1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
4 * Copyright (C) 2004-2007 Red Hat, Inc. All rights reserved.
7 #include <linux/sched.h>
8 #include <linux/slab.h>
9 #include <linux/spinlock.h>
10 #include <linux/completion.h>
11 #include <linux/buffer_head.h>
12 #include <linux/gfs2_ondisk.h>
13 #include <linux/crc32.h>
14 #include <linux/crc32c.h>
15 #include <linux/delay.h>
16 #include <linux/kthread.h>
17 #include <linux/freezer.h>
18 #include <linux/bio.h>
19 #include <linux/blkdev.h>
20 #include <linux/writeback.h>
21 #include <linux/list_sort.h>
32 #include "trace_gfs2.h"
35 static void gfs2_log_shutdown(struct gfs2_sbd *sdp);
38 * gfs2_struct2blk - compute stuff
39 * @sdp: the filesystem
40 * @nstruct: the number of structures
42 * Compute the number of log descriptor blocks needed to hold a certain number
43 * of structures of a certain size.
45 * Returns: the number of blocks needed (minimum is always 1)
48 unsigned int gfs2_struct2blk(struct gfs2_sbd *sdp, unsigned int nstruct)
51 unsigned int first, second;
53 /* The initial struct gfs2_log_descriptor block */
55 first = sdp->sd_ldptrs;
57 if (nstruct > first) {
58 /* Subsequent struct gfs2_meta_header blocks */
59 second = sdp->sd_inptrs;
60 blks += DIV_ROUND_UP(nstruct - first, second);
67 * gfs2_remove_from_ail - Remove an entry from the ail lists, updating counters
68 * @bd: The gfs2_bufdata to remove
70 * The ail lock _must_ be held when calling this function
74 void gfs2_remove_from_ail(struct gfs2_bufdata *bd)
77 list_del_init(&bd->bd_ail_st_list);
78 list_del_init(&bd->bd_ail_gl_list);
79 atomic_dec(&bd->bd_gl->gl_ail_count);
84 * gfs2_ail1_start_one - Start I/O on a transaction
85 * @sdp: The superblock
86 * @wbc: The writeback control structure
87 * @tr: The transaction to start I/O on
88 * @plug: The block plug currently active
91 static int gfs2_ail1_start_one(struct gfs2_sbd *sdp,
92 struct writeback_control *wbc,
93 struct gfs2_trans *tr, struct blk_plug *plug)
94 __releases(&sdp->sd_ail_lock)
95 __acquires(&sdp->sd_ail_lock)
97 struct gfs2_glock *gl = NULL;
98 struct address_space *mapping;
99 struct gfs2_bufdata *bd, *s;
100 struct buffer_head *bh;
103 list_for_each_entry_safe_reverse(bd, s, &tr->tr_ail1_list, bd_ail_st_list) {
106 gfs2_assert(sdp, bd->bd_tr == tr);
108 if (!buffer_busy(bh)) {
109 if (buffer_uptodate(bh)) {
110 list_move(&bd->bd_ail_st_list,
114 if (!cmpxchg(&sdp->sd_log_error, 0, -EIO)) {
115 gfs2_io_error_bh(sdp, bh);
116 gfs2_withdraw_delayed(sdp);
120 if (gfs2_withdrawn(sdp)) {
121 gfs2_remove_from_ail(bd);
124 if (!buffer_dirty(bh))
129 list_move(&bd->bd_ail_st_list, &tr->tr_ail1_list);
130 mapping = bh->b_page->mapping;
133 spin_unlock(&sdp->sd_ail_lock);
134 ret = filemap_fdatawrite_wbc(mapping, wbc);
135 if (need_resched()) {
136 blk_finish_plug(plug);
138 blk_start_plug(plug);
140 spin_lock(&sdp->sd_ail_lock);
141 if (ret == -ENODATA) /* if a jdata write into a new hole */
142 ret = 0; /* ignore it */
143 if (ret || wbc->nr_to_write <= 0)
151 static void dump_ail_list(struct gfs2_sbd *sdp)
153 struct gfs2_trans *tr;
154 struct gfs2_bufdata *bd;
155 struct buffer_head *bh;
157 list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
158 list_for_each_entry_reverse(bd, &tr->tr_ail1_list,
161 fs_err(sdp, "bd %p: blk:0x%llx bh=%p ", bd,
162 (unsigned long long)bd->bd_blkno, bh);
167 fs_err(sdp, "0x%llx up2:%d dirt:%d lkd:%d req:%d "
168 "map:%d new:%d ar:%d aw:%d delay:%d "
169 "io err:%d unwritten:%d dfr:%d pin:%d esc:%d\n",
170 (unsigned long long)bh->b_blocknr,
171 buffer_uptodate(bh), buffer_dirty(bh),
172 buffer_locked(bh), buffer_req(bh),
173 buffer_mapped(bh), buffer_new(bh),
174 buffer_async_read(bh), buffer_async_write(bh),
175 buffer_delay(bh), buffer_write_io_error(bh),
