xref: /netbsd-src/sys/kern/vfs_bio.c (revision 89c5a767f8fc7a4633b2d409966e2becbb98ff92)
1 /*	$NetBSD: vfs_bio.c,v 1.65 2000/02/14 20:12:03 thorpej 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 <vm/vm.h>
63 
64 #include <miscfs/specfs/specdev.h>
65 
66 /* Macros to clear/set/test flags. */
67 #define	SET(t, f)	(t) |= (f)
68 #define	CLR(t, f)	(t) &= ~(f)
69 #define	ISSET(t, f)	((t) & (f))
70 
71 /*
72  * Definitions for the buffer hash lists.
73  */
74 #define	BUFHASH(dvp, lbn)	\
75 	(&bufhashtbl[((long)(dvp) / sizeof(*(dvp)) + (int)(lbn)) & bufhash])
76 LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
77 u_long	bufhash;
78 struct bio_ops bioops;	/* I/O operation notification */
79 
80 /*
81  * Insq/Remq for the buffer hash lists.
82  */
83 #define	binshash(bp, dp)	LIST_INSERT_HEAD(dp, bp, b_hash)
84 #define	bremhash(bp)		LIST_REMOVE(bp, b_hash)
85 
86 /*
87  * Definitions for the buffer free lists.
88  */
89 #define	BQUEUES		4		/* number of free buffer queues */
90 
91 #define	BQ_LOCKED	0		/* super-blocks &c */
92 #define	BQ_LRU		1		/* lru, useful buffers */
93 #define	BQ_AGE		2		/* rubbish */
94 #define	BQ_EMPTY	3		/* buffer headers with no memory */
95 
96 TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES];
97 int needbuffer;
98 
99 /*
100  * Buffer pool for I/O buffers.
101  */
102 struct pool bufpool;
103 
104 /*
105  * Insq/Remq for the buffer free lists.
106  */
107 #define	binsheadfree(bp, dp)	TAILQ_INSERT_HEAD(dp, bp, b_freelist)
108 #define	binstailfree(bp, dp)	TAILQ_INSERT_TAIL(dp, bp, b_freelist)
109 
110 static __inline struct buf *bio_doread __P((struct vnode *, daddr_t, int,
111 					    struct ucred *, int));
112 int count_lock_queue __P((void));
113 
114 void
115 bremfree(bp)
116 	struct buf *bp;
117 {
118 	int s = splbio();
119 
120 	struct bqueues *dp = NULL;
121 
122 	/*
123 	 * We only calculate the head of the freelist when removing
124 	 * the last element of the list as that is the only time that
125 	 * it is needed (e.g. to reset the tail pointer).
126 	 *
127 	 * NB: This makes an assumption about how tailq's are implemented.
128 	 */
129 	if (bp->b_freelist.tqe_next == NULL) {
130 		for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++)
131 			if (dp->tqh_last == &bp->b_freelist.tqe_next)
132 				break;
133 		if (dp == &bufqueues[BQUEUES])
134 			panic("bremfree: lost tail");
135 	}
136 	TAILQ_REMOVE(dp, bp, b_freelist);
137 
138 	splx(s);
139 }
140 
141 /*
142  * Initialize buffers and hash links for buffers.
143  */
144 void
145 bufinit()
146 {
147 	register struct buf *bp;
148 	struct bqueues *dp;
149 	register int i;
150 	int base, residual;
151 
152 	/*
153 	 * Initialize the buffer pool.  This pool is used for buffers
154 	 * which are strictly I/O control blocks, not buffer cache
155 	 * buffers.
