vfs_bio.c

来自「基于组件方式开发操作系统的OSKIT源代码」· C语言 代码 · 共 993 行 · 第 1/2 页

 */
void
brelse(bp)
	struct buf *bp;
{
	struct bqueues *bufq;
	int s;

	KASSERT(ISSET(bp->b_flags, B_BUSY));

	/* Wake up any processes waiting for any buffer to become free. */
	if (needbuffer) {
		needbuffer = 0;
		wakeup(&needbuffer);
	}

	/* Block disk interrupts. */
	s = splbio();

	/* Wake up any proceeses waiting for _this_ buffer to become free. */
	if (ISSET(bp->b_flags, B_WANTED)) {
		CLR(bp->b_flags, B_WANTED|B_AGE);
		wakeup(bp);
	}

	/*
	 * Determine which queue the buffer should be on, then put it there.
	 */

	/* If it's locked, don't report an error; try again later. */
	if (ISSET(bp->b_flags, (B_LOCKED|B_ERROR)) == (B_LOCKED|B_ERROR))
		CLR(bp->b_flags, B_ERROR);

	/* If it's not cacheable, or an error, mark it invalid. */
	if (ISSET(bp->b_flags, (B_NOCACHE|B_ERROR)))
		SET(bp->b_flags, B_INVAL);

	if (ISSET(bp->b_flags, B_VFLUSH)) {
		/*
		 * This is a delayed write buffer that was just flushed to
		 * disk.  It is still on the LRU queue.  If it's become
		 * invalid, then we need to move it to a different queue;
		 * otherwise leave it in its current position.
		 */
		CLR(bp->b_flags, B_VFLUSH);
		if (!ISSET(bp->b_flags, B_ERROR|B_INVAL|B_LOCKED|B_AGE))
			goto already_queued;
		else
			bremfree(bp);
	}

	if ((bp->b_bufsize <= 0) || ISSET(bp->b_flags, B_INVAL)) {
		/*
		 * If it's invalid or empty, dissociate it from its vnode
		 * and put on the head of the appropriate queue.
		 */
		if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
			(*bioops.io_deallocate)(bp);
		CLR(bp->b_flags, B_DONE|B_DELWRI);
		if (bp->b_vp) {
			reassignbuf(bp, bp->b_vp);
			brelvp(bp);
		}
		if (bp->b_bufsize <= 0)
			/* no data */
			bufq = &bufqueues[BQ_EMPTY];
		else
			/* invalid data */
			bufq = &bufqueues[BQ_AGE];
		binsheadfree(bp, bufq);
	} else {
		/*
		 * It has valid data.  Put it on the end of the appropriate
		 * queue, so that it'll stick around for as long as possible.
		 * If buf is AGE, but has dependencies, must put it on last
		 * bufqueue to be scanned, ie LRU. This protects against the
		 * livelock where BQ_AGE only has buffers with dependencies,
		 * and we thus never get to the dependent buffers in BQ_LRU.
		 */
		if (ISSET(bp->b_flags, B_LOCKED))
			/* locked in core */
			bufq = &bufqueues[BQ_LOCKED];
		else if (!ISSET(bp->b_flags, B_AGE))
			/* valid data */
			bufq = &bufqueues[BQ_LRU];
		else {
			/* stale but valid data */
			int has_deps;

			if (LIST_FIRST(&bp->b_dep) != NULL &&
			    bioops.io_countdeps)
				has_deps = (*bioops.io_countdeps)(bp, 0);
			else
				has_deps = 0;
			bufq = has_deps ? &bufqueues[BQ_LRU] :
			    &bufqueues[BQ_AGE];
		}
		binstailfree(bp, bufq);
	}

already_queued:
	/* Unlock the buffer. */
	CLR(bp->b_flags, B_AGE|B_ASYNC|B_BUSY|B_NOCACHE|B_ORDERED);
	SET(bp->b_flags, B_CACHE);

	/* Allow disk interrupts. */
	splx(s);
}


/*
 * Determine if a block is in the cache.
 * Just look on what would be its hash chain.  If it's there, return
 * a pointer to it, unless it's marked invalid.  If it's marked invalid,
 * we normally don't return the buffer, unless the caller explicitly
 * wants us to.
 */
struct buf *
incore(vp, blkno)
	struct vnode *vp;
	daddr_t blkno;
{
	struct buf *bp;

	bp = BUFHASH(vp, blkno)->lh_first;

