blob: 038ed743619911799dfff8e27623e75f593fc60c [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
294 lock_acquire(&handle->h_lockdep_map, 0, 0, 0, 2, _THIS_IP_);
295
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/**
372 * int journal_restart() - restart a handle .
373 * @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 /*
612 * If there is already a copy-out version of this buffer, then we don't
613 * need to make another one
614 */
615 if (jh->b_frozen_data) {
616 JBUFFER_TRACE(jh, "has frozen data");
617 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
618 jh->b_next_transaction = transaction;
619 goto done;
620 }
621
622 /* Is there data here we need to preserve? */
623
624 if (jh->b_transaction && jh->b_transaction != transaction) {
625 JBUFFER_TRACE(jh, "owned by older transaction");
626 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
627 J_ASSERT_JH(jh, jh->b_transaction ==
628 journal->j_committing_transaction);
629
630 /* There is one case we have to be very careful about.
631 * If the committing transaction is currently writing
632 * this buffer out to disk and has NOT made a copy-out,
633 * then we cannot modify the buffer contents at all
634 * right now. The essence of copy-out is that it is the
635 * extra copy, not the primary copy, which gets
636 * journaled. If the primary copy is already going to
637 * disk then we cannot do copy-out here. */
638
639 if (jh->b_jlist == BJ_Shadow) {
640 DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
641 wait_queue_head_t *wqh;
642
643 wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
644
645 JBUFFER_TRACE(jh, "on shadow: sleep");
646 jbd_unlock_bh_state(bh);
647 /* commit wakes up all shadow buffers after IO */
648 for ( ; ; ) {
649 prepare_to_wait(wqh, &wait.wait,
650 TASK_UNINTERRUPTIBLE);
651 if (jh->b_jlist != BJ_Shadow)
652 break;
653 schedule();
654 }
655 finish_wait(wqh, &wait.wait);
656 goto repeat;
657 }
658
659 /* Only do the copy if the currently-owning transaction
660 * still needs it. If it is on the Forget list, the
661 * committing transaction is past that stage. The
662 * buffer had better remain locked during the kmalloc,
663 * but that should be true --- we hold the journal lock
664 * still and the buffer is already on the BUF_JOURNAL
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700665 * list so won't be flushed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 *
667 * Subtle point, though: if this is a get_undo_access,
668 * then we will be relying on the frozen_data to contain
669 * the new value of the committed_data record after the
670 * transaction, so we HAVE to force the frozen_data copy
671 * in that case. */
672
673 if (jh->b_jlist != BJ_Forget || force_copy) {
674 JBUFFER_TRACE(jh, "generate frozen data");
675 if (!frozen_buffer) {
676 JBUFFER_TRACE(jh, "allocate memory for buffer");
677 jbd_unlock_bh_state(bh);
Badari Pulavartyea817392006-08-27 01:23:52 -0700678 frozen_buffer =
Mingming Caoc089d492007-10-16 18:38:25 -0400679 jbd_alloc(jh2bh(jh)->b_size,
Badari Pulavartyea817392006-08-27 01:23:52 -0700680 GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 if (!frozen_buffer) {
682 printk(KERN_EMERG
683 "%s: OOM for frozen_buffer\n",
684 __FUNCTION__);
685 JBUFFER_TRACE(jh, "oom!");
686 error = -ENOMEM;
687 jbd_lock_bh_state(bh);
688 goto done;
689 }
690 goto repeat;
691 }
692 jh->b_frozen_data = frozen_buffer;
693 frozen_buffer = NULL;
694 need_copy = 1;
695 }
696 jh->b_next_transaction = transaction;
697 }
698
699
700 /*
701 * Finally, if the buffer is not journaled right now, we need to make
702 * sure it doesn't get written to disk before the caller actually
703 * commits the new data
704 */
705 if (!jh->b_transaction) {
706 JBUFFER_TRACE(jh, "no transaction");
707 J_ASSERT_JH(jh, !jh->b_next_transaction);
708 jh->b_transaction = transaction;
709 JBUFFER_TRACE(jh, "file as BJ_Reserved");
710 spin_lock(&journal->j_list_lock);
711 __journal_file_buffer(jh, transaction, BJ_Reserved);
712 spin_unlock(&journal->j_list_lock);
713 }
714
715done:
716 if (need_copy) {
717 struct page *page;
718 int offset;
719 char *source;
720
721 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
722 "Possible IO failure.\n");
723 page = jh2bh(jh)->b_page;
724 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
725 source = kmap_atomic(page, KM_USER0);
726 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
727 kunmap_atomic(source, KM_USER0);
728 }
729 jbd_unlock_bh_state(bh);
730
731 /*
732 * If we are about to journal a buffer, then any revoke pending on it is
733 * no longer valid
734 */
735 journal_cancel_revoke(handle, jh);
736
737out:
Andrew Morton304c4c82006-06-23 02:05:31 -0700738 if (unlikely(frozen_buffer)) /* It's usually NULL */
Mingming Caoc089d492007-10-16 18:38:25 -0400739 jbd_free(frozen_buffer, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740
741 JBUFFER_TRACE(jh, "exit");
742 return error;
743}
744
745/**
746 * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
747 * @handle: transaction to add buffer modifications to
748 * @bh: bh to be used for metadata writes
749 * @credits: variable that will receive credits for the buffer
750 *
751 * Returns an error code or 0 on success.
752 *
753 * In full data journalling mode the buffer may be of type BJ_AsyncData,
754 * because we're write()ing a buffer which is also part of a shared mapping.
755 */
756
757int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
758{
759 struct journal_head *jh = journal_add_journal_head(bh);
760 int rc;
761
762 /* We do not want to get caught playing with fields which the
763 * log thread also manipulates. Make sure that the buffer
764 * completes any outstanding IO before proceeding. */
765 rc = do_get_write_access(handle, jh, 0);
766 journal_put_journal_head(jh);
767 return rc;
768}
769
770
771/*
772 * When the user wants to journal a newly created buffer_head
773 * (ie. getblk() returned a new buffer and we are going to populate it
774 * manually rather than reading off disk), then we need to keep the
775 * buffer_head locked until it has been completely filled with new
776 * data. In this case, we should be able to make the assertion that
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700777 * the bh is not already part of an existing transaction.
778 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 * The buffer should already be locked by the caller by this point.
780 * There is no lock ranking violation: it was a newly created,
781 * unlocked buffer beforehand. */
782
783/**
784 * int journal_get_create_access () - notify intent to use newly created bh
785 * @handle: transaction to new buffer to
786 * @bh: new buffer.
787 *
788 * Call this if you create a new bh.
