outstanding_credits += nblocks;
224 transaction->t_updates++;
225 transaction->t_handle_count++;
226 jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
227 handle, nblocks, transaction->t_outstanding_credits,
228 __log_space_left(journal));
229 spin_unlock(&transaction->t_handle_lock);
230 spin_unlock(&journal->j_state_lock);
231out:
Andrew Morton304c4c82006-06-23 02:05:31 -0700232 if (unlikely(new_transaction)) /* It's usually NULL */
233 kfree(new_transaction);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 return ret;
235}
236
237/* Allocate a new handle. This should probably be in a slab... */
238static handle_t *new_handle(int nblocks)
239{
240 handle_t *handle = jbd_alloc_handle(GFP_NOFS);
241 if (!handle)
242 return NULL;
243 memset(handle, 0, sizeof(*handle));
244 handle->h_buffer_credits = nblocks;
245 handle->h_ref = 1;
246
247 return handle;
248}
249
250/**
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700251 * handle_t *journal_start() - Obtain a new handle.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 * @journal: Journal to start transaction on.
253 * @nblocks: number of block buffer we might modify
254 *
255 * We make sure that the transaction can guarantee at least nblocks of
256 * modified buffers in the log. We block until the log can guarantee
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700257 * that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 *
259 * This function is visible to journal users (like ext3fs), so is not
260 * called with the journal already locked.
261 *
262 * Return a pointer to a newly allocated handle, or NULL on failure
263 */
264handle_t *journal_start(journal_t *journal, int nblocks)
265{
266 handle_t *handle = journal_current_handle();
267 int err;
268
269 if (!journal)
270 return ERR_PTR(-EROFS);
271
272 if (handle) {
273 J_ASSERT(handle->h_transaction->t_journal == journal);
274 handle->h_ref++;
275 return handle;
276 }
277
278 handle = new_handle(nblocks);
279 if (!handle)
280 return ERR_PTR(-ENOMEM);
281
282 current->journal_info = handle;
283
284 err = start_this_handle(journal, handle);
285 if (err < 0) {
286 jbd_free_handle(handle);
287 current->journal_info = NULL;
288 handle = ERR_PTR(err);
289 }
290 return handle;
291}
292
293/**
294 * int journal_extend() - extend buffer credits.
295 * @handle: handle to 'extend'
296 * @nblocks: nr blocks to try to extend by.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700297 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 * Some transactions, such as large extends and truncates, can be done
299 * atomically all at once or in several stages. The operation requests
300 * a credit for a number of buffer modications in advance, but can
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700301 * extend its credit if it needs more.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302 *
303 * journal_extend tries to give the running handle more buffer credits.
304 * It does not guarantee that allocation - this is a best-effort only.
305 * The calling process MUST be able to deal cleanly with a failure to
306 * extend here.
307 *
308 * Return 0 on success, non-zero on failure.
309 *
310 * return code < 0 implies an error
311 * return code > 0 implies normal transaction-full status.
312 */
313int journal_extend(handle_t *handle, int nblocks)
314{
315 transaction_t *transaction = handle->h_transaction;
316 journal_t *journal = transaction->t_journal;
317 int result;
318 int wanted;
319
320 result = -EIO;
321 if (is_handle_aborted(handle))
322 goto out;
323
324 result = 1;
325
326 spin_lock(&journal->j_state_lock);
327
328 /* Don't extend a locked-down transaction! */
329 if (handle->h_transaction->t_state != T_RUNNING) {
330 jbd_debug(3, "denied handle %p %d blocks: "
331 "transaction not running\n", handle, nblocks);
332 goto error_out;
333 }
334
335 spin_lock(&transaction->t_handle_lock);
336 wanted = transaction->t_outstanding_credits + nblocks;
337
338 if (wanted > journal->j_max_transaction_buffers) {
339 jbd_debug(3, "denied handle %p %d blocks: "
340 "transaction too large\n", handle, nblocks);
341 goto unlock;
342 }
343
344 if (wanted > __log_space_left(journal)) {
345 jbd_debug(3, "denied handle %p %d blocks: "
346 "insufficient log space\n", handle, nblocks);
347 goto unlock;
348 }
349
350 handle->h_buffer_credits += nblocks;
351 transaction->t_outstanding_credits += nblocks;
352 result = 0;
353
354 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
355unlock:
356 spin_unlock(&transaction->t_handle_lock);
357error_out:
358 spin_unlock(&journal->j_state_lock);
359out:
360 return result;
361}
362
363
364/**
365 * int journal_restart() - restart a handle .
366 * @handle: handle to restart
367 * @nblocks: nr credits requested
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700368 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 * Restart a handle for a multi-transaction filesystem
370 * operation.
371 *
372 * If the journal_extend() call above fails to grant new buffer credits
373 * to a running handle, a call to journal_restart will commit the
374 * handle's transaction so far and reattach the handle to a new
375 * transaction capabable of guaranteeing the requested number of
376 * credits.
377 */
378
379int journal_restart(handle_t *handle, int nblocks)
380{
381 transaction_t *transaction = handle->h_transaction;
382 journal_t *journal = transaction->t_journal;
383 int ret;
384
385 /* If we've had an abort of any type, don't even think about
386 * actually doing the restart! */
387 if (is_handle_aborted(handle))
388 return 0;
389
390 /*
391 * First unlink the handle from its current transaction, and start the
392 * commit on that.
393 */
394 J_ASSERT(transaction->t_updates > 0);
395 J_ASSERT(journal_current_handle() == handle);
396
397 spin_lock(&journal->j_state_lock);
398 spin_lock(&transaction->t_handle_lock);
399 transaction->t_outstanding_credits -= handle->h_buffer_credits;
400 transaction->t_updates--;
401
402 if (!transaction->t_updates)
403 wake_up(&journal->j_wait_updates);
404 spin_unlock(&transaction->t_handle_lock);
405
406 jbd_debug(2, "restarting handle %p\n", handle);
407 __log_start_commit(journal, transaction->t_tid);
408 spin_unlock(&journal->j_state_lock);
409
410 handle->h_buffer_credits = nblocks;
411 ret = start_this_handle(journal, handle);
412 return ret;
413}
414
415
416/**
417 * void journal_lock_updates () - establish a transaction barrier.
418 * @journal: Journal to establish a barrier on.
419 *
420 * This locks out any further updates from being started, and blocks
421 * until all existing updates have completed, returning only once the
422 * journal is in a quiescent state with no updates running.
423 *
424 * The journal lock should not be held on entry.
425 */
426void journal_lock_updates(journal_t *journal)
427{
428 DEFINE_WAIT(wait);
429
430 spin_lock(&journal->j_state_lock);
431 ++journal->j_barrier_count;
432
433 /* Wait until there are no running updates */
434 while (1) {
435 transaction_t *transaction = journal->j_running_transaction;
436
437 if (!transaction)
438 break;
439
440 spin_lock(&transaction->t_handle_lock);
441 if (!transaction->t_updates) {
442 spin_unlock(&transaction->t_handle_lock);
443 break;
444 }
445 prepare_to_wait(&journal->j_wait_updates, &wait,
446 TASK_UNINTERRUPTIBLE);
447 spin_unlock(&transaction->t_handle_lock);
448 spin_unlock(&journal->j_state_lock);
449 schedule();
450 finish_wait(&journal->j_wait_updates, &wait);
451 spin_lock(&journal->j_state_lock);
452 }
453 spin_unlock(&journal->j_state_lock);
454
455 /*
456 * We have now established a barrier against other normal updates, but
457 * we also need to barrier against other journal_lock_updates() calls
458 * to make sure that we serialise special journal-locked operations
459 * too.
460 */
Arjan van de Ven2c68ee72006-03-23 03:00:35 -0800461 mutex_lock(&journal->j_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462}
463
464/**
465 * void journal_unlock_updates (journal_t* journal) - release barrier
466 * @journal: Journal to release the barrier on.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700467 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 * Release a transaction barrier obtained with journal_lock_updates().
469 *
470 * Should be called without the journal lock held.
