GFS2: Clean up log flush header writing
[linux-2.6.git] / fs / gfs2 / log.c
1 /*
2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
3  * Copyright (C) 2004-2007 Red Hat, Inc.  All rights reserved.
4  *
5  * This copyrighted material is made available to anyone wishing to use,
6  * modify, copy, or redistribute it subject to the terms and conditions
7  * of the GNU General Public License version 2.
8  */
9
10 #include <linux/sched.h>
11 #include <linux/slab.h>
12 #include <linux/spinlock.h>
13 #include <linux/completion.h>
14 #include <linux/buffer_head.h>
15 #include <linux/gfs2_ondisk.h>
16 #include <linux/crc32.h>
17 #include <linux/delay.h>
18 #include <linux/kthread.h>
19 #include <linux/freezer.h>
20 #include <linux/bio.h>
21 #include <linux/writeback.h>
22 #include <linux/list_sort.h>
23
24 #include "gfs2.h"
25 #include "incore.h"
26 #include "bmap.h"
27 #include "glock.h"
28 #include "log.h"
29 #include "lops.h"
30 #include "meta_io.h"
31 #include "util.h"
32 #include "dir.h"
33 #include "trace_gfs2.h"
34
35 #define PULL 1
36
37 /**
38  * gfs2_struct2blk - compute stuff
39  * @sdp: the filesystem
40  * @nstruct: the number of structures
41  * @ssize: the size of the structures
42  *
43  * Compute the number of log descriptor blocks needed to hold a certain number
44  * of structures of a certain size.
45  *
46  * Returns: the number of blocks needed (minimum is always 1)
47  */
48
49 unsigned int gfs2_struct2blk(struct gfs2_sbd *sdp, unsigned int nstruct,
50                              unsigned int ssize)
51 {
52         unsigned int blks;
53         unsigned int first, second;
54
55         blks = 1;
56         first = (sdp->sd_sb.sb_bsize - sizeof(struct gfs2_log_descriptor)) / ssize;
57
58         if (nstruct > first) {
59                 second = (sdp->sd_sb.sb_bsize -
60                           sizeof(struct gfs2_meta_header)) / ssize;
61                 blks += DIV_ROUND_UP(nstruct - first, second);
62         }
63
64         return blks;
65 }
66
67 /**
68  * gfs2_remove_from_ail - Remove an entry from the ail lists, updating counters
69  * @mapping: The associated mapping (maybe NULL)
70  * @bd: The gfs2_bufdata to remove
71  *
72  * The ail lock _must_ be held when calling this function
73  *
74  */
75
76 void gfs2_remove_from_ail(struct gfs2_bufdata *bd)
77 {
78         bd->bd_ail = NULL;
79         list_del_init(&bd->bd_ail_st_list);
80         list_del_init(&bd->bd_ail_gl_list);
81         atomic_dec(&bd->bd_gl->gl_ail_count);
82         brelse(bd->bd_bh);
83 }
84
85 /**
86  * gfs2_ail1_start_one - Start I/O on a part of the AIL
87  * @sdp: the filesystem
88  * @wbc: The writeback control structure
89  * @ai: The ail structure
90  *
91  */
92
93 static int gfs2_ail1_start_one(struct gfs2_sbd *sdp,
94                                struct writeback_control *wbc,
95                                struct gfs2_ail *ai)
96 __releases(&sdp->sd_ail_lock)
97 __acquires(&sdp->sd_ail_lock)
98 {
99         struct gfs2_glock *gl = NULL;
100         struct address_space *mapping;
101         struct gfs2_bufdata *bd, *s;
102         struct buffer_head *bh;
103
104         list_for_each_entry_safe_reverse(bd, s, &ai->ai_ail1_list, bd_ail_st_list) {
105                 bh = bd->bd_bh;
106
107                 gfs2_assert(sdp, bd->bd_ail == ai);
108
109                 if (!buffer_busy(bh)) {
110                         if (!buffer_uptodate(bh))
111                                 gfs2_io_error_bh(sdp, bh);
112                         list_move(&bd->bd_ail_st_list, &ai->ai_ail2_list);
113                         continue;
114                 }
115
116                 if (!buffer_dirty(bh))
117                         continue;
118                 if (gl == bd->bd_gl)
119                         continue;
120                 gl = bd->bd_gl;
121                 list_move(&bd->bd_ail_st_list, &ai->ai_ail1_list);
122                 mapping = bh->b_page->mapping;
123                 if (!mapping)
124                         continue;
125                 spin_unlock(&sdp->sd_ail_lock);
