aboutsummaryrefslogtreecommitdiffstats
path: root/fs/jbd/checkpoint.c
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 18:20:36 -0400
committerLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 18:20:36 -0400
commit1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 (patch)
tree0bba044c4ce775e45a88a51686b5d9f90697ea9d /fs/jbd/checkpoint.c
Linux-2.6.12-rc2v2.6.12-rc2
Initial git repository build. I'm not bothering with the full history, even though we have it. We can create a separate "historical" git archive of that later if we want to, and in the meantime it's about 3.2GB when imported into git - space that would just make the early git days unnecessarily complicated, when we don't have a lot of good infrastructure for it. Let it rip!
Diffstat (limited to 'fs/jbd/checkpoint.c')
-rw-r--r--fs/jbd/checkpoint.c636
1 files changed, 636 insertions, 0 deletions
diff --git a/fs/jbd/checkpoint.c b/fs/jbd/checkpoint.c
new file mode 100644
index 000000000000..98d830401c56
--- /dev/null
+++ b/fs/jbd/checkpoint.c
@@ -0,0 +1,636 @@
1/*
2 * linux/fs/checkpoint.c
3 *
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
5 *
6 * Copyright 1999 Red Hat Software --- 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 * Checkpoint routines for the generic filesystem journaling code.
13 * Part of the ext2fs journaling system.
14 *
15 * Checkpointing is the process of ensuring that a section of the log is
16 * committed fully to disk, so that that portion of the log can be
17 * reused.
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
26/*
27 * Unlink a buffer from a transaction.
28 *
29 * Called with j_list_lock held.
30 */
31
32static inline void __buffer_unlink(struct journal_head *jh)
33{
34 transaction_t *transaction;
35
36 transaction = jh->b_cp_transaction;
37 jh->b_cp_transaction = NULL;
38
39 jh->b_cpnext->b_cpprev = jh->b_cpprev;
40 jh->b_cpprev->b_cpnext = jh->b_cpnext;
41 if (transaction->t_checkpoint_list == jh)
42 transaction->t_checkpoint_list = jh->b_cpnext;
43 if (transaction->t_checkpoint_list == jh)
44 transaction->t_checkpoint_list = NULL;
45}
46
47/*
48 * Try to release a checkpointed buffer from its transaction.
49 * Returns 1 if we released it.
50 * Requires j_list_lock
51 * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
52 */
53static int __try_to_free_cp_buf(struct journal_head *jh)
54{
55 int ret = 0;
56 struct buffer_head *bh = jh2bh(jh);
57
58 if (jh->b_jlist == BJ_None && !buffer_locked(bh) && !buffer_dirty(bh)) {
59 JBUFFER_TRACE(jh, "remove from checkpoint list");
60 __journal_remove_checkpoint(jh);
61 jbd_unlock_bh_state(bh);
62 journal_remove_journal_head(bh);
63 BUFFER_TRACE(bh, "release");
64 __brelse(bh);
65 ret = 1;
66 } else {
67 jbd_unlock_bh_state(bh);
68 }
69 return ret;
70}
71
72/*
73 * __log_wait_for_space: wait until there is space in the journal.
74 *
75 * Called under j-state_lock *only*. It will be unlocked if we have to wait
76 * for a checkpoint to free up some space in the log.
77 */
78void __log_wait_for_space(journal_t *journal)
79{
80 int nblocks;
81 assert_spin_locked(&journal->j_state_lock);
82
83 nblocks = jbd_space_needed(journal);
84 while (__log_space_left(journal) < nblocks) {
85 if (journal->j_flags & JFS_ABORT)
86 return;
87 spin_unlock(&journal->j_state_lock);
88 down(&journal->j_checkpoint_sem);
89
90 /*
91 * Test again, another process may have checkpointed while we
92 * were waiting for the checkpoint lock
93 */
94 spin_lock(&journal->j_state_lock);
95 nblocks = jbd_space_needed(journal);
96 if (__log_space_left(journal) < nblocks) {
97 spin_unlock(&journal->j_state_lock);
98 log_do_checkpoint(journal);
99 spin_lock(&journal->j_state_lock);
100 }
101 up(&journal->j_checkpoint_sem);
102 }
103}
104
105/*
106 * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
107 * The caller must restart a list walk. Wait for someone else to run
108 * jbd_unlock_bh_state().
