blob: 5b2e4c30a2a18b81626ee90f31c6fb6f81940acf [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) {
Mingming Cao8c3478a2007-10-18 23:39:20 -0700102 new_transaction = kzalloc(sizeof(*new_transaction),
Mingming Caoa5005da2007-10-16 18:38:25 -0400103 GFP_NOFS|__GFP_NOFAIL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104 if (!new_transaction) {
105 ret = -ENOMEM;
106 goto out;
107 }
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;
248 memset(handle, 0, sizeof(*handle));
249 handle->h_buffer_credits = nblocks;
250 handle->h_ref = 1;
251
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +0200252 lockdep_init_map(&handle->h_lockdep_map, "jbd_handle", &jbd_handle_key, 0);
253
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 return handle;
255}
256
257/**
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700258 * handle_t *journal_start() - Obtain a new handle.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 * @journal: Journal to start transaction on.
260 * @nblocks: number of block buffer we might modify
261 *
262 * We make sure that the transaction can guarantee at least nblocks of
263 * modified buffers in the log. We block until the log can guarantee
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700264 * that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 *
266 * This function is visible to journal users (like ext3fs), so is not
267 * called with the journal already locked.
268 *
269 * Return a pointer to a newly allocated handle, or NULL on failure
270 */
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 *
428 * This locks out any further updates from being started, and blocks
429 * until all existing updates have completed, returning only once the
430 * journal is in a quiescent state with no updates running.
431 *
432 * The journal lock should not be held on entry.
433 */
434void journal_lock_updates(journal_t *journal)
435{
436 DEFINE_WAIT(wait);
437
438 spin_lock(&journal->j_state_lock);
439 ++journal->j_barrier_count;
440
441 /* Wait until there are no running updates */
442 while (1) {
443 transaction_t *transaction = journal->j_running_transaction;
444
445 if (!transaction)
446 break;
447
448 spin_lock(&transaction->t_handle_lock);
449 if (!transaction->t_updates) {
450 spin_unlock(&transaction->t_handle_lock);
451 break;
452 }
453 prepare_to_wait(&journal->j_wait_updates, &wait,
454 TASK_UNINTERRUPTIBLE);
455 spin_unlock(&transaction->t_handle_lock);
456 spin_unlock(&journal->j_state_lock);
457 schedule();
458 finish_wait(&journal->j_wait_updates, &wait);
459 spin_lock(&journal->j_state_lock);
460 }
461 spin_unlock(&journal->j_state_lock);
462
463 /*
464 * We have now established a barrier against other normal updates, but
465 * we also need to barrier against other journal_lock_updates() calls
466 * to make sure that we serialise special journal-locked operations
467 * too.
468 */
Arjan van de Ven2c68ee72006-03-23 03:00:35 -0800469 mutex_lock(&journal->j_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470}
471
472/**
473 * void journal_unlock_updates (journal_t* journal) - release barrier
474 * @journal: Journal to release the barrier on.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700475 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476 * Release a transaction barrier obtained with journal_lock_updates().
477 *
478 * Should be called without the journal lock held.
479 */
480void journal_unlock_updates (journal_t *journal)
481{
482 J_ASSERT(journal->j_barrier_count != 0);
483
Arjan van de Ven2c68ee72006-03-23 03:00:35 -0800484 mutex_unlock(&journal->j_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 spin_lock(&journal->j_state_lock);
486 --journal->j_barrier_count;
487 spin_unlock(&journal->j_state_lock);
488 wake_up(&journal->j_wait_transaction_locked);
489}
490
Jan Kara1e9fd532009-06-24 17:31:40 +0200491static void warn_dirty_buffer(struct buffer_head *bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492{
Jan Kara1e9fd532009-06-24 17:31:40 +0200493 char b[BDEVNAME_SIZE];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494
Jan Kara1e9fd532009-06-24 17:31:40 +0200495 printk(KERN_WARNING
496 "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
497 "There's a risk of filesystem corruption in case of system "
498 "crash.\n",
499 bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500}
501
502/*
503 * If the buffer is already part of the current transaction, then there
504 * is nothing we need to do. If it is already part of a prior
505 * transaction which we are still committing to disk, then we need to
506 * make sure that we do not overwrite the old copy: we do copy-out to
507 * preserve the copy going to disk. We also account the buffer against
508 * the handle's metadata buffer credits (unless the buffer is already
509 * part of the transaction, that is).
510 *
511 */
512static int
513do_get_write_access(handle_t *handle, struct journal_head *jh,
514 int force_copy)
515{
516 struct buffer_head *bh;
517 transaction_t *transaction;
518 journal_t *journal;
519 int error;
520 char *frozen_buffer = NULL;
521 int need_copy = 0;
522
523 if (is_handle_aborted(handle))
524 return -EROFS;
525
526 transaction = handle->h_transaction;
527 journal = transaction->t_journal;
528
Namhyung Kimb8ea49f2010-10-13 23:14:07 +0900529 jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530
531 JBUFFER_TRACE(jh, "entry");
532repeat:
533 bh = jh2bh(jh);
534
535 /* @@@ Need to check for errors here at some point. */
536
537 lock_buffer(bh);
538 jbd_lock_bh_state(bh);
539
540 /* We now hold the buffer lock so it is safe to query the buffer
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700541 * state. Is the buffer dirty?
542 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 * If so, there are two possibilities. The buffer may be
544 * non-journaled, and undergoing a quite legitimate writeback.
545 * Otherwise, it is journaled, and we don't expect dirty buffers
546 * in that state (the buffers should be marked JBD_Dirty
547 * instead.) So either the IO is being done under our own
548 * control and this is a bug, or it's a third party IO such as
549 * dump(8) (which may leave the buffer scheduled for read ---
550 * ie. locked but not dirty) or tune2fs (which may actually have
551 * the buffer dirtied, ugh.) */
552
553 if (buffer_dirty(bh)) {
554 /*
555 * First question: is this buffer already part of the current
556 * transaction or the existing committing transaction?
557 */
558 if (jh->b_transaction) {
559 J_ASSERT_JH(jh,
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700560 jh->b_transaction == transaction ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 jh->b_transaction ==
562 journal->j_committing_transaction);
563 if (jh->b_next_transaction)
564 J_ASSERT_JH(jh, jh->b_next_transaction ==
565 transaction);
Jan Kara1e9fd532009-06-24 17:31:40 +0200566 warn_dirty_buffer(bh);
Jan Kara4407c2b2005-09-06 15:19:17 -0700567 }
568 /*
569 * In any case we need to clean the dirty flag and we must
570 * do it under the buffer lock to be sure we don't race
571 * with running write-out.
572 */
Jan Kara1e9fd532009-06-24 17:31:40 +0200573 JBUFFER_TRACE(jh, "Journalling dirty buffer");
574 clear_buffer_dirty(bh);
575 set_buffer_jbddirty(bh);
Dave Kleikampe9ad5622006-09-27 01:49:35 -0700576 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577
578 unlock_buffer(bh);
579
580 error = -EROFS;
581 if (is_handle_aborted(handle)) {
582 jbd_unlock_bh_state(bh);
583 goto out;
584 }
585 error = 0;
586
587 /*
588 * The buffer is already part of this transaction if b_transaction or
589 * b_next_transaction points to it
590 */
591 if (jh->b_transaction == transaction ||
592 jh->b_next_transaction == transaction)
593 goto done;
594
595 /*
Josef Bacik5bc833f2008-04-28 02:16:10 -0700596 * this is the first time this transaction is touching this buffer,
597 * reset the modified flag
598 */
599 jh->b_modified = 0;
600
601 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 * If there is already a copy-out version of this buffer, then we don't
603 * need to make another one
604 */
605 if (jh->b_frozen_data) {
606 JBUFFER_TRACE(jh, "has frozen data");
607 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
608 jh->b_next_transaction = transaction;
609 goto done;
610 }
611
612 /* Is there data here we need to preserve? */
613
614 if (jh->b_transaction && jh->b_transaction != transaction) {
615 JBUFFER_TRACE(jh, "owned by older transaction");
616 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
617 J_ASSERT_JH(jh, jh->b_transaction ==
618 journal->j_committing_transaction);
619
620 /* There is one case we have to be very careful about.
