2 * Copyright (C) International Business Machines Corp., 2000-2004
3 * Portions Copyright (C) Christoph Hellwig, 2001-2002
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See
13 * the GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 * jfs_logmgr.c: log manager
23 * for related information, see transaction manager (jfs_txnmgr.c), and
24 * recovery manager (jfs_logredo.c).
26 * note: for detail, RTFS.
29 * special purpose buffer manager supporting log i/o requirements.
30 * per log serial pageout of logpage
31 * queuing i/o requests and redrive i/o at iodone
32 * maintain current logpage buffer
33 * no caching since append only
34 * appropriate jfs buffer cache buffers as needed
37 * transactions which wrote COMMIT records in the same in-memory
38 * log page during the pageout of previous/current log page(s) are
39 * committed together by the pageout of the page.
42 * transactions are committed asynchronously when the log page
43 * containing it COMMIT is paged out when it becomes full;
46 * . a per log lock serialize log write.
47 * . a per log lock serialize group commit.
48 * . a per log lock serialize log open/close;
51 * careful-write (ping-pong) of last logpage to recover from crash
53 * detection of split (out-of-order) write of physical sectors
54 * of last logpage via timestamp at end of each sector
55 * with its mirror data array at trailer).
58 * lsn - 64-bit monotonically increasing integer vs
59 * 32-bit lspn and page eor.
63 #include <linux/blkdev.h>
64 #include <linux/interrupt.h>
65 #include <linux/smp_lock.h>
66 #include <linux/completion.h>
67 #include <linux/buffer_head.h> /* for sync_blockdev() */
68 #include <linux/bio.h>
69 #include <linux/suspend.h>
70 #include <linux/delay.h>
71 #include "jfs_incore.h"
72 #include "jfs_filsys.h"
73 #include "jfs_metapage.h"
74 #include "jfs_txnmgr.h"
75 #include "jfs_debug.h"
79 * lbuf's ready to be redriven. Protected by log_redrive_lock (jfsIO thread)
81 static struct lbuf *log_redrive_list;
82 static DEFINE_SPINLOCK(log_redrive_lock);
83 DECLARE_WAIT_QUEUE_HEAD(jfs_IO_thread_wait);
87 * log read/write serialization (per log)
89 #define LOG_LOCK_INIT(log) init_MUTEX(&(log)->loglock)
90 #define LOG_LOCK(log) down(&((log)->loglock))
91 #define LOG_UNLOCK(log) up(&((log)->loglock))
95 * log group commit serialization (per log)
98 #define LOGGC_LOCK_INIT(log) spin_lock_init(&(log)->gclock)
99 #define LOGGC_LOCK(log) spin_lock_irq(&(log)->gclock)
100 #define LOGGC_UNLOCK(log) spin_unlock_irq(&(log)->gclock)
101 #define LOGGC_WAKEUP(tblk) wake_up_all(&(tblk)->gcwait)
104 * log sync serialization (per log)
106 #define LOGSYNC_DELTA(logsize) min((logsize)/8, 128*LOGPSIZE)
107 #define LOGSYNC_BARRIER(logsize) ((logsize)/4)
109 #define LOGSYNC_DELTA(logsize) min((logsize)/4, 256*LOGPSIZE)
110 #define LOGSYNC_BARRIER(logsize) ((logsize)/2)
115 * log buffer cache synchronization
117 static DEFINE_SPINLOCK(jfsLCacheLock);
119 #define LCACHE_LOCK(flags) spin_lock_irqsave(&jfsLCacheLock, flags)
120 #define LCACHE_UNLOCK(flags) spin_unlock_irqrestore(&jfsLCacheLock, flags)
123 * See __SLEEP_COND in jfs_locks.h
125 #define LCACHE_SLEEP_COND(wq, cond, flags) \
129 __SLEEP_COND(wq, cond, LCACHE_LOCK(flags), LCACHE_UNLOCK(flags)); \
132 #define LCACHE_WAKEUP(event) wake_up(event)
136 * lbuf buffer cache (lCache) control
138 /* log buffer manager pageout control (cumulative, inclusive) */
139 #define lbmREAD 0x0001
140 #define lbmWRITE 0x0002 /* enqueue at tail of write queue;
141 * init pageout if at head of queue;
143 #define lbmRELEASE 0x0004 /* remove from write queue
144 * at completion of pageout;
145 * do not free/recycle it yet:
146 * caller will free it;
148 #define lbmSYNC 0x0008 /* do not return to freelist
149 * when removed from write queue;
151 #define lbmFREE 0x0010 /* return to freelist
152 * at completion of pageout;
153 * the buffer may be recycled;
155 #define lbmDONE 0x0020
156 #define lbmERROR 0x0040
157 #define lbmGC 0x0080 /* lbmIODone to perform post-GC processing
160 #define lbmDIRECT 0x0100
163 * Global list of active external journals
165 static LIST_HEAD(jfs_external_logs);
166 static struct jfs_log *dummy_log = NULL;
167 static DECLARE_MUTEX(jfs_log_sem);
170 * external references
172 extern void txLazyUnlock(struct tblock * tblk);
173 extern int jfs_stop_threads;
174 extern struct completion jfsIOwait;
175 extern int jfs_tlocks_low;
180 static int lmWriteRecord(struct jfs_log * log, struct tblock * tblk,
181 struct lrd * lrd, struct tlock * tlck);
183 static int lmNextPage(struct jfs_log * log);
184 static int lmLogFileSystem(struct jfs_log * log, struct jfs_sb_info *sbi,
187 static int open_inline_log(struct super_block *sb);
188 static int open_dummy_log(struct super_block *sb);
189 static int lbmLogInit(struct jfs_log * log);
190 static void lbmLogShutdown(struct jfs_log * log);
191 static struct lbuf *lbmAllocate(struct jfs_log * log, int);
192 static void lbmFree(struct lbuf * bp);
193 static void lbmfree(struct lbuf * bp);
194 static int lbmRead(struct jfs_log * log, int pn, struct lbuf ** bpp);
195 static void lbmWrite(struct jfs_log * log, struct lbuf * bp, int flag, int cant_block);
196 static void lbmDirectWrite(struct jfs_log * log, struct lbuf * bp, int flag);
197 static int lbmIOWait(struct lbuf * bp, int flag);
198 static bio_end_io_t lbmIODone;
199 static void lbmStartIO(struct lbuf * bp);
200 static void lmGCwrite(struct jfs_log * log, int cant_block);
201 static int lmLogSync(struct jfs_log * log, int nosyncwait);
208 #ifdef CONFIG_JFS_STATISTICS
209 static struct lmStat {
210 uint commit; /* # of commit */
211 uint pagedone; /* # of page written */
212 uint submitted; /* # of pages submitted */
213 uint full_page; /* # of full pages submitted */
214 uint partial_page; /* # of partial pages submitted */
222 * FUNCTION: write a log record;
226 * RETURN: lsn - offset to the next log record to write (end-of-log);
229 * note: todo: log error handler
231 int lmLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
236 struct metapage *mp = NULL;
239 jfs_info("lmLog: log:0x%p tblk:0x%p, lrd:0x%p tlck:0x%p",
240 log, tblk, lrd, tlck);
244 /* log by (out-of-transaction) JFS ? */
248 /* log from page ? */
250 tlck->type & tlckBTROOT || (mp = tlck->mp) == NULL)
254 * initialize/update page/transaction recovery lsn
258 LOGSYNC_LOCK(log, flags);
261 * initialize page lsn if first log write of the page
268 /* insert page at tail of logsynclist */
269 list_add_tail(&mp->synclist, &log->synclist);
273 * initialize/update lsn of tblock of the page
275 * transaction inherits oldest lsn of pages associated
276 * with allocation/deallocation of resources (their
277 * log records are used to reconstruct allocation map
278 * at recovery time: inode for inode allocation map,
279 * B+-tree index of extent descriptors for block
281 * allocation map pages inherit transaction lsn at
282 * commit time to allow forwarding log syncpt past log
283 * records associated with allocation/deallocation of
284 * resources only after persistent map of these map pages
285 * have been updated and propagated to home.