	/* Search hash chain */
	for (; bp != NULL; bp = bp->b_hash.le_next) {
		if (bp->b_lblkno == blkno && bp->b_vp == vp &&
		    !ISSET(bp->b_flags, B_INVAL))
		return (bp);
	}

	return (NULL);
}

/*
 * Get a block of requested size that is associated with
 * a given vnode and block offset. If it is found in the
 * block cache, mark it as having been found, make it busy
 * and return it. Otherwise, return an empty block of the
 * correct size. It is up to the caller to insure that the
 * cached blocks be of the correct size.
 */
struct buf *
getblk(vp, blkno, size, slpflag, slptimeo)
	struct vnode *vp;
	daddr_t blkno;
	int size, slpflag, slptimeo;
{
	struct buf *bp;
	int s, err;

start:
	bp = incore(vp, blkno);
	if (bp != NULL) {
		s = splbio();
		if (ISSET(bp->b_flags, B_BUSY)) {
			if (curproc == uvm.pagedaemon_proc) {
				splx(s);
				return NULL;
			}
			SET(bp->b_flags, B_WANTED);
			err = tsleep(bp, slpflag | (PRIBIO + 1), "getblk",
				     slptimeo);
			splx(s);
			if (err)
				return (NULL);
			goto start;
		}
#ifdef DIAGNOSTIC
		if (ISSET(bp->b_flags, B_DONE|B_DELWRI) && bp->b_bcount < size)
			panic("getblk: block size invariant failed");
#endif
		SET(bp->b_flags, B_BUSY);
		bremfree(bp);
		splx(s);
	} else {
		if ((bp = getnewbuf(slpflag, slptimeo)) == NULL)
			goto start;

		binshash(bp, BUFHASH(vp, blkno));
		bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno;
		s = splbio();
		bgetvp(vp, bp);
		splx(s);
	}
	allocbuf(bp, size);
	return (bp);
}

/*
 * Get an empty, disassociated buffer of given size.
 */
struct buf *
geteblk(size)
	int size;
{
	struct buf *bp; 

	while ((bp = getnewbuf(0, 0)) == 0)
		;
	SET(bp->b_flags, B_INVAL);
	binshash(bp, &invalhash);
	allocbuf(bp, size);
	return (bp);
}

/*
 * Expand or contract the actual memory allocated to a buffer.
 *
 * If the buffer shrinks, data is lost, so it's up to the
 * caller to have written it out *first*; this routine will not
 * start a write.  If the buffer grows, it's the callers
 * responsibility to fill out the buffer's additional contents.
 */
void
allocbuf(bp, size)
	struct buf *bp;
	int size;
{
	struct buf *nbp;
	vsize_t desired_size;
	int s;

	desired_size = round_page((vsize_t)size);
	if (desired_size > MAXBSIZE)
		panic("allocbuf: buffer larger than MAXBSIZE requested");

	if (bp->b_bufsize == desired_size)
		goto out;

	/*
	 * If the buffer is smaller than the desired size, we need to snarf
	 * it from other buffers.  Get buffers (via getnewbuf()), and
	 * steal their pages.
	 */
	while (bp->b_bufsize < desired_size) {
		int amt;

		/* find a buffer */
		while ((nbp = getnewbuf(0, 0)) == NULL)
			;

		SET(nbp->b_flags, B_INVAL);
		binshash(nbp, &invalhash);

		/* and steal its pages, up to the amount we need */
		amt = min(nbp->b_bufsize, (desired_size - bp->b_bufsize));
		pagemove((nbp->b_data + nbp->b_bufsize - amt),
			 bp->b_data + bp->b_bufsize, amt);
		bp->b_bufsize += amt;
		nbp->b_bufsize -= amt;

		/* reduce transfer count if we stole some data */
		if (nbp->b_bcount > nbp->b_bufsize)
			nbp->b_bcount = nbp->b_bufsize;

#ifdef DIAGNOSTIC
		if (nbp->b_bufsize < 0)
			panic("allocbuf: negative bufsize");
#endif

		brelse(nbp);
	}

	/*
	 * If we want a buffer smaller than the current size,
	 * shrink this buffer.  Grab a buf head from the EMPTY queue,
	 * move a page onto it, and put it on front of the AGE queue.
	 * If there are no free buffer headers, leave the buffer alone.
	 */
	if (bp->b_bufsize > desired_size) {
		s = splbio();
		if ((nbp = bufqueues[BQ_EMPTY].tqh_first) == NULL) {
			/* No free buffer head */
			splx(s);
			goto out;
		}
		bremfree(nbp);
		SET(nbp->b_flags, B_BUSY);
		splx(s);

		/* move the page to it and note this change */
		pagemove(bp->b_data + desired_size,
		    nbp->b_data, bp->b_bufsize - desired_size);
		nbp->b_bufsize = bp->b_bufsize - desired_size;
		bp->b_bufsize = desired_size;
		nbp->b_bcount = 0;
		SET(nbp->b_flags, B_INVAL);

		/* release the newly-filled buffer and leave */
		brelse(nbp);
	}

out:
	bp->b_bcount = size;
}

/*
 * Find a buffer which is available for use.
 * Select something from a free list.
 * Preference is to AGE list, then LRU list.    
 */
struct buf *
getnewbuf(slpflag, slptimeo)
	int slpflag, slptimeo;
{
	struct buf *bp;
	int s;

start:
	s = splbio();
	if ((bp = bufqueues[BQ_AGE].tqh_first) != NULL ||
	    (bp = bufqueues[BQ_LRU].tqh_first) != NULL) {
		bremfree(bp);
	} else {
		/* wait for a free buffer of any kind */
		needbuffer = 1;
		tsleep(&needbuffer, slpflag|(PRIBIO+1), "getnewbuf", slptimeo);
		splx(s);
		return (NULL);
	}

	if (ISSET(bp->b_flags, B_VFLUSH)) {
		/*
		 * This is a delayed write buffer being flushed to disk.  Make
		 * sure it gets aged out of the queue when it's finished, and
		 * leave it off the LRU queue.
		 */
		CLR(bp->b_flags, B_VFLUSH);
		SET(bp->b_flags, B_AGE);
		splx(s);
		goto start;
	}

