blob: 1695ba8334a2a640ad21dfb0030eb0c63c0d2e61 [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>
Josef Bacikf420d4d2009-01-07 18:07:24 -080028#include <linux/hrtimer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029
Adrian Bunkd394e122006-12-06 20:38:26 -080030static void __journal_temp_unlink_buffer(struct journal_head *jh);
31
Linus Torvalds1da177e2005-04-16 15:20:36 -070032/*
33 * get_transaction: obtain a new transaction_t object.
34 *
35 * Simply allocate and initialise a new transaction. Create it in
36 * RUNNING state and add it to the current journal (which should not
37 * have an existing running transaction: we only make a new transaction
38 * once we have started to commit the old one).
39 *
40 * Preconditions:
41 * The journal MUST be locked. We don't perform atomic mallocs on the
42 * new transaction and we can't block without protecting against other
43 * processes trying to touch the journal while it is in transition.
44 *
45 * Called under j_state_lock
46 */
47
48static transaction_t *
49get_transaction(journal_t *journal, transaction_t *transaction)
50{
51 transaction->t_journal = journal;
52 transaction->t_state = T_RUNNING;
Josef Bacikf420d4d2009-01-07 18:07:24 -080053 transaction->t_start_time = ktime_get();
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 transaction->t_tid = journal->j_transaction_sequence++;
55 transaction->t_expires = jiffies + journal->j_commit_interval;
56 spin_lock_init(&transaction->t_handle_lock);
57
58 /* Set up the commit timer for the new transaction. */
Andreas Dilgerb449fc62009-07-30 20:09:46 +020059 journal->j_commit_timer.expires =
60 round_jiffies_up(transaction->t_expires);
Andrew Mortone3df1892006-03-25 03:06:53 -080061 add_timer(&journal->j_commit_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
63 J_ASSERT(journal->j_running_transaction == NULL);
64 journal->j_running_transaction = transaction;
65
66 return transaction;
67}
68
69/*
70 * Handle management.
71 *
72 * A handle_t is an object which represents a single atomic update to a
73 * filesystem, and which tracks all of the modifications which form part
74 * of that one update.
75 */
76
77/*
78 * start_this_handle: Given a handle, deal with any locking or stalling
79 * needed to make sure that there is enough journal space for the handle
80 * to begin. Attach the handle to a transaction and set up the
Mingming Caoae6ddcc2006-09-27 01:49:27 -070081 * transaction's buffer credits.
Linus Torvalds1da177e2005-04-16 15:20:36 -070082 */
83
84static int start_this_handle(journal_t *journal, handle_t *handle)
85{
86 transaction_t *transaction;
87 int needed;
88 int nblocks = handle->h_buffer_credits;
89 transaction_t *new_transaction = NULL;
90 int ret = 0;
91
92 if (nblocks > journal->j_max_transaction_buffers) {
93 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
94 current->comm, nblocks,
95 journal->j_max_transaction_buffers);
96 ret = -ENOSPC;
97 goto out;
98 }
99
100alloc_transaction:
101 if (!journal->j_running_transaction) {
Jan Kara1ce0aa82013-10-31 20:37:15 +0100102 new_transaction = kzalloc(sizeof(*new_transaction),
103 GFP_NOFS|__GFP_NOFAIL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104 if (!new_transaction) {
Jan Kara1ce0aa82013-10-31 20:37:15 +0100105 ret = -ENOMEM;
106 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108 }
109
110 jbd_debug(3, "New handle %p going live.\n", handle);
111
112repeat:
113
114 /*
115 * We need to hold j_state_lock until t_updates has been incremented,
116 * for proper journal barrier handling
117 */
118 spin_lock(&journal->j_state_lock);
119repeat_locked:
120 if (is_journal_aborted(journal) ||
121 (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
122 spin_unlock(&journal->j_state_lock);
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700123 ret = -EROFS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 goto out;
125 }
126
127 /* Wait on the journal's transaction barrier if necessary */
128 if (journal->j_barrier_count) {
129 spin_unlock(&journal->j_state_lock);
130 wait_event(journal->j_wait_transaction_locked,
131 journal->j_barrier_count == 0);
132 goto repeat;
133 }
134
135 if (!journal->j_running_transaction) {
136 if (!new_transaction) {
137 spin_unlock(&journal->j_state_lock);
138 goto alloc_transaction;
139 }
140 get_transaction(journal, new_transaction);
141 new_transaction = NULL;
142 }
143
144 transaction = journal->j_running_transaction;
145
146 /*
147 * If the current transaction is locked down for commit, wait for the
148 * lock to be released.
149 */
150 if (transaction->t_state == T_LOCKED) {
151 DEFINE_WAIT(wait);
152
153 prepare_to_wait(&journal->j_wait_transaction_locked,
154 &wait, TASK_UNINTERRUPTIBLE);
155 spin_unlock(&journal->j_state_lock);
156 schedule();
157 finish_wait(&journal->j_wait_transaction_locked, &wait);
158 goto repeat;
159 }
160
161 /*
162 * If there is not enough space left in the log to write all potential
163 * buffers requested by this operation, we need to stall pending a log
164 * checkpoint to free some more log space.
165 */
166 spin_lock(&transaction->t_handle_lock);
167 needed = transaction->t_outstanding_credits + nblocks;
168
169 if (needed > journal->j_max_transaction_buffers) {
170 /*
171 * If the current transaction is already too large, then start
172 * to commit it: we can then go back and attach this handle to
173 * a new transaction.
174 */
175 DEFINE_WAIT(wait);
176
177 jbd_debug(2, "Handle %p starting new commit...\n", handle);
178 spin_unlock(&transaction->t_handle_lock);
179 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
180 TASK_UNINTERRUPTIBLE);
181 __log_start_commit(journal, transaction->t_tid);
182 spin_unlock(&journal->j_state_lock);
183 schedule();
184 finish_wait(&journal->j_wait_transaction_locked, &wait);
185 goto repeat;
186 }
187
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700188 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 * The commit code assumes that it can get enough log space
190 * without forcing a checkpoint. This is *critical* for
191 * correctness: a checkpoint of a buffer which is also
192 * associated with a committing transaction creates a deadlock,
193 * so commit simply cannot force through checkpoints.
194 *
195 * We must therefore ensure the necessary space in the journal
196 * *before* starting to dirty potentially checkpointed buffers
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700197 * in the new transaction.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 *
199 * The worst part is, any transaction currently committing can
200 * reduce the free space arbitrarily. Be careful to account for
201 * those buffers when checkpointing.
202 */
203
204 /*
205 * @@@ AKPM: This seems rather over-defensive. We're giving commit
206 * a _lot_ of headroom: 1/4 of the journal plus the size of
207 * the committing transaction. Really, we only need to give it
208 * committing_transaction->t_outstanding_credits plus "enough" for
209 * the log control blocks.
Uwe Kleine-Königa34f0b32010-12-10 14:55:42 +0100210 * Also, this test is inconsistent with the matching one in
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 * journal_extend().
212 */
213 if (__log_space_left(journal) < jbd_space_needed(journal)) {
214 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
215 spin_unlock(&transaction->t_handle_lock);
216 __log_wait_for_space(journal);
217 goto repeat_locked;
218 }
219
220 /* OK, account for the buffers that this operation expects to
221 * use and add the handle to the running transaction. */
222
223 handle->h_transaction = transaction;
224 transaction->t_outstanding_credits += nblocks;
225 transaction->t_updates++;
226 transaction->t_handle_count++;
227 jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
228 handle, nblocks, transaction->t_outstanding_credits,
229 __log_space_left(journal));
230 spin_unlock(&transaction->t_handle_lock);
231 spin_unlock(&journal->j_state_lock);
Jan Kara3adae9d2009-08-11 17:27:21 +0200232
233 lock_map_acquire(&handle->h_lockdep_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234out:
Andrew Morton304c4c82006-06-23 02:05:31 -0700235 if (unlikely(new_transaction)) /* It's usually NULL */
236 kfree(new_transaction);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 return ret;
238}
239
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +0200240static struct lock_class_key jbd_handle_key;
241
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242/* Allocate a new handle. This should probably be in a slab... */
243static handle_t *new_handle(int nblocks)
244{
245 handle_t *handle = jbd_alloc_handle(GFP_NOFS);
246 if (!handle)
247 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248 handle->h_buffer_credits = nblocks;
249 handle->h_ref = 1;
250
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +0200251 lockdep_init_map(&handle->h_lockdep_map, "jbd_handle", &jbd_handle_key, 0);
252
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 return handle;
254}
255
256/**
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700257 * handle_t *journal_start() - Obtain a new handle.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 * @journal: Journal to start transaction on.
259 * @nblocks: number of block buffer we might modify
260 *
261 * We make sure that the transaction can guarantee at least nblocks of
262 * modified buffers in the log. We block until the log can guarantee
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700263 * that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264 *
265 * This function is visible to journal users (like ext3fs), so is not
266 * called with the journal already locked.
267 *
Eryu Guanc2b67732011-05-24 00:17:34 +0800268 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
269 * on failure.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 */
271handle_t *journal_start(journal_t *journal, int nblocks)
272{
273 handle_t *handle = journal_current_handle();
274 int err;
275
276 if (!journal)
277 return ERR_PTR(-EROFS);
278
279 if (handle) {
280 J_ASSERT(handle->h_transaction->t_journal == journal);
281 handle->h_ref++;
282 return handle;
283 }
284
285 handle = new_handle(nblocks);
286 if (!handle)
287 return ERR_PTR(-ENOMEM);
288
289 current->journal_info = handle;
290
291 err = start_this_handle(journal, handle);
292 if (err < 0) {
293 jbd_free_handle(handle);
294 current->journal_info = NULL;
295 handle = ERR_PTR(err);
296 }
297 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
Jan Kara3adae9d2009-08-11 17:27:21 +0200417 lock_map_release(&handle->h_lockdep_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 handle->h_buffer_credits = nblocks;
419 ret = start_this_handle(journal, handle);
420 return ret;
421}
422
423
424/**
425 * void journal_lock_updates () - establish a transaction barrier.
