| blob | bb4c1e240d1740a0b982c916b717d440f13bd506 |
1 Add support for tracking metadata blocks in the log.
3 From: Abutalib Aghayev <agayev@cs.cmu.edu>
5 This patch adds two important data structures, jmap and transaction_infos,
6 and supporting functions. Jmap is a map from a metadata block number to
7 the log block number. When a transaction commits, jmap is updated with new
8 mappings; when a block is revoked, the mapping for the block is removed
9 from the jmap. Transaction_infos is an array of transaction_info
10 structures that contain information about transactions currently present in
11 the log. It contains a linked list of live blocks in a transaction, and it
12 is updated after every commit to keep the list up-to-date.
13 Transaction_infos array will be used by the cleaner for identifying live
14 blocks and migrating them to appropriate location.
16 [ Modified by tytso to conditionalize changes on the JBD2_LAZY journal flag ]
18 Signed-off-by: Abutalib Aghayev <agayev@cs.cmu.edu>
19 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
20 ---
21 fs/jbd2/Makefile | 3 +-
22 fs/jbd2/commit.c | 23 ++++
23 fs/jbd2/jmap.c | 456 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
24 fs/jbd2/journal.c | 17 ++-
25 include/linux/jbd2.h | 14 +++
26 include/linux/jmap.h | 131 +++++++++++++++++++++
27 include/trace/events/jbd2.h | 193 +++++++++++++++++++++++++++++++
28 7 files changed, 832 insertions(+), 5 deletions(-)
30 diff --git a/fs/jbd2/Makefile b/fs/jbd2/Makefile
31 index 802a3413872a..a54f50b3a06e 100644
32 --- a/fs/jbd2/Makefile
33 +++ b/fs/jbd2/Makefile
34 @@ -4,4 +4,5 @@
36 obj-$(CONFIG_JBD2) += jbd2.o
38 -jbd2-objs := transaction.o commit.o recovery.o checkpoint.o revoke.o journal.o
39 +jbd2-objs := transaction.o commit.o recovery.o checkpoint.o revoke.o journal.o \
40 + jmap.o
41 diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c
42 index 8c514367ba5a..50e1a0b375c5 100644
43 --- a/fs/jbd2/commit.c
44 +++ b/fs/jbd2/commit.c
45 @@ -362,6 +362,8 @@ void jbd2_journal_commit_transaction(journal_t *journal)
46 int flags;
47 int err;
48 unsigned long long blocknr;
49 + struct blk_mapping *mappings = NULL;
50 + struct blk_mapping *map_ptr = NULL;
51 ktime_t start_time;
52 u64 commit_time;
53 char *tagp = NULL;
54 @@ -563,6 +565,14 @@ void jbd2_journal_commit_transaction(journal_t *journal)
55 J_ASSERT(commit_transaction->t_nr_buffers <=
56 atomic_read(&commit_transaction->t_outstanding_credits));
58 + if (journal->j_flags & JBD2_LAZY) {
59 + int nr_mappings = commit_transaction->t_nr_buffers;
60 +
61 + map_ptr = mappings = kmalloc(sizeof(*mappings) * nr_mappings, GFP_NOFS);
62 + if (!mappings)
63 + jbd2_journal_abort(journal, -ENOMEM);
64 + }
65 +
66 err = 0;
67 bufs = 0;
68 descriptor = NULL;
69 @@ -661,6 +671,11 @@ void jbd2_journal_commit_transaction(journal_t *journal)
70 continue;
71 }
72 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
73 + if (map_ptr) {
74 + map_ptr->fsblk = jh2bh(jh)->b_blocknr;
75 + map_ptr->logblk = blocknr;
76 + map_ptr++;
77 + }
79 /* Record the new block's tag in the current descriptor
80 buffer */
81 @@ -895,6 +910,14 @@ void jbd2_journal_commit_transaction(journal_t *journal)
82 transaction can be removed from any checkpoint list it was on
83 before. */
85 + if (mappings) {
86 + err = jbd2_transaction_infos_add(journal, commit_transaction,
87 + mappings, map_ptr - mappings);
88 + if (err)
89 + jbd2_journal_abort(journal, -ENOMEM);
90 + kfree(mappings);
91 + }
92 +
93 jbd_debug(3, "JBD2: commit phase 6\n");
95 J_ASSERT(list_empty(&commit_transaction->t_inode_list));
96 diff --git a/fs/jbd2/jmap.c b/fs/jbd2/jmap.c
97 new file mode 100644
98 index 000000000000..31d143ea28ee
99 --- /dev/null
100 +++ b/fs/jbd2/jmap.c
101 @@ -0,0 +1,456 @@
102 +#include <linux/blk_types.h>
103 +#include <linux/jbd2.h>
104 +#include <linux/jmap.h>
105 +#include <trace/events/jbd2.h>
106 +
107 +static struct kmem_cache *jbd2_jmap_cache;
108 +
109 +int jbd2_journal_init_jmap_cache(void)
110 +{
111 + jbd2_jmap_cache = KMEM_CACHE(jmap_entry, SLAB_RECLAIM_ACCOUNT);
112 + if (!jbd2_jmap_cache)
113 + return -ENOMEM;
114 + return 0;
115 +}
116 +
117 +void jbd2_journal_destroy_jmap_cache(void)
118 +{
119 + kmem_cache_destroy(jbd2_jmap_cache);
120 + jbd2_jmap_cache = NULL;
121 +}
122 +
123 +/*
124 + * Allocate an array of transaction_info structures and initialize the list
125 + * heads inside them.