126                 generic_writepages(mapping, wbc);
127                 spin_lock(&sdp->sd_ail_lock);
128                 if (wbc->nr_to_write <= 0)
129                         break;
130                 return 1;
131         }
132
133         return 0;
134 }
135
136
137 /**
138  * gfs2_ail1_flush - start writeback of some ail1 entries 
139  * @sdp: The super block
140  * @wbc: The writeback control structure
141  *
142  * Writes back some ail1 entries, according to the limits in the
143  * writeback control structure
144  */
145
146 void gfs2_ail1_flush(struct gfs2_sbd *sdp, struct writeback_control *wbc)
147 {
148         struct list_head *head = &sdp->sd_ail1_list;
149         struct gfs2_ail *ai;
150
151         trace_gfs2_ail_flush(sdp, wbc, 1);
152         spin_lock(&sdp->sd_ail_lock);
153 restart:
154         list_for_each_entry_reverse(ai, head, ai_list) {
155                 if (wbc->nr_to_write <= 0)
156                         break;
157                 if (gfs2_ail1_start_one(sdp, wbc, ai))
158                         goto restart;
159         }
160         spin_unlock(&sdp->sd_ail_lock);
161         trace_gfs2_ail_flush(sdp, wbc, 0);
162 }
163
164 /**
165  * gfs2_ail1_start - start writeback of all ail1 entries
166  * @sdp: The superblock
167  */
168
169 static void gfs2_ail1_start(struct gfs2_sbd *sdp)
170 {
171         struct writeback_control wbc = {
172                 .sync_mode = WB_SYNC_NONE,
173                 .nr_to_write = LONG_MAX,
174                 .range_start = 0,
175                 .range_end = LLONG_MAX,
176         };
177
178         return gfs2_ail1_flush(sdp, &wbc);
179 }
180
181 /**
182  * gfs2_ail1_empty_one - Check whether or not a trans in the AIL has been synced
183  * @sdp: the filesystem
184  * @ai: the AIL entry
185  *
186  */
187
188 static void gfs2_ail1_empty_one(struct gfs2_sbd *sdp, struct gfs2_ail *ai)
189 {
190         struct gfs2_bufdata *bd, *s;
191         struct buffer_head *bh;
192
193         list_for_each_entry_safe_reverse(bd, s, &ai->ai_ail1_list,
194                                          bd_ail_st_list) {
195                 bh = bd->bd_bh;
196                 gfs2_assert(sdp, bd->bd_ail == ai);
197                 if (buffer_busy(bh))
198                         continue;
199                 if (!buffer_uptodate(bh))
200                         gfs2_io_error_bh(sdp, bh);
201                 list_move(&bd->bd_ail_st_list, &ai->ai_ail2_list);
202         }
203
204 }
205
206 /**
207  * gfs2_ail1_empty - Try to empty the ail1 lists
208  * @sdp: The superblock
209  *
210  * Tries to empty the ail1 lists, starting with the oldest first
211  */
212
213 static int gfs2_ail1_empty(struct gfs2_sbd *sdp)
214 {
215         struct gfs2_ail *ai, *s;
216         int ret;
217
218         spin_lock(&sdp->sd_ail_lock);
219         list_for_each_entry_safe_reverse(ai, s, &sdp->sd_ail1_list, ai_list) {
220                 gfs2_ail1_empty_one(sdp, ai);
221                 if (list_empty(&ai->ai_ail1_list))
222                         list_move(&ai->ai_list, &sdp->sd_ail2_list);
223                 else
224                         break;
225         }
226         ret = list_empty(&sdp->sd_ail1_list);
227         spin_unlock(&sdp->sd_ail_lock);
228
229         return ret;
230 }
231
232 static void gfs2_ail1_wait(struct gfs2_sbd *sdp)
233 {
234         struct gfs2_ail *ai;
235         struct gfs2_bufdata *bd;
236         struct buffer_head *bh;
237
238         spin_lock(&sdp->sd_ail_lock);
239         list_for_each_entry_reverse(ai, &sdp->sd_ail1_list, ai_list) {
240                 list_for_each_entry(bd, &ai->ai_ail1_list, bd_ail_st_list) {
241                         bh = bd->bd_bh;
242                         if (!buffer_locked(bh))
243                                 continue;
244                         get_bh(bh);
245                         spin_unlock(&sdp->sd_ail_lock);
246                         wait_on_buffer(bh);
247                         brelse(bh);
248                         return;
249                 }
250         }
