vfs_cluster.c

早期freebsd实现

/*-
 * Copyright (c) 1993
 *	The Regents of the University of California.  All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 * 3. All advertising materials mentioning features or use of this software
 *    must display the following acknowledgement:
 *	This product includes software developed by the University of
 *	California, Berkeley and its contributors.
 * 4. Neither the name of the University nor the names of its contributors
 *    may be used to endorse or promote products derived from this software
 *    without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 *
 *	@(#)vfs_cluster.c	8.7 (Berkeley) 2/13/94
 */

#include <sys/param.h>
#include <sys/proc.h>
#include <sys/buf.h>
#include <sys/vnode.h>
#include <sys/mount.h>
#include <sys/trace.h>
#include <sys/malloc.h>
#include <sys/resourcevar.h>
#include <libkern/libkern.h>

#ifdef DEBUG
#include <vm/vm.h>
#include <sys/sysctl.h>
int doreallocblks = 1;
struct ctldebug debug13 = { "doreallocblks", &doreallocblks };
#else
/* XXX for cluster_write */
#define doreallocblks 1
#endif

/*
 * Local declarations
 */
struct buf *cluster_newbuf __P((struct vnode *, struct buf *, long, daddr_t,
	    daddr_t, long, int));
struct buf *cluster_rbuild __P((struct vnode *, u_quad_t, struct buf *,
	    daddr_t, daddr_t, long, int, long));
void	    cluster_wbuild __P((struct vnode *, struct buf *, long,
	    daddr_t, int, daddr_t));
struct cluster_save *cluster_collectbufs __P((struct vnode *, struct buf *));

#ifdef DIAGNOSTIC
/*
 * Set to 1 if reads of block zero should cause readahead to be done.
 * Set to 0 treats a read of block zero as a non-sequential read.
 *
 * Setting to one assumes that most reads of block zero of files are due to
 * sequential passes over the files (e.g. cat, sum) where additional blocks
 * will soon be needed.  Setting to zero assumes that the majority are
 * surgical strikes to get particular info (e.g. size, file) where readahead
 * blocks will not be used and, in fact, push out other potentially useful
 * blocks from the cache.  The former seems intuitive, but some quick tests
 * showed that the latter performed better from a system-wide point of view.
 */
int	doclusterraz = 0;
#define ISSEQREAD(vp, blk) \
	(((blk) != 0 || doclusterraz) && \
	 ((blk) == (vp)->v_lastr + 1 || (blk) == (vp)->v_lastr))
#else
#define ISSEQREAD(vp, blk) \
	((blk) != 0 && ((blk) == (vp)->v_lastr + 1 || (blk) == (vp)->v_lastr))
#endif

/*
 * This replaces bread.  If this is a bread at the beginning of a file and
 * lastr is 0, we assume this is the first read and we'll read up to two
 * blocks if they are sequential.  After that, we'll do regular read ahead
 * in clustered chunks.
 *
 * There are 4 or 5 cases depending on how you count:
 *	Desired block is in the cache:
 *	    1 Not sequential access (0 I/Os).
 *	    2 Access is sequential, do read-ahead (1 ASYNC).
 *	Desired block is not in cache:
 *	    3 Not sequential access (1 SYNC).
 *	    4 Sequential access, next block is contiguous (1 SYNC).
 *	    5 Sequential access, next block is not contiguous (1 SYNC, 1 ASYNC)
 *
 * There are potentially two buffers that require I/O.
 * 	bp is the block requested.
 *	rbp is the read-ahead block.
 *	If either is NULL, then you don't have to do the I/O.
 */
cluster_read(vp, filesize, lblkno, size, cred, bpp)
	struct vnode *vp;
	u_quad_t filesize;
	daddr_t lblkno;
	long size;
	struct ucred *cred;
	struct buf **bpp;
{
	struct buf *bp, *rbp;
	daddr_t blkno, ioblkno;
	long flags;
	int error, num_ra, alreadyincore;