176 buffer_unwritten(bh),
177 buffer_defer_completion(bh),
178 buffer_pinned(bh), buffer_escaped(bh));
184 * gfs2_ail1_flush - start writeback of some ail1 entries
185 * @sdp: The super block
186 * @wbc: The writeback control structure
188 * Writes back some ail1 entries, according to the limits in the
189 * writeback control structure
192 void gfs2_ail1_flush(struct gfs2_sbd *sdp, struct writeback_control *wbc)
194 struct list_head *head = &sdp->sd_ail1_list;
195 struct gfs2_trans *tr;
196 struct blk_plug plug;
198 unsigned long flush_start = jiffies;
200 trace_gfs2_ail_flush(sdp, wbc, 1);
201 blk_start_plug(&plug);
202 spin_lock(&sdp->sd_ail_lock);
205 if (time_after(jiffies, flush_start + (HZ * 600))) {
206 fs_err(sdp, "Error: In %s for ten minutes! t=%d\n",
207 __func__, current->journal_info ? 1 : 0);
211 list_for_each_entry_reverse(tr, head, tr_list) {
212 if (wbc->nr_to_write <= 0)
214 ret = gfs2_ail1_start_one(sdp, wbc, tr, &plug);
222 spin_unlock(&sdp->sd_ail_lock);
223 blk_finish_plug(&plug);
225 gfs2_lm(sdp, "gfs2_ail1_start_one returned: %d\n", ret);
228 trace_gfs2_ail_flush(sdp, wbc, 0);
232 * gfs2_ail1_start - start writeback of all ail1 entries
233 * @sdp: The superblock
236 static void gfs2_ail1_start(struct gfs2_sbd *sdp)
238 struct writeback_control wbc = {
239 .sync_mode = WB_SYNC_NONE,
240 .nr_to_write = LONG_MAX,
242 .range_end = LLONG_MAX,
245 return gfs2_ail1_flush(sdp, &wbc);
248 static void gfs2_log_update_flush_tail(struct gfs2_sbd *sdp)
250 unsigned int new_flush_tail = sdp->sd_log_head;
251 struct gfs2_trans *tr;
253 if (!list_empty(&sdp->sd_ail1_list)) {
254 tr = list_last_entry(&sdp->sd_ail1_list,
255 struct gfs2_trans, tr_list);
256 new_flush_tail = tr->tr_first;
258 sdp->sd_log_flush_tail = new_flush_tail;
261 static void gfs2_log_update_head(struct gfs2_sbd *sdp)
263 unsigned int new_head = sdp->sd_log_flush_head;
265 if (sdp->sd_log_flush_tail == sdp->sd_log_head)
266 sdp->sd_log_flush_tail = new_head;
267 sdp->sd_log_head = new_head;
271 * gfs2_ail_empty_tr - empty one of the ail lists of a transaction
274 static void gfs2_ail_empty_tr(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
275 struct list_head *head)
277 struct gfs2_bufdata *bd;
279 while (!list_empty(head)) {
280 bd = list_first_entry(head, struct gfs2_bufdata,
282 gfs2_assert(sdp, bd->bd_tr == tr);
283 gfs2_remove_from_ail(bd);
288 * gfs2_ail1_empty_one - Check whether or not a trans in the AIL has been synced
289 * @sdp: the filesystem
290 * @tr: the transaction
291 * @max_revokes: If nonzero, issue revokes for the bd items for written buffers
293 * returns: the transaction's count of remaining active items
296 static int gfs2_ail1_empty_one(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
299 struct gfs2_bufdata *bd, *s;
300 struct buffer_head *bh;
301 int active_count = 0;
303 list_for_each_entry_safe_reverse(bd, s, &tr->tr_ail1_list,
306 gfs2_assert(sdp, bd->bd_tr == tr);
308 * If another process flagged an io error, e.g. writing to the
309 * journal, error all other bhs and move them off the ail1 to
310 * prevent a tight loop when unmount tries to flush ail1,
311 * regardless of whether they're still busy. If no outside
312 * errors were found and the buffer is busy, move to the next.
313 * If the ail buffer is not busy and caught an error, flag it
316 if (!sdp->sd_log_error && buffer_busy(bh)) {
320 if (!buffer_uptodate(bh) &&
321 !cmpxchg(&sdp->sd_log_error, 0, -EIO)) {
322 gfs2_io_error_bh(sdp, bh);
323 gfs2_withdraw_delayed(sdp);
326 * If we have space for revokes and the bd is no longer on any
327 * buf list, we can just add a revoke for it immediately and
328 * avoid having to put it on the ail2 list, where it would need
329 * to be revoked later.
