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/kthread.h>
68 #include <linux/buffer_head.h> /* for sync_blockdev() */
69 #include <linux/bio.h>
70 #include <linux/suspend.h>
71 #include <linux/delay.h>
72 #include "jfs_incore.h"
73 #include "jfs_filsys.h"
74 #include "jfs_metapage.h"
75 #include "jfs_superblock.h"
76 #include "jfs_txnmgr.h"
77 #include "jfs_debug.h"
81 * lbuf's ready to be redriven. Protected by log_redrive_lock (jfsIO thread)
83 static struct lbuf
*log_redrive_list
;
84 static DEFINE_SPINLOCK(log_redrive_lock
);
88 * log read/write serialization (per log)
90 #define LOG_LOCK_INIT(log) mutex_init(&(log)->loglock)
91 #define LOG_LOCK(log) mutex_lock(&((log)->loglock))
92 #define LOG_UNLOCK(log) mutex_unlock(&((log)->loglock))
96 * log group commit serialization (per log)
99 #define LOGGC_LOCK_INIT(log) spin_lock_init(&(log)->gclock)
100 #define LOGGC_LOCK(log) spin_lock_irq(&(log)->gclock)
101 #define LOGGC_UNLOCK(log) spin_unlock_irq(&(log)->gclock)
102 #define LOGGC_WAKEUP(tblk) wake_up_all(&(tblk)->gcwait)
105 * log sync serialization (per log)
107 #define LOGSYNC_DELTA(logsize) min((logsize)/8, 128*LOGPSIZE)
108 #define LOGSYNC_BARRIER(logsize) ((logsize)/4)
110 #define LOGSYNC_DELTA(logsize) min((logsize)/4, 256*LOGPSIZE)
111 #define LOGSYNC_BARRIER(logsize) ((logsize)/2)
116 * log buffer cache synchronization
118 static DEFINE_SPINLOCK(jfsLCacheLock
);
120 #define LCACHE_LOCK(flags) spin_lock_irqsave(&jfsLCacheLock, flags)
121 #define LCACHE_UNLOCK(flags) spin_unlock_irqrestore(&jfsLCacheLock, flags)
124 * See __SLEEP_COND in jfs_locks.h
126 #define LCACHE_SLEEP_COND(wq, cond, flags) \
130 __SLEEP_COND(wq, cond, LCACHE_LOCK(flags), LCACHE_UNLOCK(flags)); \
133 #define LCACHE_WAKEUP(event) wake_up(event)
137 * lbuf buffer cache (lCache) control
139 /* log buffer manager pageout control (cumulative, inclusive) */
140 #define lbmREAD 0x0001
141 #define lbmWRITE 0x0002 /* enqueue at tail of write queue;
142 * init pageout if at head of queue;
144 #define lbmRELEASE 0x0004 /* remove from write queue
145 * at completion of pageout;
146 * do not free/recycle it yet:
147 * caller will free it;
149 #define lbmSYNC 0x0008 /* do not return to freelist
150 * when removed from write queue;
152 #define lbmFREE 0x0010 /* return to freelist
153 * at completion of pageout;
154 * the buffer may be recycled;
156 #define lbmDONE 0x0020
157 #define lbmERROR 0x0040
158 #define lbmGC 0x0080 /* lbmIODone to perform post-GC processing
161 #define lbmDIRECT 0x0100
164 * Global list of active external journals
166 static LIST_HEAD(jfs_external_logs
);
167 static struct jfs_log
*dummy_log
= NULL
;
168 static DECLARE_MUTEX(jfs_log_sem
);
173 static int lmWriteRecord(struct jfs_log
* log
, struct tblock
* tblk
,
174 struct lrd
* lrd
, struct tlock
* tlck
);
176 static int lmNextPage(struct jfs_log
* log
);
177 static int lmLogFileSystem(struct jfs_log
* log
, struct jfs_sb_info
*sbi
,
180 static int open_inline_log(struct super_block
*sb
);
181 static int open_dummy_log(struct super_block
*sb
);
182 static int lbmLogInit(struct jfs_log
* log
);
183 static void lbmLogShutdown(struct jfs_log
* log
);
184 static struct lbuf
*lbmAllocate(struct jfs_log
* log
, int);
185 static void lbmFree(struct lbuf
* bp
);
186 static void lbmfree(struct lbuf
* bp
);
187 static int lbmRead(struct jfs_log
* log
, int pn
, struct lbuf
** bpp
);
188 static void lbmWrite(struct jfs_log
* log
, struct lbuf
* bp
, int flag
, int cant_block
);
189 static void lbmDirectWrite(struct jfs_log
* log
, struct lbuf
* bp
, int flag
);
190 static int lbmIOWait(struct lbuf
* bp
, int flag
);
191 static bio_end_io_t lbmIODone
;
192 static void lbmStartIO(struct lbuf
* bp
);
193 static void lmGCwrite(struct jfs_log
* log
, int cant_block
);
194 static int lmLogSync(struct jfs_log
* log
, int hard_sync
);
201 #ifdef CONFIG_JFS_STATISTICS
202 static struct lmStat
{
203 uint commit
; /* # of commit */
204 uint pagedone
; /* # of page written */
205 uint submitted
; /* # of pages submitted */
206 uint full_page
; /* # of full pages submitted */
207 uint partial_page
; /* # of partial pages submitted */
215 * FUNCTION: write a log record;
219 * RETURN: lsn - offset to the next log record to write (end-of-log);
222 * note: todo: log error handler
224 int lmLog(struct jfs_log
* log
, struct tblock
* tblk
, struct lrd
* lrd
,
229 struct metapage
*mp
= NULL
;
232 jfs_info("lmLog: log:0x%p tblk:0x%p, lrd:0x%p tlck:0x%p",
233 log
, tblk
, lrd
, tlck
);
237 /* log by (out-of-transaction) JFS ? */
241 /* log from page ? */
243 tlck
->type
& tlckBTROOT
|| (mp
= tlck
->mp
) == NULL
)
247 * initialize/update page/transaction recovery lsn
251 LOGSYNC_LOCK(log
, flags
);
254 * initialize page lsn if first log write of the page
261 /* insert page at tail of logsynclist */
262 list_add_tail(&mp
->synclist
, &log
->synclist
);
266 * initialize/update lsn of tblock of the page
268 * transaction inherits oldest lsn of pages associated
269 * with allocation/deallocation of resources (their
270 * log records are used to reconstruct allocation map
271 * at recovery time: inode for inode allocation map,
272 * B+-tree index of extent descriptors for block
274 * allocation map pages inherit transaction lsn at
275 * commit time to allow forwarding log syncpt past log
276 * records associated with allocation/deallocation of
277 * resources only after persistent map of these map pages
278 * have been updated and propagated to home.