251         spin_unlock(&sdp->sd_ail_lock);
252 }
253
254 /**
255  * gfs2_ail2_empty_one - Check whether or not a trans in the AIL has been synced
256  * @sdp: the filesystem
257  * @ai: the AIL entry
258  *
259  */
260
261 static void gfs2_ail2_empty_one(struct gfs2_sbd *sdp, struct gfs2_ail *ai)
262 {
263         struct list_head *head = &ai->ai_ail2_list;
264         struct gfs2_bufdata *bd;
265
266         while (!list_empty(head)) {
267                 bd = list_entry(head->prev, struct gfs2_bufdata,
268                                 bd_ail_st_list);
269                 gfs2_assert(sdp, bd->bd_ail == ai);
270                 gfs2_remove_from_ail(bd);
271         }
272 }
273
274 static void ail2_empty(struct gfs2_sbd *sdp, unsigned int new_tail)
275 {
276         struct gfs2_ail *ai, *safe;
277         unsigned int old_tail = sdp->sd_log_tail;
278         int wrap = (new_tail < old_tail);
279         int a, b, rm;
280
281         spin_lock(&sdp->sd_ail_lock);
282
283         list_for_each_entry_safe(ai, safe, &sdp->sd_ail2_list, ai_list) {
284                 a = (old_tail <= ai->ai_first);
285                 b = (ai->ai_first < new_tail);
286                 rm = (wrap) ? (a || b) : (a && b);
287                 if (!rm)
288                         continue;
289
290                 gfs2_ail2_empty_one(sdp, ai);
291                 list_del(&ai->ai_list);
292                 gfs2_assert_warn(sdp, list_empty(&ai->ai_ail1_list));
293                 gfs2_assert_warn(sdp, list_empty(&ai->ai_ail2_list));
294                 kfree(ai);
295         }
296
297         spin_unlock(&sdp->sd_ail_lock);
298 }
299
300 /**
301  * gfs2_log_reserve - Make a log reservation
302  * @sdp: The GFS2 superblock
303  * @blks: The number of blocks to reserve
304  *
305  * Note that we never give out the last few blocks of the journal. Thats
306  * due to the fact that there is a small number of header blocks
307  * associated with each log flush. The exact number can't be known until
308  * flush time, so we ensure that we have just enough free blocks at all
309  * times to avoid running out during a log flush.
310  *
311  * We no longer flush the log here, instead we wake up logd to do that
312  * for us. To avoid the thundering herd and to ensure that we deal fairly
313  * with queued waiters, we use an exclusive wait. This means that when we
314  * get woken with enough journal space to get our reservation, we need to
315  * wake the next waiter on the list.
316  *
317  * Returns: errno
318  */
319
320 int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks)
321 {
322         unsigned reserved_blks = 6 * (4096 / sdp->sd_vfs->s_blocksize);
323         unsigned wanted = blks + reserved_blks;
324         DEFINE_WAIT(wait);
325         int did_wait = 0;
326         unsigned int free_blocks;
327
328         if (gfs2_assert_warn(sdp, blks) ||
329             gfs2_assert_warn(sdp, blks <= sdp->sd_jdesc->jd_blocks))
330                 return -EINVAL;
331 retry:
332         free_blocks = atomic_read(&sdp->sd_log_blks_free);
333         if (unlikely(free_blocks <= wanted)) {
334                 do {
335                         prepare_to_wait_exclusive(&sdp->sd_log_waitq, &wait,
336                                         TASK_UNINTERRUPTIBLE);
337                         wake_up(&sdp->sd_logd_waitq);
338                         did_wait = 1;
339                         if (atomic_read(&sdp->sd_log_blks_free) <= wanted)
340                                 io_schedule();
341                         free_blocks = atomic_read(&sdp->sd_log_blks_free);
342                 } while(free_blocks <= wanted);
343                 finish_wait(&sdp->sd_log_waitq, &wait);
344         }
345         if (atomic_cmpxchg(&sdp->sd_log_blks_free, free_blocks,
346                                 free_blocks - blks) != free_blocks)
347                 goto retry;
348         trace_gfs2_log_blocks(sdp, -blks);
349
350         /*
351          * If we waited, then so might others, wake them up _after_ we get
352          * our share of the log.
