jfs_logmgr.c

linux 内核源代码

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

	/* initialize the group commit serialization lock */
	LOGGC_LOCK_INIT(log);

	/* allocate/initialize the log write serialization lock */
	LOG_LOCK_INIT(log);

	LOGSYNC_LOCK_INIT(log);

	INIT_LIST_HEAD(&log->synclist);

	INIT_LIST_HEAD(&log->cqueue);
	log->flush_tblk = NULL;

	log->count = 0;

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

	if (!test_bit(log_INLINELOG, &log->flag))
		log->l2bsize = L2LOGPSIZE;

	/* check for disabled journaling to disk */
	if (log->no_integrity) {
		/*
		 * Journal pages will still be filled.  When the time comes
		 * to actually do the I/O, the write is not done, and the
		 * endio routine is called directly.
		 */
		bp = lbmAllocate(log , 0);
		log->bp = bp;
		bp->l_pn = bp->l_eor = 0;
	} else {
		/*
		 * validate log superblock
		 */
		if ((rc = lbmRead(log, 1, &bpsuper)))
			goto errout10;

		logsuper = (struct logsuper *) bpsuper->l_ldata;

		if (logsuper->magic != cpu_to_le32(LOGMAGIC)) {
			jfs_warn("*** Log Format Error ! ***");
			rc = -EINVAL;
			goto errout20;
		}

		/* logredo() should have been run successfully. */
		if (logsuper->state != cpu_to_le32(LOGREDONE)) {
			jfs_warn("*** Log Is Dirty ! ***");
			rc = -EINVAL;
			goto errout20;
		}

		/* initialize log from log superblock */
		if (test_bit(log_INLINELOG,&log->flag)) {
			if (log->size != le32_to_cpu(logsuper->size)) {
				rc = -EINVAL;
				goto errout20;
			}
			jfs_info("lmLogInit: inline log:0x%p base:0x%Lx "
				 "size:0x%x", log,
				 (unsigned long long) log->base, log->size);
		} else {
			if (memcmp(logsuper->uuid, log->uuid, 16)) {
				jfs_warn("wrong uuid on JFS log device");
				goto errout20;
			}
			log->size = le32_to_cpu(logsuper->size);
			log->l2bsize = le32_to_cpu(logsuper->l2bsize);
			jfs_info("lmLogInit: external log:0x%p base:0x%Lx "
				 "size:0x%x", log,
				 (unsigned long long) log->base, log->size);
		}

		log->page = le32_to_cpu(logsuper->end) / LOGPSIZE;
		log->eor = le32_to_cpu(logsuper->end) - (LOGPSIZE * log->page);

		/*
		 * initialize for log append write mode
		 */
		/* establish current/end-of-log page/buffer */
		if ((rc = lbmRead(log, log->page, &bp)))
			goto errout20;

		lp = (struct logpage *) bp->l_ldata;

		jfs_info("lmLogInit: lsn:0x%x page:%d eor:%d:%d",
			 le32_to_cpu(logsuper->end), log->page, log->eor,
			 le16_to_cpu(lp->h.eor));

		log->bp = bp;
		bp->l_pn = log->page;
		bp->l_eor = log->eor;

		/* if current page is full, move on to next page */
		if (log->eor >= LOGPSIZE - LOGPTLRSIZE)
			lmNextPage(log);

		/*
		 * initialize log syncpoint
		 */
		/*
		 * write the first SYNCPT record with syncpoint = 0
		 * (i.e., log redo up to HERE !);
		 * remove current page from lbm write queue at end of pageout
		 * (to write log superblock update), but do not release to
		 * freelist;
		 */
		lrd.logtid = 0;
		lrd.backchain = 0;
		lrd.type = cpu_to_le16(LOG_SYNCPT);
		lrd.length = 0;
		lrd.log.syncpt.sync = 0;
		lsn = lmWriteRecord(log, NULL, &lrd, NULL);
		bp = log->bp;
		bp->l_ceor = bp->l_eor;
		lp = (struct logpage *) bp->l_ldata;
		lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_eor);
		lbmWrite(log, bp, lbmWRITE | lbmSYNC, 0);
		if ((rc = lbmIOWait(bp, 0)))
			goto errout30;