621 * If the committing transaction is currently writing
622 * this buffer out to disk and has NOT made a copy-out,
623 * then we cannot modify the buffer contents at all
624 * right now. The essence of copy-out is that it is the
625 * extra copy, not the primary copy, which gets
626 * journaled. If the primary copy is already going to
627 * disk then we cannot do copy-out here. */
628
629 if (jh->b_jlist == BJ_Shadow) {
630 DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
631 wait_queue_head_t *wqh;
632
633 wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
634
635 JBUFFER_TRACE(jh, "on shadow: sleep");
636 jbd_unlock_bh_state(bh);
637 /* commit wakes up all shadow buffers after IO */
638 for ( ; ; ) {
639 prepare_to_wait(wqh, &wait.wait,
640 TASK_UNINTERRUPTIBLE);
641 if (jh->b_jlist != BJ_Shadow)
642 break;
643 schedule();
644 }
645 finish_wait(wqh, &wait.wait);
646 goto repeat;
647 }
648
649 /* Only do the copy if the currently-owning transaction
650 * still needs it. If it is on the Forget list, the
651 * committing transaction is past that stage. The
652 * buffer had better remain locked during the kmalloc,
653 * but that should be true --- we hold the journal lock
654 * still and the buffer is already on the BUF_JOURNAL
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700655 * list so won't be flushed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 *
657 * Subtle point, though: if this is a get_undo_access,
658 * then we will be relying on the frozen_data to contain
659 * the new value of the committed_data record after the
660 * transaction, so we HAVE to force the frozen_data copy
661 * in that case. */
662
663 if (jh->b_jlist != BJ_Forget || force_copy) {
664 JBUFFER_TRACE(jh, "generate frozen data");
665 if (!frozen_buffer) {
666 JBUFFER_TRACE(jh, "allocate memory for buffer");
667 jbd_unlock_bh_state(bh);
Badari Pulavartyea817392006-08-27 01:23:52 -0700668 frozen_buffer =
Mingming Caoc089d492007-10-16 18:38:25 -0400669 jbd_alloc(jh2bh(jh)->b_size,
Badari Pulavartyea817392006-08-27 01:23:52 -0700670 GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 if (!frozen_buffer) {
672 printk(KERN_EMERG
673 "%s: OOM for frozen_buffer\n",
Harvey Harrison08fc99b2008-04-28 02:16:16 -0700674 __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 JBUFFER_TRACE(jh, "oom!");
676 error = -ENOMEM;
677 jbd_lock_bh_state(bh);
678 goto done;
679 }
680 goto repeat;
681 }
682 jh->b_frozen_data = frozen_buffer;
683 frozen_buffer = NULL;
684 need_copy = 1;
685 }
686 jh->b_next_transaction = transaction;
687 }
688
689
690 /*
691 * Finally, if the buffer is not journaled right now, we need to make
692 * sure it doesn't get written to disk before the caller actually
693 * commits the new data
694 */
695 if (!jh->b_transaction) {
696 JBUFFER_TRACE(jh, "no transaction");
697 J_ASSERT_JH(jh, !jh->b_next_transaction);
698 jh->b_transaction = transaction;
699 JBUFFER_TRACE(jh, "file as BJ_Reserved");
700 spin_lock(&journal->j_list_lock);
701 __journal_file_buffer(jh, transaction, BJ_Reserved);
702 spin_unlock(&journal->j_list_lock);
703 }
704
705done:
706 if (need_copy) {
707 struct page *page;
708 int offset;
709 char *source;
710
711 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
712 "Possible IO failure.\n");
713 page = jh2bh(jh)->b_page;
Namhyung Kim8117f982010-10-05 01:03:40 +0900714 offset = offset_in_page(jh2bh(jh)->b_data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 source = kmap_atomic(page, KM_USER0);
716 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
717 kunmap_atomic(source, KM_USER0);
718 }
719 jbd_unlock_bh_state(bh);
720
721 /*
722 * If we are about to journal a buffer, then any revoke pending on it is
723 * no longer valid
724 */
725 journal_cancel_revoke(handle, jh);
726
727out:
Andrew Morton304c4c82006-06-23 02:05:31 -0700728 if (unlikely(frozen_buffer)) /* It's usually NULL */
Mingming Caoc089d492007-10-16 18:38:25 -0400729 jbd_free(frozen_buffer, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730
731 JBUFFER_TRACE(jh, "exit");
732 return error;
733}
734
735/**
736 * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
737 * @handle: transaction to add buffer modifications to
738 * @bh: bh to be used for metadata writes
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 *
740 * Returns an error code or 0 on success.
741 *
742 * In full data journalling mode the buffer may be of type BJ_AsyncData,
743 * because we're write()ing a buffer which is also part of a shared mapping.
744 */
745
746int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
747{
748 struct journal_head *jh = journal_add_journal_head(bh);
749 int rc;
750
751 /* We do not want to get caught playing with fields which the
752 * log thread also manipulates. Make sure that the buffer
753 * completes any outstanding IO before proceeding. */
754 rc = do_get_write_access(handle, jh, 0);
755 journal_put_journal_head(jh);
756 return rc;
757}
758
759
760/*
761 * When the user wants to journal a newly created buffer_head
762 * (ie. getblk() returned a new buffer and we are going to populate it
763 * manually rather than reading off disk), then we need to keep the
764 * buffer_head locked until it has been completely filled with new
765 * data. In this case, we should be able to make the assertion that
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700766 * the bh is not already part of an existing transaction.
767 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 * The buffer should already be locked by the caller by this point.
769 * There is no lock ranking violation: it was a newly created,
770 * unlocked buffer beforehand. */
771
772/**
773 * int journal_get_create_access () - notify intent to use newly created bh
774 * @handle: transaction to new buffer to
775 * @bh: new buffer.
776 *
777 * Call this if you create a new bh.
778 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700779int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780{
781 transaction_t *transaction = handle->h_transaction;
782 journal_t *journal = transaction->t_journal;
783 struct journal_head *jh = journal_add_journal_head(bh);
784 int err;
785
786 jbd_debug(5, "journal_head %p\n", jh);
787 err = -EROFS;
788 if (is_handle_aborted(handle))
789 goto out;
790 err = 0;
791
792 JBUFFER_TRACE(jh, "entry");
793 /*
794 * The buffer may already belong to this transaction due to pre-zeroing
795 * in the filesystem's new_block code. It may also be on the previous,
796 * committing transaction's lists, but it HAS to be in Forget state in
797 * that case: the transaction must have deleted the buffer for it to be
798 * reused here.
799 */
800 jbd_lock_bh_state(bh);
801 spin_lock(&journal->j_list_lock);
802 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
803 jh->b_transaction == NULL ||
804 (jh->b_transaction == journal->j_committing_transaction &&
805 jh->b_jlist == BJ_Forget)));
806
807 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
808 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
809
810 if (jh->b_transaction == NULL) {
Jan Kara1e9fd532009-06-24 17:31:40 +0200811 /*
812 * Previous journal_forget() could have left the buffer
813 * with jbddirty bit set because it was being committed. When
814 * the commit finished, we've filed the buffer for
815 * checkpointing and marked it dirty. Now we are reallocating
816 * the buffer so the transaction freeing it must have
817 * committed and so it's safe to clear the dirty bit.
818 */
819 clear_buffer_dirty(jh2bh(jh));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 jh->b_transaction = transaction;
Josef Bacik5bc833f2008-04-28 02:16:10 -0700821
822 /* first access by this transaction */
823 jh->b_modified = 0;
824
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 JBUFFER_TRACE(jh, "file as BJ_Reserved");
826 __journal_file_buffer(jh, transaction, BJ_Reserved);
827 } else if (jh->b_transaction == journal->j_committing_transaction) {
Josef Bacik5bc833f2008-04-28 02:16:10 -0700828 /* first access by this transaction */
829 jh->b_modified = 0;
830
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 JBUFFER_TRACE(jh, "set next transaction");
832 jh->b_next_transaction = transaction;
833 }
834 spin_unlock(&journal->j_list_lock);
835 jbd_unlock_bh_state(bh);
836
837 /*
838 * akpm: I added this. ext3_alloc_branch can pick up new indirect
839 * blocks which contain freed but then revoked metadata. We need
840 * to cancel the revoke in case we end up freeing it yet again
841 * and the reallocating as data - this would cause a second revoke,
842 * which hits an assertion error.