353          */
354         if (unlikely(did_wait))
355                 wake_up(&sdp->sd_log_waitq);
356
357         down_read(&sdp->sd_log_flush_lock);
358
359         return 0;
360 }
361
362 u64 gfs2_log_bmap(struct gfs2_sbd *sdp, unsigned int lbn)
363 {
364         struct gfs2_journal_extent *je;
365
366         list_for_each_entry(je, &sdp->sd_jdesc->extent_list, extent_list) {
367                 if (lbn >= je->lblock && lbn < je->lblock + je->blocks)
368                         return je->dblock + lbn - je->lblock;
369         }
370
371         return -1;
372 }
373
374 /**
375  * log_distance - Compute distance between two journal blocks
376  * @sdp: The GFS2 superblock
377  * @newer: The most recent journal block of the pair
378  * @older: The older journal block of the pair
379  *
380  *   Compute the distance (in the journal direction) between two
381  *   blocks in the journal
382  *
383  * Returns: the distance in blocks
384  */
385
386 static inline unsigned int log_distance(struct gfs2_sbd *sdp, unsigned int newer,
387                                         unsigned int older)
388 {
389         int dist;
390
391         dist = newer - older;
392         if (dist < 0)
393                 dist += sdp->sd_jdesc->jd_blocks;
394
395         return dist;
396 }
397
398 /**
399  * calc_reserved - Calculate the number of blocks to reserve when
400  *                 refunding a transaction's unused buffers.
401  * @sdp: The GFS2 superblock
402  *
403  * This is complex.  We need to reserve room for all our currently used
404  * metadata buffers (e.g. normal file I/O rewriting file time stamps) and 
405  * all our journaled data buffers for journaled files (e.g. files in the 
406  * meta_fs like rindex, or files for which chattr +j was done.)
407  * If we don't reserve enough space, gfs2_log_refund and gfs2_log_flush
408  * will count it as free space (sd_log_blks_free) and corruption will follow.
409  *
410  * We can have metadata bufs and jdata bufs in the same journal.  So each
411  * type gets its own log header, for which we need to reserve a block.
412  * In fact, each type has the potential for needing more than one header 
413  * in cases where we have more buffers than will fit on a journal page.
414  * Metadata journal entries take up half the space of journaled buffer entries.
415  * Thus, metadata entries have buf_limit (502) and journaled buffers have
416  * databuf_limit (251) before they cause a wrap around.
417  *
418  * Also, we need to reserve blocks for revoke journal entries and one for an
419  * overall header for the lot.