789 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700790int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791{
792 transaction_t *transaction = handle->h_transaction;
793 journal_t *journal = transaction->t_journal;
794 struct journal_head *jh = journal_add_journal_head(bh);
795 int err;
796
797 jbd_debug(5, "journal_head %p\n", jh);
798 err = -EROFS;
799 if (is_handle_aborted(handle))
800 goto out;
801 err = 0;
802
803 JBUFFER_TRACE(jh, "entry");
804 /*
805 * The buffer may already belong to this transaction due to pre-zeroing
806 * in the filesystem's new_block code. It may also be on the previous,
807 * committing transaction's lists, but it HAS to be in Forget state in
808 * that case: the transaction must have deleted the buffer for it to be
809 * reused here.
810 */
811 jbd_lock_bh_state(bh);
812 spin_lock(&journal->j_list_lock);
813 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
814 jh->b_transaction == NULL ||
815 (jh->b_transaction == journal->j_committing_transaction &&
816 jh->b_jlist == BJ_Forget)));
817
818 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
819 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
820
821 if (jh->b_transaction == NULL) {
822 jh->b_transaction = transaction;
823 JBUFFER_TRACE(jh, "file as BJ_Reserved");
824 __journal_file_buffer(jh, transaction, BJ_Reserved);
825 } else if (jh->b_transaction == journal->j_committing_transaction) {
826 JBUFFER_TRACE(jh, "set next transaction");
827 jh->b_next_transaction = transaction;
828 }
829 spin_unlock(&journal->j_list_lock);
830 jbd_unlock_bh_state(bh);
831
832 /*
833 * akpm: I added this. ext3_alloc_branch can pick up new indirect
834 * blocks which contain freed but then revoked metadata. We need
835 * to cancel the revoke in case we end up freeing it yet again
836 * and the reallocating as data - this would cause a second revoke,
837 * which hits an assertion error.
838 */
839 JBUFFER_TRACE(jh, "cancelling revoke");
840 journal_cancel_revoke(handle, jh);
841 journal_put_journal_head(jh);
842out:
843 return err;
844}
845
846/**
847 * int journal_get_undo_access() - Notify intent to modify metadata with
848 * non-rewindable consequences
849 * @handle: transaction
850 * @bh: buffer to undo
851 * @credits: store the number of taken credits here (if not NULL)
852 *
853 * Sometimes there is a need to distinguish between metadata which has
854 * been committed to disk and that which has not. The ext3fs code uses
855 * this for freeing and allocating space, we have to make sure that we
856 * do not reuse freed space until the deallocation has been committed,
857 * since if we overwrote that space we would make the delete
858 * un-rewindable in case of a crash.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700859 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 * To deal with that, journal_get_undo_access requests write access to a
861 * buffer for parts of non-rewindable operations such as delete
862 * operations on the bitmaps. The journaling code must keep a copy of
863 * the buffer's contents prior to the undo_access call until such time
864 * as we know that the buffer has definitely been committed to disk.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700865 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 * We never need to know which transaction the committed data is part
867 * of, buffers touched here are guaranteed to be dirtied later and so
868 * will be committed to a new transaction in due course, at which point
869 * we can discard the old committed data pointer.
870 *
871 * Returns error number or 0 on success.
872 */
873int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
874{
875 int err;
876 struct journal_head *jh = journal_add_journal_head(bh);
877 char *committed_data = NULL;
878
879 JBUFFER_TRACE(jh, "entry");
880
881 /*
882 * Do this first --- it can drop the journal lock, so we want to
883 * make sure that obtaining the committed_data is done
884 * atomically wrt. completion of any outstanding commits.
885 */
886 err = do_get_write_access(handle, jh, 1);
887 if (err)
888 goto out;
889
890repeat:
891 if (!jh->b_committed_data) {
Mingming Caoc089d492007-10-16 18:38:25 -0400892 committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 if (!committed_data) {
894 printk(KERN_EMERG "%s: No memory for committed data\n",
895 __FUNCTION__);
896 err = -ENOMEM;
897 goto out;
898 }
899 }
900
901 jbd_lock_bh_state(bh);
902 if (!jh->b_committed_data) {
903 /* Copy out the current buffer contents into the
904 * preserved, committed copy. */
905 JBUFFER_TRACE(jh, "generate b_committed data");
906 if (!committed_data) {
907 jbd_unlock_bh_state(bh);
908 goto repeat;
909 }
910
911 jh->b_committed_data = committed_data;
912 committed_data = NULL;
913 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
914 }
915 jbd_unlock_bh_state(bh);
916out:
917 journal_put_journal_head(jh);
Andrew Morton304c4c82006-06-23 02:05:31 -0700918 if (unlikely(committed_data))
Mingming Caoc089d492007-10-16 18:38:25 -0400919 jbd_free(committed_data, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 return err;
921}
922
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700923/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 * int journal_dirty_data() - mark a buffer as containing dirty data which
925 * needs to be flushed before we can commit the
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700926 * current transaction.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 * @handle: transaction
928 * @bh: bufferhead to mark
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700929 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 * The buffer is placed on the transaction's data list and is marked as
931 * belonging to the transaction.
932 *
933 * Returns error number or 0 on success.
934 *
935 * journal_dirty_data() can be called via page_launder->ext3_writepage
936 * by kswapd.
937 */
938int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
939{
940 journal_t *journal = handle->h_transaction->t_journal;
941 int need_brelse = 0;
942 struct journal_head *jh;
943
944 if (is_handle_aborted(handle))
945 return 0;
946
947 jh = journal_add_journal_head(bh);
948 JBUFFER_TRACE(jh, "entry");
949
950 /*
951 * The buffer could *already* be dirty. Writeout can start
952 * at any time.
953 */
954 jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
955
956 /*
957 * What if the buffer is already part of a running transaction?
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700958 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 * There are two cases:
960 * 1) It is part of the current running transaction. Refile it,
961 * just in case we have allocated it as metadata, deallocated
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700962 * it, then reallocated it as data.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 * 2) It is part of the previous, still-committing transaction.
964 * If all we want to do is to guarantee that the buffer will be
965 * written to disk before this new transaction commits, then
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700966 * being sure that the *previous* transaction has this same
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 * property is sufficient for us! Just leave it on its old
968 * transaction.
969 *
970 * In case (2), the buffer must not already exist as metadata
971 * --- that would violate write ordering (a transaction is free
972 * to write its data at any point, even before the previous
973 * committing transaction has committed). The caller must
974 * never, ever allow this to happen: there's nothing we can do
975 * about it in this layer.
