jfs_logmgr.c

jfs-2.4-1.1.7.tar.gz jfs 2.4－1.1.7 源码

			 * of the pages since log pages will be added
			 * continuously
			 */
			if (bp->l_wqnext == NULL)
				lbmWrite(log, bp, 0, 0);
		} else {
			/*
			 * No current GC leader, initiate group commit
			 */
			log->cflag |= logGC_PAGEOUT;
			lmGCwrite(log, 0);
		}
	}
	/* page is not bound with outstanding tblk:
	 * init write or mark it to be redriven (lbmWRITE)
	 */
	else {
		/* finalize the page */
		bp->l_ceor = bp->l_eor;
		lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_ceor);
		lbmWrite(log, bp, lbmWRITE | lbmRELEASE | lbmFREE, 0);
	}
	LOGGC_UNLOCK(log);

	/*
	 *      allocate/initialize next page
	 */
	/* if log wraps, the first data page of log is 2
	 * (0 never used, 1 is superblock).
	 */
	log->page = (pn == log->size - 1) ? 2 : pn + 1;
	log->eor = LOGPHDRSIZE;	/* ? valid page empty/full at logRedo() */

	/* allocate/initialize next log page buffer */
	nextbp = lbmAllocate(log, log->page);
	nextbp->l_eor = log->eor;
	log->bp = nextbp;

	/* initialize next log page */
	lp = (struct logpage *) nextbp->l_ldata;
	lp->h.page = lp->t.page = cpu_to_le32(lspn + 1);
	lp->h.eor = lp->t.eor = cpu_to_le16(LOGPHDRSIZE);

	return 0;
}


/*
 * NAME:	lmGroupCommit()
 *
 * FUNCTION:	group commit
 *	initiate pageout of the pages with COMMIT in the order of
 *	page number - redrive pageout of the page at the head of
 *	pageout queue until full page has been written.
 *
 * RETURN:	
 *
 * NOTE:
 *	LOGGC_LOCK serializes log group commit queue, and
 *	transaction blocks on the commit queue.
 *	N.B. LOG_LOCK is NOT held during lmGroupCommit().
 */
int lmGroupCommit(struct jfs_log * log, struct tblock * tblk)
{
	int rc = 0;

	LOGGC_LOCK(log);

	/* group committed already ? */
	if (tblk->flag & tblkGC_COMMITTED) {
		if (tblk->flag & tblkGC_ERROR)
			rc = -EIO;

		LOGGC_UNLOCK(log);
		return rc;
	}
	jfs_info("lmGroup Commit: tblk = 0x%p, gcrtc = %d", tblk, log->gcrtc);

	if (tblk->xflag & COMMIT_LAZY)
		tblk->flag |= tblkGC_LAZY;

	if ((!(log->cflag & logGC_PAGEOUT)) && log->cqueue.head &&
	    (!(tblk->xflag & COMMIT_LAZY) || test_bit(log_FLUSH, &log->flag))) {
		/*
		 * No pageout in progress
		 *
		 * start group commit as its group leader.
		 */
		log->cflag |= logGC_PAGEOUT;

		lmGCwrite(log, 0);
	}

	if (tblk->xflag & COMMIT_LAZY) {
		/*
		 * Lazy transactions can leave now
		 */
		LOGGC_UNLOCK(log);
		return 0;
	}

	/* lmGCwrite gives up LOGGC_LOCK, check again */

	if (tblk->flag & tblkGC_COMMITTED) {
		if (tblk->flag & tblkGC_ERROR)
			rc = -EIO;

		LOGGC_UNLOCK(log);
		return rc;
	}

	/* upcount transaction waiting for completion
	 */
	log->gcrtc++;
	tblk->flag |= tblkGC_READY;

	__SLEEP_COND(tblk->gcwait, (tblk->flag & tblkGC_COMMITTED),
		     LOGGC_LOCK(log), LOGGC_UNLOCK(log));

	/* removed from commit queue */
	if (tblk->flag & tblkGC_ERROR)
		rc = -EIO;

	LOGGC_UNLOCK(log);
	return rc;
}

/*
 * NAME:	lmGCwrite()
 *
 * FUNCTION:	group commit write
 *	initiate write of log page, building a group of all transactions
 *	with commit records on that page.
 *
 * RETURN:	None
 *
 * NOTE:
 *	LOGGC_LOCK must be held by caller.
 *	N.B. LOG_LOCK is NOT held during lmGroupCommit().
 */
static void lmGCwrite(struct jfs_log * log, int cant_write)
{
	struct lbuf *bp;
	struct logpage *lp;
	int gcpn;		/* group commit page number */
	struct tblock *tblk;
	struct tblock *xtblk;