471 */
472void journal_unlock_updates (journal_t *journal)
473{
474 J_ASSERT(journal->j_barrier_count != 0);
475
Arjan van de Ven2c68ee72006-03-23 03:00:35 -0800476 mutex_unlock(&journal->j_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 spin_lock(&journal->j_state_lock);
478 --journal->j_barrier_count;
479 spin_unlock(&journal->j_state_lock);
480 wake_up(&journal->j_wait_transaction_locked);
481}
482
483/*
484 * Report any unexpected dirty buffers which turn up. Normally those
485 * indicate an error, but they can occur if the user is running (say)
486 * tune2fs to modify the live filesystem, so we need the option of
487 * continuing as gracefully as possible. #
488 *
489 * The caller should already hold the journal lock and
490 * j_list_lock spinlock: most callers will need those anyway
491 * in order to probe the buffer's journaling state safely.
492 */
493static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
494{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 int jlist;
496
Jan Kara4407c2b2005-09-06 15:19:17 -0700497 /* If this buffer is one which might reasonably be dirty
498 * --- ie. data, or not part of this journal --- then
499 * we're OK to leave it alone, but otherwise we need to
500 * move the dirty bit to the journal's own internal
501 * JBDDirty bit. */
502 jlist = jh->b_jlist;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503
Jan Kara4407c2b2005-09-06 15:19:17 -0700504 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
505 jlist == BJ_Shadow || jlist == BJ_Forget) {
506 struct buffer_head *bh = jh2bh(jh);
507
508 if (test_clear_buffer_dirty(bh))
509 set_buffer_jbddirty(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 }
511}
512
513/*
514 * If the buffer is already part of the current transaction, then there
515 * is nothing we need to do. If it is already part of a prior
516 * transaction which we are still committing to disk, then we need to
517 * make sure that we do not overwrite the old copy: we do copy-out to
518 * preserve the copy going to disk. We also account the buffer against
519 * the handle's metadata buffer credits (unless the buffer is already
520 * part of the transaction, that is).
521 *
522 */
523static int
524do_get_write_access(handle_t *handle, struct journal_head *jh,
525 int force_copy)
526{
527 struct buffer_head *bh;
528 transaction_t *transaction;
529 journal_t *journal;
530 int error;
531 char *frozen_buffer = NULL;
532 int need_copy = 0;
533
534 if (is_handle_aborted(handle))
535 return -EROFS;
536
537 transaction = handle->h_transaction;
538 journal = transaction->t_journal;
539
540 jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
541
542 JBUFFER_TRACE(jh, "entry");
543repeat:
544 bh = jh2bh(jh);
545
546 /* @@@ Need to check for errors here at some point. */
547
548 lock_buffer(bh);
549 jbd_lock_bh_state(bh);
550
551 /* We now hold the buffer lock so it is safe to query the buffer
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700552 * state. Is the buffer dirty?
553 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554 * If so, there are two possibilities. The buffer may be
555 * non-journaled, and undergoing a quite legitimate writeback.
556 * Otherwise, it is journaled, and we don't expect dirty buffers
557 * in that state (the buffers should be marked JBD_Dirty
558 * instead.) So either the IO is being done under our own
559 * control and this is a bug, or it's a third party IO such as
560 * dump(8) (which may leave the buffer scheduled for read ---
561 * ie. locked but not dirty) or tune2fs (which may actually have
562 * the buffer dirtied, ugh.) */
563
564 if (buffer_dirty(bh)) {
565 /*
566 * First question: is this buffer already part of the current
567 * transaction or the existing committing transaction?
568 */
569 if (jh->b_transaction) {
570 J_ASSERT_JH(jh,
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700571 jh->b_transaction == transaction ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 jh->b_transaction ==
573 journal->j_committing_transaction);
574 if (jh->b_next_transaction)
575 J_ASSERT_JH(jh, jh->b_next_transaction ==
576 transaction);
Jan Kara4407c2b2005-09-06 15:19:17 -0700577 }
578 /*
579 * In any case we need to clean the dirty flag and we must
580 * do it under the buffer lock to be sure we don't race
581 * with running write-out.
582 */
583 JBUFFER_TRACE(jh, "Unexpected dirty buffer");
584 jbd_unexpected_dirty_buffer(jh);
Dave Kleikampe9ad5622006-09-27 01:49:35 -0700585 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586
587 unlock_buffer(bh);
588
589 error = -EROFS;
590 if (is_handle_aborted(handle)) {
591 jbd_unlock_bh_state(bh);
592 goto out;
593 }
594 error = 0;
595
596 /*
597 * The buffer is already part of this transaction if b_transaction or
598 * b_next_transaction points to it
599 */
600 if (jh->b_transaction == transaction ||
601 jh->b_next_transaction == transaction)
602 goto done;
603
604 /*
605 * If there is already a copy-out version of this buffer, then we don't
606 * need to make another one
607 */
608 if (jh->b_frozen_data) {
609 JBUFFER_TRACE(jh, "has frozen data");
610 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
611 jh->b_next_transaction = transaction;
612 goto done;
613 }
614
615 /* Is there data here we need to preserve? */
616
617 if (jh->b_transaction && jh->b_transaction != transaction) {
618 JBUFFER_TRACE(jh, "owned by older transaction");
619 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
620 J_ASSERT_JH(jh, jh->b_transaction ==
621 journal->j_committing_transaction);
622
623 /* There is one case we have to be very careful about.
624 * If the committing transaction is currently writing
625 * this buffer out to disk and has NOT made a copy-out,
626 * then we cannot modify the buffer contents at all
627 * right now. The essence of copy-out is that it is the
628 * extra copy, not the primary copy, which gets
629 * journaled. If the primary copy is already going to
630 * disk then we cannot do copy-out here. */
631
632 if (jh->b_jlist == BJ_Shadow) {
633 DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
634 wait_queue_head_t *wqh;
635
636 wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
637
638 JBUFFER_TRACE(jh, "on shadow: sleep");
639 jbd_unlock_bh_state(bh);
640 /* commit wakes up all shadow buffers after IO */
641 for ( ; ; ) {
642 prepare_to_wait(wqh, &wait.wait,
643 TASK_UNINTERRUPTIBLE);
644 if (jh->b_jlist != BJ_Shadow)
645 break;
646 schedule();
647 }
648 finish_wait(wqh, &wait.wait);
649 goto repeat;
650 }
651
652 /* Only do the copy if the currently-owning transaction
653 * still needs it. If it is on the Forget list, the
654 * committing transaction is past that stage. The
655 * buffer had better remain locked during the kmalloc,
656 * but that should be true --- we hold the journal lock
657 * still and the buffer is already on the BUF_JOURNAL
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700658 * list so won't be flushed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 *
660 * Subtle point, though: if this is a get_undo_access,
661 * then we will be relying on the frozen_data to contain
662 * the new value of the committed_data record after the
663 * transaction, so we HAVE to force the frozen_data copy
664 * in that case. */
665
666 if (jh->b_jlist != BJ_Forget || force_copy) {
667 JBUFFER_TRACE(jh, "generate frozen data");
668 if (!frozen_buffer) {
669 JBUFFER_TRACE(jh, "allocate memory for buffer");
670 jbd_unlock_bh_state(bh);
Badari Pulavartyea817392006-08-27 01:23:52 -0700671 frozen_buffer =
672 jbd_slab_alloc(jh2bh(jh)->b_size,
673 GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 if (!frozen_buffer) {
675 printk(KERN_EMERG
676 "%s: OOM for frozen_buffer\n",
677 __FUNCTION__);
678 JBUFFER_TRACE(jh, "oom!");
679 error = -ENOMEM;
680 jbd_lock_bh_state(bh);
681 goto done;
682 }
683 goto repeat;
684 }
685 jh->b_frozen_data = frozen_buffer;
686 frozen_buffer = NULL;
687 need_copy = 1;
688 }
689 jh->b_next_transaction = transaction;
690 }
691
692
693 /*
694 * Finally, if the buffer is not journaled right now, we need to make
695 * sure it doesn't get written to disk before the caller actually
696 * commits the new data
697 */
698 if (!jh->b_transaction) {
699 JBUFFER_TRACE(jh, "no transaction");
700 J_ASSERT_JH(jh, !jh->b_next_transaction);
701 jh->b_transaction = transaction;
702 JBUFFER_TRACE(jh, "file as BJ_Reserved");
703 spin_lock(&journal->j_list_lock);
704 __journal_file_buffer(jh, transaction, BJ_Reserved);
705 spin_unlock(&journal->j_list_lock);
706 }
707
708done:
709 if (need_copy) {
710 struct page *page;
711 int offset;
712 char *source;
713
714 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
715 "Possible IO failure.\n");
716 page = jh2bh(jh)->b_page;
717 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
718 source = kmap_atomic(page, KM_USER0);
719 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
720 kunmap_atomic(source, KM_USER0);
721 }
722 jbd_unlock_bh_state(bh);
723
724 /*
725 * If we are about to journal a buffer, then any revoke pending on it is
726 * no longer valid
727 */
728 journal_cancel_revoke(handle, jh);
729
730out:
Andrew Morton304c4c82006-06-23 02:05:31 -0700731 if (unlikely(frozen_buffer)) /* It's usually NULL */
Badari Pulavarty4cfb04a2006-08-31 21:27:54 -0700732 jbd_slab_free(frozen_buffer, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733
734 JBUFFER_TRACE(jh, "exit");
735 return error;
736}
737
738/**
739 * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
740 * @handle: transaction to add buffer modifications to
741 * @bh: bh to be used for metadata writes
742 * @credits: variable that will receive credits for the buffer
743 *
744 * Returns an error code or 0 on success.