288 * initialize transaction lsn:
290 if (tblk->lsn == 0) {
291 /* inherit lsn of its first page logged */
295 /* insert tblock after the page on logsynclist */
296 list_add(&tblk->synclist, &mp->synclist);
299 * update transaction lsn:
302 /* inherit oldest/smallest lsn of page */
303 logdiff(diffp, mp->lsn, log);
304 logdiff(difft, tblk->lsn, log);
306 /* update tblock lsn with page lsn */
309 /* move tblock after page on logsynclist */
310 list_move(&tblk->synclist, &mp->synclist);
314 LOGSYNC_UNLOCK(log, flags);
317 * write the log record
320 lsn = lmWriteRecord(log, tblk, lrd, tlck);
323 * forward log syncpt if log reached next syncpt trigger
325 logdiff(diffp, lsn, log);
326 if (diffp >= log->nextsync)
327 lsn = lmLogSync(log, 0);
329 /* update end-of-log lsn */
334 /* return end-of-log address */
339 * NAME: lmWriteRecord()
341 * FUNCTION: move the log record to current log page
343 * PARAMETER: cd - commit descriptor
345 * RETURN: end-of-log address
347 * serialization: LOG_LOCK() held on entry/exit
350 lmWriteRecord(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
353 int lsn = 0; /* end-of-log address */
354 struct lbuf *bp; /* dst log page buffer */
355 struct logpage *lp; /* dst log page */
356 caddr_t dst; /* destination address in log page */
357 int dstoffset; /* end-of-log offset in log page */
358 int freespace; /* free space in log page */
359 caddr_t p; /* src meta-data page */
362 int nbytes; /* number of bytes to move */
365 struct linelock *linelock;
372 /* retrieve destination log page to write */
373 bp = (struct lbuf *) log->bp;
374 lp = (struct logpage *) bp->l_ldata;
375 dstoffset = log->eor;
377 /* any log data to write ? */
382 * move log record data
384 /* retrieve source meta-data page to log */
385 if (tlck->flag & tlckPAGELOCK) {
386 p = (caddr_t) (tlck->mp->data);
387 linelock = (struct linelock *) & tlck->lock;
389 /* retrieve source in-memory inode to log */
390 else if (tlck->flag & tlckINODELOCK) {
391 if (tlck->type & tlckDTREE)
392 p = (caddr_t) &JFS_IP(tlck->ip)->i_dtroot;
394 p = (caddr_t) &JFS_IP(tlck->ip)->i_xtroot;
395 linelock = (struct linelock *) & tlck->lock;
398 else if (tlck->flag & tlckINLINELOCK) {
400 inlinelock = (struct inlinelock *) & tlck;
401 p = (caddr_t) & inlinelock->pxd;
402 linelock = (struct linelock *) & tlck;
404 #endif /* _JFS_WIP */
406 jfs_err("lmWriteRecord: UFO tlck:0x%p", tlck);
407 return 0; /* Probably should trap */
409 l2linesize = linelock->l2linesize;
412 ASSERT(linelock->index <= linelock->maxcnt);
415 for (i = 0; i < linelock->index; i++, lv++) {
420 if (dstoffset >= LOGPSIZE - LOGPTLRSIZE) {
421 /* page become full: move on to next page */
425 lp = (struct logpage *) bp->l_ldata;
426 dstoffset = LOGPHDRSIZE;
430 * move log vector data
432 src = (u8 *) p + (lv->offset << l2linesize);
433 srclen = lv->length << l2linesize;
436 freespace = (LOGPSIZE - LOGPTLRSIZE) - dstoffset;
437 nbytes = min(freespace, srclen);
438 dst = (caddr_t) lp + dstoffset;
439 memcpy(dst, src, nbytes);
442 /* is page not full ? */
443 if (dstoffset < LOGPSIZE - LOGPTLRSIZE)
446 /* page become full: move on to next page */
449 bp = (struct lbuf *) log->bp;
450 lp = (struct logpage *) bp->l_ldata;
451 dstoffset = LOGPHDRSIZE;
458 * move log vector descriptor
461 lvd = (struct lvd *) ((caddr_t) lp + dstoffset);
462 lvd->offset = cpu_to_le16(lv->offset);
463 lvd->length = cpu_to_le16(lv->length);
465 jfs_info("lmWriteRecord: lv offset:%d length:%d",
466 lv->offset, lv->length);
469 if ((i = linelock->next)) {
470 linelock = (struct linelock *) lid_to_tlock(i);
475 * move log record descriptor
478 lrd->length = cpu_to_le16(len);
484 freespace = (LOGPSIZE - LOGPTLRSIZE) - dstoffset;
485 nbytes = min(freespace, srclen);
486 dst = (caddr_t) lp + dstoffset;
487 memcpy(dst, src, nbytes);
492 /* are there more to move than freespace of page ? */
497 * end of log record descriptor
500 /* update last log record eor */
501 log->eor = dstoffset;
502 bp->l_eor = dstoffset;
503 lsn = (log->page << L2LOGPSIZE) + dstoffset;
505 if (lrd->type & cpu_to_le16(LOG_COMMIT)) {
507 jfs_info("wr: tclsn:0x%x, beor:0x%x", tblk->clsn,
510 INCREMENT(lmStat.commit); /* # of commit */
513 * enqueue tblock for group commit:
515 * enqueue tblock of non-trivial/synchronous COMMIT
516 * at tail of group commit queue
517 * (trivial/asynchronous COMMITs are ignored by
522 /* init tblock gc state */
523 tblk->flag = tblkGC_QUEUE;
525 tblk->pn = log->page;
526 tblk->eor = log->eor;
528 /* enqueue transaction to commit queue */
529 list_add_tail(&tblk->cqueue, &log->cqueue);
534 jfs_info("lmWriteRecord: lrd:0x%04x bp:0x%p pn:%d eor:0x%x",
535 le16_to_cpu(lrd->type), log->bp, log->page, dstoffset);
537 /* page not full ? */
538 if (dstoffset < LOGPSIZE - LOGPTLRSIZE)
542 /* page become full: move on to next page */
545 bp = (struct lbuf *) log->bp;
546 lp = (struct logpage *) bp->l_ldata;
547 dstoffset = LOGPHDRSIZE;
558 * FUNCTION: write current page and allocate next page.
564 * serialization: LOG_LOCK() held on entry/exit
566 static int lmNextPage(struct jfs_log * log)
569 int lspn; /* log sequence page number */
570 int pn; /* current page number */
575 /* get current log page number and log sequence page number */
578 lp = (struct logpage *) bp->l_ldata;
579 lspn = le32_to_cpu(lp->h.page);
584 * write or queue the full page at the tail of write queue
586 /* get the tail tblk on commit queue */
587 if (list_empty(&log->cqueue))
590 tblk = list_entry(log->cqueue.prev, struct tblock, cqueue);
592 /* every tblk who has COMMIT record on the current page,
593 * and has not been committed, must be on commit queue
594 * since tblk is queued at commit queueu at the time
595 * of writing its COMMIT record on the page before
596 * page becomes full (even though the tblk thread
597 * who wrote COMMIT record may have been suspended
601 /* is page bound with outstanding tail tblk ? */
602 if (tblk && tblk->pn == pn) {
603 /* mark tblk for end-of-page */
604 tblk->flag |= tblkGC_EOP;
606 if (log->cflag & logGC_PAGEOUT) {
607 /* if page is not already on write queue,
608 * just enqueue (no lbmWRITE to prevent redrive)
609 * buffer to wqueue to ensure correct serial order
610 * of the pages since log pages will be added
613 if (bp->l_wqnext == NULL)
614 lbmWrite(log, bp, 0, 0);
617 * No current GC leader, initiate group commit
619 log->cflag |= logGC_PAGEOUT;
623 /* page is not bound with outstanding tblk:
624 * init write or mark it to be redriven (lbmWRITE)
627 /* finalize the page */
628 bp->l_ceor = bp->l_eor;
629 lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_ceor);
630 lbmWrite(log, bp, lbmWRITE | lbmRELEASE | lbmFREE, 0);
635 * allocate/initialize next page
637 /* if log wraps, the first data page of log is 2
638 * (0 never used, 1 is superblock).