126 + */
127 +int jbd2_init_transaction_infos(journal_t *journal)
128 +{
129 + int i;
130 + struct transaction_infos *tis = kzalloc(sizeof(*tis), GFP_KERNEL);
131 + if (!tis)
132 + return -ENOMEM;
133 +
134 + tis->buf = kzalloc(sizeof(*tis->buf) * MAX_LIVE_TRANSACTIONS,
135 + GFP_KERNEL);
136 + if (!tis->buf) {
137 + kfree(tis);
138 + return -ENOMEM;
139 + }
140 +
141 + for (i = 0; i < MAX_LIVE_TRANSACTIONS; ++i)
142 + INIT_LIST_HEAD(&tis->buf[i].live_logblks);
143 +
144 + journal->j_transaction_infos = tis;
145 + return 0;
146 +}
147 +
148 +/*
149 + * Free the array of transaction_info structures.
150 + */
151 +void jbd2_free_transaction_infos(journal_t *journal)
152 +{
153 + struct transaction_infos *tis = journal->j_transaction_infos;
154 + if (!tis)
155 + return;
156 + kfree(tis->buf);
157 + kfree(tis);
158 +}
159 +
160 +/*
161 + * Fill an entry to be stored in jmap.
162 + */
163 +static void fill_entry(struct jmap_entry *entry, struct blk_mapping *mapping,
164 + int t_idx, struct list_head *list)
165 +{
166 + entry->mapping = *mapping;
167 + entry->fsblk_last_modified = jiffies;
168 + entry->t_idx = t_idx;
169 + list_add(&entry->list, list);
170 +}
171 +
172 +/*
173 + * A helper function for jbd2_transaction_infos_add. Scans through the mappings
174 + * array, dropping revoked entries from jmap and updating existing entries.
175 + * Moves the new mappings to the beginning of the mappings array and returns the
176 + * number of new mappings. Should be called with a write lock on j_jmap_lock.
177 + */
178 +static int process_existing_mappings(journal_t *journal,
179 + struct transaction_info *ti, int t_idx,
180 + struct blk_mapping *mappings, int nr_mappings)
181 +{
182 + struct jmap_entry *je;
183 + int i, nr_new = 0;
184 +
185 + for (i = 0; i < nr_mappings; ++i) {
186 + je = jbd2_jmap_lookup(journal, mappings[i].fsblk, __func__);
187 + if (!je) {
188 + mappings[nr_new++] = mappings[i];
189 + continue;
190 + }
191 + /*
192 + * We are either deleting the entry because it was revoked, or
193 + * we are moving it to the live blocks list of this transaction.
194 + * In either case, we remove it from its existing list.
195 + */
196 + list_del(&je->list);
197 +
198 + if (je->revoked) {
199 + rb_erase(&je->rb_node, &journal->j_jmap);
200 + kmem_cache_free(jbd2_jmap_cache, je);
201 + } else {
202 + trace_jbd2_jmap_replace(je, &mappings[i], t_idx);
203 + fill_entry(je, &mappings[i], t_idx, &ti->live_logblks);
204 + }
205 + }
206 + return nr_new;
207 +}
208 +
209 +/*
210 + * A helper function for jbd2_transaction_infos_add. Allocates an array of
211 + * jmap_entry structures and returns the pointer to array if successful.
212 + * Otherwise, returns NULL.
213 + */
214 +static struct jmap_entry **alloc_jmap_entries(int nr_entries)
215 +{
216 + struct jmap_entry **jmap_entries;
217 + int i;
218 +
219 + jmap_entries = kmalloc(sizeof(struct jmap_entry *) * nr_entries,
220 + GFP_NOFS);
221 + if (!jmap_entries)
222 + return NULL;
223 +
224 + for (i = 0; i < nr_entries; i++) {
225 + jmap_entries[i] = kmem_cache_zalloc(jbd2_jmap_cache, GFP_NOFS);
226 + if (!jmap_entries[i])
227 + goto out_err;
228 + }
229 + return jmap_entries;
230 +
231 +out_err:
232 + for (i = 0; i < nr_entries && jmap_entries[i]; ++i)
233 + kmem_cache_free(jbd2_jmap_cache, jmap_entries[i]);
234 + kfree(jmap_entries);
235 + return NULL;
236 +}
237 +
238 +/*
239 + * A helper function for jbd2_transaction_infos_add. Adds new mappings to jmap
240 + * and updates the linked list of live logblks of the new transaction. Should
241 + * be called with write lock on j_jmap_lock.