156 	 */
157 	pool_init(&bufpool, sizeof(struct buf), 0, 0, 0, "bufpl", 0,
158 	    NULL, NULL, M_DEVBUF);
159 
160 	for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++)
161 		TAILQ_INIT(dp);
162 	bufhashtbl = hashinit(nbuf, M_CACHE, M_WAITOK, &bufhash);
163 	base = bufpages / nbuf;
164 	residual = bufpages % nbuf;
165 	for (i = 0; i < nbuf; i++) {
166 		bp = &buf[i];
167 		memset((char *)bp, 0, sizeof(*bp));
168 		bp->b_dev = NODEV;
169 		bp->b_rcred = NOCRED;
170 		bp->b_wcred = NOCRED;
171 		bp->b_vnbufs.le_next = NOLIST;
172 		LIST_INIT(&bp->b_dep);
173 		bp->b_data = buffers + i * MAXBSIZE;
174 		if (i < residual)
175 			bp->b_bufsize = (base + 1) * NBPG;
176 		else
177 			bp->b_bufsize = base * NBPG;
178 		bp->b_flags = B_INVAL;
179 		dp = bp->b_bufsize ? &bufqueues[BQ_AGE] : &bufqueues[BQ_EMPTY];
180 		binsheadfree(bp, dp);
181 		binshash(bp, &invalhash);
182 	}
183 }
184 
185 static __inline struct buf *
186 bio_doread(vp, blkno, size, cred, async)
187 	struct vnode *vp;
188 	daddr_t blkno;
189 	int size;
190 	struct ucred *cred;
191 	int async;
192 {
193 	register struct buf *bp;
194 	struct proc *p = (curproc != NULL ? curproc : &proc0);	/* XXX */
195 
196 	bp = getblk(vp, blkno, size, 0, 0);
197 
198 	/*
199 	 * If buffer does not have data valid, start a read.
200 	 * Note that if buffer is B_INVAL, getblk() won't return it.
201 	 * Therefore, it's valid if it's I/O has completed or been delayed.
202 	 */
203 	if (!ISSET(bp->b_flags, (B_DONE | B_DELWRI))) {
204 		/* Start I/O for the buffer (keeping credentials). */
205 		SET(bp->b_flags, B_READ | async);
206 		if (cred != NOCRED && bp->b_rcred == NOCRED) {
207 			crhold(cred);
208 			bp->b_rcred = cred;
209 		}
210 		VOP_STRATEGY(bp);
211 
212 		/* Pay for the read. */
213 		p->p_stats->p_ru.ru_inblock++;
214 	} else if (async) {
215 		brelse(bp);
216 	}
217 
218 	return (bp);
219 }
220 
221 /*
222  * Read a disk block.
223  * This algorithm described in Bach (p.54).
224  */
225 int
226 bread(vp, blkno, size, cred, bpp)
227 	struct vnode *vp;
228 	daddr_t blkno;
229 	int size;
230 	struct ucred *cred;
231 	struct buf **bpp;
232 {
233 	register struct buf *bp;
234 
235 	/* Get buffer for block. */
236 	bp = *bpp = bio_doread(vp, blkno, size, cred, 0);
237 
238 	/*
239 	 * Delayed write buffers are found in the cache and have
240 	 * valid contents. Also, B_ERROR is not set, otherwise
241 	 * getblk() would not have returned them.
242 	 */
243 	if (ISSET(bp->b_flags, B_DONE|B_DELWRI))
244 		return (0);
245 
246 	/*
247 	 * Otherwise, we had to start a read for it; wait until
248 	 * it's valid and return the result.
249 	 */
250 	return (biowait(bp));
251 }
252 
253 /*
254  * Read-ahead multiple disk blocks. The first is sync, the rest async.
255  * Trivial modification to the breada algorithm presented in Bach (p.55).
256  */
257 int
258 breadn(vp, blkno, size, rablks, rasizes, nrablks, cred, bpp)
259 	struct vnode *vp;
260 	daddr_t blkno; int size;
261 	daddr_t rablks[]; int rasizes[];
262 	int nrablks;
263 	struct ucred *cred;
264 	struct buf **bpp;
265 {
266 	register struct buf *bp;
267 	int i;
268 
269 	bp = *bpp = bio_doread(vp, blkno, size, cred, 0);
270 
271 	/*
272 	 * For each of the read-ahead blocks, start a read, if necessary.
273 	 */
274 	for (i = 0; i < nrablks; i++) {
275 		/* If it's in the cache, just go on to next one. */
276 		if (incore(vp, rablks[i]))
277 			continue;
278 
279 		/* Get a buffer for the read-ahead block */
280 		(void) bio_doread(vp, rablks[i], rasizes[i], cred, B_ASYNC);
281 	}
282 
283 	/*
284 	 * Delayed write buffers are found in the cache and have
285 	 * valid contents. Also, B_ERROR is not set, otherwise
286 	 * getblk() would not have returned them.