420  *
421  * Returns: the number of blocks reserved
422  */
423 static unsigned int calc_reserved(struct gfs2_sbd *sdp)
424 {
425         unsigned int reserved = 0;
426         unsigned int mbuf_limit, metabufhdrs_needed;
427         unsigned int dbuf_limit, databufhdrs_needed;
428         unsigned int revokes = 0;
429
430         mbuf_limit = buf_limit(sdp);
431         metabufhdrs_needed = (sdp->sd_log_commited_buf +
432                               (mbuf_limit - 1)) / mbuf_limit;
433         dbuf_limit = databuf_limit(sdp);
434         databufhdrs_needed = (sdp->sd_log_commited_databuf +
435                               (dbuf_limit - 1)) / dbuf_limit;
436
437         if (sdp->sd_log_commited_revoke > 0)
438                 revokes = gfs2_struct2blk(sdp, sdp->sd_log_commited_revoke,
439                                           sizeof(u64));
440
441         reserved = sdp->sd_log_commited_buf + metabufhdrs_needed +
442                 sdp->sd_log_commited_databuf + databufhdrs_needed +
443                 revokes;
444         /* One for the overall header */
445         if (reserved)
446                 reserved++;
447         return reserved;
448 }
449
450 static unsigned int current_tail(struct gfs2_sbd *sdp)
451 {
452         struct gfs2_ail *ai;
453         unsigned int tail;
454
455         spin_lock(&sdp->sd_ail_lock);
456
457         if (list_empty(&sdp->sd_ail1_list)) {
458                 tail = sdp->sd_log_head;
459         } else {
460                 ai = list_entry(sdp->sd_ail1_list.prev, struct gfs2_ail, ai_list);
461                 tail = ai->ai_first;
462         }
463
464         spin_unlock(&sdp->sd_ail_lock);
465
466         return tail;
467 }
468
469 void gfs2_log_incr_head(struct gfs2_sbd *sdp)
470 {
471         BUG_ON((sdp->sd_log_flush_head == sdp->sd_log_tail) &&
472                (sdp->sd_log_flush_head != sdp->sd_log_head));
473
474         if (++sdp->sd_log_flush_head == sdp->sd_jdesc->jd_blocks) {
475                 sdp->sd_log_flush_head = 0;
476                 sdp->sd_log_flush_wrapped = 1;
477         }
478 }
479
480 static void log_pull_tail(struct gfs2_sbd *sdp, unsigned int new_tail)
481 {
482         unsigned int dist = log_distance(sdp, new_tail, sdp->sd_log_tail);
483
484         ail2_empty(sdp, new_tail);
485
486         atomic_add(dist, &sdp->sd_log_blks_free);
487         trace_gfs2_log_blocks(sdp, dist);
488         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
489                              sdp->sd_jdesc->jd_blocks);
490
491         sdp->sd_log_tail = new_tail;
492 }
493
494
495 static void log_flush_wait(struct gfs2_sbd *sdp)
496 {
497         DEFINE_WAIT(wait);
498
499         if (atomic_read(&sdp->sd_log_in_flight)) {
500                 do {
501                         prepare_to_wait(&sdp->sd_log_flush_wait, &wait,
502                                         TASK_UNINTERRUPTIBLE);
503                         if (atomic_read(&sdp->sd_log_in_flight))
504                                 io_schedule();
505                 } while(atomic_read(&sdp->sd_log_in_flight));
506                 finish_wait(&sdp->sd_log_flush_wait, &wait);
507         }
508 }
509
510 static int bd_cmp(void *priv, struct list_head *a, struct list_head *b)
511 {
512         struct gfs2_bufdata *bda, *bdb;
513
514         bda = list_entry(a, struct gfs2_bufdata, bd_le.le_list);
515         bdb = list_entry(b, struct gfs2_bufdata, bd_le.le_list);
516
517         if (bda->bd_bh->b_blocknr < bdb->bd_bh->b_blocknr)
518                 return -1;
519         if (bda->bd_bh->b_blocknr > bdb->bd_bh->b_blocknr)
520                 return 1;
521         return 0;
522 }
523
524 static void gfs2_ordered_write(struct gfs2_sbd *sdp)
525 {
526         struct gfs2_bufdata *bd;
527         struct buffer_head *bh;
528         LIST_HEAD(written);
529
530         gfs2_log_lock(sdp);
531         list_sort(NULL, &sdp->sd_log_le_ordered, &bd_cmp);
532         while (!list_empty(&sdp->sd_log_le_ordered)) {
533                 bd = list_entry(sdp->sd_log_le_ordered.next, struct gfs2_bufdata, bd_le.le_list);
534                 list_move(&bd->bd_le.le_list, &written);
535                 bh = bd->bd_bh;
536                 if (!buffer_dirty(bh))
537                         continue;
538                 get_bh(bh);