640 log->page = (pn == log->size - 1) ? 2 : pn + 1;
641 log->eor = LOGPHDRSIZE; /* ? valid page empty/full at logRedo() */
643 /* allocate/initialize next log page buffer */
644 nextbp = lbmAllocate(log, log->page);
645 nextbp->l_eor = log->eor;
648 /* initialize next log page */
649 lp = (struct logpage *) nextbp->l_ldata;
650 lp->h.page = lp->t.page = cpu_to_le32(lspn + 1);
651 lp->h.eor = lp->t.eor = cpu_to_le16(LOGPHDRSIZE);
658 * NAME: lmGroupCommit()
660 * FUNCTION: group commit
661 * initiate pageout of the pages with COMMIT in the order of
662 * page number - redrive pageout of the page at the head of
663 * pageout queue until full page has been written.
668 * LOGGC_LOCK serializes log group commit queue, and
669 * transaction blocks on the commit queue.
670 * N.B. LOG_LOCK is NOT held during lmGroupCommit().
672 int lmGroupCommit(struct jfs_log * log, struct tblock * tblk)
678 /* group committed already ? */
679 if (tblk->flag & tblkGC_COMMITTED) {
680 if (tblk->flag & tblkGC_ERROR)
686 jfs_info("lmGroup Commit: tblk = 0x%p, gcrtc = %d", tblk, log->gcrtc);
688 if (tblk->xflag & COMMIT_LAZY)
689 tblk->flag |= tblkGC_LAZY;
691 if ((!(log->cflag & logGC_PAGEOUT)) && (!list_empty(&log->cqueue)) &&
692 (!(tblk->xflag & COMMIT_LAZY) || test_bit(log_FLUSH, &log->flag)
693 || jfs_tlocks_low)) {
695 * No pageout in progress
697 * start group commit as its group leader.
699 log->cflag |= logGC_PAGEOUT;
704 if (tblk->xflag & COMMIT_LAZY) {
706 * Lazy transactions can leave now
712 /* lmGCwrite gives up LOGGC_LOCK, check again */
714 if (tblk->flag & tblkGC_COMMITTED) {
715 if (tblk->flag & tblkGC_ERROR)
722 /* upcount transaction waiting for completion
725 tblk->flag |= tblkGC_READY;
727 __SLEEP_COND(tblk->gcwait, (tblk->flag & tblkGC_COMMITTED),
728 LOGGC_LOCK(log), LOGGC_UNLOCK(log));
730 /* removed from commit queue */
731 if (tblk->flag & tblkGC_ERROR)
741 * FUNCTION: group commit write
742 * initiate write of log page, building a group of all transactions
743 * with commit records on that page.
748 * LOGGC_LOCK must be held by caller.
749 * N.B. LOG_LOCK is NOT held during lmGroupCommit().
751 static void lmGCwrite(struct jfs_log * log, int cant_write)
755 int gcpn; /* group commit page number */
757 struct tblock *xtblk = NULL;
760 * build the commit group of a log page
762 * scan commit queue and make a commit group of all
763 * transactions with COMMIT records on the same log page.
765 /* get the head tblk on the commit queue */
766 gcpn = list_entry(log->cqueue.next, struct tblock, cqueue)->pn;
768 list_for_each_entry(tblk, &log->cqueue, cqueue) {
769 if (tblk->pn != gcpn)
774 /* state transition: (QUEUE, READY) -> COMMIT */
775 tblk->flag |= tblkGC_COMMIT;
777 tblk = xtblk; /* last tblk of the page */
780 * pageout to commit transactions on the log page.
782 bp = (struct lbuf *) tblk->bp;
783 lp = (struct logpage *) bp->l_ldata;
784 /* is page already full ? */
785 if (tblk->flag & tblkGC_EOP) {
786 /* mark page to free at end of group commit of the page */
787 tblk->flag &= ~tblkGC_EOP;
788 tblk->flag |= tblkGC_FREE;
789 bp->l_ceor = bp->l_eor;
790 lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_ceor);
791 lbmWrite(log, bp, lbmWRITE | lbmRELEASE | lbmGC,
793 INCREMENT(lmStat.full_page);
795 /* page is not yet full */
797 bp->l_ceor = tblk->eor; /* ? bp->l_ceor = bp->l_eor; */
798 lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_ceor);
799 lbmWrite(log, bp, lbmWRITE | lbmGC, cant_write);
800 INCREMENT(lmStat.partial_page);
807 * FUNCTION: group commit post-processing
808 * Processes transactions after their commit records have been written
809 * to disk, redriving log I/O if necessary.
814 * This routine is called a interrupt time by lbmIODone
816 static void lmPostGC(struct lbuf * bp)
819 struct jfs_log *log = bp->l_log;
821 struct tblock *tblk, *temp;
824 spin_lock_irqsave(&log->gclock, flags);
826 * current pageout of group commit completed.
828 * remove/wakeup transactions from commit queue who were
829 * group committed with the current log page
831 list_for_each_entry_safe(tblk, temp, &log->cqueue, cqueue) {
832 if (!(tblk->flag & tblkGC_COMMIT))
834 /* if transaction was marked GC_COMMIT then
835 * it has been shipped in the current pageout
836 * and made it to disk - it is committed.
839 if (bp->l_flag & lbmERROR)
840 tblk->flag |= tblkGC_ERROR;
842 /* remove it from the commit queue */
843 list_del(&tblk->cqueue);
844 tblk->flag &= ~tblkGC_QUEUE;
846 if (tblk == log->flush_tblk) {
847 /* we can stop flushing the log now */
848 clear_bit(log_FLUSH, &log->flag);
849 log->flush_tblk = NULL;
852 jfs_info("lmPostGC: tblk = 0x%p, flag = 0x%x", tblk,
855 if (!(tblk->xflag & COMMIT_FORCE))
857 * Hand tblk over to lazy commit thread
861 /* state transition: COMMIT -> COMMITTED */
862 tblk->flag |= tblkGC_COMMITTED;
864 if (tblk->flag & tblkGC_READY)
870 /* was page full before pageout ?
871 * (and this is the last tblk bound with the page)
873 if (tblk->flag & tblkGC_FREE)
875 /* did page become full after pageout ?
876 * (and this is the last tblk bound with the page)
878 else if (tblk->flag & tblkGC_EOP) {
879 /* finalize the page */
880 lp = (struct logpage *) bp->l_ldata;
881 bp->l_ceor = bp->l_eor;
882 lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_eor);
883 jfs_info("lmPostGC: calling lbmWrite");
884 lbmWrite(log, bp, lbmWRITE | lbmRELEASE | lbmFREE,
890 /* are there any transactions who have entered lnGroupCommit()
891 * (whose COMMITs are after that of the last log page written.
892 * They are waiting for new group commit (above at (SLEEP 1))
893 * or lazy transactions are on a full (queued) log page,
894 * select the latest ready transaction as new group leader and
895 * wake her up to lead her group.
897 if ((!list_empty(&log->cqueue)) &&
898 ((log->gcrtc > 0) || (tblk->bp->l_wqnext != NULL) ||
899 test_bit(log_FLUSH, &log->flag) || jfs_tlocks_low))
901 * Call lmGCwrite with new group leader
905 /* no transaction are ready yet (transactions are only just
906 * queued (GC_QUEUE) and not entered for group commit yet).
907 * the first transaction entering group commit
908 * will elect herself as new group leader.
911 log->cflag &= ~logGC_PAGEOUT;
914 spin_unlock_irqrestore(&log->gclock, flags);
921 * FUNCTION: write log SYNCPT record for specified log
922 * if new sync address is available
923 * (normally the case if sync() is executed by back-ground
925 * if not, explicitly run jfs_blogsync() to initiate
926 * getting of new sync address.
927 * calculate new value of i_nextsync which determines when
928 * this code is called again.
930 * PARAMETERS: log - log structure
931 * nosyncwait - 1 if called asynchronously
935 * serialization: LOG_LOCK() held on entry/exit
937 static int lmLogSync(struct jfs_log * log, int nosyncwait)
940 int written; /* written since last syncpt */
941 int free; /* free space left available */
942 int delta; /* additional delta to write normally */
943 int more; /* additional write granted */
946 struct logsyncblk *lp;
947 struct jfs_sb_info *sbi;
950 /* push dirty metapages out to disk */
951 list_for_each_entry(sbi, &log->sb_list, log_list) {
952 filemap_flush(sbi->ipbmap->i_mapping);
953 filemap_flush(sbi->ipimap->i_mapping);
954 filemap_flush(sbi->direct_inode->i_mapping);
960 /* if last sync is same as last syncpt,
961 * invoke sync point forward processing to update sync.
