jfs_dmap.c

jfs-2.4-1.1.7.tar.gz jfs 2.4－1.1.7 源码

 *		the blocks will be updated in the persistent map one
 *		dmap at a time.
 *
 * PARAMETERS:
 *      ipbmap	-  pointer to in-core inode for the block map.
 *      free	- TRUE if block range is to be freed from the persistent
 *		  map; FALSE if it is to   be allocated.
 *      blkno	-  starting block number of the range.
 *      nblocks	-  number of contiguous blocks in the range.
 *      tblk	-  transaction block;
 *
 * RETURN VALUES:
 *      0	- success
 *      -EIO	- i/o error
 */
int
dbUpdatePMap(struct inode *ipbmap,
	     int free, s64 blkno, s64 nblocks, struct tblock * tblk)
{
	int nblks, dbitno, wbitno, rbits;
	int word, nbits, nwords;
	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
	s64 lblkno, rem, lastlblkno;
	u32 mask;
	struct dmap *dp;
	struct metapage *mp;
	struct jfs_log *log;
	int lsn, difft, diffp;

	/* the blocks better be within the mapsize. */
	if (blkno + nblocks > bmp->db_mapsize) {
		printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
		       (unsigned long long) blkno,
		       (unsigned long long) nblocks);
		jfs_error(ipbmap->i_sb,
			  "dbUpdatePMap: blocks are outside the map");
		return -EIO;
	}

	/* compute delta of transaction lsn from log syncpt */
	lsn = tblk->lsn;
	log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
	logdiff(difft, lsn, log);

	/*
	 * update the block state a dmap at a time.
	 */
	mp = NULL;
	lastlblkno = 0;
	for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
		/* get the buffer for the current dmap. */
		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
		if (lblkno != lastlblkno) {
			if (mp) {
				write_metapage(mp);
			}

			mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
					   0);
			if (mp == NULL)
				return -EIO;
		}
		dp = (struct dmap *) mp->data;

		/* determine the bit number and word within the dmap of
		 * the starting block.  also determine how many blocks
		 * are to be updated within this dmap.
		 */
		dbitno = blkno & (BPERDMAP - 1);
		word = dbitno >> L2DBWORD;
		nblks = min(rem, (s64)BPERDMAP - dbitno);

		/* update the bits of the dmap words. the first and last
		 * words may only have a subset of their bits updated. if
		 * this is the case, we'll work against that word (i.e.
		 * partial first and/or last) only in a single pass.  a 
		 * single pass will also be used to update all words that
		 * are to have all their bits updated.
		 */
		for (rbits = nblks; rbits > 0;
		     rbits -= nbits, dbitno += nbits) {
			/* determine the bit number within the word and
			 * the number of bits within the word.
			 */
			wbitno = dbitno & (DBWORD - 1);
			nbits = min(rbits, DBWORD - wbitno);

			/* check if only part of the word is to be updated. */
			if (nbits < DBWORD) {
				/* update (free or allocate) the bits
				 * in this word.
				 */
				mask =
				    (ONES << (DBWORD - nbits) >> wbitno);
				if (free)
					dp->pmap[word] &=
					    cpu_to_le32(~mask);
				else
					dp->pmap[word] |=
					    cpu_to_le32(mask);

				word += 1;
			} else {
				/* one or more words are to have all
				 * their bits updated.  determine how
				 * many words and how many bits.
				 */
				nwords = rbits >> L2DBWORD;
				nbits = nwords << L2DBWORD;

				/* update (free or allocate) the bits
				 * in these words.
				 */
				if (free)
					memset(&dp->pmap[word], 0,
					       nwords * 4);
				else
					memset(&dp->pmap[word], (int) ONES,
					       nwords * 4);

				word += nwords;
			}
		}

		/*
		 * update dmap lsn
		 */
		if (lblkno == lastlblkno)
			continue;

		lastlblkno = lblkno;

		if (mp->lsn != 0) {
			/* inherit older/smaller lsn */
			logdiff(diffp, mp->lsn, log);
			if (difft < diffp) {
				mp->lsn = lsn;

				/* move bp after tblock in logsync list */
				LOGSYNC_LOCK(log);
				list_del(&mp->synclist);
				list_add(&mp->synclist, &tblk->synclist);
				LOGSYNC_UNLOCK(log);
			}