843 */
844 JBUFFER_TRACE(jh, "cancelling revoke");
845 journal_cancel_revoke(handle, jh);
846 journal_put_journal_head(jh);
847out:
848 return err;
849}
850
851/**
Randy Dunlap78a4a502008-02-29 22:02:31 -0800852 * int journal_get_undo_access() - Notify intent to modify metadata with non-rewindable consequences
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 * @handle: transaction
854 * @bh: buffer to undo
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 *
856 * Sometimes there is a need to distinguish between metadata which has
857 * been committed to disk and that which has not. The ext3fs code uses
858 * this for freeing and allocating space, we have to make sure that we
859 * do not reuse freed space until the deallocation has been committed,
860 * since if we overwrote that space we would make the delete
861 * un-rewindable in case of a crash.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700862 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 * To deal with that, journal_get_undo_access requests write access to a
864 * buffer for parts of non-rewindable operations such as delete
865 * operations on the bitmaps. The journaling code must keep a copy of
866 * the buffer's contents prior to the undo_access call until such time
867 * as we know that the buffer has definitely been committed to disk.
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700868 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869 * We never need to know which transaction the committed data is part
870 * of, buffers touched here are guaranteed to be dirtied later and so
871 * will be committed to a new transaction in due course, at which point
872 * we can discard the old committed data pointer.
873 *
874 * Returns error number or 0 on success.
875 */
876int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
877{
878 int err;
879 struct journal_head *jh = journal_add_journal_head(bh);
880 char *committed_data = NULL;
881
882 JBUFFER_TRACE(jh, "entry");
883
884 /*
885 * Do this first --- it can drop the journal lock, so we want to
886 * make sure that obtaining the committed_data is done
887 * atomically wrt. completion of any outstanding commits.
888 */
889 err = do_get_write_access(handle, jh, 1);
890 if (err)
891 goto out;
892
893repeat:
894 if (!jh->b_committed_data) {
Mingming Caoc089d492007-10-16 18:38:25 -0400895 committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 if (!committed_data) {
897 printk(KERN_EMERG "%s: No memory for committed data\n",
Harvey Harrison08fc99b2008-04-28 02:16:16 -0700898 __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 err = -ENOMEM;
900 goto out;
901 }
902 }
903
904 jbd_lock_bh_state(bh);
905 if (!jh->b_committed_data) {
906 /* Copy out the current buffer contents into the
907 * preserved, committed copy. */
908 JBUFFER_TRACE(jh, "generate b_committed data");
909 if (!committed_data) {
910 jbd_unlock_bh_state(bh);
911 goto repeat;
912 }
913
914 jh->b_committed_data = committed_data;
915 committed_data = NULL;
916 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
917 }
918 jbd_unlock_bh_state(bh);
919out:
920 journal_put_journal_head(jh);
Andrew Morton304c4c82006-06-23 02:05:31 -0700921 if (unlikely(committed_data))
Mingming Caoc089d492007-10-16 18:38:25 -0400922 jbd_free(committed_data, bh->b_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 return err;
924}
925
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700926/**
Randy Dunlap78a4a502008-02-29 22:02:31 -0800927 * int journal_dirty_data() - mark a buffer as containing dirty data to be flushed
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 * @handle: transaction
929 * @bh: bufferhead to mark
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700930 *
Randy Dunlap78a4a502008-02-29 22:02:31 -0800931 * Description:
932 * Mark a buffer as containing dirty data which needs to be flushed before
933 * we can commit the current transaction.
934 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 * The buffer is placed on the transaction's data list and is marked as
936 * belonging to the transaction.
937 *
938 * Returns error number or 0 on success.
939 *
940 * journal_dirty_data() can be called via page_launder->ext3_writepage
941 * by kswapd.
942 */
943int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
944{
945 journal_t *journal = handle->h_transaction->t_journal;
946 int need_brelse = 0;
947 struct journal_head *jh;
Hidehiro Kawai960a22a2008-10-18 20:27:58 -0700948 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
950 if (is_handle_aborted(handle))
Hidehiro Kawai960a22a2008-10-18 20:27:58 -0700951 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952
953 jh = journal_add_journal_head(bh);
954 JBUFFER_TRACE(jh, "entry");
955
956 /*
957 * The buffer could *already* be dirty. Writeout can start
958 * at any time.
959 */
960 jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
961
962 /*
963 * What if the buffer is already part of a running transaction?
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700964 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 * There are two cases:
966 * 1) It is part of the current running transaction. Refile it,
967 * just in case we have allocated it as metadata, deallocated
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700968 * it, then reallocated it as data.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 * 2) It is part of the previous, still-committing transaction.
970 * If all we want to do is to guarantee that the buffer will be
971 * written to disk before this new transaction commits, then
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700972 * being sure that the *previous* transaction has this same
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 * property is sufficient for us! Just leave it on its old
974 * transaction.
975 *
976 * In case (2), the buffer must not already exist as metadata
977 * --- that would violate write ordering (a transaction is free
978 * to write its data at any point, even before the previous
979 * committing transaction has committed). The caller must
980 * never, ever allow this to happen: there's nothing we can do
981 * about it in this layer.
982 */
983 jbd_lock_bh_state(bh);
984 spin_lock(&journal->j_list_lock);
Eric Sandeenf58a74d2006-10-28 10:38:27 -0700985
986 /* Now that we have bh_state locked, are we really still mapped? */
987 if (!buffer_mapped(bh)) {
988 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
989 goto no_journal;
990 }
991
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 if (jh->b_transaction) {
993 JBUFFER_TRACE(jh, "has transaction");
994 if (jh->b_transaction != handle->h_transaction) {
995 JBUFFER_TRACE(jh, "belongs to older transaction");
996 J_ASSERT_JH(jh, jh->b_transaction ==
997 journal->j_committing_transaction);
998
999 /* @@@ IS THIS TRUE ? */
1000 /*
1001 * Not any more. Scenario: someone does a write()
1002 * in data=journal mode. The buffer's transaction has
1003 * moved into commit. Then someone does another
1004 * write() to the file. We do the frozen data copyout
1005 * and set b_next_transaction to point to j_running_t.
1006 * And while we're in that state, someone does a
1007 * writepage() in an attempt to pageout the same area
1008 * of the file via a shared mapping. At present that
1009 * calls journal_dirty_data(), and we get right here.
1010 * It may be too late to journal the data. Simply
1011 * falling through to the next test will suffice: the
1012 * data will be dirty and wil be checkpointed. The
1013 * ordering comments in the next comment block still
1014 * apply.
1015 */
1016 //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1017
1018 /*
1019 * If we're journalling data, and this buffer was
1020 * subject to a write(), it could be metadata, forget
1021 * or shadow against the committing transaction. Now,
1022 * someone has dirtied the same darn page via a mapping
1023 * and it is being writepage()'d.
1024 * We *could* just steal the page from commit, with some
1025 * fancy locking there. Instead, we just skip it -
1026 * don't tie the page's buffers to the new transaction
1027 * at all.
1028 * Implication: if we crash before the writepage() data
1029 * is written into the filesystem, recovery will replay
1030 * the write() data.
1031 */
1032 if (jh->b_jlist != BJ_None &&
1033 jh->b_jlist != BJ_SyncData &&
1034 jh->b_jlist != BJ_Locked) {
1035 JBUFFER_TRACE(jh, "Not stealing");
1036 goto no_journal;
1037 }
1038
1039 /*
1040 * This buffer may be undergoing writeout in commit. We
1041 * can't return from here and let the caller dirty it
1042 * again because that can cause the write-out loop in
1043 * commit to never terminate.
1044 */
1045 if (buffer_dirty(bh)) {
1046 get_bh(bh);
1047 spin_unlock(&journal->j_list_lock);
1048 jbd_unlock_bh_state(bh);
1049 need_brelse = 1;
1050 sync_dirty_buffer(bh);
1051 jbd_lock_bh_state(bh);
1052 spin_lock(&journal->j_list_lock);
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001053 /* Since we dropped the lock... */
1054 if (!buffer_mapped(bh)) {
1055 JBUFFER_TRACE(jh, "buffer got unmapped");
1056 goto no_journal;
1057 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058 /* The buffer may become locked again at any
1059 time if it is redirtied */
1060 }
1061
Hidehiro Kawai960a22a2008-10-18 20:27:58 -07001062 /*
1063 * We cannot remove the buffer with io error from the
1064 * committing transaction, because otherwise it would
1065 * miss the error and the commit would not abort.
