vfs_cluster.c

早期freebsd实现

	struct vnode *vp;
	struct buf *bp;
	long flags;
	daddr_t blkno;
	daddr_t lblkno;
	long size;
	int run;
{
	if (!bp) {
		bp = getblk(vp, lblkno, size, 0, 0);
		if (bp->b_flags & (B_DONE | B_DELWRI)) {
			bp->b_blkno = blkno;
			return(bp);
		}
	}
	allocbuf(bp, run * size);
	bp->b_blkno = blkno;
	bp->b_iodone = cluster_callback;
	bp->b_flags |= flags | B_CALL;
	return(bp);
}

/*
 * Cleanup after a clustered read or write.
 * This is complicated by the fact that any of the buffers might have
 * extra memory (if there were no empty buffer headers at allocbuf time)
 * that we will need to shift around.
 */
void
cluster_callback(bp)
	struct buf *bp;
{
	struct cluster_save *b_save;
	struct buf **bpp, *tbp;
	long bsize;
	caddr_t cp;
	int error = 0;

	/*
	 * Must propogate errors to all the components.
	 */
	if (bp->b_flags & B_ERROR)
		error = bp->b_error;

	b_save = (struct cluster_save *)(bp->b_saveaddr);
	bp->b_saveaddr = b_save->bs_saveaddr;

	bsize = b_save->bs_bufsize;
	cp = (char *)bp->b_data + bsize;
	/*
	 * Move memory from the large cluster buffer into the component
	 * buffers and mark IO as done on these.
	 */
	for (bpp = b_save->bs_children; b_save->bs_nchildren--; ++bpp) {
		tbp = *bpp;
		pagemove(cp, tbp->b_data, bsize);
		tbp->b_bufsize += bsize;
		tbp->b_bcount = bsize;
		if (error) {
			tbp->b_flags |= B_ERROR;
			tbp->b_error = error;
		}
		biodone(tbp);
		bp->b_bufsize -= bsize;
		cp += bsize;
	}
	/*
	 * If there was excess memory in the cluster buffer,
	 * slide it up adjacent to the remaining valid data.
	 */
	if (bp->b_bufsize != bsize) {
		if (bp->b_bufsize < bsize)
			panic("cluster_callback: too little memory");
		pagemove(cp, (char *)bp->b_data + bsize, bp->b_bufsize - bsize);
	}
	bp->b_bcount = bsize;
	bp->b_iodone = NULL;
	free(b_save, M_SEGMENT);
	if (bp->b_flags & B_ASYNC)
		brelse(bp);
	else {
		bp->b_flags &= ~B_WANTED;
		wakeup((caddr_t)bp);
	}
}

/*
 * Do clustered write for FFS.
 *
 * Three cases:
 *	1. Write is not sequential (write asynchronously)
 *	Write is sequential:
 *	2.	beginning of cluster - begin cluster
 *	3.	middle of a cluster - add to cluster
 *	4.	end of a cluster - asynchronously write cluster
 */
void
cluster_write(bp, filesize)
        struct buf *bp;
	u_quad_t filesize;
{
        struct vnode *vp;
        daddr_t lbn;
        int maxclen, cursize;

        vp = bp->b_vp;
        lbn = bp->b_lblkno;

	/* Initialize vnode to beginning of file. */
	if (lbn == 0)
		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;

        if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
	    (bp->b_blkno != vp->v_lasta + btodb(bp->b_bcount))) {
		maxclen = MAXBSIZE / vp->v_mount->mnt_stat.f_iosize - 1;
		if (vp->v_clen != 0) {
			/*
			 * Next block is not sequential.
			 *
			 * If we are not writing at end of file, the process
			 * seeked to another point in the file since its
			 * last write, or we have reached our maximum
			 * cluster size, then push the previous cluster.
			 * Otherwise try reallocating to make it sequential.
			 */
			cursize = vp->v_lastw - vp->v_cstart + 1;
			if (!doreallocblks ||
			    (lbn + 1) * bp->b_bcount != filesize ||
			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
				cluster_wbuild(vp, NULL, bp->b_bcount,
				    vp->v_cstart, cursize, lbn);
			} else {
				struct buf **bpp, **endbp;
				struct cluster_save *buflist;