			/* inherit younger/larger clsn */
			LOGSYNC_LOCK(log);
			logdiff(difft, tblk->clsn, log);
			logdiff(diffp, mp->clsn, log);
			if (difft > diffp)
				mp->clsn = tblk->clsn;
			LOGSYNC_UNLOCK(log);
		} else {
			mp->log = log;
			mp->lsn = lsn;

			/* insert bp after tblock in logsync list */
			LOGSYNC_LOCK(log);

			log->count++;
			list_add(&mp->synclist, &tblk->synclist);

			mp->clsn = tblk->clsn;
			LOGSYNC_UNLOCK(log);
		}
	}

	/* write the last buffer. */
	if (mp) {
		write_metapage(mp);
	}

	return (0);
}


/*
 * NAME:	dbNextAG()
 *
 * FUNCTION:    find the preferred allocation group for new allocations.
 *
 *		Within the allocation groups, we maintain a preferred
 *		allocation group which consists of a group with at least
 *		average free space.  It is the preferred group that we target
 *		new inode allocation towards.  The tie-in between inode
 *		allocation and block allocation occurs as we allocate the
 *		first (data) block of an inode and specify the inode (block)
 *		as the allocation hint for this block.
 *
 *		We try to avoid having more than one open file growing in
 *		an allocation group, as this will lead to fragmentation.
 *		This differs from the old OS/2 method of trying to keep
 *		empty ags around for large allocations.
 *
 * PARAMETERS:
 *      ipbmap	-  pointer to in-core inode for the block map.
 *
 * RETURN VALUES:
 *      the preferred allocation group number.
 */
int dbNextAG(struct inode *ipbmap)
{
	s64 avgfree;
	int agpref;
	s64 hwm = 0;
	int i;
	int next_best = -1;
	struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;

	BMAP_LOCK(bmp);

	/* determine the average number of free blocks within the ags. */
	avgfree = (u32)bmp->db_nfree / bmp->db_numag;

	/*
	 * if the current preferred ag does not have an active allocator
	 * and has at least average freespace, return it
	 */
	agpref = bmp->db_agpref;
	if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
	    (bmp->db_agfree[agpref] >= avgfree))
		goto unlock;

	/* From the last preferred ag, find the next one with at least
	 * average free space.
	 */
	for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
		if (agpref == bmp->db_numag)
			agpref = 0;

		if (atomic_read(&bmp->db_active[agpref]))
			/* open file is currently growing in this ag */
			continue;
		if (bmp->db_agfree[agpref] >= avgfree) {
			/* Return this one */
			bmp->db_agpref = agpref;
			goto unlock;
		} else if (bmp->db_agfree[agpref] > hwm) {
			/* Less than avg. freespace, but best so far */
			hwm = bmp->db_agfree[agpref];
			next_best = agpref;
		}
	}

	/*
	 * If no inactive ag was found with average freespace, use the
	 * next best
	 */
	if (next_best != -1)
		bmp->db_agpref = next_best;
	/* else leave db_agpref unchanged */
unlock:
	BMAP_UNLOCK(bmp);

	/* return the preferred group.
	 */
	return (bmp->db_agpref);
}

/*
 * NAME:	dbAlloc()
 *
 * FUNCTION:    attempt to allocate a specified number of contiguous free
 *		blocks from the working allocation block map.
 *
 *		the block allocation policy uses hints and a multi-step
 *		approach.
 *
 *	  	for allocation requests smaller than the number of blocks
 *		per dmap, we first try to allocate the new blocks
 *		immediately following the hint.  if these blocks are not
 *		available, we try to allocate blocks near the hint.  if
 *		no blocks near the hint are available, we next try to 
 *		allocate within the same dmap as contains the hint.
 *
 *		if no blocks are available in the dmap or the allocation
 *		request is larger than the dmap size, we try to allocate
 *		within the same allocation group as contains the hint. if
 *		this does not succeed, we finally try to allocate anywhere
 *		within the aggregate.
 *
 *		we also try to allocate anywhere within the aggregate for
 *		for allocation requests larger than the allocation group
 *		size or requests that specify no hint value.
 *
 * PARAMETERS:
 *      ip	-  pointer to in-core inode;
 *      hint	- allocation hint.
 *      nblocks	- number of contiguous blocks in the range.
 *      results	- on successful return, set to the starting block number
 *		  of the newly allocated contiguous range.
 *
 * RETURN VALUES:
 *      0	- success
 *      -ENOSPC	- insufficient disk resources
 *      -EIO	- i/o error
 */
int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
{
	int rc, agno;
	struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
	struct bmap *bmp;
	struct metapage *mp;
	s64 lblkno, blkno;
	struct dmap *dp;
	int l2nb;
	s64 mapSize;
	int writers;

