jfs_logmgr.c

jfs-2.4-1.1.7.tar.gz jfs 2.4－1.1.7 源码

/*
 *   Copyright (c) International Business Machines Corp., 2000-2003
 *   Portions Copyright (c) Christoph Hellwig, 2001-2002
 *
 *   This program is free software;  you can redistribute it and/or modify
 *   it under the terms of the GNU General Public License as published by
 *   the Free Software Foundation; either version 2 of the License, or 
 *   (at your option) any later version.
 * 
 *   This program is distributed in the hope that it will be useful,
 *   but WITHOUT ANY WARRANTY;  without even the implied warranty of
 *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See
 *   the GNU General Public License for more details.
 *
 *   You should have received a copy of the GNU General Public License
 *   along with this program;  if not, write to the Free Software 
 *   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
 */

/*
 *	jfs_logmgr.c: log manager
 *
 * for related information, see transaction manager (jfs_txnmgr.c), and
 * recovery manager (jfs_logredo.c).
 *
 * note: for detail, RTFS.
 *
 *	log buffer manager:
 * special purpose buffer manager supporting log i/o requirements.
 * per log serial pageout of logpage
 * queuing i/o requests and redrive i/o at iodone
 * maintain current logpage buffer
 * no caching since append only
 * appropriate jfs buffer cache buffers as needed
 *
 *	group commit:
 * transactions which wrote COMMIT records in the same in-memory
 * log page during the pageout of previous/current log page(s) are
 * committed together by the pageout of the page.
 *
 *	TBD lazy commit:
 * transactions are committed asynchronously when the log page
 * containing it COMMIT is paged out when it becomes full;
 *
 *	serialization:
 * . a per log lock serialize log write.
 * . a per log lock serialize group commit.
 * . a per log lock serialize log open/close;
 *
 *	TBD log integrity:
 * careful-write (ping-pong) of last logpage to recover from crash
 * in overwrite.
 * detection of split (out-of-order) write of physical sectors
 * of last logpage via timestamp at end of each sector
 * with its mirror data array at trailer).
 *
 *	alternatives:
 * lsn - 64-bit monotonically increasing integer vs
 * 32-bit lspn and page eor.
 */

#include <linux/fs.h>
#include <linux/locks.h>
#include <linux/blkdev.h>
#include <linux/interrupt.h>
#include <linux/smp_lock.h>
#include <linux/completion.h>
#include "jfs_incore.h"
#include "jfs_filsys.h"
#include "jfs_metapage.h"
#include "jfs_txnmgr.h"
#include "jfs_debug.h"


/*
 * lbuf's ready to be redriven.  Protected by log_redrive_lock (jfsIO thread)
 */
static struct lbuf *log_redrive_list;
static spinlock_t log_redrive_lock = SPIN_LOCK_UNLOCKED;
DECLARE_WAIT_QUEUE_HEAD(jfs_IO_thread_wait);


/*
 *	log read/write serialization (per log)
 */
#define LOG_LOCK_INIT(log)	init_MUTEX(&(log)->loglock)
#define LOG_LOCK(log)		down(&((log)->loglock))
#define LOG_UNLOCK(log)		up(&((log)->loglock))


/*
 *	log group commit serialization (per log)
 */

#define LOGGC_LOCK_INIT(log)	spin_lock_init(&(log)->gclock)
#define LOGGC_LOCK(log)		spin_lock_irq(&(log)->gclock)
#define LOGGC_UNLOCK(log)	spin_unlock_irq(&(log)->gclock)
#define LOGGC_WAKEUP(tblk)	wake_up_all(&(tblk)->gcwait)

/*
 *	log sync serialization (per log)
 */
#define	LOGSYNC_DELTA(logsize)		min((logsize)/8, 128*LOGPSIZE)
#define	LOGSYNC_BARRIER(logsize)	((logsize)/4)
/*
#define	LOGSYNC_DELTA(logsize)		min((logsize)/4, 256*LOGPSIZE)
#define	LOGSYNC_BARRIER(logsize)	((logsize)/2)
*/


/*
 *	log buffer cache synchronization
 */
static spinlock_t jfsLCacheLock = SPIN_LOCK_UNLOCKED;

#define	LCACHE_LOCK(flags)	spin_lock_irqsave(&jfsLCacheLock, flags)
#define	LCACHE_UNLOCK(flags)	spin_unlock_irqrestore(&jfsLCacheLock, flags)

/*
 * See __SLEEP_COND in jfs_locks.h
 */
#define LCACHE_SLEEP_COND(wq, cond, flags)	\
do {						\
	if (cond)				\
		break;				\
	__SLEEP_COND(wq, cond, LCACHE_LOCK(flags), LCACHE_UNLOCK(flags)); \
} while (0)