1066 */
1067 if (unlikely(!buffer_uptodate(bh))) {
1068 ret = -EIO;
1069 goto no_journal;
1070 }
1071
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 if (jh->b_transaction != NULL) {
1073 JBUFFER_TRACE(jh, "unfile from commit");
1074 __journal_temp_unlink_buffer(jh);
1075 /* It still points to the committing
1076 * transaction; move it to this one so
1077 * that the refile assert checks are
1078 * happy. */
1079 jh->b_transaction = handle->h_transaction;
1080 }
1081 /* The buffer will be refiled below */
1082
1083 }
1084 /*
1085 * Special case --- the buffer might actually have been
1086 * allocated and then immediately deallocated in the previous,
1087 * committing transaction, so might still be left on that
1088 * transaction's metadata lists.
1089 */
1090 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1091 JBUFFER_TRACE(jh, "not on correct data list: unfile");
1092 J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1093 __journal_temp_unlink_buffer(jh);
1094 jh->b_transaction = handle->h_transaction;
1095 JBUFFER_TRACE(jh, "file as data");
1096 __journal_file_buffer(jh, handle->h_transaction,
1097 BJ_SyncData);
1098 }
1099 } else {
1100 JBUFFER_TRACE(jh, "not on a transaction");
1101 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1102 }
1103no_journal:
1104 spin_unlock(&journal->j_list_lock);
1105 jbd_unlock_bh_state(bh);
1106 if (need_brelse) {
1107 BUFFER_TRACE(bh, "brelse");
1108 __brelse(bh);
1109 }
1110 JBUFFER_TRACE(jh, "exit");
1111 journal_put_journal_head(jh);
Hidehiro Kawai960a22a2008-10-18 20:27:58 -07001112 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113}
1114
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001115/**
Randy Dunlap78a4a502008-02-29 22:02:31 -08001116 * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 * @handle: transaction to add buffer to.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001118 * @bh: buffer to mark
1119 *
Randy Dunlap78a4a502008-02-29 22:02:31 -08001120 * Mark dirty metadata which needs to be journaled as part of the current
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 * transaction.
1122 *
1123 * The buffer is placed on the transaction's metadata list and is marked
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001124 * as belonging to the transaction.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 *
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001126 * Returns error number or 0 on success.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 *
1128 * Special care needs to be taken if the buffer already belongs to the
1129 * current committing transaction (in which case we should have frozen
1130 * data present for that commit). In that case, we don't relink the
1131 * buffer: that only gets done when the old transaction finally
1132 * completes its commit.
1133 */
1134int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1135{
1136 transaction_t *transaction = handle->h_transaction;
1137 journal_t *journal = transaction->t_journal;
1138 struct journal_head *jh = bh2jh(bh);
1139
1140 jbd_debug(5, "journal_head %p\n", jh);
1141 JBUFFER_TRACE(jh, "entry");
1142 if (is_handle_aborted(handle))
1143 goto out;
1144
1145 jbd_lock_bh_state(bh);
1146
1147 if (jh->b_modified == 0) {
1148 /*
1149 * This buffer's got modified and becoming part
1150 * of the transaction. This needs to be done
1151 * once a transaction -bzzz
1152 */
1153 jh->b_modified = 1;
1154 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1155 handle->h_buffer_credits--;
1156 }
1157
1158 /*
1159 * fastpath, to avoid expensive locking. If this buffer is already
1160 * on the running transaction's metadata list there is nothing to do.
1161 * Nobody can take it off again because there is a handle open.
1162 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1163 * result in this test being false, so we go in and take the locks.
1164 */
1165 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1166 JBUFFER_TRACE(jh, "fastpath");
1167 J_ASSERT_JH(jh, jh->b_transaction ==
1168 journal->j_running_transaction);
1169 goto out_unlock_bh;
1170 }
1171
1172 set_buffer_jbddirty(bh);
1173
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001174 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 * Metadata already on the current transaction list doesn't
1176 * need to be filed. Metadata on another transaction's list must
1177 * be committing, and will be refiled once the commit completes:
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001178 * leave it alone for now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 */
1180 if (jh->b_transaction != transaction) {
1181 JBUFFER_TRACE(jh, "already on other transaction");
1182 J_ASSERT_JH(jh, jh->b_transaction ==
1183 journal->j_committing_transaction);
1184 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1185 /* And this case is illegal: we can't reuse another
1186 * transaction's data buffer, ever. */
1187 goto out_unlock_bh;
1188 }
1189
1190 /* That test should have eliminated the following case: */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001191 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192
1193 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1194 spin_lock(&journal->j_list_lock);
1195 __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1196 spin_unlock(&journal->j_list_lock);
1197out_unlock_bh:
1198 jbd_unlock_bh_state(bh);
1199out:
1200 JBUFFER_TRACE(jh, "exit");
1201 return 0;
1202}
1203
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001204/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 * journal_release_buffer: undo a get_write_access without any buffer
1206 * updates, if the update decided in the end that it didn't need access.
1207 *
1208 */
1209void
1210journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1211{
1212 BUFFER_TRACE(bh, "entry");
1213}
1214
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001215/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 * void journal_forget() - bforget() for potentially-journaled buffers.
1217 * @handle: transaction handle
1218 * @bh: bh to 'forget'
1219 *
1220 * We can only do the bforget if there are no commits pending against the
1221 * buffer. If the buffer is dirty in the current running transaction we
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001222 * can safely unlink it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 *
1224 * bh may not be a journalled buffer at all - it may be a non-JBD
1225 * buffer which came off the hashtable. Check for this.
1226 *
1227 * Decrements bh->b_count by one.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001228 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 * Allow this call even if the handle has aborted --- it may be part of
1230 * the caller's cleanup after an abort.
1231 */
1232int journal_forget (handle_t *handle, struct buffer_head *bh)
1233{
1234 transaction_t *transaction = handle->h_transaction;
1235 journal_t *journal = transaction->t_journal;
1236 struct journal_head *jh;
1237 int drop_reserve = 0;
1238 int err = 0;
Josef Bacik5b9a4992008-04-28 02:16:12 -07001239 int was_modified = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240
1241 BUFFER_TRACE(bh, "entry");
1242
1243 jbd_lock_bh_state(bh);
1244 spin_lock(&journal->j_list_lock);
1245
1246 if (!buffer_jbd(bh))
1247 goto not_jbd;
1248 jh = bh2jh(bh);
1249
1250 /* Critical error: attempting to delete a bitmap buffer, maybe?
1251 * Don't do any jbd operations, and return an error. */
1252 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1253 "inconsistent data on disk")) {
1254 err = -EIO;
1255 goto not_jbd;
1256 }
1257
Josef Bacik5b9a4992008-04-28 02:16:12 -07001258 /* keep track of wether or not this transaction modified us */
1259 was_modified = jh->b_modified;
1260
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 /*
1262 * The buffer's going from the transaction, we must drop
1263 * all references -bzzz
1264 */
1265 jh->b_modified = 0;
1266
1267 if (jh->b_transaction == handle->h_transaction) {
1268 J_ASSERT_JH(jh, !jh->b_frozen_data);
1269
1270 /* If we are forgetting a buffer which is already part
1271 * of this transaction, then we can just drop it from
1272 * the transaction immediately. */
1273 clear_buffer_dirty(bh);
1274 clear_buffer_jbddirty(bh);
1275
1276 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1277
Josef Bacik5b9a4992008-04-28 02:16:12 -07001278 /*
1279 * we only want to drop a reference if this transaction
1280 * modified the buffer
1281 */
1282 if (was_modified)
1283 drop_reserve = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001285 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 * We are no longer going to journal this buffer.
1287 * However, the commit of this transaction is still
1288 * important to the buffer: the delete that we are now
1289 * processing might obsolete an old log entry, so by
1290 * committing, we can satisfy the buffer's checkpoint.
