blob: 60d4c32c880869719c7261a4b0370e5d52d590e9 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Uwe Kleine-König58862692007-05-09 07:51:49 +02002 * linux/fs/jbd/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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/mm.h>
27#include <linux/highmem.h>
28
Adrian Bunkd394e122006-12-06 20:38:26 -080029static void __journal_temp_unlink_buffer(struct journal_head *jh);
30
Linus Torvalds1da177e2005-04-16 15:20:36 -070031/*
32 * get_transaction: obtain a new transaction_t object.
33 *
34 * Simply allocate and initialise a new transaction. Create it in
35 * RUNNING state and add it to the current journal (which should not
36 * have an existing running transaction: we only make a new transaction
37 * once we have started to commit the old one).
38 *
39 * Preconditions:
40 * The journal MUST be locked. We don't perform atomic mallocs on the
41 * new transaction and we can't block without protecting against other
42 * processes trying to touch the journal while it is in transition.
43 *
44 * Called under j_state_lock
45 */
46
47static transaction_t *
48get_transaction(journal_t *journal, transaction_t *transaction)
49{
50 transaction->t_journal = journal;
51 transaction->t_state = T_RUNNING;
52 transaction->t_tid = journal->j_transaction_sequence++;
53 transaction->t_expires = jiffies + journal->j_commit_interval;
54 spin_lock_init(&transaction->t_handle_lock);
55
56 /* Set up the commit timer for the new transaction. */
Arjan van de Ven44d306e2006-12-10 02:21:26 -080057 journal->j_commit_timer.expires = round_jiffies(transaction->t_expires);
Andrew Mortone3df1892006-03-25 03:06:53 -080058 add_timer(&journal->j_commit_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -070059
60 J_ASSERT(journal->j_running_transaction == NULL);
61 journal->j_running_transaction = transaction;
62
63 return transaction;
64}
65
66/*
67 * Handle management.
68 *
69 * A handle_t is an object which represents a single atomic update to a
70 * filesystem, and which tracks all of the modifications which form part
71 * of that one update.
72 */
73
74/*
75 * start_this_handle: Given a handle, deal with any locking or stalling
76 * needed to make sure that there is enough journal space for the handle
77 * to begin. Attach the handle to a transaction and set up the
Mingming Caoae6ddcc2006-09-27 01:49:27 -070078 * transaction's buffer credits.
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 */
80
81static int start_this_handle(journal_t *journal, handle_t *handle)
82{
83 transaction_t *transaction;
84 int needed;
85 int nblocks = handle->h_buffer_credits;
86 transaction_t *new_transaction = NULL;
87 int ret = 0;
88
89 if (nblocks > journal->j_max_transaction_buffers) {
90 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
91 current->comm, nblocks,
92 journal->j_max_transaction_buffers);
93 ret = -ENOSPC;
94 goto out;
95 }
96
97alloc_transaction:
98 if (!journal->j_running_transaction) {
Mingming Cao8c3478a2007-10-18 23:39:20 -070099 new_transaction = kzalloc(sizeof(*new_transaction),
Mingming Caoa5005da2007-10-16 18:38:25 -0400100 GFP_NOFS|__GFP_NOFAIL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101 if (!new_transaction) {
102 ret = -ENOMEM;
103 goto out;
104 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105 }
106
107 jbd_debug(3, "New handle %p going live.\n", handle);
108
109repeat:
110
111 /*
112 * We need to hold j_state_lock until t_updates has been incremented,
113 * for proper journal barrier handling
114 */
115 spin_lock(&journal->j_state_lock);
116repeat_locked:
117 if (is_journal_aborted(journal) ||
118 (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
119 spin_unlock(&journal->j_state_lock);
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700120 ret = -EROFS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121 goto out;
122 }
123
124 /* Wait on the journal's transaction barrier if necessary */
125 if (journal->j_barrier_count) {
126 spin_unlock(&journal->j_state_lock);
127 wait_event(journal->j_wait_transaction_locked,
128 journal->j_barrier_count == 0);
129 goto repeat;
130 }
131
132 if (!journal->j_running_transaction) {
133 if (!new_transaction) {
134 spin_unlock(&journal->j_state_lock);
135 goto alloc_transaction;
136 }
137 get_transaction(journal, new_transaction);
138 new_transaction = NULL;
139 }
140
141 transaction = journal->j_running_transaction;
142
143 /*
144 * If the current transaction is locked down for commit, wait for the
145 * lock to be released.
146 */
147 if (transaction->t_state == T_LOCKED) {
148 DEFINE_WAIT(wait);
149
150 prepare_to_wait(&journal->j_wait_transaction_locked,
151 &wait, TASK_UNINTERRUPTIBLE);
152 spin_unlock(&journal->j_state_lock);
153 schedule();
154 finish_wait(&journal->j_wait_transaction_locked, &wait);
155 goto repeat;
156 }
157
158 /*
159 * If there is not enough space left in the log to write all potential
160 * buffers requested by this operation, we need to stall pending a log
161 * checkpoint to free some more log space.
162 */
163 spin_lock(&transaction->t_handle_lock);
164 needed = transaction->t_outstanding_credits + nblocks;
165
166 if (needed > journal->j_max_transaction_buffers) {
167 /*
168 * If the current transaction is already too large, then start
169 * to commit it: we can then go back and attach this handle to
170 * a new transaction.
171 */
172 DEFINE_WAIT(wait);
173
174 jbd_debug(2, "Handle %p starting new commit...\n", handle);
175 spin_unlock(&transaction->t_handle_lock);
176 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
177 TASK_UNINTERRUPTIBLE);
178 __log_start_commit(journal, transaction->t_tid);
179 spin_unlock(&journal->j_state_lock);
180 schedule();
181 finish_wait(&journal->j_wait_transaction_locked, &wait);
182 goto repeat;
183 }
184
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700185 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186 * The commit code assumes that it can get enough log space
187 * without forcing a checkpoint. This is *critical* for
188 * correctness: a checkpoint of a buffer which is also
189 * associated with a committing transaction creates a deadlock,
190 * so commit simply cannot force through checkpoints.
191 *
192 * We must therefore ensure the necessary space in the journal
193 * *before* starting to dirty potentially checkpointed buffers
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700194 * in the new transaction.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195 *
196 * The worst part is, any transaction currently committing can
197 * reduce the free space arbitrarily. Be careful to account for
198 * those buffers when checkpointing.
199 */
200
201 /*
202 * @@@ AKPM: This seems rather over-defensive. We're giving commit
203 * a _lot_ of headroom: 1/4 of the journal plus the size of
204 * the committing transaction. Really, we only need to give it
205 * committing_transaction->t_outstanding_credits plus "enough" for
206 * the log control blocks.
207 * Also, this test is inconsitent with the matching one in
208 * journal_extend().
209 */
210 if (__log_space_left(journal) < jbd_space_needed(journal)) {
211 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
212 spin_unlock(&transaction->t_handle_lock);
213 __log_wait_for_space(journal);
214 goto repeat_locked;
215 }
216
217 /* OK, account for the buffers that this operation expects to
218 * use and add the handle to the running transaction. */
219
220 handle->h_transaction = transaction;
221 transaction->t_outstanding_credits += nblocks;
222 transaction->t_updates++;
223 transaction->t_handle_count++;
224 jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
225 handle, nblocks, transaction->t_outstanding_credits,
226 __log_space_left(journal));
227 spin_unlock(&transaction->t_handle_lock);
228 spin_unlock(&journal->j_state_lock);
229out:
Andrew Morton304c4c82006-06-23 02:05:31 -0700230 if (unlikely(new_transaction)) /* It's usually NULL */
231 kfree(new_transaction);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 return ret;
233}
234
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +0200235static struct lock_class_key jbd_handle_key;
236
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237/* 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
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +0200247 lockdep_init_map(&handle->h_lockdep_map, "jbd_handle", &jbd_handle_key, 0);
248
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 return handle;
250}
251
252/**
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700253 * handle_t *journal_start() - Obtain a new handle.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 * @journal: Journal to start transaction on.
255 * @nblocks: number of block buffer we might modify
256 *
257 * We make sure that the transaction can guarantee at least nblocks of
258 * modified buffers in the log. We block until the log can guarantee
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700259 * that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 *
261 * This function is visible to journal users (like ext3fs), so is not
262 * called with the journal already locked.
263 *
264 * Return a pointer to a newly allocated handle, or NULL on failure
265 */
266handle_t *journal_start(journal_t *journal, int nblocks)
267{
268 handle_t *handle = journal_current_handle();
269 int err;
270
271 if (!journal)
272 return ERR_PTR(-EROFS);
273
274 if (handle) {
275 J_ASSERT(handle->h_transaction->t_journal == journal);
276 handle->h_ref++;
277 return handle;
278 }
279
280 handle = new_handle(nblocks);
281 if (!handle)
282 return ERR_PTR(-ENOMEM);
283
284 current->journal_info = handle;
285
286 err = start_this_handle(journal, handle);
287 if (err < 0) {
288 jbd_free_handle(handle);
289 current->journal_info = NULL;
290 handle = ERR_PTR(err);
Jonas Bonnf63dcda2008-01-17 15:21:13 -0800291 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 }
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +0200293
Ingo Molnar3295f0e2008-08-11 10:30:30 +0200294 lock_map_acquire(&handle->h_lockdep_map);
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +0200295
Jonas Bonnf63dcda2008-01-17 15:21:13 -0800296out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297 return handle;
298}
299
300/**
301 * int journal_extend() - extend buffer credits.