539                 gfs2_log_unlock(sdp);
540                 lock_buffer(bh);
541                 if (buffer_mapped(bh) && test_clear_buffer_dirty(bh)) {
542                         bh->b_end_io = end_buffer_write_sync;
543                         submit_bh(WRITE_SYNC, bh);
544                 } else {
545                         unlock_buffer(bh);
546                         brelse(bh);
547                 }
548                 gfs2_log_lock(sdp);
549         }
550         list_splice(&written, &sdp->sd_log_le_ordered);
551         gfs2_log_unlock(sdp);
552 }
553
554 static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
555 {
556         struct gfs2_bufdata *bd;
557         struct buffer_head *bh;
558
559         gfs2_log_lock(sdp);
560         while (!list_empty(&sdp->sd_log_le_ordered)) {
561                 bd = list_entry(sdp->sd_log_le_ordered.prev, struct gfs2_bufdata, bd_le.le_list);
562                 bh = bd->bd_bh;
563                 if (buffer_locked(bh)) {
564                         get_bh(bh);
565                         gfs2_log_unlock(sdp);
566                         wait_on_buffer(bh);
567                         brelse(bh);
568                         gfs2_log_lock(sdp);
569                         continue;
570                 }
571                 list_del_init(&bd->bd_le.le_list);
572         }
573         gfs2_log_unlock(sdp);
574 }
575
576 /**
577  * log_write_header - Get and initialize a journal header buffer
578  * @sdp: The GFS2 superblock
579  *
580  * Returns: the initialized log buffer descriptor
581  */
582
583 static void log_write_header(struct gfs2_sbd *sdp, u32 flags, int pull)
584 {
585         u64 blkno = gfs2_log_bmap(sdp, sdp->sd_log_flush_head);
586         struct buffer_head *bh;
587         struct gfs2_log_header *lh;
588         unsigned int tail;
589         u32 hash;
590
591         bh = sb_getblk(sdp->sd_vfs, blkno);
592         lock_buffer(bh);
593         memset(bh->b_data, 0, bh->b_size);
594         set_buffer_uptodate(bh);
595         clear_buffer_dirty(bh);
596
597         gfs2_ail1_empty(sdp);
598         tail = current_tail(sdp);
599
600         lh = (struct gfs2_log_header *)bh->b_data;
601         memset(lh, 0, sizeof(struct gfs2_log_header));
602         lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
603         lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
604         lh->lh_header.__pad0 = cpu_to_be64(0);
605         lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
606         lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
607         lh->lh_sequence = cpu_to_be64(sdp->sd_log_sequence++);
608         lh->lh_flags = cpu_to_be32(flags);
609         lh->lh_tail = cpu_to_be32(tail);
610         lh->lh_blkno = cpu_to_be32(sdp->sd_log_flush_head);
611         hash = gfs2_disk_hash(bh->b_data, sizeof(struct gfs2_log_header));
612         lh->lh_hash = cpu_to_be32(hash);
613
614         bh->b_end_io = end_buffer_write_sync;
615         get_bh(bh);
616         if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags)) {
617                 gfs2_ordered_wait(sdp);
618                 log_flush_wait(sdp);
619                 submit_bh(WRITE_SYNC | REQ_META | REQ_PRIO, bh);
620         } else {
621                 submit_bh(WRITE_FLUSH_FUA | REQ_META, bh);
622         }
623         wait_on_buffer(bh);
624
625         if (!buffer_uptodate(bh))
626                 gfs2_io_error_bh(sdp, bh);
627         brelse(bh);
628
629         if (sdp->sd_log_tail != tail)
630                 log_pull_tail(sdp, tail);
631         else
632                 gfs2_assert_withdraw(sdp, !pull);
633
634         sdp->sd_log_idle = (tail == sdp->sd_log_flush_head);
635         gfs2_log_incr_head(sdp);
636 }
637
638 /**
639  * gfs2_log_flush - flush incore transaction(s)
640  * @sdp: the filesystem
641  * @gl: The glock structure to flush.  If NULL, flush the whole incore log
642  *
643  */
644
645 void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl)
646 {
647         struct gfs2_ail *ai;
648
649         down_write(&sdp->sd_log_flush_lock);
650
651         /* Log might have been flushed while we waited for the flush lock */
652         if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags)) {
653                 up_write(&sdp->sd_log_flush_lock);
654                 return;
655         }