331 if (*max_revokes && list_empty(&bd->bd_list)) {
332 gfs2_add_revoke(sdp, bd);
336 list_move(&bd->bd_ail_st_list, &tr->tr_ail2_list);
342 * gfs2_ail1_empty - Try to empty the ail1 lists
343 * @sdp: The superblock
344 * @max_revokes: If non-zero, add revokes where appropriate
346 * Tries to empty the ail1 lists, starting with the oldest first
349 static int gfs2_ail1_empty(struct gfs2_sbd *sdp, int max_revokes)
351 struct gfs2_trans *tr, *s;
355 spin_lock(&sdp->sd_ail_lock);
356 list_for_each_entry_safe_reverse(tr, s, &sdp->sd_ail1_list, tr_list) {
357 if (!gfs2_ail1_empty_one(sdp, tr, &max_revokes) && oldest_tr)
358 list_move(&tr->tr_list, &sdp->sd_ail2_list);
362 gfs2_log_update_flush_tail(sdp);
363 ret = list_empty(&sdp->sd_ail1_list);
364 spin_unlock(&sdp->sd_ail_lock);
366 if (test_bit(SDF_WITHDRAWING, &sdp->sd_flags)) {
367 gfs2_lm(sdp, "fatal: I/O error(s)\n");
374 static void gfs2_ail1_wait(struct gfs2_sbd *sdp)
376 struct gfs2_trans *tr;
377 struct gfs2_bufdata *bd;
378 struct buffer_head *bh;
380 spin_lock(&sdp->sd_ail_lock);
381 list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
382 list_for_each_entry(bd, &tr->tr_ail1_list, bd_ail_st_list) {
384 if (!buffer_locked(bh))
387 spin_unlock(&sdp->sd_ail_lock);
393 spin_unlock(&sdp->sd_ail_lock);
396 static void __ail2_empty(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
398 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
399 list_del(&tr->tr_list);
400 gfs2_assert_warn(sdp, list_empty(&tr->tr_ail1_list));
401 gfs2_assert_warn(sdp, list_empty(&tr->tr_ail2_list));
402 gfs2_trans_free(sdp, tr);
405 static void ail2_empty(struct gfs2_sbd *sdp, unsigned int new_tail)
407 struct list_head *ail2_list = &sdp->sd_ail2_list;
408 unsigned int old_tail = sdp->sd_log_tail;
409 struct gfs2_trans *tr, *safe;
411 spin_lock(&sdp->sd_ail_lock);
412 if (old_tail <= new_tail) {
413 list_for_each_entry_safe(tr, safe, ail2_list, tr_list) {
414 if (old_tail <= tr->tr_first && tr->tr_first < new_tail)
415 __ail2_empty(sdp, tr);
418 list_for_each_entry_safe(tr, safe, ail2_list, tr_list) {
419 if (old_tail <= tr->tr_first || tr->tr_first < new_tail)
420 __ail2_empty(sdp, tr);
423 spin_unlock(&sdp->sd_ail_lock);
427 * gfs2_log_is_empty - Check if the log is empty
428 * @sdp: The GFS2 superblock
431 bool gfs2_log_is_empty(struct gfs2_sbd *sdp) {
432 return atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks;
435 static bool __gfs2_log_try_reserve_revokes(struct gfs2_sbd *sdp, unsigned int revokes)
437 unsigned int available;
439 available = atomic_read(&sdp->sd_log_revokes_available);
440 while (available >= revokes) {
441 if (atomic_try_cmpxchg(&sdp->sd_log_revokes_available,
442 &available, available - revokes))
449 * gfs2_log_release_revokes - Release a given number of revokes
450 * @sdp: The GFS2 superblock
451 * @revokes: The number of revokes to release
453 * sdp->sd_log_flush_lock must be held.
455 void gfs2_log_release_revokes(struct gfs2_sbd *sdp, unsigned int revokes)
458 atomic_add(revokes, &sdp->sd_log_revokes_available);
462 * gfs2_log_release - Release a given number of log blocks
463 * @sdp: The GFS2 superblock
464 * @blks: The number of blocks
468 void gfs2_log_release(struct gfs2_sbd *sdp, unsigned int blks)
470 atomic_add(blks, &sdp->sd_log_blks_free);
471 trace_gfs2_log_blocks(sdp, blks);
472 gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
473 sdp->sd_jdesc->jd_blocks);
474 if (atomic_read(&sdp->sd_log_blks_needed))
475 wake_up(&sdp->sd_log_waitq);
479 * __gfs2_log_try_reserve - Try to make a log reservation
480 * @sdp: The GFS2 superblock
481 * @blks: The number of blocks to reserve
482 * @taboo_blks: The number of blocks to leave free
484 * Try to do the same as __gfs2_log_reserve(), but fail if no more log
485 * space is immediately available.
487 static bool __gfs2_log_try_reserve(struct gfs2_sbd *sdp, unsigned int blks,
488 unsigned int taboo_blks)
490 unsigned wanted = blks + taboo_blks;
491 unsigned int free_blocks;
493 free_blocks = atomic_read(&sdp->sd_log_blks_free);
494 while (free_blocks >= wanted) {
495 if (atomic_try_cmpxchg(&sdp->sd_log_blks_free, &free_blocks,
496 free_blocks - blks)) {
497 trace_gfs2_log_blocks(sdp, -blks);
505 * __gfs2_log_reserve - Make a log reservation
506 * @sdp: The GFS2 superblock
507 * @blks: The number of blocks to reserve
508 * @taboo_blks: The number of blocks to leave free
510 * @taboo_blks is set to 0 for logd, and to GFS2_LOG_FLUSH_MIN_BLOCKS
511 * for all other processes. This ensures that when the log is almost full,
512 * logd will still be able to call gfs2_log_flush one more time without
513 * blocking, which will advance the tail and make some more log space
516 * We no longer flush the log here, instead we wake up logd to do that
517 * for us. To avoid the thundering herd and to ensure that we deal fairly
518 * with queued waiters, we use an exclusive wait. This means that when we
519 * get woken with enough journal space to get our reservation, we need to
520 * wake the next waiter on the list.
523 static void __gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks,
524 unsigned int taboo_blks)
526 unsigned wanted = blks + taboo_blks;
527 unsigned int free_blocks;
529 atomic_add(blks, &sdp->sd_log_blks_needed);
531 if (current != sdp->sd_logd_process)
532 wake_up(&sdp->sd_logd_waitq);
533 io_wait_event(sdp->sd_log_waitq,
534 (free_blocks = atomic_read(&sdp->sd_log_blks_free),
535 free_blocks >= wanted));
537 if (atomic_try_cmpxchg(&sdp->sd_log_blks_free,
541 } while (free_blocks >= wanted);
545 trace_gfs2_log_blocks(sdp, -blks);
546 if (atomic_sub_return(blks, &sdp->sd_log_blks_needed))
547 wake_up(&sdp->sd_log_waitq);
551 * gfs2_log_try_reserve - Try to make a log reservation
552 * @sdp: The GFS2 superblock
553 * @tr: The transaction
554 * @extra_revokes: The number of additional revokes reserved (output)
556 * This is similar to gfs2_log_reserve, but sdp->sd_log_flush_lock must be
557 * held for correct revoke accounting.
