2b9f0d9b1b287d66456361b542b47a19a0c5e2ef
[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  * log_write_header - Get and initialize a journal header buffer
496  * @sdp: The GFS2 superblock
497  *
498  * Returns: the initialized log buffer descriptor
499  */
500
501 static void log_write_header(struct gfs2_sbd *sdp, u32 flags, int pull)
502 {
503         u64 blkno = gfs2_log_bmap(sdp, sdp->sd_log_flush_head);
504         struct buffer_head *bh;
505         struct gfs2_log_header *lh;
506         unsigned int tail;
507         u32 hash;
508
509         bh = sb_getblk(sdp->sd_vfs, blkno);
510         lock_buffer(bh);
511         memset(bh->b_data, 0, bh->b_size);
512         set_buffer_uptodate(bh);
513         clear_buffer_dirty(bh);
514
515         gfs2_ail1_empty(sdp);
516         tail = current_tail(sdp);
517
518         lh = (struct gfs2_log_header *)bh->b_data;
519         memset(lh, 0, sizeof(struct gfs2_log_header));
520         lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
521         lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
522         lh->lh_header.__pad0 = cpu_to_be64(0);
523         lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
524         lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
525         lh->lh_sequence = cpu_to_be64(sdp->sd_log_sequence++);
526         lh->lh_flags = cpu_to_be32(flags);
527         lh->lh_tail = cpu_to_be32(tail);
528         lh->lh_blkno = cpu_to_be32(sdp->sd_log_flush_head);
529         hash = gfs2_disk_hash(bh->b_data, sizeof(struct gfs2_log_header));
530         lh->lh_hash = cpu_to_be32(hash);
531
532         bh->b_end_io = end_buffer_write_sync;
533         get_bh(bh);
534         if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags))
535                 submit_bh(WRITE_SYNC | REQ_META | REQ_PRIO, bh);
536         else
537                 submit_bh(WRITE_FLUSH_FUA | REQ_META, bh);
538         wait_on_buffer(bh);
539
540         if (!buffer_uptodate(bh))
541                 gfs2_io_error_bh(sdp, bh);
542         brelse(bh);
543
544         if (sdp->sd_log_tail != tail)
545                 log_pull_tail(sdp, tail);
546         else
547                 gfs2_assert_withdraw(sdp, !pull);
548
549         sdp->sd_log_idle = (tail == sdp->sd_log_flush_head);
550         gfs2_log_incr_head(sdp);
551 }
552
553 static void log_flush_commit(struct gfs2_sbd *sdp)
554 {
555         DEFINE_WAIT(wait);
556
557         if (atomic_read(&sdp->sd_log_in_flight)) {
558                 do {
559                         prepare_to_wait(&sdp->sd_log_flush_wait, &wait,
560                                         TASK_UNINTERRUPTIBLE);
561                         if (atomic_read(&sdp->sd_log_in_flight))
562                                 io_schedule();
563                 } while(atomic_read(&sdp->sd_log_in_flight));
564                 finish_wait(&sdp->sd_log_flush_wait, &wait);
565         }
566
567         log_write_header(sdp, 0, 0);
568 }
569
570 static int bd_cmp(void *priv, struct list_head *a, struct list_head *b)
571 {
572         struct gfs2_bufdata *bda, *bdb;
573
574         bda = list_entry(a, struct gfs2_bufdata, bd_le.le_list);
575         bdb = list_entry(b, struct gfs2_bufdata, bd_le.le_list);
576
577         if (bda->bd_bh->b_blocknr < bdb->bd_bh->b_blocknr)
578                 return -1;
579         if (bda->bd_bh->b_blocknr > bdb->bd_bh->b_blocknr)
580                 return 1;
581         return 0;
582 }
583
584 static void gfs2_ordered_write(struct gfs2_sbd *sdp)
585 {
586         struct gfs2_bufdata *bd;
587         struct buffer_head *bh;
588         LIST_HEAD(written);
589
590         gfs2_log_lock(sdp);
591         list_sort(NULL, &sdp->sd_log_le_ordered, &bd_cmp);
592         while (!list_empty(&sdp->sd_log_le_ordered)) {
593                 bd = list_entry(sdp->sd_log_le_ordered.next, struct gfs2_bufdata, bd_le.le_list);
594                 list_move(&bd->bd_le.le_list, &written);
595                 bh = bd->bd_bh;
596                 if (!buffer_dirty(bh))
597                         continue;
598                 get_bh(bh);
599                 gfs2_log_unlock(sdp);
600                 lock_buffer(bh);
601                 if (buffer_mapped(bh) && test_clear_buffer_dirty(bh)) {
602                         bh->b_end_io = end_buffer_write_sync;
603                         submit_bh(WRITE_SYNC, bh);
604                 } else {
605                         unlock_buffer(bh);
606                         brelse(bh);
607                 }
608                 gfs2_log_lock(sdp);
609         }
610         list_splice(&written, &sdp->sd_log_le_ordered);
611         gfs2_log_unlock(sdp);
612 }
613
614 static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
615 {
616         struct gfs2_bufdata *bd;
617         struct buffer_head *bh;
618
619         gfs2_log_lock(sdp);
620         while (!list_empty(&sdp->sd_log_le_ordered)) {
621                 bd = list_entry(sdp->sd_log_le_ordered.prev, struct gfs2_bufdata, bd_le.le_list);
622                 bh = bd->bd_bh;
623                 if (buffer_locked(bh)) {
624                         get_bh(bh);
625                         gfs2_log_unlock(sdp);
626                         wait_on_buffer(bh);
