xref: /netbsd-src/sys/kern/vfs_bio.c (revision 481fca6e59249d8ffcf24fef7cfbe7b131bfb080)
1 /*	$NetBSD: vfs_bio.c,v 1.68 2000/06/27 17:41:47 mrg Exp $	*/
2 
3 /*-
4  * Copyright (c) 1994 Christopher G. Demetriou
5  * Copyright (c) 1982, 1986, 1989, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  * (c) UNIX System Laboratories, Inc.
8  * All or some portions of this file are derived from material licensed
9  * to the University of California by American Telephone and Telegraph
10  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
11  * the permission of UNIX System Laboratories, Inc.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. All advertising materials mentioning features or use of this software
22  *    must display the following acknowledgement:
23  *	This product includes software developed by the University of
24  *	California, Berkeley and its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	@(#)vfs_bio.c	8.6 (Berkeley) 1/11/94
42  */
43 
44 /*
45  * Some references:
46  *	Bach: The Design of the UNIX Operating System (Prentice Hall, 1986)
47  *	Leffler, et al.: The Design and Implementation of the 4.3BSD
48  *		UNIX Operating System (Addison Welley, 1989)
49  */
50 
51 #include <sys/param.h>
52 #include <sys/systm.h>
53 #include <sys/proc.h>
54 #include <sys/buf.h>
55 #include <sys/vnode.h>
56 #include <sys/mount.h>
57 #include <sys/trace.h>
58 #include <sys/malloc.h>
59 #include <sys/resourcevar.h>
60 #include <sys/conf.h>
61 
62 #include <miscfs/specfs/specdev.h>
63 
64 /* Macros to clear/set/test flags. */
65 #define	SET(t, f)	(t) |= (f)
66 #define	CLR(t, f)	(t) &= ~(f)
67 #define	ISSET(t, f)	((t) & (f))
68 
69 /*
70  * Definitions for the buffer hash lists.
71  */
72 #define	BUFHASH(dvp, lbn)	\
73 	(&bufhashtbl[((long)(dvp) / sizeof(*(dvp)) + (int)(lbn)) & bufhash])
74 LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
75 u_long	bufhash;
76 struct bio_ops bioops;	/* I/O operation notification */
77 
78 /*
79  * Insq/Remq for the buffer hash lists.
80  */
81 #define	binshash(bp, dp)	LIST_INSERT_HEAD(dp, bp, b_hash)
82 #define	bremhash(bp)		LIST_REMOVE(bp, b_hash)
83 
84 /*
85  * Definitions for the buffer free lists.
86  */
87 #define	BQUEUES		4		/* number of free buffer queues */
88 
89 #define	BQ_LOCKED	0		/* super-blocks &c */
90 #define	BQ_LRU		1		/* lru, useful buffers */
91 #define	BQ_AGE		2		/* rubbish */
92 #define	BQ_EMPTY	3		/* buffer headers with no memory */
93 
94 TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES];
95 int needbuffer;
96 
97 /*
98  * Buffer pool for I/O buffers.
99  */
100 struct pool bufpool;
101 
102 /*
103  * Insq/Remq for the buffer free lists.
104  */
105 #define	binsheadfree(bp, dp)	TAILQ_INSERT_HEAD(dp, bp, b_freelist)
106 #define	binstailfree(bp, dp)	TAILQ_INSERT_TAIL(dp, bp, b_freelist)
107 
108 static __inline struct buf *bio_doread __P((struct vnode *, daddr_t, int,
109 					    struct ucred *, int));
110 int count_lock_queue __P((void));
111 
112 void
113 bremfree(bp)
114 	struct buf *bp;
115 {
116 	int s = splbio();
117 
118 	struct bqueues *dp = NULL;
119 
120 	/*
121 	 * We only calculate the head of the freelist when removing
122 	 * the last element of the list as that is the only time that
123 	 * it is needed (e.g. to reset the tail pointer).
124 	 *
125 	 * NB: This makes an assumption about how tailq's are implemented.
126 	 */
127 	if (bp->b_freelist.tqe_next == NULL) {
128 		for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++)
129 			if (dp->tqh_last == &bp->b_freelist.tqe_next)
130 				break;
131 		if (dp == &bufqueues[BQUEUES])
132 			panic("bremfree: lost tail");
133 	}
134 	TAILQ_REMOVE(dp, bp, b_freelist);
135 
136 	splx(s);
137 }
138 
139 /*
140  * Initialize buffers and hash links for buffers.