				buflist = cluster_collectbufs(vp, bp);
				endbp = &buflist->bs_children
				    [buflist->bs_nchildren - 1];
				if (VOP_REALLOCBLKS(vp, buflist)) {
					/*
					 * Failed, push the previous cluster.
					 */
					for (bpp = buflist->bs_children;
					     bpp < endbp; bpp++)
						brelse(*bpp);
					free(buflist, M_SEGMENT);
					cluster_wbuild(vp, NULL, bp->b_bcount,
					    vp->v_cstart, cursize, lbn);
				} else {
					/*
					 * Succeeded, keep building cluster.
					 */
					for (bpp = buflist->bs_children;
					     bpp <= endbp; bpp++)
						bdwrite(*bpp);
					free(buflist, M_SEGMENT);
					vp->v_lastw = lbn;
					vp->v_lasta = bp->b_blkno;
					return;
				}
			}
		}
		/*
		 * Consider beginning a cluster.
		 * If at end of file, make cluster as large as possible,
		 * otherwise find size of existing cluster.
		 */
		if ((lbn + 1) * bp->b_bcount != filesize &&
		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen) ||
		     bp->b_blkno == -1)) {
			bawrite(bp);
			vp->v_clen = 0;
			vp->v_lasta = bp->b_blkno;
			vp->v_cstart = lbn + 1;
			vp->v_lastw = lbn;
			return;
		}
                vp->v_clen = maxclen;
                if (maxclen == 0) {		/* I/O not contiguous */
			vp->v_cstart = lbn + 1;
                        bawrite(bp);
                } else {			/* Wait for rest of cluster */
			vp->v_cstart = lbn;
                        bdwrite(bp);
		}
	} else if (lbn == vp->v_cstart + vp->v_clen) {
		/*
		 * At end of cluster, write it out.
		 */
		cluster_wbuild(vp, bp, bp->b_bcount, vp->v_cstart,
		    vp->v_clen + 1, lbn);
		vp->v_clen = 0;
		vp->v_cstart = lbn + 1;
	} else
		/*
		 * In the middle of a cluster, so just delay the
		 * I/O for now.
		 */
		bdwrite(bp);
	vp->v_lastw = lbn;
	vp->v_lasta = bp->b_blkno;
}


/*
 * This is an awful lot like cluster_rbuild...wish they could be combined.
 * The last lbn argument is the current block on which I/O is being
 * performed.  Check to see that it doesn't fall in the middle of
 * the current block (if last_bp == NULL).
 */
void
cluster_wbuild(vp, last_bp, size, start_lbn, len, lbn)
	struct vnode *vp;
	struct buf *last_bp;
	long size;
	daddr_t start_lbn;
	int len;
	daddr_t	lbn;
{
	struct cluster_save *b_save;
	struct buf *bp, *tbp;
	caddr_t	cp;
	int i, s;

#ifdef DIAGNOSTIC
	if (size != vp->v_mount->mnt_stat.f_iosize)
		panic("cluster_wbuild: size %d != filesize %d\n",
			size, vp->v_mount->mnt_stat.f_iosize);
#endif
redo:
	while ((!incore(vp, start_lbn) || start_lbn == lbn) && len) {
		++start_lbn;
		--len;
	}

	/* Get more memory for current buffer */
	if (len <= 1) {
		if (last_bp) {
			bawrite(last_bp);
		} else if (len) {
			bp = getblk(vp, start_lbn, size, 0, 0);
			bawrite(bp);
		}
		return;
	}

	bp = getblk(vp, start_lbn, size, 0, 0);
	if (!(bp->b_flags & B_DELWRI)) {
		++start_lbn;
		--len;
		brelse(bp);
		goto redo;
	}