287 	 */
288 	if (ISSET(bp->b_flags, B_DONE|B_DELWRI))
289 		return (0);
290 
291 	/*
292 	 * Otherwise, we had to start a read for it; wait until
293 	 * it's valid and return the result.
294 	 */
295 	return (biowait(bp));
296 }
297 
298 /*
299  * Read with single-block read-ahead.  Defined in Bach (p.55), but
300  * implemented as a call to breadn().
301  * XXX for compatibility with old file systems.
302  */
303 int
304 breada(vp, blkno, size, rablkno, rabsize, cred, bpp)
305 	struct vnode *vp;
306 	daddr_t blkno; int size;
307 	daddr_t rablkno; int rabsize;
308 	struct ucred *cred;
309 	struct buf **bpp;
310 {
311 
312 	return (breadn(vp, blkno, size, &rablkno, &rabsize, 1, cred, bpp));
313 }
314 
315 /*
316  * Block write.  Described in Bach (p.56)
317  */
318 int
319 bwrite(bp)
320 	struct buf *bp;
321 {
322 	int rv, sync, wasdelayed, s;
323 	struct proc *p = (curproc != NULL ? curproc : &proc0);	/* XXX */
324 	struct vnode *vp;
325 	struct mount *mp;
326 
327 	/*
328 	 * Remember buffer type, to switch on it later.  If the write was
329 	 * synchronous, but the file system was mounted with MNT_ASYNC,
330 	 * convert it to a delayed write.
331 	 * XXX note that this relies on delayed tape writes being converted
332 	 * to async, not sync writes (which is safe, but ugly).
333 	 */
334 	sync = !ISSET(bp->b_flags, B_ASYNC);
335 	if (sync && bp->b_vp && bp->b_vp->v_mount &&
336 	    ISSET(bp->b_vp->v_mount->mnt_flag, MNT_ASYNC)) {
337 		bdwrite(bp);
338 		return (0);
339 	}
340 
341 	/*
342 	 * Collect statistics on synchronous and asynchronous writes.
343 	 * Writes to block devices are charged to their associated
344 	 * filesystem (if any).
345 	 */
346 	if ((vp = bp->b_vp) != NULL) {
347 		if (vp->v_type == VBLK)
348 			mp = vp->v_specmountpoint;
349 		else
350 			mp = vp->v_mount;
351 		if (mp != NULL) {
352 			if (sync)
353 				mp->mnt_stat.f_syncwrites++;
354 			else
355 				mp->mnt_stat.f_asyncwrites++;
356 		}
357 	}
358 
359 	wasdelayed = ISSET(bp->b_flags, B_DELWRI);
360 
361 	s = splbio();
362 
363 	CLR(bp->b_flags, (B_READ | B_DONE | B_ERROR | B_DELWRI));
364 
365 	/*
366 	 * Pay for the I/O operation and make sure the buf is on the correct
367 	 * vnode queue.
368 	 */
369 	if (wasdelayed)
370 		reassignbuf(bp, bp->b_vp);
371 	else
372 		p->p_stats->p_ru.ru_oublock++;
373 
374 	/* Initiate disk write.  Make sure the appropriate party is charged. */
375 	bp->b_vp->v_numoutput++;
376 	splx(s);
377 
378 	SET(bp->b_flags, B_WRITEINPROG);
379 	VOP_STRATEGY(bp);
380 
381 	if (sync) {
382 		/* If I/O was synchronous, wait for it to complete. */
383 		rv = biowait(bp);
384 
385 		/* Release the buffer. */
386 		brelse(bp);
387 
388 		return (rv);
389 	} else {
390 		return (0);
391 	}
392 }
393 
394 int
395 vn_bwrite(v)
396 	void *v;
397 {
398 	struct vop_bwrite_args *ap = v;
399 
400 	return (bwrite(ap->a_bp));
401 }
402 
403 /*
404  * Delayed write.
405  *
406  * The buffer is marked dirty, but is not queued for I/O.
407  * This routine should be used when the buffer is expected
408  * to be modified again soon, typically a small write that
409  * partially fills a buffer.
410  *
411  * NB: magnetic tapes cannot be delayed; they must be
412  * written in the order that the writes are requested.
