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
;
238 jfs_info("lmLog: log:0x%p tblk:0x%p, lrd:0x%p tlck:0x%p",
239 log
, tblk
, lrd
, tlck
);
243 /* log by (out-of-transaction) JFS ? */
247 /* log from page ? */
249 tlck
->type
& tlckBTROOT
|| (mp
= tlck
->mp
) == NULL
)
253 * initialize/update page/transaction recovery lsn
260 * initialize page lsn if first log write of the page
267 /* insert page at tail of logsynclist */
268 list_add_tail(&mp
->synclist
, &log
->synclist
);
272 * initialize/update lsn of tblock of the page
274 * transaction inherits oldest lsn of pages associated
275 * with allocation/deallocation of resources (their
276 * log records are used to reconstruct allocation map
277 * at recovery time: inode for inode allocation map,
278 * B+-tree index of extent descriptors for block
280 * allocation map pages inherit transaction lsn at
281 * commit time to allow forwarding log syncpt past log
282 * records associated with allocation/deallocation of
283 * resources only after persistent map of these map pages
284 * have been updated and propagated to home.
287 * initialize transaction lsn:
289 if (tblk
->lsn
== 0) {
290 /* inherit lsn of its first page logged */
294 /* insert tblock after the page on logsynclist */
295 list_add(&tblk
->synclist
, &mp
->synclist
);
298 * update transaction lsn:
301 /* inherit oldest/smallest lsn of page */
302 logdiff(diffp
, mp
->lsn
, log
);
303 logdiff(difft
, tblk
->lsn
, log
);
305 /* update tblock lsn with page lsn */
308 /* move tblock after page on logsynclist */
309 list_move(&tblk
->synclist
, &mp
->synclist
);
316 * write the log record
319 lsn
= lmWriteRecord(log
, tblk
, lrd
, tlck
);
322 * forward log syncpt if log reached next syncpt trigger
324 logdiff(diffp
, lsn
, log
);
325 if (diffp
>= log
->nextsync
)
326 lsn
= lmLogSync(log
, 0);
328 /* update end-of-log lsn */
333 /* 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 * this is called only from lmLog().
932 * PARAMETER: ip - pointer to logs inode.
936 * serialization: LOG_LOCK() held on entry/exit
938 static int lmLogSync(struct jfs_log
* log
, int nosyncwait
)
941 int written
; /* written since last syncpt */
942 int free
; /* free space left available */
943 int delta
; /* additional delta to write normally */
944 int more
; /* additional write granted */
947 struct logsyncblk
*lp
;
952 /* if last sync is same as last syncpt,
953 * invoke sync point forward processing to update sync.
956 if (log
->sync
== log
->syncpt
) {
958 /* ToDo: push dirty metapages out to disk */
961 if (list_empty(&log
->synclist
))
962 log
->sync
= log
->lsn
;
964 lp
= list_entry(log
->synclist
.next
,
965 struct logsyncblk
, synclist
);
972 /* if sync is different from last syncpt,
973 * write a SYNCPT record with syncpt = sync.
974 * reset syncpt = sync
976 if (log
->sync
!= log
->syncpt
) {
977 struct jfs_sb_info
*sbi
;
980 * We need to make sure all of the "written" metapages
981 * actually make it to disk
983 list_for_each_entry(sbi
, &log
->sb_list
, log_list
) {
984 if (sbi
->flag
& JFS_NOINTEGRITY
)
986 filemap_fdatawrite(sbi
->ipbmap
->i_mapping
);
987 filemap_fdatawrite(sbi
->ipimap
->i_mapping
);
988 filemap_fdatawrite(sbi
->sb
->s_bdev
->bd_inode
->i_mapping
);
990 list_for_each_entry(sbi
, &log
->sb_list
, log_list
) {
991 if (sbi
->flag
& JFS_NOINTEGRITY
)
993 filemap_fdatawait(sbi
->ipbmap
->i_mapping
);
994 filemap_fdatawait(sbi
->ipimap
->i_mapping
);
995 filemap_fdatawait(sbi
->sb
->s_bdev
->bd_inode
->i_mapping
);
1000 lrd
.type
= cpu_to_le16(LOG_SYNCPT
);
1002 lrd
.log
.syncpt
.sync
= cpu_to_le32(log
->sync
);
1003 lsn
= lmWriteRecord(log
, NULL
, &lrd
, NULL
);
1005 log
->syncpt
= log
->sync
;
1010 * setup next syncpt trigger (SWAG)
1012 logsize
= log
->logsize
;
1014 logdiff(written
, lsn
, log
);
1015 free
= logsize
- written
;
1016 delta
= LOGSYNC_DELTA(logsize
);
1017 more
= min(free
/ 2, delta
);
1018 if (more
< 2 * LOGPSIZE
) {
1019 jfs_warn("\n ... Log Wrap ... Log Wrap ... Log Wrap ...\n");
1023 * option 1 - panic ? No.!
1024 * option 2 - shutdown file systems
1025 * associated with log ?
1026 * option 3 - extend log ?
1029 * option 4 - second chance
1031 * mark log wrapped, and continue.
1032 * when all active transactions are completed,
1033 * mark log vaild for recovery.
1034 * if crashed during invalid state, log state
1035 * implies invald log, forcing fsck().
1037 /* mark log state log wrap in log superblock */
1038 /* log->state = LOGWRAP; */
1040 /* reset sync point computation */
1041 log
->syncpt
= log
->sync
= lsn
;
1042 log
->nextsync
= delta
;
1044 /* next syncpt trigger = written + more */
1045 log
->nextsync
= written
+ more
;
1047 /* return if lmLogSync() from outside of transaction, e.g., sync() */
1051 /* if number of bytes written from last sync point is more
1052 * than 1/4 of the log size, stop new transactions from
1053 * starting until all current transactions are completed
1054 * by setting syncbarrier flag.