302 * @handle: handle to 'extend'
303 * @nblocks: nr blocks to try to extend by.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700304 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 * Some transactions, such as large extends and truncates, can be done
306 * atomically all at once or in several stages. The operation requests
307 * a credit for a number of buffer modications in advance, but can
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700308 * extend its credit if it needs more.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309 *
310 * journal_extend tries to give the running handle more buffer credits.
311 * It does not guarantee that allocation - this is a best-effort only.
312 * The calling process MUST be able to deal cleanly with a failure to
313 * extend here.
314 *
315 * Return 0 on success, non-zero on failure.
316 *
317 * return code < 0 implies an error
318 * return code > 0 implies normal transaction-full status.
319 */
320int journal_extend(handle_t *handle, int nblocks)
321{
322 transaction_t *transaction = handle->h_transaction;
323 journal_t *journal = transaction->t_journal;
324 int result;
325 int wanted;
326
327 result = -EIO;
328 if (is_handle_aborted(handle))
329 goto out;
330
331 result = 1;
332
333 spin_lock(&journal->j_state_lock);
334
335 /* Don't extend a locked-down transaction! */
336 if (handle->h_transaction->t_state != T_RUNNING) {
337 jbd_debug(3, "denied handle %p %d blocks: "
338 "transaction not running\n", handle, nblocks);
339 goto error_out;
340 }
341
342 spin_lock(&transaction->t_handle_lock);
343 wanted = transaction->t_outstanding_credits + nblocks;
344
345 if (wanted > journal->j_max_transaction_buffers) {
346 jbd_debug(3, "denied handle %p %d blocks: "
347 "transaction too large\n", handle, nblocks);
348 goto unlock;
349 }
350
351 if (wanted > __log_space_left(journal)) {
352 jbd_debug(3, "denied handle %p %d blocks: "
353 "insufficient log space\n", handle, nblocks);
354 goto unlock;
355 }
356
357 handle->h_buffer_credits += nblocks;
358 transaction->t_outstanding_credits += nblocks;
359 result = 0;
360
361 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
362unlock:
363 spin_unlock(&transaction->t_handle_lock);
364error_out:
365 spin_unlock(&journal->j_state_lock);
366out:
367 return result;
368}
369
370
371/**
Randy Dunlap78a4a502008-02-29 22:02:31 -0800372 * int journal_restart() - restart a handle.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373 * @handle: handle to restart
374 * @nblocks: nr credits requested
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700375 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 * Restart a handle for a multi-transaction filesystem
377 * operation.
378 *
379 * If the journal_extend() call above fails to grant new buffer credits
380 * to a running handle, a call to journal_restart will commit the
381 * handle's transaction so far and reattach the handle to a new
382 * transaction capabable of guaranteeing the requested number of
383 * credits.
384 */
385
386int journal_restart(handle_t *handle, int nblocks)
387{
388 transaction_t *transaction = handle->h_transaction;
389 journal_t *journal = transaction->t_journal;
390 int ret;
391
392 /* If we've had an abort of any type, don't even think about
393 * actually doing the restart! */
394 if (is_handle_aborted(handle))
395 return 0;
396
397 /*
398 * First unlink the handle from its current transaction, and start the
399 * commit on that.
400 */
401 J_ASSERT(transaction->t_updates > 0);
402 J_ASSERT(journal_current_handle() == handle);
403
404 spin_lock(&journal->j_state_lock);
405 spin_lock(&transaction->t_handle_lock);
406 transaction->t_outstanding_credits -= handle->h_buffer_credits;
407 transaction->t_updates--;
408
409 if (!transaction->t_updates)
410 wake_up(&journal->j_wait_updates);
411 spin_unlock(&transaction->t_handle_lock);
412
413 jbd_debug(2, "restarting handle %p\n", handle);
414 __log_start_commit(journal, transaction->t_tid);
415 spin_unlock(&journal->j_state_lock);
416
417 handle->h_buffer_credits = nblocks;
418 ret = start_this_handle(journal, handle);
419 return ret;
420}
421
422
423/**
424 * void journal_lock_updates () - establish a transaction barrier.
425 * @journal: Journal to establish a barrier on.
426 *
427 * This locks out any further updates from being started, and blocks
428 * until all existing updates have completed, returning only once the
429 * journal is in a quiescent state with no updates running.
430 *
431 * The journal lock should not be held on entry.
432 */
433void journal_lock_updates(journal_t *journal)
434{
435 DEFINE_WAIT(wait);
436
437 spin_lock(&journal->j_state_lock);
438 ++journal->j_barrier_count;
439
440 /* Wait until there are no running updates */
441 while (1) {
442 transaction_t *transaction = journal->j_running_transaction;
443
444 if (!transaction)
445 break;
446
447 spin_lock(&transaction->t_handle_lock);
448 if (!transaction->t_updates) {
449 spin_unlock(&transaction->t_handle_lock);
450 break;
451 }
452 prepare_to_wait(&journal->j_wait_updates, &wait,
453 TASK_UNINTERRUPTIBLE);
454 spin_unlock(&transaction->t_handle_lock);
455 spin_unlock(&journal->j_state_lock);
456 schedule();
457 finish_wait(&journal->j_wait_updates, &wait);
458 spin_lock(&journal->j_state_lock);
459 }
460 spin_unlock(&journal->j_state_lock);
461
462 /*
463 * We have now established a barrier against other normal updates, but
464 * we also need to barrier against other journal_lock_updates() calls
465 * to make sure that we serialise special journal-locked operations
466 * too.
467 */
Arjan van de Ven2c68ee72006-03-23 03:00:35 -0800468 mutex_lock(&journal->j_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469}
470
471/**
472 * void journal_unlock_updates (journal_t* journal) - release barrier
473 * @journal: Journal to release the barrier on.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700474 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475 * Release a transaction barrier obtained with journal_lock_updates().
476 *
477 * Should be called without the journal lock held.
478 */
479void journal_unlock_updates (journal_t *journal)
480{
481 J_ASSERT(journal->j_barrier_count != 0);
482
Arjan van de Ven2c68ee72006-03-23 03:00:35 -0800483 mutex_unlock(&journal->j_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 spin_lock(&journal->j_state_lock);
485 --journal->j_barrier_count;
486 spin_unlock(&journal->j_state_lock);
487 wake_up(&journal->j_wait_transaction_locked);
488}
489
490/*
491 * Report any unexpected dirty buffers which turn up. Normally those
492 * indicate an error, but they can occur if the user is running (say)
493 * tune2fs to modify the live filesystem, so we need the option of
494 * continuing as gracefully as possible. #
495 *
496 * The caller should already hold the journal lock and
497 * j_list_lock spinlock: most callers will need those anyway
498 * in order to probe the buffer's journaling state safely.
499 */
500static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
501{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 int jlist;
503
Jan Kara4407c2b2005-09-06 15:19:17 -0700504 /* If this buffer is one which might reasonably be dirty
505 * --- ie. data, or not part of this journal --- then
506 * we're OK to leave it alone, but otherwise we need to
507 * move the dirty bit to the journal's own internal
508 * JBDDirty bit. */
509 jlist = jh->b_jlist;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510
Jan Kara4407c2b2005-09-06 15:19:17 -0700511 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
512 jlist == BJ_Shadow || jlist == BJ_Forget) {
513 struct buffer_head *bh = jh2bh(jh);
514
515 if (test_clear_buffer_dirty(bh))
516 set_buffer_jbddirty(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 }
518}
519
520/*
521 * If the buffer is already part of the current transaction, then there
522 * is nothing we need to do. If it is already part of a prior
523 * transaction which we are still committing to disk, then we need to
524 * make sure that we do not overwrite the old copy: we do copy-out to
525 * preserve the copy going to disk. We also account the buffer against
526 * the handle's metadata buffer credits (unless the buffer is already
527 * part of the transaction, that is).
528 *
529 */
530static int
531do_get_write_access(handle_t *handle, struct journal_head *jh,
532 int force_copy)
533{
534 struct buffer_head *bh;
535 transaction_t *transaction;
536 journal_t *journal;
537 int error;
538 char *frozen_buffer = NULL;
539 int need_copy = 0;
540
541 if (is_handle_aborted(handle))
542 return -EROFS;
543
544 transaction = handle->h_transaction;
545 journal = transaction->t_journal;
546
547 jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
548
549 JBUFFER_TRACE(jh, "entry");
550repeat:
551 bh = jh2bh(jh);
552
553 /* @@@ Need to check for errors here at some point. */
554
555 lock_buffer(bh);
556 jbd_lock_bh_state(bh);
557
558 /* We now hold the buffer lock so it is safe to query the buffer
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700559 * state. Is the buffer dirty?
560 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 * If so, there are two possibilities. The buffer may be
562 * non-journaled, and undergoing a quite legitimate writeback.
563 * Otherwise, it is journaled, and we don't expect dirty buffers
564 * in that state (the buffers should be marked JBD_Dirty
565 * instead.) So either the IO is being done under our own
566 * control and this is a bug, or it's a third party IO such as
567 * dump(8) (which may leave the buffer scheduled for read ---
568 * ie. locked but not dirty) or tune2fs (which may actually have
569 * the buffer dirtied, ugh.) */
570
571 if (buffer_dirty(bh)) {
572 /*
573 * First question: is this buffer already part of the current
574 * transaction or the existing committing transaction?
575 */
576 if (jh->b_transaction) {
577 J_ASSERT_JH(jh,
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700578 jh->b_transaction == transaction ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 jh->b_transaction ==
580 journal->j_committing_transaction);
581 if (jh->b_next_transaction)
582 J_ASSERT_JH(jh, jh->b_next_transaction ==
583 transaction);
Jan Kara4407c2b2005-09-06 15:19:17 -0700584 }
585 /*
586 * In any case we need to clean the dirty flag and we must
587 * do it under the buffer lock to be sure we don't race
588 * with running write-out.