		/*
		 * update/write superblock
		 */
		logsuper->state = cpu_to_le32(LOGMOUNT);
		log->serial = le32_to_cpu(logsuper->serial) + 1;
		logsuper->serial = cpu_to_le32(log->serial);
		lbmDirectWrite(log, bpsuper, lbmWRITE | lbmRELEASE | lbmSYNC);
		if ((rc = lbmIOWait(bpsuper, lbmFREE)))
			goto errout30;
	}

	/* initialize logsync parameters */
	log->logsize = (log->size - 2) << L2LOGPSIZE;
	log->lsn = lsn;
	log->syncpt = lsn;
	log->sync = log->syncpt;
	log->nextsync = LOGSYNC_DELTA(log->logsize);

	jfs_info("lmLogInit: lsn:0x%x syncpt:0x%x sync:0x%x",
		 log->lsn, log->syncpt, log->sync);

	/*
	 * initialize for lazy/group commit
	 */
	log->clsn = lsn;

	return 0;

	/*
	 *	unwind on error
	 */
      errout30:		/* release log page */
	log->wqueue = NULL;
	bp->l_wqnext = NULL;
	lbmFree(bp);

      errout20:		/* release log superblock */
	lbmFree(bpsuper);

      errout10:		/* unwind lbmLogInit() */
	lbmLogShutdown(log);

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


/*
 * NAME:	lmLogClose()
 *
 * FUNCTION:	remove file system <ipmnt> from active list of log <iplog>
 *		and close it on last close.
 *
 * PARAMETER:	sb	- superblock
 *
 * RETURN:	errors from subroutines
 *
 * serialization:
 */
int lmLogClose(struct super_block *sb)
{
	struct jfs_sb_info *sbi = JFS_SBI(sb);
	struct jfs_log *log = sbi->log;
	struct block_device *bdev;
	int rc = 0;

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

	mutex_lock(&jfs_log_mutex);
	LOG_LOCK(log);
	list_del(&sbi->log_list);
	LOG_UNLOCK(log);
	sbi->log = NULL;

	/*
	 * We need to make sure all of the "written" metapages
	 * actually make it to disk
	 */
	sync_blockdev(sb->s_bdev);

	if (test_bit(log_INLINELOG, &log->flag)) {
		/*
		 *	in-line log in host file system
		 */
		rc = lmLogShutdown(log);
		kfree(log);
		goto out;
	}

	if (!log->no_integrity)
		lmLogFileSystem(log, sbi, 0);

	if (!list_empty(&log->sb_list))
		goto out;

	/*
	 * TODO: ensure that the dummy_log is in a state to allow
	 * lbmLogShutdown to deallocate all the buffers and call
	 * kfree against dummy_log.  For now, leave dummy_log & its
	 * buffers in memory, and resuse if another no-integrity mount
	 * is requested.
	 */
	if (log->no_integrity)
		goto out;

	/*
	 *	external log as separate logical volume
	 */
	list_del(&log->journal_list);
	bdev = log->bdev;
	rc = lmLogShutdown(log);

	bd_release(bdev);
	blkdev_put(bdev);

	kfree(log);

      out:
	mutex_unlock(&jfs_log_mutex);
	jfs_info("lmLogClose: exit(%d)", rc);
	return rc;
}


/*
 * NAME:	jfs_flush_journal()
 *
 * FUNCTION:	initiate write of any outstanding transactions to the journal
 *		and optionally wait until they are all written to disk
 *
 *		wait == 0  flush until latest txn is committed, don't wait
 *		wait == 1  flush until latest txn is committed, wait
 *		wait > 1   flush until all txn's are complete, wait
 */
void jfs_flush_journal(struct jfs_log *log, int wait)
{
	int i;
	struct tblock *target = NULL;
	struct jfs_sb_info *sbi;

