xref: /netbsd-src/sys/kern/vfs_bio.c (revision 96fc3e30a7c3f7bba53384bf41dad5f78306fac4)
1 /*	$NetBSD: vfs_bio.c,v 1.242 2012/12/30 09:19:24 hannken Exp $	*/
2 
3 /*-
4  * Copyright (c) 2007, 2008, 2009 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran, and by Wasabi Systems, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*-
33  * Copyright (c) 1982, 1986, 1989, 1993
34  *	The Regents of the University of California.  All rights reserved.
35  * (c) UNIX System Laboratories, Inc.
36  * All or some portions of this file are derived from material licensed
37  * to the University of California by American Telephone and Telegraph
38  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
39  * the permission of UNIX System Laboratories, Inc.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. Neither the name of the University nor the names of its contributors
50  *    may be used to endorse or promote products derived from this software
51  *    without specific prior written permission.
52  *
53  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
54  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
55  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
56  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
57  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
58  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
59  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63  * SUCH DAMAGE.
64  *
65  *	@(#)vfs_bio.c	8.6 (Berkeley) 1/11/94
66  */
67 
68 /*-
69  * Copyright (c) 1994 Christopher G. Demetriou
70  *
71  * Redistribution and use in source and binary forms, with or without
72  * modification, are permitted provided that the following conditions
73  * are met:
74  * 1. Redistributions of source code must retain the above copyright
75  *    notice, this list of conditions and the following disclaimer.
76  * 2. Redistributions in binary form must reproduce the above copyright
77  *    notice, this list of conditions and the following disclaimer in the
78  *    documentation and/or other materials provided with the distribution.
79  * 3. All advertising materials mentioning features or use of this software
80  *    must display the following acknowledgement:
81  *	This product includes software developed by the University of
82  *	California, Berkeley and its contributors.
83  * 4. Neither the name of the University nor the names of its contributors
84  *    may be used to endorse or promote products derived from this software
85  *    without specific prior written permission.
86  *
87  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
88  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
89  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
90  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
91  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
92  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
93  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
94  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
95  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
96  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
97  * SUCH DAMAGE.
98  *
99  *	@(#)vfs_bio.c	8.6 (Berkeley) 1/11/94
100  */
101 
102 /*
103  * The buffer cache subsystem.
104  *
105  * Some references:
106  *	Bach: The Design of the UNIX Operating System (Prentice Hall, 1986)
107  *	Leffler, et al.: The Design and Implementation of the 4.3BSD
108  *		UNIX Operating System (Addison Welley, 1989)
109  *
110  * Locking
111  *
112  * There are three locks:
113  * - bufcache_lock: protects global buffer cache state.
114  * - BC_BUSY: a long term per-buffer lock.
115  * - buf_t::b_objlock: lock on completion (biowait vs biodone).
116  *
117  * For buffers associated with vnodes (a most common case) b_objlock points
118  * to the vnode_t::v_interlock.  Otherwise, it points to generic buffer_lock.
119  *
120  * Lock order:
121  *	bufcache_lock ->
122  *		buf_t::b_objlock
123  */
124 
125 #include <sys/cdefs.h>
126 __KERNEL_RCSID(0, "$NetBSD: vfs_bio.c,v 1.242 2012/12/30 09:19:24 hannken Exp $");
127 
128 #include "opt_bufcache.h"
129 
130 #include <sys/param.h>
131 #include <sys/systm.h>
132 #include <sys/kernel.h>
133 #include <sys/proc.h>
134 #include <sys/buf.h>
135 #include <sys/vnode.h>
136 #include <sys/mount.h>
137 #include <sys/resourcevar.h>
138 #include <sys/sysctl.h>
139 #include <sys/conf.h>
140 #include <sys/kauth.h>
141 #include <sys/fstrans.h>
142 #include <sys/intr.h>
143 #include <sys/cpu.h>
144 #include <sys/wapbl.h>
145 #include <sys/bitops.h>
146 
147 #include <uvm/uvm.h>	/* extern struct uvm uvm */
148 
149 #include <miscfs/specfs/specdev.h>
150 
151 #ifndef	BUFPAGES
152 # define BUFPAGES 0
153 #endif
154 
155 #ifdef BUFCACHE
156 # if (BUFCACHE < 5) || (BUFCACHE > 95)
157 #  error BUFCACHE is not between 5 and 95
158 # endif
159 #else
160 # define BUFCACHE 15
161 #endif
162 
163 u_int	nbuf;			/* desired number of buffer headers */
164 u_int	bufpages = BUFPAGES;	/* optional hardwired count */
165 u_int	bufcache = BUFCACHE;	/* max % of RAM to use for buffer cache */
166 
167 /* Function prototypes */
168 struct bqueue;
169 
170 static void buf_setwm(void);
171 static int buf_trim(void);
172 static void *bufpool_page_alloc(struct pool *, int);
173 static void bufpool_page_free(struct pool *, void *);
174 static buf_t *bio_doread(struct vnode *, daddr_t, int,
175     kauth_cred_t, int);
176 static buf_t *getnewbuf(int, int, int);
177 static int buf_lotsfree(void);
178 static int buf_canrelease(void);
179 static u_long buf_mempoolidx(u_long);
180 static u_long buf_roundsize(u_long);
181 static void *buf_alloc(size_t);
182 static void buf_mrelease(void *, size_t);
183 static void binsheadfree(buf_t *, struct bqueue *);
184 static void binstailfree(buf_t *, struct bqueue *);
185 #ifdef DEBUG
186 static int checkfreelist(buf_t *, struct bqueue *, int);
187 #endif
188 static void biointr(void *);
189 static void biodone2(buf_t *);
190 static void bref(buf_t *);
191 static void brele(buf_t *);
192 static void sysctl_kern_buf_setup(void);
193 static void sysctl_vm_buf_setup(void);
194 
195 /*
196  * Definitions for the buffer hash lists.