656         trace_gfs2_log_flush(sdp, 1);
657
658         ai = kzalloc(sizeof(struct gfs2_ail), GFP_NOFS | __GFP_NOFAIL);
659         INIT_LIST_HEAD(&ai->ai_ail1_list);
660         INIT_LIST_HEAD(&ai->ai_ail2_list);
661
662         if (sdp->sd_log_num_buf != sdp->sd_log_commited_buf) {
663                 printk(KERN_INFO "GFS2: log buf %u %u\n", sdp->sd_log_num_buf,
664                        sdp->sd_log_commited_buf);
665                 gfs2_assert_withdraw(sdp, 0);
666         }
667         if (sdp->sd_log_num_databuf != sdp->sd_log_commited_databuf) {
668                 printk(KERN_INFO "GFS2: log databuf %u %u\n",
669                        sdp->sd_log_num_databuf, sdp->sd_log_commited_databuf);
670                 gfs2_assert_withdraw(sdp, 0);
671         }
672         gfs2_assert_withdraw(sdp,
673                         sdp->sd_log_num_revoke == sdp->sd_log_commited_revoke);
674
675         sdp->sd_log_flush_head = sdp->sd_log_head;
676         sdp->sd_log_flush_wrapped = 0;
677         ai->ai_first = sdp->sd_log_flush_head;
678
679         gfs2_ordered_write(sdp);
680         lops_before_commit(sdp);
681
682         if (sdp->sd_log_head != sdp->sd_log_flush_head) {
683                 log_write_header(sdp, 0, 0);
684         } else if (sdp->sd_log_tail != current_tail(sdp) && !sdp->sd_log_idle){
685                 gfs2_log_lock(sdp);
686                 atomic_dec(&sdp->sd_log_blks_free); /* Adjust for unreserved buffer */
687                 trace_gfs2_log_blocks(sdp, -1);
688                 gfs2_log_unlock(sdp);
689                 log_write_header(sdp, 0, PULL);
690         }
691         lops_after_commit(sdp, ai);
692
693         gfs2_log_lock(sdp);
694         sdp->sd_log_head = sdp->sd_log_flush_head;
695         sdp->sd_log_blks_reserved = 0;
696         sdp->sd_log_commited_buf = 0;
697         sdp->sd_log_commited_databuf = 0;
698         sdp->sd_log_commited_revoke = 0;
699
700         spin_lock(&sdp->sd_ail_lock);
701         if (!list_empty(&ai->ai_ail1_list)) {
702                 list_add(&ai->ai_list, &sdp->sd_ail1_list);
703                 ai = NULL;
704         }
705         spin_unlock(&sdp->sd_ail_lock);
706         gfs2_log_unlock(sdp);
707         trace_gfs2_log_flush(sdp, 0);
708         up_write(&sdp->sd_log_flush_lock);
709
710         kfree(ai);
711 }
712
713 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
714 {
715         unsigned int reserved;
716         unsigned int unused;
717
718         gfs2_log_lock(sdp);
719
720         sdp->sd_log_commited_buf += tr->tr_num_buf_new - tr->tr_num_buf_rm;
721         sdp->sd_log_commited_databuf += tr->tr_num_databuf_new -
722                 tr->tr_num_databuf_rm;
723         gfs2_assert_withdraw(sdp, (((int)sdp->sd_log_commited_buf) >= 0) ||
724                              (((int)sdp->sd_log_commited_databuf) >= 0));
725         sdp->sd_log_commited_revoke += tr->tr_num_revoke - tr->tr_num_revoke_rm;
726         reserved = calc_reserved(sdp);
727         gfs2_assert_withdraw(sdp, sdp->sd_log_blks_reserved + tr->tr_reserved >= reserved);
728         unused = sdp->sd_log_blks_reserved - reserved + tr->tr_reserved;
729         atomic_add(unused, &sdp->sd_log_blks_free);
730         trace_gfs2_log_blocks(sdp, unused);
731         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
732                              sdp->sd_jdesc->jd_blocks);
733         sdp->sd_log_blks_reserved = reserved;
734
735         gfs2_log_unlock(sdp);
736 }
737
738 static void buf_lo_incore_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
739 {
740         struct list_head *head = &tr->tr_list_buf;
741         struct gfs2_bufdata *bd;
742
743         gfs2_log_lock(sdp);
744         while (!list_empty(head)) {
745                 bd = list_entry(head->next, struct gfs2_bufdata, bd_list_tr);
746                 list_del_init(&bd->bd_list_tr);
747                 tr->tr_num_buf--;
748         }
749         gfs2_log_unlock(sdp);
750         gfs2_assert_warn(sdp, !tr->tr_num_buf);
751 }
752
753 /**
754  * gfs2_log_commit - Commit a transaction to the log
755  * @sdp: the filesystem
756  * @tr: the transaction
757  *
758  * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
759  * or the total number of used blocks (pinned blocks plus AIL blocks)
760  * is greater than thresh2.