1291 *
1292 * So, if we have a checkpoint on the buffer, we should
1293 * now refile the buffer on our BJ_Forget list so that
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001294 * we know to remove the checkpoint after we commit.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001295 */
1296
1297 if (jh->b_cp_transaction) {
1298 __journal_temp_unlink_buffer(jh);
1299 __journal_file_buffer(jh, transaction, BJ_Forget);
1300 } else {
1301 __journal_unfile_buffer(jh);
1302 journal_remove_journal_head(bh);
1303 __brelse(bh);
1304 if (!buffer_jbd(bh)) {
1305 spin_unlock(&journal->j_list_lock);
1306 jbd_unlock_bh_state(bh);
1307 __bforget(bh);
1308 goto drop;
1309 }
1310 }
1311 } else if (jh->b_transaction) {
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001312 J_ASSERT_JH(jh, (jh->b_transaction ==
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 journal->j_committing_transaction));
1314 /* However, if the buffer is still owned by a prior
1315 * (committing) transaction, we can't drop it yet... */
1316 JBUFFER_TRACE(jh, "belongs to older transaction");
1317 /* ... but we CAN drop it from the new transaction if we
1318 * have also modified it since the original commit. */
1319
1320 if (jh->b_next_transaction) {
1321 J_ASSERT(jh->b_next_transaction == transaction);
1322 jh->b_next_transaction = NULL;
Josef Bacik5b9a4992008-04-28 02:16:12 -07001323
1324 /*
1325 * only drop a reference if this transaction modified
1326 * the buffer
1327 */
1328 if (was_modified)
1329 drop_reserve = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 }
1331 }
1332
1333not_jbd:
1334 spin_unlock(&journal->j_list_lock);
1335 jbd_unlock_bh_state(bh);
1336 __brelse(bh);
1337drop:
1338 if (drop_reserve) {
1339 /* no need to reserve log space for this block -bzzz */
1340 handle->h_buffer_credits++;
1341 }
1342 return err;
1343}
1344
1345/**
1346 * int journal_stop() - complete a transaction
1347 * @handle: tranaction to complete.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001348 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 * All done for a particular handle.
1350 *
1351 * There is not much action needed here. We just return any remaining
1352 * buffer credits to the transaction and remove the handle. The only
1353 * complication is that we need to start a commit operation if the
1354 * filesystem is marked for synchronous update.
1355 *
1356 * journal_stop itself will not usually return an error, but it may
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001357 * do so in unusual circumstances. In particular, expect it to
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 * return -EIO if a journal_abort has been executed since the
1359 * transaction began.
1360 */
1361int journal_stop(handle_t *handle)
1362{
1363 transaction_t *transaction = handle->h_transaction;
1364 journal_t *journal = transaction->t_journal;
Josef Bacikf420d4d2009-01-07 18:07:24 -08001365 int err;
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001366 pid_t pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 J_ASSERT(journal_current_handle() == handle);
1369
1370 if (is_handle_aborted(handle))
1371 err = -EIO;
OGAWA Hirofumi3e2a5322006-10-19 23:29:11 -07001372 else {
1373 J_ASSERT(transaction->t_updates > 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 err = 0;
OGAWA Hirofumi3e2a5322006-10-19 23:29:11 -07001375 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376
1377 if (--handle->h_ref > 0) {
1378 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1379 handle->h_ref);
1380 return err;
1381 }
1382
1383 jbd_debug(4, "Handle %p going down\n", handle);
1384
1385 /*
1386 * Implement synchronous transaction batching. If the handle
1387 * was synchronous, don't force a commit immediately. Let's
1388 * yield and let another thread piggyback onto this transaction.
1389 * Keep doing that while new threads continue to arrive.
1390 * It doesn't cost much - we're about to run a commit and sleep
1391 * on IO anyway. Speeds up many-threaded, many-dir operations
1392 * by 30x or more...
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001393 *
Josef Bacikf420d4d2009-01-07 18:07:24 -08001394 * We try and optimize the sleep time against what the underlying disk
1395 * can do, instead of having a static sleep time. This is usefull for
1396 * the case where our storage is so fast that it is more optimal to go
1397 * ahead and force a flush and wait for the transaction to be committed
1398 * than it is to wait for an arbitrary amount of time for new writers to
Daniel Mack3ad2f3f2010-02-03 08:01:28 +08001399 * join the transaction. We achieve this by measuring how long it takes
Josef Bacikf420d4d2009-01-07 18:07:24 -08001400 * to commit a transaction, and compare it with how long this
1401 * transaction has been running, and if run time < commit time then we
1402 * sleep for the delta and commit. This greatly helps super fast disks
1403 * that would see slowdowns as more threads started doing fsyncs.
1404 *
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001405 * But don't do this if this process was the most recent one to
1406 * perform a synchronous write. We do this to detect the case where a
1407 * single process is doing a stream of sync writes. No point in waiting
1408 * for joiners in that case.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 */
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001410 pid = current->pid;
1411 if (handle->h_sync && journal->j_last_sync_writer != pid) {
Josef Bacikf420d4d2009-01-07 18:07:24 -08001412 u64 commit_time, trans_time;
1413
Andrew Mortonfe1dcbc2006-02-04 23:27:54 -08001414 journal->j_last_sync_writer = pid;
Josef Bacikf420d4d2009-01-07 18:07:24 -08001415
1416 spin_lock(&journal->j_state_lock);
1417 commit_time = journal->j_average_commit_time;
1418 spin_unlock(&journal->j_state_lock);
1419
1420 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1421 transaction->t_start_time));
1422
1423 commit_time = min_t(u64, commit_time,
1424 1000*jiffies_to_usecs(1));
1425
1426 if (trans_time < commit_time) {
1427 ktime_t expires = ktime_add_ns(ktime_get(),
1428 commit_time);
1429 set_current_state(TASK_UNINTERRUPTIBLE);
1430 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1431 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 }
1433
Theodore Ts'o512a0042009-03-27 22:14:27 -04001434 if (handle->h_sync)
1435 transaction->t_synchronous_commit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 current->journal_info = NULL;
1437 spin_lock(&journal->j_state_lock);
1438 spin_lock(&transaction->t_handle_lock);
1439 transaction->t_outstanding_credits -= handle->h_buffer_credits;
1440 transaction->t_updates--;
1441 if (!transaction->t_updates) {
1442 wake_up(&journal->j_wait_updates);
1443 if (journal->j_barrier_count)
1444 wake_up(&journal->j_wait_transaction_locked);
1445 }
1446
1447 /*
1448 * If the handle is marked SYNC, we need to set another commit
1449 * going! We also want to force a commit if the current
1450 * transaction is occupying too much of the log, or if the
1451 * transaction is too old now.
1452 */
1453 if (handle->h_sync ||
1454 transaction->t_outstanding_credits >
1455 journal->j_max_transaction_buffers ||
Dave Kleikampe9ad5622006-09-27 01:49:35 -07001456 time_after_eq(jiffies, transaction->t_expires)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 /* Do this even for aborted journals: an abort still
1458 * completes the commit thread, it just doesn't write
1459 * anything to disk. */
1460 tid_t tid = transaction->t_tid;
1461
1462 spin_unlock(&transaction->t_handle_lock);
1463 jbd_debug(2, "transaction too old, requesting commit for "
1464 "handle %p\n", handle);
1465 /* This is non-blocking */
1466 __log_start_commit(journal, transaction->t_tid);
1467 spin_unlock(&journal->j_state_lock);
1468
1469 /*
1470 * Special case: JFS_SYNC synchronous updates require us
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001471 * to wait for the commit to complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 */
1473 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1474 err = log_wait_commit(journal, tid);
1475 } else {
1476 spin_unlock(&transaction->t_handle_lock);
1477 spin_unlock(&journal->j_state_lock);
1478 }
1479
Ingo Molnar3295f0e2008-08-11 10:30:30 +02001480 lock_map_release(&handle->h_lockdep_map);
Peter Zijlstra34a3d1e2007-10-11 22:11:12 +02001481
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 jbd_free_handle(handle);
1483 return err;
1484}
1485
Randy Dunlap0cf01f62008-03-19 17:00:44 -07001486/**
1487 * int journal_force_commit() - force any uncommitted transactions
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 * @journal: journal to force
1489 *
1490 * For synchronous operations: force any uncommitted transactions
1491 * to disk. May seem kludgy, but it reuses all the handle batching
1492 * code in a very simple manner.
1493 */
1494int journal_force_commit(journal_t *journal)
1495{
1496 handle_t *handle;
1497 int ret;
1498
1499 handle = journal_start(journal, 1);
1500 if (IS_ERR(handle)) {
1501 ret = PTR_ERR(handle);
1502 } else {
1503 handle->h_sync = 1;
1504 ret = journal_stop(handle);
1505 }
1506 return ret;
1507}
1508
1509/*
1510 *
1511 * List management code snippets: various functions for manipulating the
1512 * transaction buffer lists.