976 */
977 jbd_lock_bh_state(bh);
978 spin_lock(&journal->j_list_lock);
Eric Sandeenf58a74d2006-10-28 10:38:27 -0700979
980 /* Now that we have bh_state locked, are we really still mapped? */
981 if (!buffer_mapped(bh)) {
982 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
983 goto no_journal;
984 }
985
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 if (jh->b_transaction) {
987 JBUFFER_TRACE(jh, "has transaction");
988 if (jh->b_transaction != handle->h_transaction) {
989 JBUFFER_TRACE(jh, "belongs to older transaction");
990 J_ASSERT_JH(jh, jh->b_transaction ==
991 journal->j_committing_transaction);
992
993 /* @@@ IS THIS TRUE ? */
994 /*
995 * Not any more. Scenario: someone does a write()
996 * in data=journal mode. The buffer's transaction has
997 * moved into commit. Then someone does another
998 * write() to the file. We do the frozen data copyout
999 * and set b_next_transaction to point to j_running_t.
1000 * And while we're in that state, someone does a
1001 * writepage() in an attempt to pageout the same area
1002 * of the file via a shared mapping. At present that
1003 * calls journal_dirty_data(), and we get right here.
1004 * It may be too late to journal the data. Simply
1005 * falling through to the next test will suffice: the
1006 * data will be dirty and wil be checkpointed. The
1007 * ordering comments in the next comment block still
1008 * apply.
1009 */
1010 //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1011
1012 /*
1013 * If we're journalling data, and this buffer was
1014 * subject to a write(), it could be metadata, forget
1015 * or shadow against the committing transaction. Now,
1016 * someone has dirtied the same darn page via a mapping
1017 * and it is being writepage()'d.
1018 * We *could* just steal the page from commit, with some
1019 * fancy locking there. Instead, we just skip it -
1020 * don't tie the page's buffers to the new transaction
1021 * at all.
1022 * Implication: if we crash before the writepage() data
1023 * is written into the filesystem, recovery will replay
1024 * the write() data.
1025 */
1026 if (jh->b_jlist != BJ_None &&
1027 jh->b_jlist != BJ_SyncData &&
1028 jh->b_jlist != BJ_Locked) {
1029 JBUFFER_TRACE(jh, "Not stealing");
1030 goto no_journal;
1031 }
1032
1033 /*
1034 * This buffer may be undergoing writeout in commit. We
1035 * can't return from here and let the caller dirty it
1036 * again because that can cause the write-out loop in
1037 * commit to never terminate.
1038 */
1039 if (buffer_dirty(bh)) {
1040 get_bh(bh);
1041 spin_unlock(&journal->j_list_lock);
1042 jbd_unlock_bh_state(bh);
1043 need_brelse = 1;
1044 sync_dirty_buffer(bh);
1045 jbd_lock_bh_state(bh);
1046 spin_lock(&journal->j_list_lock);
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001047 /* Since we dropped the lock... */
1048 if (!buffer_mapped(bh)) {
1049 JBUFFER_TRACE(jh, "buffer got unmapped");
1050 goto no_journal;
1051 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 /* The buffer may become locked again at any
1053 time if it is redirtied */
1054 }
1055
1056 /* journal_clean_data_list() may have got there first */
1057 if (jh->b_transaction != NULL) {
1058 JBUFFER_TRACE(jh, "unfile from commit");
1059 __journal_temp_unlink_buffer(jh);
1060 /* It still points to the committing
1061 * transaction; move it to this one so
1062 * that the refile assert checks are
1063 * happy. */
1064 jh->b_transaction = handle->h_transaction;
1065 }
1066 /* The buffer will be refiled below */
1067
1068 }
1069 /*
1070 * Special case --- the buffer might actually have been
1071 * allocated and then immediately deallocated in the previous,
1072 * committing transaction, so might still be left on that
1073 * transaction's metadata lists.
1074 */
1075 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1076 JBUFFER_TRACE(jh, "not on correct data list: unfile");
1077 J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1078 __journal_temp_unlink_buffer(jh);
1079 jh->b_transaction = handle->h_transaction;
1080 JBUFFER_TRACE(jh, "file as data");
1081 __journal_file_buffer(jh, handle->h_transaction,
1082 BJ_SyncData);
1083 }
1084 } else {
1085 JBUFFER_TRACE(jh, "not on a transaction");
1086 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1087 }
1088no_journal:
1089 spin_unlock(&journal->j_list_lock);
1090 jbd_unlock_bh_state(bh);
1091 if (need_brelse) {
1092 BUFFER_TRACE(bh, "brelse");
1093 __brelse(bh);
1094 }
1095 JBUFFER_TRACE(jh, "exit");
1096 journal_put_journal_head(jh);
1097 return 0;
1098}
1099
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001100/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1102 * @handle: transaction to add buffer to.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001103 * @bh: buffer to mark
1104 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105 * mark dirty metadata which needs to be journaled as part of the current
1106 * transaction.
1107 *
1108 * The buffer is placed on the transaction's metadata list and is marked
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001109 * as belonging to the transaction.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 *
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001111 * Returns error number or 0 on success.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 *
1113 * Special care needs to be taken if the buffer already belongs to the
1114 * current committing transaction (in which case we should have frozen
1115 * data present for that commit). In that case, we don't relink the
1116 * buffer: that only gets done when the old transaction finally
1117 * completes its commit.
1118 */
1119int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1120{
1121 transaction_t *transaction = handle->h_transaction;
1122 journal_t *journal = transaction->t_journal;
1123 struct journal_head *jh = bh2jh(bh);
1124
1125 jbd_debug(5, "journal_head %p\n", jh);
1126 JBUFFER_TRACE(jh, "entry");
1127 if (is_handle_aborted(handle))
1128 goto out;
1129
1130 jbd_lock_bh_state(bh);
1131
1132 if (jh->b_modified == 0) {
1133 /*
1134 * This buffer's got modified and becoming part
1135 * of the transaction. This needs to be done
1136 * once a transaction -bzzz
1137 */
1138 jh->b_modified = 1;
1139 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1140 handle->h_buffer_credits--;
1141 }
1142
1143 /*
1144 * fastpath, to avoid expensive locking. If this buffer is already
1145 * on the running transaction's metadata list there is nothing to do.
1146 * Nobody can take it off again because there is a handle open.
1147 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1148 * result in this test being false, so we go in and take the locks.