281 * initialize transaction lsn:
283 if (tblk
->lsn
== 0) {
284 /* inherit lsn of its first page logged */
288 /* insert tblock after the page on logsynclist */
289 list_add(&tblk
->synclist
, &mp
->synclist
);
292 * update transaction lsn:
295 /* inherit oldest/smallest lsn of page */
296 logdiff(diffp
, mp
->lsn
, log
);
297 logdiff(difft
, tblk
->lsn
, log
);
299 /* update tblock lsn with page lsn */
302 /* move tblock after page on logsynclist */
303 list_move(&tblk
->synclist
, &mp
->synclist
);
307 LOGSYNC_UNLOCK(log
, flags
);
310 * write the log record
313 lsn
= lmWriteRecord(log
, tblk
, lrd
, tlck
);
316 * forward log syncpt if log reached next syncpt trigger
318 logdiff(diffp
, lsn
, log
);
319 if (diffp
>= log
->nextsync
)
320 lsn
= lmLogSync(log
, 0);
322 /* update end-of-log lsn */
327 /* return end-of-log address */
332 * NAME: lmWriteRecord()
334 * FUNCTION: move the log record to current log page
336 * PARAMETER: cd - commit descriptor
338 * RETURN: end-of-log address
340 * serialization: LOG_LOCK() held on entry/exit
343 lmWriteRecord(struct jfs_log
* log
, struct tblock
* tblk
, struct lrd
* lrd
,
346 int lsn
= 0; /* end-of-log address */
347 struct lbuf
*bp
; /* dst log page buffer */
348 struct logpage
*lp
; /* dst log page */
349 caddr_t dst
; /* destination address in log page */
350 int dstoffset
; /* end-of-log offset in log page */
351 int freespace
; /* free space in log page */
352 caddr_t p
; /* src meta-data page */
355 int nbytes
; /* number of bytes to move */
358 struct linelock
*linelock
;
365 /* retrieve destination log page to write */
366 bp
= (struct lbuf
*) log
->bp
;
367 lp
= (struct logpage
*) bp
->l_ldata
;
368 dstoffset
= log
->eor
;
370 /* any log data to write ? */
375 * move log record data
377 /* retrieve source meta-data page to log */
378 if (tlck
->flag
& tlckPAGELOCK
) {
379 p
= (caddr_t
) (tlck
->mp
->data
);
380 linelock
= (struct linelock
*) & tlck
->lock
;
382 /* retrieve source in-memory inode to log */
383 else if (tlck
->flag
& tlckINODELOCK
) {
384 if (tlck
->type
& tlckDTREE
)
385 p
= (caddr_t
) &JFS_IP(tlck
->ip
)->i_dtroot
;
387 p
= (caddr_t
) &JFS_IP(tlck
->ip
)->i_xtroot
;
388 linelock
= (struct linelock
*) & tlck
->lock
;
391 else if (tlck
->flag
& tlckINLINELOCK
) {
393 inlinelock
= (struct inlinelock
*) & tlck
;
394 p
= (caddr_t
) & inlinelock
->pxd
;
395 linelock
= (struct linelock
*) & tlck
;
397 #endif /* _JFS_WIP */
399 jfs_err("lmWriteRecord: UFO tlck:0x%p", tlck
);
400 return 0; /* Probably should trap */
402 l2linesize
= linelock
->l2linesize
;
405 ASSERT(linelock
->index
<= linelock
->maxcnt
);
408 for (i
= 0; i
< linelock
->index
; i
++, lv
++) {
413 if (dstoffset
>= LOGPSIZE
- LOGPTLRSIZE
) {
414 /* page become full: move on to next page */
418 lp
= (struct logpage
*) bp
->l_ldata
;
419 dstoffset
= LOGPHDRSIZE
;
423 * move log vector data
425 src
= (u8
*) p
+ (lv
->offset
<< l2linesize
);
426 srclen
= lv
->length
<< l2linesize
;
429 freespace
= (LOGPSIZE
- LOGPTLRSIZE
) - dstoffset
;
430 nbytes
= min(freespace
, srclen
);
431 dst
= (caddr_t
) lp
+ dstoffset
;
432 memcpy(dst
, src
, nbytes
);
435 /* is page not full ? */
436 if (dstoffset
< LOGPSIZE
- LOGPTLRSIZE
)
439 /* page become full: move on to next page */
442 bp
= (struct lbuf
*) log
->bp
;
443 lp
= (struct logpage
*) bp
->l_ldata
;
444 dstoffset
= LOGPHDRSIZE
;
451 * move log vector descriptor
454 lvd
= (struct lvd
*) ((caddr_t
) lp
+ dstoffset
);
455 lvd
->offset
= cpu_to_le16(lv
->offset
);
456 lvd
->length
= cpu_to_le16(lv
->length
);
458 jfs_info("lmWriteRecord: lv offset:%d length:%d",
459 lv
->offset
, lv
->length
);
462 if ((i
= linelock
->next
)) {
463 linelock
= (struct linelock
*) lid_to_tlock(i
);
468 * move log record descriptor
471 lrd
->length
= cpu_to_le16(len
);
477 freespace
= (LOGPSIZE
- LOGPTLRSIZE
) - dstoffset
;
478 nbytes
= min(freespace
, srclen
);
479 dst
= (caddr_t
) lp
+ dstoffset
;
480 memcpy(dst
, src
, nbytes
);
485 /* are there more to move than freespace of page ? */
490 * end of log record descriptor
493 /* update last log record eor */
494 log
->eor
= dstoffset
;
495 bp
->l_eor
= dstoffset
;
496 lsn
= (log
->page
<< L2LOGPSIZE
) + dstoffset
;
498 if (lrd
->type
& cpu_to_le16(LOG_COMMIT
)) {
500 jfs_info("wr: tclsn:0x%x, beor:0x%x", tblk
->clsn
,
503 INCREMENT(lmStat
.commit
); /* # of commit */
506 * enqueue tblock for group commit:
508 * enqueue tblock of non-trivial/synchronous COMMIT
509 * at tail of group commit queue
510 * (trivial/asynchronous COMMITs are ignored by
515 /* init tblock gc state */
516 tblk
->flag
= tblkGC_QUEUE
;
518 tblk
->pn
= log
->page
;
519 tblk
->eor
= log
->eor
;
521 /* enqueue transaction to commit queue */
522 list_add_tail(&tblk
->cqueue
, &log
->cqueue
);
527 jfs_info("lmWriteRecord: lrd:0x%04x bp:0x%p pn:%d eor:0x%x",
528 le16_to_cpu(lrd
->type
), log
->bp
, log
->page
, dstoffset
);
530 /* page not full ? */
531 if (dstoffset
< LOGPSIZE
- LOGPTLRSIZE
)
535 /* page become full: move on to next page */
538 bp
= (struct lbuf
*) log
->bp
;
539 lp
= (struct logpage
*) bp
->l_ldata
;
540 dstoffset
= LOGPHDRSIZE
;
551 * FUNCTION: write current page and allocate next page.
557 * serialization: LOG_LOCK() held on entry/exit
559 static int lmNextPage(struct jfs_log
* log
)
562 int lspn
; /* log sequence page number */
563 int pn
; /* current page number */
568 /* get current log page number and log sequence page number */
571 lp
= (struct logpage
*) bp
->l_ldata
;
572 lspn
= le32_to_cpu(lp
->h
.page
);
577 * write or queue the full page at the tail of write queue
579 /* get the tail tblk on commit queue */
580 if (list_empty(&log
->cqueue
))
583 tblk
= list_entry(log
->cqueue
.prev
, struct tblock
, cqueue
);
585 /* every tblk who has COMMIT record on the current page,
586 * and has not been committed, must be on commit queue
587 * since tblk is queued at commit queueu at the time
588 * of writing its COMMIT record on the page before
589 * page becomes full (even though the tblk thread
590 * who wrote COMMIT record may have been suspended
594 /* is page bound with outstanding tail tblk ? */
595 if (tblk
&& tblk
->pn
== pn
) {
596 /* mark tblk for end-of-page */
597 tblk
->flag
|= tblkGC_EOP
;
599 if (log
->cflag
& logGC_PAGEOUT
) {
600 /* if page is not already on write queue,
601 * just enqueue (no lbmWRITE to prevent redrive)
602 * buffer to wqueue to ensure correct serial order
603 * of the pages since log pages will be added
606 if (bp
->l_wqnext
== NULL
)
607 lbmWrite(log
, bp
, 0, 0);
610 * No current GC leader, initiate group commit
612 log
->cflag
|= logGC_PAGEOUT
;
616 /* page is not bound with outstanding tblk:
617 * init write or mark it to be redriven (lbmWRITE)
620 /* finalize the page */
621 bp
->l_ceor
= bp
->l_eor
;
622 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(bp
->l_ceor
);
623 lbmWrite(log
, bp
, lbmWRITE
| lbmRELEASE
| lbmFREE
, 0);
628 * allocate/initialize next page
630 /* if log wraps, the first data page of log is 2
631 * (0 never used, 1 is superblock).
633 log
->page
= (pn
== log
->size
- 1) ? 2 : pn
+ 1;
634 log
->eor
= LOGPHDRSIZE
; /* ? valid page empty/full at logRedo() */
636 /* allocate/initialize next log page buffer */
637 nextbp
= lbmAllocate(log
, log
->page
);
638 nextbp
->l_eor
= log
->eor
;
641 /* initialize next log page */
642 lp
= (struct logpage
*) nextbp
->l_ldata
;
643 lp
->h
.page
= lp
->t
.page
= cpu_to_le32(lspn
+ 1);
644 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(LOGPHDRSIZE
);
651 * NAME: lmGroupCommit()
653 * FUNCTION: group commit
654 * initiate pageout of the pages with COMMIT in the order of
655 * page number - redrive pageout of the page at the head of
656 * pageout queue until full page has been written.
661 * LOGGC_LOCK serializes log group commit queue, and
662 * transaction blocks on the commit queue.
663 * N.B. LOG_LOCK is NOT held during lmGroupCommit().