1513 *
1514 */
1515
1516/*
1517 * Append a buffer to a transaction list, given the transaction's list head
1518 * pointer.
1519 *
1520 * j_list_lock is held.
1521 *
1522 * jbd_lock_bh_state(jh2bh(jh)) is held.
1523 */
1524
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001525static inline void
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526__blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1527{
1528 if (!*list) {
1529 jh->b_tnext = jh->b_tprev = jh;
1530 *list = jh;
1531 } else {
1532 /* Insert at the tail of the list to preserve order */
1533 struct journal_head *first = *list, *last = first->b_tprev;
1534 jh->b_tprev = last;
1535 jh->b_tnext = first;
1536 last->b_tnext = first->b_tprev = jh;
1537 }
1538}
1539
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001540/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 * Remove a buffer from a transaction list, given the transaction's list
1542 * head pointer.
1543 *
1544 * Called with j_list_lock held, and the journal may not be locked.
1545 *
1546 * jbd_lock_bh_state(jh2bh(jh)) is held.
1547 */
1548
1549static inline void
1550__blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1551{
1552 if (*list == jh) {
1553 *list = jh->b_tnext;
1554 if (*list == jh)
1555 *list = NULL;
1556 }
1557 jh->b_tprev->b_tnext = jh->b_tnext;
1558 jh->b_tnext->b_tprev = jh->b_tprev;
1559}
1560
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001561/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 * Remove a buffer from the appropriate transaction list.
1563 *
1564 * Note that this function can *change* the value of
1565 * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1566 * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list. If the caller
1567 * is holding onto a copy of one of thee pointers, it could go bad.
1568 * Generally the caller needs to re-read the pointer from the transaction_t.
1569 *
1570 * Called under j_list_lock. The journal may not be locked.
1571 */
Adrian Bunkd394e122006-12-06 20:38:26 -08001572static void __journal_temp_unlink_buffer(struct journal_head *jh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573{
1574 struct journal_head **list = NULL;
1575 transaction_t *transaction;
1576 struct buffer_head *bh = jh2bh(jh);
1577
1578 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1579 transaction = jh->b_transaction;
1580 if (transaction)
1581 assert_spin_locked(&transaction->t_journal->j_list_lock);
1582
1583 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1584 if (jh->b_jlist != BJ_None)
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001585 J_ASSERT_JH(jh, transaction != NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586
1587 switch (jh->b_jlist) {
1588 case BJ_None:
1589 return;
1590 case BJ_SyncData:
1591 list = &transaction->t_sync_datalist;
1592 break;
1593 case BJ_Metadata:
1594 transaction->t_nr_buffers--;
1595 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1596 list = &transaction->t_buffers;
1597 break;
1598 case BJ_Forget:
1599 list = &transaction->t_forget;
1600 break;
1601 case BJ_IO:
1602 list = &transaction->t_iobuf_list;
1603 break;
1604 case BJ_Shadow:
1605 list = &transaction->t_shadow_list;
1606 break;
1607 case BJ_LogCtl:
1608 list = &transaction->t_log_list;
1609 break;
1610 case BJ_Reserved:
1611 list = &transaction->t_reserved_list;
1612 break;
1613 case BJ_Locked:
1614 list = &transaction->t_locked_list;
1615 break;
1616 }
1617
1618 __blist_del_buffer(list, jh);
1619 jh->b_jlist = BJ_None;
1620 if (test_clear_buffer_jbddirty(bh))
1621 mark_buffer_dirty(bh); /* Expose it to the VM */
1622}
1623
1624void __journal_unfile_buffer(struct journal_head *jh)
1625{
1626 __journal_temp_unlink_buffer(jh);
1627 jh->b_transaction = NULL;
1628}
1629
1630void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1631{
1632 jbd_lock_bh_state(jh2bh(jh));
1633 spin_lock(&journal->j_list_lock);
1634 __journal_unfile_buffer(jh);
1635 spin_unlock(&journal->j_list_lock);
1636 jbd_unlock_bh_state(jh2bh(jh));
1637}
1638
1639/*
1640 * Called from journal_try_to_free_buffers().
1641 *
1642 * Called under jbd_lock_bh_state(bh)
1643 */
1644static void
1645__journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1646{
1647 struct journal_head *jh;
1648
1649 jh = bh2jh(bh);
1650
1651 if (buffer_locked(bh) || buffer_dirty(bh))
1652 goto out;
1653
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001654 if (jh->b_next_transaction != NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 goto out;
1656
1657 spin_lock(&journal->j_list_lock);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001658 if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1660 /* A written-back ordered data buffer */
1661 JBUFFER_TRACE(jh, "release data");
1662 __journal_unfile_buffer(jh);
1663 journal_remove_journal_head(bh);
1664 __brelse(bh);
1665 }
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001666 } else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 /* written-back checkpointed metadata buffer */
1668 if (jh->b_jlist == BJ_None) {
1669 JBUFFER_TRACE(jh, "remove from checkpoint list");
1670 __journal_remove_checkpoint(jh);
1671 journal_remove_journal_head(bh);
1672 __brelse(bh);
1673 }
1674 }
1675 spin_unlock(&journal->j_list_lock);
1676out:
1677 return;
1678}
1679
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001680/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 * int journal_try_to_free_buffers() - try to free page buffers.
1682 * @journal: journal for operation
1683 * @page: to try and free
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001684 * @gfp_mask: we use the mask to detect how hard should we try to release
1685 * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1686 * release the buffers.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 *
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001688 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 * For all the buffers on this page,
1690 * if they are fully written out ordered data, move them onto BUF_CLEAN
1691 * so try_to_free_buffers() can reap them.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001692 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 * This function returns non-zero if we wish try_to_free_buffers()
1694 * to be called. We do this if the page is releasable by try_to_free_buffers().
1695 * We also do it if the page has locked or dirty buffers and the caller wants
1696 * us to perform sync or async writeout.
1697 *
1698 * This complicates JBD locking somewhat. We aren't protected by the
1699 * BKL here. We wish to remove the buffer from its committing or
1700 * running transaction's ->t_datalist via __journal_unfile_buffer.
1701 *
1702 * This may *change* the value of transaction_t->t_datalist, so anyone
1703 * who looks at t_datalist needs to lock against this function.
1704 *
1705 * Even worse, someone may be doing a journal_dirty_data on this
1706 * buffer. So we need to lock against that. journal_dirty_data()
1707 * will come out of the lock with the buffer dirty, which makes it
1708 * ineligible for release here.
1709 *
1710 * Who else is affected by this? hmm... Really the only contender
1711 * is do_get_write_access() - it could be looking at the buffer while
1712 * journal_try_to_free_buffer() is changing its state. But that
1713 * cannot happen because we never reallocate freed data as metadata
1714 * while the data is part of a transaction. Yes?
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001715 *
1716 * Return 0 on failure, 1 on success
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001718int journal_try_to_free_buffers(journal_t *journal,
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001719 struct page *page, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720{
1721 struct buffer_head *head;
1722 struct buffer_head *bh;
1723 int ret = 0;
1724
1725 J_ASSERT(PageLocked(page));
1726
1727 head = page_buffers(page);
1728 bh = head;
1729 do {
1730 struct journal_head *jh;
1731
1732 /*
1733 * We take our own ref against the journal_head here to avoid
1734 * having to add tons of locking around each instance of
1735 * journal_remove_journal_head() and journal_put_journal_head().
1736 */
1737 jh = journal_grab_journal_head(bh);
1738 if (!jh)
1739 continue;
1740
1741 jbd_lock_bh_state(bh);
1742 __journal_try_to_free_buffer(journal, bh);
1743 journal_put_journal_head(jh);
1744 jbd_unlock_bh_state(bh);
1745 if (buffer_jbd(bh))
1746 goto busy;
1747 } while ((bh = bh->b_this_page) != head);
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001748
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749 ret = try_to_free_buffers(page);
Mingming Cao3f31fdd2008-07-25 01:46:22 -07001750
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751busy:
1752 return ret;
1753}
1754
1755/*
1756 * This buffer is no longer needed. If it is on an older transaction's
1757 * checkpoint list we need to record it on this transaction's forget list
1758 * to pin this buffer (and hence its checkpointing transaction) down until
1759 * this transaction commits. If the buffer isn't on a checkpoint list, we
1760 * release it.
1761 * Returns non-zero if JBD no longer has an interest in the buffer.