560 bool gfs2_log_try_reserve(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
561 unsigned int *extra_revokes)
563 unsigned int blks = tr->tr_reserved;
564 unsigned int revokes = tr->tr_revokes;
565 unsigned int revoke_blks = 0;
568 if (revokes && !__gfs2_log_try_reserve_revokes(sdp, revokes)) {
569 revoke_blks = DIV_ROUND_UP(revokes, sdp->sd_inptrs);
570 *extra_revokes = revoke_blks * sdp->sd_inptrs - revokes;
575 if (__gfs2_log_try_reserve(sdp, blks, GFS2_LOG_FLUSH_MIN_BLOCKS))
578 gfs2_log_release_revokes(sdp, revokes);
583 * gfs2_log_reserve - Make a log reservation
584 * @sdp: The GFS2 superblock
585 * @tr: The transaction
586 * @extra_revokes: The number of additional revokes reserved (output)
588 * sdp->sd_log_flush_lock must not be held.
591 void gfs2_log_reserve(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
592 unsigned int *extra_revokes)
594 unsigned int blks = tr->tr_reserved;
595 unsigned int revokes = tr->tr_revokes;
596 unsigned int revoke_blks;
600 revoke_blks = DIV_ROUND_UP(revokes, sdp->sd_inptrs);
601 *extra_revokes = revoke_blks * sdp->sd_inptrs - revokes;
604 __gfs2_log_reserve(sdp, blks, GFS2_LOG_FLUSH_MIN_BLOCKS);
608 * log_distance - Compute distance between two journal blocks
609 * @sdp: The GFS2 superblock
610 * @newer: The most recent journal block of the pair
611 * @older: The older journal block of the pair
613 * Compute the distance (in the journal direction) between two
614 * blocks in the journal
616 * Returns: the distance in blocks
619 static inline unsigned int log_distance(struct gfs2_sbd *sdp, unsigned int newer,
624 dist = newer - older;
626 dist += sdp->sd_jdesc->jd_blocks;
632 * calc_reserved - Calculate the number of blocks to keep reserved
633 * @sdp: The GFS2 superblock
635 * This is complex. We need to reserve room for all our currently used
636 * metadata blocks (e.g. normal file I/O rewriting file time stamps) and
637 * all our journaled data blocks for journaled files (e.g. files in the
638 * meta_fs like rindex, or files for which chattr +j was done.)
639 * If we don't reserve enough space, corruption will follow.
641 * We can have metadata blocks and jdata blocks in the same journal. Each
642 * type gets its own log descriptor, for which we need to reserve a block.
643 * In fact, each type has the potential for needing more than one log descriptor
644 * in cases where we have more blocks than will fit in a log descriptor.
645 * Metadata journal entries take up half the space of journaled buffer entries.
647 * Also, we need to reserve blocks for revoke journal entries and one for an
648 * overall header for the lot.
650 * Returns: the number of blocks reserved
652 static unsigned int calc_reserved(struct gfs2_sbd *sdp)
654 unsigned int reserved = GFS2_LOG_FLUSH_MIN_BLOCKS;
656 struct gfs2_trans *tr = sdp->sd_log_tr;
659 blocks = tr->tr_num_buf_new - tr->tr_num_buf_rm;
660 reserved += blocks + DIV_ROUND_UP(blocks, buf_limit(sdp));
661 blocks = tr->tr_num_databuf_new - tr->tr_num_databuf_rm;
662 reserved += blocks + DIV_ROUND_UP(blocks, databuf_limit(sdp));
667 static void log_pull_tail(struct gfs2_sbd *sdp)
669 unsigned int new_tail = sdp->sd_log_flush_tail;
672 if (new_tail == sdp->sd_log_tail)
674 dist = log_distance(sdp, new_tail, sdp->sd_log_tail);
675 ail2_empty(sdp, new_tail);
676 gfs2_log_release(sdp, dist);
677 sdp->sd_log_tail = new_tail;
681 void log_flush_wait(struct gfs2_sbd *sdp)
685 if (atomic_read(&sdp->sd_log_in_flight)) {
687 prepare_to_wait(&sdp->sd_log_flush_wait, &wait,
688 TASK_UNINTERRUPTIBLE);
689 if (atomic_read(&sdp->sd_log_in_flight))
691 } while(atomic_read(&sdp->sd_log_in_flight));
692 finish_wait(&sdp->sd_log_flush_wait, &wait);
696 static int ip_cmp(void *priv, const struct list_head *a, const struct list_head *b)
698 struct gfs2_inode *ipa, *ipb;
700 ipa = list_entry(a, struct gfs2_inode, i_ordered);
701 ipb = list_entry(b, struct gfs2_inode, i_ordered);
703 if (ipa->i_no_addr < ipb->i_no_addr)
705 if (ipa->i_no_addr > ipb->i_no_addr)
710 static void __ordered_del_inode(struct gfs2_inode *ip)
712 if (!list_empty(&ip->i_ordered))
713 list_del_init(&ip->i_ordered);