1056 if (written
> LOGSYNC_BARRIER(logsize
) && logsize
> 32 * LOGPSIZE
) {
1057 set_bit(log_SYNCBARRIER
, &log
->flag
);
1058 jfs_info("log barrier on: lsn=0x%x syncpt=0x%x", lsn
,
1061 * We may have to initiate group commit
1063 jfs_flush_journal(log
, 0);
1073 * FUNCTION: open the log on first open;
1074 * insert filesystem in the active list of the log.
1076 * PARAMETER: ipmnt - file system mount inode
1077 * iplog - log inode (out)
1083 int lmLogOpen(struct super_block
*sb
)
1086 struct block_device
*bdev
;
1087 struct jfs_log
*log
;
1088 struct jfs_sb_info
*sbi
= JFS_SBI(sb
);
1090 if (sbi
->flag
& JFS_NOINTEGRITY
)
1091 return open_dummy_log(sb
);
1093 if (sbi
->mntflag
& JFS_INLINELOG
)
1094 return open_inline_log(sb
);
1097 list_for_each_entry(log
, &jfs_external_logs
, journal_list
) {
1098 if (log
->bdev
->bd_dev
== sbi
->logdev
) {
1099 if (memcmp(log
->uuid
, sbi
->loguuid
,
1100 sizeof(log
->uuid
))) {
1101 jfs_warn("wrong uuid on JFS journal\n");
1106 * add file system to log active file system list
1108 if ((rc
= lmLogFileSystem(log
, sbi
, 1))) {
1116 if (!(log
= kmalloc(sizeof(struct jfs_log
), GFP_KERNEL
))) {
1120 memset(log
, 0, sizeof(struct jfs_log
));
1121 INIT_LIST_HEAD(&log
->sb_list
);
1122 init_waitqueue_head(&log
->syncwait
);
1125 * external log as separate logical volume
1127 * file systems to log may have n-to-1 relationship;
1130 bdev
= open_by_devnum(sbi
->logdev
, FMODE_READ
|FMODE_WRITE
);
1132 rc
= -PTR_ERR(bdev
);
1136 if ((rc
= bd_claim(bdev
, log
))) {
1141 memcpy(log
->uuid
, sbi
->loguuid
, sizeof(log
->uuid
));
1146 if ((rc
= lmLogInit(log
)))
1149 list_add(&log
->journal_list
, &jfs_external_logs
);
1152 * add file system to log active file system list
1154 if ((rc
= lmLogFileSystem(log
, sbi
, 1)))
1159 list_add(&sbi
->log_list
, &log
->sb_list
);
1169 shutdown
: /* unwind lbmLogInit() */
1170 list_del(&log
->journal_list
);
1171 lbmLogShutdown(log
);
1176 close
: /* close external log device */
1179 free
: /* free log descriptor */
1183 jfs_warn("lmLogOpen: exit(%d)", rc
);
1187 static int open_inline_log(struct super_block
*sb
)
1189 struct jfs_log
*log
;
1192 if (!(log
= kmalloc(sizeof(struct jfs_log
), GFP_KERNEL
)))
1194 memset(log
, 0, sizeof(struct jfs_log
));
1195 INIT_LIST_HEAD(&log
->sb_list
);
1196 init_waitqueue_head(&log
->syncwait
);
1198 set_bit(log_INLINELOG
, &log
->flag
);
1199 log
->bdev
= sb
->s_bdev
;
1200 log
->base
= addressPXD(&JFS_SBI(sb
)->logpxd
);
1201 log
->size
= lengthPXD(&JFS_SBI(sb
)->logpxd
) >>
1202 (L2LOGPSIZE
- sb
->s_blocksize_bits
);
1203 log
->l2bsize
= sb
->s_blocksize_bits
;
1204 ASSERT(L2LOGPSIZE
>= sb
->s_blocksize_bits
);
1209 if ((rc
= lmLogInit(log
))) {
1211 jfs_warn("lmLogOpen: exit(%d)", rc
);
1215 list_add(&JFS_SBI(sb
)->log_list
, &log
->sb_list
);
1216 JFS_SBI(sb
)->log
= log
;
1221 static int open_dummy_log(struct super_block
*sb
)
1227 dummy_log
= kmalloc(sizeof(struct jfs_log
), GFP_KERNEL
);
1232 memset(dummy_log
, 0, sizeof(struct jfs_log
));
1233 INIT_LIST_HEAD(&dummy_log
->sb_list
);
1234 init_waitqueue_head(&dummy_log
->syncwait
);
1235 dummy_log
->no_integrity
= 1;
1236 /* Make up some stuff */
1237 dummy_log
->base
= 0;
1238 dummy_log
->size
= 1024;
1239 rc
= lmLogInit(dummy_log
);
1248 LOG_LOCK(dummy_log
);
1249 list_add(&JFS_SBI(sb
)->log_list
, &dummy_log
->sb_list
);
1250 JFS_SBI(sb
)->log
= dummy_log
;
1251 LOG_UNLOCK(dummy_log
);
1260 * FUNCTION: log initialization at first log open.
1262 * logredo() (or logformat()) should have been run previously.
1263 * initialize the log from log superblock.
1264 * set the log state in the superblock to LOGMOUNT and
1265 * write SYNCPT log record.