	/*
	 * build the commit group of a log page
	 *
	 * scan commit queue and make a commit group of all
	 * transactions with COMMIT records on the same log page.
	 */
	/* get the head tblk on the commit queue */
	tblk = xtblk = log->cqueue.head;
	gcpn = tblk->pn;

	while (tblk && tblk->pn == gcpn) {
		xtblk = tblk;

		/* state transition: (QUEUE, READY) -> COMMIT */
		tblk->flag |= tblkGC_COMMIT;
		tblk = tblk->cqnext;
	}
	tblk = xtblk;		/* last tblk of the page */

	/*
	 * pageout to commit transactions on the log page.
	 */
	bp = (struct lbuf *) tblk->bp;
	lp = (struct logpage *) bp->l_ldata;
	/* is page already full ? */
	if (tblk->flag & tblkGC_EOP) {
		/* mark page to free at end of group commit of the page */
		tblk->flag &= ~tblkGC_EOP;
		tblk->flag |= tblkGC_FREE;
		bp->l_ceor = bp->l_eor;
		lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_ceor);
		lbmWrite(log, bp, lbmWRITE | lbmRELEASE | lbmGC,
			 cant_write);
		INCREMENT(lmStat.full_page);
	}
	/* page is not yet full */
	else {
		bp->l_ceor = tblk->eor;	/* ? bp->l_ceor = bp->l_eor; */
		lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_ceor);
		lbmWrite(log, bp, lbmWRITE | lbmGC, cant_write);
		INCREMENT(lmStat.partial_page);
	}
}

/*
 * NAME:	lmPostGC()
 *
 * FUNCTION:	group commit post-processing
 *	Processes transactions after their commit records have been written
 *	to disk, redriving log I/O if necessary.
 *
 * RETURN:	None
 *
 * NOTE:
 *	This routine is called a interrupt time by lbmIODone
 */
static void lmPostGC(struct lbuf * bp)
{
	unsigned long flags;
	struct jfs_log *log = bp->l_log;
	struct logpage *lp;
	struct tblock *tblk;

	//LOGGC_LOCK(log);
	spin_lock_irqsave(&log->gclock, flags);
	/*
	 * current pageout of group commit completed.
	 *
	 * remove/wakeup transactions from commit queue who were
	 * group committed with the current log page
	 */
	while ((tblk = log->cqueue.head) && (tblk->flag & tblkGC_COMMIT)) {
		/* if transaction was marked GC_COMMIT then
		 * it has been shipped in the current pageout
		 * and made it to disk - it is committed.
		 */

		if (bp->l_flag & lbmERROR)
			tblk->flag |= tblkGC_ERROR;

		/* remove it from the commit queue */
		log->cqueue.head = tblk->cqnext;
		if (log->cqueue.head == NULL)
			log->cqueue.tail = NULL;
		tblk->flag &= ~tblkGC_QUEUE;
		tblk->cqnext = 0;

		if (tblk == log->flush_tblk) {
			/* we can stop flushing the log now */
			clear_bit(log_FLUSH, &log->flag);
			log->flush_tblk = NULL;
		}

		jfs_info("lmPostGC: tblk = 0x%p, flag = 0x%x", tblk,
			 tblk->flag);

		if (!(tblk->xflag & COMMIT_FORCE))
			/*
			 * Hand tblk over to lazy commit thread
			 */
			txLazyUnlock(tblk);
		else {
			/* state transition: COMMIT -> COMMITTED */
			tblk->flag |= tblkGC_COMMITTED;

			if (tblk->flag & tblkGC_READY)
				log->gcrtc--;

			LOGGC_WAKEUP(tblk);
		}

		/* was page full before pageout ?
		 * (and this is the last tblk bound with the page)
		 */
		if (tblk->flag & tblkGC_FREE)
			lbmFree(bp);
		/* did page become full after pageout ?
		 * (and this is the last tblk bound with the page)
		 */
		else if (tblk->flag & tblkGC_EOP) {
			/* finalize the page */
			lp = (struct logpage *) bp->l_ldata;
			bp->l_ceor = bp->l_eor;
			lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_eor);
			jfs_info("lmPostGC: calling lbmWrite");
			lbmWrite(log, bp, lbmWRITE | lbmRELEASE | lbmFREE,
				 1);
		}