242 + */
243 +static void add_new_mappings(journal_t *journal, struct transaction_info *ti,
244 + int t_idx, struct blk_mapping *mappings,
245 + struct jmap_entry **new_entries, int nr_new)
246 +{
247 + struct rb_node **p;
248 + struct rb_node *parent = NULL;
249 + struct jmap_entry *je;
250 + int i;
251 +
252 + for (i = 0; i < nr_new; ++i) {
253 + p = &journal->j_jmap.rb_node;
254 + while (*p) {
255 + parent = *p;
256 + je = rb_entry(parent, struct jmap_entry, rb_node);
257 +
258 + if (mappings[i].fsblk < je->mapping.fsblk)
259 + p = &(*p)->rb_left;
260 + else if (mappings[i].fsblk > je->mapping.fsblk)
261 + p = &(*p)->rb_right;
262 + else
263 + BUG_ON(1);
264 + }
265 + fill_entry(new_entries[i], &mappings[i], t_idx,
266 + &ti->live_logblks);
267 + rb_link_node(&new_entries[i]->rb_node, parent, p);
268 + rb_insert_color(&new_entries[i]->rb_node, &journal->j_jmap);
269 + trace_jbd2_jmap_insert(&mappings[i], t_idx);
270 + }
271 +}
272 +
273 +/*
274 + * This function is called after a transaction commits. It adds new
275 + * transaction_info structure to transaction_infos and populates jmap map with
276 + * the new mappings that are part of the committed transaction. It also adds
277 + * all the mappings to the linked list that is part of the transaction_info
278 + * structure.
279 + */
280 +int jbd2_transaction_infos_add(journal_t *journal, transaction_t *transaction,
281 + struct blk_mapping *mappings, int nr_mappings)
282 +{
283 + struct transaction_infos *tis = journal->j_transaction_infos;
284 + int t_idx = tis->head;
285 + struct transaction_info *ti = &tis->buf[t_idx];
286 + struct jmap_entry **new_entries = NULL;
287 + int nr_new = 0;
288 +
289 + /*
290 + * We are possibly reusing space of an old transaction_info. The old
291 + * transaction should not have any live blocks in it.
292 + */
293 + BUG_ON(!list_empty(&ti->live_logblks));
294 +
295 + write_lock(&journal->j_jmap_lock);
296 + nr_new = process_existing_mappings(journal, ti, t_idx, mappings,
297 + nr_mappings);
298 + write_unlock(&journal->j_jmap_lock);
299 +
300 + if (nr_new == 0)
301 + goto move_head;
302 +
303 + new_entries = alloc_jmap_entries(nr_new);
304 + if (!new_entries)
305 + return -ENOMEM;
306 +
307 + write_lock(&journal->j_jmap_lock);
308 + add_new_mappings(journal, ti, t_idx, mappings, new_entries, nr_new);
309 + write_unlock(&journal->j_jmap_lock);
310 +
311 + kfree(new_entries);
312 +
313 +move_head:
314 + write_lock(&journal->j_jmap_lock);
315 + ti->tid = transaction->t_tid;
316 + ti->offset = transaction->t_log_start;
317 + tis->head = (tis->head + 1) & (MAX_LIVE_TRANSACTIONS - 1);
318 + write_unlock(&journal->j_jmap_lock);
319 +
320 + trace_jbd2_transaction_infos_add(t_idx, ti, nr_mappings);
321 + return 0;
322 +}
323 +
324 +/*
325 + * Look up fsblk in the jmap and return the corresponding jmap entry if found.
326 + * Should be called with a read lock on j_jmap_lock.
327 + */
328 +struct jmap_entry *jbd2_jmap_lookup(journal_t *journal, sector_t fsblk,
329 + const char *func)
330 +{
331 + struct rb_node *p;
332 +
333 + BUG_ON(!journal);
334 +
335 + for (p = journal->j_jmap.rb_node; p; ) {
336 + struct jmap_entry *je = rb_entry(p, struct jmap_entry, rb_node);
337 + if (je->mapping.fsblk > fsblk)
338 + p = p->rb_left;
339 + else if (je->mapping.fsblk < fsblk)
340 + p = p->rb_right;
341 + else {
342 + trace_jbd2_jmap_lookup(fsblk, je->mapping.logblk, func);
343 + return je;
344 + }
345 + }
346 + trace_jbd2_jmap_lookup(fsblk, 0, func);
347 + return NULL;
348 +}
349 +
350 +/*
351 + * Revoke a mapping for the fsblk in the jmap. A lookup for fsblk will return
352 + * NULL and the mapping will be removed from the jmap during commit, unless
353 + * fsblk is reallocated as a metadata block.