426 * @journal: Journal to establish a barrier on.
427 *
Jan Kara00482782011-12-22 14:52:21 +0100428 * This locks out any further updates from being started, and blocks until all
429 * existing updates have completed, returning only once the journal is in a
430 * quiescent state with no updates running.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431 *
Jan Kara00482782011-12-22 14:52:21 +0100432 * We do not use simple mutex for synchronization as there are syscalls which
433 * want to return with filesystem locked and that trips up lockdep. Also
434 * hibernate needs to lock filesystem but locked mutex then blocks hibernation.
435 * Since locking filesystem is rare operation, we use simple counter and
436 * waitqueue for locking.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 */
438void journal_lock_updates(journal_t *journal)
439{
440 DEFINE_WAIT(wait);
441
Jan Kara00482782011-12-22 14:52:21 +0100442wait:
443 /* Wait for previous locked operation to finish */
444 wait_event(journal->j_wait_transaction_locked,
445 journal->j_barrier_count == 0);
446
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 spin_lock(&journal->j_state_lock);
Jan Kara00482782011-12-22 14:52:21 +0100448 /*
449 * Check reliably under the lock whether we are the ones winning the race
450 * and locking the journal
451 */
452 if (journal->j_barrier_count > 0) {
453 spin_unlock(&journal->j_state_lock);
454 goto wait;
455 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 ++journal->j_barrier_count;
457
458 /* Wait until there are no running updates */
459 while (1) {
460 transaction_t *transaction = journal->j_running_transaction;
461
462 if (!transaction)
463 break;
464
465 spin_lock(&transaction->t_handle_lock);
466 if (!transaction->t_updates) {
467 spin_unlock(&transaction->t_handle_lock);
468 break;
469 }
470 prepare_to_wait(&journal->j_wait_updates, &wait,
471 TASK_UNINTERRUPTIBLE);
472 spin_unlock(&transaction->t_handle_lock);
473 spin_unlock(&journal->j_state_lock);
474 schedule();
475 finish_wait(&journal->j_wait_updates, &wait);
476 spin_lock(&journal->j_state_lock);
477 }
478 spin_unlock(&journal->j_state_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479}
480
481/**
482 * void journal_unlock_updates (journal_t* journal) - release barrier
483 * @journal: Journal to release the barrier on.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700484 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 * Release a transaction barrier obtained with journal_lock_updates().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 */
487void journal_unlock_updates (journal_t *journal)
488{
489 J_ASSERT(journal->j_barrier_count != 0);
490
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 spin_lock(&journal->j_state_lock);
492 --journal->j_barrier_count;
493 spin_unlock(&journal->j_state_lock);
494 wake_up(&journal->j_wait_transaction_locked);
495}
496
Jan Kara1e9fd532009-06-24 17:31:40 +0200497static void warn_dirty_buffer(struct buffer_head *bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498{
Jan Kara1e9fd532009-06-24 17:31:40 +0200499 char b[BDEVNAME_SIZE];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500
Jan Kara1e9fd532009-06-24 17:31:40 +0200501 printk(KERN_WARNING
502 "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
503 "There's a risk of filesystem corruption in case of system "
504 "crash.\n",
505 bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506}
507
508/*
509 * If the buffer is already part of the current transaction, then there
510 * is nothing we need to do. If it is already part of a prior
511 * transaction which we are still committing to disk, then we need to
512 * make sure that we do not overwrite the old copy: we do copy-out to
513 * preserve the copy going to disk. We also account the buffer against
514 * the handle's metadata buffer credits (unless the buffer is already
515 * part of the transaction, that is).
516 *
517 */
518static int
519do_get_write_access(handle_t *handle, struct journal_head *jh,
520 int force_copy)
521{
522 struct buffer_head *bh;
523 transaction_t *transaction;
524 journal_t *journal;
525 int error;
526 char *frozen_buffer = NULL;
527 int need_copy = 0;
528
529 if (is_handle_aborted(handle))
530 return -EROFS;
531
532 transaction = handle->h_transaction;
533 journal = transaction->t_journal;
534
Namhyung Kimb8ea49f2010-10-13 23:14:07 +0900535 jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536
537 JBUFFER_TRACE(jh, "entry");
538repeat:
539 bh = jh2bh(jh);
540
541 /* @@@ Need to check for errors here at some point. */
542
543 lock_buffer(bh);
544 jbd_lock_bh_state(bh);
545
546 /* We now hold the buffer lock so it is safe to query the buffer
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700547 * state. Is the buffer dirty?
548 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 * If so, there are two possibilities. The buffer may be
550 * non-journaled, and undergoing a quite legitimate writeback.
551 * Otherwise, it is journaled, and we don't expect dirty buffers
552 * in that state (the buffers should be marked JBD_Dirty
553 * instead.) So either the IO is being done under our own
554 * control and this is a bug, or it's a third party IO such as
555 * dump(8) (which may leave the buffer scheduled for read ---
556 * ie. locked but not dirty) or tune2fs (which may actually have
557 * the buffer dirtied, ugh.) */
558
559 if (buffer_dirty(bh)) {
560 /*
561 * First question: is this buffer already part of the current
562 * transaction or the existing committing transaction?
563 */
564 if (jh->b_transaction) {
565 J_ASSERT_JH(jh,
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700566 jh->b_transaction == transaction ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 jh->b_transaction ==
568 journal->j_committing_transaction);
569 if (jh->b_next_transaction)
570 J_ASSERT_JH(jh, jh->b_next_transaction ==
571 transaction);
Jan Kara1e9fd532009-06-24 17:31:40 +0200572 warn_dirty_buffer(bh);
Jan Kara4407c2b2005-09-06 15:19:17 -0700573 }
574 /*
575 * In any case we need to clean the dirty flag and we must
576 * do it under the buffer lock to be sure we don't race
577 * with running write-out.
578 */
Jan Kara1e9fd532009-06-24 17:31:40 +0200579 JBUFFER_TRACE(jh, "Journalling dirty buffer");
580 clear_buffer_dirty(bh);
581 set_buffer_jbddirty(bh);
Dave Kleikampe9ad5622006-09-27 01:49:35 -0700582 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583
584 unlock_buffer(bh);
585
586 error = -EROFS;
587 if (is_handle_aborted(handle)) {
588 jbd_unlock_bh_state(bh);
589 goto out;
590 }
591 error = 0;
592
593 /*
594 * The buffer is already part of this transaction if b_transaction or
595 * b_next_transaction points to it
596 */
597 if (jh->b_transaction == transaction ||
598 jh->b_next_transaction == transaction)
599 goto done;
600
601 /*
Josef Bacik5bc833f2008-04-28 02:16:10 -0700602 * this is the first time this transaction is touching this buffer,
603 * reset the modified flag
604 */
605 jh->b_modified = 0;
606
607 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608 * If there is already a copy-out version of this buffer, then we don't
609 * need to make another one
610 */
611 if (jh->b_frozen_data) {
612 JBUFFER_TRACE(jh, "has frozen data");
613 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
614 jh->b_next_transaction = transaction;
615 goto done;
616 }
617
618 /* Is there data here we need to preserve? */
619
620 if (jh->b_transaction && jh->b_transaction != transaction) {
621 JBUFFER_TRACE(jh, "owned by older transaction");
622 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
623 J_ASSERT_JH(jh, jh->b_transaction ==
624 journal->j_committing_transaction);
625
626 /* There is one case we have to be very careful about.
627 * If the committing transaction is currently writing
628 * this buffer out to disk and has NOT made a copy-out,
629 * then we cannot modify the buffer contents at all
630 * right now. The essence of copy-out is that it is the
631 * extra copy, not the primary copy, which gets
632 * journaled. If the primary copy is already going to
633 * disk then we cannot do copy-out here. */
634
635 if (jh->b_jlist == BJ_Shadow) {
636 DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
637 wait_queue_head_t *wqh;
638
639 wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
640
641 JBUFFER_TRACE(jh, "on shadow: sleep");
642 jbd_unlock_bh_state(bh);
643 /* commit wakes up all shadow buffers after IO */
644 for ( ; ; ) {
645 prepare_to_wait(wqh, &wait.wait,
646 TASK_UNINTERRUPTIBLE);
647 if (jh->b_jlist != BJ_Shadow)
648 break;
649 schedule();
650 }
651 finish_wait(wqh, &wait.wait);
652 goto repeat;
653 }
654
655 /* Only do the copy if the currently-owning transaction
656 * still needs it. If it is on the Forget list, the
657 * committing transaction is past that stage. The
658 * buffer had better remain locked during the kmalloc,
659 * but that should be true --- we hold the journal lock
660 * still and the buffer is already on the BUF_JOURNAL
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700661 * list so won't be flushed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 *
663 * Subtle point, though: if this is a get_undo_access,
664 * then we will be relying on the frozen_data to contain
665 * the new value of the committed_data record after the
666 * transaction, so we HAVE to force the frozen_data copy
667 * in that case. */
668
669 if (jh->b_jlist != BJ_Forget || force_copy) {
670 JBUFFER_TRACE(jh, "generate frozen data");
671 if (!frozen_buffer) {
672 JBUFFER_TRACE(jh, "allocate memory for buffer");
673 jbd_unlock_bh_state(bh);
Badari Pulavartyea817392006-08-27 01:23:52 -0700674 frozen_buffer =
Mingming Caoc089d492007-10-16 18:38:25 -0400675 jbd_alloc(jh2bh(jh)->b_size,
Badari Pulavartyea817392006-08-27 01:23:52 -0700676 GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 if (!frozen_buffer) {
Jan Kara301d4c92013-11-25 23:30:26 +0100678 printk(KERN_ERR
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 "%s: OOM for frozen_buffer\n",
Harvey Harrison08fc99b2008-04-28 02:16:16 -0700680 __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 JBUFFER_TRACE(jh, "oom!");
682 error = -ENOMEM;
683 jbd_lock_bh_state(bh);
684 goto done;
685 }
686 goto repeat;
687 }
688 jh->b_frozen_data = frozen_buffer;
689 frozen_buffer = NULL;
690 need_copy = 1;
691 }
692 jh->b_next_transaction = transaction;
693 }
694
695
696 /*
697 * Finally, if the buffer is not journaled right now, we need to make
698 * sure it doesn't get written to disk before the caller actually
699 * commits the new data
700 */
701 if (!jh->b_transaction) {
702 JBUFFER_TRACE(jh, "no transaction");
703 J_ASSERT_JH(jh, !jh->b_next_transaction);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 JBUFFER_TRACE(jh, "file as BJ_Reserved");
705 spin_lock(&journal->j_list_lock);
706 __journal_file_buffer(jh, transaction, BJ_Reserved);
707 spin_unlock(&journal->j_list_lock);
708 }
709
710done:
711 if (need_copy) {
712 struct page *page;
713 int offset;
714 char *source;
715
716 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
717 "Possible IO failure.\n");
718 page = jh2bh(jh)->b_page;
Namhyung Kim8117f982010-10-05 01:03:40 +0900719 offset = offset_in_page(jh2bh(jh)->b_data);
Cong Wang8fb53c42011-11-25 23:14:31 +0800720 source = kmap_atomic(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
Cong Wang8fb53c42011-11-25 23:14:31 +0800722 kunmap_atomic(source);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 }
724 jbd_unlock_bh_state(bh);
725
726 /*
727 * If we are about to journal a buffer, then any revoke pending on it is
728 * no longer valid
729 */
730 journal_cancel_revoke(handle, jh);
731
732out:
Andrew Morton304c4c82006-06-23 02:05:31 -0700733 if (unlikely(frozen_buffer)) /* It's usually NULL */
Mingming Caoc089d492007-10-16 18:38:25 -0400734 jbd_free(frozen_buffer, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735
736 JBUFFER_TRACE(jh, "exit");
737 return error;
738}
739
740/**
741 * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
742 * @handle: transaction to add buffer modifications to
743 * @bh: bh to be used for metadata writes
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 *
745 * Returns an error code or 0 on success.