413  *
414  * Described in Leffler, et al. (pp. 208-213).
415  */
416 void
417 bdwrite(bp)
418 	struct buf *bp;
419 {
420 	struct proc *p = (curproc != NULL ? curproc : &proc0);	/* XXX */
421 	int s;
422 
423 	/* If this is a tape block, write the block now. */
424 	/* XXX NOTE: the memory filesystem usurpes major device */
425 	/* XXX       number 255, which is a bad idea.		*/
426 	if (bp->b_dev != NODEV &&
427 	    major(bp->b_dev) != 255 &&	/* XXX - MFS buffers! */
428 	    bdevsw[major(bp->b_dev)].d_type == D_TAPE) {
429 		bawrite(bp);
430 		return;
431 	}
432 
433 	/*
434 	 * If the block hasn't been seen before:
435 	 *	(1) Mark it as having been seen,
436 	 *	(2) Charge for the write,
437 	 *	(3) Make sure it's on its vnode's correct block list.
438 	 */
439 	s = splbio();
440 
441 	if (!ISSET(bp->b_flags, B_DELWRI)) {
442 		SET(bp->b_flags, B_DELWRI);
443 		p->p_stats->p_ru.ru_oublock++;
444 		reassignbuf(bp, bp->b_vp);
445 	}
446 
447 	/* Otherwise, the "write" is done, so mark and release the buffer. */
448 	CLR(bp->b_flags, B_NEEDCOMMIT|B_DONE);
449 	splx(s);
450 
451 	brelse(bp);
452 }
453 
454 /*
455  * Asynchronous block write; just an asynchronous bwrite().
456  */
457 void
458 bawrite(bp)
459 	struct buf *bp;
460 {
461 
462 	SET(bp->b_flags, B_ASYNC);
463 	VOP_BWRITE(bp);
464 }
465 
466 /*
467  * Ordered block write; asynchronous, but I/O will occur in order queued.
468  */
469 void
470 bowrite(bp)
471 	struct buf *bp;
472 {
473 
474 	SET(bp->b_flags, B_ASYNC | B_ORDERED);
475 	VOP_BWRITE(bp);
476 }
477 
478 /*
479  * Same as first half of bdwrite, mark buffer dirty, but do not release it.
480  */
481 void
482 bdirty(bp)
483 	struct buf *bp;
484 {
485 	struct proc *p = (curproc != NULL ? curproc : &proc0);	/* XXX */
486 	int s;
487 
488 	s = splbio();
489 
490 	CLR(bp->b_flags, B_AGE);
491 
492 	if (!ISSET(bp->b_flags, B_DELWRI)) {
493 		SET(bp->b_flags, B_DELWRI);
494 		p->p_stats->p_ru.ru_oublock++;
495 		reassignbuf(bp, bp->b_vp);
496 	}
497 
498 	splx(s);
499 }
500 
501 /*
502  * Release a buffer on to the free lists.
503  * Described in Bach (p. 46).
504  */
505 void
506 brelse(bp)
507 	struct buf *bp;
508 {
509 	struct bqueues *bufq;
510 	int s;
511 
512 	/* Wake up any processes waiting for any buffer to become free. */
513 	if (needbuffer) {
514 		needbuffer = 0;
515 		wakeup(&needbuffer);
516 	}
517 
518 	/* Block disk interrupts. */
519 	s = splbio();
520 
521 	/* Wake up any proceeses waiting for _this_ buffer to become free. */
522 	if (ISSET(bp->b_flags, B_WANTED)) {
523 		CLR(bp->b_flags, B_WANTED|B_AGE);
524 		wakeup(bp);
525 	}
526 
527 	/*
528 	 * Determine which queue the buffer should be on, then put it there.
529 	 */
530 
531 	/* If it's locked, don't report an error; try again later. */
532 	if (ISSET(bp->b_flags, (B_LOCKED|B_ERROR)) == (B_LOCKED|B_ERROR))
533 		CLR(bp->b_flags, B_ERROR);
534 
535 	/* If it's not cacheable, or an error, mark it invalid. */
536 	if (ISSET(bp->b_flags, (B_NOCACHE|B_ERROR)))
537 		SET(bp->b_flags, B_INVAL);
538 
539 	if (ISSET(bp->b_flags, B_VFLUSH)) {
540 		/*
541 		 * This is a delayed write buffer that was just flushed to
542 		 * disk.  It is still on the LRU queue.  If it's become
543 		 * invalid, then we need to move it to a different queue;
544 		 * otherwise leave it in its current position.