1762 *
1763 * Called under j_list_lock.
1764 *
1765 * Called under jbd_lock_bh_state(bh).
1766 */
1767static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1768{
1769 int may_free = 1;
1770 struct buffer_head *bh = jh2bh(jh);
1771
1772 __journal_unfile_buffer(jh);
1773
1774 if (jh->b_cp_transaction) {
1775 JBUFFER_TRACE(jh, "on running+cp transaction");
Jan Kara1e9fd532009-06-24 17:31:40 +02001776 /*
1777 * We don't want to write the buffer anymore, clear the
1778 * bit so that we don't confuse checks in
1779 * __journal_file_buffer
1780 */
1781 clear_buffer_dirty(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782 __journal_file_buffer(jh, transaction, BJ_Forget);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 may_free = 0;
1784 } else {
1785 JBUFFER_TRACE(jh, "on running transaction");
1786 journal_remove_journal_head(bh);
1787 __brelse(bh);
1788 }
1789 return may_free;
1790}
1791
1792/*
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001793 * journal_invalidatepage
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 *
1795 * This code is tricky. It has a number of cases to deal with.
1796 *
1797 * There are two invariants which this code relies on:
1798 *
1799 * i_size must be updated on disk before we start calling invalidatepage on the
1800 * data.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001801 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802 * This is done in ext3 by defining an ext3_setattr method which
1803 * updates i_size before truncate gets going. By maintaining this
1804 * invariant, we can be sure that it is safe to throw away any buffers
1805 * attached to the current transaction: once the transaction commits,
1806 * we know that the data will not be needed.
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001807 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808 * Note however that we can *not* throw away data belonging to the
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001809 * previous, committing transaction!
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 *
1811 * Any disk blocks which *are* part of the previous, committing
1812 * transaction (and which therefore cannot be discarded immediately) are
1813 * not going to be reused in the new running transaction
1814 *
1815 * The bitmap committed_data images guarantee this: any block which is
1816 * allocated in one transaction and removed in the next will be marked
1817 * as in-use in the committed_data bitmap, so cannot be reused until
1818 * the next transaction to delete the block commits. This means that
1819 * leaving committing buffers dirty is quite safe: the disk blocks
1820 * cannot be reallocated to a different file and so buffer aliasing is
1821 * not possible.
1822 *
1823 *
1824 * The above applies mainly to ordered data mode. In writeback mode we
1825 * don't make guarantees about the order in which data hits disk --- in
1826 * particular we don't guarantee that new dirty data is flushed before
1827 * transaction commit --- so it is always safe just to discard data
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001828 * immediately in that mode. --sct
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 */
1830
1831/*
1832 * The journal_unmap_buffer helper function returns zero if the buffer
1833 * concerned remains pinned as an anonymous buffer belonging to an older
1834 * transaction.
1835 *
1836 * We're outside-transaction here. Either or both of j_running_transaction
1837 * and j_committing_transaction may be NULL.
1838 */
1839static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1840{
1841 transaction_t *transaction;
1842 struct journal_head *jh;
1843 int may_free = 1;
1844 int ret;
1845
1846 BUFFER_TRACE(bh, "entry");
1847
1848 /*
1849 * It is safe to proceed here without the j_list_lock because the
1850 * buffers cannot be stolen by try_to_free_buffers as long as we are
1851 * holding the page lock. --sct
1852 */
1853
1854 if (!buffer_jbd(bh))
1855 goto zap_buffer_unlocked;
1856
1857 spin_lock(&journal->j_state_lock);
1858 jbd_lock_bh_state(bh);
1859 spin_lock(&journal->j_list_lock);
1860
1861 jh = journal_grab_journal_head(bh);
1862 if (!jh)
1863 goto zap_buffer_no_jh;
1864
Jan Kara86963912010-02-16 20:37:12 +01001865 /*
1866 * We cannot remove the buffer from checkpoint lists until the
1867 * transaction adding inode to orphan list (let's call it T)
1868 * is committed. Otherwise if the transaction changing the
1869 * buffer would be cleaned from the journal before T is
1870 * committed, a crash will cause that the correct contents of
1871 * the buffer will be lost. On the other hand we have to
1872 * clear the buffer dirty bit at latest at the moment when the
1873 * transaction marking the buffer as freed in the filesystem
1874 * structures is committed because from that moment on the
1875 * buffer can be reallocated and used by a different page.
1876 * Since the block hasn't been freed yet but the inode has
1877 * already been added to orphan list, it is safe for us to add
1878 * the buffer to BJ_Forget list of the newest transaction.
1879 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880 transaction = jh->b_transaction;
1881 if (transaction == NULL) {
1882 /* First case: not on any transaction. If it
1883 * has no checkpoint link, then we can zap it:
1884 * it's a writeback-mode buffer so we don't care
1885 * if it hits disk safely. */
1886 if (!jh->b_cp_transaction) {
1887 JBUFFER_TRACE(jh, "not on any transaction: zap");
1888 goto zap_buffer;
1889 }
1890
1891 if (!buffer_dirty(bh)) {
1892 /* bdflush has written it. We can drop it now */
1893 goto zap_buffer;
1894 }
1895
1896 /* OK, it must be in the journal but still not
1897 * written fully to disk: it's metadata or
1898 * journaled data... */
1899
1900 if (journal->j_running_transaction) {
1901 /* ... and once the current transaction has
1902 * committed, the buffer won't be needed any
1903 * longer. */
1904 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1905 ret = __dispose_buffer(jh,
1906 journal->j_running_transaction);
1907 journal_put_journal_head(jh);
1908 spin_unlock(&journal->j_list_lock);
1909 jbd_unlock_bh_state(bh);
1910 spin_unlock(&journal->j_state_lock);
1911 return ret;
1912 } else {
1913 /* There is no currently-running transaction. So the
1914 * orphan record which we wrote for this file must have
1915 * passed into commit. We must attach this buffer to
1916 * the committing transaction, if it exists. */
1917 if (journal->j_committing_transaction) {
1918 JBUFFER_TRACE(jh, "give to committing trans");
1919 ret = __dispose_buffer(jh,
1920 journal->j_committing_transaction);
1921 journal_put_journal_head(jh);
1922 spin_unlock(&journal->j_list_lock);
1923 jbd_unlock_bh_state(bh);
1924 spin_unlock(&journal->j_state_lock);
1925 return ret;
1926 } else {
1927 /* The orphan record's transaction has
1928 * committed. We can cleanse this buffer */
1929 clear_buffer_jbddirty(bh);
1930 goto zap_buffer;
1931 }
1932 }
1933 } else if (transaction == journal->j_committing_transaction) {
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001934 JBUFFER_TRACE(jh, "on committing transaction");
akpm@osdl.orgd13df842005-04-16 15:26:36 -07001935 if (jh->b_jlist == BJ_Locked) {
1936 /*
1937 * The buffer is on the committing transaction's locked
1938 * list. We have the buffer locked, so I/O has
1939 * completed. So we can nail the buffer now.
1940 */
1941 may_free = __dispose_buffer(jh, transaction);
1942 goto zap_buffer;
1943 }
1944 /*
Jan Kara86963912010-02-16 20:37:12 +01001945 * The buffer is committing, we simply cannot touch
1946 * it. So we just set j_next_transaction to the
1947 * running transaction (if there is one) and mark
1948 * buffer as freed so that commit code knows it should
1949 * clear dirty bits when it is done with the buffer.
1950 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 set_buffer_freed(bh);
Jan Kara86963912010-02-16 20:37:12 +01001952 if (journal->j_running_transaction && buffer_jbddirty(bh))
1953 jh->b_next_transaction = journal->j_running_transaction;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 journal_put_journal_head(jh);
1955 spin_unlock(&journal->j_list_lock);
1956 jbd_unlock_bh_state(bh);
1957 spin_unlock(&journal->j_state_lock);
1958 return 0;
1959 } else {
1960 /* Good, the buffer belongs to the running transaction.