	}

	/* are there any transactions who have entered lnGroupCommit()
	 * (whose COMMITs are after that of the last log page written.
	 * They are waiting for new group commit (above at (SLEEP 1))
	 * or lazy transactions are on a full (queued) log page,
	 * select the latest ready transaction as new group leader and
	 * wake her up to lead her group.
	 */
	if ((tblk = log->cqueue.head) &&
	    ((log->gcrtc > 0) || (tblk->bp->l_wqnext != NULL) ||
	     test_bit(log_FLUSH, &log->flag)))
		/*
		 * Call lmGCwrite with new group leader
		 */
		lmGCwrite(log, 1);

	/* no transaction are ready yet (transactions are only just
	 * queued (GC_QUEUE) and not entered for group commit yet).
	 * the first transaction entering group commit
	 * will elect herself as new group leader.
	 */
	else
		log->cflag &= ~logGC_PAGEOUT;

	//LOGGC_UNLOCK(log);
	spin_unlock_irqrestore(&log->gclock, flags);
	return;
}

/*
 * NAME:	lmLogSync()
 *
 * FUNCTION:	write log SYNCPT record for specified log
 *	if new sync address is available
 *	(normally the case if sync() is executed by back-ground
 *	process).
 *	if not, explicitly run jfs_blogsync() to initiate
 *	getting of new sync address.
 *	calculate new value of i_nextsync which determines when
 *	this code is called again.
 *
 *	this is called only from lmLog().
 *
 * PARAMETER:	ip	- pointer to logs inode.
 *
 * RETURN:	0
 *			
 * serialization: LOG_LOCK() held on entry/exit
 */
static int lmLogSync(struct jfs_log * log, int nosyncwait)
{
	int logsize;
	int written;		/* written since last syncpt */
	int free;		/* free space left available */
	int delta;		/* additional delta to write normally */
	int more;		/* additional write granted */
	struct lrd lrd;
	int lsn;
	struct logsyncblk *lp;

	/*
	 *      forward syncpt
	 */
	/* if last sync is same as last syncpt,
	 * invoke sync point forward processing to update sync.
	 */

	if (log->sync == log->syncpt) {
		LOGSYNC_LOCK(log);
		/* ToDo: push dirty metapages out to disk */
//              bmLogSync(log);

		if (list_empty(&log->synclist))
			log->sync = log->lsn;
		else {
			lp = list_entry(log->synclist.next,
					struct logsyncblk, synclist);
			log->sync = lp->lsn;
		}
		LOGSYNC_UNLOCK(log);

	}

	/* if sync is different from last syncpt,
	 * write a SYNCPT record with syncpt = sync.
	 * reset syncpt = sync
	 */
	if (log->sync != log->syncpt) {
		struct super_block *sb = log->sb;
		struct jfs_sb_info *sbi = JFS_SBI(sb);

		/*
		 * We need to make sure all of the "written" metapages
		 * actually make it to disk
		 */
		fsync_inode_data_buffers(sbi->ipbmap);
		fsync_inode_data_buffers(sbi->ipimap);
		fsync_inode_data_buffers(sb->s_bdev->bd_inode);

		lrd.logtid = 0;
		lrd.backchain = 0;
		lrd.type = cpu_to_le16(LOG_SYNCPT);
		lrd.length = 0;
		lrd.log.syncpt.sync = cpu_to_le32(log->sync);
		lsn = lmWriteRecord(log, NULL, &lrd, NULL);

		log->syncpt = log->sync;
	} else
		lsn = log->lsn;

	/*
	 *      setup next syncpt trigger (SWAG)
	 */
	logsize = log->logsize;

	logdiff(written, lsn, log);
	free = logsize - written;
	delta = LOGSYNC_DELTA(logsize);
	more = min(free / 2, delta);
	if (more < 2 * LOGPSIZE) {
		jfs_warn("\n ... Log Wrap ... Log Wrap ... Log Wrap ...\n");
		/*
		 *      log wrapping
		 *
		 * option 1 - panic ? No.!
		 * option 2 - shutdown file systems
		 *            associated with log ?
		 * option 3 - extend log ?
		 */
		/*
		 * option 4 - second chance
		 *
		 * mark log wrapped, and continue.
		 * when all active transactions are completed,
		 * mark log vaild for recovery.
		 * if crashed during invalid state, log state
		 * implies invald log, forcing fsck().
		 */
		/* mark log state log wrap in log superblock */
		/* log->state = LOGWRAP; */

		/* reset sync point computation */
		log->syncpt = log->sync = lsn;
		log->nextsync = delta;
	} else
		/* next syncpt trigger = written + more */
		log->nextsync = written + more;

	/* return if lmLogSync() from outside of transaction, e.g., sync() */
	if (nosyncwait)
		return lsn;