	/*
	 * Extra memory in the buffer, punt on this buffer.
	 * XXX we could handle this in most cases, but we would have to
	 * push the extra memory down to after our max possible cluster
	 * size and then potentially pull it back up if the cluster was
	 * terminated prematurely--too much hassle.
	 */
	if (bp->b_bcount != bp->b_bufsize) {
		++start_lbn;
		--len;
		bawrite(bp);
		goto redo;
	}

	--len;
	b_save = malloc(sizeof(struct buf *) * len + sizeof(struct cluster_save),
	    M_SEGMENT, M_WAITOK);
	b_save->bs_bcount = bp->b_bcount;
	b_save->bs_bufsize = bp->b_bufsize;
	b_save->bs_nchildren = 0;
	b_save->bs_children = (struct buf **)(b_save + 1);
	b_save->bs_saveaddr = bp->b_saveaddr;
	bp->b_saveaddr = (caddr_t) b_save;

	bp->b_flags |= B_CALL;
	bp->b_iodone = cluster_callback;
	cp = (char *)bp->b_data + size;
	for (++start_lbn, i = 0; i < len; ++i, ++start_lbn) {
		/*
		 * Block is not in core or the non-sequential block
		 * ending our cluster was part of the cluster (in which
		 * case we don't want to write it twice).
		 */
		if (!incore(vp, start_lbn) ||
		    last_bp == NULL && start_lbn == lbn)
			break;

		/*
		 * Get the desired block buffer (unless it is the final
		 * sequential block whose buffer was passed in explictly
		 * as last_bp).
		 */
		if (last_bp == NULL || start_lbn != lbn) {
			tbp = getblk(vp, start_lbn, size, 0, 0);
			if (!(tbp->b_flags & B_DELWRI)) {
				brelse(tbp);
				break;
			}
		} else
			tbp = last_bp;

		++b_save->bs_nchildren;

		/* Move memory from children to parent */
		if (tbp->b_blkno != (bp->b_blkno + btodb(bp->b_bufsize))) {
			printf("Clustered Block: %d addr %x bufsize: %d\n",
			    bp->b_lblkno, bp->b_blkno, bp->b_bufsize);
			printf("Child Block: %d addr: %x\n", tbp->b_lblkno,
			    tbp->b_blkno);
			panic("Clustered write to wrong blocks");
		}

		pagemove(tbp->b_data, cp, size);
		bp->b_bcount += size;
		bp->b_bufsize += size;

		tbp->b_bufsize -= size;
		tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI);
		tbp->b_flags |= (B_ASYNC | B_AGE);
		s = splbio();
		reassignbuf(tbp, tbp->b_vp);		/* put on clean list */
		++tbp->b_vp->v_numoutput;
		splx(s);
		b_save->bs_children[i] = tbp;

		cp += size;
	}

	if (i == 0) {
		/* None to cluster */
		bp->b_saveaddr = b_save->bs_saveaddr;
		bp->b_flags &= ~B_CALL;
		bp->b_iodone = NULL;
		free(b_save, M_SEGMENT);
	}
	bawrite(bp);
	if (i < len) {
		len -= i + 1;
		start_lbn += 1;
		goto redo;
	}
}

/*
 * Collect together all the buffers in a cluster.
 * Plus add one additional buffer.
 */
struct cluster_save *
cluster_collectbufs(vp, last_bp)
	struct vnode *vp;
	struct buf *last_bp;
{
	struct cluster_save *buflist;
	daddr_t	lbn;
	int i, len;

	len = vp->v_lastw - vp->v_cstart + 1;
	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
	    M_SEGMENT, M_WAITOK);
	buflist->bs_nchildren = 0;
	buflist->bs_children = (struct buf **)(buflist + 1);
	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++)
		    (void)bread(vp, lbn, last_bp->b_bcount, NOCRED,
			&buflist->bs_children[i]);
	buflist->bs_children[i] = last_bp;
	buflist->bs_nchildren = i + 1;
	return (buflist);
}