665 int lmGroupCommit(struct jfs_log
* log
, struct tblock
* tblk
)
671 /* group committed already ? */
672 if (tblk
->flag
& tblkGC_COMMITTED
) {
673 if (tblk
->flag
& tblkGC_ERROR
)
679 jfs_info("lmGroup Commit: tblk = 0x%p, gcrtc = %d", tblk
, log
->gcrtc
);
681 if (tblk
->xflag
& COMMIT_LAZY
)
682 tblk
->flag
|= tblkGC_LAZY
;
684 if ((!(log
->cflag
& logGC_PAGEOUT
)) && (!list_empty(&log
->cqueue
)) &&
685 (!(tblk
->xflag
& COMMIT_LAZY
) || test_bit(log_FLUSH
, &log
->flag
)
686 || jfs_tlocks_low
)) {
688 * No pageout in progress
690 * start group commit as its group leader.
692 log
->cflag
|= logGC_PAGEOUT
;
697 if (tblk
->xflag
& COMMIT_LAZY
) {
699 * Lazy transactions can leave now
705 /* lmGCwrite gives up LOGGC_LOCK, check again */
707 if (tblk
->flag
& tblkGC_COMMITTED
) {
708 if (tblk
->flag
& tblkGC_ERROR
)
715 /* upcount transaction waiting for completion
718 tblk
->flag
|= tblkGC_READY
;
720 __SLEEP_COND(tblk
->gcwait
, (tblk
->flag
& tblkGC_COMMITTED
),
721 LOGGC_LOCK(log
), LOGGC_UNLOCK(log
));
723 /* removed from commit queue */
724 if (tblk
->flag
& tblkGC_ERROR
)
734 * FUNCTION: group commit write
735 * initiate write of log page, building a group of all transactions
736 * with commit records on that page.
741 * LOGGC_LOCK must be held by caller.
742 * N.B. LOG_LOCK is NOT held during lmGroupCommit().
744 static void lmGCwrite(struct jfs_log
* log
, int cant_write
)
748 int gcpn
; /* group commit page number */
750 struct tblock
*xtblk
= NULL
;
753 * build the commit group of a log page
755 * scan commit queue and make a commit group of all
756 * transactions with COMMIT records on the same log page.
758 /* get the head tblk on the commit queue */
759 gcpn
= list_entry(log
->cqueue
.next
, struct tblock
, cqueue
)->pn
;
761 list_for_each_entry(tblk
, &log
->cqueue
, cqueue
) {
762 if (tblk
->pn
!= gcpn
)
767 /* state transition: (QUEUE, READY) -> COMMIT */
768 tblk
->flag
|= tblkGC_COMMIT
;
770 tblk
= xtblk
; /* last tblk of the page */
773 * pageout to commit transactions on the log page.
775 bp
= (struct lbuf
*) tblk
->bp
;
776 lp
= (struct logpage
*) bp
->l_ldata
;
777 /* is page already full ? */
778 if (tblk
->flag
& tblkGC_EOP
) {
779 /* mark page to free at end of group commit of the page */
780 tblk
->flag
&= ~tblkGC_EOP
;
781 tblk
->flag
|= tblkGC_FREE
;
782 bp
->l_ceor
= bp
->l_eor
;
783 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(bp
->l_ceor
);
784 lbmWrite(log
, bp
, lbmWRITE
| lbmRELEASE
| lbmGC
,
786 INCREMENT(lmStat
.full_page
);
788 /* page is not yet full */
790 bp
->l_ceor
= tblk
->eor
; /* ? bp->l_ceor = bp->l_eor; */
791 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(bp
->l_ceor
);
792 lbmWrite(log
, bp
, lbmWRITE
| lbmGC
, cant_write
);
793 INCREMENT(lmStat
.partial_page
);
800 * FUNCTION: group commit post-processing
801 * Processes transactions after their commit records have been written
802 * to disk, redriving log I/O if necessary.
807 * This routine is called a interrupt time by lbmIODone
809 static void lmPostGC(struct lbuf
* bp
)
812 struct jfs_log
*log
= bp
->l_log
;
814 struct tblock
*tblk
, *temp
;
817 spin_lock_irqsave(&log
->gclock
, flags
);
819 * current pageout of group commit completed.
821 * remove/wakeup transactions from commit queue who were
822 * group committed with the current log page
824 list_for_each_entry_safe(tblk
, temp
, &log
->cqueue
, cqueue
) {
825 if (!(tblk
->flag
& tblkGC_COMMIT
))
827 /* if transaction was marked GC_COMMIT then
828 * it has been shipped in the current pageout
829 * and made it to disk - it is committed.
832 if (bp
->l_flag
& lbmERROR
)
833 tblk
->flag
|= tblkGC_ERROR
;
835 /* remove it from the commit queue */
836 list_del(&tblk
->cqueue
);
837 tblk
->flag
&= ~tblkGC_QUEUE
;
839 if (tblk
== log
->flush_tblk
) {
840 /* we can stop flushing the log now */
841 clear_bit(log_FLUSH
, &log
->flag
);
842 log
->flush_tblk
= NULL
;
845 jfs_info("lmPostGC: tblk = 0x%p, flag = 0x%x", tblk
,
848 if (!(tblk
->xflag
& COMMIT_FORCE
))
850 * Hand tblk over to lazy commit thread
854 /* state transition: COMMIT -> COMMITTED */
855 tblk
->flag
|= tblkGC_COMMITTED
;
857 if (tblk
->flag
& tblkGC_READY
)
863 /* was page full before pageout ?
864 * (and this is the last tblk bound with the page)
866 if (tblk
->flag
& tblkGC_FREE
)
868 /* did page become full after pageout ?
869 * (and this is the last tblk bound with the page)
871 else if (tblk
->flag
& tblkGC_EOP
) {
872 /* finalize the page */
873 lp
= (struct logpage
*) bp
->l_ldata
;
874 bp
->l_ceor
= bp
->l_eor
;
875 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(bp
->l_eor
);
876 jfs_info("lmPostGC: calling lbmWrite");
877 lbmWrite(log
, bp
, lbmWRITE
| lbmRELEASE
| lbmFREE
,
883 /* are there any transactions who have entered lnGroupCommit()
884 * (whose COMMITs are after that of the last log page written.
885 * They are waiting for new group commit (above at (SLEEP 1))
886 * or lazy transactions are on a full (queued) log page,
887 * select the latest ready transaction as new group leader and
888 * wake her up to lead her group.
890 if ((!list_empty(&log
->cqueue
)) &&
891 ((log
->gcrtc
> 0) || (tblk
->bp
->l_wqnext
!= NULL
) ||
892 test_bit(log_FLUSH
, &log
->flag
) || jfs_tlocks_low
))
894 * Call lmGCwrite with new group leader
898 /* no transaction are ready yet (transactions are only just
899 * queued (GC_QUEUE) and not entered for group commit yet).
900 * the first transaction entering group commit
901 * will elect herself as new group leader.
904 log
->cflag
&= ~logGC_PAGEOUT
;
907 spin_unlock_irqrestore(&log
->gclock
, flags
);
914 * FUNCTION: write log SYNCPT record for specified log
915 * if new sync address is available
916 * (normally the case if sync() is executed by back-ground
918 * calculate new value of i_nextsync which determines when
919 * this code is called again.
921 * PARAMETERS: log - log structure
922 * hard_sync - 1 to force all metadata to be written
926 * serialization: LOG_LOCK() held on entry/exit
928 static int lmLogSync(struct jfs_log
* log
, int hard_sync
)
931 int written
; /* written since last syncpt */
932 int free
; /* free space left available */
933 int delta
; /* additional delta to write normally */
934 int more
; /* additional write granted */
937 struct logsyncblk
*lp
;
938 struct jfs_sb_info
*sbi
;
941 /* push dirty metapages out to disk */
943 list_for_each_entry(sbi
, &log
->sb_list
, log_list
) {
944 filemap_fdatawrite(sbi
->ipbmap
->i_mapping
);
945 filemap_fdatawrite(sbi
->ipimap
->i_mapping
);
946 filemap_fdatawrite(sbi
->direct_inode
->i_mapping
);
949 list_for_each_entry(sbi
, &log
->sb_list
, log_list
) {
950 filemap_flush(sbi
->ipbmap
->i_mapping
);
951 filemap_flush(sbi
->ipimap
->i_mapping
);
952 filemap_flush(sbi
->direct_inode
->i_mapping
);
958 /* if last sync is same as last syncpt,
959 * invoke sync point forward processing to update sync.
962 if (log
->sync
== log
->syncpt
) {
963 LOGSYNC_LOCK(log
, flags
);
964 if (list_empty(&log
->synclist
))
965 log
->sync
= log
->lsn
;
967 lp
= list_entry(log
->synclist
.next
,
968 struct logsyncblk
, synclist
);
971 LOGSYNC_UNLOCK(log
, flags
);
975 /* if sync is different from last syncpt,
976 * write a SYNCPT record with syncpt = sync.