	/* jfs_write_inode may call us during read-only mount */
	if (!log)
		return;

	jfs_info("jfs_flush_journal: log:0x%p wait=%d", log, wait);

	LOGGC_LOCK(log);

	if (!list_empty(&log->cqueue)) {
		/*
		 * This ensures that we will keep writing to the journal as long
		 * as there are unwritten commit records
		 */
		target = list_entry(log->cqueue.prev, struct tblock, cqueue);

		if (test_bit(log_FLUSH, &log->flag)) {
			/*
			 * We're already flushing.
			 * if flush_tblk is NULL, we are flushing everything,
			 * so leave it that way.  Otherwise, update it to the
			 * latest transaction
			 */
			if (log->flush_tblk)
				log->flush_tblk = target;
		} else {
			/* Only flush until latest transaction is committed */
			log->flush_tblk = target;
			set_bit(log_FLUSH, &log->flag);

			/*
			 * Initiate I/O on outstanding transactions
			 */
			if (!(log->cflag & logGC_PAGEOUT)) {
				log->cflag |= logGC_PAGEOUT;
				lmGCwrite(log, 0);
			}
		}
	}
	if ((wait > 1) || test_bit(log_SYNCBARRIER, &log->flag)) {
		/* Flush until all activity complete */
		set_bit(log_FLUSH, &log->flag);
		log->flush_tblk = NULL;
	}

	if (wait && target && !(target->flag & tblkGC_COMMITTED)) {
		DECLARE_WAITQUEUE(__wait, current);

		add_wait_queue(&target->gcwait, &__wait);
		set_current_state(TASK_UNINTERRUPTIBLE);
		LOGGC_UNLOCK(log);
		schedule();
		__set_current_state(TASK_RUNNING);
		LOGGC_LOCK(log);
		remove_wait_queue(&target->gcwait, &__wait);
	}
	LOGGC_UNLOCK(log);

	if (wait < 2)
		return;

	list_for_each_entry(sbi, &log->sb_list, log_list) {
		filemap_fdatawrite(sbi->ipbmap->i_mapping);
		filemap_fdatawrite(sbi->ipimap->i_mapping);
		filemap_fdatawrite(sbi->direct_inode->i_mapping);
	}

	/*
	 * If there was recent activity, we may need to wait
	 * for the lazycommit thread to catch up
	 */
	if ((!list_empty(&log->cqueue)) || !list_empty(&log->synclist)) {
		for (i = 0; i < 200; i++) {	/* Too much? */
			msleep(250);
			if (list_empty(&log->cqueue) &&
			    list_empty(&log->synclist))
				break;
		}
	}
	assert(list_empty(&log->cqueue));

#ifdef CONFIG_JFS_DEBUG
	if (!list_empty(&log->synclist)) {
		struct logsyncblk *lp;

		printk(KERN_ERR "jfs_flush_journal: synclist not empty\n");
		list_for_each_entry(lp, &log->synclist, synclist) {
			if (lp->xflag & COMMIT_PAGE) {
				struct metapage *mp = (struct metapage *)lp;
				print_hex_dump(KERN_ERR, "metapage: ",
					       DUMP_PREFIX_ADDRESS, 16, 4,
					       mp, sizeof(struct metapage), 0);
				print_hex_dump(KERN_ERR, "page: ",
					       DUMP_PREFIX_ADDRESS, 16,
					       sizeof(long), mp->page,
					       sizeof(struct page), 0);
			} else
				print_hex_dump(KERN_ERR, "tblock:",
					       DUMP_PREFIX_ADDRESS, 16, 4,
					       lp, sizeof(struct tblock), 0);
		}
	}
#else
	WARN_ON(!list_empty(&log->synclist));
#endif
	clear_bit(log_FLUSH, &log->flag);
}