745 *
746 * In full data journalling mode the buffer may be of type BJ_AsyncData,
747 * because we're write()ing a buffer which is also part of a shared mapping.
748 */
749
750int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
751{
752 struct journal_head *jh = journal_add_journal_head(bh);
753 int rc;
754
755 /* We do not want to get caught playing with fields which the
756 * log thread also manipulates. Make sure that the buffer
757 * completes any outstanding IO before proceeding. */
758 rc = do_get_write_access(handle, jh, 0);
759 journal_put_journal_head(jh);
760 return rc;
761}
762
763
764/*
765 * When the user wants to journal a newly created buffer_head
766 * (ie. getblk() returned a new buffer and we are going to populate it
767 * manually rather than reading off disk), then we need to keep the
768 * buffer_head locked until it has been completely filled with new
769 * data. In this case, we should be able to make the assertion that
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700770 * the bh is not already part of an existing transaction.
771 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 * The buffer should already be locked by the caller by this point.
773 * There is no lock ranking violation: it was a newly created,
774 * unlocked buffer beforehand. */
775
776/**
777 * int journal_get_create_access () - notify intent to use newly created bh
778 * @handle: transaction to new buffer to
779 * @bh: new buffer.
780 *
781 * Call this if you create a new bh.
782 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700783int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784{
785 transaction_t *transaction = handle->h_transaction;
786 journal_t *journal = transaction->t_journal;
787 struct journal_head *jh = journal_add_journal_head(bh);
788 int err;
789
790 jbd_debug(5, "journal_head %p\n", jh);
791 err = -EROFS;
792 if (is_handle_aborted(handle))
793 goto out;
794 err = 0;
795
796 JBUFFER_TRACE(jh, "entry");
797 /*
798 * The buffer may already belong to this transaction due to pre-zeroing
799 * in the filesystem's new_block code. It may also be on the previous,
800 * committing transaction's lists, but it HAS to be in Forget state in
801 * that case: the transaction must have deleted the buffer for it to be
802 * reused here.
803 */
804 jbd_lock_bh_state(bh);
805 spin_lock(&journal->j_list_lock);
806 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
807 jh->b_transaction == NULL ||
808 (jh->b_transaction == journal->j_committing_transaction &&
809 jh->b_jlist == BJ_Forget)));
810
811 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
812 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
813
814 if (jh->b_transaction == NULL) {
815 jh->b_transaction = transaction;
816 JBUFFER_TRACE(jh, "file as BJ_Reserved");
817 __journal_file_buffer(jh, transaction, BJ_Reserved);
818 } else if (jh->b_transaction == journal->j_committing_transaction) {
819 JBUFFER_TRACE(jh, "set next transaction");
820 jh->b_next_transaction = transaction;
821 }
822 spin_unlock(&journal->j_list_lock);
823 jbd_unlock_bh_state(bh);
824
825 /*
826 * akpm: I added this. ext3_alloc_branch can pick up new indirect
827 * blocks which contain freed but then revoked metadata. We need
828 * to cancel the revoke in case we end up freeing it yet again
829 * and the reallocating as data - this would cause a second revoke,
830 * which hits an assertion error.
831 */
832 JBUFFER_TRACE(jh, "cancelling revoke");
833 journal_cancel_revoke(handle, jh);
834 journal_put_journal_head(jh);
835out:
836 return err;
837}
838
839/**
840 * int journal_get_undo_access() - Notify intent to modify metadata with
841 * non-rewindable consequences
842 * @handle: transaction
843 * @bh: buffer to undo
844 * @credits: store the number of taken credits here (if not NULL)
845 *
846 * Sometimes there is a need to distinguish between metadata which has
847 * been committed to disk and that which has not. The ext3fs code uses
848 * this for freeing and allocating space, we have to make sure that we
849 * do not reuse freed space until the deallocation has been committed,
850 * since if we overwrote that space we would make the delete
851 * un-rewindable in case of a crash.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700852 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 * To deal with that, journal_get_undo_access requests write access to a
854 * buffer for parts of non-rewindable operations such as delete
855 * operations on the bitmaps. The journaling code must keep a copy of
856 * the buffer's contents prior to the undo_access call until such time
857 * as we know that the buffer has definitely been committed to disk.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700858 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 * We never need to know which transaction the committed data is part
860 * of, buffers touched here are guaranteed to be dirtied later and so
861 * will be committed to a new transaction in due course, at which point
862 * we can discard the old committed data pointer.
863 *
864 * Returns error number or 0 on success.
865 */
866int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
867{
868 int err;
869 struct journal_head *jh = journal_add_journal_head(bh);
870 char *committed_data = NULL;
871
872 JBUFFER_TRACE(jh, "entry");
873
874 /*
875 * Do this first --- it can drop the journal lock, so we want to
876 * make sure that obtaining the committed_data is done
877 * atomically wrt. completion of any outstanding commits.
878 */
879 err = do_get_write_access(handle, jh, 1);
880 if (err)
881 goto out;
882
883repeat:
884 if (!jh->b_committed_data) {
Badari Pulavartyea817392006-08-27 01:23:52 -0700885 committed_data = jbd_slab_alloc(jh2bh(jh)->b_size, GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 if (!committed_data) {
887 printk(KERN_EMERG "%s: No memory for committed data\n",
888 __FUNCTION__);
889 err = -ENOMEM;
890 goto out;
891 }
892 }
893
894 jbd_lock_bh_state(bh);
895 if (!jh->b_committed_data) {
896 /* Copy out the current buffer contents into the
897 * preserved, committed copy. */
898 JBUFFER_TRACE(jh, "generate b_committed data");
899 if (!committed_data) {
900 jbd_unlock_bh_state(bh);
901 goto repeat;
902 }
903
904 jh->b_committed_data = committed_data;
905 committed_data = NULL;
906 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
907 }
908 jbd_unlock_bh_state(bh);
909out:
910 journal_put_journal_head(jh);
Andrew Morton304c4c82006-06-23 02:05:31 -0700911 if (unlikely(committed_data))
Badari Pulavartyea817392006-08-27 01:23:52 -0700912 jbd_slab_free(committed_data, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 return err;
914}
915
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700916/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 * int journal_dirty_data() - mark a buffer as containing dirty data which
918 * needs to be flushed before we can commit the
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700919 * current transaction.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 * @handle: transaction
921 * @bh: bufferhead to mark
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700922 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 * The buffer is placed on the transaction's data list and is marked as
924 * belonging to the transaction.
925 *
926 * Returns error number or 0 on success.
927 *
928 * journal_dirty_data() can be called via page_launder->ext3_writepage
929 * by kswapd.