627                         brelse(bh);
628                         gfs2_log_lock(sdp);
629                         continue;
630                 }
631                 list_del_init(&bd->bd_le.le_list);
632         }
633         gfs2_log_unlock(sdp);
634 }
635
636 /**
637  * gfs2_log_flush - flush incore transaction(s)
638  * @sdp: the filesystem
639  * @gl: The glock structure to flush.  If NULL, flush the whole incore log
640  *
641  */
642
643 void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl)
644 {
645         struct gfs2_ail *ai;
646
647         down_write(&sdp->sd_log_flush_lock);
648
649         /* Log might have been flushed while we waited for the flush lock */
650         if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags)) {
651                 up_write(&sdp->sd_log_flush_lock);
652                 return;
653         }
654         trace_gfs2_log_flush(sdp, 1);
655
656         ai = kzalloc(sizeof(struct gfs2_ail), GFP_NOFS | __GFP_NOFAIL);
657         INIT_LIST_HEAD(&ai->ai_ail1_list);
658         INIT_LIST_HEAD(&ai->ai_ail2_list);
659
660         if (sdp->sd_log_num_buf != sdp->sd_log_commited_buf) {
661                 printk(KERN_INFO "GFS2: log buf %u %u\n", sdp->sd_log_num_buf,
662                        sdp->sd_log_commited_buf);
663                 gfs2_assert_withdraw(sdp, 0);
664         }
665         if (sdp->sd_log_num_databuf != sdp->sd_log_commited_databuf) {
666                 printk(KERN_INFO "GFS2: log databuf %u %u\n",
667                        sdp->sd_log_num_databuf, sdp->sd_log_commited_databuf);
668                 gfs2_assert_withdraw(sdp, 0);
669         }
670         gfs2_assert_withdraw(sdp,
671                         sdp->sd_log_num_revoke == sdp->sd_log_commited_revoke);
672
673         sdp->sd_log_flush_head = sdp->sd_log_head;
674         sdp->sd_log_flush_wrapped = 0;
675         ai->ai_first = sdp->sd_log_flush_head;
676
677         gfs2_ordered_write(sdp);
678         lops_before_commit(sdp);
679         gfs2_ordered_wait(sdp);
680
681         if (sdp->sd_log_head != sdp->sd_log_flush_head)
682                 log_flush_commit(sdp);
683         else if (sdp->sd_log_tail != current_tail(sdp) && !sdp->sd_log_idle){
684                 gfs2_log_lock(sdp);
685                 atomic_dec(&sdp->sd_log_blks_free); /* Adjust for unreserved buffer */
686                 trace_gfs2_log_blocks(sdp, -1);
687                 gfs2_log_unlock(sdp);
688                 log_write_header(sdp, 0, PULL);
689         }
690         lops_after_commit(sdp, ai);
691
692         gfs2_log_lock(sdp);
693         sdp->sd_log_head = sdp->sd_log_flush_head;
694         sdp->sd_log_blks_reserved = 0;
695         sdp->sd_log_commited_buf = 0;
696         sdp->sd_log_commited_databuf = 0;
697         sdp->sd_log_commited_revoke = 0;
698
699         spin_lock(&sdp->sd_ail_lock);
700         if (!list_empty(&ai->ai_ail1_list)) {
701                 list_add(&ai->ai_list, &sdp->sd_ail1_list);
702                 ai = NULL;
703         }
704         spin_unlock(&sdp->sd_ail_lock);
705         gfs2_log_unlock(sdp);
706         trace_gfs2_log_flush(sdp, 0);
707         up_write(&sdp->sd_log_flush_lock);
708
709         kfree(ai);
710 }
711
712 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
713 {
714         unsigned int reserved;
715         unsigned int unused;
716
717         gfs2_log_lock(sdp);
718
719         sdp->sd_log_commited_buf += tr->tr_num_buf_new - tr->tr_num_buf_rm;
720         sdp->sd_log_commited_databuf += tr->tr_num_databuf_new -
721                 tr->tr_num_databuf_rm;
722         gfs2_assert_withdraw(sdp, (((int)sdp->sd_log_commited_buf) >= 0) ||
723                              (((int)sdp->sd_log_commited_databuf) >= 0));
724         sdp->sd_log_commited_revoke += tr->tr_num_revoke - tr->tr_num_revoke_rm;
725         reserved = calc_reserved(sdp);
726         gfs2_assert_withdraw(sdp, sdp->sd_log_blks_reserved + tr->tr_reserved >= reserved);
727         unused = sdp->sd_log_blks_reserved - reserved + tr->tr_reserved;
728         atomic_add(unused, &sdp->sd_log_blks_free);
729         trace_gfs2_log_blocks(sdp, unused);
730         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
731                              sdp->sd_jdesc->jd_blocks);
732         sdp->sd_log_blks_reserved = reserved;
733
734         gfs2_log_unlock(sdp);
735 }
736
737 static void buf_lo_incore_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
738 {
739         struct list_head *head = &tr->tr_list_buf;
740         struct gfs2_bufdata *bd;
741
742         gfs2_log_lock(sdp);
743         while (!list_empty(head)) {
744                 bd = list_entry(head->next, struct gfs2_bufdata, bd_list_tr);
745                 list_del_init(&bd->bd_list_tr);
746                 tr->tr_num_buf--;
747         }
748         gfs2_log_unlock(sdp);
749         gfs2_assert_warn(sdp, !tr->tr_num_buf);
750 }
751
752 /**
753  * gfs2_log_commit - Commit a transaction to the log
754  * @sdp: the filesystem
755  * @tr: the transaction
756  *
757  * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
758  * or the total number of used blocks (pinned blocks plus AIL blocks)
759  * is greater than thresh2.