/*
 * NAME:	lmLogShutdown()
 *
 * FUNCTION:	log shutdown at last LogClose().
 *
 *		write log syncpt record.
 *		update super block to set redone flag to 0.
 *
 * PARAMETER:	log	- log inode
 *
 * RETURN:	0	- success
 *
 * serialization: single last close thread
 */
int lmLogShutdown(struct jfs_log * log)
{
	int rc;
	struct lrd lrd;
	int lsn;
	struct logsuper *logsuper;
	struct lbuf *bpsuper;
	struct lbuf *bp;
	struct logpage *lp;

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

	jfs_flush_journal(log, 2);

	/*
	 * write the last SYNCPT record with syncpoint = 0
	 * (i.e., log redo up to HERE !)
	 */
	lrd.logtid = 0;
	lrd.backchain = 0;
	lrd.type = cpu_to_le16(LOG_SYNCPT);
	lrd.length = 0;
	lrd.log.syncpt.sync = 0;

	lsn = lmWriteRecord(log, NULL, &lrd, NULL);
	bp = log->bp;
	lp = (struct logpage *) bp->l_ldata;
	lp->h.eor = lp->t.eor = cpu_to_le16(bp->l_eor);
	lbmWrite(log, log->bp, lbmWRITE | lbmRELEASE | lbmSYNC, 0);
	lbmIOWait(log->bp, lbmFREE);
	log->bp = NULL;

	/*
	 * synchronous update log superblock
	 * mark log state as shutdown cleanly
	 * (i.e., Log does not need to be replayed).
	 */
	if ((rc = lbmRead(log, 1, &bpsuper)))
		goto out;

	logsuper = (struct logsuper *) bpsuper->l_ldata;
	logsuper->state = cpu_to_le32(LOGREDONE);
	logsuper->end = cpu_to_le32(lsn);
	lbmDirectWrite(log, bpsuper, lbmWRITE | lbmRELEASE | lbmSYNC);
	rc = lbmIOWait(bpsuper, lbmFREE);

	jfs_info("lmLogShutdown: lsn:0x%x page:%d eor:%d",
		 lsn, log->page, log->eor);

      out:
	/*
	 * shutdown per log i/o
	 */
	lbmLogShutdown(log);

	if (rc) {
		jfs_warn("lmLogShutdown: exit(%d)", rc);
	}
	return rc;
}


/*
 * NAME:	lmLogFileSystem()
 *
 * FUNCTION:	insert (<activate> = true)/remove (<activate> = false)
 *	file system into/from log active file system list.
 *
 * PARAMETE:	log	- pointer to logs inode.
 *		fsdev	- kdev_t of filesystem.
 *		serial	- pointer to returned log serial number
 *		activate - insert/remove device from active list.
 *
 * RETURN:	0	- success
 *		errors returned by vms_iowait().
 */
static int lmLogFileSystem(struct jfs_log * log, struct jfs_sb_info *sbi,
			   int activate)
{
	int rc = 0;
	int i;
	struct logsuper *logsuper;
	struct lbuf *bpsuper;
	char *uuid = sbi->uuid;

	/*
	 * insert/remove file system device to log active file system list.
	 */
	if ((rc = lbmRead(log, 1, &bpsuper)))
		return rc;

	logsuper = (struct logsuper *) bpsuper->l_ldata;
	if (activate) {
		for (i = 0; i < MAX_ACTIVE; i++)
			if (!memcmp(logsuper->active[i].uuid, NULL_UUID, 16)) {
				memcpy(logsuper->active[i].uuid, uuid, 16);
				sbi->aggregate = i;
				break;
			}
		if (i == MAX_ACTIVE) {
			jfs_warn("Too many file systems sharing journal!");
			lbmFree(bpsuper);
			return -EMFILE;	/* Is there a better rc? */
		}
	} else {
		for (i = 0; i < MAX_ACTIVE; i++)
			if (!memcmp(logsuper->active[i].uuid, uuid, 16)) {
				memcpy(logsuper->active[i].uuid, NULL_UUID, 16);
				break;
			}
		if (i == MAX_ACTIVE) {
			jfs_warn("Somebody stomped on the journal!");
			lbmFree(bpsuper);
			return -EIO;
		}