746 *
747 * In full data journalling mode the buffer may be of type BJ_AsyncData,
748 * because we're write()ing a buffer which is also part of a shared mapping.
749 */
750
751int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
752{
753 struct journal_head *jh = journal_add_journal_head(bh);
754 int rc;
755
756 /* We do not want to get caught playing with fields which the
757 * log thread also manipulates. Make sure that the buffer
758 * completes any outstanding IO before proceeding. */
759 rc = do_get_write_access(handle, jh, 0);
760 journal_put_journal_head(jh);
761 return rc;
762}
763
764
765/*
766 * When the user wants to journal a newly created buffer_head
767 * (ie. getblk() returned a new buffer and we are going to populate it
768 * manually rather than reading off disk), then we need to keep the
769 * buffer_head locked until it has been completely filled with new
770 * data. In this case, we should be able to make the assertion that
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700771 * the bh is not already part of an existing transaction.
772 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 * The buffer should already be locked by the caller by this point.
774 * There is no lock ranking violation: it was a newly created,
775 * unlocked buffer beforehand. */
776
777/**
778 * int journal_get_create_access () - notify intent to use newly created bh
779 * @handle: transaction to new buffer to
780 * @bh: new buffer.
781 *
782 * Call this if you create a new bh.
783 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700784int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785{
786 transaction_t *transaction = handle->h_transaction;
787 journal_t *journal = transaction->t_journal;
788 struct journal_head *jh = journal_add_journal_head(bh);
789 int err;
790
791 jbd_debug(5, "journal_head %p\n", jh);
792 err = -EROFS;
793 if (is_handle_aborted(handle))
794 goto out;
795 err = 0;
796
797 JBUFFER_TRACE(jh, "entry");
798 /*
799 * The buffer may already belong to this transaction due to pre-zeroing
800 * in the filesystem's new_block code. It may also be on the previous,
801 * committing transaction's lists, but it HAS to be in Forget state in
802 * that case: the transaction must have deleted the buffer for it to be
803 * reused here.
804 */
805 jbd_lock_bh_state(bh);
806 spin_lock(&journal->j_list_lock);
807 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
808 jh->b_transaction == NULL ||
809 (jh->b_transaction == journal->j_committing_transaction &&
810 jh->b_jlist == BJ_Forget)));
811
812 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
813 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
814
815 if (jh->b_transaction == NULL) {
Jan Kara1e9fd532009-06-24 17:31:40 +0200816 /*
817 * Previous journal_forget() could have left the buffer
818 * with jbddirty bit set because it was being committed. When
819 * the commit finished, we've filed the buffer for
820 * checkpointing and marked it dirty. Now we are reallocating
821 * the buffer so the transaction freeing it must have
822 * committed and so it's safe to clear the dirty bit.
823 */
824 clear_buffer_dirty(jh2bh(jh));
Josef Bacik5bc833f2008-04-28 02:16:10 -0700825
826 /* first access by this transaction */
827 jh->b_modified = 0;
828
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 JBUFFER_TRACE(jh, "file as BJ_Reserved");
830 __journal_file_buffer(jh, transaction, BJ_Reserved);
831 } else if (jh->b_transaction == journal->j_committing_transaction) {
Josef Bacik5bc833f2008-04-28 02:16:10 -0700832 /* first access by this transaction */
833 jh->b_modified = 0;
834
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 JBUFFER_TRACE(jh, "set next transaction");
836 jh->b_next_transaction = transaction;
837 }
838 spin_unlock(&journal->j_list_lock);
839 jbd_unlock_bh_state(bh);
840
841 /*
842 * akpm: I added this. ext3_alloc_branch can pick up new indirect
843 * blocks which contain freed but then revoked metadata. We need
844 * to cancel the revoke in case we end up freeing it yet again
845 * and the reallocating as data - this would cause a second revoke,
846 * which hits an assertion error.
847 */
848 JBUFFER_TRACE(jh, "cancelling revoke");
849 journal_cancel_revoke(handle, jh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850out:
Ding Dinghuabd5c9e12011-05-26 10:29:01 +0800851 journal_put_journal_head(jh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 return err;
853}
854
855/**
Randy Dunlap78a4a502008-02-29 22:02:31 -0800856 * int journal_get_undo_access() - Notify intent to modify metadata with non-rewindable consequences
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 * @handle: transaction
858 * @bh: buffer to undo
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 *
860 * Sometimes there is a need to distinguish between metadata which has
861 * been committed to disk and that which has not. The ext3fs code uses
862 * this for freeing and allocating space, we have to make sure that we
863 * do not reuse freed space until the deallocation has been committed,
864 * since if we overwrote that space we would make the delete
865 * un-rewindable in case of a crash.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700866 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 * To deal with that, journal_get_undo_access requests write access to a
868 * buffer for parts of non-rewindable operations such as delete
869 * operations on the bitmaps. The journaling code must keep a copy of
870 * the buffer's contents prior to the undo_access call until such time
871 * as we know that the buffer has definitely been committed to disk.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700872 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873 * We never need to know which transaction the committed data is part
874 * of, buffers touched here are guaranteed to be dirtied later and so
875 * will be committed to a new transaction in due course, at which point
876 * we can discard the old committed data pointer.
877 *
878 * Returns error number or 0 on success.
879 */
880int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
881{
882 int err;
883 struct journal_head *jh = journal_add_journal_head(bh);
884 char *committed_data = NULL;
885
886 JBUFFER_TRACE(jh, "entry");
887
888 /*
889 * Do this first --- it can drop the journal lock, so we want to
890 * make sure that obtaining the committed_data is done
891 * atomically wrt. completion of any outstanding commits.
892 */
893 err = do_get_write_access(handle, jh, 1);
894 if (err)
895 goto out;
896
897repeat:
898 if (!jh->b_committed_data) {
Mingming Caoc089d492007-10-16 18:38:25 -0400899 committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 if (!committed_data) {
Jan Kara301d4c92013-11-25 23:30:26 +0100901 printk(KERN_ERR "%s: No memory for committed data\n",
Harvey Harrison08fc99b2008-04-28 02:16:16 -0700902 __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 err = -ENOMEM;
904 goto out;
905 }
906 }
907
908 jbd_lock_bh_state(bh);
909 if (!jh->b_committed_data) {
910 /* Copy out the current buffer contents into the
911 * preserved, committed copy. */
912 JBUFFER_TRACE(jh, "generate b_committed data");
913 if (!committed_data) {
914 jbd_unlock_bh_state(bh);
915 goto repeat;
916 }
917
918 jh->b_committed_data = committed_data;
919 committed_data = NULL;
920 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
921 }
922 jbd_unlock_bh_state(bh);
923out:
924 journal_put_journal_head(jh);
Andrew Morton304c4c82006-06-23 02:05:31 -0700925 if (unlikely(committed_data))
Mingming Caoc089d492007-10-16 18:38:25 -0400926 jbd_free(committed_data, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 return err;
928}
929
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700930/**
Randy Dunlap78a4a502008-02-29 22:02:31 -0800931 * int journal_dirty_data() - mark a buffer as containing dirty data to be flushed
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 * @handle: transaction
933 * @bh: bufferhead to mark
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700934 *
Randy Dunlap78a4a502008-02-29 22:02:31 -0800935 * Description:
936 * Mark a buffer as containing dirty data which needs to be flushed before
937 * we can commit the current transaction.
938 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 * The buffer is placed on the transaction's data list and is marked as
940 * belonging to the transaction.
941 *
942 * Returns error number or 0 on success.
943 *
944 * journal_dirty_data() can be called via page_launder->ext3_writepage
945 * by kswapd.