964 if (log->sync == log->syncpt) {
965 LOGSYNC_LOCK(log, flags);
966 if (list_empty(&log->synclist))
967 log->sync = log->lsn;
969 lp = list_entry(log->synclist.next,
970 struct logsyncblk, synclist);
973 LOGSYNC_UNLOCK(log, flags);
977 /* if sync is different from last syncpt,
978 * write a SYNCPT record with syncpt = sync.
979 * reset syncpt = sync
981 if (log->sync != log->syncpt) {
984 lrd.type = cpu_to_le16(LOG_SYNCPT);
986 lrd.log.syncpt.sync = cpu_to_le32(log->sync);
987 lsn = lmWriteRecord(log, NULL, &lrd, NULL);
989 log->syncpt = log->sync;
994 * setup next syncpt trigger (SWAG)
996 logsize = log->logsize;
998 logdiff(written, lsn, log);
999 free = logsize - written;
1000 delta = LOGSYNC_DELTA(logsize);
1001 more = min(free / 2, delta);
1002 if (more < 2 * LOGPSIZE) {
1003 jfs_warn("\n ... Log Wrap ... Log Wrap ... Log Wrap ...\n");
1007 * option 1 - panic ? No.!
1008 * option 2 - shutdown file systems
1009 * associated with log ?
1010 * option 3 - extend log ?
1013 * option 4 - second chance
1015 * mark log wrapped, and continue.
1016 * when all active transactions are completed,
1017 * mark log vaild for recovery.
1018 * if crashed during invalid state, log state
1019 * implies invald log, forcing fsck().
1021 /* mark log state log wrap in log superblock */
1022 /* log->state = LOGWRAP; */
1024 /* reset sync point computation */
1025 log->syncpt = log->sync = lsn;
1026 log->nextsync = delta;
1028 /* next syncpt trigger = written + more */
1029 log->nextsync = written + more;
1031 /* return if lmLogSync() from outside of transaction, e.g., sync() */
1035 /* if number of bytes written from last sync point is more
1036 * than 1/4 of the log size, stop new transactions from
1037 * starting until all current transactions are completed
1038 * by setting syncbarrier flag.
1040 if (written > LOGSYNC_BARRIER(logsize) && logsize > 32 * LOGPSIZE) {
1041 set_bit(log_SYNCBARRIER, &log->flag);
1042 jfs_info("log barrier on: lsn=0x%x syncpt=0x%x", lsn,
1045 * We may have to initiate group commit
1047 jfs_flush_journal(log, 0);
1056 * FUNCTION: write log SYNCPT record for specified log
1058 * PARAMETERS: log - log structure
1060 void jfs_syncpt(struct jfs_log *log)
1069 * FUNCTION: open the log on first open;
1070 * insert filesystem in the active list of the log.
1072 * PARAMETER: ipmnt - file system mount inode
1073 * iplog - log inode (out)
1079 int lmLogOpen(struct super_block *sb)
1082 struct block_device *bdev;
1083 struct jfs_log *log;
1084 struct jfs_sb_info *sbi = JFS_SBI(sb);
1086 if (sbi->flag & JFS_NOINTEGRITY)
1087 return open_dummy_log(sb);
1089 if (sbi->mntflag & JFS_INLINELOG)
1090 return open_inline_log(sb);
1093 list_for_each_entry(log, &jfs_external_logs, journal_list) {
1094 if (log->bdev->bd_dev == sbi->logdev) {
1095 if (memcmp(log->uuid, sbi->loguuid,
1096 sizeof(log->uuid))) {
1097 jfs_warn("wrong uuid on JFS journal\n");
1102 * add file system to log active file system list
1104 if ((rc = lmLogFileSystem(log, sbi, 1))) {
1112 if (!(log = kmalloc(sizeof(struct jfs_log), GFP_KERNEL))) {
1116 memset(log, 0, sizeof(struct jfs_log));
1117 INIT_LIST_HEAD(&log->sb_list);
1118 init_waitqueue_head(&log->syncwait);
1121 * external log as separate logical volume
1123 * file systems to log may have n-to-1 relationship;
1126 bdev = open_by_devnum(sbi->logdev, FMODE_READ|FMODE_WRITE);
1128 rc = -PTR_ERR(bdev);
1132 if ((rc = bd_claim(bdev, log))) {
1137 memcpy(log->uuid, sbi->loguuid, sizeof(log->uuid));
1142 if ((rc = lmLogInit(log)))
1145 list_add(&log->journal_list, &jfs_external_logs);
1148 * add file system to log active file system list
1150 if ((rc = lmLogFileSystem(log, sbi, 1)))
1155 list_add(&sbi->log_list, &log->sb_list);
1165 shutdown: /* unwind lbmLogInit() */
1166 list_del(&log->journal_list);
1167 lbmLogShutdown(log);
1172 close: /* close external log device */
1175 free: /* free log descriptor */
1179 jfs_warn("lmLogOpen: exit(%d)", rc);
1183 static int open_inline_log(struct super_block *sb)
1185 struct jfs_log *log;
1188 if (!(log = kmalloc(sizeof(struct jfs_log), GFP_KERNEL)))
1190 memset(log, 0, sizeof(struct jfs_log));
1191 INIT_LIST_HEAD(&log->sb_list);
1192 init_waitqueue_head(&log->syncwait);
1194 set_bit(log_INLINELOG, &log->flag);
1195 log->bdev = sb->s_bdev;
1196 log->base = addressPXD(&JFS_SBI(sb)->logpxd);
1197 log->size = lengthPXD(&JFS_SBI(sb)->logpxd) >>
1198 (L2LOGPSIZE - sb->s_blocksize_bits);
1199 log->l2bsize = sb->s_blocksize_bits;
1200 ASSERT(L2LOGPSIZE >= sb->s_blocksize_bits);
1205 if ((rc = lmLogInit(log))) {
1207 jfs_warn("lmLogOpen: exit(%d)", rc);
1211 list_add(&JFS_SBI(sb)->log_list, &log->sb_list);
1212 JFS_SBI(sb)->log = log;
1217 static int open_dummy_log(struct super_block *sb)
1223 dummy_log = kmalloc(sizeof(struct jfs_log), GFP_KERNEL);
1228 memset(dummy_log, 0, sizeof(struct jfs_log));
1229 INIT_LIST_HEAD(&dummy_log->sb_list);
1230 init_waitqueue_head(&dummy_log->syncwait);
1231 dummy_log->no_integrity = 1;
1232 /* Make up some stuff */
1233 dummy_log->base = 0;
1234 dummy_log->size = 1024;
1235 rc = lmLogInit(dummy_log);
1244 LOG_LOCK(dummy_log);
1245 list_add(&JFS_SBI(sb)->log_list, &dummy_log->sb_list);
1246 JFS_SBI(sb)->log = dummy_log;
1247 LOG_UNLOCK(dummy_log);
1256 * FUNCTION: log initialization at first log open.
1258 * logredo() (or logformat()) should have been run previously.
1259 * initialize the log from log superblock.
1260 * set the log state in the superblock to LOGMOUNT and
1261 * write SYNCPT log record.
1263 * PARAMETER: log - log structure
1266 * -EINVAL - bad log magic number or superblock dirty
1267 * error returned from logwait()
1269 * serialization: single first open thread
1271 int lmLogInit(struct jfs_log * log)
1275 struct logsuper *logsuper;
1276 struct lbuf *bpsuper;
1281 jfs_info("lmLogInit: log:0x%p", log);
1283 /* initialize the group commit serialization lock */
1284 LOGGC_LOCK_INIT(log);
1286 /* allocate/initialize the log write serialization lock */
1289 LOGSYNC_LOCK_INIT(log);
1291 INIT_LIST_HEAD(&log->synclist);
1293 INIT_LIST_HEAD(&log->cqueue);
1294 log->flush_tblk = NULL;
1299 * initialize log i/o
1301 if ((rc = lbmLogInit(log)))
1304 if (!test_bit(log_INLINELOG, &log->flag))
1305 log->l2bsize = L2LOGPSIZE;
1307 /* check for disabled journaling to disk */
1308 if (log->no_integrity) {
1310 * Journal pages will still be filled. When the time comes
1311 * to actually do the I/O, the write is not done, and the
1312 * endio routine is called directly.