977 * reset syncpt = sync
979 if (log
->sync
!= log
->syncpt
) {
982 lrd
.type
= cpu_to_le16(LOG_SYNCPT
);
984 lrd
.log
.syncpt
.sync
= cpu_to_le32(log
->sync
);
985 lsn
= lmWriteRecord(log
, NULL
, &lrd
, NULL
);
987 log
->syncpt
= log
->sync
;
992 * setup next syncpt trigger (SWAG)
994 logsize
= log
->logsize
;
996 logdiff(written
, lsn
, log
);
997 free
= logsize
- written
;
998 delta
= LOGSYNC_DELTA(logsize
);
999 more
= min(free
/ 2, delta
);
1000 if (more
< 2 * LOGPSIZE
) {
1001 jfs_warn("\n ... Log Wrap ... Log Wrap ... Log Wrap ...\n");
1005 * option 1 - panic ? No.!
1006 * option 2 - shutdown file systems
1007 * associated with log ?
1008 * option 3 - extend log ?
1011 * option 4 - second chance
1013 * mark log wrapped, and continue.
1014 * when all active transactions are completed,
1015 * mark log vaild for recovery.
1016 * if crashed during invalid state, log state
1017 * implies invald log, forcing fsck().
1019 /* mark log state log wrap in log superblock */
1020 /* log->state = LOGWRAP; */
1022 /* reset sync point computation */
1023 log
->syncpt
= log
->sync
= lsn
;
1024 log
->nextsync
= delta
;
1026 /* next syncpt trigger = written + more */
1027 log
->nextsync
= written
+ more
;
1029 /* if number of bytes written from last sync point is more
1030 * than 1/4 of the log size, stop new transactions from
1031 * starting until all current transactions are completed
1032 * by setting syncbarrier flag.
1034 if (!test_bit(log_SYNCBARRIER
, &log
->flag
) &&
1035 (written
> LOGSYNC_BARRIER(logsize
)) && log
->active
) {
1036 set_bit(log_SYNCBARRIER
, &log
->flag
);
1037 jfs_info("log barrier on: lsn=0x%x syncpt=0x%x", lsn
,
1040 * We may have to initiate group commit
1042 jfs_flush_journal(log
, 0);
1051 * FUNCTION: write log SYNCPT record for specified log
1053 * PARAMETERS: log - log structure
1054 * hard_sync - set to 1 to force metadata to be written
1056 void jfs_syncpt(struct jfs_log
*log
, int hard_sync
)
1058 lmLogSync(log
, hard_sync
);
1065 * FUNCTION: open the log on first open;
1066 * insert filesystem in the active list of the log.
1068 * PARAMETER: ipmnt - file system mount inode
1069 * iplog - log inode (out)
1075 int lmLogOpen(struct super_block
*sb
)
1078 struct block_device
*bdev
;
1079 struct jfs_log
*log
;
1080 struct jfs_sb_info
*sbi
= JFS_SBI(sb
);
1082 if (sbi
->flag
& JFS_NOINTEGRITY
)
1083 return open_dummy_log(sb
);
1085 if (sbi
->mntflag
& JFS_INLINELOG
)
1086 return open_inline_log(sb
);
1089 list_for_each_entry(log
, &jfs_external_logs
, journal_list
) {
1090 if (log
->bdev
->bd_dev
== sbi
->logdev
) {
1091 if (memcmp(log
->uuid
, sbi
->loguuid
,
1092 sizeof(log
->uuid
))) {
1093 jfs_warn("wrong uuid on JFS journal\n");
1098 * add file system to log active file system list
1100 if ((rc
= lmLogFileSystem(log
, sbi
, 1))) {
1108 if (!(log
= kzalloc(sizeof(struct jfs_log
), GFP_KERNEL
))) {
1112 INIT_LIST_HEAD(&log
->sb_list
);
1113 init_waitqueue_head(&log
->syncwait
);
1116 * external log as separate logical volume
1118 * file systems to log may have n-to-1 relationship;
1121 bdev
= open_by_devnum(sbi
->logdev
, FMODE_READ
|FMODE_WRITE
);
1123 rc
= -PTR_ERR(bdev
);
1127 if ((rc
= bd_claim(bdev
, log
))) {
1132 memcpy(log
->uuid
, sbi
->loguuid
, sizeof(log
->uuid
));
1137 if ((rc
= lmLogInit(log
)))
1140 list_add(&log
->journal_list
, &jfs_external_logs
);
1143 * add file system to log active file system list
1145 if ((rc
= lmLogFileSystem(log
, sbi
, 1)))
1150 list_add(&sbi
->log_list
, &log
->sb_list
);
1160 shutdown
: /* unwind lbmLogInit() */
1161 list_del(&log
->journal_list
);
1162 lbmLogShutdown(log
);
1167 close
: /* close external log device */
1170 free
: /* free log descriptor */
1174 jfs_warn("lmLogOpen: exit(%d)", rc
);
1178 static int open_inline_log(struct super_block
*sb
)
1180 struct jfs_log
*log
;
1183 if (!(log
= kzalloc(sizeof(struct jfs_log
), GFP_KERNEL
)))
1185 INIT_LIST_HEAD(&log
->sb_list
);
1186 init_waitqueue_head(&log
->syncwait
);
1188 set_bit(log_INLINELOG
, &log
->flag
);
1189 log
->bdev
= sb
->s_bdev
;
1190 log
->base
= addressPXD(&JFS_SBI(sb
)->logpxd
);
1191 log
->size
= lengthPXD(&JFS_SBI(sb
)->logpxd
) >>
1192 (L2LOGPSIZE
- sb
->s_blocksize_bits
);
1193 log
->l2bsize
= sb
->s_blocksize_bits
;
1194 ASSERT(L2LOGPSIZE
>= sb
->s_blocksize_bits
);
1199 if ((rc
= lmLogInit(log
))) {
1201 jfs_warn("lmLogOpen: exit(%d)", rc
);
1205 list_add(&JFS_SBI(sb
)->log_list
, &log
->sb_list
);
1206 JFS_SBI(sb
)->log
= log
;
1211 static int open_dummy_log(struct super_block
*sb
)
1217 dummy_log
= kzalloc(sizeof(struct jfs_log
), GFP_KERNEL
);
1222 INIT_LIST_HEAD(&dummy_log
->sb_list
);
1223 init_waitqueue_head(&dummy_log
->syncwait
);
1224 dummy_log
->no_integrity
= 1;
1225 /* Make up some stuff */
1226 dummy_log
->base
= 0;
1227 dummy_log
->size
= 1024;
1228 rc
= lmLogInit(dummy_log
);
1237 LOG_LOCK(dummy_log
);
1238 list_add(&JFS_SBI(sb
)->log_list
, &dummy_log
->sb_list
);
1239 JFS_SBI(sb
)->log
= dummy_log
;
1240 LOG_UNLOCK(dummy_log
);
1249 * FUNCTION: log initialization at first log open.
1251 * logredo() (or logformat()) should have been run previously.
1252 * initialize the log from log superblock.
1253 * set the log state in the superblock to LOGMOUNT and
1254 * write SYNCPT log record.
1256 * PARAMETER: log - log structure
1259 * -EINVAL - bad log magic number or superblock dirty
1260 * error returned from logwait()
1262 * serialization: single first open thread
1264 int lmLogInit(struct jfs_log
* log
)
1268 struct logsuper
*logsuper
;
1269 struct lbuf
*bpsuper
;
1274 jfs_info("lmLogInit: log:0x%p", log
);
1276 /* initialize the group commit serialization lock */
1277 LOGGC_LOCK_INIT(log
);
1279 /* allocate/initialize the log write serialization lock */
1282 LOGSYNC_LOCK_INIT(log
);
1284 INIT_LIST_HEAD(&log
->synclist
);
1286 INIT_LIST_HEAD(&log
->cqueue
);
1287 log
->flush_tblk
= NULL
;
1292 * initialize log i/o
1294 if ((rc
= lbmLogInit(log
)))
1297 if (!test_bit(log_INLINELOG
, &log
->flag
))
1298 log
->l2bsize
= L2LOGPSIZE
;
1300 /* check for disabled journaling to disk */
1301 if (log
->no_integrity
) {
1303 * Journal pages will still be filled. When the time comes
1304 * to actually do the I/O, the write is not done, and the
1305 * endio routine is called directly.