#define	LCACHE_WAKEUP(event)	wake_up(event)


/*
 *	lbuf buffer cache (lCache) control
 */
/* log buffer manager pageout control (cumulative, inclusive) */
#define	lbmREAD		0x0001
#define	lbmWRITE	0x0002	/* enqueue at tail of write queue;
				 * init pageout if at head of queue;
				 */
#define	lbmRELEASE	0x0004	/* remove from write queue
				 * at completion of pageout;
				 * do not free/recycle it yet:
				 * caller will free it;
				 */
#define	lbmSYNC		0x0008	/* do not return to freelist
				 * when removed from write queue;
				 */
#define lbmFREE		0x0010	/* return to freelist
				 * at completion of pageout;
				 * the buffer may be recycled;
				 */
#define	lbmDONE		0x0020
#define	lbmERROR	0x0040
#define lbmGC		0x0080	/* lbmIODone to perform post-GC processing
				 * of log page
				 */
#define lbmDIRECT	0x0100

/*
 * external references
 */
extern void txLazyUnlock(struct tblock * tblk);
extern int jfs_stop_threads;
extern struct completion jfsIOwait;

/*
 * forward references
 */
static int lmWriteRecord(struct jfs_log * log, struct tblock * tblk,
			 struct lrd * lrd, struct tlock * tlck);

static int lmNextPage(struct jfs_log * log);
static int lmLogFileSystem(struct jfs_log * log, char *uuid, int activate);

static int lbmLogInit(struct jfs_log * log);
static void lbmLogShutdown(struct jfs_log * log);
static struct lbuf *lbmAllocate(struct jfs_log * log, int);
static void lbmFree(struct lbuf * bp);
static void lbmfree(struct lbuf * bp);
static int lbmRead(struct jfs_log * log, int pn, struct lbuf ** bpp);
static void lbmWrite(struct jfs_log * log, struct lbuf * bp, int flag,
		     int cant_block);
static void lbmDirectWrite(struct jfs_log * log, struct lbuf * bp, int flag);
static int lbmIOWait(struct lbuf * bp, int flag);
static void lbmIODone(struct buffer_head *bh, int);
static void lbmStartIO(struct lbuf * bp);
static void lmGCwrite(struct jfs_log * log, int cant_block);
static int lmLogSync(struct jfs_log * log, int nosyncwait);


/*
 *	statistics
 */
#ifdef CONFIG_JFS_STATISTICS
struct lmStat {
	uint commit;		/* # of commit */
	uint pagedone;		/* # of page written */
	uint submitted;		/* # of pages submitted */
	uint full_page;		/* # of full pages submitted */
	uint partial_page;	/* # of partial pages submitted */
} lmStat;
#endif


/*
 * NAME:	lmLog()
 *
 * FUNCTION:	write a log record;
 *
 * PARAMETER:
 *
 * RETURN:	lsn - offset to the next log record to write (end-of-log);
 *		-1  - error;
 *
 * note: todo: log error handler
 */
int lmLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
	  struct tlock * tlck)
{
	int lsn;
	int diffp, difft;
	struct metapage *mp = NULL;

	jfs_info("lmLog: log:0x%p tblk:0x%p, lrd:0x%p tlck:0x%p",
		 log, tblk, lrd, tlck);

	LOG_LOCK(log);

	/* log by (out-of-transaction) JFS ? */
	if (tblk == NULL)
		goto writeRecord;

	/* log from page ? */
	if (tlck == NULL ||
	    tlck->type & tlckBTROOT || (mp = tlck->mp) == NULL)
		goto writeRecord;

	/*
	 *      initialize/update page/transaction recovery lsn
	 */
	lsn = log->lsn;

	LOGSYNC_LOCK(log);

	/*
	 * initialize page lsn if first log write of the page
	 */
	if (mp->lsn == 0) {
		mp->log = log;
		mp->lsn = lsn;
		log->count++;

		/* insert page at tail of logsynclist */
		list_add_tail(&mp->synclist, &log->synclist);
	}

	/*
	 *      initialize/update lsn of tblock of the page
	 *
	 * transaction inherits oldest lsn of pages associated
	 * with allocation/deallocation of resources (their
	 * log records are used to reconstruct allocation map
	 * at recovery time: inode for inode allocation map,
	 * B+-tree index of extent descriptors for block
	 * allocation map);
	 * allocation map pages inherit transaction lsn at
	 * commit time to allow forwarding log syncpt past log
	 * records associated with allocation/deallocation of
	 * resources only after persistent map of these map pages
	 * have been updated and propagated to home.
	 */
	/*
	 * initialize transaction lsn:
	 */
	if (tblk->lsn == 0) {
		/* inherit lsn of its first page logged */
		tblk->lsn = mp->lsn;
		log->count++;