1149 */
1150 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1151 JBUFFER_TRACE(jh, "fastpath");
1152 J_ASSERT_JH(jh, jh->b_transaction ==
1153 journal->j_running_transaction);
1154 goto out_unlock_bh;
1155 }
1156
1157 set_buffer_jbddirty(bh);
1158
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001159 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160 * Metadata already on the current transaction list doesn't
1161 * need to be filed. Metadata on another transaction's list must
1162 * be committing, and will be refiled once the commit completes:
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001163 * leave it alone for now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164 */
1165 if (jh->b_transaction != transaction) {
1166 JBUFFER_TRACE(jh, "already on other transaction");
1167 J_ASSERT_JH(jh, jh->b_transaction ==
1168 journal->j_committing_transaction);
1169 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1170 /* And this case is illegal: we can't reuse another
1171 * transaction's data buffer, ever. */
1172 goto out_unlock_bh;
1173 }
1174
1175 /* That test should have eliminated the following case: */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001176 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177
1178 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1179 spin_lock(&journal->j_list_lock);
1180 __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1181 spin_unlock(&journal->j_list_lock);
1182out_unlock_bh:
1183 jbd_unlock_bh_state(bh);
1184out:
1185 JBUFFER_TRACE(jh, "exit");
1186 return 0;
1187}
1188
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001189/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190 * journal_release_buffer: undo a get_write_access without any buffer
1191 * updates, if the update decided in the end that it didn't need access.
1192 *
1193 */
1194void
1195journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1196{
1197 BUFFER_TRACE(bh, "entry");
1198}
1199
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001200/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201 * void journal_forget() - bforget() for potentially-journaled buffers.
1202 * @handle: transaction handle
1203 * @bh: bh to 'forget'
1204 *
1205 * We can only do the bforget if there are no commits pending against the
1206 * buffer. If the buffer is dirty in the current running transaction we
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001207 * can safely unlink it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 *
1209 * bh may not be a journalled buffer at all - it may be a non-JBD
1210 * buffer which came off the hashtable. Check for this.
1211 *
1212 * Decrements bh->b_count by one.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001213 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214 * Allow this call even if the handle has aborted --- it may be part of
1215 * the caller's cleanup after an abort.
1216 */
1217int journal_forget (handle_t *handle, struct buffer_head *bh)
1218{
1219 transaction_t *transaction = handle->h_transaction;
1220 journal_t *journal = transaction->t_journal;
1221 struct journal_head *jh;
1222 int drop_reserve = 0;
1223 int err = 0;
1224
1225 BUFFER_TRACE(bh, "entry");
1226
1227 jbd_lock_bh_state(bh);
1228 spin_lock(&journal->j_list_lock);
1229
1230 if (!buffer_jbd(bh))
1231 goto not_jbd;
1232 jh = bh2jh(bh);
1233
1234 /* Critical error: attempting to delete a bitmap buffer, maybe?
1235 * Don't do any jbd operations, and return an error. */
1236 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1237 "inconsistent data on disk")) {
1238 err = -EIO;
1239 goto not_jbd;
1240 }
1241
1242 /*
1243 * The buffer's going from the transaction, we must drop
1244 * all references -bzzz
1245 */
1246 jh->b_modified = 0;
1247
1248 if (jh->b_transaction == handle->h_transaction) {
1249 J_ASSERT_JH(jh, !jh->b_frozen_data);
1250
1251 /* If we are forgetting a buffer which is already part
1252 * of this transaction, then we can just drop it from
1253 * the transaction immediately. */
1254 clear_buffer_dirty(bh);
1255 clear_buffer_jbddirty(bh);
1256
1257 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1258
1259 drop_reserve = 1;
1260
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001261 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262 * We are no longer going to journal this buffer.
1263 * However, the commit of this transaction is still
1264 * important to the buffer: the delete that we are now
1265 * processing might obsolete an old log entry, so by
1266 * committing, we can satisfy the buffer's checkpoint.
1267 *
1268 * So, if we have a checkpoint on the buffer, we should
1269 * now refile the buffer on our BJ_Forget list so that
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001270 * we know to remove the checkpoint after we commit.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271 */
1272
1273 if (jh->b_cp_transaction) {
1274 __journal_temp_unlink_buffer(jh);
1275 __journal_file_buffer(jh, transaction, BJ_Forget);
1276 } else {
1277 __journal_unfile_buffer(jh);
1278 journal_remove_journal_head(bh);
1279 __brelse(bh);
1280 if (!buffer_jbd(bh)) {
1281 spin_unlock(&journal->j_list_lock);
1282 jbd_unlock_bh_state(bh);
1283 __bforget(bh);
1284 goto drop;
1285 }
1286 }
1287 } else if (jh->b_transaction) {
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001288 J_ASSERT_JH(jh, (jh->b_transaction ==
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 journal->j_committing_transaction));
1290 /* However, if the buffer is still owned by a prior
1291 * (committing) transaction, we can't drop it yet... */
1292 JBUFFER_TRACE(jh, "belongs to older transaction");
1293 /* ... but we CAN drop it from the new transaction if we
1294 * have also modified it since the original commit. */
1295
1296 if (jh->b_next_transaction) {
1297 J_ASSERT(jh->b_next_transaction == transaction);
1298 jh->b_next_transaction = NULL;
1299 drop_reserve = 1;
1300 }
1301 }
1302
1303not_jbd:
1304 spin_unlock(&journal->j_list_lock);
1305 jbd_unlock_bh_state(bh);
1306 __brelse(bh);
1307drop:
1308 if (drop_reserve) {
1309 /* no need to reserve log space for this block -bzzz */
1310 handle->h_buffer_credits++;
1311 }
1312 return err;
1313}
1314
1315/**
1316 * int journal_stop() - complete a transaction
1317 * @handle: tranaction to complete.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001318 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 * All done for a particular handle.
1320 *
1321 * There is not much action needed here. We just return any remaining
1322 * buffer credits to the transaction and remove the handle. The only
1323 * complication is that we need to start a commit operation if the
1324 * filesystem is marked for synchronous update.
1325 *
1326 * journal_stop itself will not usually return an error, but it may
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001327 * do so in unusual circumstances. In particular, expect it to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 * return -EIO if a journal_abort has been executed since the
1329 * transaction began.
1330 */
1331int journal_stop(handle_t *handle)
1332{
1333 transaction_t *transaction = handle->h_transaction;
1334 journal_t *journal = transaction->t_journal;
1335 int old_handle_count, err;
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001336 pid_t pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 J_ASSERT(journal_current_handle() == handle);
1339
1340 if (is_handle_aborted(handle))
1341 err = -EIO;
OGAWA Hirofumi3e2a5322006-10-19 23:29:11 -07001342 else {
1343 J_ASSERT(transaction->t_updates > 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 err = 0;
OGAWA Hirofumi3e2a5322006-10-19 23:29:11 -07001345 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346
1347 if (--handle->h_ref > 0) {
1348 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1349 handle->h_ref);
1350 return err;
1351 }
1352
1353 jbd_debug(4, "Handle %p going down\n", handle);
1354
1355 /*
1356 * Implement synchronous transaction batching. If the handle
1357 * was synchronous, don't force a commit immediately. Let's
1358 * yield and let another thread piggyback onto this transaction.