1267 * PARAMETER: log - log structure
1270 * -EINVAL - bad log magic number or superblock dirty
1271 * error returned from logwait()
1273 * serialization: single first open thread
1275 int lmLogInit(struct jfs_log
* log
)
1279 struct logsuper
*logsuper
;
1280 struct lbuf
*bpsuper
;
1285 jfs_info("lmLogInit: log:0x%p", log
);
1287 /* initialize the group commit serialization lock */
1288 LOGGC_LOCK_INIT(log
);
1290 /* allocate/initialize the log write serialization lock */
1293 LOGSYNC_LOCK_INIT(log
);
1295 INIT_LIST_HEAD(&log
->synclist
);
1297 INIT_LIST_HEAD(&log
->cqueue
);
1298 log
->flush_tblk
= NULL
;
1303 * initialize log i/o
1305 if ((rc
= lbmLogInit(log
)))
1308 if (!test_bit(log_INLINELOG
, &log
->flag
))
1309 log
->l2bsize
= L2LOGPSIZE
;
1311 /* check for disabled journaling to disk */
1312 if (log
->no_integrity
) {
1314 * Journal pages will still be filled. When the time comes
1315 * to actually do the I/O, the write is not done, and the
1316 * endio routine is called directly.
1318 bp
= lbmAllocate(log
, 0);
1320 bp
->l_pn
= bp
->l_eor
= 0;
1323 * validate log superblock
1325 if ((rc
= lbmRead(log
, 1, &bpsuper
)))
1328 logsuper
= (struct logsuper
*) bpsuper
->l_ldata
;
1330 if (logsuper
->magic
!= cpu_to_le32(LOGMAGIC
)) {
1331 jfs_warn("*** Log Format Error ! ***");
1336 /* logredo() should have been run successfully. */
1337 if (logsuper
->state
!= cpu_to_le32(LOGREDONE
)) {
1338 jfs_warn("*** Log Is Dirty ! ***");
1343 /* initialize log from log superblock */
1344 if (test_bit(log_INLINELOG
,&log
->flag
)) {
1345 if (log
->size
!= le32_to_cpu(logsuper
->size
)) {
1349 jfs_info("lmLogInit: inline log:0x%p base:0x%Lx "
1351 (unsigned long long) log
->base
, log
->size
);
1353 if (memcmp(logsuper
->uuid
, log
->uuid
, 16)) {
1354 jfs_warn("wrong uuid on JFS log device");
1357 log
->size
= le32_to_cpu(logsuper
->size
);
1358 log
->l2bsize
= le32_to_cpu(logsuper
->l2bsize
);
1359 jfs_info("lmLogInit: external log:0x%p base:0x%Lx "
1361 (unsigned long long) log
->base
, log
->size
);
1364 log
->page
= le32_to_cpu(logsuper
->end
) / LOGPSIZE
;
1365 log
->eor
= le32_to_cpu(logsuper
->end
) - (LOGPSIZE
* log
->page
);
1368 * initialize for log append write mode
1370 /* establish current/end-of-log page/buffer */
1371 if ((rc
= lbmRead(log
, log
->page
, &bp
)))
1374 lp
= (struct logpage
*) bp
->l_ldata
;
1376 jfs_info("lmLogInit: lsn:0x%x page:%d eor:%d:%d",
1377 le32_to_cpu(logsuper
->end
), log
->page
, log
->eor
,
1378 le16_to_cpu(lp
->h
.eor
));
1381 bp
->l_pn
= log
->page
;
1382 bp
->l_eor
= log
->eor
;
1384 /* if current page is full, move on to next page */
1385 if (log
->eor
>= LOGPSIZE
- LOGPTLRSIZE
)
1389 * initialize log syncpoint
1392 * write the first SYNCPT record with syncpoint = 0
1393 * (i.e., log redo up to HERE !);
1394 * remove current page from lbm write queue at end of pageout
1395 * (to write log superblock update), but do not release to
1400 lrd
.type
= cpu_to_le16(LOG_SYNCPT
);
1402 lrd
.log
.syncpt
.sync
= 0;
1403 lsn
= lmWriteRecord(log
, NULL
, &lrd
, NULL
);
1405 bp
->l_ceor
= bp
->l_eor
;
1406 lp
= (struct logpage
*) bp
->l_ldata
;
1407 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(bp
->l_eor
);
1408 lbmWrite(log
, bp
, lbmWRITE
| lbmSYNC
, 0);
1409 if ((rc
= lbmIOWait(bp
, 0)))
1413 * update/write superblock
1415 logsuper
->state
= cpu_to_le32(LOGMOUNT
);
1416 log
->serial
= le32_to_cpu(logsuper
->serial
) + 1;
1417 logsuper
->serial
= cpu_to_le32(log
->serial
);
1418 lbmDirectWrite(log
, bpsuper
, lbmWRITE
| lbmRELEASE
| lbmSYNC
);
1419 if ((rc
= lbmIOWait(bpsuper
, lbmFREE
)))
1423 /* initialize logsync parameters */
1424 log
->logsize
= (log
->size
- 2) << L2LOGPSIZE
;
1427 log
->sync
= log
->syncpt
;
1428 log
->nextsync
= LOGSYNC_DELTA(log
->logsize
);
1430 jfs_info("lmLogInit: lsn:0x%x syncpt:0x%x sync:0x%x",
1431 log
->lsn
, log
->syncpt
, log
->sync
);
1434 * initialize for lazy/group commit
1443 errout30
: /* release log page */
1445 bp
->l_wqnext
= NULL
;
1448 errout20
: /* release log superblock */
1451 errout10
: /* unwind lbmLogInit() */
1452 lbmLogShutdown(log
);
1454 jfs_warn("lmLogInit: exit(%d)", rc
);
1460 * NAME: lmLogClose()
1462 * FUNCTION: remove file system <ipmnt> from active list of log <iplog>
1463 * and close it on last close.