545 		 */
546 		CLR(bp->b_flags, B_VFLUSH);
547 		if (!ISSET(bp->b_flags, B_ERROR|B_INVAL|B_LOCKED|B_AGE))
548 			goto already_queued;
549 		else
550 			bremfree(bp);
551 	}
552 
553 	if ((bp->b_bufsize <= 0) || ISSET(bp->b_flags, B_INVAL)) {
554 		/*
555 		 * If it's invalid or empty, dissociate it from its vnode
556 		 * and put on the head of the appropriate queue.
557 		 */
558 		if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
559 			(*bioops.io_deallocate)(bp);
560 		CLR(bp->b_flags, B_DONE|B_DELWRI);
561 		if (bp->b_vp) {
562 			reassignbuf(bp, bp->b_vp);
563 			brelvp(bp);
564 		}
565 		if (bp->b_bufsize <= 0)
566 			/* no data */
567 			bufq = &bufqueues[BQ_EMPTY];
568 		else
569 			/* invalid data */
570 			bufq = &bufqueues[BQ_AGE];
571 		binsheadfree(bp, bufq);
572 	} else {
573 		/*
574 		 * It has valid data.  Put it on the end of the appropriate
575 		 * queue, so that it'll stick around for as long as possible.
576 		 */
577 		if (ISSET(bp->b_flags, B_LOCKED))
578 			/* locked in core */
579 			bufq = &bufqueues[BQ_LOCKED];
580 		else if (ISSET(bp->b_flags, B_AGE))
581 			/* stale but valid data */
582 			bufq = &bufqueues[BQ_AGE];
583 		else
584 			/* valid data */
585 			bufq = &bufqueues[BQ_LRU];
586 		binstailfree(bp, bufq);
587 	}
588 
589 already_queued:
590 	/* Unlock the buffer. */
591 	CLR(bp->b_flags, B_AGE|B_ASYNC|B_BUSY|B_NOCACHE|B_ORDERED);
592 
593 	/* Allow disk interrupts. */
594 	splx(s);
595 }
596 
597 /*
598  * Determine if a block is in the cache.
599  * Just look on what would be its hash chain.  If it's there, return
600  * a pointer to it, unless it's marked invalid.  If it's marked invalid,
601  * we normally don't return the buffer, unless the caller explicitly
602  * wants us to.
603  */
604 struct buf *
605 incore(vp, blkno)
606 	struct vnode *vp;
607 	daddr_t blkno;
608 {
609 	struct buf *bp;
610 
611 	bp = BUFHASH(vp, blkno)->lh_first;
612 
613 	/* Search hash chain */
614 	for (; bp != NULL; bp = bp->b_hash.le_next) {
615 		if (bp->b_lblkno == blkno && bp->b_vp == vp &&
616 		    !ISSET(bp->b_flags, B_INVAL))
617 		return (bp);
618 	}
619 
620 	return (0);
621 }
622 
623 /*
624  * Get a block of requested size that is associated with
625  * a given vnode and block offset. If it is found in the
626  * block cache, mark it as having been found, make it busy
627  * and return it. Otherwise, return an empty block of the
628  * correct size. It is up to the caller to insure that the
629  * cached blocks be of the correct size.
630  */
631 struct buf *
632 getblk(vp, blkno, size, slpflag, slptimeo)
633 	register struct vnode *vp;
634 	daddr_t blkno;
635 	int size, slpflag, slptimeo;
636 {
637 	struct bufhashhdr *bh;
638 	struct buf *bp;
639 	int s, err;
640 
641 	/*
642 	 * XXX
643 	 * The following is an inlined version of 'incore()', but with
644 	 * the 'invalid' test moved to after the 'busy' test.  It's
645 	 * necessary because there are some cases in which the NFS
646 	 * code sets B_INVAL prior to writing data to the server, but
647 	 * in which the buffers actually contain valid data.  In this
648 	 * case, we can't allow the system to allocate a new buffer for
649 	 * the block until the write is finished.