141  */
142 void
143 bufinit()
144 {
145 	struct buf *bp;
146 	struct bqueues *dp;
147 	int i;
148 	int base, residual;
149 
150 	/*
151 	 * Initialize the buffer pool.  This pool is used for buffers
152 	 * which are strictly I/O control blocks, not buffer cache
153 	 * buffers.
154 	 */
155 	pool_init(&bufpool, sizeof(struct buf), 0, 0, 0, "bufpl", 0,
156 	    NULL, NULL, M_DEVBUF);
157 
158 	for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++)
159 		TAILQ_INIT(dp);
160 	bufhashtbl = hashinit(nbuf, M_CACHE, M_WAITOK, &bufhash);
161 	base = bufpages / nbuf;
162 	residual = bufpages % nbuf;
163 	for (i = 0; i < nbuf; i++) {
164 		bp = &buf[i];
165 		memset((char *)bp, 0, sizeof(*bp));
166 		bp->b_dev = NODEV;
167 		bp->b_rcred = NOCRED;
168 		bp->b_wcred = NOCRED;
169 		bp->b_vnbufs.le_next = NOLIST;
170 		LIST_INIT(&bp->b_dep);
171 		bp->b_data = buffers + i * MAXBSIZE;
172 		if (i < residual)
173 			bp->b_bufsize = (base + 1) * NBPG;
174 		else
175 			bp->b_bufsize = base * NBPG;
176 		bp->b_flags = B_INVAL;
177 		dp = bp->b_bufsize ? &bufqueues[BQ_AGE] : &bufqueues[BQ_EMPTY];
178 		binsheadfree(bp, dp);
179 		binshash(bp, &invalhash);
180 	}
181 }
182 
183 static __inline struct buf *
184 bio_doread(vp, blkno, size, cred, async)
185 	struct vnode *vp;
186 	daddr_t blkno;
187 	int size;
188 	struct ucred *cred;
189 	int async;
190 {
191 	struct buf *bp;
192 	struct proc *p = (curproc != NULL ? curproc : &proc0);	/* XXX */
193 
194 	bp = getblk(vp, blkno, size, 0, 0);
195 
196 	/*
197 	 * If buffer does not have data valid, start a read.
198 	 * Note that if buffer is B_INVAL, getblk() won't return it.
199 	 * Therefore, it's valid if it's I/O has completed or been delayed.
200 	 */
201 	if (!ISSET(bp->b_flags, (B_DONE | B_DELWRI))) {
202 		/* Start I/O for the buffer (keeping credentials). */
203 		SET(bp->b_flags, B_READ | async);
204 		if (cred != NOCRED && bp->b_rcred == NOCRED) {
205 			crhold(cred);
206 			bp->b_rcred = cred;
207 		}
208 		VOP_STRATEGY(bp);
209 
210 		/* Pay for the read. */
211 		p->p_stats->p_ru.ru_inblock++;
212 	} else if (async) {
213 		brelse(bp);
214 	}
215 
216 	return (bp);
217 }
218 
219 /*
220  * Read a disk block.
221  * This algorithm described in Bach (p.54).
222  */
223 int
224 bread(vp, blkno, size, cred, bpp)
225 	struct vnode *vp;
226 	daddr_t blkno;
227 	int size;
228 	struct ucred *cred;
229 	struct buf **bpp;
230 {
231 	struct buf *bp;
232 
233 	/* Get buffer for block. */
234 	bp = *bpp = bio_doread(vp, blkno, size, cred, 0);
235 
236 	/*
237 	 * Delayed write buffers are found in the cache and have
238 	 * valid contents. Also, B_ERROR is not set, otherwise
239 	 * getblk() would not have returned them.
240 	 */
241 	if (ISSET(bp->b_flags, B_DONE|B_DELWRI))
242 		return (0);
243 
244 	/*
245 	 * Otherwise, we had to start a read for it; wait until
246 	 * it's valid and return the result.
247 	 */
248 	return (biowait(bp));
249 }
250 
251 /*
252  * Read-ahead multiple disk blocks. The first is sync, the rest async.
253  * Trivial modification to the breada algorithm presented in Bach (p.55).