1465 * PARAMETER: sb - superblock
1467 * RETURN: errors from subroutines
1471 int lmLogClose(struct super_block
*sb
)
1473 struct jfs_sb_info
*sbi
= JFS_SBI(sb
);
1474 struct jfs_log
*log
= sbi
->log
;
1475 struct block_device
*bdev
;
1478 jfs_info("lmLogClose: log:0x%p", log
);
1482 list_del(&sbi
->log_list
);
1487 * We need to make sure all of the "written" metapages
1488 * actually make it to disk
1490 sync_blockdev(sb
->s_bdev
);
1492 if (test_bit(log_INLINELOG
, &log
->flag
)) {
1494 * in-line log in host file system
1496 rc
= lmLogShutdown(log
);
1501 if (!log
->no_integrity
)
1502 lmLogFileSystem(log
, sbi
, 0);
1504 if (!list_empty(&log
->sb_list
))
1508 * TODO: ensure that the dummy_log is in a state to allow
1509 * lbmLogShutdown to deallocate all the buffers and call
1510 * kfree against dummy_log. For now, leave dummy_log & its
1511 * buffers in memory, and resuse if another no-integrity mount
1514 if (log
->no_integrity
)
1518 * external log as separate logical volume
1520 list_del(&log
->journal_list
);
1522 rc
= lmLogShutdown(log
);
1531 jfs_info("lmLogClose: exit(%d)", rc
);
1537 * NAME: jfs_flush_journal()
1539 * FUNCTION: initiate write of any outstanding transactions to the journal
1540 * and optionally wait until they are all written to disk
1542 * wait == 0 flush until latest txn is committed, don't wait
1543 * wait == 1 flush until latest txn is committed, wait
1544 * wait > 1 flush until all txn's are complete, wait
1546 void jfs_flush_journal(struct jfs_log
*log
, int wait
)
1549 struct tblock
*target
= NULL
;
1551 /* jfs_write_inode may call us during read-only mount */
1555 jfs_info("jfs_flush_journal: log:0x%p wait=%d", log
, wait
);
1559 if (!list_empty(&log
->cqueue
)) {
1561 * This ensures that we will keep writing to the journal as long
1562 * as there are unwritten commit records
1564 target
= list_entry(log
->cqueue
.prev
, struct tblock
, cqueue
);
1566 if (test_bit(log_FLUSH
, &log
->flag
)) {
1568 * We're already flushing.
1569 * if flush_tblk is NULL, we are flushing everything,
1570 * so leave it that way. Otherwise, update it to the
1571 * latest transaction
1573 if (log
->flush_tblk
)
1574 log
->flush_tblk
= target
;
1576 /* Only flush until latest transaction is committed */
1577 log
->flush_tblk
= target
;
1578 set_bit(log_FLUSH
, &log
->flag
);
1581 * Initiate I/O on outstanding transactions
1583 if (!(log
->cflag
& logGC_PAGEOUT
)) {
1584 log
->cflag
|= logGC_PAGEOUT
;
1589 if ((wait
> 1) || test_bit(log_SYNCBARRIER
, &log
->flag
)) {
1590 /* Flush until all activity complete */
1591 set_bit(log_FLUSH
, &log
->flag
);
1592 log
->flush_tblk
= NULL
;
1595 if (wait
&& target
&& !(target
->flag
& tblkGC_COMMITTED
)) {
1596 DECLARE_WAITQUEUE(__wait
, current
);
1598 add_wait_queue(&target
->gcwait
, &__wait
);
1599 set_current_state(TASK_UNINTERRUPTIBLE
);
1602 current
->state
= TASK_RUNNING
;
1604 remove_wait_queue(&target
->gcwait
, &__wait
);
1612 * If there was recent activity, we may need to wait
1613 * for the lazycommit thread to catch up
1615 if ((!list_empty(&log
->cqueue
)) || !list_empty(&log
->synclist
)) {
1616 for (i
= 0; i
< 800; i
++) { /* Too much? */
1618 if (list_empty(&log
->cqueue
) &&
1619 list_empty(&log
->synclist
))
1623 assert(list_empty(&log
->cqueue
));
1624 assert(list_empty(&log
->synclist
));
1625 clear_bit(log_FLUSH
, &log
->flag
);
1629 * NAME: lmLogShutdown()
1631 * FUNCTION: log shutdown at last LogClose().
1633 * write log syncpt record.
1634 * update super block to set redone flag to 0.
1636 * PARAMETER: log - log inode
1638 * RETURN: 0 - success
1640 * serialization: single last close thread
1642 int lmLogShutdown(struct jfs_log
* log
)
1647 struct logsuper
*logsuper
;
1648 struct lbuf
*bpsuper
;
1652 jfs_info("lmLogShutdown: log:0x%p", log
);
1654 jfs_flush_journal(log
, 2);
1657 * write the last SYNCPT record with syncpoint = 0
1658 * (i.e., log redo up to HERE !)
1662 lrd
.type
= cpu_to_le16(LOG_SYNCPT
);
1664 lrd
.log
.syncpt
.sync
= 0;
1666 lsn
= lmWriteRecord(log
, NULL
, &lrd
, NULL
);
1668 lp
= (struct logpage
*) bp
->l_ldata
;
1669 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(bp
->l_eor
);
1670 lbmWrite(log
, log
->bp
, lbmWRITE
| lbmRELEASE
| lbmSYNC
, 0);
1671 lbmIOWait(log
->bp
, lbmFREE
);
1674 * synchronous update log superblock
1675 * mark log state as shutdown cleanly
1676 * (i.e., Log does not need to be replayed).