109 */
110static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
111{
112 get_bh(bh);
113 spin_unlock(&journal->j_list_lock);
114 jbd_lock_bh_state(bh);
115 jbd_unlock_bh_state(bh);
116 put_bh(bh);
117}
118
119/*
120 * Clean up a transaction's checkpoint list.
121 *
122 * We wait for any pending IO to complete and make sure any clean
123 * buffers are removed from the transaction.
124 *
125 * Return 1 if we performed any actions which might have destroyed the
126 * checkpoint. (journal_remove_checkpoint() deletes the transaction when
127 * the last checkpoint buffer is cleansed)
128 *
129 * Called with j_list_lock held.
130 */
131static int __cleanup_transaction(journal_t *journal, transaction_t *transaction)
132{
133 struct journal_head *jh, *next_jh, *last_jh;
134 struct buffer_head *bh;
135 int ret = 0;
136
137 assert_spin_locked(&journal->j_list_lock);
138 jh = transaction->t_checkpoint_list;
139 if (!jh)
140 return 0;
141
142 last_jh = jh->b_cpprev;
143 next_jh = jh;
144 do {
145 jh = next_jh;
146 bh = jh2bh(jh);
147 if (buffer_locked(bh)) {
148 atomic_inc(&bh->b_count);
149 spin_unlock(&journal->j_list_lock);
150 wait_on_buffer(bh);
151 /* the journal_head may have gone by now */
152 BUFFER_TRACE(bh, "brelse");
153 __brelse(bh);
154 goto out_return_1;
155 }
156
157 /*
158 * This is foul
159 */
160 if (!jbd_trylock_bh_state(bh)) {
161 jbd_sync_bh(journal, bh);
162 goto out_return_1;
163 }
164
165 if (jh->b_transaction != NULL) {
166 transaction_t *t = jh->b_transaction;
167 tid_t tid = t->t_tid;
168
169 spin_unlock(&journal->j_list_lock);
170 jbd_unlock_bh_state(bh);
171 log_start_commit(journal, tid);
172 log_wait_commit(journal, tid);
173 goto out_return_1;
174 }
175
176 /*
177 * AKPM: I think the buffer_jbddirty test is redundant - it
178 * shouldn't have NULL b_transaction?
179 */
180 next_jh = jh->b_cpnext;
181 if (!buffer_dirty(bh) && !buffer_jbddirty(bh)) {
182 BUFFER_TRACE(bh, "remove from checkpoint");
183 __journal_remove_checkpoint(jh);
184 jbd_unlock_bh_state(bh);
185 journal_remove_journal_head(bh);
186 __brelse(bh);
187 ret = 1;
188 } else {
189 jbd_unlock_bh_state(bh);
190 }
191 jh = next_jh;
192 } while (jh != last_jh);
193
194 return ret;
195out_return_1:
196 spin_lock(&journal->j_list_lock);
197 return 1;
198}
199
200#define NR_BATCH 64
201
202static void
203__flush_batch(journal_t *journal, struct buffer_head **bhs, int *batch_count)
204{
205 int i;
206
207 spin_unlock(&journal->j_list_lock);
208 ll_rw_block(WRITE, *batch_count, bhs);
209 spin_lock(&journal->j_list_lock);
210 for (i = 0; i < *batch_count; i++) {
211 struct buffer_head *bh = bhs[i];
212 clear_buffer_jwrite(bh);
213 BUFFER_TRACE(bh, "brelse");
214 __brelse(bh);
215 }
216 *batch_count = 0;
217}
218
219/*
220 * Try to flush one buffer from the checkpoint list to disk.
221 *
222 * Return 1 if something happened which requires us to abort the current
223 * scan of the checkpoint list.
224 *
225 * Called with j_list_lock held.
226 * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
227 */
228static int __flush_buffer(journal_t *journal, struct journal_head *jh,
229 struct buffer_head **bhs, int *batch_count,
230 int *drop_count)
231{
232 struct buffer_head *bh = jh2bh(jh);
233 int ret = 0;
234
235 if (buffer_dirty(bh) && !buffer_locked(bh) && jh->b_jlist == BJ_None) {
236 J_ASSERT_JH(jh, jh->b_transaction == NULL);
237
238 /*
239 * Important: we are about to write the buffer, and
240 * possibly block, while still holding the journal lock.
241 * We cannot afford to let the transaction logic start
242 * messing around with this buffer before we write it to
243 * disk, as that would break recoverability.