254  */
255 int
256 breadn(vp, blkno, size, rablks, rasizes, nrablks, cred, bpp)
257 	struct vnode *vp;
258 	daddr_t blkno; int size;
259 	daddr_t rablks[]; int rasizes[];
260 	int nrablks;
261 	struct ucred *cred;
262 	struct buf **bpp;
263 {
264 	struct buf *bp;
265 	int i;
266 
267 	bp = *bpp = bio_doread(vp, blkno, size, cred, 0);
268 
269 	/*
270 	 * For each of the read-ahead blocks, start a read, if necessary.
271 	 */
272 	for (i = 0; i < nrablks; i++) {
273 		/* If it's in the cache, just go on to next one. */
274 		if (incore(vp, rablks[i]))
275 			continue;
276 
277 		/* Get a buffer for the read-ahead block */
278 		(void) bio_doread(vp, rablks[i], rasizes[i], cred, B_ASYNC);
279 	}
280 
281 	/*
282 	 * Delayed write buffers are found in the cache and have
283 	 * valid contents. Also, B_ERROR is not set, otherwise
284 	 * getblk() would not have returned them.
285 	 */
286 	if (ISSET(bp->b_flags, B_DONE|B_DELWRI))
287 		return (0);
288 
289 	/*
290 	 * Otherwise, we had to start a read for it; wait until
291 	 * it's valid and return the result.
292 	 */
293 	return (biowait(bp));
294 }
295 
296 /*
297  * Read with single-block read-ahead.  Defined in Bach (p.55), but
298  * implemented as a call to breadn().
299  * XXX for compatibility with old file systems.
300  */
301 int
302 breada(vp, blkno, size, rablkno, rabsize, cred, bpp)
303 	struct vnode *vp;
304 	daddr_t blkno; int size;
305 	daddr_t rablkno; int rabsize;
306 	struct ucred *cred;
307 	struct buf **bpp;
308 {
309 
310 	return (breadn(vp, blkno, size, &rablkno, &rabsize, 1, cred, bpp));
311 }
312 
313 /*
314  * Block write.  Described in Bach (p.56)
315  */
316 int
317 bwrite(bp)
318 	struct buf *bp;
319 {
320 	int rv, sync, wasdelayed, s;
321 	struct proc *p = (curproc != NULL ? curproc : &proc0);	/* XXX */
322 	struct vnode *vp;
323 	struct mount *mp;
324 
325 	/*
326 	 * Remember buffer type, to switch on it later.  If the write was
327 	 * synchronous, but the file system was mounted with MNT_ASYNC,
328 	 * convert it to a delayed write.
329 	 * XXX note that this relies on delayed tape writes being converted
330 	 * to async, not sync writes (which is safe, but ugly).
331 	 */
332 	sync = !ISSET(bp->b_flags, B_ASYNC);
333 	if (sync && bp->b_vp && bp->b_vp->v_mount &&
334 	    ISSET(bp->b_vp->v_mount->mnt_flag, MNT_ASYNC)) {
335 		bdwrite(bp);
336 		return (0);
337 	}
338 
339 	/*
340 	 * Collect statistics on synchronous and asynchronous writes.
341 	 * Writes to block devices are charged to their associated
342 	 * filesystem (if any).
343 	 */
344 	if ((vp = bp->b_vp) != NULL) {
345 		if (vp->v_type == VBLK)
346 			mp = vp->v_specmountpoint;
347 		else
348 			mp = vp->v_mount;
349 		if (mp != NULL) {
350 			if (sync)
351 				mp->mnt_stat.f_syncwrites++;
352 			else
353 				mp->mnt_stat.f_asyncwrites++;
354 		}
355 	}
356 
357 	wasdelayed = ISSET(bp->b_flags, B_DELWRI);
358 
359 	s = splbio();
360 
361 	CLR(bp->b_flags, (B_READ | B_DONE | B_ERROR | B_DELWRI));
362 
363 	/*
364 	 * Pay for the I/O operation and make sure the buf is on the correct
365 	 * vnode queue.
366 	 */
367 	if (wasdelayed)
368 		reassignbuf(bp, bp->b_vp);
369 	else
370 		p->p_stats->p_ru.ru_oublock++;
371 
372 	/* Initiate disk write.  Make sure the appropriate party is charged. */
373 	bp->b_vp->v_numoutput++;
374 	splx(s);
375 
376 	SET(bp->b_flags, B_WRITEINPROG);
377 	VOP_STRATEGY(bp);
378 
379 	if (sync) {
380 		/* If I/O was synchronous, wait for it to complete. */
381 		rv = biowait(bp);
382 
383 		/* Release the buffer. */
384 		brelse(bp);
385 
386 		return (rv);
387 	} else {
388 		return (0);
389 	}
390 }
391 
392 int
393 vn_bwrite(v)
394 	void *v;
395 {
396 	struct vop_bwrite_args *ap = v;
397 
398 	return (bwrite(ap->a_bp));
399 }
400 
401 /*
402  * Delayed write.