930 */
931int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
932{
933 journal_t *journal = handle->h_transaction->t_journal;
934 int need_brelse = 0;
935 struct journal_head *jh;
936
937 if (is_handle_aborted(handle))
938 return 0;
939
940 jh = journal_add_journal_head(bh);
941 JBUFFER_TRACE(jh, "entry");
942
943 /*
944 * The buffer could *already* be dirty. Writeout can start
945 * at any time.
946 */
947 jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
948
949 /*
950 * What if the buffer is already part of a running transaction?
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700951 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 * There are two cases:
953 * 1) It is part of the current running transaction. Refile it,
954 * just in case we have allocated it as metadata, deallocated
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700955 * it, then reallocated it as data.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 * 2) It is part of the previous, still-committing transaction.
957 * If all we want to do is to guarantee that the buffer will be
958 * written to disk before this new transaction commits, then
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700959 * being sure that the *previous* transaction has this same
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 * property is sufficient for us! Just leave it on its old
961 * transaction.
962 *
963 * In case (2), the buffer must not already exist as metadata
964 * --- that would violate write ordering (a transaction is free
965 * to write its data at any point, even before the previous
966 * committing transaction has committed). The caller must
967 * never, ever allow this to happen: there's nothing we can do
968 * about it in this layer.
969 */
970 jbd_lock_bh_state(bh);
971 spin_lock(&journal->j_list_lock);
Eric Sandeenf58a74d2006-10-28 10:38:27 -0700972
973 /* Now that we have bh_state locked, are we really still mapped? */
974 if (!buffer_mapped(bh)) {
975 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
976 goto no_journal;
977 }
978
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 if (jh->b_transaction) {
980 JBUFFER_TRACE(jh, "has transaction");
981 if (jh->b_transaction != handle->h_transaction) {
982 JBUFFER_TRACE(jh, "belongs to older transaction");
983 J_ASSERT_JH(jh, jh->b_transaction ==
984 journal->j_committing_transaction);
985
986 /* @@@ IS THIS TRUE ? */
987 /*
988 * Not any more. Scenario: someone does a write()
989 * in data=journal mode. The buffer's transaction has
990 * moved into commit. Then someone does another
991 * write() to the file. We do the frozen data copyout
992 * and set b_next_transaction to point to j_running_t.
993 * And while we're in that state, someone does a
994 * writepage() in an attempt to pageout the same area
995 * of the file via a shared mapping. At present that
996 * calls journal_dirty_data(), and we get right here.
997 * It may be too late to journal the data. Simply
998 * falling through to the next test will suffice: the
999 * data will be dirty and wil be checkpointed. The
1000 * ordering comments in the next comment block still
1001 * apply.
1002 */
1003 //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1004
1005 /*
1006 * If we're journalling data, and this buffer was
1007 * subject to a write(), it could be metadata, forget
1008 * or shadow against the committing transaction. Now,
1009 * someone has dirtied the same darn page via a mapping
1010 * and it is being writepage()'d.
1011 * We *could* just steal the page from commit, with some
1012 * fancy locking there. Instead, we just skip it -
1013 * don't tie the page's buffers to the new transaction
1014 * at all.
1015 * Implication: if we crash before the writepage() data
1016 * is written into the filesystem, recovery will replay
1017 * the write() data.
1018 */
1019 if (jh->b_jlist != BJ_None &&
1020 jh->b_jlist != BJ_SyncData &&
1021 jh->b_jlist != BJ_Locked) {
1022 JBUFFER_TRACE(jh, "Not stealing");
1023 goto no_journal;
1024 }
1025
1026 /*
1027 * This buffer may be undergoing writeout in commit. We
1028 * can't return from here and let the caller dirty it
1029 * again because that can cause the write-out loop in
1030 * commit to never terminate.
1031 */
1032 if (buffer_dirty(bh)) {
1033 get_bh(bh);
1034 spin_unlock(&journal->j_list_lock);
1035 jbd_unlock_bh_state(bh);
1036 need_brelse = 1;
1037 sync_dirty_buffer(bh);
1038 jbd_lock_bh_state(bh);
1039 spin_lock(&journal->j_list_lock);
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001040 /* Since we dropped the lock... */
1041 if (!buffer_mapped(bh)) {
1042 JBUFFER_TRACE(jh, "buffer got unmapped");
1043 goto no_journal;
1044 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 /* The buffer may become locked again at any
1046 time if it is redirtied */
1047 }
1048
1049 /* journal_clean_data_list() may have got there first */
1050 if (jh->b_transaction != NULL) {
1051 JBUFFER_TRACE(jh, "unfile from commit");
1052 __journal_temp_unlink_buffer(jh);
1053 /* It still points to the committing
1054 * transaction; move it to this one so
1055 * that the refile assert checks are
1056 * happy. */
1057 jh->b_transaction = handle->h_transaction;
1058 }
1059 /* The buffer will be refiled below */
1060
1061 }
1062 /*
1063 * Special case --- the buffer might actually have been
1064 * allocated and then immediately deallocated in the previous,
1065 * committing transaction, so might still be left on that
1066 * transaction's metadata lists.
1067 */
1068 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1069 JBUFFER_TRACE(jh, "not on correct data list: unfile");
1070 J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1071 __journal_temp_unlink_buffer(jh);
1072 jh->b_transaction = handle->h_transaction;
1073 JBUFFER_TRACE(jh, "file as data");
1074 __journal_file_buffer(jh, handle->h_transaction,
1075 BJ_SyncData);
1076 }
1077 } else {
1078 JBUFFER_TRACE(jh, "not on a transaction");
1079 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1080 }
1081no_journal:
1082 spin_unlock(&journal->j_list_lock);
1083 jbd_unlock_bh_state(bh);
1084 if (need_brelse) {
1085 BUFFER_TRACE(bh, "brelse");
1086 __brelse(bh);
1087 }
1088 JBUFFER_TRACE(jh, "exit");
1089 journal_put_journal_head(jh);
1090 return 0;
1091}
1092
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001093/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1095 * @handle: transaction to add buffer to.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001096 * @bh: buffer to mark
1097 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 * mark dirty metadata which needs to be journaled as part of the current
1099 * transaction.
1100 *
1101 * The buffer is placed on the transaction's metadata list and is marked
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001102 * as belonging to the transaction.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 *
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001104 * Returns error number or 0 on success.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105 *
1106 * Special care needs to be taken if the buffer already belongs to the
1107 * current committing transaction (in which case we should have frozen
1108 * data present for that commit). In that case, we don't relink the
1109 * buffer: that only gets done when the old transaction finally
1110 * completes its commit.
1111 */
1112int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1113{
1114 transaction_t *transaction = handle->h_transaction;
1115 journal_t *journal = transaction->t_journal;
1116 struct journal_head *jh = bh2jh(bh);
1117
1118 jbd_debug(5, "journal_head %p\n", jh);
1119 JBUFFER_TRACE(jh, "entry");
1120 if (is_handle_aborted(handle))
1121 goto out;
1122
1123 jbd_lock_bh_state(bh);
1124
1125 if (jh->b_modified == 0) {
1126 /*
1127 * This buffer's got modified and becoming part
1128 * of the transaction. This needs to be done
1129 * once a transaction -bzzz
1130 */
1131 jh->b_modified = 1;
1132 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1133 handle->h_buffer_credits--;
1134 }
1135
1136 /*
1137 * fastpath, to avoid expensive locking. If this buffer is already
1138 * on the running transaction's metadata list there is nothing to do.
1139 * Nobody can take it off again because there is a handle open.
1140 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1141 * result in this test being false, so we go in and take the locks.
1142 */
1143 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1144 JBUFFER_TRACE(jh, "fastpath");
1145 J_ASSERT_JH(jh, jh->b_transaction ==
1146 journal->j_running_transaction);
1147 goto out_unlock_bh;
1148 }
1149
1150 set_buffer_jbddirty(bh);
1151
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001152 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153 * Metadata already on the current transaction list doesn't
1154 * need to be filed. Metadata on another transaction's list must
1155 * be committing, and will be refiled once the commit completes:
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001156 * leave it alone for now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 */
1158 if (jh->b_transaction != transaction) {
1159 JBUFFER_TRACE(jh, "already on other transaction");
1160 J_ASSERT_JH(jh, jh->b_transaction ==
1161 journal->j_committing_transaction);
1162 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1163 /* And this case is illegal: we can't reuse another
1164 * transaction's data buffer, ever. */
1165 goto out_unlock_bh;
1166 }
1167
1168 /* That test should have eliminated the following case: */
1169 J_ASSERT_JH(jh, jh->b_frozen_data == 0);
1170
1171 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1172 spin_lock(&journal->j_list_lock);
1173 __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1174 spin_unlock(&journal->j_list_lock);
1175out_unlock_bh:
1176 jbd_unlock_bh_state(bh);
1177out:
1178 JBUFFER_TRACE(jh, "exit");
1179 return 0;
1180}
1181
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001182/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 * journal_release_buffer: undo a get_write_access without any buffer
1184 * updates, if the update decided in the end that it didn't need access.