1961 * We are writing our own transaction's data, not any
1962 * previous one's, so it is safe to throw it away
1963 * (remember that we expect the filesystem to have set
1964 * i_size already for this truncate so recovery will not
1965 * expose the disk blocks we are discarding here.) */
1966 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
Eric Sandeenf58a74d2006-10-28 10:38:27 -07001967 JBUFFER_TRACE(jh, "on running transaction");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968 may_free = __dispose_buffer(jh, transaction);
1969 }
1970
1971zap_buffer:
1972 journal_put_journal_head(jh);
1973zap_buffer_no_jh:
1974 spin_unlock(&journal->j_list_lock);
1975 jbd_unlock_bh_state(bh);
1976 spin_unlock(&journal->j_state_lock);
1977zap_buffer_unlocked:
1978 clear_buffer_dirty(bh);
1979 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1980 clear_buffer_mapped(bh);
1981 clear_buffer_req(bh);
1982 clear_buffer_new(bh);
1983 bh->b_bdev = NULL;
1984 return may_free;
1985}
1986
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001987/**
Randy Dunlapa6b91912008-03-19 17:01:00 -07001988 * void journal_invalidatepage() - invalidate a journal page
1989 * @journal: journal to use for flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 * @page: page to flush
1991 * @offset: length of page to invalidate.
1992 *
1993 * Reap page buffers containing data after offset in page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 */
NeilBrown2ff28e22006-03-26 01:37:18 -08001995void journal_invalidatepage(journal_t *journal,
Mingming Caoae6ddcc2006-09-27 01:49:27 -07001996 struct page *page,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997 unsigned long offset)
1998{
1999 struct buffer_head *head, *bh, *next;
2000 unsigned int curr_off = 0;
2001 int may_free = 1;
2002
2003 if (!PageLocked(page))
2004 BUG();
2005 if (!page_has_buffers(page))
NeilBrown2ff28e22006-03-26 01:37:18 -08002006 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007
2008 /* We will potentially be playing with lists other than just the
2009 * data lists (especially for journaled data mode), so be
2010 * cautious in our locking. */
2011
2012 head = bh = page_buffers(page);
2013 do {
2014 unsigned int next_off = curr_off + bh->b_size;
2015 next = bh->b_this_page;
2016
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017 if (offset <= curr_off) {
Dave Kleikampe9ad5622006-09-27 01:49:35 -07002018 /* This block is wholly outside the truncation point */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 lock_buffer(bh);
2020 may_free &= journal_unmap_buffer(journal, bh);
2021 unlock_buffer(bh);
2022 }
2023 curr_off = next_off;
2024 bh = next;
2025
2026 } while (bh != head);
2027
2028 if (!offset) {
NeilBrown2ff28e22006-03-26 01:37:18 -08002029 if (may_free && try_to_free_buffers(page))
2030 J_ASSERT(!page_has_buffers(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032}
2033
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002034/*
2035 * File a buffer on the given transaction list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 */
2037void __journal_file_buffer(struct journal_head *jh,
2038 transaction_t *transaction, int jlist)
2039{
2040 struct journal_head **list = NULL;
2041 int was_dirty = 0;
2042 struct buffer_head *bh = jh2bh(jh);
2043
2044 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2045 assert_spin_locked(&transaction->t_journal->j_list_lock);
2046
2047 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2048 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07002049 jh->b_transaction == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050
2051 if (jh->b_transaction && jh->b_jlist == jlist)
2052 return;
2053
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002054 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055 jlist == BJ_Shadow || jlist == BJ_Forget) {
Jan Kara1e9fd532009-06-24 17:31:40 +02002056 /*
2057 * For metadata buffers, we track dirty bit in buffer_jbddirty
2058 * instead of buffer_dirty. We should not see a dirty bit set
2059 * here because we clear it in do_get_write_access but e.g.
2060 * tune2fs can modify the sb and set the dirty bit at any time
2061 * so we try to gracefully handle that.
2062 */
2063 if (buffer_dirty(bh))
2064 warn_dirty_buffer(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 if (test_clear_buffer_dirty(bh) ||
2066 test_clear_buffer_jbddirty(bh))
2067 was_dirty = 1;
2068 }
2069
2070 if (jh->b_transaction)
2071 __journal_temp_unlink_buffer(jh);
2072 jh->b_transaction = transaction;
2073
2074 switch (jlist) {
2075 case BJ_None:
2076 J_ASSERT_JH(jh, !jh->b_committed_data);
2077 J_ASSERT_JH(jh, !jh->b_frozen_data);
2078 return;
2079 case BJ_SyncData:
2080 list = &transaction->t_sync_datalist;
2081 break;
2082 case BJ_Metadata:
2083 transaction->t_nr_buffers++;
2084 list = &transaction->t_buffers;
2085 break;
2086 case BJ_Forget:
2087 list = &transaction->t_forget;
2088 break;
2089 case BJ_IO:
2090 list = &transaction->t_iobuf_list;
2091 break;
2092 case BJ_Shadow:
2093 list = &transaction->t_shadow_list;
2094 break;
2095 case BJ_LogCtl:
2096 list = &transaction->t_log_list;
2097 break;
2098 case BJ_Reserved:
2099 list = &transaction->t_reserved_list;
2100 break;
2101 case BJ_Locked:
2102 list = &transaction->t_locked_list;
2103 break;
2104 }
2105
2106 __blist_add_buffer(list, jh);
2107 jh->b_jlist = jlist;
2108
2109 if (was_dirty)
2110 set_buffer_jbddirty(bh);
2111}
2112
2113void journal_file_buffer(struct journal_head *jh,
2114 transaction_t *transaction, int jlist)
2115{
2116 jbd_lock_bh_state(jh2bh(jh));
2117 spin_lock(&transaction->t_journal->j_list_lock);
2118 __journal_file_buffer(jh, transaction, jlist);
2119 spin_unlock(&transaction->t_journal->j_list_lock);
2120 jbd_unlock_bh_state(jh2bh(jh));
2121}
2122
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002123/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124 * Remove a buffer from its current buffer list in preparation for
2125 * dropping it from its current transaction entirely. If the buffer has
2126 * already started to be used by a subsequent transaction, refile the
2127 * buffer on that transaction's metadata list.
2128 *
2129 * Called under journal->j_list_lock
2130 *
2131 * Called under jbd_lock_bh_state(jh2bh(jh))
2132 */
2133void __journal_refile_buffer(struct journal_head *jh)
2134{
Jan Kara86963912010-02-16 20:37:12 +01002135 int was_dirty, jlist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 struct buffer_head *bh = jh2bh(jh);
2137
2138 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2139 if (jh->b_transaction)
2140 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2141
2142 /* If the buffer is now unused, just drop it. */
2143 if (jh->b_next_transaction == NULL) {
2144 __journal_unfile_buffer(jh);
2145 return;
2146 }
2147
2148 /*
2149 * It has been modified by a later transaction: add it to the new
2150 * transaction's metadata list.
2151 */
2152
2153 was_dirty = test_clear_buffer_jbddirty(bh);
2154 __journal_temp_unlink_buffer(jh);
2155 jh->b_transaction = jh->b_next_transaction;
2156 jh->b_next_transaction = NULL;
Jan Kara86963912010-02-16 20:37:12 +01002157 if (buffer_freed(bh))
2158 jlist = BJ_Forget;
2159 else if (jh->b_modified)
2160 jlist = BJ_Metadata;
2161 else
2162 jlist = BJ_Reserved;
2163 __journal_file_buffer(jh, jh->b_transaction, jlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2165
2166 if (was_dirty)
2167 set_buffer_jbddirty(bh);
2168}
2169
2170/*
2171 * For the unlocked version of this call, also make sure that any
2172 * hanging journal_head is cleaned up if necessary.
2173 *
2174 * __journal_refile_buffer is usually called as part of a single locked
2175 * operation on a buffer_head, in which the caller is probably going to
2176 * be hooking the journal_head onto other lists. In that case it is up
2177 * to the caller to remove the journal_head if necessary. For the
2178 * unlocked journal_refile_buffer call, the caller isn't going to be
2179 * doing anything else to the buffer so we need to do the cleanup
Mingming Caoae6ddcc2006-09-27 01:49:27 -07002180 * ourselves to avoid a jh leak.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181 *
2182 * *** The journal_head may be freed by this call! ***
2183 */
2184void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2185{
2186 struct buffer_head *bh = jh2bh(jh);
2187
2188 jbd_lock_bh_state(bh);
2189 spin_lock(&journal->j_list_lock);
2190
2191 __journal_refile_buffer(jh);
2192 jbd_unlock_bh_state(bh);
2193 journal_remove_journal_head(bh);
2194
2195 spin_unlock(&journal->j_list_lock);
2196 __brelse(bh);
2197}