1314 bp = lbmAllocate(log , 0);
1316 bp->l_pn = bp->l_eor = 0;
1319 * validate log superblock
1321 if ((rc = lbmRead(log, 1, &bpsuper)))
1324 logsuper = (struct logsuper *) bpsuper->l_ldata;
1326 if (logsuper->magic != cpu_to_le32(LOGMAGIC)) {
1327 jfs_warn("*** Log Format Error ! ***");
1332 /* logredo() should have been run successfully. */
1333 if (logsuper->state != cpu_to_le32(LOGREDONE)) {
1334 jfs_warn("*** Log Is Dirty ! ***");
1339 /* initialize log from log superblock */
1340 if (test_bit(log_INLINELOG,&log->flag)) {
1341 if (log->size != le32_to_cpu(logsuper->size)) {
1345 jfs_info("lmLogInit: inline log:0x%p base:0x%Lx "
1347 (unsigned long long) log->base, log->size);
1349 if (memcmp(logsuper->uuid, log->uuid, 16)) {
1350 jfs_warn("wrong uuid on JFS log device");
1353 log->size = le32_to_cpu(logsuper->size);
1354 log->l2bsize = le32_to_cpu(logsuper->l2bsize);
1355 jfs_info("lmLogInit: external log:0x%p base:0x%Lx "
1357 (unsigned long long) log->base, log->size);
1360 log->page = le32_to_cpu(logsuper->end) / LOGPSIZE;
1361 log->eor = le32_to_cpu(logsuper->end) - (LOGPSIZE * log->page);
1364 * initialize for log append write mode
1366 /* establish current/end-of-log page/buffer */
1367 if ((rc = lbmRead(log, log->page, &bp)))
1370 lp = (struct logpage *) bp->l_ldata;
1372 jfs_info("lmLogInit: lsn:0x%x page:%d eor:%d:%d",
1373 le32_to_cpu(logsuper->end), log->page, log->eor,
1374 le16_to_cpu(lp->h.eor));
1377 bp->l_pn = log->page;
1378 bp->l_eor = log->eor;
1380 /* if current page is full, move on to next page */
1381 if (log->eor >= LOGPSIZE - LOGPTLRSIZE)
1385 * initialize log syncpoint
1388 * write the first SYNCPT record with syncpoint = 0
1389 * (i.e., log redo up to HERE !);
1390 * remove current page from lbm write queue at end of pageout
1391 * (to write log superblock update), but do not release to
1396 lrd.type = cpu_to_le16(LOG_SYNCPT);
1398 lrd.log.syncpt.sync = 0;
1399 lsn = lmWriteRecord(log, NULL, &lrd, NULL);
1401 bp->l_ceor = bp->l_eor;
1402 lp = (struct logpage *) bp->l_ldata;
1403 lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_eor);
1404 lbmWrite(log, bp, lbmWRITE | lbmSYNC, 0);
1405 if ((rc = lbmIOWait(bp, 0)))
1409 * update/write superblock
1411 logsuper->state = cpu_to_le32(LOGMOUNT);
1412 log->serial = le32_to_cpu(logsuper->serial) + 1;
1413 logsuper->serial = cpu_to_le32(log->serial);
1414 lbmDirectWrite(log, bpsuper, lbmWRITE | lbmRELEASE | lbmSYNC);
1415 if ((rc = lbmIOWait(bpsuper, lbmFREE)))
1419 /* initialize logsync parameters */
1420 log->logsize = (log->size - 2) << L2LOGPSIZE;
1423 log->sync = log->syncpt;
1424 log->nextsync = LOGSYNC_DELTA(log->logsize);
1426 jfs_info("lmLogInit: lsn:0x%x syncpt:0x%x sync:0x%x",
1427 log->lsn, log->syncpt, log->sync);
1430 * initialize for lazy/group commit
1439 errout30: /* release log page */
1441 bp->l_wqnext = NULL;
1444 errout20: /* release log superblock */
1447 errout10: /* unwind lbmLogInit() */
1448 lbmLogShutdown(log);
1450 jfs_warn("lmLogInit: exit(%d)", rc);
1456 * NAME: lmLogClose()
1458 * FUNCTION: remove file system <ipmnt> from active list of log <iplog>
1459 * and close it on last close.
1461 * PARAMETER: sb - superblock
1463 * RETURN: errors from subroutines
1467 int lmLogClose(struct super_block *sb)
1469 struct jfs_sb_info *sbi = JFS_SBI(sb);
1470 struct jfs_log *log = sbi->log;
1471 struct block_device *bdev;
1474 jfs_info("lmLogClose: log:0x%p", log);
1478 list_del(&sbi->log_list);
1483 * We need to make sure all of the "written" metapages
1484 * actually make it to disk
1486 sync_blockdev(sb->s_bdev);
1488 if (test_bit(log_INLINELOG, &log->flag)) {
1490 * in-line log in host file system
1492 rc = lmLogShutdown(log);
1497 if (!log->no_integrity)
1498 lmLogFileSystem(log, sbi, 0);
1500 if (!list_empty(&log->sb_list))
1504 * TODO: ensure that the dummy_log is in a state to allow
1505 * lbmLogShutdown to deallocate all the buffers and call
1506 * kfree against dummy_log. For now, leave dummy_log & its
1507 * buffers in memory, and resuse if another no-integrity mount
1510 if (log->no_integrity)
1514 * external log as separate logical volume
1516 list_del(&log->journal_list);
1518 rc = lmLogShutdown(log);
1527 jfs_info("lmLogClose: exit(%d)", rc);
1533 * NAME: jfs_flush_journal()
1535 * FUNCTION: initiate write of any outstanding transactions to the journal
1536 * and optionally wait until they are all written to disk
1538 * wait == 0 flush until latest txn is committed, don't wait
1539 * wait == 1 flush until latest txn is committed, wait
1540 * wait > 1 flush until all txn's are complete, wait
1542 void jfs_flush_journal(struct jfs_log *log, int wait)
1545 struct tblock *target = NULL;
1546 struct jfs_sb_info *sbi;
1548 /* jfs_write_inode may call us during read-only mount */
1552 jfs_info("jfs_flush_journal: log:0x%p wait=%d", log, wait);
1556 if (!list_empty(&log->cqueue)) {
1558 * This ensures that we will keep writing to the journal as long
1559 * as there are unwritten commit records
1561 target = list_entry(log->cqueue.prev, struct tblock, cqueue);
1563 if (test_bit(log_FLUSH, &log->flag)) {
1565 * We're already flushing.