403  *
404  * The buffer is marked dirty, but is not queued for I/O.
405  * This routine should be used when the buffer is expected
406  * to be modified again soon, typically a small write that
407  * partially fills a buffer.
408  *
409  * NB: magnetic tapes cannot be delayed; they must be
410  * written in the order that the writes are requested.
411  *
412  * Described in Leffler, et al. (pp. 208-213).
413  */
414 void
415 bdwrite(bp)
416 	struct buf *bp;
417 {
418 	struct proc *p = (curproc != NULL ? curproc : &proc0);	/* XXX */
419 	int s;
420 
421 	/* If this is a tape block, write the block now. */
422 	/* XXX NOTE: the memory filesystem usurpes major device */
423 	/* XXX       number 255, which is a bad idea.		*/
424 	if (bp->b_dev != NODEV &&
425 	    major(bp->b_dev) != 255 &&	/* XXX - MFS buffers! */
426 	    bdevsw[major(bp->b_dev)].d_type == D_TAPE) {
427 		bawrite(bp);
428 		return;
429 	}
430 
431 	/*
432 	 * If the block hasn't been seen before:
433 	 *	(1) Mark it as having been seen,
434 	 *	(2) Charge for the write,
435 	 *	(3) Make sure it's on its vnode's correct block list.
436 	 */
437 	s = splbio();
438 
439 	if (!ISSET(bp->b_flags, B_DELWRI)) {
440 		SET(bp->b_flags, B_DELWRI);
441 		p->p_stats->p_ru.ru_oublock++;
442 		reassignbuf(bp, bp->b_vp);
443 	}
444 
445 	/* Otherwise, the "write" is done, so mark and release the buffer. */
446 	CLR(bp->b_flags, B_NEEDCOMMIT|B_DONE);
447 	splx(s);
448 
449 	brelse(bp);
450 }
451 
452 /*
453  * Asynchronous block write; just an asynchronous bwrite().
454  */
455 void
456 bawrite(bp)
457 	struct buf *bp;
458 {
459 
460 	SET(bp->b_flags, B_ASYNC);
461 	VOP_BWRITE(bp);
462 }
463 
464 /*
465  * Ordered block write; asynchronous, but I/O will occur in order queued.
466  */
467 void
468 bowrite(bp)
469 	struct buf *bp;
470 {
471 
472 	SET(bp->b_flags, B_ASYNC | B_ORDERED);
473 	VOP_BWRITE(bp);
474 }
475 
476 /*
477  * Same as first half of bdwrite, mark buffer dirty, but do not release it.
478  */
479 void
480 bdirty(bp)
481 	struct buf *bp;
482 {
483 	struct proc *p = (curproc != NULL ? curproc : &proc0);	/* XXX */
484 	int s;
485 
486 	s = splbio();
487 
488 	CLR(bp->b_flags, B_AGE);
489 
490 	if (!ISSET(bp->b_flags, B_DELWRI)) {
491 		SET(bp->b_flags, B_DELWRI);
492 		p->p_stats->p_ru.ru_oublock++;
493 		reassignbuf(bp, bp->b_vp);
494 	}
495 
496 	splx(s);
497 }
498 
499 /*
500  * Release a buffer on to the free lists.
501  * Described in Bach (p. 46).
502  */
503 void
504 brelse(bp)
505 	struct buf *bp;
506 {
507 	struct bqueues *bufq;
508 	int s;
509 
510 	/* Wake up any processes waiting for any buffer to become free. */
511 	if (needbuffer) {
512 		needbuffer = 0;
513 		wakeup(&needbuffer);
514 	}
515 
516 	/* Block disk interrupts. */
517 	s = splbio();
518 
519 	/* Wake up any proceeses waiting for _this_ buffer to become free. */
520 	if (ISSET(bp->b_flags, B_WANTED)) {
521 		CLR(bp->b_flags, B_WANTED|B_AGE);
522 		wakeup(bp);
523 	}
524 
525 	/*
526 	 * Determine which queue the buffer should be on, then put it there.