244 */
245 BUFFER_TRACE(bh, "queue");
246 get_bh(bh);
247 J_ASSERT_BH(bh, !buffer_jwrite(bh));
248 set_buffer_jwrite(bh);
249 bhs[*batch_count] = bh;
250 jbd_unlock_bh_state(bh);
251 (*batch_count)++;
252 if (*batch_count == NR_BATCH) {
253 __flush_batch(journal, bhs, batch_count);
254 ret = 1;
255 }
256 } else {
257 int last_buffer = 0;
258 if (jh->b_cpnext == jh) {
259 /* We may be about to drop the transaction. Tell the
260 * caller that the lists have changed.
261 */
262 last_buffer = 1;
263 }
264 if (__try_to_free_cp_buf(jh)) {
265 (*drop_count)++;
266 ret = last_buffer;
267 }
268 }
269 return ret;
270}
271
272/*
273 * Perform an actual checkpoint. We don't write out only enough to
274 * satisfy the current blocked requests: rather we submit a reasonably
275 * sized chunk of the outstanding data to disk at once for
276 * efficiency. __log_wait_for_space() will retry if we didn't free enough.
277 *
278 * However, we _do_ take into account the amount requested so that once
279 * the IO has been queued, we can return as soon as enough of it has
280 * completed to disk.
281 *
282 * The journal should be locked before calling this function.
283 */
284int log_do_checkpoint(journal_t *journal)
285{
286 int result;
287 int batch_count = 0;
288 struct buffer_head *bhs[NR_BATCH];
289
290 jbd_debug(1, "Start checkpoint\n");
291
292 /*
293 * First thing: if there are any transactions in the log which
294 * don't need checkpointing, just eliminate them from the
295 * journal straight away.
296 */
297 result = cleanup_journal_tail(journal);
298 jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
299 if (result <= 0)
300 return result;
301
302 /*
303 * OK, we need to start writing disk blocks. Try to free up a
304 * quarter of the log in a single checkpoint if we can.
305 */
306 /*
307 * AKPM: check this code. I had a feeling a while back that it
308 * degenerates into a busy loop at unmount time.
309 */
310 spin_lock(&journal->j_list_lock);
311 while (journal->j_checkpoint_transactions) {
312 transaction_t *transaction;
313 struct journal_head *jh, *last_jh, *next_jh;
314 int drop_count = 0;
315 int cleanup_ret, retry = 0;
316 tid_t this_tid;
317
318 transaction = journal->j_checkpoint_transactions;
319 this_tid = transaction->t_tid;
320 jh = transaction->t_checkpoint_list;
321 last_jh = jh->b_cpprev;
322 next_jh = jh;
323 do {
324 struct buffer_head *bh;
325
326 jh = next_jh;
327 next_jh = jh->b_cpnext;
328 bh = jh2bh(jh);
329 if (!jbd_trylock_bh_state(bh)) {
330 jbd_sync_bh(journal, bh);
331 spin_lock(&journal->j_list_lock);
332 retry = 1;
333 break;
334 }
335 retry = __flush_buffer(journal, jh, bhs, &batch_count, &drop_count);
336 if (cond_resched_lock(&journal->j_list_lock)) {
337 retry = 1;
338 break;
339 }
340 } while (jh != last_jh && !retry);
341
342 if (batch_count)
343 __flush_batch(journal, bhs, &batch_count);
344
345 /*
346 * If someone cleaned up this transaction while we slept, we're
347 * done
348 */
349 if (journal->j_checkpoint_transactions != transaction)
350 break;
351 if (retry)
352 continue;
353 /*
354 * Maybe it's a new transaction, but it fell at the same
355 * address
356 */
357 if (transaction->t_tid != this_tid)
358 continue;
359 /*
360 * We have walked the whole transaction list without
361 * finding anything to write to disk. We had better be
362 * able to make some progress or we are in trouble.
363 */
364 cleanup_ret = __cleanup_transaction(journal, transaction);
365 J_ASSERT(drop_count != 0 || cleanup_ret != 0);
366 if (journal->j_checkpoint_transactions != transaction)
367 break;
368 }
369 spin_unlock(&journal->j_list_lock);
370 result = cleanup_journal_tail(journal);
371 if (result < 0)
372 return result;
373
374 return 0;
375}
376
377/*
378 * Check the list of checkpoint transactions for the journal to see if
379 * we have already got rid of any since the last update of the log tail
380 * in the journal superblock. If so, we can instantly roll the
381 * superblock forward to remove those transactions from the log.
