1/*
2 * linux/fs/jbd/transaction.c
3 *
4 * 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
13 * journaling system.
14 *
15 * This file manages transactions (compound commits managed by the
16 * journaling code) and handles (individual atomic operations by the
17 * filesystem).
18 */
19
20#include <linux/time.h>
21#include <linux/fs.h>
22#include <linux/jbd.h>
23#include <linux/errno.h>
24#include <linux/slab.h>
25#include <linux/timer.h>
26#include <linux/mm.h>
27#include <linux/highmem.h>
28#include <linux/hrtimer.h>
29
30static void __journal_temp_unlink_buffer(struct journal_head *jh);
31
32/*
33 * get_transaction: obtain a new transaction_t object.
34 *
35 * Simply allocate and initialise a new transaction.  Create it in
36 * RUNNING state and add it to the current journal (which should not
37 * have an existing running transaction: we only make a new transaction
38 * once we have started to commit the old one).
39 *
40 * Preconditions:
41 *	The journal MUST be locked.  We don't perform atomic mallocs on the
42 *	new transaction	and we can't block without protecting against other
43 *	processes trying to touch the journal while it is in transition.
44 *
45 * Called under j_state_lock
46 */
47
48static transaction_t *
49get_transaction(journal_t *journal, transaction_t *transaction)
50{
51	transaction->t_journal = journal;
52	transaction->t_state = T_RUNNING;
53	transaction->t_start_time = ktime_get();
54	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. */
59	journal->j_commit_timer.expires =
60				round_jiffies_up(transaction->t_expires);
61	add_timer(&journal->j_commit_timer);
62
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
81 * transaction's buffer credits.
82 */
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) {
102		new_transaction = kzalloc(sizeof(*new_transaction),
103						GFP_NOFS|__GFP_NOFAIL);
104		if (!new_transaction) {
105			ret = -ENOMEM;
106			goto out;
107		}
108	}
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);
123		ret = -EROFS;
124		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
188	/*
189	 * 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
197	 * in the new transaction.
198	 *
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.
210	 * Also, this test is inconsistent with the matching one in
211	 * 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);
232
233	lock_map_acquire(&handle->h_lockdep_map);
234out:
235	if (unlikely(new_transaction))		/* It's usually NULL */
236		kfree(new_transaction);
237	return ret;
238}
239
240static struct lock_class_key jbd_handle_key;
241
242/* 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	handle->h_buffer_credits = nblocks;
249	handle->h_ref = 1;
250
251	lockdep_init_map(&handle->h_lockdep_map, "jbd_handle", &jbd_handle_key, 0);
252
253	return handle;
254}
255
256/**
257 * handle_t *journal_start() - Obtain a new handle.
258 * @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
263 * that much space.
264 *
265 * This function is visible to journal users (like ext3fs), so is not
266 * called with the journal already locked.
267 *
268 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
269 * 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.
304 *
305 * 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
308 * extend its credit if it needs more.
309 *
310 * journal_extend tries to give the running handle more buffer credits.
311 * It does not guarantee that allocation - this is a best-effort only.
312 * The calling process MUST be able to deal cleanly with a failure to
313 * extend here.
314 *
315 * Return 0 on success, non-zero on failure.
316 *
317 * return code < 0 implies an error
318 * return code > 0 implies normal transaction-full status.
319 */
320int journal_extend(handle_t *handle, int nblocks)
321{
322	transaction_t *transaction = handle->h_transaction;
323	journal_t *journal = transaction->t_journal;
324	int result;
325	int wanted;
326
327	result = -EIO;
328	if (is_handle_aborted(handle))
329		goto out;
330
331	result = 1;
332
333	spin_lock(&journal->j_state_lock);
334
335	/* Don't extend a locked-down transaction! */
336	if (handle->h_transaction->t_state != T_RUNNING) {
337		jbd_debug(3, "denied handle %p %d blocks: "
338			  "transaction not running\n", handle, nblocks);
339		goto error_out;
340	}
341
342	spin_lock(&transaction->t_handle_lock);
343	wanted = transaction->t_outstanding_credits + nblocks;
344
345	if (wanted > journal->j_max_transaction_buffers) {
346		jbd_debug(3, "denied handle %p %d blocks: "
347			  "transaction too large\n", handle, nblocks);
348		goto unlock;
349	}
350
351	if (wanted > __log_space_left(journal)) {
352		jbd_debug(3, "denied handle %p %d blocks: "
353			  "insufficient log space\n", handle, nblocks);
354		goto unlock;
355	}
356
357	handle->h_buffer_credits += nblocks;
358	transaction->t_outstanding_credits += nblocks;
359	result = 0;
360
361	jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
362unlock:
363	spin_unlock(&transaction->t_handle_lock);
364error_out:
365	spin_unlock(&journal->j_state_lock);
366out:
367	return result;
368}
369
370
371/**
372 * int journal_restart() - restart a handle.