#ifdef DIAGNOSTIC
	if (size == 0)
		panic("cluster_read: size = 0");
#endif

	error = 0;
	flags = B_READ;
	*bpp = bp = getblk(vp, lblkno, size, 0, 0);
	if (bp->b_flags & B_CACHE) {
		/*
		 * Desired block is in cache; do any readahead ASYNC.
		 * Case 1, 2.
		 */
		trace(TR_BREADHIT, pack(vp, size), lblkno);
		flags |= B_ASYNC;
		ioblkno = lblkno + (vp->v_ralen ? vp->v_ralen : 1);
		alreadyincore = (int)incore(vp, ioblkno);
		bp = NULL;
	} else {
		/* Block wasn't in cache, case 3, 4, 5. */
		trace(TR_BREADMISS, pack(vp, size), lblkno);
		bp->b_flags |= B_READ;
		ioblkno = lblkno;
		alreadyincore = 0;
		curproc->p_stats->p_ru.ru_inblock++;		/* XXX */
	}
	/*
	 * XXX
	 * Replace 1 with a window size based on some permutation of
	 * maxcontig and rot_delay.  This will let you figure out how
	 * many blocks you should read-ahead (case 2, 4, 5).
	 *
	 * If the access isn't sequential, reset the window to 1.
	 * Note that a read to the same block is considered sequential.
	 * This catches the case where the file is being read sequentially,
	 * but at smaller than the filesystem block size.
	 */
	rbp = NULL;
	if (!ISSEQREAD(vp, lblkno)) {
		vp->v_ralen = 0;
		vp->v_maxra = lblkno;
	} else if ((ioblkno + 1) * size <= filesize && !alreadyincore &&
	    !(error = VOP_BMAP(vp, ioblkno, NULL, &blkno, &num_ra)) &&
	    blkno != -1) {
		/*
		 * Reading sequentially, and the next block is not in the
		 * cache.  We are going to try reading ahead.
		 */
		if (num_ra) {
			/*
			 * If our desired readahead block had been read
			 * in a previous readahead but is no longer in
			 * core, then we may be reading ahead too far
			 * or are not using our readahead very rapidly.
			 * In this case we scale back the window.
			 */
			if (!alreadyincore && ioblkno <= vp->v_maxra)
				vp->v_ralen = max(vp->v_ralen >> 1, 1);
			/*
			 * There are more sequential blocks than our current
			 * window allows, scale up.  Ideally we want to get
			 * in sync with the filesystem maxcontig value.
			 */
			else if (num_ra > vp->v_ralen && lblkno != vp->v_lastr)
				vp->v_ralen = vp->v_ralen ?
					min(num_ra, vp->v_ralen << 1) : 1;

			if (num_ra > vp->v_ralen)
				num_ra = vp->v_ralen;
		}

		if (num_ra)				/* case 2, 4 */
			rbp = cluster_rbuild(vp, filesize,
			    bp, ioblkno, blkno, size, num_ra, flags);
		else if (ioblkno == lblkno) {
			bp->b_blkno = blkno;
			/* Case 5: check how many blocks to read ahead */
			++ioblkno;
			if ((ioblkno + 1) * size > filesize ||
			    incore(vp, ioblkno) || (error = VOP_BMAP(vp,
			     ioblkno, NULL, &blkno, &num_ra)) || blkno == -1)
				goto skip_readahead;
			/*
			 * Adjust readahead as above
			 */
			if (num_ra) {
				if (!alreadyincore && ioblkno <= vp->v_maxra)
					vp->v_ralen = max(vp->v_ralen >> 1, 1);
				else if (num_ra > vp->v_ralen &&
					 lblkno != vp->v_lastr)
					vp->v_ralen = vp->v_ralen ?
						min(num_ra,vp->v_ralen<<1) : 1;
				if (num_ra > vp->v_ralen)
					num_ra = vp->v_ralen;
			}
			flags |= B_ASYNC;
			if (num_ra)
				rbp = cluster_rbuild(vp, filesize,
				    NULL, ioblkno, blkno, size, num_ra, flags);
			else {
				rbp = getblk(vp, ioblkno, size, 0, 0);
				rbp->b_flags |= flags;
				rbp->b_blkno = blkno;
			}
		} else {
			/* case 2; read ahead single block */
			rbp = getblk(vp, ioblkno, size, 0, 0);
			rbp->b_flags |= flags;
			rbp->b_blkno = blkno;
		}

		if (rbp == bp)			/* case 4 */
			rbp = NULL;
		else if (rbp) {			/* case 2, 5 */
			trace(TR_BREADMISSRA,
			    pack(vp, (num_ra + 1) * size), ioblkno);
			curproc->p_stats->p_ru.ru_inblock++;	/* XXX */
		}
	}