	/* assert that nblocks is valid */
	assert(nblocks > 0);

#ifdef _STILL_TO_PORT
	/* DASD limit check                                     F226941 */
	if (OVER_LIMIT(ip, nblocks))
		return -ENOSPC;
#endif				/* _STILL_TO_PORT */

	/* get the log2 number of blocks to be allocated.
	 * if the number of blocks is not a log2 multiple, 
	 * it will be rounded up to the next log2 multiple.
	 */
	l2nb = BLKSTOL2(nblocks);

	bmp = JFS_SBI(ip->i_sb)->bmap;

//retry:        /* serialize w.r.t.extendfs() */
	mapSize = bmp->db_mapsize;

	/* the hint should be within the map */
	if (hint >= mapSize) {
		jfs_error(ip->i_sb, "dbAlloc: the hint is outside the map");
		return -EIO;
	}

	/* if the number of blocks to be allocated is greater than the
	 * allocation group size, try to allocate anywhere.
	 */
	if (l2nb > bmp->db_agl2size) {
		IWRITE_LOCK(ipbmap);

		rc = dbAllocAny(bmp, nblocks, l2nb, results);
		if (rc == 0) {
			DBALLOC(bmp->db_DBmap, bmp->db_mapsize, *results,
				nblocks);
		}

		goto write_unlock;
	}

	/*
	 * If no hint, let dbNextAG recommend an allocation group
	 */
	if (hint == 0)
		goto pref_ag;

	/* we would like to allocate close to the hint.  adjust the
	 * hint to the block following the hint since the allocators
	 * will start looking for free space starting at this point.
	 */
	blkno = hint + 1;

	if (blkno >= bmp->db_mapsize)
		goto pref_ag;

	agno = blkno >> bmp->db_agl2size;

	/* check if blkno crosses over into a new allocation group.
	 * if so, check if we should allow allocations within this
	 * allocation group.
	 */
	if ((blkno & (bmp->db_agsize - 1)) == 0)
		/* check if the AG is currenly being written to.
		 * if so, call dbNextAG() to find a non-busy
		 * AG with sufficient free space.
		 */
		if (atomic_read(&bmp->db_active[agno]))
			goto pref_ag;

	/* check if the allocation request size can be satisfied from a
	 * single dmap.  if so, try to allocate from the dmap containing
	 * the hint using a tiered strategy.
	 */
	if (nblocks <= BPERDMAP) {
		IREAD_LOCK(ipbmap);

		/* get the buffer for the dmap containing the hint.
		 */
		rc = -EIO;
		lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
		mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
		if (mp == NULL)
			goto read_unlock;

		dp = (struct dmap *) mp->data;

		/* first, try to satisfy the allocation request with the
		 * blocks beginning at the hint.
		 */
		if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
		    != -ENOSPC) {
			if (rc == 0) {
				*results = blkno;
				DBALLOC(bmp->db_DBmap, bmp->db_mapsize,
					*results, nblocks);
				mark_metapage_dirty(mp);
			}

			release_metapage(mp);
			goto read_unlock;
		}

		writers = atomic_read(&bmp->db_active[agno]);
		if ((writers > 1) ||
		    ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
			/*
			 * Someone else is writing in this allocation
			 * group.  To avoid fragmenting, try another ag
			 */
			release_metapage(mp);
			IREAD_UNLOCK(ipbmap);
			goto pref_ag;
		}

		/* next, try to satisfy the allocation request with blocks
		 * near the hint.
		 */
		if ((rc =
		     dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
		    != -ENOSPC) {
			if (rc == 0) {
				DBALLOC(bmp->db_DBmap, bmp->db_mapsize,
					*results, nblocks);
				mark_metapage_dirty(mp);
			}

			release_metapage(mp);
			goto read_unlock;
		}

		/* try to satisfy the allocation request with blocks within
		 * the same dmap as the hint.
		 */
		if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
		    != -ENOSPC) {