373 * @handle:  handle to restart
374 * @nblocks: nr credits requested
375 *
376 * Restart a handle for a multi-transaction filesystem
377 * operation.
378 *
379 * If the journal_extend() call above fails to grant new buffer credits
380 * to a running handle, a call to journal_restart will commit the
381 * handle's transaction so far and reattach the handle to a new
382 * transaction capabable of guaranteeing the requested number of
383 * credits.
384 */
385
386int journal_restart(handle_t *handle, int nblocks)
387{
388	transaction_t *transaction = handle->h_transaction;
389	journal_t *journal = transaction->t_journal;
390	int ret;
391
392	/* If we've had an abort of any type, don't even think about
393	 * actually doing the restart! */
394	if (is_handle_aborted(handle))
395		return 0;
396
397	/*
398	 * First unlink the handle from its current transaction, and start the
399	 * commit on that.
400	 */
401	J_ASSERT(transaction->t_updates > 0);
402	J_ASSERT(journal_current_handle() == handle);
403
404	spin_lock(&journal->j_state_lock);
405	spin_lock(&transaction->t_handle_lock);
406	transaction->t_outstanding_credits -= handle->h_buffer_credits;
407	transaction->t_updates--;
408
409	if (!transaction->t_updates)
410		wake_up(&journal->j_wait_updates);
411	spin_unlock(&transaction->t_handle_lock);
412
413	jbd_debug(2, "restarting handle %p\n", handle);
414	__log_start_commit(journal, transaction->t_tid);
415	spin_unlock(&journal->j_state_lock);
416
417	lock_map_release(&handle->h_lockdep_map);
418	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 until all
429 * existing updates have completed, returning only once the journal is in a
430 * quiescent state with no updates running.
431 *
432 * 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.
437 */
438void journal_lock_updates(journal_t *journal)
439{
440	DEFINE_WAIT(wait);
441
442wait:
443	/* Wait for previous locked operation to finish */
444	wait_event(journal->j_wait_transaction_locked,
445		   journal->j_barrier_count == 0);
446
447	spin_lock(&journal->j_state_lock);
448	/*
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	}
456	++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);
479}
480
481/**
482 * void journal_unlock_updates (journal_t* journal) - release barrier
483 * @journal:  Journal to release the barrier on.
484 *
485 * Release a transaction barrier obtained with journal_lock_updates().
486 */
487void journal_unlock_updates (journal_t *journal)
488{
489	J_ASSERT(journal->j_barrier_count != 0);
490
491	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
497static void warn_dirty_buffer(struct buffer_head *bh)
498{
499	char b[BDEVNAME_SIZE];
500
501	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);
506}
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
535	jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
536
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
547	 * state.  Is the buffer dirty?
548	 *
549	 * 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,
566				jh->b_transaction == transaction ||
567				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);
572			warn_dirty_buffer(bh);
573		}
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		 */
579		JBUFFER_TRACE(jh, "Journalling dirty buffer");
580		clear_buffer_dirty(bh);
581		set_buffer_jbddirty(bh);
582	}
583
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	/*
602	 * this is the first time this transaction is touching this buffer,
603	 * reset the modified flag
604	 */
605	jh->b_modified = 0;
606
607	/*
608	 * 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
661		 * list so won't be flushed.