1678 if ((rc
= lbmRead(log
, 1, &bpsuper
)))
1681 logsuper
= (struct logsuper
*) bpsuper
->l_ldata
;
1682 logsuper
->state
= cpu_to_le32(LOGREDONE
);
1683 logsuper
->end
= cpu_to_le32(lsn
);
1684 lbmDirectWrite(log
, bpsuper
, lbmWRITE
| lbmRELEASE
| lbmSYNC
);
1685 rc
= lbmIOWait(bpsuper
, lbmFREE
);
1687 jfs_info("lmLogShutdown: lsn:0x%x page:%d eor:%d",
1688 lsn
, log
->page
, log
->eor
);
1692 * shutdown per log i/o
1694 lbmLogShutdown(log
);
1697 jfs_warn("lmLogShutdown: exit(%d)", rc
);
1704 * NAME: lmLogFileSystem()
1706 * FUNCTION: insert (<activate> = true)/remove (<activate> = false)
1707 * file system into/from log active file system list.
1709 * PARAMETE: log - pointer to logs inode.
1710 * fsdev - kdev_t of filesystem.
1711 * serial - pointer to returned log serial number
1712 * activate - insert/remove device from active list.
1714 * RETURN: 0 - success
1715 * errors returned by vms_iowait().
1717 static int lmLogFileSystem(struct jfs_log
* log
, struct jfs_sb_info
*sbi
,
1722 struct logsuper
*logsuper
;
1723 struct lbuf
*bpsuper
;
1724 char *uuid
= sbi
->uuid
;
1727 * insert/remove file system device to log active file system list.
1729 if ((rc
= lbmRead(log
, 1, &bpsuper
)))
1732 logsuper
= (struct logsuper
*) bpsuper
->l_ldata
;
1734 for (i
= 0; i
< MAX_ACTIVE
; i
++)
1735 if (!memcmp(logsuper
->active
[i
].uuid
, NULL_UUID
, 16)) {
1736 memcpy(logsuper
->active
[i
].uuid
, uuid
, 16);
1740 if (i
== MAX_ACTIVE
) {
1741 jfs_warn("Too many file systems sharing journal!");
1743 return -EMFILE
; /* Is there a better rc? */
1746 for (i
= 0; i
< MAX_ACTIVE
; i
++)
1747 if (!memcmp(logsuper
->active
[i
].uuid
, uuid
, 16)) {
1748 memcpy(logsuper
->active
[i
].uuid
, NULL_UUID
, 16);
1751 if (i
== MAX_ACTIVE
) {
1752 jfs_warn("Somebody stomped on the journal!");
1760 * synchronous write log superblock:
1762 * write sidestream bypassing write queue:
1763 * at file system mount, log super block is updated for
1764 * activation of the file system before any log record
1765 * (MOUNT record) of the file system, and at file system
1766 * unmount, all meta data for the file system has been
1767 * flushed before log super block is updated for deactivation
1768 * of the file system.
1770 lbmDirectWrite(log
, bpsuper
, lbmWRITE
| lbmRELEASE
| lbmSYNC
);
1771 rc
= lbmIOWait(bpsuper
, lbmFREE
);
1777 * log buffer manager (lbm)
1778 * ------------------------
1780 * special purpose buffer manager supporting log i/o requirements.
1782 * per log write queue:
1783 * log pageout occurs in serial order by fifo write queue and
1784 * restricting to a single i/o in pregress at any one time.
1785 * a circular singly-linked list
1786 * (log->wrqueue points to the tail, and buffers are linked via
1787 * bp->wrqueue field), and
1788 * maintains log page in pageout ot waiting for pageout in serial pageout.
1794 * initialize per log I/O setup at lmLogInit()
1796 static int lbmLogInit(struct jfs_log
* log
)
1801 jfs_info("lbmLogInit: log:0x%p", log
);
1803 /* initialize current buffer cursor */
1806 /* initialize log device write queue */
1810 * Each log has its own buffer pages allocated to it. These are
1811 * not managed by the page cache. This ensures that a transaction
1812 * writing to the log does not block trying to allocate a page from
1813 * the page cache (for the log). This would be bad, since page
1814 * allocation waits on the kswapd thread that may be committing inodes
1815 * which would cause log activity. Was that clear? I'm trying to
1816 * avoid deadlock here.
1818 init_waitqueue_head(&log
->free_wait
);
1820 log
->lbuf_free
= NULL
;
1822 for (i
= 0; i
< LOGPAGES
; i
++) {
1823 lbuf
= kmalloc(sizeof(struct lbuf
), GFP_KERNEL
);
1826 lbuf
->l_ldata
= (char *) get_zeroed_page(GFP_KERNEL
);
1827 if (lbuf
->l_ldata
== 0) {
1832 init_waitqueue_head(&lbuf
->l_ioevent
);
1834 lbuf
->l_freelist
= log
->lbuf_free
;
1835 log
->lbuf_free
= lbuf
;
1841 lbmLogShutdown(log
);
1849 * finalize per log I/O setup at lmLogShutdown()
1851 static void lbmLogShutdown(struct jfs_log
* log
)
1855 jfs_info("lbmLogShutdown: log:0x%p", log
);
1857 lbuf
= log
->lbuf_free
;
1859 struct lbuf
*next
= lbuf
->l_freelist
;
1860 free_page((unsigned long) lbuf
->l_ldata
);
1872 * allocate an empty log buffer
1874 static struct lbuf
*lbmAllocate(struct jfs_log
* log
, int pn
)
1877 unsigned long flags
;
1880 * recycle from log buffer freelist if any
1883 LCACHE_SLEEP_COND(log
->free_wait
, (bp
= log
->lbuf_free
), flags
);
1884 log
->lbuf_free
= bp
->l_freelist
;
1885 LCACHE_UNLOCK(flags
);
1889 bp
->l_wqnext
= NULL
;
1890 bp
->l_freelist
= NULL
;
1893 bp
->l_blkno
= log
->base
+ (pn
<< (L2LOGPSIZE
- log
->l2bsize
));
1903 * release a log buffer to freelist
1905 static void lbmFree(struct lbuf
* bp
)
1907 unsigned long flags
;
1913 LCACHE_UNLOCK(flags
);
1916 static void lbmfree(struct lbuf
* bp
)
1918 struct jfs_log
*log
= bp
->l_log
;
1920 assert(bp
->l_wqnext
== NULL
);
1923 * return the buffer to head of freelist
1925 bp
->l_freelist
= log
->lbuf_free
;
1926 log
->lbuf_free
= bp
;
1928 wake_up(&log
->free_wait
);
1936 * FUNCTION: add a log buffer to the the log redrive list
1942 * Takes log_redrive_lock.