589 */
590 JBUFFER_TRACE(jh, "Unexpected dirty buffer");
591 jbd_unexpected_dirty_buffer(jh);
Dave Kleikampe9ad5622006-09-27 01:49:35 -0700592 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593
594 unlock_buffer(bh);
595
596 error = -EROFS;
597 if (is_handle_aborted(handle)) {
598 jbd_unlock_bh_state(bh);
599 goto out;
600 }
601 error = 0;
602
603 /*
604 * The buffer is already part of this transaction if b_transaction or
605 * b_next_transaction points to it
606 */
607 if (jh->b_transaction == transaction ||
608 jh->b_next_transaction == transaction)
609 goto done;
610
611 /*
Josef Bacik5bc833f2008-04-28 02:16:10 -0700612 * this is the first time this transaction is touching this buffer,
613 * reset the modified flag
614 */
615 jh->b_modified = 0;
616
617 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 * If there is already a copy-out version of this buffer, then we don't
619 * need to make another one
620 */
621 if (jh->b_frozen_data) {
622 JBUFFER_TRACE(jh, "has frozen data");
623 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
624 jh->b_next_transaction = transaction;
625 goto done;
626 }
627
628 /* Is there data here we need to preserve? */
629
630 if (jh->b_transaction && jh->b_transaction != transaction) {
631 JBUFFER_TRACE(jh, "owned by older transaction");
632 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
633 J_ASSERT_JH(jh, jh->b_transaction ==
634 journal->j_committing_transaction);
635
636 /* There is one case we have to be very careful about.
637 * If the committing transaction is currently writing
638 * this buffer out to disk and has NOT made a copy-out,
639 * then we cannot modify the buffer contents at all
640 * right now. The essence of copy-out is that it is the
641 * extra copy, not the primary copy, which gets
642 * journaled. If the primary copy is already going to
643 * disk then we cannot do copy-out here. */
644
645 if (jh->b_jlist == BJ_Shadow) {
646 DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
647 wait_queue_head_t *wqh;
648
649 wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
650
651 JBUFFER_TRACE(jh, "on shadow: sleep");
652 jbd_unlock_bh_state(bh);
653 /* commit wakes up all shadow buffers after IO */
654 for ( ; ; ) {
655 prepare_to_wait(wqh, &wait.wait,
656 TASK_UNINTERRUPTIBLE);
657 if (jh->b_jlist != BJ_Shadow)
658 break;
659 schedule();
660 }
661 finish_wait(wqh, &wait.wait);
662 goto repeat;
663 }
664
665 /* Only do the copy if the currently-owning transaction
666 * still needs it. If it is on the Forget list, the
667 * committing transaction is past that stage. The
668 * buffer had better remain locked during the kmalloc,
669 * but that should be true --- we hold the journal lock
670 * still and the buffer is already on the BUF_JOURNAL
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700671 * list so won't be flushed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 *
673 * Subtle point, though: if this is a get_undo_access,
674 * then we will be relying on the frozen_data to contain
675 * the new value of the committed_data record after the
676 * transaction, so we HAVE to force the frozen_data copy
677 * in that case. */
678
679 if (jh->b_jlist != BJ_Forget || force_copy) {
680 JBUFFER_TRACE(jh, "generate frozen data");
681 if (!frozen_buffer) {
682 JBUFFER_TRACE(jh, "allocate memory for buffer");
683 jbd_unlock_bh_state(bh);
Badari Pulavartyea817392006-08-27 01:23:52 -0700684 frozen_buffer =
Mingming Caoc089d492007-10-16 18:38:25 -0400685 jbd_alloc(jh2bh(jh)->b_size,
Badari Pulavartyea817392006-08-27 01:23:52 -0700686 GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 if (!frozen_buffer) {
688 printk(KERN_EMERG
689 "%s: OOM for frozen_buffer\n",
Harvey Harrison08fc99b2008-04-28 02:16:16 -0700690 __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 JBUFFER_TRACE(jh, "oom!");
692 error = -ENOMEM;
693 jbd_lock_bh_state(bh);
694 goto done;
695 }
696 goto repeat;
697 }
698 jh->b_frozen_data = frozen_buffer;
699 frozen_buffer = NULL;
700 need_copy = 1;
701 }
702 jh->b_next_transaction = transaction;
703 }
704
705
706 /*
707 * Finally, if the buffer is not journaled right now, we need to make
708 * sure it doesn't get written to disk before the caller actually
709 * commits the new data
710 */
711 if (!jh->b_transaction) {
712 JBUFFER_TRACE(jh, "no transaction");
713 J_ASSERT_JH(jh, !jh->b_next_transaction);
714 jh->b_transaction = transaction;
715 JBUFFER_TRACE(jh, "file as BJ_Reserved");
716 spin_lock(&journal->j_list_lock);
717 __journal_file_buffer(jh, transaction, BJ_Reserved);
718 spin_unlock(&journal->j_list_lock);
719 }
720
721done:
722 if (need_copy) {
723 struct page *page;
724 int offset;
725 char *source;
726
727 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
728 "Possible IO failure.\n");
729 page = jh2bh(jh)->b_page;
730 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
731 source = kmap_atomic(page, KM_USER0);
732 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
733 kunmap_atomic(source, KM_USER0);
734 }
735 jbd_unlock_bh_state(bh);
736
737 /*
738 * If we are about to journal a buffer, then any revoke pending on it is
739 * no longer valid
740 */
741 journal_cancel_revoke(handle, jh);
742
743out:
Andrew Morton304c4c82006-06-23 02:05:31 -0700744 if (unlikely(frozen_buffer)) /* It's usually NULL */
Mingming Caoc089d492007-10-16 18:38:25 -0400745 jbd_free(frozen_buffer, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746
747 JBUFFER_TRACE(jh, "exit");
748 return error;
749}
750
751/**
752 * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
753 * @handle: transaction to add buffer modifications to
754 * @bh: bh to be used for metadata writes
755 * @credits: variable that will receive credits for the buffer
756 *
757 * Returns an error code or 0 on success.
758 *
759 * In full data journalling mode the buffer may be of type BJ_AsyncData,
760 * because we're write()ing a buffer which is also part of a shared mapping.
761 */
762
763int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
764{
765 struct journal_head *jh = journal_add_journal_head(bh);
766 int rc;
767
768 /* We do not want to get caught playing with fields which the
769 * log thread also manipulates. Make sure that the buffer
770 * completes any outstanding IO before proceeding. */
771 rc = do_get_write_access(handle, jh, 0);
772 journal_put_journal_head(jh);
773 return rc;
774}
775
776
777/*
778 * When the user wants to journal a newly created buffer_head
779 * (ie. getblk() returned a new buffer and we are going to populate it
780 * manually rather than reading off disk), then we need to keep the
781 * buffer_head locked until it has been completely filled with new
782 * data. In this case, we should be able to make the assertion that
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700783 * the bh is not already part of an existing transaction.
784 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 * The buffer should already be locked by the caller by this point.
786 * There is no lock ranking violation: it was a newly created,
787 * unlocked buffer beforehand. */
788
789/**
790 * int journal_get_create_access () - notify intent to use newly created bh
791 * @handle: transaction to new buffer to
792 * @bh: new buffer.
793 *
794 * Call this if you create a new bh.
795 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700796int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797{
798 transaction_t *transaction = handle->h_transaction;
799 journal_t *journal = transaction->t_journal;
800 struct journal_head *jh = journal_add_journal_head(bh);
801 int err;
802
803 jbd_debug(5, "journal_head %p\n", jh);
804 err = -EROFS;
805 if (is_handle_aborted(handle))
806 goto out;
807 err = 0;
808
809 JBUFFER_TRACE(jh, "entry");
810 /*
811 * The buffer may already belong to this transaction due to pre-zeroing
812 * in the filesystem's new_block code. It may also be on the previous,
813 * committing transaction's lists, but it HAS to be in Forget state in
814 * that case: the transaction must have deleted the buffer for it to be
815 * reused here.
816 */
817 jbd_lock_bh_state(bh);
818 spin_lock(&journal->j_list_lock);
819 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
820 jh->b_transaction == NULL ||
821 (jh->b_transaction == journal->j_committing_transaction &&
822 jh->b_jlist == BJ_Forget)));
823
824 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
825 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
826
827 if (jh->b_transaction == NULL) {
828 jh->b_transaction = transaction;
Josef Bacik5bc833f2008-04-28 02:16:10 -0700829
830 /* first access by this transaction */
831 jh->b_modified = 0;
832
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 JBUFFER_TRACE(jh, "file as BJ_Reserved");
834 __journal_file_buffer(jh, transaction, BJ_Reserved);
835 } else if (jh->b_transaction == journal->j_committing_transaction) {
Josef Bacik5bc833f2008-04-28 02:16:10 -0700836 /* first access by this transaction */
837 jh->b_modified = 0;
838
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 JBUFFER_TRACE(jh, "set next transaction");
840 jh->b_next_transaction = transaction;
841 }
842 spin_unlock(&journal->j_list_lock);
843 jbd_unlock_bh_state(bh);
844
845 /*
846 * akpm: I added this. ext3_alloc_branch can pick up new indirect
847 * blocks which contain freed but then revoked metadata. We need
848 * to cancel the revoke in case we end up freeing it yet again
849 * and the reallocating as data - this would cause a second revoke,
850 * which hits an assertion error.