946 */
947int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
948{
949 journal_t *journal = handle->h_transaction->t_journal;
950 int need_brelse = 0;
951 struct journal_head *jh;
Hidehiro Kawai960a22a2008-10-18 20:27:58 -0700952 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953
954 if (is_handle_aborted(handle))
Hidehiro Kawai960a22a2008-10-18 20:27:58 -0700955 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956
957 jh = journal_add_journal_head(bh);
958 JBUFFER_TRACE(jh, "entry");
959
960 /*
961 * The buffer could *already* be dirty. Writeout can start
962 * at any time.
963 */
964 jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
965
966 /*
967 * What if the buffer is already part of a running transaction?
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700968 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 * There are two cases:
970 * 1) It is part of the current running transaction. Refile it,
971 * just in case we have allocated it as metadata, deallocated
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700972 * it, then reallocated it as data.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 * 2) It is part of the previous, still-committing transaction.
974 * If all we want to do is to guarantee that the buffer will be
975 * written to disk before this new transaction commits, then
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700976 * being sure that the *previous* transaction has this same
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 * property is sufficient for us! Just leave it on its old
978 * transaction.
979 *
980 * In case (2), the buffer must not already exist as metadata
981 * --- that would violate write ordering (a transaction is free
982 * to write its data at any point, even before the previous
983 * committing transaction has committed). The caller must
984 * never, ever allow this to happen: there's nothing we can do
985 * about it in this layer.
986 */
987 jbd_lock_bh_state(bh);
988 spin_lock(&journal->j_list_lock);
Eric Sandeenf58a74d2006-10-28 10:38:27 -0700989
990 /* Now that we have bh_state locked, are we really still mapped? */
991 if (!buffer_mapped(bh)) {
992 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
993 goto no_journal;
994 }
995
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 if (jh->b_transaction) {
997 JBUFFER_TRACE(jh, "has transaction");
998 if (jh->b_transaction != handle->h_transaction) {
999 JBUFFER_TRACE(jh, "belongs to older transaction");
1000 J_ASSERT_JH(jh, jh->b_transaction ==
1001 journal->j_committing_transaction);
1002
1003 /* @@@ IS THIS TRUE ? */
1004 /*
1005 * Not any more. Scenario: someone does a write()
1006 * in data=journal mode. The buffer's transaction has
1007 * moved into commit. Then someone does another
1008 * write() to the file. We do the frozen data copyout
1009 * and set b_next_transaction to point to j_running_t.
1010 * And while we're in that state, someone does a
1011 * writepage() in an attempt to pageout the same area
1012 * of the file via a shared mapping. At present that
1013 * calls journal_dirty_data(), and we get right here.
1014 * It may be too late to journal the data. Simply
1015 * falling through to the next test will suffice: the
1016 * data will be dirty and wil be checkpointed. The
1017 * ordering comments in the next comment block still
1018 * apply.
1019 */
1020 //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1021
1022 /*
1023 * If we're journalling data, and this buffer was
1024 * subject to a write(), it could be metadata, forget
1025 * or shadow against the committing transaction. Now,
1026 * someone has dirtied the same darn page via a mapping
1027 * and it is being writepage()'d.
1028 * We *could* just steal the page from commit, with some
1029 * fancy locking there. Instead, we just skip it -
1030 * don't tie the page's buffers to the new transaction
1031 * at all.
1032 * Implication: if we crash before the writepage() data
1033 * is written into the filesystem, recovery will replay
1034 * the write() data.
1035 */
1036 if (jh->b_jlist != BJ_None &&
1037 jh->b_jlist != BJ_SyncData &&
1038 jh->b_jlist != BJ_Locked) {
1039 JBUFFER_TRACE(jh, "Not stealing");
1040 goto no_journal;
1041 }
1042
1043 /*
1044 * This buffer may be undergoing writeout in commit. We
1045 * can't return from here and let the caller dirty it
1046 * again because that can cause the write-out loop in
1047 * commit to never terminate.
1048 */
1049 if (buffer_dirty(bh)) {
1050 get_bh(bh);
1051 spin_unlock(&journal->j_list_lock);
1052 jbd_unlock_bh_state(bh);
1053 need_brelse = 1;
1054 sync_dirty_buffer(bh);
1055 jbd_lock_bh_state(bh);
1056 spin_lock(&journal->j_list_lock);
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001057 /* Since we dropped the lock... */
1058 if (!buffer_mapped(bh)) {
1059 JBUFFER_TRACE(jh, "buffer got unmapped");
1060 goto no_journal;
1061 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 /* The buffer may become locked again at any
1063 time if it is redirtied */
1064 }
1065
Hidehiro Kawai960a22a2008-10-18 20:27:58 -07001066 /*
1067 * We cannot remove the buffer with io error from the
1068 * committing transaction, because otherwise it would
1069 * miss the error and the commit would not abort.
1070 */
1071 if (unlikely(!buffer_uptodate(bh))) {
1072 ret = -EIO;
1073 goto no_journal;
1074 }
Jan Karabb189242011-06-24 23:11:59 +02001075 /* We might have slept so buffer could be refiled now */
1076 if (jh->b_transaction != NULL &&
1077 jh->b_transaction != handle->h_transaction) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078 JBUFFER_TRACE(jh, "unfile from commit");
1079 __journal_temp_unlink_buffer(jh);
1080 /* It still points to the committing
1081 * transaction; move it to this one so
1082 * that the refile assert checks are
1083 * happy. */
1084 jh->b_transaction = handle->h_transaction;
1085 }
1086 /* The buffer will be refiled below */
1087
1088 }
1089 /*
1090 * Special case --- the buffer might actually have been
1091 * allocated and then immediately deallocated in the previous,
1092 * committing transaction, so might still be left on that
1093 * transaction's metadata lists.
1094 */
1095 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1096 JBUFFER_TRACE(jh, "not on correct data list: unfile");
1097 J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 JBUFFER_TRACE(jh, "file as data");
1099 __journal_file_buffer(jh, handle->h_transaction,
1100 BJ_SyncData);
1101 }
1102 } else {
1103 JBUFFER_TRACE(jh, "not on a transaction");
1104 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1105 }
1106no_journal:
1107 spin_unlock(&journal->j_list_lock);
1108 jbd_unlock_bh_state(bh);
1109 if (need_brelse) {
1110 BUFFER_TRACE(bh, "brelse");
1111 __brelse(bh);
1112 }
1113 JBUFFER_TRACE(jh, "exit");
1114 journal_put_journal_head(jh);
Hidehiro Kawai960a22a2008-10-18 20:27:58 -07001115 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116}
1117
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001118/**
Randy Dunlap78a4a502008-02-29 22:02:31 -08001119 * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 * @handle: transaction to add buffer to.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001121 * @bh: buffer to mark
1122 *
Randy Dunlap78a4a502008-02-29 22:02:31 -08001123 * Mark dirty metadata which needs to be journaled as part of the current
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 * transaction.
1125 *
1126 * The buffer is placed on the transaction's metadata list and is marked
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001127 * as belonging to the transaction.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 *
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001129 * Returns error number or 0 on success.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 *
1131 * Special care needs to be taken if the buffer already belongs to the
1132 * current committing transaction (in which case we should have frozen
1133 * data present for that commit). In that case, we don't relink the
1134 * buffer: that only gets done when the old transaction finally
1135 * completes its commit.
1136 */
1137int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1138{
1139 transaction_t *transaction = handle->h_transaction;
1140 journal_t *journal = transaction->t_journal;
1141 struct journal_head *jh = bh2jh(bh);
1142
1143 jbd_debug(5, "journal_head %p\n", jh);
1144 JBUFFER_TRACE(jh, "entry");
1145 if (is_handle_aborted(handle))
1146 goto out;
1147
1148 jbd_lock_bh_state(bh);
1149
1150 if (jh->b_modified == 0) {
1151 /*
1152 * This buffer's got modified and becoming part
1153 * of the transaction. This needs to be done
1154 * once a transaction -bzzz
1155 */
1156 jh->b_modified = 1;
1157 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1158 handle->h_buffer_credits--;
1159 }
1160
1161 /*
1162 * fastpath, to avoid expensive locking. If this buffer is already
1163 * on the running transaction's metadata list there is nothing to do.
1164 * Nobody can take it off again because there is a handle open.
1165 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1166 * result in this test being false, so we go in and take the locks.
1167 */
1168 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1169 JBUFFER_TRACE(jh, "fastpath");
1170 J_ASSERT_JH(jh, jh->b_transaction ==
1171 journal->j_running_transaction);
1172 goto out_unlock_bh;
1173 }
1174
1175 set_buffer_jbddirty(bh);
1176
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001177 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 * Metadata already on the current transaction list doesn't
1179 * need to be filed. Metadata on another transaction's list must
1180 * be committing, and will be refiled once the commit completes:
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001181 * leave it alone for now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 */
1183 if (jh->b_transaction != transaction) {
1184 JBUFFER_TRACE(jh, "already on other transaction");
1185 J_ASSERT_JH(jh, jh->b_transaction ==
1186 journal->j_committing_transaction);
1187 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1188 /* And this case is illegal: we can't reuse another
1189 * transaction's data buffer, ever. */
1190 goto out_unlock_bh;
1191 }
1192
1193 /* That test should have eliminated the following case: */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001194 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195
1196 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1197 spin_lock(&journal->j_list_lock);
1198 __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1199 spin_unlock(&journal->j_list_lock);
1200out_unlock_bh:
1201 jbd_unlock_bh_state(bh);
1202out:
1203 JBUFFER_TRACE(jh, "exit");
1204 return 0;
1205}
1206
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001207/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 * journal_release_buffer: undo a get_write_access without any buffer
1209 * updates, if the update decided in the end that it didn't need access.
1210 *
1211 */
1212void
1213journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1214{
1215 BUFFER_TRACE(bh, "entry");
1216}
1217
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001218/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 * void journal_forget() - bforget() for potentially-journaled buffers.