1359 * Keep doing that while new threads continue to arrive.
1360 * It doesn't cost much - we're about to run a commit and sleep
1361 * on IO anyway. Speeds up many-threaded, many-dir operations
1362 * by 30x or more...
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001363 *
1364 * But don't do this if this process was the most recent one to
1365 * perform a synchronous write. We do this to detect the case where a
1366 * single process is doing a stream of sync writes. No point in waiting
1367 * for joiners in that case.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 */
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001369 pid = current->pid;
1370 if (handle->h_sync && journal->j_last_sync_writer != pid) {
1371 journal->j_last_sync_writer = pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 do {
1373 old_handle_count = transaction->t_handle_count;
Nishanth Aravamudan041e0e32005-09-10 00:27:23 -07001374 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 } while (old_handle_count != transaction->t_handle_count);
1376 }
1377
1378 current->journal_info = NULL;
1379 spin_lock(&journal->j_state_lock);
1380 spin_lock(&transaction->t_handle_lock);
1381 transaction->t_outstanding_credits -= handle->h_buffer_credits;
1382 transaction->t_updates--;
1383 if (!transaction->t_updates) {
1384 wake_up(&journal->j_wait_updates);
1385 if (journal->j_barrier_count)
1386 wake_up(&journal->j_wait_transaction_locked);
1387 }
1388
1389 /*
1390 * If the handle is marked SYNC, we need to set another commit
1391 * going! We also want to force a commit if the current
1392 * transaction is occupying too much of the log, or if the
1393 * transaction is too old now.
1394 */
1395 if (handle->h_sync ||
1396 transaction->t_outstanding_credits >
1397 journal->j_max_transaction_buffers ||
Dave Kleikampe9ad5622006-09-27 01:49:35 -07001398 time_after_eq(jiffies, transaction->t_expires)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 /* Do this even for aborted journals: an abort still
1400 * completes the commit thread, it just doesn't write
1401 * anything to disk. */
1402 tid_t tid = transaction->t_tid;
1403
1404 spin_unlock(&transaction->t_handle_lock);
1405 jbd_debug(2, "transaction too old, requesting commit for "
1406 "handle %p\n", handle);
1407 /* This is non-blocking */
1408 __log_start_commit(journal, transaction->t_tid);
1409 spin_unlock(&journal->j_state_lock);
1410
1411 /*
1412 * Special case: JFS_SYNC synchronous updates require us
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001413 * to wait for the commit to complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 */
1415 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1416 err = log_wait_commit(journal, tid);
1417 } else {
1418 spin_unlock(&transaction->t_handle_lock);
1419 spin_unlock(&journal->j_state_lock);
1420 }
1421
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +02001422 lock_release(&handle->h_lockdep_map, 1, _THIS_IP_);
1423
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424 jbd_free_handle(handle);
1425 return err;
1426}
1427
1428/**int journal_force_commit() - force any uncommitted transactions
1429 * @journal: journal to force
1430 *
1431 * For synchronous operations: force any uncommitted transactions
1432 * to disk. May seem kludgy, but it reuses all the handle batching
1433 * code in a very simple manner.
1434 */
1435int journal_force_commit(journal_t *journal)
1436{
1437 handle_t *handle;
1438 int ret;
1439
1440 handle = journal_start(journal, 1);
1441 if (IS_ERR(handle)) {
1442 ret = PTR_ERR(handle);
1443 } else {
1444 handle->h_sync = 1;
1445 ret = journal_stop(handle);
1446 }
1447 return ret;
1448}
1449
1450/*
1451 *
1452 * List management code snippets: various functions for manipulating the
1453 * transaction buffer lists.
1454 *
1455 */
1456
1457/*
1458 * Append a buffer to a transaction list, given the transaction's list head
1459 * pointer.
1460 *
1461 * j_list_lock is held.
1462 *
1463 * jbd_lock_bh_state(jh2bh(jh)) is held.
1464 */
1465
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001466static inline void
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467__blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1468{
1469 if (!*list) {
1470 jh->b_tnext = jh->b_tprev = jh;
1471 *list = jh;
1472 } else {
1473 /* Insert at the tail of the list to preserve order */
1474 struct journal_head *first = *list, *last = first->b_tprev;
1475 jh->b_tprev = last;
1476 jh->b_tnext = first;
1477 last->b_tnext = first->b_tprev = jh;
1478 }
1479}
1480
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001481/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 * Remove a buffer from a transaction list, given the transaction's list
1483 * head pointer.
1484 *
1485 * Called with j_list_lock held, and the journal may not be locked.
1486 *
1487 * jbd_lock_bh_state(jh2bh(jh)) is held.
1488 */
1489
1490static inline void
1491__blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1492{
1493 if (*list == jh) {
1494 *list = jh->b_tnext;
1495 if (*list == jh)
1496 *list = NULL;
1497 }
1498 jh->b_tprev->b_tnext = jh->b_tnext;
1499 jh->b_tnext->b_tprev = jh->b_tprev;
1500}
1501
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001502/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 * Remove a buffer from the appropriate transaction list.
1504 *
1505 * Note that this function can *change* the value of
1506 * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1507 * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list. If the caller
1508 * is holding onto a copy of one of thee pointers, it could go bad.
1509 * Generally the caller needs to re-read the pointer from the transaction_t.
1510 *
1511 * Called under j_list_lock. The journal may not be locked.
1512 */
Adrian Bunkd394e122006-12-06 20:38:26 -08001513static void __journal_temp_unlink_buffer(struct journal_head *jh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514{
1515 struct journal_head **list = NULL;
1516 transaction_t *transaction;
1517 struct buffer_head *bh = jh2bh(jh);
1518
1519 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1520 transaction = jh->b_transaction;
1521 if (transaction)
1522 assert_spin_locked(&transaction->t_journal->j_list_lock);
1523
1524 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1525 if (jh->b_jlist != BJ_None)
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001526 J_ASSERT_JH(jh, transaction != NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527
1528 switch (jh->b_jlist) {
1529 case BJ_None:
1530 return;
1531 case BJ_SyncData:
1532 list = &transaction->t_sync_datalist;
1533 break;
1534 case BJ_Metadata:
1535 transaction->t_nr_buffers--;
1536 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1537 list = &transaction->t_buffers;
1538 break;
1539 case BJ_Forget:
1540 list = &transaction->t_forget;
1541 break;
1542 case BJ_IO:
1543 list = &transaction->t_iobuf_list;
1544 break;
1545 case BJ_Shadow:
1546 list = &transaction->t_shadow_list;
1547 break;
1548 case BJ_LogCtl:
1549 list = &transaction->t_log_list;
1550 break;
1551 case BJ_Reserved:
1552 list = &transaction->t_reserved_list;
1553 break;
1554 case BJ_Locked:
1555 list = &transaction->t_locked_list;
1556 break;
1557 }
1558
1559 __blist_del_buffer(list, jh);
1560 jh->b_jlist = BJ_None;
1561 if (test_clear_buffer_jbddirty(bh))
1562 mark_buffer_dirty(bh); /* Expose it to the VM */
1563}
1564
1565void __journal_unfile_buffer(struct journal_head *jh)
1566{
1567 __journal_temp_unlink_buffer(jh);
1568 jh->b_transaction = NULL;
1569}
1570
1571void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1572{
1573 jbd_lock_bh_state(jh2bh(jh));
1574 spin_lock(&journal->j_list_lock);
1575 __journal_unfile_buffer(jh);
1576 spin_unlock(&journal->j_list_lock);
1577 jbd_unlock_bh_state(jh2bh(jh));
1578}
1579
1580/*
1581 * Called from journal_try_to_free_buffers().