851 */
852 JBUFFER_TRACE(jh, "cancelling revoke");
853 journal_cancel_revoke(handle, jh);
854 journal_put_journal_head(jh);
855out:
856 return err;
857}
858
859/**
Randy Dunlap78a4a502008-02-29 22:02:31 -0800860 * int journal_get_undo_access() - Notify intent to modify metadata with non-rewindable consequences
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 * @handle: transaction
862 * @bh: buffer to undo
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 *
864 * Sometimes there is a need to distinguish between metadata which has
865 * been committed to disk and that which has not. The ext3fs code uses
866 * this for freeing and allocating space, we have to make sure that we
867 * do not reuse freed space until the deallocation has been committed,
868 * since if we overwrote that space we would make the delete
869 * un-rewindable in case of a crash.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700870 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 * To deal with that, journal_get_undo_access requests write access to a
872 * buffer for parts of non-rewindable operations such as delete
873 * operations on the bitmaps. The journaling code must keep a copy of
874 * the buffer's contents prior to the undo_access call until such time
875 * as we know that the buffer has definitely been committed to disk.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700876 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 * We never need to know which transaction the committed data is part
878 * of, buffers touched here are guaranteed to be dirtied later and so
879 * will be committed to a new transaction in due course, at which point
880 * we can discard the old committed data pointer.
881 *
882 * Returns error number or 0 on success.
883 */
884int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
885{
886 int err;
887 struct journal_head *jh = journal_add_journal_head(bh);
888 char *committed_data = NULL;
889
890 JBUFFER_TRACE(jh, "entry");
891
892 /*
893 * Do this first --- it can drop the journal lock, so we want to
894 * make sure that obtaining the committed_data is done
895 * atomically wrt. completion of any outstanding commits.
896 */
897 err = do_get_write_access(handle, jh, 1);
898 if (err)
899 goto out;
900
901repeat:
902 if (!jh->b_committed_data) {
Mingming Caoc089d492007-10-16 18:38:25 -0400903 committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 if (!committed_data) {
905 printk(KERN_EMERG "%s: No memory for committed data\n",
Harvey Harrison08fc99b2008-04-28 02:16:16 -0700906 __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907 err = -ENOMEM;
908 goto out;
909 }
910 }
911
912 jbd_lock_bh_state(bh);
913 if (!jh->b_committed_data) {
914 /* Copy out the current buffer contents into the
915 * preserved, committed copy. */
916 JBUFFER_TRACE(jh, "generate b_committed data");
917 if (!committed_data) {
918 jbd_unlock_bh_state(bh);
919 goto repeat;
920 }
921
922 jh->b_committed_data = committed_data;
923 committed_data = NULL;
924 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
925 }
926 jbd_unlock_bh_state(bh);
927out:
928 journal_put_journal_head(jh);
Andrew Morton304c4c82006-06-23 02:05:31 -0700929 if (unlikely(committed_data))
Mingming Caoc089d492007-10-16 18:38:25 -0400930 jbd_free(committed_data, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 return err;
932}
933
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700934/**
Randy Dunlap78a4a502008-02-29 22:02:31 -0800935 * int journal_dirty_data() - mark a buffer as containing dirty data to be flushed
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936 * @handle: transaction
937 * @bh: bufferhead to mark
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700938 *
Randy Dunlap78a4a502008-02-29 22:02:31 -0800939 * Description:
940 * Mark a buffer as containing dirty data which needs to be flushed before
941 * we can commit the current transaction.
942 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 * The buffer is placed on the transaction's data list and is marked as
944 * belonging to the transaction.
945 *
946 * Returns error number or 0 on success.
947 *
948 * journal_dirty_data() can be called via page_launder->ext3_writepage
949 * by kswapd.
950 */
951int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
952{
953 journal_t *journal = handle->h_transaction->t_journal;
954 int need_brelse = 0;
955 struct journal_head *jh;
Hidehiro Kawai960a22a2008-10-18 20:27:58 -0700956 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957
958 if (is_handle_aborted(handle))
Hidehiro Kawai960a22a2008-10-18 20:27:58 -0700959 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960
961 jh = journal_add_journal_head(bh);
962 JBUFFER_TRACE(jh, "entry");
963
964 /*
965 * The buffer could *already* be dirty. Writeout can start
966 * at any time.
967 */
968 jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
969
970 /*
971 * What if the buffer is already part of a running transaction?
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700972 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 * There are two cases:
974 * 1) It is part of the current running transaction. Refile it,
975 * just in case we have allocated it as metadata, deallocated
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700976 * it, then reallocated it as data.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 * 2) It is part of the previous, still-committing transaction.
978 * If all we want to do is to guarantee that the buffer will be
979 * written to disk before this new transaction commits, then
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700980 * being sure that the *previous* transaction has this same
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 * property is sufficient for us! Just leave it on its old
982 * transaction.
983 *
984 * In case (2), the buffer must not already exist as metadata
985 * --- that would violate write ordering (a transaction is free
986 * to write its data at any point, even before the previous
987 * committing transaction has committed). The caller must
988 * never, ever allow this to happen: there's nothing we can do
989 * about it in this layer.
990 */
991 jbd_lock_bh_state(bh);
992 spin_lock(&journal->j_list_lock);
Eric Sandeenf58a74d2006-10-28 10:38:27 -0700993
994 /* Now that we have bh_state locked, are we really still mapped? */
995 if (!buffer_mapped(bh)) {
996 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
997 goto no_journal;
998 }
999
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 if (jh->b_transaction) {
1001 JBUFFER_TRACE(jh, "has transaction");
1002 if (jh->b_transaction != handle->h_transaction) {
1003 JBUFFER_TRACE(jh, "belongs to older transaction");
1004 J_ASSERT_JH(jh, jh->b_transaction ==
1005 journal->j_committing_transaction);
1006
1007 /* @@@ IS THIS TRUE ? */
1008 /*
1009 * Not any more. Scenario: someone does a write()
1010 * in data=journal mode. The buffer's transaction has
1011 * moved into commit. Then someone does another
1012 * write() to the file. We do the frozen data copyout
1013 * and set b_next_transaction to point to j_running_t.
1014 * And while we're in that state, someone does a
1015 * writepage() in an attempt to pageout the same area
1016 * of the file via a shared mapping. At present that
1017 * calls journal_dirty_data(), and we get right here.
1018 * It may be too late to journal the data. Simply
1019 * falling through to the next test will suffice: the
1020 * data will be dirty and wil be checkpointed. The
1021 * ordering comments in the next comment block still
1022 * apply.
1023 */
1024 //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1025
1026 /*
1027 * If we're journalling data, and this buffer was
1028 * subject to a write(), it could be metadata, forget
1029 * or shadow against the committing transaction. Now,
1030 * someone has dirtied the same darn page via a mapping
1031 * and it is being writepage()'d.
1032 * We *could* just steal the page from commit, with some
1033 * fancy locking there. Instead, we just skip it -
1034 * don't tie the page's buffers to the new transaction
1035 * at all.
1036 * Implication: if we crash before the writepage() data
1037 * is written into the filesystem, recovery will replay
1038 * the write() data.
1039 */
1040 if (jh->b_jlist != BJ_None &&
1041 jh->b_jlist != BJ_SyncData &&
1042 jh->b_jlist != BJ_Locked) {
1043 JBUFFER_TRACE(jh, "Not stealing");
1044 goto no_journal;
1045 }
1046
1047 /*
1048 * This buffer may be undergoing writeout in commit. We
1049 * can't return from here and let the caller dirty it
1050 * again because that can cause the write-out loop in
1051 * commit to never terminate.
1052 */
1053 if (buffer_dirty(bh)) {
1054 get_bh(bh);
1055 spin_unlock(&journal->j_list_lock);
1056 jbd_unlock_bh_state(bh);
1057 need_brelse = 1;
1058 sync_dirty_buffer(bh);
1059 jbd_lock_bh_state(bh);
1060 spin_lock(&journal->j_list_lock);
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001061 /* Since we dropped the lock... */
1062 if (!buffer_mapped(bh)) {
1063 JBUFFER_TRACE(jh, "buffer got unmapped");
1064 goto no_journal;
1065 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066 /* The buffer may become locked again at any
1067 time if it is redirtied */
1068 }
1069
Hidehiro Kawai960a22a2008-10-18 20:27:58 -07001070 /*
1071 * We cannot remove the buffer with io error from the
1072 * committing transaction, because otherwise it would
1073 * miss the error and the commit would not abort.
1074 */
1075 if (unlikely(!buffer_uptodate(bh))) {
1076 ret = -EIO;
1077 goto no_journal;
1078 }
1079
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 if (jh->b_transaction != NULL) {
1081 JBUFFER_TRACE(jh, "unfile from commit");
1082 __journal_temp_unlink_buffer(jh);
1083 /* It still points to the committing
1084 * transaction; move it to this one so
1085 * that the refile assert checks are
1086 * happy. */
1087 jh->b_transaction = handle->h_transaction;
1088 }
1089 /* The buffer will be refiled below */
1090
1091 }
1092 /*
1093 * Special case --- the buffer might actually have been
1094 * allocated and then immediately deallocated in the previous,
1095 * committing transaction, so might still be left on that
1096 * transaction's metadata lists.
1097 */
1098 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1099 JBUFFER_TRACE(jh, "not on correct data list: unfile");
1100 J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1101 __journal_temp_unlink_buffer(jh);
1102 jh->b_transaction = handle->h_transaction;
1103 JBUFFER_TRACE(jh, "file as data");
1104 __journal_file_buffer(jh, handle->h_transaction,
1105 BJ_SyncData);
1106 }
1107 } else {
1108 JBUFFER_TRACE(jh, "not on a transaction");
1109 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1110 }
1111no_journal:
1112 spin_unlock(&journal->j_list_lock);
1113 jbd_unlock_bh_state(bh);
1114 if (need_brelse) {
1115 BUFFER_TRACE(bh, "brelse");
1116 __brelse(bh);
1117 }
1118 JBUFFER_TRACE(jh, "exit");
1119 journal_put_journal_head(jh);
Hidehiro Kawai960a22a2008-10-18 20:27:58 -07001120 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121}
1122
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001123/**
Randy Dunlap78a4a502008-02-29 22:02:31 -08001124 * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 * @handle: transaction to add buffer to.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001126 * @bh: buffer to mark
1127 *
Randy Dunlap78a4a502008-02-29 22:02:31 -08001128 * Mark dirty metadata which needs to be journaled as part of the current
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 * transaction.