1220 * @handle: transaction handle
1221 * @bh: bh to 'forget'
1222 *
1223 * We can only do the bforget if there are no commits pending against the
1224 * buffer. If the buffer is dirty in the current running transaction we
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001225 * can safely unlink it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 *
1227 * bh may not be a journalled buffer at all - it may be a non-JBD
1228 * buffer which came off the hashtable. Check for this.
1229 *
1230 * Decrements bh->b_count by one.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001231 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 * Allow this call even if the handle has aborted --- it may be part of
1233 * the caller's cleanup after an abort.
1234 */
1235int journal_forget (handle_t *handle, struct buffer_head *bh)
1236{
1237 transaction_t *transaction = handle->h_transaction;
1238 journal_t *journal = transaction->t_journal;
1239 struct journal_head *jh;
1240 int drop_reserve = 0;
1241 int err = 0;
Josef Bacik5b9a4992008-04-28 02:16:12 -07001242 int was_modified = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243
1244 BUFFER_TRACE(bh, "entry");
1245
1246 jbd_lock_bh_state(bh);
1247 spin_lock(&journal->j_list_lock);
1248
1249 if (!buffer_jbd(bh))
1250 goto not_jbd;
1251 jh = bh2jh(bh);
1252
1253 /* Critical error: attempting to delete a bitmap buffer, maybe?
1254 * Don't do any jbd operations, and return an error. */
1255 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1256 "inconsistent data on disk")) {
1257 err = -EIO;
1258 goto not_jbd;
1259 }
1260
Adam Buchbinder48fc7f72012-09-19 21:48:00 -04001261 /* keep track of whether or not this transaction modified us */
Josef Bacik5b9a4992008-04-28 02:16:12 -07001262 was_modified = jh->b_modified;
1263
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 /*
1265 * The buffer's going from the transaction, we must drop
1266 * all references -bzzz
1267 */
1268 jh->b_modified = 0;
1269
1270 if (jh->b_transaction == handle->h_transaction) {
1271 J_ASSERT_JH(jh, !jh->b_frozen_data);
1272
1273 /* If we are forgetting a buffer which is already part
1274 * of this transaction, then we can just drop it from
1275 * the transaction immediately. */
1276 clear_buffer_dirty(bh);
1277 clear_buffer_jbddirty(bh);
1278
1279 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1280
Josef Bacik5b9a4992008-04-28 02:16:12 -07001281 /*
1282 * we only want to drop a reference if this transaction
1283 * modified the buffer
1284 */
1285 if (was_modified)
1286 drop_reserve = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001288 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 * We are no longer going to journal this buffer.
1290 * However, the commit of this transaction is still
1291 * important to the buffer: the delete that we are now
1292 * processing might obsolete an old log entry, so by
1293 * committing, we can satisfy the buffer's checkpoint.
1294 *
1295 * So, if we have a checkpoint on the buffer, we should
1296 * now refile the buffer on our BJ_Forget list so that
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001297 * we know to remove the checkpoint after we commit.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298 */
1299
1300 if (jh->b_cp_transaction) {
1301 __journal_temp_unlink_buffer(jh);
1302 __journal_file_buffer(jh, transaction, BJ_Forget);
1303 } else {
1304 __journal_unfile_buffer(jh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 if (!buffer_jbd(bh)) {
1306 spin_unlock(&journal->j_list_lock);
1307 jbd_unlock_bh_state(bh);
1308 __bforget(bh);
1309 goto drop;
1310 }
1311 }
1312 } else if (jh->b_transaction) {
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001313 J_ASSERT_JH(jh, (jh->b_transaction ==
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 journal->j_committing_transaction));
1315 /* However, if the buffer is still owned by a prior
1316 * (committing) transaction, we can't drop it yet... */
1317 JBUFFER_TRACE(jh, "belongs to older transaction");
1318 /* ... but we CAN drop it from the new transaction if we
1319 * have also modified it since the original commit. */
1320
1321 if (jh->b_next_transaction) {
1322 J_ASSERT(jh->b_next_transaction == transaction);
1323 jh->b_next_transaction = NULL;
Josef Bacik5b9a4992008-04-28 02:16:12 -07001324
1325 /*
1326 * only drop a reference if this transaction modified
1327 * the buffer
1328 */
1329 if (was_modified)
1330 drop_reserve = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 }
1332 }
1333
1334not_jbd:
1335 spin_unlock(&journal->j_list_lock);
1336 jbd_unlock_bh_state(bh);
1337 __brelse(bh);
1338drop:
1339 if (drop_reserve) {
1340 /* no need to reserve log space for this block -bzzz */
1341 handle->h_buffer_credits++;
1342 }
1343 return err;
1344}
1345
1346/**
1347 * int journal_stop() - complete a transaction
1348 * @handle: tranaction to complete.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001349 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 * All done for a particular handle.
1351 *
1352 * There is not much action needed here. We just return any remaining
1353 * buffer credits to the transaction and remove the handle. The only
1354 * complication is that we need to start a commit operation if the
1355 * filesystem is marked for synchronous update.
1356 *
1357 * journal_stop itself will not usually return an error, but it may
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001358 * do so in unusual circumstances. In particular, expect it to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 * return -EIO if a journal_abort has been executed since the
1360 * transaction began.
1361 */
1362int journal_stop(handle_t *handle)
1363{
1364 transaction_t *transaction = handle->h_transaction;
1365 journal_t *journal = transaction->t_journal;
Josef Bacikf420d4d2009-01-07 18:07:24 -08001366 int err;
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001367 pid_t pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 J_ASSERT(journal_current_handle() == handle);
1370
1371 if (is_handle_aborted(handle))
1372 err = -EIO;
OGAWA Hirofumi3e2a5322006-10-19 23:29:11 -07001373 else {
1374 J_ASSERT(transaction->t_updates > 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 err = 0;
OGAWA Hirofumi3e2a5322006-10-19 23:29:11 -07001376 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377
1378 if (--handle->h_ref > 0) {
1379 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1380 handle->h_ref);
1381 return err;
1382 }
1383
1384 jbd_debug(4, "Handle %p going down\n", handle);
1385
1386 /*
1387 * Implement synchronous transaction batching. If the handle
1388 * was synchronous, don't force a commit immediately. Let's
1389 * yield and let another thread piggyback onto this transaction.
1390 * Keep doing that while new threads continue to arrive.
1391 * It doesn't cost much - we're about to run a commit and sleep
1392 * on IO anyway. Speeds up many-threaded, many-dir operations
1393 * by 30x or more...
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001394 *
Josef Bacikf420d4d2009-01-07 18:07:24 -08001395 * We try and optimize the sleep time against what the underlying disk
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001396 * can do, instead of having a static sleep time. This is useful for
Josef Bacikf420d4d2009-01-07 18:07:24 -08001397 * the case where our storage is so fast that it is more optimal to go
1398 * ahead and force a flush and wait for the transaction to be committed
1399 * than it is to wait for an arbitrary amount of time for new writers to
Daniel Mack3ad2f3f2010-02-03 08:01:28 +08001400 * join the transaction. We achieve this by measuring how long it takes
Josef Bacikf420d4d2009-01-07 18:07:24 -08001401 * to commit a transaction, and compare it with how long this
1402 * transaction has been running, and if run time < commit time then we
1403 * sleep for the delta and commit. This greatly helps super fast disks
1404 * that would see slowdowns as more threads started doing fsyncs.
1405 *
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001406 * But don't do this if this process was the most recent one to
1407 * perform a synchronous write. We do this to detect the case where a
1408 * single process is doing a stream of sync writes. No point in waiting
1409 * for joiners in that case.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 */
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001411 pid = current->pid;
1412 if (handle->h_sync && journal->j_last_sync_writer != pid) {
Josef Bacikf420d4d2009-01-07 18:07:24 -08001413 u64 commit_time, trans_time;
1414
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001415 journal->j_last_sync_writer = pid;
Josef Bacikf420d4d2009-01-07 18:07:24 -08001416
1417 spin_lock(&journal->j_state_lock);
1418 commit_time = journal->j_average_commit_time;
1419 spin_unlock(&journal->j_state_lock);
1420
1421 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1422 transaction->t_start_time));
1423
1424 commit_time = min_t(u64, commit_time,
1425 1000*jiffies_to_usecs(1));
1426
1427 if (trans_time < commit_time) {
1428 ktime_t expires = ktime_add_ns(ktime_get(),
1429 commit_time);
1430 set_current_state(TASK_UNINTERRUPTIBLE);
1431 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1432 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 }
1434
1435 current->journal_info = NULL;
1436 spin_lock(&journal->j_state_lock);
1437 spin_lock(&transaction->t_handle_lock);
1438 transaction->t_outstanding_credits -= handle->h_buffer_credits;
1439 transaction->t_updates--;
1440 if (!transaction->t_updates) {
1441 wake_up(&journal->j_wait_updates);
1442 if (journal->j_barrier_count)
1443 wake_up(&journal->j_wait_transaction_locked);
1444 }
1445
1446 /*
1447 * If the handle is marked SYNC, we need to set another commit
1448 * going! We also want to force a commit if the current
1449 * transaction is occupying too much of the log, or if the
1450 * transaction is too old now.
1451 */
1452 if (handle->h_sync ||
1453 transaction->t_outstanding_credits >
1454 journal->j_max_transaction_buffers ||
Dave Kleikampe9ad5622006-09-27 01:49:35 -07001455 time_after_eq(jiffies, transaction->t_expires)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 /* Do this even for aborted journals: an abort still
1457 * completes the commit thread, it just doesn't write
1458 * anything to disk. */
1459 tid_t tid = transaction->t_tid;
1460
1461 spin_unlock(&transaction->t_handle_lock);
1462 jbd_debug(2, "transaction too old, requesting commit for "
1463 "handle %p\n", handle);
1464 /* This is non-blocking */
1465 __log_start_commit(journal, transaction->t_tid);
1466 spin_unlock(&journal->j_state_lock);
1467
1468 /*
1469 * Special case: JFS_SYNC synchronous updates require us
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001470 * to wait for the commit to complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 */
1472 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1473 err = log_wait_commit(journal, tid);
1474 } else {
1475 spin_unlock(&transaction->t_handle_lock);
1476 spin_unlock(&journal->j_state_lock);
1477 }
1478
Ingo Molnar3295f0e2008-08-11 10:30:30 +02001479 lock_map_release(&handle->h_lockdep_map);
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +02001480
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 jbd_free_handle(handle);
1482 return err;
1483}
1484
Randy Dunlap0cf01f62008-03-19 17:00:44 -07001485/**
1486 * int journal_force_commit() - force any uncommitted transactions
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 * @journal: journal to force
1488 *
1489 * For synchronous operations: force any uncommitted transactions
1490 * to disk. May seem kludgy, but it reuses all the handle batching
1491 * code in a very simple manner.