	/* if number of bytes written from last sync point is more
	 * than 1/4 of the log size, stop new transactions from
	 * starting until all current transactions are completed
	 * by setting syncbarrier flag.
	 */
	if (written > LOGSYNC_BARRIER(logsize) && logsize > 32 * LOGPSIZE) {
		set_bit(log_SYNCBARRIER, &log->flag);
		jfs_info("log barrier on: lsn=0x%x syncpt=0x%x", lsn,
			 log->syncpt);
		/*
		 * We may have to initiate group commit
		 */
		jfs_flush_journal(log, 0);
	}

	return lsn;
}


/*
 * NAME:	lmLogOpen()
 *
 * FUNCTION:    open the log on first open;
 *	insert filesystem in the active list of the log.
 *
 * PARAMETER:	ipmnt	- file system mount inode
 *		iplog 	- log inode (out)
 *
 * RETURN:
 *
 * serialization:
 */
int lmLogOpen(struct super_block *sb, struct jfs_log ** logptr)
{
	int rc;
	struct block_device *bdev;
	struct jfs_log *log;

	if (!(log = kmalloc(sizeof(struct jfs_log), GFP_KERNEL)))
		return -ENOMEM;
	memset(log, 0, sizeof(struct jfs_log));
	init_waitqueue_head(&log->syncwait);

	log->sb = sb;		/* This should be a list */

	if (!(JFS_SBI(sb)->mntflag & JFS_INLINELOG))
		goto externalLog;

	/*
	 *      in-line log in host file system
	 *
	 * file system to log have 1-to-1 relationship;
	 */

	set_bit(log_INLINELOG, &log->flag);
	log->bdev = sb->s_bdev;
	log->base = addressPXD(&JFS_SBI(sb)->logpxd);
	log->size = lengthPXD(&JFS_SBI(sb)->logpxd) >>
	    (L2LOGPSIZE - sb->s_blocksize_bits);
	log->l2bsize = sb->s_blocksize_bits;
	ASSERT(L2LOGPSIZE >= sb->s_blocksize_bits);

	/*
	 * initialize log.
	 */
	if ((rc = lmLogInit(log)))
		goto free;
	goto out;

	/*
	 *      external log as separate logical volume
	 *
	 * file systems to log may have n-to-1 relationship;
	 */
      externalLog:

	/*
	 * TODO: Check for already opened log devices
	 */

	if (!(bdev = bdget(kdev_t_to_nr(JFS_SBI(sb)->logdev)))) {
		rc = -ENODEV;
		goto free;
	}

	if ((rc = blkdev_get(bdev, FMODE_READ|FMODE_WRITE, 0, BDEV_FS)))
		goto free;

	log->bdev = bdev;
	memcpy(log->uuid, JFS_SBI(sb)->loguuid, sizeof(log->uuid));
	
	/*
	 * initialize log:
	 */
	if ((rc = lmLogInit(log)))
		goto close;

	/*
	 * add file system to log active file system list
	 */
	if ((rc = lmLogFileSystem(log, JFS_SBI(sb)->uuid, 1)))
		goto shutdown;

      out:
	*logptr = log;
	return 0;

	/*
	 *      unwind on error
	 */
      shutdown:		/* unwind lbmLogInit() */
	lbmLogShutdown(log);

      close:		/* close external log device */
	blkdev_put(bdev, BDEV_FS);

      free:		/* free log descriptor */
	kfree(log);

	jfs_warn("lmLogOpen: exit(%d)", rc);
	return rc;
}


/*
 * NAME:	lmLogInit()
 *
 * FUNCTION:	log initialization at first log open.
 *
 *	logredo() (or logformat()) should have been run previously.
 *	initialize the log inode from log superblock.
 *	set the log state in the superblock to LOGMOUNT and
 *	write SYNCPT log record.
 *		
 * PARAMETER:	log	- log structure
 *
 * RETURN:	0	- if ok
 *		-EINVAL	- bad log magic number or superblock dirty
 *		error returned from logwait()
 *			
 * serialization: single first open thread
 */
int lmLogInit(struct jfs_log * log)
{
	int rc = 0;
	struct lrd lrd;
	struct logsuper *logsuper;
	struct lbuf *bpsuper;
	struct lbuf *bp;
	struct logpage *lp;
	int lsn;

	jfs_info("lmLogInit: log:0x%p", log);

	/*
	 * log inode is overlaid on generic inode where
	 * dinode have been zeroed out by iRead();
	 */

	/*
	 * initialize log i/o
	 */
	if ((rc = lbmLogInit(log)))
		return rc;

	/*
	 * validate log superblock
	 */