1130 *
1131 * The buffer is placed on the transaction's metadata list and is marked
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001132 * as belonging to the transaction.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 *
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001134 * Returns error number or 0 on success.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 *
1136 * Special care needs to be taken if the buffer already belongs to the
1137 * current committing transaction (in which case we should have frozen
1138 * data present for that commit). In that case, we don't relink the
1139 * buffer: that only gets done when the old transaction finally
1140 * completes its commit.
1141 */
1142int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1143{
1144 transaction_t *transaction = handle->h_transaction;
1145 journal_t *journal = transaction->t_journal;
1146 struct journal_head *jh = bh2jh(bh);
1147
1148 jbd_debug(5, "journal_head %p\n", jh);
1149 JBUFFER_TRACE(jh, "entry");
1150 if (is_handle_aborted(handle))
1151 goto out;
1152
1153 jbd_lock_bh_state(bh);
1154
1155 if (jh->b_modified == 0) {
1156 /*
1157 * This buffer's got modified and becoming part
1158 * of the transaction. This needs to be done
1159 * once a transaction -bzzz
1160 */
1161 jh->b_modified = 1;
1162 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1163 handle->h_buffer_credits--;
1164 }
1165
1166 /*
1167 * fastpath, to avoid expensive locking. If this buffer is already
1168 * on the running transaction's metadata list there is nothing to do.
1169 * Nobody can take it off again because there is a handle open.
1170 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1171 * result in this test being false, so we go in and take the locks.
1172 */
1173 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1174 JBUFFER_TRACE(jh, "fastpath");
1175 J_ASSERT_JH(jh, jh->b_transaction ==
1176 journal->j_running_transaction);
1177 goto out_unlock_bh;
1178 }
1179
1180 set_buffer_jbddirty(bh);
1181
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001182 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 * Metadata already on the current transaction list doesn't
1184 * need to be filed. Metadata on another transaction's list must
1185 * be committing, and will be refiled once the commit completes:
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001186 * leave it alone for now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187 */
1188 if (jh->b_transaction != transaction) {
1189 JBUFFER_TRACE(jh, "already on other transaction");
1190 J_ASSERT_JH(jh, jh->b_transaction ==
1191 journal->j_committing_transaction);
1192 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1193 /* And this case is illegal: we can't reuse another
1194 * transaction's data buffer, ever. */
1195 goto out_unlock_bh;
1196 }
1197
1198 /* That test should have eliminated the following case: */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001199 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200
1201 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1202 spin_lock(&journal->j_list_lock);
1203 __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1204 spin_unlock(&journal->j_list_lock);
1205out_unlock_bh:
1206 jbd_unlock_bh_state(bh);
1207out:
1208 JBUFFER_TRACE(jh, "exit");
1209 return 0;
1210}
1211
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001212/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 * journal_release_buffer: undo a get_write_access without any buffer
1214 * updates, if the update decided in the end that it didn't need access.
1215 *
1216 */
1217void
1218journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1219{
1220 BUFFER_TRACE(bh, "entry");
1221}
1222
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001223/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 * void journal_forget() - bforget() for potentially-journaled buffers.
1225 * @handle: transaction handle
1226 * @bh: bh to 'forget'
1227 *
1228 * We can only do the bforget if there are no commits pending against the
1229 * buffer. If the buffer is dirty in the current running transaction we
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001230 * can safely unlink it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 *
1232 * bh may not be a journalled buffer at all - it may be a non-JBD
1233 * buffer which came off the hashtable. Check for this.
1234 *
1235 * Decrements bh->b_count by one.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001236 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 * Allow this call even if the handle has aborted --- it may be part of
1238 * the caller's cleanup after an abort.
1239 */
1240int journal_forget (handle_t *handle, struct buffer_head *bh)
1241{
1242 transaction_t *transaction = handle->h_transaction;
1243 journal_t *journal = transaction->t_journal;
1244 struct journal_head *jh;
1245 int drop_reserve = 0;
1246 int err = 0;
Josef Bacik5b9a4992008-04-28 02:16:12 -07001247 int was_modified = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248
1249 BUFFER_TRACE(bh, "entry");
1250
1251 jbd_lock_bh_state(bh);
1252 spin_lock(&journal->j_list_lock);
1253
1254 if (!buffer_jbd(bh))
1255 goto not_jbd;
1256 jh = bh2jh(bh);
1257
1258 /* Critical error: attempting to delete a bitmap buffer, maybe?
1259 * Don't do any jbd operations, and return an error. */
1260 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1261 "inconsistent data on disk")) {
1262 err = -EIO;
1263 goto not_jbd;
1264 }
1265
Josef Bacik5b9a4992008-04-28 02:16:12 -07001266 /* keep track of wether or not this transaction modified us */
1267 was_modified = jh->b_modified;
1268
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 /*
1270 * The buffer's going from the transaction, we must drop
1271 * all references -bzzz
1272 */
1273 jh->b_modified = 0;
1274
1275 if (jh->b_transaction == handle->h_transaction) {
1276 J_ASSERT_JH(jh, !jh->b_frozen_data);
1277
1278 /* If we are forgetting a buffer which is already part
1279 * of this transaction, then we can just drop it from
1280 * the transaction immediately. */
1281 clear_buffer_dirty(bh);
1282 clear_buffer_jbddirty(bh);
1283
1284 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1285
Josef Bacik5b9a4992008-04-28 02:16:12 -07001286 /*
1287 * we only want to drop a reference if this transaction
1288 * modified the buffer
1289 */
1290 if (was_modified)
1291 drop_reserve = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001293 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 * We are no longer going to journal this buffer.
1295 * However, the commit of this transaction is still
1296 * important to the buffer: the delete that we are now
1297 * processing might obsolete an old log entry, so by
1298 * committing, we can satisfy the buffer's checkpoint.
1299 *
1300 * So, if we have a checkpoint on the buffer, we should
1301 * now refile the buffer on our BJ_Forget list so that
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001302 * we know to remove the checkpoint after we commit.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303 */
1304
1305 if (jh->b_cp_transaction) {
1306 __journal_temp_unlink_buffer(jh);
1307 __journal_file_buffer(jh, transaction, BJ_Forget);
1308 } else {
1309 __journal_unfile_buffer(jh);
1310 journal_remove_journal_head(bh);
1311 __brelse(bh);
1312 if (!buffer_jbd(bh)) {
1313 spin_unlock(&journal->j_list_lock);
1314 jbd_unlock_bh_state(bh);
1315 __bforget(bh);
1316 goto drop;
1317 }
1318 }
1319 } else if (jh->b_transaction) {
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001320 J_ASSERT_JH(jh, (jh->b_transaction ==
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321 journal->j_committing_transaction));
1322 /* However, if the buffer is still owned by a prior
1323 * (committing) transaction, we can't drop it yet... */
1324 JBUFFER_TRACE(jh, "belongs to older transaction");
1325 /* ... but we CAN drop it from the new transaction if we
1326 * have also modified it since the original commit. */
1327
1328 if (jh->b_next_transaction) {
1329 J_ASSERT(jh->b_next_transaction == transaction);
1330 jh->b_next_transaction = NULL;
Josef Bacik5b9a4992008-04-28 02:16:12 -07001331
1332 /*
1333 * only drop a reference if this transaction modified
1334 * the buffer
1335 */
1336 if (was_modified)
1337 drop_reserve = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 }
1339 }
1340
1341not_jbd:
1342 spin_unlock(&journal->j_list_lock);
1343 jbd_unlock_bh_state(bh);
1344 __brelse(bh);
1345drop:
1346 if (drop_reserve) {
1347 /* no need to reserve log space for this block -bzzz */
1348 handle->h_buffer_credits++;
1349 }
1350 return err;
1351}
1352
1353/**
1354 * int journal_stop() - complete a transaction
1355 * @handle: tranaction to complete.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001356 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 * All done for a particular handle.
1358 *
1359 * There is not much action needed here. We just return any remaining
1360 * buffer credits to the transaction and remove the handle. The only
1361 * complication is that we need to start a commit operation if the
1362 * filesystem is marked for synchronous update.
1363 *
1364 * journal_stop itself will not usually return an error, but it may
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001365 * do so in unusual circumstances. In particular, expect it to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 * return -EIO if a journal_abort has been executed since the
1367 * transaction began.
1368 */
1369int journal_stop(handle_t *handle)
1370{
1371 transaction_t *transaction = handle->h_transaction;
1372 journal_t *journal = transaction->t_journal;
1373 int old_handle_count, err;
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001374 pid_t pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 J_ASSERT(journal_current_handle() == handle);
1377
1378 if (is_handle_aborted(handle))
1379 err = -EIO;
OGAWA Hirofumi3e2a5322006-10-19 23:29:11 -07001380 else {
1381 J_ASSERT(transaction->t_updates > 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 err = 0;
OGAWA Hirofumi3e2a5322006-10-19 23:29:11 -07001383 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384
1385 if (--handle->h_ref > 0) {
1386 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1387 handle->h_ref);
1388 return err;
1389 }
1390
1391 jbd_debug(4, "Handle %p going down\n", handle);
1392
1393 /*
1394 * Implement synchronous transaction batching. If the handle
1395 * was synchronous, don't force a commit immediately. Let's
1396 * yield and let another thread piggyback onto this transaction.