650 	 */
651 	bh = BUFHASH(vp, blkno);
652 start:
653         bp = bh->lh_first;
654         for (; bp != NULL; bp = bp->b_hash.le_next) {
655                 if (bp->b_lblkno != blkno || bp->b_vp != vp)
656 			continue;
657 
658 		s = splbio();
659 		if (ISSET(bp->b_flags, B_BUSY)) {
660 			SET(bp->b_flags, B_WANTED);
661 			err = tsleep(bp, slpflag | (PRIBIO + 1), "getblk",
662 			    slptimeo);
663 			splx(s);
664 			if (err)
665 				return (NULL);
666 			goto start;
667 		}
668 
669 		if (!ISSET(bp->b_flags, B_INVAL)) {
670 #ifdef DIAGNOSTIC
671 			if (ISSET(bp->b_flags, B_DONE|B_DELWRI) &&
672 			    bp->b_bcount < size)
673 				panic("getblk: block size invariant failed");
674 #endif
675 			SET(bp->b_flags, B_BUSY);
676 			bremfree(bp);
677 			splx(s);
678 			break;
679 		}
680 		splx(s);
681         }
682 
683 	if (bp == NULL) {
684 		if ((bp = getnewbuf(slpflag, slptimeo)) == NULL)
685 			goto start;
686 		binshash(bp, bh);
687 		bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno;
688 		s = splbio();
689 		bgetvp(vp, bp);
690 		splx(s);
691 	}
692 	allocbuf(bp, size);
693 	return (bp);
694 }
695 
696 /*
697  * Get an empty, disassociated buffer of given size.
698  */
699 struct buf *
700 geteblk(size)
701 	int size;
702 {
703 	struct buf *bp;
704 
705 	while ((bp = getnewbuf(0, 0)) == 0)
706 		;
707 	SET(bp->b_flags, B_INVAL);
708 	binshash(bp, &invalhash);
709 	allocbuf(bp, size);
710 
711 	return (bp);
712 }
713 
714 /*
715  * Expand or contract the actual memory allocated to a buffer.
716  *
717  * If the buffer shrinks, data is lost, so it's up to the
718  * caller to have written it out *first*; this routine will not
719  * start a write.  If the buffer grows, it's the callers
720  * responsibility to fill out the buffer's additional contents.
721  */
722 void
723 allocbuf(bp, size)
724 	struct buf *bp;
725 	int size;
726 {
727 	struct buf      *nbp;
728 	vsize_t       desired_size;
729 	int	     s;
730 
731 	desired_size = roundup(size, NBPG);
732 	if (desired_size > MAXBSIZE)
733 		panic("allocbuf: buffer larger than MAXBSIZE requested");
734 
735 	if (bp->b_bufsize == desired_size)
736 		goto out;
737 
738 	/*
739 	 * If the buffer is smaller than the desired size, we need to snarf
740 	 * it from other buffers.  Get buffers (via getnewbuf()), and
741 	 * steal their pages.
742 	 */
743 	while (bp->b_bufsize < desired_size) {
744 		int amt;
745 
746 		/* find a buffer */
747 		while ((nbp = getnewbuf(0, 0)) == NULL)
748 			;
749 		SET(nbp->b_flags, B_INVAL);
750 		binshash(nbp, &invalhash);
751 
752 		/* and steal its pages, up to the amount we need */
753 		amt = min(nbp->b_bufsize, (desired_size - bp->b_bufsize));
754 		pagemove((nbp->b_data + nbp->b_bufsize - amt),
755 			 bp->b_data + bp->b_bufsize, amt);
756 		bp->b_bufsize += amt;
757 		nbp->b_bufsize -= amt;
758 
759 		/* reduce transfer count if we stole some data */
760 		if (nbp->b_bcount > nbp->b_bufsize)
761 			nbp->b_bcount = nbp->b_bufsize;
762 
763 #ifdef DIAGNOSTIC
764 		if (nbp->b_bufsize < 0)
765 			panic("allocbuf: negative bufsize");
766 #endif
767 
768 		brelse(nbp);
769 	}
770 
771 	/*
772 	 * If we want a buffer smaller than the current size,
773 	 * shrink this buffer.  Grab a buf head from the EMPTY queue,
774 	 * move a page onto it, and put it on front of the AGE queue.