662		 *
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);
674				frozen_buffer =
675					jbd_alloc(jh2bh(jh)->b_size,
676							 GFP_NOFS);
677				if (!frozen_buffer) {
678					printk(KERN_ERR
679					       "%s: OOM for frozen_buffer\n",
680					       __func__);
681					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);
704		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;
719		offset = offset_in_page(jh2bh(jh)->b_data);
720		source = kmap_atomic(page);
721		memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
722		kunmap_atomic(source);
723	}
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:
733	if (unlikely(frozen_buffer))	/* It's usually NULL */
734		jbd_free(frozen_buffer, bh->b_size);
735
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
744 *
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
771 * the bh is not already part of an existing transaction.
772 *
773 * 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 */
784int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
785{
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) {
816		/*
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));
825
826		/* first access by this transaction */
827		jh->b_modified = 0;
828
829		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) {
832		/* first access by this transaction */
833		jh->b_modified = 0;
834
835		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);
850out:
851	journal_put_journal_head(jh);
852	return err;
853}
854
855/**
856 * int journal_get_undo_access() - Notify intent to modify metadata with non-rewindable consequences
857 * @handle: transaction
858 * @bh: buffer to undo
859 *
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.
866 *
867 * 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.
872 *
873 * 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) {
899		committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
900		if (!committed_data) {
901			printk(KERN_ERR "%s: No memory for committed data\n",
902				__func__);
903			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);
925	if (unlikely(committed_data))
926		jbd_free(committed_data, bh->b_size);
927	return err;
928}
929
930/**
931 * int journal_dirty_data() - mark a buffer as containing dirty data to be flushed
932 * @handle: transaction
933 * @bh: bufferhead to mark
934 *
935 * Description:
936 * Mark a buffer as containing dirty data which needs to be flushed before
937 * we can commit the current transaction.
938 *
939 * 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;
952	int ret = 0;
953
954	if (is_handle_aborted(handle))
955		return ret;
956
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?
968	 *
969	 * 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
972	 *    it, then reallocated it as data.
973	 * 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
976	 *    being sure that the *previous* transaction has this same
977	 *    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);
989
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
996	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);
1057				/* Since we dropped the lock... */
1058				if (!buffer_mapped(bh)) {
1059					JBUFFER_TRACE(jh, "buffer got unmapped");
1060					goto no_journal;
1061				}
1062				/* The buffer may become locked again at any
1063				   time if it is redirtied */
1064			}
1065
1066			/*
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			}
1075			/* We might have slept so buffer could be refiled now */
1076			if (jh->b_transaction != NULL &&
1077			    jh->b_transaction != handle->h_transaction) {
1078				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);
1098			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);
1115	return ret;
1116}
1117
1118/**
1119 * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1120 * @handle: transaction to add buffer to.
1121 * @bh: buffer to mark
1122 *
1123 * Mark dirty metadata which needs to be journaled as part of the current
1124 * transaction.
1125 *
1126 * The buffer is placed on the transaction's metadata list and is marked
1127 * as belonging to the transaction.
1128 *
1129 * Returns error number or 0 on success.
1130 *
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
1177	/*
1178	 * 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:
1181	 * leave it alone for now.
1182	 */
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: */
1194	J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1195
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
1207/*
1208 * 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
1218/**
1219 * 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
1225 * can safely unlink it.
1226 *
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.
1231 *
1232 * 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;
1242	int was_modified = 0;
1243
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
1261	/* keep track of whether or not this transaction modified us */
1262	was_modified = jh->b_modified;
1263
1264	/*
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
1281		/*
1282		 * we only want to drop a reference if this transaction
1283		 * modified the buffer
1284		 */
1285		if (was_modified)
1286			drop_reserve = 1;
1287
1288		/*
1289		 * 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
1297		 * we know to remove the checkpoint after we commit.
1298		 */
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);
1305			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) {
1313		J_ASSERT_JH(jh, (jh->b_transaction ==
1314				 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;
1324
1325			/*
1326			 * only drop a reference if this transaction modified
1327			 * the buffer
1328			 */
1329			if (was_modified)
1330				drop_reserve = 1;
1331		}
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.