1582 *
1583 * Called under jbd_lock_bh_state(bh)
1584 */
1585static void
1586__journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1587{
1588 struct journal_head *jh;
1589
1590 jh = bh2jh(bh);
1591
1592 if (buffer_locked(bh) || buffer_dirty(bh))
1593 goto out;
1594
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001595 if (jh->b_next_transaction != NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 goto out;
1597
1598 spin_lock(&journal->j_list_lock);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001599 if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1601 /* A written-back ordered data buffer */
1602 JBUFFER_TRACE(jh, "release data");
1603 __journal_unfile_buffer(jh);
1604 journal_remove_journal_head(bh);
1605 __brelse(bh);
1606 }
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001607 } else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 /* written-back checkpointed metadata buffer */
1609 if (jh->b_jlist == BJ_None) {
1610 JBUFFER_TRACE(jh, "remove from checkpoint list");
1611 __journal_remove_checkpoint(jh);
1612 journal_remove_journal_head(bh);
1613 __brelse(bh);
1614 }
1615 }
1616 spin_unlock(&journal->j_list_lock);
1617out:
1618 return;
1619}
1620
1621
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001622/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 * int journal_try_to_free_buffers() - try to free page buffers.
1624 * @journal: journal for operation
1625 * @page: to try and free
1626 * @unused_gfp_mask: unused
1627 *
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001628 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 * For all the buffers on this page,
1630 * if they are fully written out ordered data, move them onto BUF_CLEAN
1631 * so try_to_free_buffers() can reap them.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001632 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 * This function returns non-zero if we wish try_to_free_buffers()
1634 * to be called. We do this if the page is releasable by try_to_free_buffers().
1635 * We also do it if the page has locked or dirty buffers and the caller wants
1636 * us to perform sync or async writeout.
1637 *
1638 * This complicates JBD locking somewhat. We aren't protected by the
1639 * BKL here. We wish to remove the buffer from its committing or
1640 * running transaction's ->t_datalist via __journal_unfile_buffer.
1641 *
1642 * This may *change* the value of transaction_t->t_datalist, so anyone
1643 * who looks at t_datalist needs to lock against this function.
1644 *
1645 * Even worse, someone may be doing a journal_dirty_data on this
1646 * buffer. So we need to lock against that. journal_dirty_data()
1647 * will come out of the lock with the buffer dirty, which makes it
1648 * ineligible for release here.
1649 *
1650 * Who else is affected by this? hmm... Really the only contender
1651 * is do_get_write_access() - it could be looking at the buffer while
1652 * journal_try_to_free_buffer() is changing its state. But that
1653 * cannot happen because we never reallocate freed data as metadata
1654 * while the data is part of a transaction. Yes?
1655 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001656int journal_try_to_free_buffers(journal_t *journal,
Al Viro27496a82005-10-21 03:20:48 -04001657 struct page *page, gfp_t unused_gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658{
1659 struct buffer_head *head;
1660 struct buffer_head *bh;
1661 int ret = 0;
1662
1663 J_ASSERT(PageLocked(page));
1664
1665 head = page_buffers(page);
1666 bh = head;
1667 do {
1668 struct journal_head *jh;
1669
1670 /*
1671 * We take our own ref against the journal_head here to avoid
1672 * having to add tons of locking around each instance of
1673 * journal_remove_journal_head() and journal_put_journal_head().
1674 */
1675 jh = journal_grab_journal_head(bh);
1676 if (!jh)
1677 continue;
1678
1679 jbd_lock_bh_state(bh);
1680 __journal_try_to_free_buffer(journal, bh);
1681 journal_put_journal_head(jh);
1682 jbd_unlock_bh_state(bh);
1683 if (buffer_jbd(bh))
1684 goto busy;
1685 } while ((bh = bh->b_this_page) != head);
1686 ret = try_to_free_buffers(page);
1687busy:
1688 return ret;
1689}
1690
1691/*
1692 * This buffer is no longer needed. If it is on an older transaction's
1693 * checkpoint list we need to record it on this transaction's forget list
1694 * to pin this buffer (and hence its checkpointing transaction) down until
1695 * this transaction commits. If the buffer isn't on a checkpoint list, we
1696 * release it.
1697 * Returns non-zero if JBD no longer has an interest in the buffer.
1698 *
1699 * Called under j_list_lock.
1700 *
1701 * Called under jbd_lock_bh_state(bh).
1702 */
1703static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1704{
1705 int may_free = 1;
1706 struct buffer_head *bh = jh2bh(jh);
1707
1708 __journal_unfile_buffer(jh);
1709
1710 if (jh->b_cp_transaction) {
1711 JBUFFER_TRACE(jh, "on running+cp transaction");
1712 __journal_file_buffer(jh, transaction, BJ_Forget);
1713 clear_buffer_jbddirty(bh);
1714 may_free = 0;
1715 } else {
1716 JBUFFER_TRACE(jh, "on running transaction");
1717 journal_remove_journal_head(bh);
1718 __brelse(bh);
1719 }
1720 return may_free;
1721}
1722
1723/*
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001724 * journal_invalidatepage
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 *
1726 * This code is tricky. It has a number of cases to deal with.
1727 *
1728 * There are two invariants which this code relies on:
1729 *
1730 * i_size must be updated on disk before we start calling invalidatepage on the
1731 * data.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001732 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733 * This is done in ext3 by defining an ext3_setattr method which
1734 * updates i_size before truncate gets going. By maintaining this
1735 * invariant, we can be sure that it is safe to throw away any buffers
1736 * attached to the current transaction: once the transaction commits,
1737 * we know that the data will not be needed.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001738 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739 * Note however that we can *not* throw away data belonging to the
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001740 * previous, committing transaction!