1185 *
1186 */
1187void
1188journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1189{
1190 BUFFER_TRACE(bh, "entry");
1191}
1192
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001193/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 * void journal_forget() - bforget() for potentially-journaled buffers.
1195 * @handle: transaction handle
1196 * @bh: bh to 'forget'
1197 *
1198 * We can only do the bforget if there are no commits pending against the
1199 * buffer. If the buffer is dirty in the current running transaction we
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001200 * can safely unlink it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201 *
1202 * bh may not be a journalled buffer at all - it may be a non-JBD
1203 * buffer which came off the hashtable. Check for this.
1204 *
1205 * Decrements bh->b_count by one.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001206 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207 * Allow this call even if the handle has aborted --- it may be part of
1208 * the caller's cleanup after an abort.
1209 */
1210int journal_forget (handle_t *handle, struct buffer_head *bh)
1211{
1212 transaction_t *transaction = handle->h_transaction;
1213 journal_t *journal = transaction->t_journal;
1214 struct journal_head *jh;
1215 int drop_reserve = 0;
1216 int err = 0;
1217
1218 BUFFER_TRACE(bh, "entry");
1219
1220 jbd_lock_bh_state(bh);
1221 spin_lock(&journal->j_list_lock);
1222
1223 if (!buffer_jbd(bh))
1224 goto not_jbd;
1225 jh = bh2jh(bh);
1226
1227 /* Critical error: attempting to delete a bitmap buffer, maybe?
1228 * Don't do any jbd operations, and return an error. */
1229 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1230 "inconsistent data on disk")) {
1231 err = -EIO;
1232 goto not_jbd;
1233 }
1234
1235 /*
1236 * The buffer's going from the transaction, we must drop
1237 * all references -bzzz
1238 */
1239 jh->b_modified = 0;
1240
1241 if (jh->b_transaction == handle->h_transaction) {
1242 J_ASSERT_JH(jh, !jh->b_frozen_data);
1243
1244 /* If we are forgetting a buffer which is already part
1245 * of this transaction, then we can just drop it from
1246 * the transaction immediately. */
1247 clear_buffer_dirty(bh);
1248 clear_buffer_jbddirty(bh);
1249
1250 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1251
1252 drop_reserve = 1;
1253
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001254 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 * We are no longer going to journal this buffer.
1256 * However, the commit of this transaction is still
1257 * important to the buffer: the delete that we are now
1258 * processing might obsolete an old log entry, so by
1259 * committing, we can satisfy the buffer's checkpoint.
1260 *
1261 * So, if we have a checkpoint on the buffer, we should
1262 * now refile the buffer on our BJ_Forget list so that
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001263 * we know to remove the checkpoint after we commit.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 */
1265
1266 if (jh->b_cp_transaction) {
1267 __journal_temp_unlink_buffer(jh);
1268 __journal_file_buffer(jh, transaction, BJ_Forget);
1269 } else {
1270 __journal_unfile_buffer(jh);
1271 journal_remove_journal_head(bh);
1272 __brelse(bh);
1273 if (!buffer_jbd(bh)) {
1274 spin_unlock(&journal->j_list_lock);
1275 jbd_unlock_bh_state(bh);
1276 __bforget(bh);
1277 goto drop;
1278 }
1279 }
1280 } else if (jh->b_transaction) {
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001281 J_ASSERT_JH(jh, (jh->b_transaction ==
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 journal->j_committing_transaction));
1283 /* However, if the buffer is still owned by a prior
1284 * (committing) transaction, we can't drop it yet... */
1285 JBUFFER_TRACE(jh, "belongs to older transaction");
1286 /* ... but we CAN drop it from the new transaction if we
1287 * have also modified it since the original commit. */
1288
1289 if (jh->b_next_transaction) {
1290 J_ASSERT(jh->b_next_transaction == transaction);
1291 jh->b_next_transaction = NULL;
1292 drop_reserve = 1;
1293 }
1294 }
1295
1296not_jbd:
1297 spin_unlock(&journal->j_list_lock);
1298 jbd_unlock_bh_state(bh);
1299 __brelse(bh);
1300drop:
1301 if (drop_reserve) {
1302 /* no need to reserve log space for this block -bzzz */
1303 handle->h_buffer_credits++;
1304 }
1305 return err;
1306}
1307
1308/**
1309 * int journal_stop() - complete a transaction
1310 * @handle: tranaction to complete.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001311 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 * All done for a particular handle.
1313 *
1314 * There is not much action needed here. We just return any remaining
1315 * buffer credits to the transaction and remove the handle. The only
1316 * complication is that we need to start a commit operation if the
1317 * filesystem is marked for synchronous update.
1318 *
1319 * journal_stop itself will not usually return an error, but it may
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001320 * do so in unusual circumstances. In particular, expect it to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321 * return -EIO if a journal_abort has been executed since the
1322 * transaction began.
1323 */
1324int journal_stop(handle_t *handle)
1325{
1326 transaction_t *transaction = handle->h_transaction;
1327 journal_t *journal = transaction->t_journal;
1328 int old_handle_count, err;
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001329 pid_t pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 J_ASSERT(journal_current_handle() == handle);
1332
1333 if (is_handle_aborted(handle))
1334 err = -EIO;
OGAWA Hirofumi3e2a5322006-10-19 23:29:11 -07001335 else {
1336 J_ASSERT(transaction->t_updates > 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 err = 0;
OGAWA Hirofumi3e2a5322006-10-19 23:29:11 -07001338 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339
1340 if (--handle->h_ref > 0) {
1341 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1342 handle->h_ref);
1343 return err;
1344 }
1345
1346 jbd_debug(4, "Handle %p going down\n", handle);
1347
1348 /*
1349 * Implement synchronous transaction batching. If the handle
1350 * was synchronous, don't force a commit immediately. Let's
1351 * yield and let another thread piggyback onto this transaction.
1352 * Keep doing that while new threads continue to arrive.
1353 * It doesn't cost much - we're about to run a commit and sleep
1354 * on IO anyway. Speeds up many-threaded, many-dir operations
1355 * by 30x or more...
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001356 *
1357 * But don't do this if this process was the most recent one to
1358 * perform a synchronous write. We do this to detect the case where a
1359 * single process is doing a stream of sync writes. No point in waiting
1360 * for joiners in that case.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 */
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001362 pid = current->pid;
1363 if (handle->h_sync && journal->j_last_sync_writer != pid) {
1364 journal->j_last_sync_writer = pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 do {
1366 old_handle_count = transaction->t_handle_count;
Nishanth Aravamudan041e0e32005-09-10 00:27:23 -07001367 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 } while (old_handle_count != transaction->t_handle_count);
1369 }
1370
1371 current->journal_info = NULL;
1372 spin_lock(&journal->j_state_lock);
1373 spin_lock(&transaction->t_handle_lock);
1374 transaction->t_outstanding_credits -= handle->h_buffer_credits;
1375 transaction->t_updates--;
1376 if (!transaction->t_updates) {
1377 wake_up(&journal->j_wait_updates);
1378 if (journal->j_barrier_count)
1379 wake_up(&journal->j_wait_transaction_locked);
1380 }
1381
1382 /*
1383 * If the handle is marked SYNC, we need to set another commit
1384 * going! We also want to force a commit if the current
1385 * transaction is occupying too much of the log, or if the
1386 * transaction is too old now.