527 	 */
528 
529 	/* If it's locked, don't report an error; try again later. */
530 	if (ISSET(bp->b_flags, (B_LOCKED|B_ERROR)) == (B_LOCKED|B_ERROR))
531 		CLR(bp->b_flags, B_ERROR);
532 
533 	/* If it's not cacheable, or an error, mark it invalid. */
534 	if (ISSET(bp->b_flags, (B_NOCACHE|B_ERROR)))
535 		SET(bp->b_flags, B_INVAL);
536 
537 	if (ISSET(bp->b_flags, B_VFLUSH)) {
538 		/*
539 		 * This is a delayed write buffer that was just flushed to
540 		 * disk.  It is still on the LRU queue.  If it's become
541 		 * invalid, then we need to move it to a different queue;
542 		 * otherwise leave it in its current position.
543 		 */
544 		CLR(bp->b_flags, B_VFLUSH);
545 		if (!ISSET(bp->b_flags, B_ERROR|B_INVAL|B_LOCKED|B_AGE))
546 			goto already_queued;
547 		else
548 			bremfree(bp);
549 	}
550 
551 	if ((bp->b_bufsize <= 0) || ISSET(bp->b_flags, B_INVAL)) {
552 		/*
553 		 * If it's invalid or empty, dissociate it from its vnode
554 		 * and put on the head of the appropriate queue.
555 		 */
556 		if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
557 			(*bioops.io_deallocate)(bp);
558 		CLR(bp->b_flags, B_DONE|B_DELWRI);
559 		if (bp->b_vp) {
560 			reassignbuf(bp, bp->b_vp);
561 			brelvp(bp);
562 		}
563 		if (bp->b_bufsize <= 0)
564 			/* no data */
565 			bufq = &bufqueues[BQ_EMPTY];
566 		else
567 			/* invalid data */
568 			bufq = &bufqueues[BQ_AGE];
569 		binsheadfree(bp, bufq);
570 	} else {
571 		/*
572 		 * It has valid data.  Put it on the end of the appropriate
573 		 * queue, so that it'll stick around for as long as possible.
574 		 * If buf is AGE, but has dependencies, must put it on last
575 		 * bufqueue to be scanned, ie LRU. This protects against the
576 		 * livelock where BQ_AGE only has buffers with dependencies,
577 		 * and we thus never get to the dependent buffers in BQ_LRU.
578 		 */
579 		if (ISSET(bp->b_flags, B_LOCKED))
580 			/* locked in core */
581 			bufq = &bufqueues[BQ_LOCKED];
582 		else if (!ISSET(bp->b_flags, B_AGE))
583 			/* valid data */
584 			bufq = &bufqueues[BQ_LRU];
585 		else {
586 			/* stale but valid data */
587 			int has_deps;
588 
589 			if (LIST_FIRST(&bp->b_dep) != NULL &&
590 			    bioops.io_countdeps)
591 				has_deps = (*bioops.io_countdeps)(bp, 0);
592 			else
593 				has_deps = 0;
594 			bufq = has_deps ? &bufqueues[BQ_LRU] :
595 			    &bufqueues[BQ_AGE];
596 		}
597 		binstailfree(bp, bufq);
598 	}
599 
600 already_queued:
601 	/* Unlock the buffer. */
602 	CLR(bp->b_flags, B_AGE|B_ASYNC|B_BUSY|B_NOCACHE|B_ORDERED);
603 
604 	/* Allow disk interrupts. */
605 	splx(s);
606 }
607 
608 /*
609  * Determine if a block is in the cache.
610  * Just look on what would be its hash chain.  If it's there, return
611  * a pointer to it, unless it's marked invalid.  If it's marked invalid,
612  * we normally don't return the buffer, unless the caller explicitly
613  * wants us to.
614  */
615 struct buf *
616 incore(vp, blkno)
617 	struct vnode *vp;
618 	daddr_t blkno;
619 {
620 	struct buf *bp;
621 
622 	bp = BUFHASH(vp, blkno)->lh_first;
623 
624 	/* Search hash chain */
625 	for (; bp != NULL; bp = bp->b_hash.le_next) {
626 		if (bp->b_lblkno == blkno && bp->b_vp == vp &&
627 		    !ISSET(bp->b_flags, B_INVAL))
628 		return (bp);
629 	}
630 
631 	return (0);
632 }
633 
634 /*
635  * Get a block of requested size that is associated with
636  * a given vnode and block offset. If it is found in the
637  * block cache, mark it as having been found, make it busy
638  * and return it. Otherwise, return an empty block of the
639  * correct size. It is up to the caller to insure that the
640  * cached blocks be of the correct size.