354 + */
355 +void jbd2_jmap_revoke(journal_t *journal, sector_t fsblk)
356 +{
357 + struct jmap_entry *je;
358 +
359 + write_lock(&journal->j_jmap_lock);
360 + je = jbd2_jmap_lookup(journal, fsblk, __func__);
361 + /*
362 + * For now, since we do not construct jmap from the journal, it is
363 + * possible that a metadata block that was revoked is not in the jmap.
364 + * Eventually, this should not be the case and we should have a
365 + * BUG_ON(!je) here.
366 + */
367 + if (je) {
368 + BUG_ON(je->revoked);
369 + je->revoked = true;
370 + }
371 + write_unlock(&journal->j_jmap_lock);
372 +}
373 +
374 +/*
375 + * Cancel a revoke for the fsblk in the jmap.
376 + */
377 +void jbd2_jmap_cancel_revoke(journal_t *journal, sector_t fsblk)
378 +{
379 + struct jmap_entry *je;
380 +
381 + write_lock(&journal->j_jmap_lock);
382 + je = jbd2_jmap_lookup(journal, fsblk, __func__);
383 + BUG_ON(!je);
384 + BUG_ON(!je->revoked);
385 + je->revoked = false;
386 + write_unlock(&journal->j_jmap_lock);
387 +}
388 +
389 +/*
390 + * Read bh from its most up-to-date location, either from the file system or
391 + * from the log.
392 + *
393 + * If there is no mapping for the bh in jmap, this function acts like submit_bh.
394 + * Otherwise, it submits a read for the block pointed by the mapping located in
395 + * the log. Upon completion, bh will be filled with the contents of the block
396 + * read from the log.
397 + */
398 +void jbd2_submit_bh(journal_t *journal, int rw, int op_flags,
399 + struct buffer_head *bh, const char *func)
400 +{
401 + sector_t fsblk = bh->b_blocknr;
402 + sector_t logblk;
403 + struct jmap_entry *je;
404 +
405 + BUG_ON(!buffer_locked(bh));
406 +
407 + if (!journal || !(journal->j_flags & JBD2_LAZY)) {
408 + submit_bh(rw, op_flags, bh);
409 + return;
410 + }
411 +
412 + read_lock(&journal->j_jmap_lock);
413 + je = jbd2_jmap_lookup(journal, fsblk, func);
414 + if (!je) {
415 + read_unlock(&journal->j_jmap_lock);
416 + submit_bh(rw, op_flags, bh);
417 + return;
418 + }
419 + logblk = je->mapping.logblk;
420 + read_unlock(&journal->j_jmap_lock);
421 +
422 + BUG_ON(rw == WRITE);
423 + read_block_from_log(journal, bh, op_flags, logblk);
424 +}
425 +EXPORT_SYMBOL(jbd2_submit_bh);
426 +
427 +/*
428 + * End_io handler for read_block_from_log that copies the contents of
429 + * log_bh read from log to the embedded bh.
430 + */
431 +static void jbd2_end_log_read(struct buffer_head *log_bh, int uptodate)
432 +{
433 + struct buffer_head *bh = log_bh->b_private;
434 +
435 + trace_jbd2_jmap_read_from_log(bh->b_blocknr, log_bh->b_blocknr,
436 + uptodate);
437 + if (uptodate)
438 + memcpy(bh->b_data, log_bh->b_data, log_bh->b_size);
439 +
440 + unlock_buffer(log_bh);
441 + put_bh(log_bh);
442 + brelse(log_bh);
443 +
444 + bh->b_end_io(bh, uptodate);
445 +}
446 +
447 +/*
448 + * This function fills |bh| with the contents of the |blk|. Assume
449 + * jmap maps metadata block 123 to log block 100123. To read the
450 + * metadata block 123, we obtain a buffer head for it and call
451 + * read_block_from_log passing the obtained buffer head as |bh| and
452 + * 100123 as |blk|. If block 100123 is cached, then we copy the
453 + * contents to |bh| and return. Otherwise, we submit a request and
454 + * end_io handler copies the contents of block 100123 to |bh|.
455 + * Returns -ENOMEM if getblk fails, 1 if block is not cached, 0 if
456 + * block is cached.
457 + */
458 +int read_block_from_log(journal_t *journal, struct buffer_head *bh,
459 + int op_flags, sector_t blk)
460 +{
461 + struct buffer_head *log_bh;
462 +
463 + BUG_ON(!buffer_locked(bh));
464 +
465 + log_bh = __getblk(journal->j_fs_dev, blk, bh->b_size);
466 + if (unlikely(!log_bh)) {
467 + bh->b_end_io(bh, 0);
468 + return -ENOMEM;
469 + }
470 +
471 + lock_buffer(log_bh);
472 + if (buffer_uptodate(log_bh)) {
473 + memcpy(bh->b_data, log_bh->b_data, bh->b_size);
474 + unlock_buffer(log_bh);
475 + brelse(log_bh);
476 + bh->b_end_io(bh, 1);
477 + return 0;
478 + }
479 +
480 + log_bh->b_end_io = jbd2_end_log_read;
481 + log_bh->b_private = bh;
482 + get_bh(log_bh);
483 + submit_bh(READ, op_flags, log_bh);
484 + return 1;
485 +}
486 +
487 +/*
488 + * Copy of ll_rw_block that uses jbd2_submit_bh instead of submit_bh.