		/* insert tblock after the page on logsynclist */
		list_add(&tblk->synclist, &mp->synclist);
	}
	/*
	 * update transaction lsn:
	 */
	else {
		/* inherit oldest/smallest lsn of page */
		logdiff(diffp, mp->lsn, log);
		logdiff(difft, tblk->lsn, log);
		if (diffp < difft) {
			/* update tblock lsn with page lsn */
			tblk->lsn = mp->lsn;

			/* move tblock after page on logsynclist */
			list_del(&tblk->synclist);
			list_add(&tblk->synclist, &mp->synclist);
		}
	}

	LOGSYNC_UNLOCK(log);

	/*
	 *      write the log record
	 */
      writeRecord:
	lsn = lmWriteRecord(log, tblk, lrd, tlck);

	/*
	 * forward log syncpt if log reached next syncpt trigger
	 */
	logdiff(diffp, lsn, log);
	if (diffp >= log->nextsync)
		lsn = lmLogSync(log, 0);

	/* update end-of-log lsn */
	log->lsn = lsn;

	LOG_UNLOCK(log);

	/* return end-of-log address */
	return lsn;
}


/*
 * NAME:	lmWriteRecord()
 *
 * FUNCTION:	move the log record to current log page
 *
 * PARAMETER:	cd	- commit descriptor
 *
 * RETURN:	end-of-log address
 *			
 * serialization: LOG_LOCK() held on entry/exit
 */
static int
lmWriteRecord(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
	      struct tlock * tlck)
{
	int lsn = 0;		/* end-of-log address */
	struct lbuf *bp;	/* dst log page buffer */
	struct logpage *lp;	/* dst log page */
	caddr_t dst;		/* destination address in log page */
	int dstoffset;		/* end-of-log offset in log page */
	int freespace;		/* free space in log page */
	caddr_t p;		/* src meta-data page */
	caddr_t src;
	int srclen;
	int nbytes;		/* number of bytes to move */
	int i;
	int len;
	struct linelock *linelock;
	struct lv *lv;
	struct lvd *lvd;
	int l2linesize;

	len = 0;

	/* retrieve destination log page to write */
	bp = (struct lbuf *) log->bp;
	lp = (struct logpage *) bp->l_ldata;
	dstoffset = log->eor;

	/* any log data to write ? */
	if (tlck == NULL)
		goto moveLrd;

	/*
	 *      move log record data
	 */
	/* retrieve source meta-data page to log */
	if (tlck->flag & tlckPAGELOCK) {
		p = (caddr_t) (tlck->mp->data);
		linelock = (struct linelock *) & tlck->lock;
	}
	/* retrieve source in-memory inode to log */
	else if (tlck->flag & tlckINODELOCK) {
		if (tlck->type & tlckDTREE)
			p = (caddr_t) &JFS_IP(tlck->ip)->i_dtroot;
		else
			p = (caddr_t) &JFS_IP(tlck->ip)->i_xtroot;
		linelock = (struct linelock *) & tlck->lock;
	}
#ifdef	_JFS_WIP
	else if (tlck->flag & tlckINLINELOCK) {

		inlinelock = (struct inlinelock *) & tlck;
		p = (caddr_t) & inlinelock->pxd;
		linelock = (struct linelock *) & tlck;
	}
#endif				/* _JFS_WIP */
	else {
		jfs_err("lmWriteRecord: UFO tlck:0x%p", tlck);
		return 0;	/* Probably should trap */
	}
	l2linesize = linelock->l2linesize;

      moveData:
	ASSERT(linelock->index <= linelock->maxcnt);

	lv = linelock->lv;
	for (i = 0; i < linelock->index; i++, lv++) {
		if (lv->length == 0)
			continue;

		/* is page full ? */
		if (dstoffset >= LOGPSIZE - LOGPTLRSIZE) {
			/* page become full: move on to next page */
			lmNextPage(log);

			bp = log->bp;
			lp = (struct logpage *) bp->l_ldata;
			dstoffset = LOGPHDRSIZE;
		}

		/*
		 * move log vector data
		 */
		src = (u8 *) p + (lv->offset << l2linesize);
		srclen = lv->length << l2linesize;
		len += srclen;
		while (srclen > 0) {
			freespace = (LOGPSIZE - LOGPTLRSIZE) - dstoffset;
			nbytes = min(freespace, srclen);
			dst = (caddr_t) lp + dstoffset;
			memcpy(dst, src, nbytes);
			dstoffset += nbytes;

			/* is page not full ? */
			if (dstoffset < LOGPSIZE - LOGPTLRSIZE)
				break;