641  */
642 struct buf *
643 getblk(vp, blkno, size, slpflag, slptimeo)
644 	struct vnode *vp;
645 	daddr_t blkno;
646 	int size, slpflag, slptimeo;
647 {
648 	struct bufhashhdr *bh;
649 	struct buf *bp;
650 	int s, err;
651 
652 	/*
653 	 * XXX
654 	 * The following is an inlined version of 'incore()', but with
655 	 * the 'invalid' test moved to after the 'busy' test.  It's
656 	 * necessary because there are some cases in which the NFS
657 	 * code sets B_INVAL prior to writing data to the server, but
658 	 * in which the buffers actually contain valid data.  In this
659 	 * case, we can't allow the system to allocate a new buffer for
660 	 * the block until the write is finished.
661 	 */
662 	bh = BUFHASH(vp, blkno);
663 start:
664         bp = bh->lh_first;
665         for (; bp != NULL; bp = bp->b_hash.le_next) {
666                 if (bp->b_lblkno != blkno || bp->b_vp != vp)
667 			continue;
668 
669 		s = splbio();
670 		if (ISSET(bp->b_flags, B_BUSY)) {
671 			SET(bp->b_flags, B_WANTED);
672 			err = tsleep(bp, slpflag | (PRIBIO + 1), "getblk",
673 			    slptimeo);
674 			splx(s);
675 			if (err)
676 				return (NULL);
677 			goto start;
678 		}
679 
680 		if (!ISSET(bp->b_flags, B_INVAL)) {
681 #ifdef DIAGNOSTIC
682 			if (ISSET(bp->b_flags, B_DONE|B_DELWRI) &&
683 			    bp->b_bcount < size)
684 				panic("getblk: block size invariant failed");
685 #endif
686 			SET(bp->b_flags, B_BUSY);
687 			bremfree(bp);
688 			splx(s);
689 			break;
690 		}
691 		splx(s);
692         }
693 
694 	if (bp == NULL) {
695 		if ((bp = getnewbuf(slpflag, slptimeo)) == NULL)
696 			goto start;
697 		binshash(bp, bh);
698 		bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno;
699 		s = splbio();
700 		bgetvp(vp, bp);
701 		splx(s);
702 	}
703 	allocbuf(bp, size);
704 	return (bp);
705 }
706 
707 /*
708  * Get an empty, disassociated buffer of given size.
709  */
710 struct buf *
711 geteblk(size)
712 	int size;
713 {
714 	struct buf *bp;
715 
716 	while ((bp = getnewbuf(0, 0)) == 0)
717 		;
718 	SET(bp->b_flags, B_INVAL);
719 	binshash(bp, &invalhash);
720 	allocbuf(bp, size);
721 
722 	return (bp);
723 }
724 
725 /*
726  * Expand or contract the actual memory allocated to a buffer.
727  *
728  * If the buffer shrinks, data is lost, so it's up to the
729  * caller to have written it out *first*; this routine will not
730  * start a write.  If the buffer grows, it's the callers
731  * responsibility to fill out the buffer's additional contents.
732  */
733 void
734 allocbuf(bp, size)
735 	struct buf *bp;
736 	int size;
737 {
738 	struct buf      *nbp;
739 	vsize_t       desired_size;
740 	int	     s;
741 
742 	desired_size = roundup(size, NBPG);
743 	if (desired_size > MAXBSIZE)
744 		panic("allocbuf: buffer larger than MAXBSIZE requested");
745 
746 	if (bp->b_bufsize == desired_size)
747 		goto out;
748 
749 	/*
750 	 * If the buffer is smaller than the desired size, we need to snarf
751 	 * it from other buffers.  Get buffers (via getnewbuf()), and
752 	 * steal their pages.
753 	 */
754 	while (bp->b_bufsize < desired_size) {
755 		int amt;
756 
757 		/* find a buffer */
758 		while ((nbp = getnewbuf(0, 0)) == NULL)
759 			;
760 		SET(nbp->b_flags, B_INVAL);
761 		binshash(nbp, &invalhash);
762 
763 		/* and steal its pages, up to the amount we need */
764 		amt = min(nbp->b_bufsize, (desired_size - bp->b_bufsize));
765 		pagemove((nbp->b_data + nbp->b_bufsize - amt),
766 			 bp->b_data + bp->b_bufsize, amt);
767 		bp->b_bufsize += amt;
768 		nbp->b_bufsize -= amt;
769 
770 		/* reduce transfer count if we stole some data */
771 		if (nbp->b_bcount > nbp->b_bufsize)
772 			nbp->b_bcount = nbp->b_bufsize;
773 
774 #ifdef DIAGNOSTIC
775 		if (nbp->b_bufsize < 0)
776 			panic("allocbuf: negative bufsize");
777 #endif
778 
779 		brelse(nbp);
780 	}
781 
782 	/*
783 	 * If we want a buffer smaller than the current size,
784 	 * shrink this buffer.  Grab a buf head from the EMPTY queue,
785 	 * move a page onto it, and put it on front of the AGE queue.