716 static void gfs2_ordered_write(struct gfs2_sbd *sdp)
718 struct gfs2_inode *ip;
721 spin_lock(&sdp->sd_ordered_lock);
722 list_sort(NULL, &sdp->sd_log_ordered, &ip_cmp);
723 while (!list_empty(&sdp->sd_log_ordered)) {
724 ip = list_first_entry(&sdp->sd_log_ordered, struct gfs2_inode, i_ordered);
725 if (ip->i_inode.i_mapping->nrpages == 0) {
726 __ordered_del_inode(ip);
729 list_move(&ip->i_ordered, &written);
730 spin_unlock(&sdp->sd_ordered_lock);
731 filemap_fdatawrite(ip->i_inode.i_mapping);
732 spin_lock(&sdp->sd_ordered_lock);
734 list_splice(&written, &sdp->sd_log_ordered);
735 spin_unlock(&sdp->sd_ordered_lock);
738 static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
740 struct gfs2_inode *ip;
742 spin_lock(&sdp->sd_ordered_lock);
743 while (!list_empty(&sdp->sd_log_ordered)) {
744 ip = list_first_entry(&sdp->sd_log_ordered, struct gfs2_inode, i_ordered);
745 __ordered_del_inode(ip);
746 if (ip->i_inode.i_mapping->nrpages == 0)
748 spin_unlock(&sdp->sd_ordered_lock);
749 filemap_fdatawait(ip->i_inode.i_mapping);
750 spin_lock(&sdp->sd_ordered_lock);
752 spin_unlock(&sdp->sd_ordered_lock);
755 void gfs2_ordered_del_inode(struct gfs2_inode *ip)
757 struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
759 spin_lock(&sdp->sd_ordered_lock);
760 __ordered_del_inode(ip);
761 spin_unlock(&sdp->sd_ordered_lock);
764 void gfs2_add_revoke(struct gfs2_sbd *sdp, struct gfs2_bufdata *bd)
766 struct buffer_head *bh = bd->bd_bh;
767 struct gfs2_glock *gl = bd->bd_gl;
769 sdp->sd_log_num_revoke++;
770 if (atomic_inc_return(&gl->gl_revokes) == 1)
772 bh->b_private = NULL;
773 bd->bd_blkno = bh->b_blocknr;
774 gfs2_remove_from_ail(bd); /* drops ref on bh */
776 set_bit(GLF_LFLUSH, &gl->gl_flags);
777 list_add(&bd->bd_list, &sdp->sd_log_revokes);
780 void gfs2_glock_remove_revoke(struct gfs2_glock *gl)
782 if (atomic_dec_return(&gl->gl_revokes) == 0) {
783 clear_bit(GLF_LFLUSH, &gl->gl_flags);
784 gfs2_glock_queue_put(gl);
789 * gfs2_flush_revokes - Add as many revokes to the system transaction as we can
790 * @sdp: The GFS2 superblock
792 * Our usual strategy is to defer writing revokes as much as we can in the hope
793 * that we'll eventually overwrite the journal, which will make those revokes
794 * go away. This changes when we flush the log: at that point, there will
795 * likely be some left-over space in the last revoke block of that transaction.
796 * We can fill that space with additional revokes for blocks that have already
797 * been written back. This will basically come at no cost now, and will save
798 * us from having to keep track of those blocks on the AIL2 list later.
800 void gfs2_flush_revokes(struct gfs2_sbd *sdp)
802 /* number of revokes we still have room for */
803 unsigned int max_revokes = atomic_read(&sdp->sd_log_revokes_available);
806 gfs2_ail1_empty(sdp, max_revokes);
807 gfs2_log_unlock(sdp);
811 * gfs2_write_log_header - Write a journal log header buffer at lblock
812 * @sdp: The GFS2 superblock
813 * @jd: journal descriptor of the journal to which we are writing
814 * @seq: sequence number
815 * @tail: tail of the log
816 * @lblock: value for lh_blkno (block number relative to start of journal)
817 * @flags: log header flags GFS2_LOG_HEAD_*
818 * @op_flags: flags to pass to the bio
820 * Returns: the initialized log buffer descriptor
823 void gfs2_write_log_header(struct gfs2_sbd *sdp, struct gfs2_jdesc *jd,
824 u64 seq, u32 tail, u32 lblock, u32 flags,
827 struct gfs2_log_header *lh;
830 struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
831 struct timespec64 tv;
832 struct super_block *sb = sdp->sd_vfs;
835 if (gfs2_withdrawn(sdp))
838 page = mempool_alloc(gfs2_page_pool, GFP_NOIO);
839 lh = page_address(page);
842 lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
843 lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
844 lh->lh_header.__pad0 = cpu_to_be64(0);
845 lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
846 lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
847 lh->lh_sequence = cpu_to_be64(seq);
848 lh->lh_flags = cpu_to_be32(flags);
849 lh->lh_tail = cpu_to_be32(tail);
850 lh->lh_blkno = cpu_to_be32(lblock);