1397 * Keep doing that while new threads continue to arrive.
1398 * It doesn't cost much - we're about to run a commit and sleep
1399 * on IO anyway. Speeds up many-threaded, many-dir operations
1400 * by 30x or more...
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001401 *
1402 * But don't do this if this process was the most recent one to
1403 * perform a synchronous write. We do this to detect the case where a
1404 * single process is doing a stream of sync writes. No point in waiting
1405 * for joiners in that case.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 */
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001407 pid = current->pid;
1408 if (handle->h_sync && journal->j_last_sync_writer != pid) {
1409 journal->j_last_sync_writer = pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 do {
1411 old_handle_count = transaction->t_handle_count;
Nishanth Aravamudan041e0e32005-09-10 00:27:23 -07001412 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 } while (old_handle_count != transaction->t_handle_count);
1414 }
1415
1416 current->journal_info = NULL;
1417 spin_lock(&journal->j_state_lock);
1418 spin_lock(&transaction->t_handle_lock);
1419 transaction->t_outstanding_credits -= handle->h_buffer_credits;
1420 transaction->t_updates--;
1421 if (!transaction->t_updates) {
1422 wake_up(&journal->j_wait_updates);
1423 if (journal->j_barrier_count)
1424 wake_up(&journal->j_wait_transaction_locked);
1425 }
1426
1427 /*
1428 * If the handle is marked SYNC, we need to set another commit
1429 * going! We also want to force a commit if the current
1430 * transaction is occupying too much of the log, or if the
1431 * transaction is too old now.
1432 */
1433 if (handle->h_sync ||
1434 transaction->t_outstanding_credits >
1435 journal->j_max_transaction_buffers ||
Dave Kleikampe9ad5622006-09-27 01:49:35 -07001436 time_after_eq(jiffies, transaction->t_expires)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 /* Do this even for aborted journals: an abort still
1438 * completes the commit thread, it just doesn't write
1439 * anything to disk. */
1440 tid_t tid = transaction->t_tid;
1441
1442 spin_unlock(&transaction->t_handle_lock);
1443 jbd_debug(2, "transaction too old, requesting commit for "
1444 "handle %p\n", handle);
1445 /* This is non-blocking */
1446 __log_start_commit(journal, transaction->t_tid);
1447 spin_unlock(&journal->j_state_lock);
1448
1449 /*
1450 * Special case: JFS_SYNC synchronous updates require us
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001451 * to wait for the commit to complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452 */
1453 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1454 err = log_wait_commit(journal, tid);
1455 } else {
1456 spin_unlock(&transaction->t_handle_lock);
1457 spin_unlock(&journal->j_state_lock);
1458 }
1459
Ingo Molnar3295f0e2008-08-11 10:30:30 +02001460 lock_map_release(&handle->h_lockdep_map);
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +02001461
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462 jbd_free_handle(handle);
1463 return err;
1464}
1465
Randy Dunlap0cf01f62008-03-19 17:00:44 -07001466/**
1467 * int journal_force_commit() - force any uncommitted transactions
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 * @journal: journal to force
1469 *
1470 * For synchronous operations: force any uncommitted transactions
1471 * to disk. May seem kludgy, but it reuses all the handle batching
1472 * code in a very simple manner.
1473 */
1474int journal_force_commit(journal_t *journal)
1475{
1476 handle_t *handle;
1477 int ret;
1478
1479 handle = journal_start(journal, 1);
1480 if (IS_ERR(handle)) {
1481 ret = PTR_ERR(handle);
1482 } else {
1483 handle->h_sync = 1;
1484 ret = journal_stop(handle);
1485 }
1486 return ret;
1487}
1488
1489/*
1490 *
1491 * List management code snippets: various functions for manipulating the
1492 * transaction buffer lists.
1493 *
1494 */
1495
1496/*
1497 * Append a buffer to a transaction list, given the transaction's list head
1498 * pointer.
1499 *
1500 * j_list_lock is held.
1501 *
1502 * jbd_lock_bh_state(jh2bh(jh)) is held.
1503 */
1504
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001505static inline void
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506__blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1507{
1508 if (!*list) {
1509 jh->b_tnext = jh->b_tprev = jh;
1510 *list = jh;
1511 } else {
1512 /* Insert at the tail of the list to preserve order */
1513 struct journal_head *first = *list, *last = first->b_tprev;
1514 jh->b_tprev = last;
1515 jh->b_tnext = first;
1516 last->b_tnext = first->b_tprev = jh;
1517 }
1518}
1519
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001520/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 * Remove a buffer from a transaction list, given the transaction's list
1522 * head pointer.
1523 *
1524 * Called with j_list_lock held, and the journal may not be locked.
1525 *
1526 * jbd_lock_bh_state(jh2bh(jh)) is held.
1527 */
1528
1529static inline void
1530__blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1531{
1532 if (*list == jh) {
1533 *list = jh->b_tnext;
1534 if (*list == jh)
1535 *list = NULL;
1536 }
1537 jh->b_tprev->b_tnext = jh->b_tnext;
1538 jh->b_tnext->b_tprev = jh->b_tprev;
1539}
1540
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001541/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 * Remove a buffer from the appropriate transaction list.
1543 *
1544 * Note that this function can *change* the value of
1545 * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1546 * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list. If the caller
1547 * is holding onto a copy of one of thee pointers, it could go bad.
1548 * Generally the caller needs to re-read the pointer from the transaction_t.
1549 *
1550 * Called under j_list_lock. The journal may not be locked.
1551 */
Adrian Bunkd394e122006-12-06 20:38:26 -08001552static void __journal_temp_unlink_buffer(struct journal_head *jh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553{
1554 struct journal_head **list = NULL;
1555 transaction_t *transaction;
1556 struct buffer_head *bh = jh2bh(jh);
1557
1558 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1559 transaction = jh->b_transaction;
1560 if (transaction)
1561 assert_spin_locked(&transaction->t_journal->j_list_lock);
1562
1563 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1564 if (jh->b_jlist != BJ_None)
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001565 J_ASSERT_JH(jh, transaction != NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566
1567 switch (jh->b_jlist) {
1568 case BJ_None:
1569 return;
1570 case BJ_SyncData:
1571 list = &transaction->t_sync_datalist;
1572 break;
1573 case BJ_Metadata:
1574 transaction->t_nr_buffers--;
1575 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1576 list = &transaction->t_buffers;
1577 break;
1578 case BJ_Forget:
1579 list = &transaction->t_forget;
1580 break;
1581 case BJ_IO:
1582 list = &transaction->t_iobuf_list;
1583 break;
1584 case BJ_Shadow:
1585 list = &transaction->t_shadow_list;
1586 break;
1587 case BJ_LogCtl:
1588 list = &transaction->t_log_list;
1589 break;
1590 case BJ_Reserved:
1591 list = &transaction->t_reserved_list;
1592 break;
1593 case BJ_Locked:
1594 list = &transaction->t_locked_list;
1595 break;
1596 }
1597
1598 __blist_del_buffer(list, jh);
1599 jh->b_jlist = BJ_None;
1600 if (test_clear_buffer_jbddirty(bh))
1601 mark_buffer_dirty(bh); /* Expose it to the VM */
1602}
1603
1604void __journal_unfile_buffer(struct journal_head *jh)
1605{
1606 __journal_temp_unlink_buffer(jh);
1607 jh->b_transaction = NULL;
1608}
1609
1610void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1611{
1612 jbd_lock_bh_state(jh2bh(jh));
1613 spin_lock(&journal->j_list_lock);
1614 __journal_unfile_buffer(jh);
1615 spin_unlock(&journal->j_list_lock);
1616 jbd_unlock_bh_state(jh2bh(jh));
1617}
1618
1619/*
1620 * Called from journal_try_to_free_buffers().
1621 *
1622 * Called under jbd_lock_bh_state(bh)
1623 */
1624static void
1625__journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1626{
1627 struct journal_head *jh;
1628
1629 jh = bh2jh(bh);
1630
1631 if (buffer_locked(bh) || buffer_dirty(bh))
1632 goto out;
1633
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001634 if (jh->b_next_transaction != NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 goto out;
1636
1637 spin_lock(&journal->j_list_lock);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001638 if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1640 /* A written-back ordered data buffer */
1641 JBUFFER_TRACE(jh, "release data");
1642 __journal_unfile_buffer(jh);
1643 journal_remove_journal_head(bh);
1644 __brelse(bh);
1645 }
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001646 } else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647 /* written-back checkpointed metadata buffer */
1648 if (jh->b_jlist == BJ_None) {
1649 JBUFFER_TRACE(jh, "remove from checkpoint list");
1650 __journal_remove_checkpoint(jh);
1651 journal_remove_journal_head(bh);
1652 __brelse(bh);
1653 }
1654 }
1655 spin_unlock(&journal->j_list_lock);
1656out:
1657 return;
1658}
1659
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001660/*
1661 * journal_try_to_free_buffers() could race with journal_commit_transaction()
1662 * The latter might still hold the a count on buffers when inspecting
1663 * them on t_syncdata_list or t_locked_list.
1664 *
1665 * journal_try_to_free_buffers() will call this function to
1666 * wait for the current transaction to finish syncing data buffers, before
1667 * tryinf to free that buffer.
1668 *
1669 * Called with journal->j_state_lock held.