1492 */
1493int journal_force_commit(journal_t *journal)
1494{
1495 handle_t *handle;
1496 int ret;
1497
1498 handle = journal_start(journal, 1);
1499 if (IS_ERR(handle)) {
1500 ret = PTR_ERR(handle);
1501 } else {
1502 handle->h_sync = 1;
1503 ret = journal_stop(handle);
1504 }
1505 return ret;
1506}
1507
1508/*
1509 *
1510 * List management code snippets: various functions for manipulating the
1511 * transaction buffer lists.
1512 *
1513 */
1514
1515/*
1516 * Append a buffer to a transaction list, given the transaction's list head
1517 * pointer.
1518 *
1519 * j_list_lock is held.
1520 *
1521 * jbd_lock_bh_state(jh2bh(jh)) is held.
1522 */
1523
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001524static inline void
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525__blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1526{
1527 if (!*list) {
1528 jh->b_tnext = jh->b_tprev = jh;
1529 *list = jh;
1530 } else {
1531 /* Insert at the tail of the list to preserve order */
1532 struct journal_head *first = *list, *last = first->b_tprev;
1533 jh->b_tprev = last;
1534 jh->b_tnext = first;
1535 last->b_tnext = first->b_tprev = jh;
1536 }
1537}
1538
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001539/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 * Remove a buffer from a transaction list, given the transaction's list
1541 * head pointer.
1542 *
1543 * Called with j_list_lock held, and the journal may not be locked.
1544 *
1545 * jbd_lock_bh_state(jh2bh(jh)) is held.
1546 */
1547
1548static inline void
1549__blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1550{
1551 if (*list == jh) {
1552 *list = jh->b_tnext;
1553 if (*list == jh)
1554 *list = NULL;
1555 }
1556 jh->b_tprev->b_tnext = jh->b_tnext;
1557 jh->b_tnext->b_tprev = jh->b_tprev;
1558}
1559
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001560/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 * Remove a buffer from the appropriate transaction list.
1562 *
1563 * Note that this function can *change* the value of
1564 * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1565 * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list. If the caller
1566 * is holding onto a copy of one of thee pointers, it could go bad.
1567 * Generally the caller needs to re-read the pointer from the transaction_t.
1568 *
1569 * Called under j_list_lock. The journal may not be locked.
1570 */
Adrian Bunkd394e122006-12-06 20:38:26 -08001571static void __journal_temp_unlink_buffer(struct journal_head *jh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572{
1573 struct journal_head **list = NULL;
1574 transaction_t *transaction;
1575 struct buffer_head *bh = jh2bh(jh);
1576
1577 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1578 transaction = jh->b_transaction;
1579 if (transaction)
1580 assert_spin_locked(&transaction->t_journal->j_list_lock);
1581
1582 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1583 if (jh->b_jlist != BJ_None)
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001584 J_ASSERT_JH(jh, transaction != NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585
1586 switch (jh->b_jlist) {
1587 case BJ_None:
1588 return;
1589 case BJ_SyncData:
1590 list = &transaction->t_sync_datalist;
1591 break;
1592 case BJ_Metadata:
1593 transaction->t_nr_buffers--;
1594 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1595 list = &transaction->t_buffers;
1596 break;
1597 case BJ_Forget:
1598 list = &transaction->t_forget;
1599 break;
1600 case BJ_IO:
1601 list = &transaction->t_iobuf_list;
1602 break;
1603 case BJ_Shadow:
1604 list = &transaction->t_shadow_list;
1605 break;
1606 case BJ_LogCtl:
1607 list = &transaction->t_log_list;
1608 break;
1609 case BJ_Reserved:
1610 list = &transaction->t_reserved_list;
1611 break;
1612 case BJ_Locked:
1613 list = &transaction->t_locked_list;
1614 break;
1615 }
1616
1617 __blist_del_buffer(list, jh);
1618 jh->b_jlist = BJ_None;
1619 if (test_clear_buffer_jbddirty(bh))
1620 mark_buffer_dirty(bh); /* Expose it to the VM */
1621}
1622
Jan Karabb189242011-06-24 23:11:59 +02001623/*
1624 * Remove buffer from all transactions.
1625 *
1626 * Called with bh_state lock and j_list_lock
1627 *
1628 * jh and bh may be already freed when this function returns.
1629 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630void __journal_unfile_buffer(struct journal_head *jh)
1631{
1632 __journal_temp_unlink_buffer(jh);
1633 jh->b_transaction = NULL;
Jan Karabb189242011-06-24 23:11:59 +02001634 journal_put_journal_head(jh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635}
1636
1637void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1638{
Jan Karabb189242011-06-24 23:11:59 +02001639 struct buffer_head *bh = jh2bh(jh);
1640
1641 /* Get reference so that buffer cannot be freed before we unlock it */
1642 get_bh(bh);
1643 jbd_lock_bh_state(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 spin_lock(&journal->j_list_lock);
1645 __journal_unfile_buffer(jh);
1646 spin_unlock(&journal->j_list_lock);
Jan Karabb189242011-06-24 23:11:59 +02001647 jbd_unlock_bh_state(bh);
1648 __brelse(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649}
1650
1651/*
1652 * Called from journal_try_to_free_buffers().
1653 *
1654 * Called under jbd_lock_bh_state(bh)
1655 */
1656static void
1657__journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1658{
1659 struct journal_head *jh;
1660
1661 jh = bh2jh(bh);
1662
1663 if (buffer_locked(bh) || buffer_dirty(bh))
1664 goto out;
1665
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001666 if (jh->b_next_transaction != NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 goto out;
1668
1669 spin_lock(&journal->j_list_lock);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001670 if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1672 /* A written-back ordered data buffer */
1673 JBUFFER_TRACE(jh, "release data");
1674 __journal_unfile_buffer(jh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675 }
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001676 } else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 /* written-back checkpointed metadata buffer */
1678 if (jh->b_jlist == BJ_None) {
1679 JBUFFER_TRACE(jh, "remove from checkpoint list");
1680 __journal_remove_checkpoint(jh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 }
1682 }
1683 spin_unlock(&journal->j_list_lock);
1684out:
1685 return;
1686}
1687
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001688/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 * int journal_try_to_free_buffers() - try to free page buffers.
1690 * @journal: journal for operation
1691 * @page: to try and free
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001692 * @gfp_mask: we use the mask to detect how hard should we try to release
1693 * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1694 * release the buffers.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 *
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001696 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 * For all the buffers on this page,
1698 * if they are fully written out ordered data, move them onto BUF_CLEAN
1699 * so try_to_free_buffers() can reap them.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001700 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 * This function returns non-zero if we wish try_to_free_buffers()
1702 * to be called. We do this if the page is releasable by try_to_free_buffers().
1703 * We also do it if the page has locked or dirty buffers and the caller wants
1704 * us to perform sync or async writeout.
1705 *
1706 * This complicates JBD locking somewhat. We aren't protected by the
1707 * BKL here. We wish to remove the buffer from its committing or
1708 * running transaction's ->t_datalist via __journal_unfile_buffer.
1709 *
1710 * This may *change* the value of transaction_t->t_datalist, so anyone
1711 * who looks at t_datalist needs to lock against this function.
1712 *
1713 * Even worse, someone may be doing a journal_dirty_data on this
1714 * buffer. So we need to lock against that. journal_dirty_data()
1715 * will come out of the lock with the buffer dirty, which makes it
1716 * ineligible for release here.
1717 *
1718 * Who else is affected by this? hmm... Really the only contender
1719 * is do_get_write_access() - it could be looking at the buffer while
1720 * journal_try_to_free_buffer() is changing its state. But that
1721 * cannot happen because we never reallocate freed data as metadata
1722 * while the data is part of a transaction. Yes?
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001723 *
1724 * Return 0 on failure, 1 on success
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001726int journal_try_to_free_buffers(journal_t *journal,
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001727 struct page *page, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728{
1729 struct buffer_head *head;
1730 struct buffer_head *bh;
1731 int ret = 0;
1732
1733 J_ASSERT(PageLocked(page));
1734
1735 head = page_buffers(page);
1736 bh = head;
1737 do {
1738 struct journal_head *jh;
1739
1740 /*
1741 * We take our own ref against the journal_head here to avoid
1742 * having to add tons of locking around each instance of
Jan Karabb189242011-06-24 23:11:59 +02001743 * journal_put_journal_head().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 */
1745 jh = journal_grab_journal_head(bh);
1746 if (!jh)
1747 continue;
1748
1749 jbd_lock_bh_state(bh);
1750 __journal_try_to_free_buffer(journal, bh);
1751 journal_put_journal_head(jh);
1752 jbd_unlock_bh_state(bh);
1753 if (buffer_jbd(bh))
1754 goto busy;
1755 } while ((bh = bh->b_this_page) != head);
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001756
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 ret = try_to_free_buffers(page);
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001758
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759busy:
1760 return ret;
1761}
1762
1763/*
1764 * This buffer is no longer needed. If it is on an older transaction's
1765 * checkpoint list we need to record it on this transaction's forget list
1766 * to pin this buffer (and hence its checkpointing transaction) down until
1767 * this transaction commits. If the buffer isn't on a checkpoint list, we
1768 * release it.