1349 *
1350 * 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
1358 * do so in unusual circumstances.  In particular, expect it to
1359 * 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;
1366	int err;
1367	pid_t pid;
1368
1369	J_ASSERT(journal_current_handle() == handle);
1370
1371	if (is_handle_aborted(handle))
1372		err = -EIO;
1373	else {
1374		J_ASSERT(transaction->t_updates > 0);
1375		err = 0;
1376	}
1377
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...
1394	 *
1395	 * We try and optimize the sleep time against what the underlying disk
1396	 * can do, instead of having a static sleep time.  This is useful for
1397	 * 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
1400	 * join the transaction.  We achieve this by measuring how long it takes
1401	 * 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	 *
1406	 * 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.
1410	 */
1411	pid = current->pid;
1412	if (handle->h_sync && journal->j_last_sync_writer != pid) {
1413		u64 commit_time, trans_time;
1414
1415		journal->j_last_sync_writer = pid;
1416
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		}
1433	}
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 ||
1455			time_after_eq(jiffies, transaction->t_expires)) {
1456		/* 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
1470		 * to wait for the commit to complete.
1471		 */
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
1479	lock_map_release(&handle->h_lockdep_map);
1480
1481	jbd_free_handle(handle);
1482	return err;
1483}
1484
1485/**
1486 * int journal_force_commit() - force any uncommitted transactions
1487 * @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
1524static inline void
1525__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
1539/*
1540 * 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
1560/*
1561 * 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 */
1571static void __journal_temp_unlink_buffer(struct journal_head *jh)
1572{
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)
1584		J_ASSERT_JH(jh, transaction != NULL);
1585
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
1623/*
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 */
1630void __journal_unfile_buffer(struct journal_head *jh)
1631{
1632	__journal_temp_unlink_buffer(jh);
1633	jh->b_transaction = NULL;
1634	journal_put_journal_head(jh);
1635}
1636
1637void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1638{
1639	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);
1644	spin_lock(&journal->j_list_lock);
1645	__journal_unfile_buffer(jh);
1646	spin_unlock(&journal->j_list_lock);
1647	jbd_unlock_bh_state(bh);
1648	__brelse(bh);
1649}
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
1666	if (jh->b_next_transaction != NULL)
1667		goto out;
1668
1669	spin_lock(&journal->j_list_lock);
1670	if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
1671		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);
1675		}
1676	} else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1677		/* 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);
1681		}
1682	}
1683	spin_unlock(&journal->j_list_lock);
1684out:
1685	return;
1686}
1687
1688/**
1689 * int journal_try_to_free_buffers() - try to free page buffers.
1690 * @journal: journal for operation
1691 * @page: to try and free
1692 * @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.
1695 *
1696 *
1697 * 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.
1700 *
1701 * 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?
1723 *
1724 * Return 0 on failure, 1 on success
1725 */
1726int journal_try_to_free_buffers(journal_t *journal,
1727				struct page *page, gfp_t gfp_mask)
1728{
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
1743		 * journal_put_journal_head().
1744		 */
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);
1756
1757	ret = try_to_free_buffers(page);
1758
1759busy:
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
1780	if (jh->b_cp_transaction) {
1781		JBUFFER_TRACE(jh, "on running+cp transaction");
1782		__journal_temp_unlink_buffer(jh);
1783		/*
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);
1789		__journal_file_buffer(jh, transaction, BJ_Forget);
1790		may_free = 0;
1791	} else {
1792		JBUFFER_TRACE(jh, "on running transaction");
1793		__journal_unfile_buffer(jh);
1794	}
1795	return may_free;
1796}
1797
1798/*
1799 * journal_invalidatepage
1800 *
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.
1807 *
1808 *  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.
1813 *
1814 *  Note however that we can *not* throw away data belonging to the
1815 *  previous, committing transaction!
1816 *
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
1834 * immediately in that mode.  --sct
1835 */
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 */
1845static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
1846				int partial_page)
1847{
1848	transaction_t *transaction;
1849	struct journal_head *jh;
1850	int may_free = 1;
1851
1852	BUFFER_TRACE(bh, "entry");
1853
1854retry:
1855	/*
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
1872	/*
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
1882	 * block can be reallocated and used by a different page.
1883	 * 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.
1886	 *
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.