382 *
383 * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
384 *
385 * Called with the journal lock held.
386 *
387 * This is the only part of the journaling code which really needs to be
388 * aware of transaction aborts. Checkpointing involves writing to the
389 * main filesystem area rather than to the journal, so it can proceed
390 * even in abort state, but we must not update the journal superblock if
391 * we have an abort error outstanding.
392 */
393
394int cleanup_journal_tail(journal_t *journal)
395{
396 transaction_t * transaction;
397 tid_t first_tid;
398 unsigned long blocknr, freed;
399
400 /* OK, work out the oldest transaction remaining in the log, and
401 * the log block it starts at.
402 *
403 * If the log is now empty, we need to work out which is the
404 * next transaction ID we will write, and where it will
405 * start. */
406
407 spin_lock(&journal->j_state_lock);
408 spin_lock(&journal->j_list_lock);
409 transaction = journal->j_checkpoint_transactions;
410 if (transaction) {
411 first_tid = transaction->t_tid;
412 blocknr = transaction->t_log_start;
413 } else if ((transaction = journal->j_committing_transaction) != NULL) {
414 first_tid = transaction->t_tid;
415 blocknr = transaction->t_log_start;
416 } else if ((transaction = journal->j_running_transaction) != NULL) {
417 first_tid = transaction->t_tid;
418 blocknr = journal->j_head;
419 } else {
420 first_tid = journal->j_transaction_sequence;
421 blocknr = journal->j_head;
422 }
423 spin_unlock(&journal->j_list_lock);
424 J_ASSERT(blocknr != 0);
425
426 /* If the oldest pinned transaction is at the tail of the log
427 already then there's not much we can do right now. */
428 if (journal->j_tail_sequence == first_tid) {
429 spin_unlock(&journal->j_state_lock);
430 return 1;
431 }
432
433 /* OK, update the superblock to recover the freed space.
434 * Physical blocks come first: have we wrapped beyond the end of
435 * the log? */
436 freed = blocknr - journal->j_tail;
437 if (blocknr < journal->j_tail)
438 freed = freed + journal->j_last - journal->j_first;
439
440 jbd_debug(1,
441 "Cleaning journal tail from %d to %d (offset %lu), "
442 "freeing %lu\n",
443 journal->j_tail_sequence, first_tid, blocknr, freed);
444
445 journal->j_free += freed;
446 journal->j_tail_sequence = first_tid;
447 journal->j_tail = blocknr;
448 spin_unlock(&journal->j_state_lock);
449 if (!(journal->j_flags & JFS_ABORT))
450 journal_update_superblock(journal, 1);
451 return 0;
452}
453
454
455/* Checkpoint list management */
456
457/*
458 * journal_clean_checkpoint_list
459 *
460 * Find all the written-back checkpoint buffers in the journal and release them.
461 *
462 * Called with the journal locked.
463 * Called with j_list_lock held.
464 * Returns number of bufers reaped (for debug)
465 */
466
467int __journal_clean_checkpoint_list(journal_t *journal)
468{
469 transaction_t *transaction, *last_transaction, *next_transaction;
470 int ret = 0;
471
472 transaction = journal->j_checkpoint_transactions;
473 if (transaction == 0)
474 goto out;
475
476 last_transaction = transaction->t_cpprev;
477 next_transaction = transaction;
478 do {
479 struct journal_head *jh;
480
481 transaction = next_transaction;
482 next_transaction = transaction->t_cpnext;
483 jh = transaction->t_checkpoint_list;
484 if (jh) {
485 struct journal_head *last_jh = jh->b_cpprev;
486 struct journal_head *next_jh = jh;
487
488 do {
489 jh = next_jh;
490 next_jh = jh->b_cpnext;
491 /* Use trylock because of the ranknig */
492 if (jbd_trylock_bh_state(jh2bh(jh)))
493 ret += __try_to_free_cp_buf(jh);
494 /*
495 * This function only frees up some memory
496 * if possible so we dont have an obligation
497 * to finish processing. Bail out if preemption
498 * requested:
499 */
500 if (need_resched())
501 goto out;
502 } while (jh != last_jh);
503 }
504 } while (transaction != last_transaction);
505out:
506 return ret;
507}
508
509/*
510 * journal_remove_checkpoint: called after a buffer has been committed
511 * to disk (either by being write-back flushed to disk, or being
512 * committed to the log).