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 *
1742 * Any disk blocks which *are* part of the previous, committing
1743 * transaction (and which therefore cannot be discarded immediately) are
1744 * not going to be reused in the new running transaction
1745 *
1746 * The bitmap committed_data images guarantee this: any block which is
1747 * allocated in one transaction and removed in the next will be marked
1748 * as in-use in the committed_data bitmap, so cannot be reused until
1749 * the next transaction to delete the block commits. This means that
1750 * leaving committing buffers dirty is quite safe: the disk blocks
1751 * cannot be reallocated to a different file and so buffer aliasing is
1752 * not possible.
1753 *
1754 *
1755 * The above applies mainly to ordered data mode. In writeback mode we
1756 * don't make guarantees about the order in which data hits disk --- in
1757 * particular we don't guarantee that new dirty data is flushed before
1758 * transaction commit --- so it is always safe just to discard data
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001759 * immediately in that mode. --sct
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 */
1761
1762/*
1763 * The journal_unmap_buffer helper function returns zero if the buffer
1764 * concerned remains pinned as an anonymous buffer belonging to an older
1765 * transaction.
1766 *
1767 * We're outside-transaction here. Either or both of j_running_transaction
1768 * and j_committing_transaction may be NULL.
1769 */
1770static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1771{
1772 transaction_t *transaction;
1773 struct journal_head *jh;
1774 int may_free = 1;
1775 int ret;
1776
1777 BUFFER_TRACE(bh, "entry");
1778
1779 /*
1780 * It is safe to proceed here without the j_list_lock because the
1781 * buffers cannot be stolen by try_to_free_buffers as long as we are
1782 * holding the page lock. --sct
1783 */
1784
1785 if (!buffer_jbd(bh))
1786 goto zap_buffer_unlocked;
1787
1788 spin_lock(&journal->j_state_lock);
1789 jbd_lock_bh_state(bh);
1790 spin_lock(&journal->j_list_lock);
1791
1792 jh = journal_grab_journal_head(bh);
1793 if (!jh)
1794 goto zap_buffer_no_jh;
1795
1796 transaction = jh->b_transaction;
1797 if (transaction == NULL) {
1798 /* First case: not on any transaction. If it
1799 * has no checkpoint link, then we can zap it:
1800 * it's a writeback-mode buffer so we don't care
1801 * if it hits disk safely. */
1802 if (!jh->b_cp_transaction) {
1803 JBUFFER_TRACE(jh, "not on any transaction: zap");
1804 goto zap_buffer;
1805 }
1806
1807 if (!buffer_dirty(bh)) {
1808 /* bdflush has written it. We can drop it now */
1809 goto zap_buffer;
1810 }
1811
1812 /* OK, it must be in the journal but still not
1813 * written fully to disk: it's metadata or
1814 * journaled data... */
1815
1816 if (journal->j_running_transaction) {
1817 /* ... and once the current transaction has
1818 * committed, the buffer won't be needed any
1819 * longer. */
1820 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1821 ret = __dispose_buffer(jh,
1822 journal->j_running_transaction);
1823 journal_put_journal_head(jh);
1824 spin_unlock(&journal->j_list_lock);
1825 jbd_unlock_bh_state(bh);
1826 spin_unlock(&journal->j_state_lock);
1827 return ret;
1828 } else {
1829 /* There is no currently-running transaction. So the
1830 * orphan record which we wrote for this file must have
1831 * passed into commit. We must attach this buffer to
1832 * the committing transaction, if it exists. */
1833 if (journal->j_committing_transaction) {
1834 JBUFFER_TRACE(jh, "give to committing trans");
1835 ret = __dispose_buffer(jh,
1836 journal->j_committing_transaction);
1837 journal_put_journal_head(jh);
1838 spin_unlock(&journal->j_list_lock);
1839 jbd_unlock_bh_state(bh);
1840 spin_unlock(&journal->j_state_lock);
1841 return ret;
1842 } else {
1843 /* The orphan record's transaction has
1844 * committed. We can cleanse this buffer */
1845 clear_buffer_jbddirty(bh);
1846 goto zap_buffer;
1847 }
1848 }
1849 } else if (transaction == journal->j_committing_transaction) {
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001850 JBUFFER_TRACE(jh, "on committing transaction");
akpm@osdl.orgd13df842005-04-16 15:26:36 -07001851 if (jh->b_jlist == BJ_Locked) {
1852 /*
1853 * The buffer is on the committing transaction's locked
1854 * list. We have the buffer locked, so I/O has
1855 * completed. So we can nail the buffer now.
1856 */
1857 may_free = __dispose_buffer(jh, transaction);
1858 goto zap_buffer;
1859 }
1860 /*
1861 * If it is committing, we simply cannot touch it. We
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 * can remove it's next_transaction pointer from the
1863 * running transaction if that is set, but nothing
1864 * else. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 set_buffer_freed(bh);
1866 if (jh->b_next_transaction) {
1867 J_ASSERT(jh->b_next_transaction ==
1868 journal->j_running_transaction);
1869 jh->b_next_transaction = NULL;
1870 }
1871 journal_put_journal_head(jh);
1872 spin_unlock(&journal->j_list_lock);
1873 jbd_unlock_bh_state(bh);
1874 spin_unlock(&journal->j_state_lock);
1875 return 0;
1876 } else {
1877 /* Good, the buffer belongs to the running transaction.
1878 * We are writing our own transaction's data, not any
1879 * previous one's, so it is safe to throw it away
1880 * (remember that we expect the filesystem to have set
1881 * i_size already for this truncate so recovery will not
1882 * expose the disk blocks we are discarding here.) */
1883 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001884 JBUFFER_TRACE(jh, "on running transaction");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 may_free = __dispose_buffer(jh, transaction);
1886 }
1887
1888zap_buffer:
1889 journal_put_journal_head(jh);
1890zap_buffer_no_jh:
1891 spin_unlock(&journal->j_list_lock);
1892 jbd_unlock_bh_state(bh);
1893 spin_unlock(&journal->j_state_lock);
1894zap_buffer_unlocked:
1895 clear_buffer_dirty(bh);
1896 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1897 clear_buffer_mapped(bh);
1898 clear_buffer_req(bh);
1899 clear_buffer_new(bh);
1900 bh->b_bdev = NULL;
1901 return may_free;
1902}
1903
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001904/**
NeilBrown2ff28e22006-03-26 01:37:18 -08001905 * void journal_invalidatepage()
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001906 * @journal: journal to use for flush...
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 * @page: page to flush
1908 * @offset: length of page to invalidate.
1909 *
1910 * Reap page buffers containing data after offset in page.