1307 bp
= lbmAllocate(log
, 0);
1309 bp
->l_pn
= bp
->l_eor
= 0;
1312 * validate log superblock
1314 if ((rc
= lbmRead(log
, 1, &bpsuper
)))
1317 logsuper
= (struct logsuper
*) bpsuper
->l_ldata
;
1319 if (logsuper
->magic
!= cpu_to_le32(LOGMAGIC
)) {
1320 jfs_warn("*** Log Format Error ! ***");
1325 /* logredo() should have been run successfully. */
1326 if (logsuper
->state
!= cpu_to_le32(LOGREDONE
)) {
1327 jfs_warn("*** Log Is Dirty ! ***");
1332 /* initialize log from log superblock */
1333 if (test_bit(log_INLINELOG
,&log
->flag
)) {
1334 if (log
->size
!= le32_to_cpu(logsuper
->size
)) {
1338 jfs_info("lmLogInit: inline log:0x%p base:0x%Lx "
1340 (unsigned long long) log
->base
, log
->size
);
1342 if (memcmp(logsuper
->uuid
, log
->uuid
, 16)) {
1343 jfs_warn("wrong uuid on JFS log device");
1346 log
->size
= le32_to_cpu(logsuper
->size
);
1347 log
->l2bsize
= le32_to_cpu(logsuper
->l2bsize
);
1348 jfs_info("lmLogInit: external log:0x%p base:0x%Lx "
1350 (unsigned long long) log
->base
, log
->size
);
1353 log
->page
= le32_to_cpu(logsuper
->end
) / LOGPSIZE
;
1354 log
->eor
= le32_to_cpu(logsuper
->end
) - (LOGPSIZE
* log
->page
);
1357 * initialize for log append write mode
1359 /* establish current/end-of-log page/buffer */
1360 if ((rc
= lbmRead(log
, log
->page
, &bp
)))
1363 lp
= (struct logpage
*) bp
->l_ldata
;
1365 jfs_info("lmLogInit: lsn:0x%x page:%d eor:%d:%d",
1366 le32_to_cpu(logsuper
->end
), log
->page
, log
->eor
,
1367 le16_to_cpu(lp
->h
.eor
));
1370 bp
->l_pn
= log
->page
;
1371 bp
->l_eor
= log
->eor
;
1373 /* if current page is full, move on to next page */
1374 if (log
->eor
>= LOGPSIZE
- LOGPTLRSIZE
)
1378 * initialize log syncpoint
1381 * write the first SYNCPT record with syncpoint = 0
1382 * (i.e., log redo up to HERE !);
1383 * remove current page from lbm write queue at end of pageout
1384 * (to write log superblock update), but do not release to
1389 lrd
.type
= cpu_to_le16(LOG_SYNCPT
);
1391 lrd
.log
.syncpt
.sync
= 0;
1392 lsn
= lmWriteRecord(log
, NULL
, &lrd
, NULL
);
1394 bp
->l_ceor
= bp
->l_eor
;
1395 lp
= (struct logpage
*) bp
->l_ldata
;
1396 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(bp
->l_eor
);
1397 lbmWrite(log
, bp
, lbmWRITE
| lbmSYNC
, 0);
1398 if ((rc
= lbmIOWait(bp
, 0)))
1402 * update/write superblock
1404 logsuper
->state
= cpu_to_le32(LOGMOUNT
);
1405 log
->serial
= le32_to_cpu(logsuper
->serial
) + 1;
1406 logsuper
->serial
= cpu_to_le32(log
->serial
);
1407 lbmDirectWrite(log
, bpsuper
, lbmWRITE
| lbmRELEASE
| lbmSYNC
);
1408 if ((rc
= lbmIOWait(bpsuper
, lbmFREE
)))
1412 /* initialize logsync parameters */
1413 log
->logsize
= (log
->size
- 2) << L2LOGPSIZE
;
1416 log
->sync
= log
->syncpt
;
1417 log
->nextsync
= LOGSYNC_DELTA(log
->logsize
);
1419 jfs_info("lmLogInit: lsn:0x%x syncpt:0x%x sync:0x%x",
1420 log
->lsn
, log
->syncpt
, log
->sync
);
1423 * initialize for lazy/group commit
1432 errout30
: /* release log page */
1434 bp
->l_wqnext
= NULL
;
1437 errout20
: /* release log superblock */
1440 errout10
: /* unwind lbmLogInit() */
1441 lbmLogShutdown(log
);
1443 jfs_warn("lmLogInit: exit(%d)", rc
);
1449 * NAME: lmLogClose()
1451 * FUNCTION: remove file system <ipmnt> from active list of log <iplog>
1452 * and close it on last close.
1454 * PARAMETER: sb - superblock
1456 * RETURN: errors from subroutines
1460 int lmLogClose(struct super_block
*sb
)
1462 struct jfs_sb_info
*sbi
= JFS_SBI(sb
);
1463 struct jfs_log
*log
= sbi
->log
;
1464 struct block_device
*bdev
;
1467 jfs_info("lmLogClose: log:0x%p", log
);
1471 list_del(&sbi
->log_list
);
1476 * We need to make sure all of the "written" metapages
1477 * actually make it to disk
1479 sync_blockdev(sb
->s_bdev
);
1481 if (test_bit(log_INLINELOG
, &log
->flag
)) {
1483 * in-line log in host file system
1485 rc
= lmLogShutdown(log
);
1490 if (!log
->no_integrity
)
1491 lmLogFileSystem(log
, sbi
, 0);
1493 if (!list_empty(&log
->sb_list
))
1497 * TODO: ensure that the dummy_log is in a state to allow
1498 * lbmLogShutdown to deallocate all the buffers and call
1499 * kfree against dummy_log. For now, leave dummy_log & its
1500 * buffers in memory, and resuse if another no-integrity mount
1503 if (log
->no_integrity
)
1507 * external log as separate logical volume
1509 list_del(&log
->journal_list
);
1511 rc
= lmLogShutdown(log
);
1520 jfs_info("lmLogClose: exit(%d)", rc
);
1526 * NAME: jfs_flush_journal()
1528 * FUNCTION: initiate write of any outstanding transactions to the journal
1529 * and optionally wait until they are all written to disk
1531 * wait == 0 flush until latest txn is committed, don't wait
1532 * wait == 1 flush until latest txn is committed, wait
1533 * wait > 1 flush until all txn's are complete, wait
1535 void jfs_flush_journal(struct jfs_log
*log
, int wait
)
1538 struct tblock
*target
= NULL
;
1539 struct jfs_sb_info
*sbi
;
1541 /* jfs_write_inode may call us during read-only mount */
1545 jfs_info("jfs_flush_journal: log:0x%p wait=%d", log
, wait
);
1549 if (!list_empty(&log
->cqueue
)) {
1551 * This ensures that we will keep writing to the journal as long
1552 * as there are unwritten commit records
1554 target
= list_entry(log
->cqueue
.prev
, struct tblock
, cqueue
);
1556 if (test_bit(log_FLUSH
, &log
->flag
)) {
1558 * We're already flushing.
1559 * if flush_tblk is NULL, we are flushing everything,
1560 * so leave it that way. Otherwise, update it to the
1561 * latest transaction
1563 if (log
->flush_tblk
)
1564 log
->flush_tblk
= target
;
1566 /* Only flush until latest transaction is committed */
1567 log
->flush_tblk
= target
;
1568 set_bit(log_FLUSH
, &log
->flag
);
1571 * Initiate I/O on outstanding transactions
1573 if (!(log
->cflag
& logGC_PAGEOUT
)) {
1574 log
->cflag
|= logGC_PAGEOUT
;
1579 if ((wait
> 1) || test_bit(log_SYNCBARRIER
, &log
->flag
)) {
1580 /* Flush until all activity complete */
1581 set_bit(log_FLUSH
, &log
->flag
);
1582 log
->flush_tblk
= NULL
;
1585 if (wait
&& target
&& !(target
->flag
& tblkGC_COMMITTED
)) {
1586 DECLARE_WAITQUEUE(__wait
, current
);
1588 add_wait_queue(&target
->gcwait
, &__wait
);
1589 set_current_state(TASK_UNINTERRUPTIBLE
);
1592 current
->state
= TASK_RUNNING
;
1594 remove_wait_queue(&target
->gcwait
, &__wait
);
1601 list_for_each_entry(sbi
, &log
->sb_list
, log_list
) {
1602 filemap_fdatawrite(sbi
->ipbmap
->i_mapping
);
1603 filemap_fdatawrite(sbi
->ipimap
->i_mapping
);
1604 filemap_fdatawrite(sbi
->direct_inode
->i_mapping
);
1608 * If there was recent activity, we may need to wait
1609 * for the lazycommit thread to catch up
1611 if ((!list_empty(&log
->cqueue
)) || !list_empty(&log
->synclist
)) {
1612 for (i
= 0; i
< 200; i
++) { /* Too much? */
1614 if (list_empty(&log
->cqueue
) &&
1615 list_empty(&log
->synclist
))
1619 assert(list_empty(&log
->cqueue
));
1621 #ifdef CONFIG_JFS_DEBUG
1622 if (!list_empty(&log
->synclist
)) {
1623 struct logsyncblk
*lp
;
1625 list_for_each_entry(lp
, &log
->synclist
, synclist
) {
1626 if (lp
->xflag
& COMMIT_PAGE
) {
1627 struct metapage
*mp
= (struct metapage
*)lp
;
1628 dump_mem("orphan metapage", lp
,
1629 sizeof(struct metapage
));
1630 dump_mem("page", mp
->page
, sizeof(struct page
));
1633 dump_mem("orphan tblock", lp
,
1634 sizeof(struct tblock
));
1638 //assert(list_empty(&log->synclist));
1639 clear_bit(log_FLUSH
, &log
->flag
);
1643 * NAME: lmLogShutdown()
1645 * FUNCTION: log shutdown at last LogClose().