1769 * Returns non-zero if JBD no longer has an interest in the buffer.
1770 *
1771 * Called under j_list_lock.
1772 *
1773 * Called under jbd_lock_bh_state(bh).
1774 */
1775static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1776{
1777 int may_free = 1;
1778 struct buffer_head *bh = jh2bh(jh);
1779
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 if (jh->b_cp_transaction) {
1781 JBUFFER_TRACE(jh, "on running+cp transaction");
Jan Karabb189242011-06-24 23:11:59 +02001782 __journal_temp_unlink_buffer(jh);
Jan Kara1e9fd532009-06-24 17:31:40 +02001783 /*
1784 * We don't want to write the buffer anymore, clear the
1785 * bit so that we don't confuse checks in
1786 * __journal_file_buffer
1787 */
1788 clear_buffer_dirty(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 __journal_file_buffer(jh, transaction, BJ_Forget);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 may_free = 0;
1791 } else {
1792 JBUFFER_TRACE(jh, "on running transaction");
Jan Karabb189242011-06-24 23:11:59 +02001793 __journal_unfile_buffer(jh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 }
1795 return may_free;
1796}
1797
1798/*
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001799 * journal_invalidatepage
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 *
1801 * This code is tricky. It has a number of cases to deal with.
1802 *
1803 * There are two invariants which this code relies on:
1804 *
1805 * i_size must be updated on disk before we start calling invalidatepage on the
1806 * data.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001807 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808 * This is done in ext3 by defining an ext3_setattr method which
1809 * updates i_size before truncate gets going. By maintaining this
1810 * invariant, we can be sure that it is safe to throw away any buffers
1811 * attached to the current transaction: once the transaction commits,
1812 * we know that the data will not be needed.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001813 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814 * Note however that we can *not* throw away data belonging to the
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001815 * previous, committing transaction!
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 *
1817 * Any disk blocks which *are* part of the previous, committing
1818 * transaction (and which therefore cannot be discarded immediately) are
1819 * not going to be reused in the new running transaction
1820 *
1821 * The bitmap committed_data images guarantee this: any block which is
1822 * allocated in one transaction and removed in the next will be marked
1823 * as in-use in the committed_data bitmap, so cannot be reused until
1824 * the next transaction to delete the block commits. This means that
1825 * leaving committing buffers dirty is quite safe: the disk blocks
1826 * cannot be reallocated to a different file and so buffer aliasing is
1827 * not possible.
1828 *
1829 *
1830 * The above applies mainly to ordered data mode. In writeback mode we
1831 * don't make guarantees about the order in which data hits disk --- in
1832 * particular we don't guarantee that new dirty data is flushed before
1833 * transaction commit --- so it is always safe just to discard data
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001834 * immediately in that mode. --sct
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835 */
1836
1837/*
1838 * The journal_unmap_buffer helper function returns zero if the buffer
1839 * concerned remains pinned as an anonymous buffer belonging to an older
1840 * transaction.
1841 *
1842 * We're outside-transaction here. Either or both of j_running_transaction
1843 * and j_committing_transaction may be NULL.
1844 */
Jan Kara09e05d42012-07-11 23:16:25 +02001845static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
1846 int partial_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847{
1848 transaction_t *transaction;
1849 struct journal_head *jh;
1850 int may_free = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851
1852 BUFFER_TRACE(bh, "entry");
1853
Jan Kara09e05d42012-07-11 23:16:25 +02001854retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 /*
1856 * It is safe to proceed here without the j_list_lock because the
1857 * buffers cannot be stolen by try_to_free_buffers as long as we are
1858 * holding the page lock. --sct
1859 */
1860
1861 if (!buffer_jbd(bh))
1862 goto zap_buffer_unlocked;
1863
1864 spin_lock(&journal->j_state_lock);
1865 jbd_lock_bh_state(bh);
1866 spin_lock(&journal->j_list_lock);
1867
1868 jh = journal_grab_journal_head(bh);
1869 if (!jh)
1870 goto zap_buffer_no_jh;
1871
Jan Kara86963912010-02-16 20:37:12 +01001872 /*
1873 * We cannot remove the buffer from checkpoint lists until the
1874 * transaction adding inode to orphan list (let's call it T)
1875 * is committed. Otherwise if the transaction changing the
1876 * buffer would be cleaned from the journal before T is
1877 * committed, a crash will cause that the correct contents of
1878 * the buffer will be lost. On the other hand we have to
1879 * clear the buffer dirty bit at latest at the moment when the
1880 * transaction marking the buffer as freed in the filesystem
1881 * structures is committed because from that moment on the
Jan Kara09e05d42012-07-11 23:16:25 +02001882 * block can be reallocated and used by a different page.
Jan Kara86963912010-02-16 20:37:12 +01001883 * Since the block hasn't been freed yet but the inode has
1884 * already been added to orphan list, it is safe for us to add
1885 * the buffer to BJ_Forget list of the newest transaction.
Jan Kara09e05d42012-07-11 23:16:25 +02001886 *
1887 * Also we have to clear buffer_mapped flag of a truncated buffer
1888 * because the buffer_head may be attached to the page straddling
1889 * i_size (can happen only when blocksize < pagesize) and thus the
1890 * buffer_head can be reused when the file is extended again. So we end
1891 * up keeping around invalidated buffers attached to transactions'
1892 * BJ_Forget list just to stop checkpointing code from cleaning up
1893 * the transaction this buffer was modified in.
Jan Kara86963912010-02-16 20:37:12 +01001894 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 transaction = jh->b_transaction;
1896 if (transaction == NULL) {
1897 /* First case: not on any transaction. If it
1898 * has no checkpoint link, then we can zap it:
1899 * it's a writeback-mode buffer so we don't care
1900 * if it hits disk safely. */
1901 if (!jh->b_cp_transaction) {
1902 JBUFFER_TRACE(jh, "not on any transaction: zap");
1903 goto zap_buffer;
1904 }
1905
1906 if (!buffer_dirty(bh)) {
1907 /* bdflush has written it. We can drop it now */
1908 goto zap_buffer;
1909 }
1910
1911 /* OK, it must be in the journal but still not
1912 * written fully to disk: it's metadata or
1913 * journaled data... */
1914
1915 if (journal->j_running_transaction) {
1916 /* ... and once the current transaction has
1917 * committed, the buffer won't be needed any
1918 * longer. */
1919 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
Jan Kara09e05d42012-07-11 23:16:25 +02001920 may_free = __dispose_buffer(jh,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921 journal->j_running_transaction);
Jan Kara09e05d42012-07-11 23:16:25 +02001922 goto zap_buffer;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 } else {
1924 /* There is no currently-running transaction. So the
1925 * orphan record which we wrote for this file must have
1926 * passed into commit. We must attach this buffer to
1927 * the committing transaction, if it exists. */
1928 if (journal->j_committing_transaction) {
1929 JBUFFER_TRACE(jh, "give to committing trans");
Jan Kara09e05d42012-07-11 23:16:25 +02001930 may_free = __dispose_buffer(jh,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931 journal->j_committing_transaction);
Jan Kara09e05d42012-07-11 23:16:25 +02001932 goto zap_buffer;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 } else {
1934 /* The orphan record's transaction has
1935 * committed. We can cleanse this buffer */
1936 clear_buffer_jbddirty(bh);
1937 goto zap_buffer;
1938 }
1939 }
1940 } else if (transaction == journal->j_committing_transaction) {
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001941 JBUFFER_TRACE(jh, "on committing transaction");
akpm@osdl.orgd13df842005-04-16 15:26:36 -07001942 if (jh->b_jlist == BJ_Locked) {
1943 /*
1944 * The buffer is on the committing transaction's locked
1945 * list. We have the buffer locked, so I/O has
1946 * completed. So we can nail the buffer now.
1947 */
1948 may_free = __dispose_buffer(jh, transaction);
1949 goto zap_buffer;
1950 }
1951 /*
Jan Kara86963912010-02-16 20:37:12 +01001952 * The buffer is committing, we simply cannot touch
Jan Kara09e05d42012-07-11 23:16:25 +02001953 * it. If the page is straddling i_size we have to wait
1954 * for commit and try again.
1955 */
1956 if (partial_page) {
1957 tid_t tid = journal->j_committing_transaction->t_tid;
1958
1959 journal_put_journal_head(jh);
1960 spin_unlock(&journal->j_list_lock);
1961 jbd_unlock_bh_state(bh);
1962 spin_unlock(&journal->j_state_lock);
Jan Kara25389bb2012-11-23 14:03:04 +01001963 unlock_buffer(bh);
Jan Kara09e05d42012-07-11 23:16:25 +02001964 log_wait_commit(journal, tid);
Jan Kara25389bb2012-11-23 14:03:04 +01001965 lock_buffer(bh);
Jan Kara09e05d42012-07-11 23:16:25 +02001966 goto retry;
1967 }
1968 /*
1969 * OK, buffer won't be reachable after truncate. We just set
1970 * j_next_transaction to the running transaction (if there is
1971 * one) and mark buffer as freed so that commit code knows it
1972 * should clear dirty bits when it is done with the buffer.
Jan Kara86963912010-02-16 20:37:12 +01001973 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974 set_buffer_freed(bh);
Jan Kara86963912010-02-16 20:37:12 +01001975 if (journal->j_running_transaction && buffer_jbddirty(bh))
1976 jh->b_next_transaction = journal->j_running_transaction;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 journal_put_journal_head(jh);
1978 spin_unlock(&journal->j_list_lock);
1979 jbd_unlock_bh_state(bh);
1980 spin_unlock(&journal->j_state_lock);
1981 return 0;
1982 } else {
1983 /* Good, the buffer belongs to the running transaction.