1911 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 */
NeilBrown2ff28e22006-03-26 01:37:18 -08001913void journal_invalidatepage(journal_t *journal,
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001914 struct page *page,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 unsigned long offset)
1916{
1917 struct buffer_head *head, *bh, *next;
1918 unsigned int curr_off = 0;
1919 int may_free = 1;
1920
1921 if (!PageLocked(page))
1922 BUG();
1923 if (!page_has_buffers(page))
NeilBrown2ff28e22006-03-26 01:37:18 -08001924 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925
1926 /* We will potentially be playing with lists other than just the
1927 * data lists (especially for journaled data mode), so be
1928 * cautious in our locking. */
1929
1930 head = bh = page_buffers(page);
1931 do {
1932 unsigned int next_off = curr_off + bh->b_size;
1933 next = bh->b_this_page;
1934
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 if (offset <= curr_off) {
Dave Kleikampe9ad5622006-09-27 01:49:35 -07001936 /* This block is wholly outside the truncation point */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937 lock_buffer(bh);
1938 may_free &= journal_unmap_buffer(journal, bh);
1939 unlock_buffer(bh);
1940 }
1941 curr_off = next_off;
1942 bh = next;
1943
1944 } while (bh != head);
1945
1946 if (!offset) {
NeilBrown2ff28e22006-03-26 01:37:18 -08001947 if (may_free && try_to_free_buffers(page))
1948 J_ASSERT(!page_has_buffers(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950}
1951
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001952/*
1953 * File a buffer on the given transaction list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 */
1955void __journal_file_buffer(struct journal_head *jh,
1956 transaction_t *transaction, int jlist)
1957{
1958 struct journal_head **list = NULL;
1959 int was_dirty = 0;
1960 struct buffer_head *bh = jh2bh(jh);
1961
1962 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1963 assert_spin_locked(&transaction->t_journal->j_list_lock);
1964
1965 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1966 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001967 jh->b_transaction == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968
1969 if (jh->b_transaction && jh->b_jlist == jlist)
1970 return;
1971
1972 /* The following list of buffer states needs to be consistent
1973 * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1974 * state. */
1975
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001976 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 jlist == BJ_Shadow || jlist == BJ_Forget) {
1978 if (test_clear_buffer_dirty(bh) ||
1979 test_clear_buffer_jbddirty(bh))
1980 was_dirty = 1;
1981 }
1982
1983 if (jh->b_transaction)
1984 __journal_temp_unlink_buffer(jh);
1985 jh->b_transaction = transaction;
1986
1987 switch (jlist) {
1988 case BJ_None:
1989 J_ASSERT_JH(jh, !jh->b_committed_data);
1990 J_ASSERT_JH(jh, !jh->b_frozen_data);
1991 return;
1992 case BJ_SyncData:
1993 list = &transaction->t_sync_datalist;
1994 break;
1995 case BJ_Metadata:
1996 transaction->t_nr_buffers++;
1997 list = &transaction->t_buffers;
1998 break;
1999 case BJ_Forget:
2000 list = &transaction->t_forget;
2001 break;
2002 case BJ_IO:
2003 list = &transaction->t_iobuf_list;
2004 break;
2005 case BJ_Shadow:
2006 list = &transaction->t_shadow_list;
2007 break;
2008 case BJ_LogCtl:
2009 list = &transaction->t_log_list;
2010 break;
2011 case BJ_Reserved:
2012 list = &transaction->t_reserved_list;
2013 break;
2014 case BJ_Locked:
2015 list = &transaction->t_locked_list;
2016 break;
2017 }
2018
2019 __blist_add_buffer(list, jh);
2020 jh->b_jlist = jlist;
2021
2022 if (was_dirty)
2023 set_buffer_jbddirty(bh);
2024}
2025
2026void journal_file_buffer(struct journal_head *jh,
2027 transaction_t *transaction, int jlist)
2028{
2029 jbd_lock_bh_state(jh2bh(jh));
2030 spin_lock(&transaction->t_journal->j_list_lock);
2031 __journal_file_buffer(jh, transaction, jlist);
2032 spin_unlock(&transaction->t_journal->j_list_lock);
2033 jbd_unlock_bh_state(jh2bh(jh));
2034}
2035
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002036/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037 * Remove a buffer from its current buffer list in preparation for
2038 * dropping it from its current transaction entirely. If the buffer has
2039 * already started to be used by a subsequent transaction, refile the
2040 * buffer on that transaction's metadata list.
2041 *
2042 * Called under journal->j_list_lock
2043 *
2044 * Called under jbd_lock_bh_state(jh2bh(jh))
2045 */
2046void __journal_refile_buffer(struct journal_head *jh)
2047{
2048 int was_dirty;
2049 struct buffer_head *bh = jh2bh(jh);
2050
2051 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2052 if (jh->b_transaction)
2053 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2054
2055 /* If the buffer is now unused, just drop it. */
2056 if (jh->b_next_transaction == NULL) {
2057 __journal_unfile_buffer(jh);
2058 return;
2059 }
2060
2061 /*
2062 * It has been modified by a later transaction: add it to the new
2063 * transaction's metadata list.
2064 */
2065
2066 was_dirty = test_clear_buffer_jbddirty(bh);
2067 __journal_temp_unlink_buffer(jh);
2068 jh->b_transaction = jh->b_next_transaction;
2069 jh->b_next_transaction = NULL;
Jan Kara9ada7342006-06-23 02:05:25 -07002070 __journal_file_buffer(jh, jh->b_transaction,
2071 was_dirty ? BJ_Metadata : BJ_Reserved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2073
2074 if (was_dirty)
2075 set_buffer_jbddirty(bh);
2076}
2077
2078/*
2079 * For the unlocked version of this call, also make sure that any
2080 * hanging journal_head is cleaned up if necessary.
2081 *
2082 * __journal_refile_buffer is usually called as part of a single locked
2083 * operation on a buffer_head, in which the caller is probably going to
2084 * be hooking the journal_head onto other lists. In that case it is up
2085 * to the caller to remove the journal_head if necessary. For the
2086 * unlocked journal_refile_buffer call, the caller isn't going to be
2087 * doing anything else to the buffer so we need to do the cleanup
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002088 * ourselves to avoid a jh leak.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089 *
2090 * *** The journal_head may be freed by this call! ***
2091 */
2092void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2093{
2094 struct buffer_head *bh = jh2bh(jh);
2095
2096 jbd_lock_bh_state(bh);
2097 spin_lock(&journal->j_list_lock);
2098
2099 __journal_refile_buffer(jh);
2100 jbd_unlock_bh_state(bh);
2101 journal_remove_journal_head(bh);
2102
2103 spin_unlock(&journal->j_list_lock);
2104 __brelse(bh);
2105}