1647 * write log syncpt record.
1648 * update super block to set redone flag to 0.
1650 * PARAMETER: log - log inode
1652 * RETURN: 0 - success
1654 * serialization: single last close thread
1656 int lmLogShutdown(struct jfs_log
* log
)
1661 struct logsuper
*logsuper
;
1662 struct lbuf
*bpsuper
;
1666 jfs_info("lmLogShutdown: log:0x%p", log
);
1668 jfs_flush_journal(log
, 2);
1671 * write the last SYNCPT record with syncpoint = 0
1672 * (i.e., log redo up to HERE !)
1676 lrd
.type
= cpu_to_le16(LOG_SYNCPT
);
1678 lrd
.log
.syncpt
.sync
= 0;
1680 lsn
= lmWriteRecord(log
, NULL
, &lrd
, NULL
);
1682 lp
= (struct logpage
*) bp
->l_ldata
;
1683 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(bp
->l_eor
);
1684 lbmWrite(log
, log
->bp
, lbmWRITE
| lbmRELEASE
| lbmSYNC
, 0);
1685 lbmIOWait(log
->bp
, lbmFREE
);
1689 * synchronous update log superblock
1690 * mark log state as shutdown cleanly
1691 * (i.e., Log does not need to be replayed).
1693 if ((rc
= lbmRead(log
, 1, &bpsuper
)))
1696 logsuper
= (struct logsuper
*) bpsuper
->l_ldata
;
1697 logsuper
->state
= cpu_to_le32(LOGREDONE
);
1698 logsuper
->end
= cpu_to_le32(lsn
);
1699 lbmDirectWrite(log
, bpsuper
, lbmWRITE
| lbmRELEASE
| lbmSYNC
);
1700 rc
= lbmIOWait(bpsuper
, lbmFREE
);
1702 jfs_info("lmLogShutdown: lsn:0x%x page:%d eor:%d",
1703 lsn
, log
->page
, log
->eor
);
1707 * shutdown per log i/o
1709 lbmLogShutdown(log
);
1712 jfs_warn("lmLogShutdown: exit(%d)", rc
);
1719 * NAME: lmLogFileSystem()
1721 * FUNCTION: insert (<activate> = true)/remove (<activate> = false)
1722 * file system into/from log active file system list.
1724 * PARAMETE: log - pointer to logs inode.
1725 * fsdev - kdev_t of filesystem.
1726 * serial - pointer to returned log serial number
1727 * activate - insert/remove device from active list.
1729 * RETURN: 0 - success
1730 * errors returned by vms_iowait().
1732 static int lmLogFileSystem(struct jfs_log
* log
, struct jfs_sb_info
*sbi
,
1737 struct logsuper
*logsuper
;
1738 struct lbuf
*bpsuper
;
1739 char *uuid
= sbi
->uuid
;
1742 * insert/remove file system device to log active file system list.
1744 if ((rc
= lbmRead(log
, 1, &bpsuper
)))
1747 logsuper
= (struct logsuper
*) bpsuper
->l_ldata
;
1749 for (i
= 0; i
< MAX_ACTIVE
; i
++)
1750 if (!memcmp(logsuper
->active
[i
].uuid
, NULL_UUID
, 16)) {
1751 memcpy(logsuper
->active
[i
].uuid
, uuid
, 16);
1755 if (i
== MAX_ACTIVE
) {
1756 jfs_warn("Too many file systems sharing journal!");
1758 return -EMFILE
; /* Is there a better rc? */
1761 for (i
= 0; i
< MAX_ACTIVE
; i
++)
1762 if (!memcmp(logsuper
->active
[i
].uuid
, uuid
, 16)) {
1763 memcpy(logsuper
->active
[i
].uuid
, NULL_UUID
, 16);
1766 if (i
== MAX_ACTIVE
) {
1767 jfs_warn("Somebody stomped on the journal!");
1775 * synchronous write log superblock:
1777 * write sidestream bypassing write queue:
1778 * at file system mount, log super block is updated for
1779 * activation of the file system before any log record
1780 * (MOUNT record) of the file system, and at file system
1781 * unmount, all meta data for the file system has been
1782 * flushed before log super block is updated for deactivation
1783 * of the file system.
1785 lbmDirectWrite(log
, bpsuper
, lbmWRITE
| lbmRELEASE
| lbmSYNC
);
1786 rc
= lbmIOWait(bpsuper
, lbmFREE
);
1792 * log buffer manager (lbm)
1793 * ------------------------
1795 * special purpose buffer manager supporting log i/o requirements.
1797 * per log write queue:
1798 * log pageout occurs in serial order by fifo write queue and
1799 * restricting to a single i/o in pregress at any one time.
1800 * a circular singly-linked list
1801 * (log->wrqueue points to the tail, and buffers are linked via
1802 * bp->wrqueue field), and
1803 * maintains log page in pageout ot waiting for pageout in serial pageout.
1809 * initialize per log I/O setup at lmLogInit()
1811 static int lbmLogInit(struct jfs_log
* log
)
1816 jfs_info("lbmLogInit: log:0x%p", log
);
1818 /* initialize current buffer cursor */
1821 /* initialize log device write queue */
1825 * Each log has its own buffer pages allocated to it. These are
1826 * not managed by the page cache. This ensures that a transaction
1827 * writing to the log does not block trying to allocate a page from
1828 * the page cache (for the log). This would be bad, since page
1829 * allocation waits on the kswapd thread that may be committing inodes
1830 * which would cause log activity. Was that clear? I'm trying to
1831 * avoid deadlock here.
1833 init_waitqueue_head(&log
->free_wait
);
1835 log
->lbuf_free
= NULL
;
1837 for (i
= 0; i
< LOGPAGES
;) {
1842 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
1845 page
= virt_to_page(buffer
);
1846 for (offset
= 0; offset
< PAGE_SIZE
; offset
+= LOGPSIZE
) {
1847 lbuf
= kmalloc(sizeof(struct lbuf
), GFP_KERNEL
);
1850 free_page((unsigned long) buffer
);
1853 if (offset
) /* we already have one reference */
1855 lbuf
->l_offset
= offset
;
1856 lbuf
->l_ldata
= buffer
+ offset
;
1857 lbuf
->l_page
= page
;
1859 init_waitqueue_head(&lbuf
->l_ioevent
);
1861 lbuf
->l_freelist
= log
->lbuf_free
;
1862 log
->lbuf_free
= lbuf
;
1870 lbmLogShutdown(log
);
1878 * finalize per log I/O setup at lmLogShutdown()
1880 static void lbmLogShutdown(struct jfs_log
* log
)
1884 jfs_info("lbmLogShutdown: log:0x%p", log
);
1886 lbuf
= log
->lbuf_free
;
1888 struct lbuf
*next
= lbuf
->l_freelist
;
1889 __free_page(lbuf
->l_page
);
1899 * allocate an empty log buffer
1901 static struct lbuf
*lbmAllocate(struct jfs_log
* log
, int pn
)
1904 unsigned long flags
;
1907 * recycle from log buffer freelist if any
1910 LCACHE_SLEEP_COND(log
->free_wait
, (bp
= log
->lbuf_free
), flags
);
1911 log
->lbuf_free
= bp
->l_freelist
;
1912 LCACHE_UNLOCK(flags
);
1916 bp
->l_wqnext
= NULL
;
1917 bp
->l_freelist
= NULL
;
1920 bp
->l_blkno
= log
->base
+ (pn
<< (L2LOGPSIZE
- log
->l2bsize
));
1930 * release a log buffer to freelist
1932 static void lbmFree(struct lbuf
* bp
)
1934 unsigned long flags
;
1940 LCACHE_UNLOCK(flags
);
1943 static void lbmfree(struct lbuf
* bp
)
1945 struct jfs_log
*log
= bp
->l_log
;
1947 assert(bp
->l_wqnext
== NULL
);
1950 * return the buffer to head of freelist
1952 bp
->l_freelist
= log
->lbuf_free
;
1953 log
->lbuf_free
= bp
;
1955 wake_up(&log
->free_wait
);
1963 * FUNCTION: add a log buffer to the the log redrive list
1969 * Takes log_redrive_lock.