775 	 * If there are no free buffer headers, leave the buffer alone.
776 	 */
777 	if (bp->b_bufsize > desired_size) {
778 		s = splbio();
779 		if ((nbp = bufqueues[BQ_EMPTY].tqh_first) == NULL) {
780 			/* No free buffer head */
781 			splx(s);
782 			goto out;
783 		}
784 		bremfree(nbp);
785 		SET(nbp->b_flags, B_BUSY);
786 		splx(s);
787 
788 		/* move the page to it and note this change */
789 		pagemove(bp->b_data + desired_size,
790 		    nbp->b_data, bp->b_bufsize - desired_size);
791 		nbp->b_bufsize = bp->b_bufsize - desired_size;
792 		bp->b_bufsize = desired_size;
793 		nbp->b_bcount = 0;
794 		SET(nbp->b_flags, B_INVAL);
795 
796 		/* release the newly-filled buffer and leave */
797 		brelse(nbp);
798 	}
799 
800 out:
801 	bp->b_bcount = size;
802 }
803 
804 /*
805  * Find a buffer which is available for use.
806  * Select something from a free list.
807  * Preference is to AGE list, then LRU list.
808  */
809 struct buf *
810 getnewbuf(slpflag, slptimeo)
811 	int slpflag, slptimeo;
812 {
813 	register struct buf *bp;
814 	int s;
815 
816 start:
817 	s = splbio();
818 	if ((bp = bufqueues[BQ_AGE].tqh_first) != NULL ||
819 	    (bp = bufqueues[BQ_LRU].tqh_first) != NULL) {
820 		bremfree(bp);
821 	} else {
822 		/* wait for a free buffer of any kind */
823 		needbuffer = 1;
824 		tsleep(&needbuffer, slpflag|(PRIBIO+1), "getnewbuf", slptimeo);
825 		splx(s);
826 		return (0);
827 	}
828 
829 	if (ISSET(bp->b_flags, B_VFLUSH)) {
830 		/*
831 		 * This is a delayed write buffer being flushed to disk.  Make
832 		 * sure it gets aged out of the queue when it's finished, and
833 		 * leave it off the LRU queue.
834 		 */
835 		CLR(bp->b_flags, B_VFLUSH);
836 		SET(bp->b_flags, B_AGE);
837 		splx(s);
838 		goto start;
839 	}
840 
841 	/* Buffer is no longer on free lists. */
842 	SET(bp->b_flags, B_BUSY);
843 
844 	/* If buffer was a delayed write, start it, and go back to the top. */
845 	if (ISSET(bp->b_flags, B_DELWRI)) {
846 		splx(s);
847 		/*
848 		 * This buffer has gone through the LRU, so make sure it gets
849 		 * reused ASAP.
850 		 */
851 		SET(bp->b_flags, B_AGE);
852 		bawrite(bp);
853 		goto start;
854 	}
855 
856 	/* disassociate us from our vnode, if we had one... */
857 	if (bp->b_vp)
858 		brelvp(bp);
859 	splx(s);
860 
861 	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
862 		(*bioops.io_deallocate)(bp);
863 
864 	/* clear out various other fields */
865 	bp->b_flags = B_BUSY;
866 	bp->b_dev = NODEV;
867 	bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = 0;
868 	bp->b_iodone = 0;
869 	bp->b_error = 0;
870 	bp->b_resid = 0;
871 	bp->b_bcount = 0;
872 	bp->b_dirtyoff = bp->b_dirtyend = 0;
873 	bp->b_validoff = bp->b_validend = 0;
874 
875 	/* nuke any credentials we were holding */
876 	if (bp->b_rcred != NOCRED) {
877 		crfree(bp->b_rcred);
878 		bp->b_rcred = NOCRED;
879 	}
880 	if (bp->b_wcred != NOCRED) {
881 		crfree(bp->b_wcred);
882 		bp->b_wcred = NOCRED;
883 	}
884 
885 	bremhash(bp);
886 	return (bp);
887 }
888 
889 /*
890  * Wait for operations on the buffer to complete.