1944 static inline void lbmRedrive(struct lbuf
*bp
)
1946 unsigned long flags
;
1948 spin_lock_irqsave(&log_redrive_lock
, flags
);
1949 bp
->l_redrive_next
= log_redrive_list
;
1950 log_redrive_list
= bp
;
1951 spin_unlock_irqrestore(&log_redrive_lock
, flags
);
1953 wake_up(&jfs_IO_thread_wait
);
1960 static int lbmRead(struct jfs_log
* log
, int pn
, struct lbuf
** bpp
)
1966 * allocate a log buffer
1968 *bpp
= bp
= lbmAllocate(log
, pn
);
1969 jfs_info("lbmRead: bp:0x%p pn:0x%x", bp
, pn
);
1971 bp
->l_flag
|= lbmREAD
;
1973 bio
= bio_alloc(GFP_NOFS
, 1);
1975 bio
->bi_sector
= bp
->l_blkno
<< (log
->l2bsize
- 9);
1976 bio
->bi_bdev
= log
->bdev
;
1977 bio
->bi_io_vec
[0].bv_page
= virt_to_page(bp
->l_ldata
);
1978 bio
->bi_io_vec
[0].bv_len
= LOGPSIZE
;
1979 bio
->bi_io_vec
[0].bv_offset
= 0;
1983 bio
->bi_size
= LOGPSIZE
;
1985 bio
->bi_end_io
= lbmIODone
;
1986 bio
->bi_private
= bp
;
1987 submit_bio(READ_SYNC
, bio
);
1989 wait_event(bp
->l_ioevent
, (bp
->l_flag
!= lbmREAD
));
1998 * buffer at head of pageout queue stays after completion of
1999 * partial-page pageout and redriven by explicit initiation of
2000 * pageout by caller until full-page pageout is completed and
2003 * device driver i/o done redrives pageout of new buffer at
2004 * head of pageout queue when current buffer at head of pageout
2005 * queue is released at the completion of its full-page pageout.
2007 * LOGGC_LOCK() serializes lbmWrite() by lmNextPage() and lmGroupCommit().
2008 * LCACHE_LOCK() serializes xflag between lbmWrite() and lbmIODone()
2010 static void lbmWrite(struct jfs_log
* log
, struct lbuf
* bp
, int flag
,
2014 unsigned long flags
;
2016 jfs_info("lbmWrite: bp:0x%p flag:0x%x pn:0x%x", bp
, flag
, bp
->l_pn
);
2018 /* map the logical block address to physical block address */
2020 log
->base
+ (bp
->l_pn
<< (L2LOGPSIZE
- log
->l2bsize
));
2022 LCACHE_LOCK(flags
); /* disable+lock */
2025 * initialize buffer for device driver
2030 * insert bp at tail of write queue associated with log
2032 * (request is either for bp already/currently at head of queue
2033 * or new bp to be inserted at tail)
2037 /* is buffer not already on write queue ? */
2038 if (bp
->l_wqnext
== NULL
) {
2039 /* insert at tail of wqueue */
2045 bp
->l_wqnext
= tail
->l_wqnext
;
2046 tail
->l_wqnext
= bp
;
2052 /* is buffer at head of wqueue and for write ? */
2053 if ((bp
!= tail
->l_wqnext
) || !(flag
& lbmWRITE
)) {
2054 LCACHE_UNLOCK(flags
); /* unlock+enable */
2058 LCACHE_UNLOCK(flags
); /* unlock+enable */
2062 else if (flag
& lbmSYNC
)
2075 * initiate pageout bypassing write queue for sidestream
2076 * (e.g., log superblock) write;
2078 static void lbmDirectWrite(struct jfs_log
* log
, struct lbuf
* bp
, int flag
)
2080 jfs_info("lbmDirectWrite: bp:0x%p flag:0x%x pn:0x%x",
2081 bp
, flag
, bp
->l_pn
);
2084 * initialize buffer for device driver
2086 bp
->l_flag
= flag
| lbmDIRECT
;
2088 /* map the logical block address to physical block address */
2090 log
->base
+ (bp
->l_pn
<< (L2LOGPSIZE
- log
->l2bsize
));
2093 * initiate pageout of the page
2100 * NAME: lbmStartIO()
2102 * FUNCTION: Interface to DD strategy routine
2106 * serialization: LCACHE_LOCK() is NOT held during log i/o;
2108 static void lbmStartIO(struct lbuf
* bp
)
2111 struct jfs_log
*log
= bp
->l_log
;
2113 jfs_info("lbmStartIO\n");
2115 bio
= bio_alloc(GFP_NOFS
, 1);
2116 bio
->bi_sector
= bp
->l_blkno
<< (log
->l2bsize
- 9);
2117 bio
->bi_bdev
= log
->bdev
;
2118 bio
->bi_io_vec
[0].bv_page
= virt_to_page(bp
->l_ldata
);
2119 bio
->bi_io_vec
[0].bv_len
= LOGPSIZE
;
2120 bio
->bi_io_vec
[0].bv_offset
= 0;
2124 bio
->bi_size
= LOGPSIZE
;
2126 bio
->bi_end_io
= lbmIODone
;
2127 bio
->bi_private
= bp
;
2129 /* check if journaling to disk has been disabled */
2130 if (!log
->no_integrity
) {
2131 submit_bio(WRITE_SYNC
, bio
);
2132 INCREMENT(lmStat
.submitted
);
2136 lbmIODone(bio
, 0, 0); /* 2nd argument appears to not be used => 0
2137 * 3rd argument appears to not be used => 0
2146 static int lbmIOWait(struct lbuf
* bp
, int flag
)
2148 unsigned long flags
;
2151 jfs_info("lbmIOWait1: bp:0x%p flag:0x%x:0x%x", bp
, bp
->l_flag
, flag
);
2153 LCACHE_LOCK(flags
); /* disable+lock */
2155 LCACHE_SLEEP_COND(bp
->l_ioevent
, (bp
->l_flag
& lbmDONE
), flags
);
2157 rc
= (bp
->l_flag
& lbmERROR
) ? -EIO
: 0;
2162 LCACHE_UNLOCK(flags