1670 */
1671static void journal_wait_for_transaction_sync_data(journal_t *journal)
1672{
1673 transaction_t *transaction = NULL;
1674 tid_t tid;
1675
1676 spin_lock(&journal->j_state_lock);
1677 transaction = journal->j_committing_transaction;
1678
1679 if (!transaction) {
1680 spin_unlock(&journal->j_state_lock);
1681 return;
1682 }
1683
1684 tid = transaction->t_tid;
1685 spin_unlock(&journal->j_state_lock);
1686 log_wait_commit(journal, tid);
1687}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001689/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690 * int journal_try_to_free_buffers() - try to free page buffers.
1691 * @journal: journal for operation
1692 * @page: to try and free
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001693 * @gfp_mask: we use the mask to detect how hard should we try to release
1694 * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1695 * release the buffers.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 *
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001697 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698 * For all the buffers on this page,
1699 * if they are fully written out ordered data, move them onto BUF_CLEAN
1700 * so try_to_free_buffers() can reap them.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001701 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702 * This function returns non-zero if we wish try_to_free_buffers()
1703 * to be called. We do this if the page is releasable by try_to_free_buffers().
1704 * We also do it if the page has locked or dirty buffers and the caller wants
1705 * us to perform sync or async writeout.
1706 *
1707 * This complicates JBD locking somewhat. We aren't protected by the
1708 * BKL here. We wish to remove the buffer from its committing or
1709 * running transaction's ->t_datalist via __journal_unfile_buffer.
1710 *
1711 * This may *change* the value of transaction_t->t_datalist, so anyone
1712 * who looks at t_datalist needs to lock against this function.
1713 *
1714 * Even worse, someone may be doing a journal_dirty_data on this
1715 * buffer. So we need to lock against that. journal_dirty_data()
1716 * will come out of the lock with the buffer dirty, which makes it
1717 * ineligible for release here.
1718 *
1719 * Who else is affected by this? hmm... Really the only contender
1720 * is do_get_write_access() - it could be looking at the buffer while
1721 * journal_try_to_free_buffer() is changing its state. But that
1722 * cannot happen because we never reallocate freed data as metadata
1723 * while the data is part of a transaction. Yes?
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001724 *
1725 * Return 0 on failure, 1 on success
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001727int journal_try_to_free_buffers(journal_t *journal,
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001728 struct page *page, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729{
1730 struct buffer_head *head;
1731 struct buffer_head *bh;
1732 int ret = 0;
1733
1734 J_ASSERT(PageLocked(page));
1735
1736 head = page_buffers(page);
1737 bh = head;
1738 do {
1739 struct journal_head *jh;
1740
1741 /*
1742 * We take our own ref against the journal_head here to avoid
1743 * having to add tons of locking around each instance of
1744 * journal_remove_journal_head() and journal_put_journal_head().
1745 */
1746 jh = journal_grab_journal_head(bh);
1747 if (!jh)
1748 continue;
1749
1750 jbd_lock_bh_state(bh);
1751 __journal_try_to_free_buffer(journal, bh);
1752 journal_put_journal_head(jh);
1753 jbd_unlock_bh_state(bh);
1754 if (buffer_jbd(bh))
1755 goto busy;
1756 } while ((bh = bh->b_this_page) != head);
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001757
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 ret = try_to_free_buffers(page);
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001759
1760 /*
1761 * There are a number of places where journal_try_to_free_buffers()
1762 * could race with journal_commit_transaction(), the later still
1763 * holds the reference to the buffers to free while processing them.
1764 * try_to_free_buffers() failed to free those buffers. Some of the
1765 * caller of releasepage() request page buffers to be dropped, otherwise
1766 * treat the fail-to-free as errors (such as generic_file_direct_IO())
1767 *
1768 * So, if the caller of try_to_release_page() wants the synchronous
1769 * behaviour(i.e make sure buffers are dropped upon return),
1770 * let's wait for the current transaction to finish flush of
1771 * dirty data buffers, then try to free those buffers again,
1772 * with the journal locked.
1773 */
1774 if (ret == 0 && (gfp_mask & __GFP_WAIT) && (gfp_mask & __GFP_FS)) {
1775 journal_wait_for_transaction_sync_data(journal);
1776 ret = try_to_free_buffers(page);
1777 }
1778
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779busy:
1780 return ret;
1781}
1782
1783/*
1784 * This buffer is no longer needed. If it is on an older transaction's
1785 * checkpoint list we need to record it on this transaction's forget list
1786 * to pin this buffer (and hence its checkpointing transaction) down until
1787 * this transaction commits. If the buffer isn't on a checkpoint list, we
1788 * release it.
1789 * Returns non-zero if JBD no longer has an interest in the buffer.
1790 *
1791 * Called under j_list_lock.
1792 *
1793 * Called under jbd_lock_bh_state(bh).
1794 */
1795static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1796{
1797 int may_free = 1;
1798 struct buffer_head *bh = jh2bh(jh);
1799
1800 __journal_unfile_buffer(jh);
1801
1802 if (jh->b_cp_transaction) {
1803 JBUFFER_TRACE(jh, "on running+cp transaction");
1804 __journal_file_buffer(jh, transaction, BJ_Forget);
1805 clear_buffer_jbddirty(bh);
1806 may_free = 0;
1807 } else {
1808 JBUFFER_TRACE(jh, "on running transaction");
1809 journal_remove_journal_head(bh);
1810 __brelse(bh);
1811 }
1812 return may_free;
1813}
1814
1815/*
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001816 * journal_invalidatepage
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 *
1818 * This code is tricky. It has a number of cases to deal with.
1819 *
1820 * There are two invariants which this code relies on:
1821 *
1822 * i_size must be updated on disk before we start calling invalidatepage on the
1823 * data.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001824 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 * This is done in ext3 by defining an ext3_setattr method which
1826 * updates i_size before truncate gets going. By maintaining this
1827 * invariant, we can be sure that it is safe to throw away any buffers
1828 * attached to the current transaction: once the transaction commits,
1829 * we know that the data will not be needed.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001830 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 * Note however that we can *not* throw away data belonging to the
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001832 * previous, committing transaction!
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833 *
1834 * Any disk blocks which *are* part of the previous, committing
1835 * transaction (and which therefore cannot be discarded immediately) are
1836 * not going to be reused in the new running transaction
1837 *
1838 * The bitmap committed_data images guarantee this: any block which is
1839 * allocated in one transaction and removed in the next will be marked
1840 * as in-use in the committed_data bitmap, so cannot be reused until
1841 * the next transaction to delete the block commits. This means that
1842 * leaving committing buffers dirty is quite safe: the disk blocks
1843 * cannot be reallocated to a different file and so buffer aliasing is
1844 * not possible.
1845 *
1846 *
1847 * The above applies mainly to ordered data mode. In writeback mode we
1848 * don't make guarantees about the order in which data hits disk --- in
1849 * particular we don't guarantee that new dirty data is flushed before
1850 * transaction commit --- so it is always safe just to discard data
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001851 * immediately in that mode. --sct
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 */
1853
1854/*
1855 * The journal_unmap_buffer helper function returns zero if the buffer
1856 * concerned remains pinned as an anonymous buffer belonging to an older
1857 * transaction.
1858 *
1859 * We're outside-transaction here. Either or both of j_running_transaction
1860 * and j_committing_transaction may be NULL.
1861 */
1862static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1863{
1864 transaction_t *transaction;
1865 struct journal_head *jh;
1866 int may_free = 1;
1867 int ret;
1868
1869 BUFFER_TRACE(bh, "entry");
1870
1871 /*
1872 * It is safe to proceed here without the j_list_lock because the
1873 * buffers cannot be stolen by try_to_free_buffers as long as we are
1874 * holding the page lock. --sct
1875 */
1876
1877 if (!buffer_jbd(bh))
1878 goto zap_buffer_unlocked;
1879
1880 spin_lock(&journal->j_state_lock);
1881 jbd_lock_bh_state(bh);
1882 spin_lock(&journal->j_list_lock);
1883
1884 jh = journal_grab_journal_head(bh);
1885 if (!jh)
1886 goto zap_buffer_no_jh;
1887
1888 transaction = jh->b_transaction;
1889 if (transaction == NULL) {
1890 /* First case: not on any transaction. If it
1891 * has no checkpoint link, then we can zap it:
1892 * it's a writeback-mode buffer so we don't care
1893 * if it hits disk safely. */
1894 if (!jh->b_cp_transaction) {
1895 JBUFFER_TRACE(jh, "not on any transaction: zap");
1896 goto zap_buffer;
1897 }
1898
1899 if (!buffer_dirty(bh)) {
1900 /* bdflush has written it. We can drop it now */
1901 goto zap_buffer;
1902 }
1903
1904 /* OK, it must be in the journal but still not
1905 * written fully to disk: it's metadata or
1906 * journaled data... */
1907
1908 if (journal->j_running_transaction) {
1909 /* ... and once the current transaction has
1910 * committed, the buffer won't be needed any
1911 * longer. */
1912 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1913 ret = __dispose_buffer(jh,
1914 journal->j_running_transaction);
1915 journal_put_journal_head(jh);
1916 spin_unlock(&journal->j_list_lock);
1917 jbd_unlock_bh_state(bh);
1918 spin_unlock(&journal->j_state_lock);
1919 return ret;
1920 } else {
1921 /* There is no currently-running transaction. So the
1922 * orphan record which we wrote for this file must have
1923 * passed into commit. We must attach this buffer to
1924 * the committing transaction, if it exists. */
1925 if (journal->j_committing_transaction) {
1926 JBUFFER_TRACE(jh, "give to committing trans");
1927 ret = __dispose_buffer(jh,
1928 journal->j_committing_transaction);
1929 journal_put_journal_head(jh);
1930 spin_unlock(&journal->j_list_lock);
1931 jbd_unlock_bh_state(bh);
1932 spin_unlock(&journal->j_state_lock);
1933 return ret;
1934 } else {
1935 /* The orphan record's transaction has
1936 * committed. We can cleanse this buffer */
1937 clear_buffer_jbddirty(bh);
1938 goto zap_buffer;
1939 }
1940 }
1941 } else if (transaction == journal->j_committing_transaction) {
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001942 JBUFFER_TRACE(jh, "on committing transaction");
akpm@osdl.orgd13df842005-04-16 15:26:36 -07001943 if (jh->b_jlist == BJ_Locked) {
1944 /*
1945 * The buffer is on the committing transaction's locked
1946 * list. We have the buffer locked, so I/O has
1947 * completed. So we can nail the buffer now.