	}

	/*
	 * synchronous write log superblock:
	 *
	 * write sidestream bypassing write queue:
	 * at file system mount, log super block is updated for
	 * activation of the file system before any log record
	 * (MOUNT record) of the file system, and at file system
	 * unmount, all meta data for the file system has been
	 * flushed before log super block is updated for deactivation
	 * of the file system.
	 */
	lbmDirectWrite(log, bpsuper, lbmWRITE | lbmRELEASE | lbmSYNC);
	rc = lbmIOWait(bpsuper, lbmFREE);

	return rc;
}

/*
 *		log buffer manager (lbm)
 *		------------------------
 *
 * special purpose buffer manager supporting log i/o requirements.
 *
 * per log write queue:
 * log pageout occurs in serial order by fifo write queue and
 * restricting to a single i/o in pregress at any one time.
 * a circular singly-linked list
 * (log->wrqueue points to the tail, and buffers are linked via
 * bp->wrqueue field), and
 * maintains log page in pageout ot waiting for pageout in serial pageout.
 */

/*
 *	lbmLogInit()
 *
 * initialize per log I/O setup at lmLogInit()
 */
static int lbmLogInit(struct jfs_log * log)
{				/* log inode */
	int i;
	struct lbuf *lbuf;

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

	/* initialize current buffer cursor */
	log->bp = NULL;

	/* initialize log device write queue */
	log->wqueue = NULL;

	/*
	 * Each log has its own buffer pages allocated to it.  These are
	 * not managed by the page cache.  This ensures that a transaction
	 * writing to the log does not block trying to allocate a page from
	 * the page cache (for the log).  This would be bad, since page
	 * allocation waits on the kswapd thread that may be committing inodes
	 * which would cause log activity.  Was that clear?  I'm trying to
	 * avoid deadlock here.
	 */
	init_waitqueue_head(&log->free_wait);

	log->lbuf_free = NULL;

	for (i = 0; i < LOGPAGES;) {
		char *buffer;
		uint offset;
		struct page *page;

		buffer = (char *) get_zeroed_page(GFP_KERNEL);
		if (buffer == NULL)
			goto error;
		page = virt_to_page(buffer);
		for (offset = 0; offset < PAGE_SIZE; offset += LOGPSIZE) {
			lbuf = kmalloc(sizeof(struct lbuf), GFP_KERNEL);
			if (lbuf == NULL) {
				if (offset == 0)
					free_page((unsigned long) buffer);
				goto error;
			}
			if (offset) /* we already have one reference */
				get_page(page);
			lbuf->l_offset = offset;
			lbuf->l_ldata = buffer + offset;
			lbuf->l_page = page;
			lbuf->l_log = log;
			init_waitqueue_head(&lbuf->l_ioevent);

			lbuf->l_freelist = log->lbuf_free;
			log->lbuf_free = lbuf;
			i++;
		}
	}

	return (0);

      error:
	lbmLogShutdown(log);
	return -ENOMEM;
}


/*
 *	lbmLogShutdown()
 *
 * finalize per log I/O setup at lmLogShutdown()
 */
static void lbmLogShutdown(struct jfs_log * log)
{
	struct lbuf *lbuf;

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

	lbuf = log->lbuf_free;
	while (lbuf) {
		struct lbuf *next = lbuf->l_freelist;
		__free_page(lbuf->l_page);
		kfree(lbuf);
		lbuf = next;
	}
}


/*
 *	lbmAllocate()
 *
 * allocate an empty log buffer
 */
static struct lbuf *lbmAllocate(struct jfs_log * log, int pn)
{