1387 */
1388 if (handle->h_sync ||
1389 transaction->t_outstanding_credits >
1390 journal->j_max_transaction_buffers ||
Dave Kleikampe9ad5622006-09-27 01:49:35 -07001391 time_after_eq(jiffies, transaction->t_expires)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 /* Do this even for aborted journals: an abort still
1393 * completes the commit thread, it just doesn't write
1394 * anything to disk. */
1395 tid_t tid = transaction->t_tid;
1396
1397 spin_unlock(&transaction->t_handle_lock);
1398 jbd_debug(2, "transaction too old, requesting commit for "
1399 "handle %p\n", handle);
1400 /* This is non-blocking */
1401 __log_start_commit(journal, transaction->t_tid);
1402 spin_unlock(&journal->j_state_lock);
1403
1404 /*
1405 * Special case: JFS_SYNC synchronous updates require us
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001406 * to wait for the commit to complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407 */
1408 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1409 err = log_wait_commit(journal, tid);
1410 } else {
1411 spin_unlock(&transaction->t_handle_lock);
1412 spin_unlock(&journal->j_state_lock);
1413 }
1414
1415 jbd_free_handle(handle);
1416 return err;
1417}
1418
1419/**int journal_force_commit() - force any uncommitted transactions
1420 * @journal: journal to force
1421 *
1422 * For synchronous operations: force any uncommitted transactions
1423 * to disk. May seem kludgy, but it reuses all the handle batching
1424 * code in a very simple manner.
1425 */
1426int journal_force_commit(journal_t *journal)
1427{
1428 handle_t *handle;
1429 int ret;
1430
1431 handle = journal_start(journal, 1);
1432 if (IS_ERR(handle)) {
1433 ret = PTR_ERR(handle);
1434 } else {
1435 handle->h_sync = 1;
1436 ret = journal_stop(handle);
1437 }
1438 return ret;
1439}
1440
1441/*
1442 *
1443 * List management code snippets: various functions for manipulating the
1444 * transaction buffer lists.
1445 *
1446 */
1447
1448/*
1449 * Append a buffer to a transaction list, given the transaction's list head
1450 * pointer.
1451 *
1452 * j_list_lock is held.
1453 *
1454 * jbd_lock_bh_state(jh2bh(jh)) is held.
1455 */
1456
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001457static inline void
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458__blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1459{
1460 if (!*list) {
1461 jh->b_tnext = jh->b_tprev = jh;
1462 *list = jh;
1463 } else {
1464 /* Insert at the tail of the list to preserve order */
1465 struct journal_head *first = *list, *last = first->b_tprev;
1466 jh->b_tprev = last;
1467 jh->b_tnext = first;
1468 last->b_tnext = first->b_tprev = jh;
1469 }
1470}
1471
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001472/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 * Remove a buffer from a transaction list, given the transaction's list
1474 * head pointer.
1475 *
1476 * Called with j_list_lock held, and the journal may not be locked.
1477 *
1478 * jbd_lock_bh_state(jh2bh(jh)) is held.
1479 */
1480
1481static inline void
1482__blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1483{
1484 if (*list == jh) {
1485 *list = jh->b_tnext;
1486 if (*list == jh)
1487 *list = NULL;
1488 }
1489 jh->b_tprev->b_tnext = jh->b_tnext;
1490 jh->b_tnext->b_tprev = jh->b_tprev;
1491}
1492
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001493/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 * Remove a buffer from the appropriate transaction list.
1495 *
1496 * Note that this function can *change* the value of
1497 * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1498 * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list. If the caller
1499 * is holding onto a copy of one of thee pointers, it could go bad.
1500 * Generally the caller needs to re-read the pointer from the transaction_t.
1501 *
1502 * Called under j_list_lock. The journal may not be locked.
1503 */
Adrian Bunkd394e122006-12-06 20:38:26 -08001504static void __journal_temp_unlink_buffer(struct journal_head *jh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505{
1506 struct journal_head **list = NULL;
1507 transaction_t *transaction;
1508 struct buffer_head *bh = jh2bh(jh);
1509
1510 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1511 transaction = jh->b_transaction;
1512 if (transaction)
1513 assert_spin_locked(&transaction->t_journal->j_list_lock);
1514
1515 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1516 if (jh->b_jlist != BJ_None)
1517 J_ASSERT_JH(jh, transaction != 0);
1518
1519 switch (jh->b_jlist) {
1520 case BJ_None:
1521 return;
1522 case BJ_SyncData:
1523 list = &transaction->t_sync_datalist;
1524 break;
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 case BJ_Locked:
1546 list = &transaction->t_locked_list;
1547 break;
1548 }
1549
1550 __blist_del_buffer(list, jh);
1551 jh->b_jlist = BJ_None;
1552 if (test_clear_buffer_jbddirty(bh))
1553 mark_buffer_dirty(bh); /* Expose it to the VM */
1554}
1555
1556void __journal_unfile_buffer(struct journal_head *jh)
1557{
1558 __journal_temp_unlink_buffer(jh);
1559 jh->b_transaction = NULL;
1560}
1561
1562void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1563{
1564 jbd_lock_bh_state(jh2bh(jh));
1565 spin_lock(&journal->j_list_lock);
1566 __journal_unfile_buffer(jh);
1567 spin_unlock(&journal->j_list_lock);
1568 jbd_unlock_bh_state(jh2bh(jh));
1569}
1570
1571/*
1572 * Called from journal_try_to_free_buffers().
1573 *
1574 * Called under jbd_lock_bh_state(bh)
1575 */
1576static void
1577__journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1578{
1579 struct journal_head *jh;
1580
1581 jh = bh2jh(bh);
1582
1583 if (buffer_locked(bh) || buffer_dirty(bh))
1584 goto out;
1585
1586 if (jh->b_next_transaction != 0)
1587 goto out;
1588
1589 spin_lock(&journal->j_list_lock);
1590 if (jh->b_transaction != 0 && jh->b_cp_transaction == 0) {
1591 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1592 /* A written-back ordered data buffer */
1593 JBUFFER_TRACE(jh, "release data");
1594 __journal_unfile_buffer(jh);
1595 journal_remove_journal_head(bh);
1596 __brelse(bh);
1597 }
1598 } else if (jh->b_cp_transaction != 0 && jh->b_transaction == 0) {
1599 /* written-back checkpointed metadata buffer */
1600 if (jh->b_jlist == BJ_None) {
1601 JBUFFER_TRACE(jh, "remove from checkpoint list");
1602 __journal_remove_checkpoint(jh);
1603 journal_remove_journal_head(bh);
1604 __brelse(bh);
1605 }
1606 }
1607 spin_unlock(&journal->j_list_lock);
1608out:
1609 return;
1610}
1611
1612
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001613/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 * int journal_try_to_free_buffers() - try to free page buffers.
1615 * @journal: journal for operation
1616 * @page: to try and free
1617 * @unused_gfp_mask: unused
1618 *
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001619 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620 * For all the buffers on this page,
1621 * if they are fully written out ordered data, move them onto BUF_CLEAN
1622 * so try_to_free_buffers() can reap them.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001623 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624 * This function returns non-zero if we wish try_to_free_buffers()
1625 * to be called. We do this if the page is releasable by try_to_free_buffers().
1626 * We also do it if the page has locked or dirty buffers and the caller wants
1627 * us to perform sync or async writeout.
1628 *
1629 * This complicates JBD locking somewhat. We aren't protected by the
1630 * BKL here. We wish to remove the buffer from its committing or
1631 * running transaction's ->t_datalist via __journal_unfile_buffer.
1632 *
1633 * This may *change* the value of transaction_t->t_datalist, so anyone
1634 * who looks at t_datalist needs to lock against this function.
1635 *
1636 * Even worse, someone may be doing a journal_dirty_data on this
1637 * buffer. So we need to lock against that. journal_dirty_data()
1638 * will come out of the lock with the buffer dirty, which makes it
1639 * ineligible for release here.
1640 *
1641 * Who else is affected by this? hmm... Really the only contender
1642 * is do_get_write_access() - it could be looking at the buffer while
1643 * journal_try_to_free_buffer() is changing its state. But that
1644 * cannot happen because we never reallocate freed data as metadata
1645 * while the data is part of a transaction. Yes?