1971 static inline void lbmRedrive(struct lbuf
*bp
)
1973 unsigned long flags
;
1975 spin_lock_irqsave(&log_redrive_lock
, flags
);
1976 bp
->l_redrive_next
= log_redrive_list
;
1977 log_redrive_list
= bp
;
1978 spin_unlock_irqrestore(&log_redrive_lock
, flags
);
1980 wake_up_process(jfsIOthread
);
1987 static int lbmRead(struct jfs_log
* log
, int pn
, struct lbuf
** bpp
)
1993 * allocate a log buffer
1995 *bpp
= bp
= lbmAllocate(log
, pn
);
1996 jfs_info("lbmRead: bp:0x%p pn:0x%x", bp
, pn
);
1998 bp
->l_flag
|= lbmREAD
;
2000 bio
= bio_alloc(GFP_NOFS
, 1);
2002 bio
->bi_sector
= bp
->l_blkno
<< (log
->l2bsize
- 9);
2003 bio
->bi_bdev
= log
->bdev
;
2004 bio
->bi_io_vec
[0].bv_page
= bp
->l_page
;
2005 bio
->bi_io_vec
[0].bv_len
= LOGPSIZE
;
2006 bio
->bi_io_vec
[0].bv_offset
= bp
->l_offset
;
2010 bio
->bi_size
= LOGPSIZE
;
2012 bio
->bi_end_io
= lbmIODone
;
2013 bio
->bi_private
= bp
;
2014 submit_bio(READ_SYNC
, bio
);
2016 wait_event(bp
->l_ioevent
, (bp
->l_flag
!= lbmREAD
));
2025 * buffer at head of pageout queue stays after completion of
2026 * partial-page pageout and redriven by explicit initiation of
2027 * pageout by caller until full-page pageout is completed and
2030 * device driver i/o done redrives pageout of new buffer at
2031 * head of pageout queue when current buffer at head of pageout
2032 * queue is released at the completion of its full-page pageout.
2034 * LOGGC_LOCK() serializes lbmWrite() by lmNextPage() and lmGroupCommit().
2035 * LCACHE_LOCK() serializes xflag between lbmWrite() and lbmIODone()
2037 static void lbmWrite(struct jfs_log
* log
, struct lbuf
* bp
, int flag
,
2041 unsigned long flags
;
2043 jfs_info("lbmWrite: bp:0x%p flag:0x%x pn:0x%x", bp
, flag
, bp
->l_pn
);
2045 /* map the logical block address to physical block address */
2047 log
->base
+ (bp
->l_pn
<< (L2LOGPSIZE
- log
->l2bsize
));
2049 LCACHE_LOCK(flags
); /* disable+lock */
2052 * initialize buffer for device driver
2057 * insert bp at tail of write queue associated with log
2059 * (request is either for bp already/currently at head of queue
2060 * or new bp to be inserted at tail)
2064 /* is buffer not already on write queue ? */
2065 if (bp
->l_wqnext
== NULL
) {
2066 /* insert at tail of wqueue */
2072 bp
->l_wqnext
= tail
->l_wqnext
;
2073 tail
->l_wqnext
= bp
;
2079 /* is buffer at head of wqueue and for write ? */
2080 if ((bp
!= tail
->l_wqnext
) || !(flag
& lbmWRITE
)) {
2081 LCACHE_UNLOCK(flags
); /* unlock+enable */
2085 LCACHE_UNLOCK(flags
); /* unlock+enable */
2089 else if (flag
& lbmSYNC
)
2102 * initiate pageout bypassing write queue for sidestream
2103 * (e.g., log superblock) write;
2105 static void lbmDirectWrite(struct jfs_log
* log
, struct lbuf
* bp
, int flag
)
2107 jfs_info("lbmDirectWrite: bp:0x%p flag:0x%x pn:0x%x",
2108 bp
, flag
, bp
->l_pn
);
2111 * initialize buffer for device driver
2113 bp
->l_flag
= flag
| lbmDIRECT
;
2115 /* map the logical block address to physical block address */
2117 log
->base
+ (bp
->l_pn
<< (L2LOGPSIZE
- log
->l2bsize
));
2120 * initiate pageout of the page
2127 * NAME: lbmStartIO()
2129 * FUNCTION: Interface to DD strategy routine
2133 * serialization: LCACHE_LOCK() is NOT held during log i/o;
2135 static void lbmStartIO(struct lbuf
* bp
)
2138 struct jfs_log
*log
= bp
->l_log
;
2140 jfs_info("lbmStartIO\n");
2142 bio
= bio_alloc(GFP_NOFS
, 1);
2143 bio
->bi_sector
= bp
->l_blkno
<< (log
->l2bsize
- 9);
2144 bio
->bi_bdev
= log
->bdev
;
2145 bio
->bi_io_vec
[0].bv_page
= bp
->l_page
;
2146 bio
->bi_io_vec
[0].bv_len
= LOGPSIZE
;
2147 bio
->bi_io_vec
[0].bv_offset
= bp
->l_offset
;
2151 bio
->bi_size
= LOGPSIZE
;
2153 bio
->bi_end_io
= lbmIODone
;
2154 bio
->bi_private
= bp
;
2156 /* check if journaling to disk has been disabled */
2157 if (log
->no_integrity
) {
2159 lbmIODone(bio
, 0, 0);
2161 submit_bio(WRITE_SYNC
, bio
);
2162 INCREMENT(lmStat
.submitted
);
2170 static int lbmIOWait(struct lbuf
* bp
, int flag
)
2172 unsigned long flags
;
2175 jfs_info("lbmIOWait1: bp:0x%p flag:0x%x:0x%x", bp
, bp
->l_flag
, flag
);
2177 LCACHE_LOCK(flags
); /* disable+lock */
2179 LCACHE_SLEEP_COND(bp
->l_ioevent
, (bp
->l_flag
& lbmDONE
), flags
);
2181 rc
= (bp
->l_flag
& lbmERROR
) ? -EIO
: 0;
2186 LCACHE_UNLOCK(flags
); /* unlock+enable */
2188 jfs_info("lbmIOWait2: bp:0x%p flag:0x%x:0x%x", bp
, bp
->l_flag
, flag
);
2195 * executed at INTIODONE level
2197 static int lbmIODone(struct bio
*bio
, unsigned int bytes_done
, int error
)
2199 struct lbuf
*bp
= bio
->bi_private
;
2200 struct lbuf
*nextbp
, *tail
;
2201 struct jfs_log
*log
;
2202 unsigned long flags
;
2208 * get back jfs buffer bound to the i/o buffer
2210 jfs_info("lbmIODone: bp:0x%p flag:0x%x", bp
, bp
->l_flag
);
2212 LCACHE_LOCK(flags
); /* disable+lock */
2214 bp
->l_flag
|= lbmDONE
;
2216 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
)) {
2217 bp
->l_flag
|= lbmERROR
;
2219 jfs_err("lbmIODone: I/O error in JFS log");
2227 if (bp
->l_flag
& lbmREAD
) {
2228 bp
->l_flag
&= ~lbmREAD
;
2230 LCACHE_UNLOCK(flags
); /* unlock+enable */
2232 /* wakeup I/O initiator */
2233 LCACHE_WAKEUP(&bp
->l_ioevent
);
2239 * pageout completion
2241 * the bp at the head of write queue has completed pageout.
2243 * if single-commit/full-page pageout, remove the current buffer
2244 * from head of pageout queue, and redrive pageout with
2245 * the new buffer at head of pageout queue;
2246 * otherwise, the partial-page pageout buffer stays at
2247 * the head of pageout queue to be redriven for pageout
2248 * by lmGroupCommit() until full-page pageout is completed.