851 hash = ~crc32(~0, lh, LH_V1_SIZE);
852 lh->lh_hash = cpu_to_be32(hash);
854 ktime_get_coarse_real_ts64(&tv);
855 lh->lh_nsec = cpu_to_be32(tv.tv_nsec);
856 lh->lh_sec = cpu_to_be64(tv.tv_sec);
857 if (!list_empty(&jd->extent_list))
858 dblock = gfs2_log_bmap(jd, lblock);
864 ret = gfs2_get_extent(jd->jd_inode, lblock, &dblock, &extlen);
865 if (gfs2_assert_withdraw(sdp, ret == 0))
868 lh->lh_addr = cpu_to_be64(dblock);
869 lh->lh_jinode = cpu_to_be64(GFS2_I(jd->jd_inode)->i_no_addr);
871 /* We may only write local statfs, quota, etc., when writing to our
872 own journal. The values are left 0 when recovering a journal
873 different from our own. */
874 if (!(flags & GFS2_LOG_HEAD_RECOVERY)) {
876 cpu_to_be64(GFS2_I(sdp->sd_sc_inode)->i_no_addr);
878 cpu_to_be64(GFS2_I(sdp->sd_qc_inode)->i_no_addr);
880 spin_lock(&sdp->sd_statfs_spin);
881 lh->lh_local_total = cpu_to_be64(l_sc->sc_total);
882 lh->lh_local_free = cpu_to_be64(l_sc->sc_free);
883 lh->lh_local_dinodes = cpu_to_be64(l_sc->sc_dinodes);
884 spin_unlock(&sdp->sd_statfs_spin);
887 BUILD_BUG_ON(offsetof(struct gfs2_log_header, lh_crc) != LH_V1_SIZE);
889 crc = crc32c(~0, (void *)lh + LH_V1_SIZE + 4,
890 sb->s_blocksize - LH_V1_SIZE - 4);
891 lh->lh_crc = cpu_to_be32(crc);
893 gfs2_log_write(sdp, jd, page, sb->s_blocksize, 0, dblock);
894 gfs2_log_submit_bio(&jd->jd_log_bio, REQ_OP_WRITE | op_flags);
898 * log_write_header - Get and initialize a journal header buffer
899 * @sdp: The GFS2 superblock
900 * @flags: The log header flags, including log header origin
902 * Returns: the initialized log buffer descriptor
905 static void log_write_header(struct gfs2_sbd *sdp, u32 flags)
907 blk_opf_t op_flags = REQ_PREFLUSH | REQ_FUA | REQ_META | REQ_SYNC;
908 enum gfs2_freeze_state state = atomic_read(&sdp->sd_freeze_state);
910 gfs2_assert_withdraw(sdp, (state != SFS_FROZEN));
912 if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags)) {
913 gfs2_ordered_wait(sdp);
915 op_flags = REQ_SYNC | REQ_META | REQ_PRIO;
917 sdp->sd_log_idle = (sdp->sd_log_flush_tail == sdp->sd_log_flush_head);
918 gfs2_write_log_header(sdp, sdp->sd_jdesc, sdp->sd_log_sequence++,
919 sdp->sd_log_flush_tail, sdp->sd_log_flush_head,
921 gfs2_log_incr_head(sdp);
924 gfs2_log_update_head(sdp);
928 * gfs2_ail_drain - drain the ail lists after a withdraw
929 * @sdp: Pointer to GFS2 superblock
931 void gfs2_ail_drain(struct gfs2_sbd *sdp)
933 struct gfs2_trans *tr;
935 spin_lock(&sdp->sd_ail_lock);
937 * For transactions on the sd_ail1_list we need to drain both the
938 * ail1 and ail2 lists. That's because function gfs2_ail1_start_one
939 * (temporarily) moves items from its tr_ail1 list to tr_ail2 list
940 * before revokes are sent for that block. Items on the sd_ail2_list
941 * should have already gotten beyond that point, so no need.
943 while (!list_empty(&sdp->sd_ail1_list)) {
944 tr = list_first_entry(&sdp->sd_ail1_list, struct gfs2_trans,
946 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail1_list);
947 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
948 list_del(&tr->tr_list);
949 gfs2_trans_free(sdp, tr);
951 while (!list_empty(&sdp->sd_ail2_list)) {
952 tr = list_first_entry(&sdp->sd_ail2_list, struct gfs2_trans,
954 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
955 list_del(&tr->tr_list);
956 gfs2_trans_free(sdp, tr);
958 gfs2_drain_revokes(sdp);
959 spin_unlock(&sdp->sd_ail_lock);
963 * empty_ail1_list - try to start IO and empty the ail1 list
964 * @sdp: Pointer to GFS2 superblock
966 static void empty_ail1_list(struct gfs2_sbd *sdp)
968 unsigned long start = jiffies;
971 if (time_after(jiffies, start + (HZ * 600))) {
972 fs_err(sdp, "Error: In %s for 10 minutes! t=%d\n",
973 __func__, current->journal_info ? 1 : 0);
977 gfs2_ail1_start(sdp);
979 if (gfs2_ail1_empty(sdp, 0))
985 * trans_drain - drain the buf and databuf queue for a failed transaction
986 * @tr: the transaction to drain
988 * When this is called, we're taking an error exit for a log write that failed
989 * but since we bypassed the after_commit functions, we need to remove the
990 * items from the buf and databuf queue.