489 + */
490 +void jbd2_ll_rw_block(journal_t *journal, int rw, int op_flags,
491 + int nr, struct buffer_head *bhs[], const char *func)
492 +{
493 + int i;
494 +
495 + for (i = 0; i < nr; i++) {
496 + struct buffer_head *bh = bhs[i];
497 +
498 + if (!trylock_buffer(bh))
499 + continue;
500 + BUG_ON(rw == WRITE);
501 + if (!buffer_uptodate(bh)) {
502 + bh->b_end_io = end_buffer_read_sync;
503 + get_bh(bh);
504 + jbd2_submit_bh(journal, rw, op_flags, bh, func);
505 + continue;
506 + }
507 + unlock_buffer(bh);
508 + }
509 +}
510 +EXPORT_SYMBOL(jbd2_ll_rw_block);
511 +
512 +/*
513 + * Copy of bh_submit_read that uses jbd2_submit_bh instead of submit_bh.
514 + */
515 +int jbd2_bh_submit_read(journal_t *journal, struct buffer_head *bh,
516 + const char *func)
517 +{
518 + BUG_ON(!buffer_locked(bh));
519 +
520 + if (buffer_uptodate(bh)) {
521 + unlock_buffer(bh);
522 + return 0;
523 + }
524 +
525 + get_bh(bh);
526 + bh->b_end_io = end_buffer_read_sync;
527 + jbd2_submit_bh(journal, READ, 0, bh, func);
528 + wait_on_buffer(bh);
529 + if (buffer_uptodate(bh))
530 + return 0;
531 + return -EIO;
532 +}
533 +EXPORT_SYMBOL(jbd2_bh_submit_read);
534 +
535 +int jbd2_smr_journal_init(journal_t *journal)
536 +{
537 + journal->j_jmap = RB_ROOT;
538 + rwlock_init(&journal->j_jmap_lock);
539 + return jbd2_init_transaction_infos(journal);
540 +}
541 +
542 +void jbd2_smr_journal_exit(journal_t *journal)
543 +{
544 + jbd2_free_transaction_infos(journal);
545 +}
546 +
547 +void jbd2_sb_breadahead(journal_t *journal, struct super_block *sb,
548 + sector_t block)
549 +{
550 + struct buffer_head *bh = __getblk(sb->s_bdev, block, sb->s_blocksize);
551 + if (likely(bh)) {
552 + jbd2_ll_rw_block(journal, REQ_OP_READ, REQ_RAHEAD, 1,
553 + &bh, __func__);
554 + brelse(bh);
555 + }
556 +}
557 +EXPORT_SYMBOL(jbd2_sb_breadahead);
558 diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c
559 index eae93cdfbaa7..77a44c86ccc2 100644
560 --- a/fs/jbd2/journal.c
561 +++ b/fs/jbd2/journal.c
562 @@ -1120,15 +1120,17 @@ static journal_t *journal_init_common(struct block_device *bdev,
563 journal->j_max_batch_time = 15000; /* 15ms */
564 atomic_set(&journal->j_reserved_credits, 0);
566 + err = jbd2_smr_journal_init(journal);
567 + if (err)
568 + goto out_err;
569 +
570 /* The journal is marked for error until we succeed with recovery! */
571 journal->j_flags = JBD2_ABORT;
573 /* Set up a default-sized revoke table for the new mount. */
574 err = jbd2_journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
575 - if (err) {
576 - kfree(journal);
577 - return NULL;
578 - }
579 + if (err)
580 + goto out_err;
582 spin_lock_init(&journal->j_history_lock);
584 @@ -1162,6 +1164,9 @@ static journal_t *journal_init_common(struct block_device *bdev,
585 journal->j_superblock = (journal_superblock_t *)bh->b_data;
587 return journal;
588 +out_err:
589 + kfree(journal);
590 + return NULL;
591 }
593 /* jbd2_journal_init_dev and jbd2_journal_init_inode:
594 @@ -1741,6 +1746,7 @@ int jbd2_journal_destroy(journal_t *journal)
595 jbd2_journal_destroy_revoke(journal);
596 if (journal->j_chksum_driver)
597 crypto_free_shash(journal->j_chksum_driver);
598 + jbd2_smr_journal_exit(journal);
599 kfree(journal->j_wbuf);
600 kfree(journal);
602 @@ -2641,6 +2647,8 @@ static int __init journal_init_caches(void)
603 ret = jbd2_journal_init_handle_cache();
604 if (ret == 0)
605 ret = jbd2_journal_init_transaction_cache();
606 + if (ret == 0)
607 + ret = jbd2_journal_init_jmap_cache();
608 return ret;
609 }
611 @@ -2650,6 +2658,7 @@ static void jbd2_journal_destroy_caches(void)
612 jbd2_journal_destroy_journal_head_cache();
613 jbd2_journal_destroy_handle_cache();
614 jbd2_journal_destroy_transaction_cache();
615 + jbd2_journal_destroy_jmap_cache();
616 jbd2_journal_destroy_slabs();
617 }
619 diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h
620 index 9a07b0485784..771588026353 100644
621 --- a/include/linux/jbd2.h
622 +++ b/include/linux/jbd2.h
623 @@ -25,6 +25,7 @@
624 #include <linux/types.h>
625 #include <linux/buffer_head.h>
626 #include <linux/journal-head.h>
627 +#include <linux/jmap.h>
628 #include <linux/stddef.h>
629 #include <linux/mutex.h>
630 #include <linux/timer.h>
631 @@ -732,6 +733,9 @@ jbd2_time_diff(unsigned long start, unsigned long end)
632 * prior abort)?