513 *
514 * We cannot safely clean a transaction out of the log until all of the
515 * buffer updates committed in that transaction have safely been stored
516 * elsewhere on disk. To achieve this, all of the buffers in a
517 * transaction need to be maintained on the transaction's checkpoint
518 * list until they have been rewritten, at which point this function is
519 * called to remove the buffer from the existing transaction's
520 * checkpoint list.
521 *
522 * This function is called with the journal locked.
523 * This function is called with j_list_lock held.
524 */
525
526void __journal_remove_checkpoint(struct journal_head *jh)
527{
528 transaction_t *transaction;
529 journal_t *journal;
530
531 JBUFFER_TRACE(jh, "entry");
532
533 if ((transaction = jh->b_cp_transaction) == NULL) {
534 JBUFFER_TRACE(jh, "not on transaction");
535 goto out;
536 }
537 journal = transaction->t_journal;
538
539 __buffer_unlink(jh);
540
541 if (transaction->t_checkpoint_list != NULL)
542 goto out;
543 JBUFFER_TRACE(jh, "transaction has no more buffers");
544
545 /*
546 * There is one special case to worry about: if we have just pulled the
547 * buffer off a committing transaction's forget list, then even if the
548 * checkpoint list is empty, the transaction obviously cannot be
549 * dropped!
550 *
551 * The locking here around j_committing_transaction is a bit sleazy.
552 * See the comment at the end of journal_commit_transaction().
553 */
554 if (transaction == journal->j_committing_transaction) {
555 JBUFFER_TRACE(jh, "belongs to committing transaction");
556 goto out;
557 }
558
559 /* OK, that was the last buffer for the transaction: we can now
560 safely remove this transaction from the log */
561
562 __journal_drop_transaction(journal, transaction);
563
564 /* Just in case anybody was waiting for more transactions to be
565 checkpointed... */
566 wake_up(&journal->j_wait_logspace);
567out:
568 JBUFFER_TRACE(jh, "exit");
569}
570
571/*
572 * journal_insert_checkpoint: put a committed buffer onto a checkpoint
573 * list so that we know when it is safe to clean the transaction out of
574 * the log.
575 *
576 * Called with the journal locked.
577 * Called with j_list_lock held.
578 */
579void __journal_insert_checkpoint(struct journal_head *jh,
580 transaction_t *transaction)
581{
582 JBUFFER_TRACE(jh, "entry");
583 J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
584 J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
585
586 jh->b_cp_transaction = transaction;
587
588 if (!transaction->t_checkpoint_list) {
589 jh->b_cpnext = jh->b_cpprev = jh;
590 } else {
591 jh->b_cpnext = transaction->t_checkpoint_list;
592 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
593 jh->b_cpprev->b_cpnext = jh;
594 jh->b_cpnext->b_cpprev = jh;
595 }
596 transaction->t_checkpoint_list = jh;
597}
598
599/*
600 * We've finished with this transaction structure: adios...
601 *
602 * The transaction must have no links except for the checkpoint by this
603 * point.
604 *
605 * Called with the journal locked.
606 * Called with j_list_lock held.
607 */
608
609void __journal_drop_transaction(journal_t *journal, transaction_t *transaction)
610{
611 assert_spin_locked(&journal->j_list_lock);
612 if (transaction->t_cpnext) {
613 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
614 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
615 if (journal->j_checkpoint_transactions == transaction)
616 journal->j_checkpoint_transactions =
617 transaction->t_cpnext;
618 if (journal->j_checkpoint_transactions == transaction)
619 journal->j_checkpoint_transactions = NULL;
620 }
621
622 J_ASSERT(transaction->t_state == T_FINISHED);
623 J_ASSERT(transaction->t_buffers == NULL);
624 J_ASSERT(transaction->t_sync_datalist == NULL);
625 J_ASSERT(transaction->t_forget == NULL);
626 J_ASSERT(transaction->t_iobuf_list == NULL);
627 J_ASSERT(transaction->t_shadow_list == NULL);
628 J_ASSERT(transaction->t_log_list == NULL);
629 J_ASSERT(transaction->t_checkpoint_list == NULL);
630 J_ASSERT(transaction->t_updates == 0);
631 J_ASSERT(journal->j_committing_transaction != transaction);
632 J_ASSERT(journal->j_running_transaction != transaction);
633
634 jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
635 kfree(transaction);
636}