1566 * if flush_tblk is NULL, we are flushing everything,
1567 * so leave it that way. Otherwise, update it to the
1568 * latest transaction
1570 if (log->flush_tblk)
1571 log->flush_tblk = target;
1573 /* Only flush until latest transaction is committed */
1574 log->flush_tblk = target;
1575 set_bit(log_FLUSH, &log->flag);
1578 * Initiate I/O on outstanding transactions
1580 if (!(log->cflag & logGC_PAGEOUT)) {
1581 log->cflag |= logGC_PAGEOUT;
1586 if ((wait > 1) || test_bit(log_SYNCBARRIER, &log->flag)) {
1587 /* Flush until all activity complete */
1588 set_bit(log_FLUSH, &log->flag);
1589 log->flush_tblk = NULL;
1592 if (wait && target && !(target->flag & tblkGC_COMMITTED)) {
1593 DECLARE_WAITQUEUE(__wait, current);
1595 add_wait_queue(&target->gcwait, &__wait);
1596 set_current_state(TASK_UNINTERRUPTIBLE);
1599 current->state = TASK_RUNNING;
1601 remove_wait_queue(&target->gcwait, &__wait);
1608 list_for_each_entry(sbi, &log->sb_list, log_list) {
1609 filemap_fdatawrite(sbi->ipbmap->i_mapping);
1610 filemap_fdatawrite(sbi->ipimap->i_mapping);
1611 filemap_fdatawrite(sbi->direct_inode->i_mapping);
1615 * If there was recent activity, we may need to wait
1616 * for the lazycommit thread to catch up
1618 if ((!list_empty(&log->cqueue)) || !list_empty(&log->synclist)) {
1619 for (i = 0; i < 200; i++) { /* Too much? */
1621 if (list_empty(&log->cqueue) &&
1622 list_empty(&log->synclist))
1626 assert(list_empty(&log->cqueue));
1627 if (!list_empty(&log->synclist)) {
1628 struct logsyncblk *lp;
1630 list_for_each_entry(lp, &log->synclist, synclist) {
1631 if (lp->xflag & COMMIT_PAGE) {
1632 struct metapage *mp = (struct metapage *)lp;
1633 dump_mem("orphan metapage", lp,
1634 sizeof(struct metapage));
1635 dump_mem("page", mp->page, sizeof(struct page));
1638 dump_mem("orphan tblock", lp,
1639 sizeof(struct tblock));
1641 // current->state = TASK_INTERRUPTIBLE;
1644 //assert(list_empty(&log->synclist));
1645 clear_bit(log_FLUSH, &log->flag);
1649 * NAME: lmLogShutdown()
1651 * FUNCTION: log shutdown at last LogClose().
1653 * write log syncpt record.
1654 * update super block to set redone flag to 0.
1656 * PARAMETER: log - log inode
1658 * RETURN: 0 - success
1660 * serialization: single last close thread
1662 int lmLogShutdown(struct jfs_log * log)
1667 struct logsuper *logsuper;
1668 struct lbuf *bpsuper;
1672 jfs_info("lmLogShutdown: log:0x%p", log);
1674 jfs_flush_journal(log, 2);
1677 * write the last SYNCPT record with syncpoint = 0
1678 * (i.e., log redo up to HERE !)
1682 lrd.type = cpu_to_le16(LOG_SYNCPT);
1684 lrd.log.syncpt.sync = 0;
1686 lsn = lmWriteRecord(log, NULL, &lrd, NULL);
1688 lp = (struct logpage *) bp->l_ldata;
1689 lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_eor);
1690 lbmWrite(log, log->bp, lbmWRITE | lbmRELEASE | lbmSYNC, 0);
1691 lbmIOWait(log->bp, lbmFREE);
1695 * synchronous update log superblock
1696 * mark log state as shutdown cleanly
1697 * (i.e., Log does not need to be replayed).
1699 if ((rc = lbmRead(log, 1, &bpsuper)))
1702 logsuper = (struct logsuper *) bpsuper->l_ldata;
1703 logsuper->state = cpu_to_le32(LOGREDONE);
1704 logsuper->end = cpu_to_le32(lsn);
1705 lbmDirectWrite(log, bpsuper, lbmWRITE | lbmRELEASE | lbmSYNC);
1706 rc = lbmIOWait(bpsuper, lbmFREE);
1708 jfs_info("lmLogShutdown: lsn:0x%x page:%d eor:%d",
1709 lsn, log->page, log->eor);
1713 * shutdown per log i/o
1715 lbmLogShutdown(log);
1718 jfs_warn("lmLogShutdown: exit(%d)", rc);
1725 * NAME: lmLogFileSystem()
1727 * FUNCTION: insert (<activate> = true)/remove (<activate> = false)
1728 * file system into/from log active file system list.
1730 * PARAMETE: log - pointer to logs inode.
1731 * fsdev - kdev_t of filesystem.
1732 * serial - pointer to returned log serial number
1733 * activate - insert/remove device from active list.
1735 * RETURN: 0 - success
1736 * errors returned by vms_iowait().
1738 static int lmLogFileSystem(struct jfs_log * log, struct jfs_sb_info *sbi,
1743 struct logsuper *logsuper;
1744 struct lbuf *bpsuper;
1745 char *uuid = sbi->uuid;
1748 * insert/remove file system device to log active file system list.
1750 if ((rc = lbmRead(log, 1, &bpsuper)))
1753 logsuper = (struct logsuper *) bpsuper->l_ldata;
1755 for (i = 0; i < MAX_ACTIVE; i++)
1756 if (!memcmp(logsuper->active[i].uuid, NULL_UUID, 16)) {
1757 memcpy(logsuper->active[i].uuid, uuid, 16);
1761 if (i == MAX_ACTIVE) {
1762 jfs_warn("Too many file systems sharing journal!");
1764 return -EMFILE; /* Is there a better rc? */
1767 for (i = 0; i < MAX_ACTIVE; i++)
1768 if (!memcmp(logsuper->active[i].uuid, uuid, 16)) {
1769 memcpy(logsuper->active[i].uuid, NULL_UUID, 16);
1772 if (i == MAX_ACTIVE) {
1773 jfs_warn("Somebody stomped on the journal!");
1781 * synchronous write log superblock:
1783 * write sidestream bypassing write queue:
1784 * at file system mount, log super block is updated for
1785 * activation of the file system before any log record
1786 * (MOUNT record) of the file system, and at file system
1787 * unmount, all meta data for the file system has been
1788 * flushed before log super block is updated for deactivation
1789 * of the file system.
1791 lbmDirectWrite(log, bpsuper, lbmWRITE | lbmRELEASE | lbmSYNC);
1792 rc = lbmIOWait(bpsuper, lbmFREE);
1798 * log buffer manager (lbm)
1799 * ------------------------
1801 * special purpose buffer manager supporting log i/o requirements.
1803 * per log write queue:
1804 * log pageout occurs in serial order by fifo write queue and
1805 * restricting to a single i/o in pregress at any one time.
1806 * a circular singly-linked list
1807 * (log->wrqueue points to the tail, and buffers are linked via
1808 * bp->wrqueue field), and
1809 * maintains log page in pageout ot waiting for pageout in serial pageout.
1815 * initialize per log I/O setup at lmLogInit()
1817 static int lbmLogInit(struct jfs_log * log)
1822 jfs_info("lbmLogInit: log:0x%p", log);
1824 /* initialize current buffer cursor */
1827 /* initialize log device write queue */
1831 * Each log has its own buffer pages allocated to it. These are
1832 * not managed by the page cache. This ensures that a transaction
1833 * writing to the log does not block trying to allocate a page from
1834 * the page cache (for the log). This would be bad, since page
1835 * allocation waits on the kswapd thread that may be committing inodes
1836 * which would cause log activity. Was that clear? I'm trying to
1837 * avoid deadlock here.
1839 init_waitqueue_head(&log->free_wait);
1841 log->lbuf_free = NULL;
1843 for (i = 0; i < LOGPAGES;) {
1848 buffer = (char *) get_zeroed_page(GFP_KERNEL);
1851 page = virt_to_page(buffer);
1852 for (offset = 0; offset < PAGE_SIZE; offset += LOGPSIZE) {
1853 lbuf = kmalloc(sizeof(struct lbuf), GFP_KERNEL);
1856 free_page((unsigned long) buffer);
1859 if (offset) /* we already have one reference */
1861 lbuf->l_offset = offset;
1862 lbuf->l_ldata = buffer + offset;
1863 lbuf->l_page = page;
1865 init_waitqueue_head(&lbuf->l_ioevent);
1867 lbuf->l_freelist = log->lbuf_free;
1868 log->lbuf_free = lbuf;
1876 lbmLogShutdown(log);
1884 * finalize per log I/O setup at lmLogShutdown()
1886 static void lbmLogShutdown(struct jfs_log * log)
1890 jfs_info("lbmLogShutdown: log:0x%p", log);
1892 lbuf = log->lbuf_free;
1894 struct lbuf *next = lbuf->l_freelist;
1895 __free_page(lbuf->l_page);
1905 * allocate an empty log buffer
1907 static struct lbuf *lbmAllocate(struct jfs_log * log, int pn)
1910 unsigned long flags;
1913 * recycle from log buffer freelist if any
1916 LCACHE_SLEEP_COND(log->free_wait, (bp = log->lbuf_free), flags);
1917 log->lbuf_free = bp->l_freelist;
1918 LCACHE_UNLOCK(flags);
1922 bp->l_wqnext = NULL;
1923 bp->l_freelist = NULL;
1926 bp->l_blkno = log->base + (pn << (L2LOGPSIZE - log->l2bsize));
1936 * release a log buffer to freelist
1938 static void lbmFree(struct lbuf * bp)
1940 unsigned long flags;
1946 LCACHE_UNLOCK(flags);
1949 static void lbmfree(struct lbuf * bp)
1951 struct jfs_log *log = bp->l_log;
1953 assert(bp->l_wqnext == NULL);
1956 * return the buffer to head of freelist
1958 bp->l_freelist = log->lbuf_free;
1959 log->lbuf_free = bp;
1961 wake_up(&log->free_wait);
1969 * FUNCTION: add a log buffer to the the log redrive list
1975 * Takes log_redrive_lock.