633 * @j_sb_buffer: First part of superblock buffer
634 * @j_superblock: Second part of superblock buffer
635 + * @j_map: A map from file system blocks to log blocks
636 + * @j_transaction_infos: An array of information structures per live transaction
637 + * @j_map_lock: Protect j_jmap and j_transaction_infos
638 * @j_format_version: Version of the superblock format
639 * @j_state_lock: Protect the various scalars in the journal
640 * @j_barrier_count: Number of processes waiting to create a barrier lock
641 @@ -807,6 +811,15 @@ struct journal_s
642 struct buffer_head *j_sb_buffer;
643 journal_superblock_t *j_superblock;
645 + /* A map from file system blocks to journal blocks */
646 + struct rb_root j_jmap;
647 +
648 + /* An array of housekeeping information about live transactions */
649 + struct transaction_infos *j_transaction_infos;
650 +
651 + /* Protect j_jmap and j_transaction_infos */
652 + rwlock_t j_jmap_lock;
653 +
654 /* Version of the superblock format */
655 int j_format_version;
657 @@ -1129,6 +1142,7 @@ JBD2_FEATURE_INCOMPAT_FUNCS(csum3, CSUM_V3)
658 * mode */
659 #define JBD2_REC_ERR 0x080 /* The errno in the sb has been recorded */
660 #define JBD2_NO_CLEANUP 0x100 /* Don't flush empty the journal on shutdown */
661 +#define JBD2_LAZY 0x200 /* Do lazy journalling */
663 /*
664 * Function declarations for the journaling transaction and buffer
665 diff --git a/include/linux/jmap.h b/include/linux/jmap.h
666 new file mode 100644
667 index 000000000000..a602ece7cc89
668 --- /dev/null
669 +++ b/include/linux/jmap.h
670 @@ -0,0 +1,131 @@
671 +#ifndef _LINUX_JMAP_H
672 +#define _LINUX_JMAP_H
673 +
674 +#include <linux/buffer_head.h>
675 +#include <linux/journal-head.h>
676 +#include <linux/list.h>
677 +#include <linux/circ_buf.h>
678 +
679 +/*
680 + * Maximum number of transactions. This guides the size of the circular buffer
681 + * in which we store housekeeping information per transaction. We start
682 + * cleaning either when the circular buffer is full or when we hit the free
683 + * space threshold, whichever happens first. For starters, we make this
684 + * constant large to make sure that we start cleaning only when we hit the free
685 + * space threshold. Later we can empirically determine a sensible value.
686 + */
687 +#define MAX_LIVE_TRANSACTIONS 65536
688 +
689 +/*
690 + * Forward declaration for journal_t so that we don't get circular dependency
691 + * between jbd2.h and jmap.h
692 + */
693 +struct journal_s;
694 +typedef struct journal_s journal_t;
695 +
696 +/*
697 + * A mapping from file system block to log block.
698 + */
699 +struct blk_mapping {
700 + sector_t fsblk;
701 + sector_t logblk;
702 +};
703 +
704 +/*
705 + * An RB-tree entry wrapper for blk_mapping with extra housekeeping information.
706 + */
707 +struct jmap_entry {
708 + struct rb_node rb_node;
709 +
710 + /* The actual mapping information. */
711 + struct blk_mapping mapping;
712 +
713 + /*
714 + * If a block that is mapped gets deleted, the revoked bit is set. A
715 + * lookup for a deleted block fails. If a deleted block gets
716 + * re-allocated as a metadata block, the mapping is updated and revoked
717 + * bit is cleared.
718 + */
719 + bool revoked;
720 +
721 + /*
722 + * All log blocks that are part of the same transaction in the log are
723 + * chained with a linked list. The root of the list is stored in the
724 + * transaction_info structure described below.
725 + */
726 + struct list_head list;
727 +
728 + /*
729 + * The last time when fsblk was written again to the journal and
730 + * therefore was remapped to a different log block.
731 + */
732 + unsigned long fsblk_last_modified;
733 +
734 + /*
735 + * Index of the transaction in the transaction_info_buffer (described
736 + * below) of which the log block is part of.