761  *
762  * At mount time thresh1 is 1/3rd of journal size, thresh2 is 2/3rd of
763  * journal size.
764  *
765  * Returns: errno
766  */
767
768 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
769 {
770         log_refund(sdp, tr);
771         buf_lo_incore_commit(sdp, tr);
772
773         up_read(&sdp->sd_log_flush_lock);
774
775         if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
776             ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
777             atomic_read(&sdp->sd_log_thresh2)))
778                 wake_up(&sdp->sd_logd_waitq);
779 }
780
781 /**
782  * gfs2_log_shutdown - write a shutdown header into a journal
783  * @sdp: the filesystem
784  *
785  */
786
787 void gfs2_log_shutdown(struct gfs2_sbd *sdp)
788 {
789         down_write(&sdp->sd_log_flush_lock);
790
791         gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
792         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_buf);
793         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
794         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_rg);
795         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_databuf);
796         gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
797
798         sdp->sd_log_flush_head = sdp->sd_log_head;
799         sdp->sd_log_flush_wrapped = 0;
800
801         log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT,
802                          (sdp->sd_log_tail == current_tail(sdp)) ? 0 : PULL);
803
804         gfs2_assert_warn(sdp, atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks);
805         gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
806         gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
807
808         sdp->sd_log_head = sdp->sd_log_flush_head;
809         sdp->sd_log_tail = sdp->sd_log_head;
810
811         up_write(&sdp->sd_log_flush_lock);
812 }
813
814
815 /**
816  * gfs2_meta_syncfs - sync all the buffers in a filesystem
817  * @sdp: the filesystem
818  *
819  */
820
821 void gfs2_meta_syncfs(struct gfs2_sbd *sdp)
822 {
823         gfs2_log_flush(sdp, NULL);
824         for (;;) {
825                 gfs2_ail1_start(sdp);
826                 gfs2_ail1_wait(sdp);
827                 if (gfs2_ail1_empty(sdp))
828                         break;
829         }
830         gfs2_log_flush(sdp, NULL);
831 }
832
833 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
834 {
835         return (atomic_read(&sdp->sd_log_pinned) >= atomic_read(&sdp->sd_log_thresh1));
836 }
837
838 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
839 {
840         unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
841         return used_blocks >= atomic_read(&sdp->sd_log_thresh2);
842 }
843
844 /**
845  * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
846  * @sdp: Pointer to GFS2 superblock
847  *
848  * Also, periodically check to make sure that we're using the most recent
849  * journal index.
850  */
851
852 int gfs2_logd(void *data)
853 {
854         struct gfs2_sbd *sdp = data;
855         unsigned long t = 1;
856         DEFINE_WAIT(wait);
857         unsigned preflush;
858
859         while (!kthread_should_stop()) {
860
861                 preflush = atomic_read(&sdp->sd_log_pinned);
862                 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
863                         gfs2_ail1_empty(sdp);
864                         gfs2_log_flush(sdp, NULL);
865                 }
866
867                 if (gfs2_ail_flush_reqd(sdp)) {
868                         gfs2_ail1_start(sdp);
869                         gfs2_ail1_wait(sdp);
870                         gfs2_ail1_empty(sdp);
871                         gfs2_log_flush(sdp, NULL);
872                 }
873
874                 if (!gfs2_ail_flush_reqd(sdp))
875                         wake_up(&sdp->sd_log_waitq);
876
877                 t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
878
879                 try_to_freeze();
880
881                 do {
882                         prepare_to_wait(&sdp->sd_logd_waitq, &wait,
883                                         TASK_INTERRUPTIBLE);
884                         if (!gfs2_ail_flush_reqd(sdp) &&
885                             !gfs2_jrnl_flush_reqd(sdp) &&
886                             !kthread_should_stop())
887                                 t = schedule_timeout(t);
888                 } while(t && !gfs2_ail_flush_reqd(sdp) &&
889                         !gfs2_jrnl_flush_reqd(sdp) &&
890                         !kthread_should_stop());
891                 finish_wait(&sdp->sd_logd_waitq, &wait);
892         }
893
894         return 0;
895 }
896