1646 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001647int journal_try_to_free_buffers(journal_t *journal,
Al Viro27496a82005-10-21 03:20:48 -04001648 struct page *page, gfp_t unused_gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649{
1650 struct buffer_head *head;
1651 struct buffer_head *bh;
1652 int ret = 0;
1653
1654 J_ASSERT(PageLocked(page));
1655
1656 head = page_buffers(page);
1657 bh = head;
1658 do {
1659 struct journal_head *jh;
1660
1661 /*
1662 * We take our own ref against the journal_head here to avoid
1663 * having to add tons of locking around each instance of
1664 * journal_remove_journal_head() and journal_put_journal_head().
1665 */
1666 jh = journal_grab_journal_head(bh);
1667 if (!jh)
1668 continue;
1669
1670 jbd_lock_bh_state(bh);
1671 __journal_try_to_free_buffer(journal, bh);
1672 journal_put_journal_head(jh);
1673 jbd_unlock_bh_state(bh);
1674 if (buffer_jbd(bh))
1675 goto busy;
1676 } while ((bh = bh->b_this_page) != head);
1677 ret = try_to_free_buffers(page);
1678busy:
1679 return ret;
1680}
1681
1682/*
1683 * This buffer is no longer needed. If it is on an older transaction's
1684 * checkpoint list we need to record it on this transaction's forget list
1685 * to pin this buffer (and hence its checkpointing transaction) down until
1686 * this transaction commits. If the buffer isn't on a checkpoint list, we
1687 * release it.
1688 * Returns non-zero if JBD no longer has an interest in the buffer.
1689 *
1690 * Called under j_list_lock.
1691 *
1692 * Called under jbd_lock_bh_state(bh).
1693 */
1694static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1695{
1696 int may_free = 1;
1697 struct buffer_head *bh = jh2bh(jh);
1698
1699 __journal_unfile_buffer(jh);
1700
1701 if (jh->b_cp_transaction) {
1702 JBUFFER_TRACE(jh, "on running+cp transaction");
1703 __journal_file_buffer(jh, transaction, BJ_Forget);
1704 clear_buffer_jbddirty(bh);
1705 may_free = 0;
1706 } else {
1707 JBUFFER_TRACE(jh, "on running transaction");
1708 journal_remove_journal_head(bh);
1709 __brelse(bh);
1710 }
1711 return may_free;
1712}
1713
1714/*
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001715 * journal_invalidatepage
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716 *
1717 * This code is tricky. It has a number of cases to deal with.
1718 *
1719 * There are two invariants which this code relies on:
1720 *
1721 * i_size must be updated on disk before we start calling invalidatepage on the
1722 * data.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001723 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 * This is done in ext3 by defining an ext3_setattr method which
1725 * updates i_size before truncate gets going. By maintaining this
1726 * invariant, we can be sure that it is safe to throw away any buffers
1727 * attached to the current transaction: once the transaction commits,
1728 * we know that the data will not be needed.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001729 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730 * Note however that we can *not* throw away data belonging to the
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001731 * previous, committing transaction!
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 *
1733 * Any disk blocks which *are* part of the previous, committing
1734 * transaction (and which therefore cannot be discarded immediately) are
1735 * not going to be reused in the new running transaction
1736 *
1737 * The bitmap committed_data images guarantee this: any block which is
1738 * allocated in one transaction and removed in the next will be marked
1739 * as in-use in the committed_data bitmap, so cannot be reused until
1740 * the next transaction to delete the block commits. This means that
1741 * leaving committing buffers dirty is quite safe: the disk blocks
1742 * cannot be reallocated to a different file and so buffer aliasing is
1743 * not possible.
1744 *
1745 *
1746 * The above applies mainly to ordered data mode. In writeback mode we
1747 * don't make guarantees about the order in which data hits disk --- in
1748 * particular we don't guarantee that new dirty data is flushed before
1749 * transaction commit --- so it is always safe just to discard data
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001750 * immediately in that mode. --sct
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 */
1752
1753/*
1754 * The journal_unmap_buffer helper function returns zero if the buffer
1755 * concerned remains pinned as an anonymous buffer belonging to an older
1756 * transaction.
1757 *
1758 * We're outside-transaction here. Either or both of j_running_transaction
1759 * and j_committing_transaction may be NULL.
1760 */
1761static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1762{
1763 transaction_t *transaction;
1764 struct journal_head *jh;
1765 int may_free = 1;
1766 int ret;
1767
1768 BUFFER_TRACE(bh, "entry");
1769
1770 /*
1771 * It is safe to proceed here without the j_list_lock because the
1772 * buffers cannot be stolen by try_to_free_buffers as long as we are
1773 * holding the page lock. --sct
1774 */
1775
1776 if (!buffer_jbd(bh))
1777 goto zap_buffer_unlocked;
1778
1779 spin_lock(&journal->j_state_lock);
1780 jbd_lock_bh_state(bh);
1781 spin_lock(&journal->j_list_lock);
1782
1783 jh = journal_grab_journal_head(bh);
1784 if (!jh)
1785 goto zap_buffer_no_jh;
1786
1787 transaction = jh->b_transaction;
1788 if (transaction == NULL) {
1789 /* First case: not on any transaction. If it
1790 * has no checkpoint link, then we can zap it:
1791 * it's a writeback-mode buffer so we don't care
1792 * if it hits disk safely. */
1793 if (!jh->b_cp_transaction) {
1794 JBUFFER_TRACE(jh, "not on any transaction: zap");
1795 goto zap_buffer;
1796 }
1797
1798 if (!buffer_dirty(bh)) {
1799 /* bdflush has written it. We can drop it now */
1800 goto zap_buffer;
1801 }
1802
1803 /* OK, it must be in the journal but still not
1804 * written fully to disk: it's metadata or
1805 * journaled data... */
1806
1807 if (journal->j_running_transaction) {
1808 /* ... and once the current transaction has
1809 * committed, the buffer won't be needed any
1810 * longer. */
1811 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1812 ret = __dispose_buffer(jh,
1813 journal->j_running_transaction);
1814 journal_put_journal_head(jh);
1815 spin_unlock(&journal->j_list_lock);
1816 jbd_unlock_bh_state(bh);
1817 spin_unlock(&journal->j_state_lock);
1818 return ret;
1819 } else {
1820 /* There is no currently-running transaction. So the
1821 * orphan record which we wrote for this file must have
1822 * passed into commit. We must attach this buffer to
1823 * the committing transaction, if it exists. */
1824 if (journal->j_committing_transaction) {
1825 JBUFFER_TRACE(jh, "give to committing trans");
1826 ret = __dispose_buffer(jh,
1827 journal->j_committing_transaction);
1828 journal_put_journal_head(jh);
1829 spin_unlock(&journal->j_list_lock);
1830 jbd_unlock_bh_state(bh);
1831 spin_unlock(&journal->j_state_lock);
1832 return ret;
1833 } else {
1834 /* The orphan record's transaction has
1835 * committed. We can cleanse this buffer */
1836 clear_buffer_jbddirty(bh);
1837 goto zap_buffer;
1838 }
1839 }
1840 } else if (transaction == journal->j_committing_transaction) {
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001841 JBUFFER_TRACE(jh, "on committing transaction");
akpm@osdl.orgd13df842005-04-16 15:26:36 -07001842 if (jh->b_jlist == BJ_Locked) {
1843 /*
1844 * The buffer is on the committing transaction's locked
1845 * list. We have the buffer locked, so I/O has
1846 * completed. So we can nail the buffer now.
1847 */
1848 may_free = __dispose_buffer(jh, transaction);
1849 goto zap_buffer;
1850 }
1851 /*
1852 * If it is committing, we simply cannot touch it. We
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 * can remove it's next_transaction pointer from the
1854 * running transaction if that is set, but nothing
1855 * else. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 set_buffer_freed(bh);
1857 if (jh->b_next_transaction) {
1858 J_ASSERT(jh->b_next_transaction ==
1859 journal->j_running_transaction);
1860 jh->b_next_transaction = NULL;
1861 }
1862 journal_put_journal_head(jh);
1863 spin_unlock(&journal->j_list_lock);
1864 jbd_unlock_bh_state(bh);
1865 spin_unlock(&journal->j_state_lock);
1866 return 0;
1867 } else {
1868 /* Good, the buffer belongs to the running transaction.