	/* XXX Kirk, do we need to make sure the bp has creds? */
skip_readahead:
	if (bp)
		if (bp->b_flags & (B_DONE | B_DELWRI))
			panic("cluster_read: DONE bp");
		else 
			error = VOP_STRATEGY(bp);

	if (rbp)
		if (error || rbp->b_flags & (B_DONE | B_DELWRI)) {
			rbp->b_flags &= ~(B_ASYNC | B_READ);
			brelse(rbp);
		} else
			(void) VOP_STRATEGY(rbp);

	/*
	 * Recalculate our maximum readahead
	 */
	if (rbp == NULL)
		rbp = bp;
	if (rbp)
		vp->v_maxra = rbp->b_lblkno + (rbp->b_bufsize / size) - 1;

	if (bp)
		return(biowait(bp));
	return(error);
}

/*
 * If blocks are contiguous on disk, use this to provide clustered
 * read ahead.  We will read as many blocks as possible sequentially
 * and then parcel them up into logical blocks in the buffer hash table.
 */
struct buf *
cluster_rbuild(vp, filesize, bp, lbn, blkno, size, run, flags)
	struct vnode *vp;
	u_quad_t filesize;
	struct buf *bp;
	daddr_t lbn;
	daddr_t blkno;
	long size;
	int run;
	long flags;
{
	struct cluster_save *b_save;
	struct buf *tbp;
	daddr_t bn;
	int i, inc;

#ifdef DIAGNOSTIC
	if (size != vp->v_mount->mnt_stat.f_iosize)
		panic("cluster_rbuild: size %d != filesize %d\n",
			size, vp->v_mount->mnt_stat.f_iosize);
#endif
	if (size * (lbn + run + 1) > filesize)
		--run;
	if (run == 0) {
		if (!bp) {
			bp = getblk(vp, lbn, size, 0, 0);
			bp->b_blkno = blkno;
			bp->b_flags |= flags;
		}
		return(bp);
	}

	bp = cluster_newbuf(vp, bp, flags, blkno, lbn, size, run + 1);
	if (bp->b_flags & (B_DONE | B_DELWRI))
		return (bp);

	b_save = malloc(sizeof(struct buf *) * run + sizeof(struct cluster_save),
	    M_SEGMENT, M_WAITOK);
	b_save->bs_bufsize = b_save->bs_bcount = size;
	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;

	inc = btodb(size);
	for (bn = blkno + inc, i = 1; i <= run; ++i, bn += inc) {
		if (incore(vp, lbn + i)) {
			if (i == 1) {
				bp->b_saveaddr = b_save->bs_saveaddr;
				bp->b_flags &= ~B_CALL;
				bp->b_iodone = NULL;
				allocbuf(bp, size);
				free(b_save, M_SEGMENT);
			} else
				allocbuf(bp, size * i);
			break;
		}
		tbp = getblk(vp, lbn + i, 0, 0, 0);
		/*
		 * getblk may return some memory in the buffer if there were
		 * no empty buffers to shed it to.  If there is currently
		 * memory in the buffer, we move it down size bytes to make
		 * room for the valid pages that cluster_callback will insert.
		 * We do this now so we don't have to do it at interrupt time
		 * in the callback routine.
		 */
		if (tbp->b_bufsize != 0) {
			caddr_t bdata = (char *)tbp->b_data;

			if (tbp->b_bufsize + size > MAXBSIZE)
				panic("cluster_rbuild: too much memory");
			if (tbp->b_bufsize > size) {
				/*
				 * XXX if the source and destination regions
				 * overlap we have to copy backward to avoid
				 * clobbering any valid pages (i.e. pagemove
				 * implementations typically can't handle
				 * overlap).
				 */
				bdata += tbp->b_bufsize;
				while (bdata > (char *)tbp->b_data) {
					bdata -= CLBYTES;
					pagemove(bdata, bdata + size, CLBYTES);
				}
			} else 
				pagemove(bdata, bdata + size, tbp->b_bufsize);
		}
		tbp->b_blkno = bn;
		tbp->b_flags |= flags | B_READ | B_ASYNC;
		++b_save->bs_nchildren;
		b_save->bs_children[i - 1] = tbp;
	}
	return(bp);
}

/*
 * Either get a new buffer or grow the existing one.
 */
struct buf *
cluster_newbuf(vp, bp, flags, blkno, lblkno, size, run)