1984 * We are writing our own transaction's data, not any
1985 * previous one's, so it is safe to throw it away
1986 * (remember that we expect the filesystem to have set
1987 * i_size already for this truncate so recovery will not
1988 * expose the disk blocks we are discarding here.) */
1989 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001990 JBUFFER_TRACE(jh, "on running transaction");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 may_free = __dispose_buffer(jh, transaction);
1992 }
1993
1994zap_buffer:
Jan Kara09e05d42012-07-11 23:16:25 +02001995 /*
1996 * This is tricky. Although the buffer is truncated, it may be reused
1997 * if blocksize < pagesize and it is attached to the page straddling
1998 * EOF. Since the buffer might have been added to BJ_Forget list of the
1999 * running transaction, journal_get_write_access() won't clear
2000 * b_modified and credit accounting gets confused. So clear b_modified
2001 * here. */
2002 jh->b_modified = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 journal_put_journal_head(jh);
2004zap_buffer_no_jh:
2005 spin_unlock(&journal->j_list_lock);
2006 jbd_unlock_bh_state(bh);
2007 spin_unlock(&journal->j_state_lock);
2008zap_buffer_unlocked:
2009 clear_buffer_dirty(bh);
2010 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2011 clear_buffer_mapped(bh);
2012 clear_buffer_req(bh);
2013 clear_buffer_new(bh);
2014 bh->b_bdev = NULL;
2015 return may_free;
2016}
2017
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002018/**
Randy Dunlapa6b91912008-03-19 17:01:00 -07002019 * void journal_invalidatepage() - invalidate a journal page
2020 * @journal: journal to use for flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021 * @page: page to flush
Lukas Czernerd8c89002013-05-21 23:26:36 -04002022 * @offset: offset of the range to invalidate
2023 * @length: length of the range to invalidate
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024 *
Lukas Czernerd8c89002013-05-21 23:26:36 -04002025 * Reap page buffers containing data in specified range in page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 */
NeilBrown2ff28e22006-03-26 01:37:18 -08002027void journal_invalidatepage(journal_t *journal,
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002028 struct page *page,
Lukas Czernerd8c89002013-05-21 23:26:36 -04002029 unsigned int offset,
2030 unsigned int length)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031{
2032 struct buffer_head *head, *bh, *next;
Lukas Czernerd8c89002013-05-21 23:26:36 -04002033 unsigned int stop = offset + length;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 unsigned int curr_off = 0;
Lukas Czernerd8c89002013-05-21 23:26:36 -04002035 int partial_page = (offset || length < PAGE_CACHE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 int may_free = 1;
2037
2038 if (!PageLocked(page))
2039 BUG();
2040 if (!page_has_buffers(page))
NeilBrown2ff28e22006-03-26 01:37:18 -08002041 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042
Lukas Czernerd8c89002013-05-21 23:26:36 -04002043 BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
2044
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 /* We will potentially be playing with lists other than just the
2046 * data lists (especially for journaled data mode), so be
2047 * cautious in our locking. */
2048
2049 head = bh = page_buffers(page);
2050 do {
2051 unsigned int next_off = curr_off + bh->b_size;
2052 next = bh->b_this_page;
2053
Lukas Czernerd8c89002013-05-21 23:26:36 -04002054 if (next_off > stop)
2055 return;
2056
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057 if (offset <= curr_off) {
Dave Kleikampe9ad5622006-09-27 01:49:35 -07002058 /* This block is wholly outside the truncation point */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 lock_buffer(bh);
Jan Kara09e05d42012-07-11 23:16:25 +02002060 may_free &= journal_unmap_buffer(journal, bh,
Lukas Czernerd8c89002013-05-21 23:26:36 -04002061 partial_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062 unlock_buffer(bh);
2063 }
2064 curr_off = next_off;
2065 bh = next;
2066
2067 } while (bh != head);
2068
Lukas Czernerd8c89002013-05-21 23:26:36 -04002069 if (!partial_page) {
NeilBrown2ff28e22006-03-26 01:37:18 -08002070 if (may_free && try_to_free_buffers(page))
2071 J_ASSERT(!page_has_buffers(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073}
2074
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002075/*
2076 * File a buffer on the given transaction list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 */
2078void __journal_file_buffer(struct journal_head *jh,
2079 transaction_t *transaction, int jlist)
2080{
2081 struct journal_head **list = NULL;
2082 int was_dirty = 0;
2083 struct buffer_head *bh = jh2bh(jh);
2084
2085 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2086 assert_spin_locked(&transaction->t_journal->j_list_lock);
2087
2088 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2089 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07002090 jh->b_transaction == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091
2092 if (jh->b_transaction && jh->b_jlist == jlist)
2093 return;
2094
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002095 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 jlist == BJ_Shadow || jlist == BJ_Forget) {
Jan Kara1e9fd532009-06-24 17:31:40 +02002097 /*
2098 * For metadata buffers, we track dirty bit in buffer_jbddirty
2099 * instead of buffer_dirty. We should not see a dirty bit set
2100 * here because we clear it in do_get_write_access but e.g.
2101 * tune2fs can modify the sb and set the dirty bit at any time
2102 * so we try to gracefully handle that.
2103 */
2104 if (buffer_dirty(bh))
2105 warn_dirty_buffer(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 if (test_clear_buffer_dirty(bh) ||
2107 test_clear_buffer_jbddirty(bh))
2108 was_dirty = 1;
2109 }
2110
2111 if (jh->b_transaction)
2112 __journal_temp_unlink_buffer(jh);
Jan Karabb189242011-06-24 23:11:59 +02002113 else
2114 journal_grab_journal_head(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 jh->b_transaction = transaction;
2116
2117 switch (jlist) {
2118 case BJ_None:
2119 J_ASSERT_JH(jh, !jh->b_committed_data);
2120 J_ASSERT_JH(jh, !jh->b_frozen_data);
2121 return;
2122 case BJ_SyncData:
2123 list = &transaction->t_sync_datalist;
2124 break;
2125 case BJ_Metadata:
2126 transaction->t_nr_buffers++;
2127 list = &transaction->t_buffers;
2128 break;
2129 case BJ_Forget:
2130 list = &transaction->t_forget;
2131 break;
2132 case BJ_IO:
2133 list = &transaction->t_iobuf_list;
2134 break;
2135 case BJ_Shadow:
2136 list = &transaction->t_shadow_list;
2137 break;
2138 case BJ_LogCtl:
2139 list = &transaction->t_log_list;
2140 break;
2141 case BJ_Reserved:
2142 list = &transaction->t_reserved_list;
2143 break;
2144 case BJ_Locked:
2145 list = &transaction->t_locked_list;
2146 break;
2147 }
2148
2149 __blist_add_buffer(list, jh);
2150 jh->b_jlist = jlist;
2151
2152 if (was_dirty)
2153 set_buffer_jbddirty(bh);
2154}
2155
2156void journal_file_buffer(struct journal_head *jh,
2157 transaction_t *transaction, int jlist)
2158{
2159 jbd_lock_bh_state(jh2bh(jh));
2160 spin_lock(&transaction->t_journal->j_list_lock);
2161 __journal_file_buffer(jh, transaction, jlist);
2162 spin_unlock(&transaction->t_journal->j_list_lock);
2163 jbd_unlock_bh_state(jh2bh(jh));
2164}
2165
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002166/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 * Remove a buffer from its current buffer list in preparation for
2168 * dropping it from its current transaction entirely. If the buffer has
2169 * already started to be used by a subsequent transaction, refile the
2170 * buffer on that transaction's metadata list.
2171 *
Jan Karabb189242011-06-24 23:11:59 +02002172 * Called under j_list_lock
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173 * Called under jbd_lock_bh_state(jh2bh(jh))
Jan Karabb189242011-06-24 23:11:59 +02002174 *
2175 * jh and bh may be already free when this function returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 */
2177void __journal_refile_buffer(struct journal_head *jh)
2178{
Jan Kara86963912010-02-16 20:37:12 +01002179 int was_dirty, jlist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 struct buffer_head *bh = jh2bh(jh);
2181
2182 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2183 if (jh->b_transaction)
2184 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2185
2186 /* If the buffer is now unused, just drop it. */
2187 if (jh->b_next_transaction == NULL) {
2188 __journal_unfile_buffer(jh);
2189 return;
2190 }
2191
2192 /*
2193 * It has been modified by a later transaction: add it to the new
2194 * transaction's metadata list.
2195 */
2196
2197 was_dirty = test_clear_buffer_jbddirty(bh);
2198 __journal_temp_unlink_buffer(jh);
Jan Karabb189242011-06-24 23:11:59 +02002199 /*
2200 * We set b_transaction here because b_next_transaction will inherit
2201 * our jh reference and thus __journal_file_buffer() must not take a
2202 * new one.
2203 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 jh->b_transaction = jh->b_next_transaction;
2205 jh->b_next_transaction = NULL;
Jan Kara86963912010-02-16 20:37:12 +01002206 if (buffer_freed(bh))
2207 jlist = BJ_Forget;
2208 else if (jh->b_modified)
2209 jlist = BJ_Metadata;
2210 else
2211 jlist = BJ_Reserved;
2212 __journal_file_buffer(jh, jh->b_transaction, jlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2214
2215 if (was_dirty)
2216 set_buffer_jbddirty(bh);
2217}
2218
2219/*
Jan Karabb189242011-06-24 23:11:59 +02002220 * __journal_refile_buffer() with necessary locking added. We take our bh
2221 * reference so that we can safely unlock bh.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222 *
Jan Karabb189242011-06-24 23:11:59 +02002223 * The jh and bh may be freed by this call.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 */
2225void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2226{
2227 struct buffer_head *bh = jh2bh(jh);
2228
Jan Karabb189242011-06-24 23:11:59 +02002229 /* Get reference so that buffer cannot be freed before we unlock it */
2230 get_bh(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 jbd_lock_bh_state(bh);
2232 spin_lock(&journal->j_list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 __journal_refile_buffer(jh);
2234 jbd_unlock_bh_state(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 spin_unlock(&journal->j_list_lock);
2236 __brelse(bh);
2237}