1869 * We are writing our own transaction's data, not any
1870 * previous one's, so it is safe to throw it away
1871 * (remember that we expect the filesystem to have set
1872 * i_size already for this truncate so recovery will not
1873 * expose the disk blocks we are discarding here.) */
1874 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001875 JBUFFER_TRACE(jh, "on running transaction");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 may_free = __dispose_buffer(jh, transaction);
1877 }
1878
1879zap_buffer:
1880 journal_put_journal_head(jh);
1881zap_buffer_no_jh:
1882 spin_unlock(&journal->j_list_lock);
1883 jbd_unlock_bh_state(bh);
1884 spin_unlock(&journal->j_state_lock);
1885zap_buffer_unlocked:
1886 clear_buffer_dirty(bh);
1887 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1888 clear_buffer_mapped(bh);
1889 clear_buffer_req(bh);
1890 clear_buffer_new(bh);
1891 bh->b_bdev = NULL;
1892 return may_free;
1893}
1894
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001895/**
NeilBrown2ff28e22006-03-26 01:37:18 -08001896 * void journal_invalidatepage()
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001897 * @journal: journal to use for flush...
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 * @page: page to flush
1899 * @offset: length of page to invalidate.
1900 *
1901 * Reap page buffers containing data after offset in page.
1902 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 */
NeilBrown2ff28e22006-03-26 01:37:18 -08001904void journal_invalidatepage(journal_t *journal,
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001905 struct page *page,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906 unsigned long offset)
1907{
1908 struct buffer_head *head, *bh, *next;
1909 unsigned int curr_off = 0;
1910 int may_free = 1;
1911
1912 if (!PageLocked(page))
1913 BUG();
1914 if (!page_has_buffers(page))
NeilBrown2ff28e22006-03-26 01:37:18 -08001915 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916
1917 /* We will potentially be playing with lists other than just the
1918 * data lists (especially for journaled data mode), so be
1919 * cautious in our locking. */
1920
1921 head = bh = page_buffers(page);
1922 do {
1923 unsigned int next_off = curr_off + bh->b_size;
1924 next = bh->b_this_page;
1925
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 if (offset <= curr_off) {
Dave Kleikampe9ad5622006-09-27 01:49:35 -07001927 /* This block is wholly outside the truncation point */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928 lock_buffer(bh);
1929 may_free &= journal_unmap_buffer(journal, bh);
1930 unlock_buffer(bh);
1931 }
1932 curr_off = next_off;
1933 bh = next;
1934
1935 } while (bh != head);
1936
1937 if (!offset) {
NeilBrown2ff28e22006-03-26 01:37:18 -08001938 if (may_free && try_to_free_buffers(page))
1939 J_ASSERT(!page_has_buffers(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941}
1942
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001943/*
1944 * File a buffer on the given transaction list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945 */
1946void __journal_file_buffer(struct journal_head *jh,
1947 transaction_t *transaction, int jlist)
1948{
1949 struct journal_head **list = NULL;
1950 int was_dirty = 0;
1951 struct buffer_head *bh = jh2bh(jh);
1952
1953 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1954 assert_spin_locked(&transaction->t_journal->j_list_lock);
1955
1956 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1957 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1958 jh->b_transaction == 0);
1959
1960 if (jh->b_transaction && jh->b_jlist == jlist)
1961 return;
1962
1963 /* The following list of buffer states needs to be consistent
1964 * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1965 * state. */
1966
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001967 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968 jlist == BJ_Shadow || jlist == BJ_Forget) {
1969 if (test_clear_buffer_dirty(bh) ||
1970 test_clear_buffer_jbddirty(bh))
1971 was_dirty = 1;
1972 }
1973
1974 if (jh->b_transaction)
1975 __journal_temp_unlink_buffer(jh);
1976 jh->b_transaction = transaction;
1977
1978 switch (jlist) {
1979 case BJ_None:
1980 J_ASSERT_JH(jh, !jh->b_committed_data);
1981 J_ASSERT_JH(jh, !jh->b_frozen_data);
1982 return;
1983 case BJ_SyncData:
1984 list = &transaction->t_sync_datalist;
1985 break;
1986 case BJ_Metadata:
1987 transaction->t_nr_buffers++;
1988 list = &transaction->t_buffers;
1989 break;
1990 case BJ_Forget:
1991 list = &transaction->t_forget;
1992 break;
1993 case BJ_IO:
1994 list = &transaction->t_iobuf_list;
1995 break;
1996 case BJ_Shadow:
1997 list = &transaction->t_shadow_list;
1998 break;
1999 case BJ_LogCtl:
2000 list = &transaction->t_log_list;
2001 break;
2002 case BJ_Reserved:
2003 list = &transaction->t_reserved_list;
2004 break;
2005 case BJ_Locked:
2006 list = &transaction->t_locked_list;
2007 break;
2008 }
2009
2010 __blist_add_buffer(list, jh);
2011 jh->b_jlist = jlist;
2012
2013 if (was_dirty)
2014 set_buffer_jbddirty(bh);
2015}
2016
2017void journal_file_buffer(struct journal_head *jh,
2018 transaction_t *transaction, int jlist)
2019{
2020 jbd_lock_bh_state(jh2bh(jh));
2021 spin_lock(&transaction->t_journal->j_list_lock);
2022 __journal_file_buffer(jh, transaction, jlist);
2023 spin_unlock(&transaction->t_journal->j_list_lock);
2024 jbd_unlock_bh_state(jh2bh(jh));
2025}
2026
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002027/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 * Remove a buffer from its current buffer list in preparation for
2029 * dropping it from its current transaction entirely. If the buffer has
2030 * already started to be used by a subsequent transaction, refile the
2031 * buffer on that transaction's metadata list.
2032 *
2033 * Called under journal->j_list_lock
2034 *
2035 * Called under jbd_lock_bh_state(jh2bh(jh))
2036 */
2037void __journal_refile_buffer(struct journal_head *jh)
2038{
2039 int was_dirty;
2040 struct buffer_head *bh = jh2bh(jh);
2041
2042 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2043 if (jh->b_transaction)
2044 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2045
2046 /* If the buffer is now unused, just drop it. */
2047 if (jh->b_next_transaction == NULL) {
2048 __journal_unfile_buffer(jh);
2049 return;
2050 }
2051
2052 /*
2053 * It has been modified by a later transaction: add it to the new
2054 * transaction's metadata list.
2055 */
2056
2057 was_dirty = test_clear_buffer_jbddirty(bh);
2058 __journal_temp_unlink_buffer(jh);
2059 jh->b_transaction = jh->b_next_transaction;
2060 jh->b_next_transaction = NULL;
Jan Kara9ada7342006-06-23 02:05:25 -07002061 __journal_file_buffer(jh, jh->b_transaction,
2062 was_dirty ? BJ_Metadata : BJ_Reserved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2064
2065 if (was_dirty)
2066 set_buffer_jbddirty(bh);
2067}
2068
2069/*
2070 * For the unlocked version of this call, also make sure that any
2071 * hanging journal_head is cleaned up if necessary.
2072 *
2073 * __journal_refile_buffer is usually called as part of a single locked
2074 * operation on a buffer_head, in which the caller is probably going to
2075 * be hooking the journal_head onto other lists. In that case it is up
2076 * to the caller to remove the journal_head if necessary. For the
2077 * unlocked journal_refile_buffer call, the caller isn't going to be
2078 * doing anything else to the buffer so we need to do the cleanup
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002079 * ourselves to avoid a jh leak.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 *
2081 * *** The journal_head may be freed by this call! ***
2082 */
2083void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2084{
2085 struct buffer_head *bh = jh2bh(jh);
2086
2087 jbd_lock_bh_state(bh);
2088 spin_lock(&journal->j_list_lock);
2089
2090 __journal_refile_buffer(jh);
2091 jbd_unlock_bh_state(bh);
2092 journal_remove_journal_head(bh);
2093
2094 spin_unlock(&journal->j_list_lock);
2095 __brelse(bh);
2096}