992 static void trans_drain(struct gfs2_trans *tr)
994 struct gfs2_bufdata *bd;
995 struct list_head *head;
1001 while (!list_empty(head)) {
1002 bd = list_first_entry(head, struct gfs2_bufdata, bd_list);
1003 list_del_init(&bd->bd_list);
1004 if (!list_empty(&bd->bd_ail_st_list))
1005 gfs2_remove_from_ail(bd);
1006 kmem_cache_free(gfs2_bufdata_cachep, bd);
1008 head = &tr->tr_databuf;
1009 while (!list_empty(head)) {
1010 bd = list_first_entry(head, struct gfs2_bufdata, bd_list);
1011 list_del_init(&bd->bd_list);
1012 if (!list_empty(&bd->bd_ail_st_list))
1013 gfs2_remove_from_ail(bd);
1014 kmem_cache_free(gfs2_bufdata_cachep, bd);
1019 * gfs2_log_flush - flush incore transaction(s)
1020 * @sdp: The filesystem
1021 * @gl: The glock structure to flush. If NULL, flush the whole incore log
1022 * @flags: The log header flags: GFS2_LOG_HEAD_FLUSH_* and debug flags
1026 void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl, u32 flags)
1028 struct gfs2_trans *tr = NULL;
1029 unsigned int reserved_blocks = 0, used_blocks = 0;
1030 enum gfs2_freeze_state state = atomic_read(&sdp->sd_freeze_state);
1031 unsigned int first_log_head;
1032 unsigned int reserved_revokes = 0;
1034 down_write(&sdp->sd_log_flush_lock);
1035 trace_gfs2_log_flush(sdp, 1, flags);
1039 * Do this check while holding the log_flush_lock to prevent new
1040 * buffers from being added to the ail via gfs2_pin()
1042 if (gfs2_withdrawn(sdp) || !test_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags))
1045 /* Log might have been flushed while we waited for the flush lock */
1046 if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags))
1049 first_log_head = sdp->sd_log_head;
1050 sdp->sd_log_flush_head = first_log_head;
1052 tr = sdp->sd_log_tr;
1053 if (tr || sdp->sd_log_num_revoke) {
1054 if (reserved_blocks)
1055 gfs2_log_release(sdp, reserved_blocks);
1056 reserved_blocks = sdp->sd_log_blks_reserved;
1057 reserved_revokes = sdp->sd_log_num_revoke;
1059 sdp->sd_log_tr = NULL;
1060 tr->tr_first = first_log_head;
1061 if (unlikely (state == SFS_FROZEN)) {
1062 if (gfs2_assert_withdraw_delayed(sdp,
1063 !tr->tr_num_buf_new && !tr->tr_num_databuf_new))
1067 } else if (!reserved_blocks) {
1068 unsigned int taboo_blocks = GFS2_LOG_FLUSH_MIN_BLOCKS;
1070 reserved_blocks = GFS2_LOG_FLUSH_MIN_BLOCKS;
1071 if (current == sdp->sd_logd_process)
1074 if (!__gfs2_log_try_reserve(sdp, reserved_blocks, taboo_blocks)) {
1075 up_write(&sdp->sd_log_flush_lock);
1076 __gfs2_log_reserve(sdp, reserved_blocks, taboo_blocks);
1077 down_write(&sdp->sd_log_flush_lock);
1080 BUG_ON(sdp->sd_log_num_revoke);
1083 if (flags & GFS2_LOG_HEAD_FLUSH_SHUTDOWN)
1084 clear_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags);
1086 if (unlikely(state == SFS_FROZEN))
1087 if (gfs2_assert_withdraw_delayed(sdp, !reserved_revokes))
1090 gfs2_ordered_write(sdp);
1091 if (gfs2_withdrawn(sdp))
1093 lops_before_commit(sdp, tr);
1094 if (gfs2_withdrawn(sdp))
1096 gfs2_log_submit_bio(&sdp->sd_jdesc->jd_log_bio, REQ_OP_WRITE);
1097 if (gfs2_withdrawn(sdp))
1100 if (sdp->sd_log_head != sdp->sd_log_flush_head) {
1101 log_write_header(sdp, flags);
1102 } else if (sdp->sd_log_tail != sdp->sd_log_flush_tail && !sdp->sd_log_idle) {
1103 log_write_header(sdp, flags);
1105 if (gfs2_withdrawn(sdp))
1107 lops_after_commit(sdp, tr);
1110 sdp->sd_log_blks_reserved = 0;
1112 spin_lock(&sdp->sd_ail_lock);
1113 if (tr && !list_empty(&tr->tr_ail1_list)) {
1114 list_add(&tr->tr_list, &sdp->sd_ail1_list);
1117 spin_unlock(&sdp->sd_ail_lock);
1118 gfs2_log_unlock(sdp);
1120 if (!(flags & GFS2_LOG_HEAD_FLUSH_NORMAL)) {
1121 if (!sdp->sd_log_idle) {
1122 empty_ail1_list(sdp);
1123 if (gfs2_withdrawn(sdp))
1125 log_write_header(sdp, flags);
1127 if (flags & (GFS2_LOG_HEAD_FLUSH_SHUTDOWN |
1128 GFS2_LOG_HEAD_FLUSH_FREEZE))
1129 gfs2_log_shutdown(sdp);
1130 if (flags & GFS2_LOG_HEAD_FLUSH_FREEZE)
1131 atomic_set(&sdp->sd_freeze_state, SFS_FROZEN);
1135 used_blocks = log_distance(sdp, sdp->sd_log_flush_head, first_log_head);
1136 reserved_revokes += atomic_read(&sdp->sd_log_revokes_available);
1137 atomic_set(&sdp->sd_log_revokes_available, sdp->sd_ldptrs);
1138 gfs2_assert_withdraw(sdp, reserved_revokes % sdp->sd_inptrs == sdp->sd_ldptrs);
1139 if (reserved_revokes > sdp->sd_ldptrs)
1140 reserved_blocks += (reserved_revokes - sdp->sd_ldptrs) / sdp->sd_inptrs;
1142 if (used_blocks != reserved_blocks) {
1143 gfs2_assert_withdraw_delayed(sdp, used_blocks < reserved_blocks);
1144 gfs2_log_release(sdp, reserved_blocks - used_blocks);
1146 up_write(&sdp->sd_log_flush_lock);
1147 gfs2_trans_free(sdp, tr);
1148 if (gfs2_withdrawing(sdp))
1150 trace_gfs2_log_flush(sdp, 0, flags);
1156 * If the tr_list is empty, we're withdrawing during a log
1157 * flush that targets a transaction, but the transaction was
1158 * never queued onto any of the ail lists. Here we add it to
1159 * ail1 just so that ail_drain() will find and free it.