); /* unlock+enable */
2164 jfs_info("lbmIOWait2: bp:0x%p flag:0x%x:0x%x", bp
, bp
->l_flag
, flag
);
2171 * executed at INTIODONE level
2173 static int lbmIODone(struct bio
*bio
, unsigned int bytes_done
, int error
)
2175 struct lbuf
*bp
= bio
->bi_private
;
2176 struct lbuf
*nextbp
, *tail
;
2177 struct jfs_log
*log
;
2178 unsigned long flags
;
2184 * get back jfs buffer bound to the i/o buffer
2186 jfs_info("lbmIODone: bp:0x%p flag:0x%x", bp
, bp
->l_flag
);
2188 LCACHE_LOCK(flags
); /* disable+lock */
2190 bp
->l_flag
|= lbmDONE
;
2192 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2193 bp
->l_flag
|= lbmERROR
;
2195 jfs_err("lbmIODone: I/O error in JFS log");
2203 if (bp
->l_flag
& lbmREAD
) {
2204 bp
->l_flag
&= ~lbmREAD
;
2206 LCACHE_UNLOCK(flags
); /* unlock+enable */
2208 /* wakeup I/O initiator */
2209 LCACHE_WAKEUP(&bp
->l_ioevent
);
2215 * pageout completion
2217 * the bp at the head of write queue has completed pageout.
2219 * if single-commit/full-page pageout, remove the current buffer
2220 * from head of pageout queue, and redrive pageout with
2221 * the new buffer at head of pageout queue;
2222 * otherwise, the partial-page pageout buffer stays at
2223 * the head of pageout queue to be redriven for pageout
2224 * by lmGroupCommit() until full-page pageout is completed.
2226 bp
->l_flag
&= ~lbmWRITE
;
2227 INCREMENT(lmStat
.pagedone
);
2229 /* update committed lsn */
2231 log
->clsn
= (bp
->l_pn
<< L2LOGPSIZE
) + bp
->l_ceor
;
2233 if (bp
->l_flag
& lbmDIRECT
) {
2234 LCACHE_WAKEUP(&bp
->l_ioevent
);
2235 LCACHE_UNLOCK(flags
);
2241 /* single element queue */
2243 /* remove head buffer of full-page pageout
2244 * from log device write queue
2246 if (bp
->l_flag
& lbmRELEASE
) {
2248 bp
->l_wqnext
= NULL
;
2251 /* multi element queue */
2253 /* remove head buffer of full-page pageout
2254 * from log device write queue
2256 if (bp
->l_flag
& lbmRELEASE
) {
2257 nextbp
= tail
->l_wqnext
= bp
->l_wqnext
;
2258 bp
->l_wqnext
= NULL
;
2261 * redrive pageout of next page at head of write queue:
2262 * redrive next page without any bound tblk
2263 * (i.e., page w/o any COMMIT records), or
2264 * first page of new group commit which has been
2265 * queued after current page (subsequent pageout
2266 * is performed synchronously, except page without
2267 * any COMMITs) by lmGroupCommit() as indicated
2270 if (nextbp
->l_flag
& lbmWRITE
) {
2272 * We can't do the I/O at interrupt time.
2273 * The jfsIO thread can do it
2281 * synchronous pageout:
2283 * buffer has not necessarily been removed from write queue
2284 * (e.g., synchronous write of partial-page with COMMIT):
2285 * leave buffer for i/o initiator to dispose
2287 if (bp
->l_flag
& lbmSYNC
) {
2288 LCACHE_UNLOCK(flags
); /* unlock+enable */
2290 /* wakeup I/O initiator */
2291 LCACHE_WAKEUP(&bp
->l_ioevent
);
2295 * Group Commit pageout:
2297 else if (bp
->l_flag
& lbmGC
) {
2298 LCACHE_UNLOCK(flags
);
2303 * asynchronous pageout:
2305 * buffer must have been removed from write queue:
2306 * insert buffer at head of freelist where it can be recycled
2309 assert(bp
->l_flag
& lbmRELEASE
);
2310 assert(bp
->l_flag
& lbmFREE
);
2313 LCACHE_UNLOCK(flags
); /* unlock+enable */
2319 int jfsIOWait(void *arg
)
2325 complete(&jfsIOwait
);
2328 DECLARE_WAITQUEUE(wq
, current
);
2330 spin_lock_irq(&log_redrive_lock
);
2331 while ((bp
= log_redrive_list
) != 0) {
2332 log_redrive_list
= bp
->l_redrive_next
;
2333 bp
->l_redrive_next
= NULL
;
2334 spin_unlock_irq(&log_redrive_lock
);
2336 spin_lock_irq(&log_redrive_lock
);
2338 if (current
->flags
& PF_FREEZE
) {
2339 spin_unlock_irq(&log_redrive_lock
);
2340 refrigerator(PF_FREEZE
);
2342 add_wait_queue(&jfs_IO_thread_wait
, &wq
);
2343 set_current_state(TASK_INTERRUPTIBLE
);
2344 spin_unlock_irq(&log_redrive_lock
);
2346 current
->state
= TASK_RUNNING
;
2347 remove_wait_queue(&jfs_IO_thread_wait
, &wq
);
2349 } while (!jfs_stop_threads
);
2351 jfs_info("jfsIOWait being killed!");
2352 complete_and_exit(&jfsIOwait
, 0);
2356 * NAME: lmLogFormat()/jfs_logform()
2358 * FUNCTION: format file system log
2362 * logAddress - start address of log space in FS block
2363 * logSize - length of log space in FS block;
2365 * RETURN: 0 - success
2368 * XXX: We're synchronously writing one page at a time. This needs to
2369 * be improved by writing multiple pages at once.