1977 static inline void lbmRedrive(struct lbuf *bp)
1979 unsigned long flags;
1981 spin_lock_irqsave(&log_redrive_lock, flags);
1982 bp->l_redrive_next = log_redrive_list;
1983 log_redrive_list = bp;
1984 spin_unlock_irqrestore(&log_redrive_lock, flags);
1986 wake_up(&jfs_IO_thread_wait);
1993 static int lbmRead(struct jfs_log * log, int pn, struct lbuf ** bpp)
1999 * allocate a log buffer
2001 *bpp = bp = lbmAllocate(log, pn);
2002 jfs_info("lbmRead: bp:0x%p pn:0x%x", bp, pn);
2004 bp->l_flag |= lbmREAD;
2006 bio = bio_alloc(GFP_NOFS, 1);
2008 bio->bi_sector = bp->l_blkno << (log->l2bsize - 9);
2009 bio->bi_bdev = log->bdev;
2010 bio->bi_io_vec[0].bv_page = bp->l_page;
2011 bio->bi_io_vec[0].bv_len = LOGPSIZE;
2012 bio->bi_io_vec[0].bv_offset = bp->l_offset;
2016 bio->bi_size = LOGPSIZE;
2018 bio->bi_end_io = lbmIODone;
2019 bio->bi_private = bp;
2020 submit_bio(READ_SYNC, bio);
2022 wait_event(bp->l_ioevent, (bp->l_flag != lbmREAD));
2031 * buffer at head of pageout queue stays after completion of
2032 * partial-page pageout and redriven by explicit initiation of
2033 * pageout by caller until full-page pageout is completed and
2036 * device driver i/o done redrives pageout of new buffer at
2037 * head of pageout queue when current buffer at head of pageout
2038 * queue is released at the completion of its full-page pageout.
2040 * LOGGC_LOCK() serializes lbmWrite() by lmNextPage() and lmGroupCommit().
2041 * LCACHE_LOCK() serializes xflag between lbmWrite() and lbmIODone()
2043 static void lbmWrite(struct jfs_log * log, struct lbuf * bp, int flag,
2047 unsigned long flags;
2049 jfs_info("lbmWrite: bp:0x%p flag:0x%x pn:0x%x", bp, flag, bp->l_pn);
2051 /* map the logical block address to physical block address */
2053 log->base + (bp->l_pn << (L2LOGPSIZE - log->l2bsize));
2055 LCACHE_LOCK(flags); /* disable+lock */
2058 * initialize buffer for device driver
2063 * insert bp at tail of write queue associated with log
2065 * (request is either for bp already/currently at head of queue
2066 * or new bp to be inserted at tail)
2070 /* is buffer not already on write queue ? */
2071 if (bp->l_wqnext == NULL) {
2072 /* insert at tail of wqueue */
2078 bp->l_wqnext = tail->l_wqnext;
2079 tail->l_wqnext = bp;
2085 /* is buffer at head of wqueue and for write ? */
2086 if ((bp != tail->l_wqnext) || !(flag & lbmWRITE)) {
2087 LCACHE_UNLOCK(flags); /* unlock+enable */
2091 LCACHE_UNLOCK(flags); /* unlock+enable */
2095 else if (flag & lbmSYNC)
2108 * initiate pageout bypassing write queue for sidestream
2109 * (e.g., log superblock) write;
2111 static void lbmDirectWrite(struct jfs_log * log, struct lbuf * bp, int flag)
2113 jfs_info("lbmDirectWrite: bp:0x%p flag:0x%x pn:0x%x",
2114 bp, flag, bp->l_pn);
2117 * initialize buffer for device driver
2119 bp->l_flag = flag | lbmDIRECT;
2121 /* map the logical block address to physical block address */
2123 log->base + (bp->l_pn << (L2LOGPSIZE - log->l2bsize));
2126 * initiate pageout of the page
2133 * NAME: lbmStartIO()
2135 * FUNCTION: Interface to DD strategy routine
2139 * serialization: LCACHE_LOCK() is NOT held during log i/o;
2141 static void lbmStartIO(struct lbuf * bp)
2144 struct jfs_log *log = bp->l_log;
2146 jfs_info("lbmStartIO\n");
2148 bio = bio_alloc(GFP_NOFS, 1);
2149 bio->bi_sector = bp->l_blkno << (log->l2bsize - 9);
2150 bio->bi_bdev = log->bdev;
2151 bio->bi_io_vec[0].bv_page = bp->l_page;
2152 bio->bi_io_vec[0].bv_len = LOGPSIZE;
2153 bio->bi_io_vec[0].bv_offset = bp->l_offset;
2157 bio->bi_size = LOGPSIZE;
2159 bio->bi_end_io = lbmIODone;
2160 bio->bi_private = bp;
2162 /* check if journaling to disk has been disabled */
2163 if (log->no_integrity) {
2165 lbmIODone(bio, 0, 0);
2167 submit_bio(WRITE_SYNC, bio);
2168 INCREMENT(lmStat.submitted);
2176 static int lbmIOWait(struct lbuf * bp, int flag)
2178 unsigned long flags;
2181 jfs_info("lbmIOWait1: bp:0x%p flag:0x%x:0x%x", bp, bp->l_flag, flag);
2183 LCACHE_LOCK(flags); /* disable+lock */
2185 LCACHE_SLEEP_COND(bp->l_ioevent, (bp->l_flag & lbmDONE), flags);
2187 rc = (bp->l_flag & lbmERROR) ? -EIO : 0;
2192 LCACHE_UNLOCK(flags); /* unlock+enable */
2194 jfs_info("lbmIOWait2: bp:0x%p flag:0x%x:0x%x", bp, bp->l_flag, flag);
2201 * executed at INTIODONE level
2203 static int lbmIODone(struct bio *bio, unsigned int bytes_done, int error)
2205 struct lbuf *bp = bio->bi_private;
2206 struct lbuf *nextbp, *tail;
2207 struct jfs_log *log;
2208 unsigned long flags;
2214 * get back jfs buffer bound to the i/o buffer
2216 jfs_info("lbmIODone: bp:0x%p flag:0x%x", bp, bp->l_flag);
2218 LCACHE_LOCK(flags); /* disable+lock */
2220 bp->l_flag |= lbmDONE;
2222 if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2223 bp->l_flag |= lbmERROR;
2225 jfs_err("lbmIODone: I/O error in JFS log");
2233 if (bp->l_flag & lbmREAD) {
2234 bp->l_flag &= ~lbmREAD;
2236 LCACHE_UNLOCK(flags); /* unlock+enable */
2238 /* wakeup I/O initiator */
2239 LCACHE_WAKEUP(&bp->l_ioevent);
2245 * pageout completion
2247 * the bp at the head of write queue has completed pageout.
2249 * if single-commit/full-page pageout, remove the current buffer
2250 * from head of pageout queue, and redrive pageout with
2251 * the new buffer at head of pageout queue;
2252 * otherwise, the partial-page pageout buffer stays at
2253 * the head of pageout queue to be redriven for pageout
2254 * by lmGroupCommit() until full-page pageout is completed.