2250 bp
->l_flag
&= ~lbmWRITE
;
2251 INCREMENT(lmStat
.pagedone
);
2253 /* update committed lsn */
2255 log
->clsn
= (bp
->l_pn
<< L2LOGPSIZE
) + bp
->l_ceor
;
2257 if (bp
->l_flag
& lbmDIRECT
) {
2258 LCACHE_WAKEUP(&bp
->l_ioevent
);
2259 LCACHE_UNLOCK(flags
);
2265 /* single element queue */
2267 /* remove head buffer of full-page pageout
2268 * from log device write queue
2270 if (bp
->l_flag
& lbmRELEASE
) {
2272 bp
->l_wqnext
= NULL
;
2275 /* multi element queue */
2277 /* remove head buffer of full-page pageout
2278 * from log device write queue
2280 if (bp
->l_flag
& lbmRELEASE
) {
2281 nextbp
= tail
->l_wqnext
= bp
->l_wqnext
;
2282 bp
->l_wqnext
= NULL
;
2285 * redrive pageout of next page at head of write queue:
2286 * redrive next page without any bound tblk
2287 * (i.e., page w/o any COMMIT records), or
2288 * first page of new group commit which has been
2289 * queued after current page (subsequent pageout
2290 * is performed synchronously, except page without
2291 * any COMMITs) by lmGroupCommit() as indicated
2294 if (nextbp
->l_flag
& lbmWRITE
) {
2296 * We can't do the I/O at interrupt time.
2297 * The jfsIO thread can do it
2305 * synchronous pageout:
2307 * buffer has not necessarily been removed from write queue
2308 * (e.g., synchronous write of partial-page with COMMIT):
2309 * leave buffer for i/o initiator to dispose
2311 if (bp
->l_flag
& lbmSYNC
) {
2312 LCACHE_UNLOCK(flags
); /* unlock+enable */
2314 /* wakeup I/O initiator */
2315 LCACHE_WAKEUP(&bp
->l_ioevent
);
2319 * Group Commit pageout:
2321 else if (bp
->l_flag
& lbmGC
) {
2322 LCACHE_UNLOCK(flags
);
2327 * asynchronous pageout:
2329 * buffer must have been removed from write queue:
2330 * insert buffer at head of freelist where it can be recycled
2333 assert(bp
->l_flag
& lbmRELEASE
);
2334 assert(bp
->l_flag
& lbmFREE
);
2337 LCACHE_UNLOCK(flags
); /* unlock+enable */
2343 int jfsIOWait(void *arg
)
2348 spin_lock_irq(&log_redrive_lock
);
2349 while ((bp
= log_redrive_list
) != 0) {
2350 log_redrive_list
= bp
->l_redrive_next
;
2351 bp
->l_redrive_next
= NULL
;
2352 spin_unlock_irq(&log_redrive_lock
);
2354 spin_lock_irq(&log_redrive_lock
);
2356 spin_unlock_irq(&log_redrive_lock
);
2358 if (freezing(current
)) {
2361 set_current_state(TASK_INTERRUPTIBLE
);
2363 current
->state
= TASK_RUNNING
;
2365 } while (!kthread_should_stop());
2367 jfs_info("jfsIOWait being killed!");
2372 * NAME: lmLogFormat()/jfs_logform()
2374 * FUNCTION: format file system log
2378 * logAddress - start address of log space in FS block
2379 * logSize - length of log space in FS block;
2381 * RETURN: 0 - success
2384 * XXX: We're synchronously writing one page at a time. This needs to
2385 * be improved by writing multiple pages at once.
2387 int lmLogFormat(struct jfs_log
*log
, s64 logAddress
, int logSize
)
2390 struct jfs_sb_info
*sbi
;
2391 struct logsuper
*logsuper
;
2393 int lspn
; /* log sequence page number */
2394 struct lrd
*lrd_ptr
;
2398 jfs_info("lmLogFormat: logAddress:%Ld logSize:%d",
2399 (long long)logAddress
, logSize
);
2401 sbi
= list_entry(log
->sb_list
.next
, struct jfs_sb_info
, log_list
);
2403 /* allocate a log buffer */
2404 bp
= lbmAllocate(log
, 1);
2406 npages
= logSize
>> sbi
->l2nbperpage
;
2411 * page 0 - reserved;
2412 * page 1 - log superblock;
2413 * page 2 - log data page: A SYNC log record is written
2414 * into this page at logform time;
2415 * pages 3-N - log data page: set to empty log data pages;
2418 * init log superblock: log page 1
2420 logsuper
= (struct logsuper
*) bp
->l_ldata
;
2422 logsuper
->magic
= cpu_to_le32(LOGMAGIC
);
2423 logsuper
->version
= cpu_to_le32(LOGVERSION
);
2424 logsuper
->state
= cpu_to_le32(LOGREDONE
);
2425 logsuper
->flag
= cpu_to_le32(sbi
->mntflag
); /* ? */
2426 logsuper
->size
= cpu_to_le32(npages
);
2427 logsuper
->bsize
= cpu_to_le32(sbi
->bsize
);
2428 logsuper
->l2bsize
= cpu_to_le32(sbi
->l2bsize
);
2429 logsuper
->end
= cpu_to_le32(2 * LOGPSIZE
+ LOGPHDRSIZE
+ LOGRDSIZE
);
2431 bp
->l_flag
= lbmWRITE
| lbmSYNC
| lbmDIRECT
;
2432 bp
->l_blkno
= logAddress
+ sbi
->nbperpage
;
2434 if ((rc
= lbmIOWait(bp
, 0)))
2438 * init pages 2 to npages-1 as log data pages:
2440 * log page sequence number (lpsn) initialization:
2443 * +-----+-----+=====+=====+===.....===+=====+
2445 * <--- N page circular file ---->
2447 * the N (= npages-2) data pages of the log is maintained as
2448 * a circular file for the log records;
2449 * lpsn grows by 1 monotonically as each log page is written
2450 * to the circular file of the log;
2451 * and setLogpage() will not reset the page number even if
2452 * the eor is equal to LOGPHDRSIZE. In order for binary search
2453 * still work in find log end process, we have to simulate the
2454 * log wrap situation at the log format time.
2455 * The 1st log page written will have the highest lpsn. Then
2456 * the succeeding log pages will have ascending order of
2457 * the lspn starting from 0, ... (N-2)
2459 lp
= (struct logpage
*) bp
->l_ldata
;
2461 * initialize 1st log page to be written: lpsn = N - 1,
2462 * write a SYNCPT log record is written to this page
2464 lp
->h
.page
= lp
->t
.page
= cpu_to_le32(npages
- 3);
2465 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(LOGPHDRSIZE
+ LOGRDSIZE
);
2467 lrd_ptr
= (struct lrd
*) &lp
->data
;
2468 lrd_ptr
->logtid
= 0;
2469 lrd_ptr
->backchain
= 0;
2470 lrd_ptr
->type
= cpu_to_le16(LOG_SYNCPT
);
2471 lrd_ptr
->length
= 0;
2472 lrd_ptr
->log
.syncpt
.sync
= 0;
2474 bp
->l_blkno
+= sbi
->nbperpage
;
2475 bp
->l_flag
= lbmWRITE
| lbmSYNC
| lbmDIRECT
;
2477 if ((rc
= lbmIOWait(bp
, 0)))
2481 * initialize succeeding log pages: lpsn = 0, 1, ..., (N-2)
2483 for (lspn
= 0; lspn
< npages
- 3; lspn
++) {
2484 lp
->h
.page
= lp
->t
.page
= cpu_to_le32(lspn
);
2485 lp
->h
.eor
= lp
->t
.eor
= cpu_to_le16(LOGPHDRSIZE
);
2487 bp
->l_blkno
+= sbi
->nbperpage
;
2488 bp
->l_flag
= lbmWRITE
| lbmSYNC
| lbmDIRECT
;
2490 if ((rc
= lbmIOWait(bp
, 0)))
2499 /* release the buffer */
2505 #ifdef CONFIG_JFS_STATISTICS
2506 int jfs_lmstats_read(char *buffer
, char **start
, off_t offset
, int length
,
2507 int *eof
, void *data
)
2512 len
+= sprintf(buffer
,
2513 "JFS Logmgr stats\n"
2514 "================\n"
2516 "writes submitted = %d\n"
2517 "writes completed = %d\n"
2518 "full pages submitted = %d\n"
2519 "partial pages submitted = %d\n",
2524 lmStat
.partial_page
);
2527 *start
= buffer
+ begin
;
2540 #endif /* CONFIG_JFS_STATISTICS */