2371 int lmLogFormat(struct jfs_log
*log
, s64 logAddress
, int logSize
)
2374 struct jfs_sb_info
*sbi
;
2375 struct logsuper
*logsuper
;
2377 int lspn
; /* log sequence page number */
2378 struct lrd
*lrd_ptr
;
2382 jfs_info("lmLogFormat: logAddress:%Ld logSize:%d",
2383 (long long)logAddress
, logSize
);
2385 sbi
= list_entry(log
->sb_list
.next
, struct jfs_sb_info
, log_list
);
2387 /* allocate a log buffer */
2388 bp
= lbmAllocate(log
, 1);
2390 npages
= logSize
>> sbi
->l2nbperpage
;
2395 * page 0 - reserved;
2396 * page 1 - log superblock;
2397 * page 2 - log data page: A SYNC log record is written
2398 * into this page at logform time;
2399 * pages 3-N - log data page: set to empty log data pages;
2402 * init log superblock: log page 1
2404 logsuper
= (struct logsuper
*) bp
->l_ldata
;
2406 logsuper
->magic
= cpu_to_le32(LOGMAGIC
);
2407 logsuper
->version
= cpu_to_le32(LOGVERSION
);
2408 logsuper
->state
= cpu_to_le32(LOGREDONE
);
2409 logsuper
->flag
= cpu_to_le32(sbi
->mntflag
); /* ? */
2410 logsuper
->size
= cpu_to_le32(npages
);
2411 logsuper
->bsize
= cpu_to_le32(sbi
->bsize
);
2412 logsuper
->l2bsize
= cpu_to_le32(sbi
->l2bsize
);
2413 logsuper
->end
= cpu_to_le32(2 * LOGPSIZE
+ LOGPHDRSIZE
+ LOGRDSIZE
);
2415 bp
->l_flag
= lbmWRITE
| lbmSYNC
| lbmDIRECT
;
2416 bp
->l_blkno
= logAddress
+ sbi
->nbperpage
;
2418 if ((rc
= lbmIOWait(bp
, 0)))
2422 * init pages 2 to npages-1 as log data pages:
2424 * log page sequence number (lpsn) initialization:
2427 * +-----+-----+=====+=====+===.....===+=====+
2429 * <--- N page circular file ---->
2431 * the N (= npages-2) data pages of the log is maintained as
2432 * a circular file for the log records;
2433 * lpsn grows by 1 monotonically as each log page is written
2434 * to the circular file of the log;
2435 * and setLogpage() will not reset the page number even if
2436 * the eor is equal to LOGPHDRSIZE. In order for binary search
2437 * still work in find log end process, we have to simulate the
2438 * log wrap situation at the log format time.
2439 * The 1st log page written will have the highest lpsn. Then
2440 * the succeeding log pages will have ascending order of
2441 * the lspn starting from 0, ... (N-2)
2443 lp
= (struct logpage
*) bp
->l_ldata
;
2445 * initialize 1st log page to be written: lpsn = N - 1,
2446 * write a SYNCPT log record is written to this page
2448 lp
->h
.page
= lp
->t
.page
= cpu_to_le32(npages
- 3);
2449 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(LOGPHDRSIZE
+ LOGRDSIZE
);
2451 lrd_ptr
= (struct lrd
*) &lp
->data
;
2452 lrd_ptr
->logtid
= 0;
2453 lrd_ptr
->backchain
= 0;
2454 lrd_ptr
->type
= cpu_to_le16(LOG_SYNCPT
);
2455 lrd_ptr
->length
= 0;
2456 lrd_ptr
->log
.syncpt
.sync
= 0;
2458 bp
->l_blkno
+= sbi
->nbperpage
;
2459 bp
->l_flag
= lbmWRITE
| lbmSYNC
| lbmDIRECT
;
2461 if ((rc
= lbmIOWait(bp
, 0)))
2465 * initialize succeeding log pages: lpsn = 0, 1, ..., (N-2)
2467 for (lspn
= 0; lspn
< npages
- 3; lspn
++) {
2468 lp
->h
.page
= lp
->t
.page
= cpu_to_le32(lspn
);
2469 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(LOGPHDRSIZE
);
2471 bp
->l_blkno
+= sbi
->nbperpage
;
2472 bp
->l_flag
= lbmWRITE
| lbmSYNC
| lbmDIRECT
;
2474 if ((rc
= lbmIOWait(bp
, 0)))
2483 /* release the buffer */
2489 #ifdef CONFIG_JFS_STATISTICS
2490 int jfs_lmstats_read(char *buffer
, char **start
, off_t offset
, int length
,
2491 int *eof
, void *data
)
2496 len
+= sprintf(buffer
,
2497 "JFS Logmgr stats\n"
2498 "================\n"
2500 "writes submitted = %d\n"
2501 "writes completed = %d\n"
2502 "full pages submitted = %d\n"
2503 "partial pages submitted = %d\n",
2508 lmStat
.partial_page
);
2511 *start
= buffer
+ begin
;
2524 #endif /* CONFIG_JFS_STATISTICS */