786 	 * If there are no free buffer headers, leave the buffer alone.
787 	 */
788 	if (bp->b_bufsize > desired_size) {
789 		s = splbio();
790 		if ((nbp = bufqueues[BQ_EMPTY].tqh_first) == NULL) {
791 			/* No free buffer head */
792 			splx(s);
793 			goto out;
794 		}
795 		bremfree(nbp);
796 		SET(nbp->b_flags, B_BUSY);
797 		splx(s);
798 
799 		/* move the page to it and note this change */
800 		pagemove(bp->b_data + desired_size,
801 		    nbp->b_data, bp->b_bufsize - desired_size);
802 		nbp->b_bufsize = bp->b_bufsize - desired_size;
803 		bp->b_bufsize = desired_size;
804 		nbp->b_bcount = 0;
805 		SET(nbp->b_flags, B_INVAL);
806 
807 		/* release the newly-filled buffer and leave */
808 		brelse(nbp);
809 	}
810 
811 out:
812 	bp->b_bcount = size;
813 }
814 
815 /*
816  * Find a buffer which is available for use.
817  * Select something from a free list.
818  * Preference is to AGE list, then LRU list.
819  */
820 struct buf *
821 getnewbuf(slpflag, slptimeo)
822 	int slpflag, slptimeo;
823 {
824 	struct buf *bp;
825 	int s;
826 
827 start:
828 	s = splbio();
829 	if ((bp = bufqueues[BQ_AGE].tqh_first) != NULL ||
830 	    (bp = bufqueues[BQ_LRU].tqh_first) != NULL) {
831 		bremfree(bp);
832 	} else {
833 		/* wait for a free buffer of any kind */
834 		needbuffer = 1;
835 		tsleep(&needbuffer, slpflag|(PRIBIO+1), "getnewbuf", slptimeo);
836 		splx(s);
837 		return (0);
838 	}
839 
840 	if (ISSET(bp->b_flags, B_VFLUSH)) {
841 		/*
842 		 * This is a delayed write buffer being flushed to disk.  Make
843 		 * sure it gets aged out of the queue when it's finished, and
844 		 * leave it off the LRU queue.
845 		 */
846 		CLR(bp->b_flags, B_VFLUSH);
847 		SET(bp->b_flags, B_AGE);
848 		splx(s);
849 		goto start;
850 	}
851 
852 	/* Buffer is no longer on free lists. */
853 	SET(bp->b_flags, B_BUSY);
854 
855 	/* If buffer was a delayed write, start it, and go back to the top. */
856 	if (ISSET(bp->b_flags, B_DELWRI)) {
857 		splx(s);
858 		/*
859 		 * This buffer has gone through the LRU, so make sure it gets
860 		 * reused ASAP.
861 		 */
862 		SET(bp->b_flags, B_AGE);
863 		bawrite(bp);
864 		goto start;
865 	}
866 
867 	/* disassociate us from our vnode, if we had one... */
868 	if (bp->b_vp)
869 		brelvp(bp);
870 	splx(s);
871 
872 	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
873 		(*bioops.io_deallocate)(bp);
874 
875 	/* clear out various other fields */
876 	bp->b_flags = B_BUSY;
877 	bp->b_dev = NODEV;
878 	bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = 0;
879 	bp->b_iodone = 0;
880 	bp->b_error = 0;
881 	bp->b_resid = 0;
882 	bp->b_bcount = 0;
883 	bp->b_dirtyoff = bp->b_dirtyend = 0;
884 	bp->b_validoff = bp->b_validend = 0;
885 
886 	/* nuke any credentials we were holding */
887 	if (bp->b_rcred != NOCRED) {
888 		crfree(bp->b_rcred);
889 		bp->b_rcred = NOCRED;
890 	}
891 	if (bp->b_wcred != NOCRED) {
892 		crfree(bp->b_wcred);
893 		bp->b_wcred = NOCRED;
894 	}
895 
896 	bremhash(bp);
897 	return (bp);
898 }
899 
900 /*
901  * Wait for operations on the buffer to complete.