1894	 */
1895	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");
1920			may_free = __dispose_buffer(jh,
1921					journal->j_running_transaction);
1922			goto zap_buffer;
1923		} 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");
1930				may_free = __dispose_buffer(jh,
1931					journal->j_committing_transaction);
1932				goto zap_buffer;
1933			} 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) {
1941		JBUFFER_TRACE(jh, "on committing transaction");
1942		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		/*
1952		 * The buffer is committing, we simply cannot touch
1953		 * 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);
1963			unlock_buffer(bh);
1964			log_wait_commit(journal, tid);
1965			lock_buffer(bh);
1966			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.
1973		 */
1974		set_buffer_freed(bh);
1975		if (journal->j_running_transaction && buffer_jbddirty(bh))
1976			jh->b_next_transaction = journal->j_running_transaction;
1977		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);
1990		JBUFFER_TRACE(jh, "on running transaction");
1991		may_free = __dispose_buffer(jh, transaction);
1992	}
1993
1994zap_buffer:
1995	/*
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;
2003	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
2018/**
2019 * void journal_invalidatepage() - invalidate a journal page
2020 * @journal: journal to use for flush
2021 * @page:    page to flush
2022 * @offset:  offset of the range to invalidate
2023 * @length:  length of the range to invalidate
2024 *
2025 * Reap page buffers containing data in specified range in page.
2026 */
2027void journal_invalidatepage(journal_t *journal,
2028		      struct page *page,
2029		      unsigned int offset,
2030		      unsigned int length)
2031{
2032	struct buffer_head *head, *bh, *next;
2033	unsigned int stop = offset + length;
2034	unsigned int curr_off = 0;
2035	int partial_page = (offset || length < PAGE_CACHE_SIZE);
2036	int may_free = 1;
2037
2038	if (!PageLocked(page))
2039		BUG();
2040	if (!page_has_buffers(page))
2041		return;
2042
2043	BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
2044
2045	/* 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
2054		if (next_off > stop)
2055			return;
2056
2057		if (offset <= curr_off) {
2058			/* This block is wholly outside the truncation point */
2059			lock_buffer(bh);
2060			may_free &= journal_unmap_buffer(journal, bh,
2061							 partial_page);
2062			unlock_buffer(bh);
2063		}
2064		curr_off = next_off;
2065		bh = next;
2066
2067	} while (bh != head);
2068
2069	if (!partial_page) {
2070		if (may_free && try_to_free_buffers(page))
2071			J_ASSERT(!page_has_buffers(page));
2072	}
2073}
2074
2075/*
2076 * File a buffer on the given transaction list.
2077 */
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 ||
2090				jh->b_transaction == NULL);
2091
2092	if (jh->b_transaction && jh->b_jlist == jlist)
2093		return;
2094
2095	if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2096	    jlist == BJ_Shadow || jlist == BJ_Forget) {
2097		/*
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);
2106		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);
2113	else
2114		journal_grab_journal_head(bh);
2115	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
2166/*
2167 * 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 *
2172 * Called under j_list_lock
2173 * Called under jbd_lock_bh_state(jh2bh(jh))
2174 *
2175 * jh and bh may be already free when this function returns
2176 */
2177void __journal_refile_buffer(struct journal_head *jh)
2178{
2179	int was_dirty, jlist;
2180	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);
2199	/*
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	 */
2204	jh->b_transaction = jh->b_next_transaction;
2205	jh->b_next_transaction = NULL;
2206	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);
2213	J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2214
2215	if (was_dirty)
2216		set_buffer_jbddirty(bh);
2217}
2218
2219/*
2220 * __journal_refile_buffer() with necessary locking added. We take our bh
2221 * reference so that we can safely unlock bh.
2222 *
2223 * The jh and bh may be freed by this call.
2224 */
2225void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2226{
2227	struct buffer_head *bh = jh2bh(jh);
2228
2229	/* Get reference so that buffer cannot be freed before we unlock it */
2230	get_bh(bh);
2231	jbd_lock_bh_state(bh);
2232	spin_lock(&journal->j_list_lock);
2233	__journal_refile_buffer(jh);
2234	jbd_unlock_bh_state(bh);
2235	spin_unlock(&journal->j_list_lock);
2236	__brelse(bh);
2237}
2238