2256 bp->l_flag &= ~lbmWRITE;
2257 INCREMENT(lmStat.pagedone);
2259 /* update committed lsn */
2261 log->clsn = (bp->l_pn << L2LOGPSIZE) + bp->l_ceor;
2263 if (bp->l_flag & lbmDIRECT) {
2264 LCACHE_WAKEUP(&bp->l_ioevent);
2265 LCACHE_UNLOCK(flags);
2271 /* single element queue */
2273 /* remove head buffer of full-page pageout
2274 * from log device write queue
2276 if (bp->l_flag & lbmRELEASE) {
2278 bp->l_wqnext = NULL;
2281 /* multi element queue */
2283 /* remove head buffer of full-page pageout
2284 * from log device write queue
2286 if (bp->l_flag & lbmRELEASE) {
2287 nextbp = tail->l_wqnext = bp->l_wqnext;
2288 bp->l_wqnext = NULL;
2291 * redrive pageout of next page at head of write queue:
2292 * redrive next page without any bound tblk
2293 * (i.e., page w/o any COMMIT records), or
2294 * first page of new group commit which has been
2295 * queued after current page (subsequent pageout
2296 * is performed synchronously, except page without
2297 * any COMMITs) by lmGroupCommit() as indicated
2300 if (nextbp->l_flag & lbmWRITE) {
2302 * We can't do the I/O at interrupt time.
2303 * The jfsIO thread can do it
2311 * synchronous pageout:
2313 * buffer has not necessarily been removed from write queue
2314 * (e.g., synchronous write of partial-page with COMMIT):
2315 * leave buffer for i/o initiator to dispose
2317 if (bp->l_flag & lbmSYNC) {
2318 LCACHE_UNLOCK(flags); /* unlock+enable */
2320 /* wakeup I/O initiator */
2321 LCACHE_WAKEUP(&bp->l_ioevent);
2325 * Group Commit pageout:
2327 else if (bp->l_flag & lbmGC) {
2328 LCACHE_UNLOCK(flags);
2333 * asynchronous pageout:
2335 * buffer must have been removed from write queue:
2336 * insert buffer at head of freelist where it can be recycled
2339 assert(bp->l_flag & lbmRELEASE);
2340 assert(bp->l_flag & lbmFREE);
2343 LCACHE_UNLOCK(flags); /* unlock+enable */
2349 int jfsIOWait(void *arg)
2355 complete(&jfsIOwait);
2358 DECLARE_WAITQUEUE(wq, current);
2360 spin_lock_irq(&log_redrive_lock);
2361 while ((bp = log_redrive_list) != 0) {
2362 log_redrive_list = bp->l_redrive_next;
2363 bp->l_redrive_next = NULL;
2364 spin_unlock_irq(&log_redrive_lock);
2366 spin_lock_irq(&log_redrive_lock);
2368 if (current->flags & PF_FREEZE) {
2369 spin_unlock_irq(&log_redrive_lock);
2370 refrigerator(PF_FREEZE);
2372 add_wait_queue(&jfs_IO_thread_wait, &wq);
2373 set_current_state(TASK_INTERRUPTIBLE);
2374 spin_unlock_irq(&log_redrive_lock);
2376 current->state = TASK_RUNNING;
2377 remove_wait_queue(&jfs_IO_thread_wait, &wq);
2379 } while (!jfs_stop_threads);
2381 jfs_info("jfsIOWait being killed!");
2382 complete_and_exit(&jfsIOwait, 0);
2386 * NAME: lmLogFormat()/jfs_logform()
2388 * FUNCTION: format file system log
2392 * logAddress - start address of log space in FS block
2393 * logSize - length of log space in FS block;
2395 * RETURN: 0 - success
2398 * XXX: We're synchronously writing one page at a time. This needs to
2399 * be improved by writing multiple pages at once.
2401 int lmLogFormat(struct jfs_log *log, s64 logAddress, int logSize)
2404 struct jfs_sb_info *sbi;
2405 struct logsuper *logsuper;
2407 int lspn; /* log sequence page number */
2408 struct lrd *lrd_ptr;
2412 jfs_info("lmLogFormat: logAddress:%Ld logSize:%d",
2413 (long long)logAddress, logSize);
2415 sbi = list_entry(log->sb_list.next, struct jfs_sb_info, log_list);
2417 /* allocate a log buffer */
2418 bp = lbmAllocate(log, 1);
2420 npages = logSize >> sbi->l2nbperpage;
2425 * page 0 - reserved;
2426 * page 1 - log superblock;
2427 * page 2 - log data page: A SYNC log record is written
2428 * into this page at logform time;
2429 * pages 3-N - log data page: set to empty log data pages;
2432 * init log superblock: log page 1
2434 logsuper = (struct logsuper *) bp->l_ldata;
2436 logsuper->magic = cpu_to_le32(LOGMAGIC);
2437 logsuper->version = cpu_to_le32(LOGVERSION);
2438 logsuper->state = cpu_to_le32(LOGREDONE);
2439 logsuper->flag = cpu_to_le32(sbi->mntflag); /* ? */
2440 logsuper->size = cpu_to_le32(npages);
2441 logsuper->bsize = cpu_to_le32(sbi->bsize);
2442 logsuper->l2bsize = cpu_to_le32(sbi->l2bsize);
2443 logsuper->end = cpu_to_le32(2 * LOGPSIZE + LOGPHDRSIZE + LOGRDSIZE);
2445 bp->l_flag = lbmWRITE | lbmSYNC | lbmDIRECT;
2446 bp->l_blkno = logAddress + sbi->nbperpage;
2448 if ((rc = lbmIOWait(bp, 0)))
2452 * init pages 2 to npages-1 as log data pages:
2454 * log page sequence number (lpsn) initialization:
2457 * +-----+-----+=====+=====+===.....===+=====+
2459 * <--- N page circular file ---->
2461 * the N (= npages-2) data pages of the log is maintained as
2462 * a circular file for the log records;
2463 * lpsn grows by 1 monotonically as each log page is written
2464 * to the circular file of the log;
2465 * and setLogpage() will not reset the page number even if
2466 * the eor is equal to LOGPHDRSIZE. In order for binary search
2467 * still work in find log end process, we have to simulate the
2468 * log wrap situation at the log format time.
2469 * The 1st log page written will have the highest lpsn. Then
2470 * the succeeding log pages will have ascending order of
2471 * the lspn starting from 0, ... (N-2)
2473 lp = (struct logpage *) bp->l_ldata;
2475 * initialize 1st log page to be written: lpsn = N - 1,
2476 * write a SYNCPT log record is written to this page
2478 lp->h.page = lp->t.page = cpu_to_le32(npages - 3);
2479 lp->h.eor = lp->t.eor = cpu_to_le16(LOGPHDRSIZE + LOGRDSIZE);
2481 lrd_ptr = (struct lrd *) &lp->data;
2482 lrd_ptr->logtid = 0;
2483 lrd_ptr->backchain = 0;
2484 lrd_ptr->type = cpu_to_le16(LOG_SYNCPT);
2485 lrd_ptr->length = 0;
2486 lrd_ptr->log.syncpt.sync = 0;
2488 bp->l_blkno += sbi->nbperpage;
2489 bp->l_flag = lbmWRITE | lbmSYNC | lbmDIRECT;
2491 if ((rc = lbmIOWait(bp, 0)))
2495 * initialize succeeding log pages: lpsn = 0, 1, ..., (N-2)
2497 for (lspn = 0; lspn < npages - 3; lspn++) {
2498 lp->h.page = lp->t.page = cpu_to_le32(lspn);
2499 lp->h.eor = lp->t.eor = cpu_to_le16(LOGPHDRSIZE);
2501 bp->l_blkno += sbi->nbperpage;
2502 bp->l_flag = lbmWRITE | lbmSYNC | lbmDIRECT;
2504 if ((rc = lbmIOWait(bp, 0)))
2513 /* release the buffer */
2519 #ifdef CONFIG_JFS_STATISTICS
2520 int jfs_lmstats_read(char *buffer, char **start, off_t offset, int length,
2521 int *eof, void *data)
2526 len += sprintf(buffer,
2527 "JFS Logmgr stats\n"
2528 "================\n"
2530 "writes submitted = %d\n"
2531 "writes completed = %d\n"
2532 "full pages submitted = %d\n"
2533 "partial pages submitted = %d\n",
2538 lmStat.partial_page);
2541 *start = buffer + begin;
2554 #endif /* CONFIG_JFS_STATISTICS */