737 + */
738 + int t_idx;
739 +};
740 +
741 +/*
742 + * Housekeeping information about committed transaction.
743 + */
744 +struct transaction_info {
745 + /* Id of the transaction */
746 + tid_t tid;
747 +
748 + /* Offset where the transaction starts in the log */
749 + sector_t offset;
750 +
751 + /*
752 + * A list of live log blocks referenced in the RB-tree that belong to
753 + * this transaction. It is used during cleaning to locate live blocks
754 + * and migrate them to appropriate location. If this list is empty,
755 + * then the transaction does not contain any live blocks and we can
756 + * reuse its space. If this list is not empty, then we can quickly
757 + * locate all the live blocks in this transaction.
758 + */
759 + struct list_head live_logblks;
760 +};
761 +
762 +/*
763 + * An array of transaction_info structures about all the transactions in the
764 + * log. Since there can only be a limited number of transactions in the log, we
765 + * use a circular buffer to store housekeeping information about transactions.
766 + */
767 +struct transaction_infos {
768 + struct transaction_info *buf;
769 + int head;
770 + int tail;
771 +};
772 +
773 +extern int jbd2_smr_journal_init(journal_t *journal);
774 +extern void jbd2_smr_journal_exit(journal_t *journal);
775 +
776 +extern int jbd2_journal_init_jmap_cache(void);
777 +extern void jbd2_journal_destroy_jmap_cache(void);
778 +
779 +extern int jbd2_init_transaction_infos(journal_t *journal);
780 +extern void jbd2_free_transaction_infos(journal_t *journal);
781 +extern int jbd2_transaction_infos_add(journal_t *journal,
782 + transaction_t *transaction,
783 + struct blk_mapping *mappings,
784 + int nr_mappings);
785 +
786 +extern struct jmap_entry *jbd2_jmap_lookup(journal_t *journal, sector_t fsblk,
787 + const char *func);
788 +extern void jbd2_jmap_revoke(journal_t *journal, sector_t fsblk);
789 +extern void jbd2_jmap_cancel_revoke(journal_t *journal, sector_t fsblk);
790 +extern void jbd2_submit_bh(journal_t *journal, int rw, int op_flags,
791 + struct buffer_head *bh, const char *func);
792 +extern int read_block_from_log(journal_t *journal, struct buffer_head *bh,
793 + int op_flags, sector_t blk);
794 +extern void jbd2_ll_rw_block(journal_t *journal, int rw, int op_flags, int nr,
795 + struct buffer_head *bhs[], const char *func);
796 +extern int jbd2_bh_submit_read(journal_t *journal, struct buffer_head *bh,
797 + const char *func);
798 +extern void jbd2_sb_breadahead(journal_t *journal, struct super_block *sb,
799 + sector_t block);
800 +
801 +#endif
802 diff --git a/include/trace/events/jbd2.h b/include/trace/events/jbd2.h
803 index c1d1f3eb242d..6d0619bc99af 100644
804 --- a/include/trace/events/jbd2.h
805 +++ b/include/trace/events/jbd2.h
806 @@ -379,6 +379,199 @@ TRACE_EVENT(jbd2_lock_buffer_stall,
807 __entry->stall_ms)
808 );
810 +TRACE_EVENT(jbd2_jmap_replace,
811 +
812 + TP_PROTO(struct jmap_entry *jentry, struct blk_mapping *mapping, \
813 + int t_idx),
814 +
815 + TP_ARGS(jentry, mapping, t_idx),
816 +
817 + TP_STRUCT__entry(
818 + __field(sector_t, fsblk )
819 + __field(sector_t, old_logblk )
820 + __field(sector_t, new_logblk )
821 + __field(int, old_t_idx )
822 + __field(int, new_t_idx )
823 + ),
824 +
825 + TP_fast_assign(
826 + __entry->fsblk = mapping->fsblk;
827 + __entry->old_logblk = jentry->mapping.logblk;
828 + __entry->new_logblk = mapping->logblk;
829 + __entry->old_t_idx = jentry->t_idx;
830 + __entry->new_t_idx = t_idx;
831 + ),
832 +
833 + TP_printk("remap %llu from %llu to %llu, move from transaction at index %d to transaction at index %d",
834 + (unsigned long long) __entry->fsblk,
835 + (unsigned long long) __entry->old_logblk,
836 + (unsigned long long) __entry->new_logblk,
837 + __entry->old_t_idx,
838 + __entry->new_t_idx)
839 +);
840 +
841 +TRACE_EVENT(jbd2_jmap_insert,