1161 spin_lock(&sdp->sd_ail_lock);
1162 if (tr && list_empty(&tr->tr_list))
1163 list_add(&tr->tr_list, &sdp->sd_ail1_list);
1164 spin_unlock(&sdp->sd_ail_lock);
1170 * gfs2_merge_trans - Merge a new transaction into a cached transaction
1171 * @sdp: the filesystem
1172 * @new: New transaction to be merged
1175 static void gfs2_merge_trans(struct gfs2_sbd *sdp, struct gfs2_trans *new)
1177 struct gfs2_trans *old = sdp->sd_log_tr;
1179 WARN_ON_ONCE(!test_bit(TR_ATTACHED, &old->tr_flags));
1181 old->tr_num_buf_new += new->tr_num_buf_new;
1182 old->tr_num_databuf_new += new->tr_num_databuf_new;
1183 old->tr_num_buf_rm += new->tr_num_buf_rm;
1184 old->tr_num_databuf_rm += new->tr_num_databuf_rm;
1185 old->tr_revokes += new->tr_revokes;
1186 old->tr_num_revoke += new->tr_num_revoke;
1188 list_splice_tail_init(&new->tr_databuf, &old->tr_databuf);
1189 list_splice_tail_init(&new->tr_buf, &old->tr_buf);
1191 spin_lock(&sdp->sd_ail_lock);
1192 list_splice_tail_init(&new->tr_ail1_list, &old->tr_ail1_list);
1193 list_splice_tail_init(&new->tr_ail2_list, &old->tr_ail2_list);
1194 spin_unlock(&sdp->sd_ail_lock);
1197 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
1199 unsigned int reserved;
1200 unsigned int unused;
1201 unsigned int maxres;
1205 if (sdp->sd_log_tr) {
1206 gfs2_merge_trans(sdp, tr);
1207 } else if (tr->tr_num_buf_new || tr->tr_num_databuf_new) {
1208 gfs2_assert_withdraw(sdp, !test_bit(TR_ONSTACK, &tr->tr_flags));
1209 sdp->sd_log_tr = tr;
1210 set_bit(TR_ATTACHED, &tr->tr_flags);
1213 reserved = calc_reserved(sdp);
1214 maxres = sdp->sd_log_blks_reserved + tr->tr_reserved;
1215 gfs2_assert_withdraw(sdp, maxres >= reserved);
1216 unused = maxres - reserved;
1218 gfs2_log_release(sdp, unused);
1219 sdp->sd_log_blks_reserved = reserved;
1221 gfs2_log_unlock(sdp);
1225 * gfs2_log_commit - Commit a transaction to the log
1226 * @sdp: the filesystem
1227 * @tr: the transaction
1229 * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
1230 * or the total number of used blocks (pinned blocks plus AIL blocks)
1231 * is greater than thresh2.
1233 * At mount time thresh1 is 2/5ths of journal size, thresh2 is 4/5ths of
1239 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
1241 log_refund(sdp, tr);
1243 if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
1244 ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
1245 atomic_read(&sdp->sd_log_thresh2)))
1246 wake_up(&sdp->sd_logd_waitq);
1250 * gfs2_log_shutdown - write a shutdown header into a journal
1251 * @sdp: the filesystem
1255 static void gfs2_log_shutdown(struct gfs2_sbd *sdp)
1257 gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
1258 gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
1259 gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
1261 log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT | GFS2_LFC_SHUTDOWN);
1264 gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
1265 gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
1268 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
1270 return (atomic_read(&sdp->sd_log_pinned) +
1271 atomic_read(&sdp->sd_log_blks_needed) >=
1272 atomic_read(&sdp->sd_log_thresh1));
1275 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
1277 unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
1279 if (test_and_clear_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags))
1282 return used_blocks + atomic_read(&sdp->sd_log_blks_needed) >=
1283 atomic_read(&sdp->sd_log_thresh2);
1287 * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
1288 * @data: Pointer to GFS2 superblock
1290 * Also, periodically check to make sure that we're using the most recent
1294 int gfs2_logd(void *data)
1296 struct gfs2_sbd *sdp = data;
1297 unsigned long t = 1;
1300 while (!kthread_should_stop()) {
1302 if (gfs2_withdrawn(sdp)) {
1303 msleep_interruptible(HZ);
1306 /* Check for errors writing to the journal */
1307 if (sdp->sd_log_error) {
1309 "GFS2: fsid=%s: error %d: "
1310 "withdrawing the file system to "
1311 "prevent further damage.\n",
1312 sdp->sd_fsname, sdp->sd_log_error);
1317 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
1318 gfs2_ail1_empty(sdp, 0);
1319 gfs2_log_flush(sdp, NULL, GFS2_LOG_HEAD_FLUSH_NORMAL |
1320 GFS2_LFC_LOGD_JFLUSH_REQD);
1323 if (gfs2_ail_flush_reqd(sdp)) {
1324 gfs2_ail1_start(sdp);
1325 gfs2_ail1_wait(sdp);
1326 gfs2_ail1_empty(sdp, 0);
1327 gfs2_log_flush(sdp, NULL, GFS2_LOG_HEAD_FLUSH_NORMAL |
1328 GFS2_LFC_LOGD_AIL_FLUSH_REQD);
1331 t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
1336 prepare_to_wait(&sdp->sd_logd_waitq, &wait,
1337 TASK_INTERRUPTIBLE);
1338 if (!gfs2_ail_flush_reqd(sdp) &&
1339 !gfs2_jrnl_flush_reqd(sdp) &&
1340 !kthread_should_stop())
1341 t = schedule_timeout(t);
1342 } while(t && !gfs2_ail_flush_reqd(sdp) &&
1343 !gfs2_jrnl_flush_reqd(sdp) &&
1344 !kthread_should_stop());
1345 finish_wait(&sdp->sd_logd_waitq, &wait);