1948 */
1949 may_free = __dispose_buffer(jh, transaction);
1950 goto zap_buffer;
1951 }
1952 /*
1953 * If it is committing, we simply cannot touch it. We
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 * can remove it's next_transaction pointer from the
1955 * running transaction if that is set, but nothing
1956 * else. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 set_buffer_freed(bh);
1958 if (jh->b_next_transaction) {
1959 J_ASSERT(jh->b_next_transaction ==
1960 journal->j_running_transaction);
1961 jh->b_next_transaction = NULL;
1962 }
1963 journal_put_journal_head(jh);
1964 spin_unlock(&journal->j_list_lock);
1965 jbd_unlock_bh_state(bh);
1966 spin_unlock(&journal->j_state_lock);
1967 return 0;
1968 } else {
1969 /* Good, the buffer belongs to the running transaction.
1970 * We are writing our own transaction's data, not any
1971 * previous one's, so it is safe to throw it away
1972 * (remember that we expect the filesystem to have set
1973 * i_size already for this truncate so recovery will not
1974 * expose the disk blocks we are discarding here.) */
1975 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001976 JBUFFER_TRACE(jh, "on running transaction");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 may_free = __dispose_buffer(jh, transaction);
1978 }
1979
1980zap_buffer:
1981 journal_put_journal_head(jh);
1982zap_buffer_no_jh:
1983 spin_unlock(&journal->j_list_lock);
1984 jbd_unlock_bh_state(bh);
1985 spin_unlock(&journal->j_state_lock);
1986zap_buffer_unlocked:
1987 clear_buffer_dirty(bh);
1988 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1989 clear_buffer_mapped(bh);
1990 clear_buffer_req(bh);
1991 clear_buffer_new(bh);
1992 bh->b_bdev = NULL;
1993 return may_free;
1994}
1995
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001996/**
Randy Dunlapa6b91912008-03-19 17:01:00 -07001997 * void journal_invalidatepage() - invalidate a journal page
1998 * @journal: journal to use for flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 * @page: page to flush
2000 * @offset: length of page to invalidate.
2001 *
2002 * Reap page buffers containing data after offset in page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 */
NeilBrown2ff28e22006-03-26 01:37:18 -08002004void journal_invalidatepage(journal_t *journal,
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002005 struct page *page,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 unsigned long offset)
2007{
2008 struct buffer_head *head, *bh, *next;
2009 unsigned int curr_off = 0;
2010 int may_free = 1;
2011
2012 if (!PageLocked(page))
2013 BUG();
2014 if (!page_has_buffers(page))
NeilBrown2ff28e22006-03-26 01:37:18 -08002015 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016
2017 /* We will potentially be playing with lists other than just the
2018 * data lists (especially for journaled data mode), so be
2019 * cautious in our locking. */
2020
2021 head = bh = page_buffers(page);
2022 do {
2023 unsigned int next_off = curr_off + bh->b_size;
2024 next = bh->b_this_page;
2025
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 if (offset <= curr_off) {
Dave Kleikampe9ad5622006-09-27 01:49:35 -07002027 /* This block is wholly outside the truncation point */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 lock_buffer(bh);
2029 may_free &= journal_unmap_buffer(journal, bh);
2030 unlock_buffer(bh);
2031 }
2032 curr_off = next_off;
2033 bh = next;
2034
2035 } while (bh != head);
2036
2037 if (!offset) {
NeilBrown2ff28e22006-03-26 01:37:18 -08002038 if (may_free && try_to_free_buffers(page))
2039 J_ASSERT(!page_has_buffers(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041}
2042
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002043/*
2044 * File a buffer on the given transaction list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 */
2046void __journal_file_buffer(struct journal_head *jh,
2047 transaction_t *transaction, int jlist)
2048{
2049 struct journal_head **list = NULL;
2050 int was_dirty = 0;
2051 struct buffer_head *bh = jh2bh(jh);
2052
2053 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2054 assert_spin_locked(&transaction->t_journal->j_list_lock);
2055
2056 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2057 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07002058 jh->b_transaction == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059
2060 if (jh->b_transaction && jh->b_jlist == jlist)
2061 return;
2062
2063 /* The following list of buffer states needs to be consistent
2064 * with __jbd_unexpected_dirty_buffer()'s handling of dirty
2065 * state. */
2066
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002067 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 jlist == BJ_Shadow || jlist == BJ_Forget) {
2069 if (test_clear_buffer_dirty(bh) ||
2070 test_clear_buffer_jbddirty(bh))
2071 was_dirty = 1;
2072 }
2073
2074 if (jh->b_transaction)
2075 __journal_temp_unlink_buffer(jh);
2076 jh->b_transaction = transaction;
2077
2078 switch (jlist) {
2079 case BJ_None:
2080 J_ASSERT_JH(jh, !jh->b_committed_data);
2081 J_ASSERT_JH(jh, !jh->b_frozen_data);
2082 return;
2083 case BJ_SyncData:
2084 list = &transaction->t_sync_datalist;
2085 break;
2086 case BJ_Metadata:
2087 transaction->t_nr_buffers++;
2088 list = &transaction->t_buffers;
2089 break;
2090 case BJ_Forget:
2091 list = &transaction->t_forget;
2092 break;
2093 case BJ_IO:
2094 list = &transaction->t_iobuf_list;
2095 break;
2096 case BJ_Shadow:
2097 list = &transaction->t_shadow_list;
2098 break;
2099 case BJ_LogCtl:
2100 list = &transaction->t_log_list;
2101 break;
2102 case BJ_Reserved:
2103 list = &transaction->t_reserved_list;
2104 break;
2105 case BJ_Locked:
2106 list = &transaction->t_locked_list;
2107 break;
2108 }
2109
2110 __blist_add_buffer(list, jh);
2111 jh->b_jlist = jlist;
2112
2113 if (was_dirty)
2114 set_buffer_jbddirty(bh);
2115}
2116
2117void journal_file_buffer(struct journal_head *jh,
2118 transaction_t *transaction, int jlist)
2119{
2120 jbd_lock_bh_state(jh2bh(jh));
2121 spin_lock(&transaction->t_journal->j_list_lock);
2122 __journal_file_buffer(jh, transaction, jlist);
2123 spin_unlock(&transaction->t_journal->j_list_lock);
2124 jbd_unlock_bh_state(jh2bh(jh));
2125}
2126
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002127/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 * Remove a buffer from its current buffer list in preparation for
2129 * dropping it from its current transaction entirely. If the buffer has
2130 * already started to be used by a subsequent transaction, refile the
2131 * buffer on that transaction's metadata list.
2132 *
2133 * Called under journal->j_list_lock
2134 *
2135 * Called under jbd_lock_bh_state(jh2bh(jh))
2136 */
2137void __journal_refile_buffer(struct journal_head *jh)
2138{
2139 int was_dirty;
2140 struct buffer_head *bh = jh2bh(jh);
2141
2142 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2143 if (jh->b_transaction)
2144 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2145
2146 /* If the buffer is now unused, just drop it. */
2147 if (jh->b_next_transaction == NULL) {
2148 __journal_unfile_buffer(jh);
2149 return;
2150 }
2151
2152 /*
2153 * It has been modified by a later transaction: add it to the new
2154 * transaction's metadata list.
2155 */
2156
2157 was_dirty = test_clear_buffer_jbddirty(bh);
2158 __journal_temp_unlink_buffer(jh);
2159 jh->b_transaction = jh->b_next_transaction;
2160 jh->b_next_transaction = NULL;
Jan Kara9ada7342006-06-23 02:05:25 -07002161 __journal_file_buffer(jh, jh->b_transaction,
Josef Bacik5b9a4992008-04-28 02:16:12 -07002162 jh->b_modified ? BJ_Metadata : BJ_Reserved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2164
2165 if (was_dirty)
2166 set_buffer_jbddirty(bh);
2167}
2168
2169/*
2170 * For the unlocked version of this call, also make sure that any
2171 * hanging journal_head is cleaned up if necessary.
2172 *
2173 * __journal_refile_buffer is usually called as part of a single locked
2174 * operation on a buffer_head, in which the caller is probably going to
2175 * be hooking the journal_head onto other lists. In that case it is up
2176 * to the caller to remove the journal_head if necessary. For the
2177 * unlocked journal_refile_buffer call, the caller isn't going to be
2178 * doing anything else to the buffer so we need to do the cleanup
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002179 * ourselves to avoid a jh leak.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 *
2181 * *** The journal_head may be freed by this call! ***
2182 */
2183void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2184{
2185 struct buffer_head *bh = jh2bh(jh);
2186
2187 jbd_lock_bh_state(bh);
2188 spin_lock(&journal->j_list_lock);
2189
2190 __journal_refile_buffer(jh);
2191 jbd_unlock_bh_state(bh);
2192 journal_remove_journal_head(bh);
2193
2194 spin_unlock(&journal->j_list_lock);
2195 __brelse(bh);
2196}