	/* Buffer is no longer on free lists. */
	SET(bp->b_flags, B_BUSY);

	/* If buffer was a delayed write, start it, and go back to the top. */
	if (ISSET(bp->b_flags, B_DELWRI)) {
		splx(s);
		/*
		 * This buffer has gone through the LRU, so make sure it gets
		 * reused ASAP.
		 */
		SET(bp->b_flags, B_AGE);
		bawrite(bp);
		goto start;
	}

	/* disassociate us from our vnode, if we had one... */
	if (bp->b_vp)
		brelvp(bp);
	splx(s);

	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
		(*bioops.io_deallocate)(bp);

	/* clear out various other fields */
	bp->b_flags = B_BUSY;
	bp->b_dev = NODEV;
	bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = 0;
	bp->b_iodone = 0;
	bp->b_error = 0;
	bp->b_resid = 0;
	bp->b_bcount = 0;
	
	bremhash(bp);
	return (bp); 
}
#endif /*OSKIT*/

/*
 * Wait for operations on the buffer to complete.
 * When they do, extract and return the I/O's error value.
 */
int
biowait(bp)
	struct buf *bp;
{
	int s;
	
	s = splbio();
	while (!ISSET(bp->b_flags, B_DONE))
		tsleep(bp, PRIBIO + 1, "biowait", 0);
	splx(s);

	/* check for interruption of I/O (e.g. via NFS), then errors. */
	if (ISSET(bp->b_flags, B_EINTR)) {
		CLR(bp->b_flags, B_EINTR);
		return (EINTR);
	} else if (ISSET(bp->b_flags, B_ERROR))
		return (bp->b_error ? bp->b_error : EIO);
	else
		return (0);
}

/*
 * Mark I/O complete on a buffer.
 *
 * If a callback has been requested, e.g. the pageout
 * daemon, do so. Otherwise, awaken waiting processes.
 *
 * [ Leffler, et al., says on p.247:
 *	"This routine wakes up the blocked process, frees the buffer
 *	for an asynchronous write, or, for a request by the pagedaemon
 *	process, invokes a procedure specified in the buffer structure" ]
 *
 * In real life, the pagedaemon (or other system processes) wants
 * to do async stuff to, and doesn't want the buffer brelse()'d.
 * (for swap pager, that puts swap buffers on the free lists (!!!),
 * for the vn device, that puts malloc'd buffers on the free lists!)
 */
void
biodone(bp)
	struct buf *bp;
{
	int s = splbio();

	if (ISSET(bp->b_flags, B_DONE))
		panic("biodone already");
	SET(bp->b_flags, B_DONE);		/* note that it's done */

	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
		(*bioops.io_complete)(bp);

	if (!ISSET(bp->b_flags, B_READ))	/* wake up reader */
		vwakeup(bp);

	if (ISSET(bp->b_flags, B_CALL)) {	/* if necessary, call out */
		CLR(bp->b_flags, B_CALL);	/* but note callout done */
		(*bp->b_iodone)(bp);
	} else {
		if (ISSET(bp->b_flags, B_ASYNC))	/* if async, release */
			brelse(bp);
		else {				/* or just wakeup the buffer */
			CLR(bp->b_flags, B_WANTED);
			wakeup(bp);
		}
	}

	splx(s);
}

#ifndef OSKIT
/*
 * Return a count of buffers on the "locked" queue.
 */
int
count_lock_queue()
{
	struct buf *bp;
	int n = 0;

	for (bp = bufqueues[BQ_LOCKED].tqh_first; bp;
	    bp = bp->b_freelist.tqe_next)
		n++;
	return (n);
}

#ifdef DEBUG
/*
 * Print out statistics on the current allocation of the buffer pool.
 * Can be enabled to print out on every ``sync'' by setting "syncprt"
 * in vfs_syscalls.c using sysctl.
 */
void
vfs_bufstats()
{
	int s, i, j, count;
	struct buf *bp;
	struct bqueues *dp;
	int counts[(MAXBSIZE / PAGE_SIZE) + 1];
	static char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE", "EMPTY" };

	for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) {
		count = 0;
		for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
			counts[j] = 0;
		s = splbio();
		for (bp = dp->tqh_first; bp; bp = bp->b_freelist.tqe_next) {
			counts[bp->b_bufsize/PAGE_SIZE]++;
			count++;
		}
		splx(s);
		printf("%s: total-%d", bname[i], count);
		for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
			if (counts[j] != 0)
				printf(", %d-%d", j * PAGE_SIZE, counts[j]);
		printf("\n");
	}
}
#endif /* DEBUG */
#endif /*OSKIT*/