760  *
761  * At mount time thresh1 is 1/3rd of journal size, thresh2 is 2/3rd of
762  * journal size.
763  *
764  * Returns: errno
765  */
766
767 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
768 {
769         log_refund(sdp, tr);
770         buf_lo_incore_commit(sdp, tr);
771
772         up_read(&sdp->sd_log_flush_lock);
773
774         if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
775             ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
776             atomic_read(&sdp->sd_log_thresh2)))
777                 wake_up(&sdp->sd_logd_waitq);
778 }
779
780 /**
781  * gfs2_log_shutdown - write a shutdown header into a journal
782  * @sdp: the filesystem
783  *
784  */
785
786 void gfs2_log_shutdown(struct gfs2_sbd *sdp)
787 {
788         down_write(&sdp->sd_log_flush_lock);
789
790         gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
791         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_buf);
792         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
793         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_rg);
794         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_databuf);
795         gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
796
797         sdp->sd_log_flush_head = sdp->sd_log_head;
798         sdp->sd_log_flush_wrapped = 0;
799
800         log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT,
801                          (sdp->sd_log_tail == current_tail(sdp)) ? 0 : PULL);
802
803         gfs2_assert_warn(sdp, atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks);
804         gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
805         gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
806
807         sdp->sd_log_head = sdp->sd_log_flush_head;
808         sdp->sd_log_tail = sdp->sd_log_head;
809
810         up_write(&sdp->sd_log_flush_lock);
811 }
812
813
814 /**
815  * gfs2_meta_syncfs - sync all the buffers in a filesystem
816  * @sdp: the filesystem
817  *
818  */
819
820 void gfs2_meta_syncfs(struct gfs2_sbd *sdp)
821 {
822         gfs2_log_flush(sdp, NULL);
823         for (;;) {
824                 gfs2_ail1_start(sdp);
825                 gfs2_ail1_wait(sdp);
826                 if (gfs2_ail1_empty(sdp))
827                         break;
828         }
829         gfs2_log_flush(sdp, NULL);
830 }
831
832 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
833 {
834         return (atomic_read(&sdp->sd_log_pinned) >= atomic_read(&sdp->sd_log_thresh1));
835 }
836
837 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
838 {
839         unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
840         return used_blocks >= atomic_read(&sdp->sd_log_thresh2);
841 }
842
843 /**
844  * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
845  * @sdp: Pointer to GFS2 superblock
846  *
847  * Also, periodically check to make sure that we're using the most recent
848  * journal index.
849  */
850
851 int gfs2_logd(void *data)
852 {
853         struct gfs2_sbd *sdp = data;
854         unsigned long t = 1;
855         DEFINE_WAIT(wait);
856         unsigned preflush;
857
858         while (!kthread_should_stop()) {
859
860                 preflush = atomic_read(&sdp->sd_log_pinned);
861                 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
862                         gfs2_ail1_empty(sdp);
863                         gfs2_log_flush(sdp, NULL);
864                 }
865
866                 if (gfs2_ail_flush_reqd(sdp)) {
867                         gfs2_ail1_start(sdp);
868                         gfs2_ail1_wait(sdp);
869                         gfs2_ail1_empty(sdp);
870                         gfs2_log_flush(sdp, NULL);
871                 }
872
873                 if (!gfs2_ail_flush_reqd(sdp))
874                         wake_up(&sdp->sd_log_waitq);
875
876                 t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
877
878                 try_to_freeze();
879
880                 do {
881                         prepare_to_wait(&sdp->sd_logd_waitq, &wait,
882                                         TASK_INTERRUPTIBLE);
883                         if (!gfs2_ail_flush_reqd(sdp) &&
884                             !gfs2_jrnl_flush_reqd(sdp) &&
885                             !kthread_should_stop())
886                                 t = schedule_timeout(t);
887                 } while(t && !gfs2_ail_flush_reqd(sdp) &&
888                         !gfs2_jrnl_flush_reqd(sdp) &&
889                         !kthread_should_stop());
890                 finish_wait(&sdp->sd_logd_waitq, &wait);
891         }
892
893         return 0;
894 }
895