891  * When they do, extract and return the I/O's error value.
892  */
893 int
894 biowait(bp)
895 	struct buf *bp;
896 {
897 	int s;
898 
899 	s = splbio();
900 	while (!ISSET(bp->b_flags, B_DONE))
901 		tsleep(bp, PRIBIO + 1, "biowait", 0);
902 	splx(s);
903 
904 	/* check for interruption of I/O (e.g. via NFS), then errors. */
905 	if (ISSET(bp->b_flags, B_EINTR)) {
906 		CLR(bp->b_flags, B_EINTR);
907 		return (EINTR);
908 	} else if (ISSET(bp->b_flags, B_ERROR))
909 		return (bp->b_error ? bp->b_error : EIO);
910 	else
911 		return (0);
912 }
913 
914 /*
915  * Mark I/O complete on a buffer.
916  *
917  * If a callback has been requested, e.g. the pageout
918  * daemon, do so. Otherwise, awaken waiting processes.
919  *
920  * [ Leffler, et al., says on p.247:
921  *	"This routine wakes up the blocked process, frees the buffer
922  *	for an asynchronous write, or, for a request by the pagedaemon
923  *	process, invokes a procedure specified in the buffer structure" ]
924  *
925  * In real life, the pagedaemon (or other system processes) wants
926  * to do async stuff to, and doesn't want the buffer brelse()'d.
927  * (for swap pager, that puts swap buffers on the free lists (!!!),
928  * for the vn device, that puts malloc'd buffers on the free lists!)
929  */
930 void
931 biodone(bp)
932 	struct buf *bp;
933 {
934 	int s = splbio();
935 
936 	if (ISSET(bp->b_flags, B_DONE))
937 		panic("biodone already");
938 	SET(bp->b_flags, B_DONE);		/* note that it's done */
939 
940 	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
941 		(*bioops.io_complete)(bp);
942 
943 	if (!ISSET(bp->b_flags, B_READ))	/* wake up reader */
944 		vwakeup(bp);
945 
946 	if (ISSET(bp->b_flags, B_CALL)) {	/* if necessary, call out */
947 		CLR(bp->b_flags, B_CALL);	/* but note callout done */
948 		(*bp->b_iodone)(bp);
949 	} else {
950 		if (ISSET(bp->b_flags, B_ASYNC))	/* if async, release */
951 			brelse(bp);
952 		else {				/* or just wakeup the buffer */
953 			CLR(bp->b_flags, B_WANTED);
954 			wakeup(bp);
955 		}
956 	}
957 
958 	splx(s);
959 }
960 
961 /*
962  * Return a count of buffers on the "locked" queue.
963  */
964 int
965 count_lock_queue()
966 {
967 	register struct buf *bp;
968 	register int n = 0;
969 
970 	for (bp = bufqueues[BQ_LOCKED].tqh_first; bp;
971 	    bp = bp->b_freelist.tqe_next)
972 		n++;
973 	return (n);
974 }
975 
976 #ifdef DEBUG
977 /*
978  * Print out statistics on the current allocation of the buffer pool.
979  * Can be enabled to print out on every ``sync'' by setting "syncprt"
980  * in vfs_syscalls.c using sysctl.
981  */
982 void
983 vfs_bufstats()
984 {
985 	int s, i, j, count;
986 	register struct buf *bp;
987 	register struct bqueues *dp;
988 	int counts[MAXBSIZE/NBPG+1];
989 	static char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE", "EMPTY" };
990 
991 	for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) {
992 		count = 0;
993 		for (j = 0; j <= MAXBSIZE/NBPG; j++)
994 			counts[j] = 0;
995 		s = splbio();
996 		for (bp = dp->tqh_first; bp; bp = bp->b_freelist.tqe_next) {
997 			counts[bp->b_bufsize/NBPG]++;
998 			count++;
999 		}
1000 		splx(s);
1001 		printf("%s: total-%d", bname[i], count);
1002 		for (j = 0; j <= MAXBSIZE/NBPG; j++)
1003 			if (counts[j] != 0)
1004 				printf(", %d-%d", j * NBPG, counts[j]);
1005 		printf("\n");
1006 	}
1007 }
1008 #endif /* DEBUG */
1009