			/* page become full: move on to next page */
			lmNextPage(log);

			bp = (struct lbuf *) log->bp;
			lp = (struct logpage *) bp->l_ldata;
			dstoffset = LOGPHDRSIZE;

			srclen -= nbytes;
			src += nbytes;
		}

		/*
		 * move log vector descriptor
		 */
		len += 4;
		lvd = (struct lvd *) ((caddr_t) lp + dstoffset);
		lvd->offset = cpu_to_le16(lv->offset);
		lvd->length = cpu_to_le16(lv->length);
		dstoffset += 4;
		jfs_info("lmWriteRecord: lv offset:%d length:%d",
			 lv->offset, lv->length);
	}

	if ((i = linelock->next)) {
		linelock = (struct linelock *) lid_to_tlock(i);
		goto moveData;
	}

	/*
	 *      move log record descriptor
	 */
      moveLrd:
	lrd->length = cpu_to_le16(len);

	src = (caddr_t) lrd;
	srclen = LOGRDSIZE;

	while (srclen > 0) {
		freespace = (LOGPSIZE - LOGPTLRSIZE) - dstoffset;
		nbytes = min(freespace, srclen);
		dst = (caddr_t) lp + dstoffset;
		memcpy(dst, src, nbytes);

		dstoffset += nbytes;
		srclen -= nbytes;

		/* are there more to move than freespace of page ? */
		if (srclen)
			goto pageFull;

		/*
		 * end of log record descriptor
		 */

		/* update last log record eor */
		log->eor = dstoffset;
		bp->l_eor = dstoffset;
		lsn = (log->page << L2LOGPSIZE) + dstoffset;

		if (lrd->type & cpu_to_le16(LOG_COMMIT)) {
			tblk->clsn = lsn;
			jfs_info("wr: tclsn:0x%x, beor:0x%x", tblk->clsn,
				 bp->l_eor);

			INCREMENT(lmStat.commit);	/* # of commit */

			/*
			 * enqueue tblock for group commit:
			 *
			 * enqueue tblock of non-trivial/synchronous COMMIT
			 * at tail of group commit queue
			 * (trivial/asynchronous COMMITs are ignored by
			 * group commit.)
			 */
			LOGGC_LOCK(log);

			/* init tblock gc state */
			tblk->flag = tblkGC_QUEUE;
			tblk->bp = log->bp;
			tblk->pn = log->page;
			tblk->eor = log->eor;

			/* enqueue transaction to commit queue */
			tblk->cqnext = NULL;
			if (log->cqueue.head) {
				log->cqueue.tail->cqnext = tblk;
				log->cqueue.tail = tblk;
			} else
				log->cqueue.head = log->cqueue.tail = tblk;

			LOGGC_UNLOCK(log);
		}

		jfs_info("lmWriteRecord: lrd:0x%04x bp:0x%p pn:%d eor:0x%x",
			le16_to_cpu(lrd->type), log->bp, log->page, dstoffset);

		/* page not full ? */
		if (dstoffset < LOGPSIZE - LOGPTLRSIZE)
			return lsn;

	      pageFull:
		/* page become full: move on to next page */
		lmNextPage(log);

		bp = (struct lbuf *) log->bp;
		lp = (struct logpage *) bp->l_ldata;
		dstoffset = LOGPHDRSIZE;
		src += nbytes;
	}

	return lsn;
}


/*
 * NAME:	lmNextPage()
 *
 * FUNCTION:	write current page and allocate next page.
 *
 * PARAMETER:	log
 *
 * RETURN:	0
 *			
 * serialization: LOG_LOCK() held on entry/exit
 */
static int lmNextPage(struct jfs_log * log)
{
	struct logpage *lp;
	int lspn;		/* log sequence page number */
	int pn;			/* current page number */
	struct lbuf *bp;
	struct lbuf *nextbp;
	struct tblock *tblk;

	/* get current log page number and log sequence page number */
	pn = log->page;
	bp = log->bp;
	lp = (struct logpage *) bp->l_ldata;
	lspn = le32_to_cpu(lp->h.page);

	LOGGC_LOCK(log);

	/*
	 *      write or queue the full page at the tail of write queue
	 */
	/* get the tail tblk on commit queue */
	tblk = log->cqueue.tail;

	/* every tblk who has COMMIT record on the current page,
	 * and has not been committed, must be on commit queue
	 * since tblk is queued at commit queueu at the time
	 * of writing its COMMIT record on the page before
	 * page becomes full (even though the tblk thread
	 * who wrote COMMIT record may have been suspended
	 * currently);
	 */

	/* is page bound with outstanding tail tblk ? */
	if (tblk && tblk->pn == pn) {
		/* mark tblk for end-of-page */
		tblk->flag |= tblkGC_EOP;

		if (log->cflag & logGC_PAGEOUT) {
			/* if page is not already on write queue,
			 * just enqueue (no lbmWRITE to prevent redrive)
			 * buffer to wqueue to ensure correct serial order