842 +
843 + TP_PROTO(struct blk_mapping *mapping, int t_idx),
844 +
845 + TP_ARGS(mapping, t_idx),
846 +
847 + TP_STRUCT__entry(
848 + __field(sector_t, fsblk )
849 + __field(sector_t, logblk)
850 + __field(int, t_idx)
851 + ),
852 +
853 + TP_fast_assign(
854 + __entry->fsblk = mapping->fsblk;
855 + __entry->logblk = mapping->logblk;
856 + __entry->t_idx = t_idx;
857 + ),
858 +
859 + TP_printk("map %llu to %llu, insert to transaction %d",
860 + (unsigned long long) __entry->fsblk,
861 + (unsigned long long) __entry->logblk,
862 + __entry->t_idx)
863 +);
864 +
865 +TRACE_EVENT(jbd2_jmap_lookup,
866 +
867 + TP_PROTO(sector_t fsblk, sector_t logblk, const char *func),
868 +
869 + TP_ARGS(fsblk, logblk, func),
870 +
871 + TP_STRUCT__entry(
872 + __field(sector_t, fsblk )
873 + __field(sector_t, logblk)
874 + __string(func, func)
875 + ),
876 +
877 + TP_fast_assign(
878 + __entry->fsblk = fsblk;
879 + __entry->logblk = logblk;
880 + __assign_str(func, func);
881 + ),
882 +
883 + TP_printk("%s: lookup %llu -> %llu",
884 + __get_str(func),
885 + (unsigned long long) __entry->fsblk,
886 + (unsigned long long) __entry->logblk)
887 +);
888 +
889 +TRACE_EVENT(jbd2_jmap_read_from_log,
890 +
891 + TP_PROTO(sector_t fsblk, sector_t logblk, int uptodate),
892 +
893 + TP_ARGS(fsblk, logblk, uptodate),
894 +
895 + TP_STRUCT__entry(
896 + __field(sector_t, fsblk )
897 + __field(sector_t, logblk)
898 + __field(int, uptodate)
899 + ),
900 +
901 + TP_fast_assign(
902 + __entry->fsblk = fsblk;
903 + __entry->logblk = logblk;
904 + __entry->uptodate = uptodate;
905 + ),
906 +
907 + TP_printk("fsblk %llu logblk %llu uptodate %d",
908 + (unsigned long long) __entry->fsblk,
909 + (unsigned long long) __entry->logblk,
910 + __entry->uptodate)
911 +);
912 +
913 +TRACE_EVENT(jbd2_jmap_printf,
914 +
915 + TP_PROTO(const char *s),
916 +
917 + TP_ARGS(s),
918 +
919 + TP_STRUCT__entry(
920 + __string(s, s)
921 + ),
922 +
923 + TP_fast_assign(
924 + __assign_str(s, s);
925 + ),
926 +
927 + TP_printk("%s",
928 + __get_str(s))
929 +);
930 +
931 +TRACE_EVENT(jbd2_jmap_printf1,
932 +
933 + TP_PROTO(const char *s, sector_t fsblk),
934 +
935 + TP_ARGS(s, fsblk),
936 +
937 + TP_STRUCT__entry(
938 + __string(s, s)
939 + __field(sector_t, fsblk )
940 + ),
941 +
942 + TP_fast_assign(
943 + __assign_str(s, s);
944 + __entry->fsblk = fsblk;
945 + ),
946 +
947 + TP_printk("%s: %llu",
948 + __get_str(s),
949 + (unsigned long long) __entry->fsblk)
950 +);
951 +
952 +TRACE_EVENT(jbd2_jmap_printf2,
953 +
954 + TP_PROTO(const char *s, sector_t fsblk, sector_t logblk),
955 +
956 + TP_ARGS(s, fsblk, logblk),
957 +
958 + TP_STRUCT__entry(
959 + __string(s, s)
960 + __field(sector_t, fsblk )
961 + __field(sector_t, logblk)
962 + ),
963 +
964 + TP_fast_assign(
965 + __assign_str(s, s);
966 + __entry->fsblk = fsblk;
967 + __entry->logblk = logblk;
968 + ),
969 +
970 + TP_printk("%s: %llu:%llu",
971 + __get_str(s),
972 + (unsigned long long) __entry->fsblk,
973 + (unsigned long long) __entry->logblk)
974 +);
975 +
976 +TRACE_EVENT(jbd2_transaction_infos_add,
977 +
978 + TP_PROTO(int t_idx, struct transaction_info *ti, int nr_mappings),
979 +
980 + TP_ARGS(t_idx, ti, nr_mappings),
981 +
982 + TP_STRUCT__entry(
983 + __field(int, t_idx )
984 + __field(tid_t, tid )
985 + __field(sector_t, offset)
986 + __field(int, nr_mappings)
987 + ),
988 +
989 + TP_fast_assign(
990 + __entry->t_idx = t_idx;
991 + __entry->tid = ti->tid;
992 + __entry->offset = ti->offset;
993 + __entry->nr_mappings = nr_mappings;
994 + ),
995 +
996 + TP_printk("inserted transaction %u (offset %llu) at index %d with %d mappings",
997 + __entry->tid,
998 + (unsigned long long) __entry->offset,
999 + __entry->t_idx,
1000 + __entry->nr_mappings)
1001 +);
1002 +
1003 #endif /* _TRACE_JBD2_H */
1005 /* This part must be outside protection */