902  * When they do, extract and return the I/O's error value.
903  */
904 int
905 biowait(bp)
906 	struct buf *bp;
907 {
908 	int s;
909 
910 	s = splbio();
911 	while (!ISSET(bp->b_flags, B_DONE))
912 		tsleep(bp, PRIBIO + 1, "biowait", 0);
913 	splx(s);
914 
915 	/* check for interruption of I/O (e.g. via NFS), then errors. */
916 	if (ISSET(bp->b_flags, B_EINTR)) {
917 		CLR(bp->b_flags, B_EINTR);
918 		return (EINTR);
919 	} else if (ISSET(bp->b_flags, B_ERROR))
920 		return (bp->b_error ? bp->b_error : EIO);
921 	else
922 		return (0);
923 }
924 
925 /*
926  * Mark I/O complete on a buffer.
927  *
928  * If a callback has been requested, e.g. the pageout
929  * daemon, do so. Otherwise, awaken waiting processes.
930  *
931  * [ Leffler, et al., says on p.247:
932  *	"This routine wakes up the blocked process, frees the buffer
933  *	for an asynchronous write, or, for a request by the pagedaemon
934  *	process, invokes a procedure specified in the buffer structure" ]
935  *
936  * In real life, the pagedaemon (or other system processes) wants
937  * to do async stuff to, and doesn't want the buffer brelse()'d.
938  * (for swap pager, that puts swap buffers on the free lists (!!!),
939  * for the vn device, that puts malloc'd buffers on the free lists!)
940  */
941 void
942 biodone(bp)
943 	struct buf *bp;
944 {
945 	int s = splbio();
946 
947 	if (ISSET(bp->b_flags, B_DONE))
948 		panic("biodone already");
949 	SET(bp->b_flags, B_DONE);		/* note that it's done */
950 
951 	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
952 		(*bioops.io_complete)(bp);
953 
954 	if (!ISSET(bp->b_flags, B_READ))	/* wake up reader */
955 		vwakeup(bp);
956 
957 	if (ISSET(bp->b_flags, B_CALL)) {	/* if necessary, call out */
958 		CLR(bp->b_flags, B_CALL);	/* but note callout done */
959 		(*bp->b_iodone)(bp);
960 	} else {
961 		if (ISSET(bp->b_flags, B_ASYNC))	/* if async, release */
962 			brelse(bp);
963 		else {				/* or just wakeup the buffer */
964 			CLR(bp->b_flags, B_WANTED);
965 			wakeup(bp);
966 		}
967 	}
968 
969 	splx(s);
970 }
971 
972 /*
973  * Return a count of buffers on the "locked" queue.
974  */
975 int
976 count_lock_queue()
977 {
978 	struct buf *bp;
979 	int n = 0;
980 
981 	for (bp = bufqueues[BQ_LOCKED].tqh_first; bp;
982 	    bp = bp->b_freelist.tqe_next)
983 		n++;
984 	return (n);
985 }
986 
987 #ifdef DEBUG
988 /*
989  * Print out statistics on the current allocation of the buffer pool.
990  * Can be enabled to print out on every ``sync'' by setting "syncprt"
991  * in vfs_syscalls.c using sysctl.
992  */
993 void
994 vfs_bufstats()
995 {
996 	int s, i, j, count;
997 	struct buf *bp;
998 	struct bqueues *dp;
999 	int counts[MAXBSIZE/NBPG+1];
1000 	static char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE", "EMPTY" };
1001 
1002 	for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) {
1003 		count = 0;
1004 		for (j = 0; j <= MAXBSIZE/NBPG; j++)
1005 			counts[j] = 0;
1006 		s = splbio();
1007 		for (bp = dp->tqh_first; bp; bp = bp->b_freelist.tqe_next) {
1008 			counts[bp->b_bufsize/NBPG]++;
1009 			count++;
1010 		}
1011 		splx(s);
1012 		printf("%s: total-%d", bname[i], count);
1013 		for (j = 0; j <= MAXBSIZE/NBPG; j++)
1014 			if (counts[j] != 0)
1015 				printf(", %d-%d", j * NBPG, counts[j]);
1016 		printf("\n");
1017 	}
1018 }
1019 #endif /* DEBUG */
1020