197  */
198 #define	BUFHASH(dvp, lbn)	\
199 	(&bufhashtbl[(((long)(dvp) >> 8) + (int)(lbn)) & bufhash])
200 LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
201 u_long	bufhash;
202 struct bqueue bufqueues[BQUEUES];
203 
204 static kcondvar_t needbuffer_cv;
205 
206 /*
207  * Buffer queue lock.
208  */
209 kmutex_t bufcache_lock;
210 kmutex_t buffer_lock;
211 
212 /* Software ISR for completed transfers. */
213 static void *biodone_sih;
214 
215 /* Buffer pool for I/O buffers. */
216 static pool_cache_t buf_cache;
217 static pool_cache_t bufio_cache;
218 
219 #define MEMPOOL_INDEX_OFFSET (ilog2(DEV_BSIZE))	/* smallest pool is 512 bytes */
220 #define NMEMPOOLS (ilog2(MAXBSIZE) - MEMPOOL_INDEX_OFFSET + 1)
221 __CTASSERT((1 << (NMEMPOOLS + MEMPOOL_INDEX_OFFSET - 1)) == MAXBSIZE);
222 
223 /* Buffer memory pools */
224 static struct pool bmempools[NMEMPOOLS];
225 
226 static struct vm_map *buf_map;
227 
228 /*
229  * Buffer memory pool allocator.
230  */
231 static void *
232 bufpool_page_alloc(struct pool *pp, int flags)
233 {
234 
235 	return (void *)uvm_km_alloc(buf_map,
236 	    MAXBSIZE, MAXBSIZE,
237 	    ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT|UVM_KMF_TRYLOCK)
238 	    | UVM_KMF_WIRED);
239 }
240 
241 static void
242 bufpool_page_free(struct pool *pp, void *v)
243 {
244 
245 	uvm_km_free(buf_map, (vaddr_t)v, MAXBSIZE, UVM_KMF_WIRED);
246 }
247 
248 static struct pool_allocator bufmempool_allocator = {
249 	.pa_alloc = bufpool_page_alloc,
250 	.pa_free = bufpool_page_free,
251 	.pa_pagesz = MAXBSIZE,
252 };
253 
254 /* Buffer memory management variables */
255 u_long bufmem_valimit;
256 u_long bufmem_hiwater;
257 u_long bufmem_lowater;
258 u_long bufmem;
259 
260 /*
261  * MD code can call this to set a hard limit on the amount
262  * of virtual memory used by the buffer cache.
263  */
264 int
265 buf_setvalimit(vsize_t sz)
266 {
267 
268 	/* We need to accommodate at least NMEMPOOLS of MAXBSIZE each */
269 	if (sz < NMEMPOOLS * MAXBSIZE)
270 		return EINVAL;
271 
272 	bufmem_valimit = sz;
273 	return 0;
274 }
275 
276 static void
277 buf_setwm(void)
278 {
279 
280 	bufmem_hiwater = buf_memcalc();
281 	/* lowater is approx. 2% of memory (with bufcache = 15) */
282 #define	BUFMEM_WMSHIFT	3
283 #define	BUFMEM_HIWMMIN	(64 * 1024 << BUFMEM_WMSHIFT)
284 	if (bufmem_hiwater < BUFMEM_HIWMMIN)
285 		/* Ensure a reasonable minimum value */
286 		bufmem_hiwater = BUFMEM_HIWMMIN;
287 	bufmem_lowater = bufmem_hiwater >> BUFMEM_WMSHIFT;
288 }
289 
290 #ifdef DEBUG
291 int debug_verify_freelist = 0;
292 static int
293 checkfreelist(buf_t *bp, struct bqueue *dp, int ison)
294 {
295 	buf_t *b;
296 
297 	if (!debug_verify_freelist)
298 		return 1;
299 
300 	TAILQ_FOREACH(b, &dp->bq_queue, b_freelist) {
301 		if (b == bp)
302 			return ison ? 1 : 0;
303 	}
304 
305 	return ison ? 0 : 1;
306 }
307 #endif
308 
309 /*
310  * Insq/Remq for the buffer hash lists.
311  * Call with buffer queue locked.
312  */
313 static void
314 binsheadfree(buf_t *bp, struct bqueue *dp)
315 {
316 
317 	KASSERT(mutex_owned(&bufcache_lock));
318 	KASSERT(bp->b_freelistindex == -1);
319 	TAILQ_INSERT_HEAD(&dp->bq_queue, bp, b_freelist);
320 	dp->bq_bytes += bp->b_bufsize;
321 	bp->b_freelistindex = dp - bufqueues;
322 }
323 
324 static void
325 binstailfree(buf_t *bp, struct bqueue *dp)
326 {
327 
328 	KASSERT(mutex_owned(&bufcache_lock));
329 	KASSERT(bp->b_freelistindex == -1);
330 	TAILQ_INSERT_TAIL(&dp->bq_queue, bp, b_freelist);
331 	dp->bq_bytes += bp->b_bufsize;
332 	bp->b_freelistindex = dp - bufqueues;
333 }
334 
335 void
336 bremfree(buf_t *bp)
337 {
338 	struct bqueue *dp;
339 	int bqidx = bp->b_freelistindex;
340 
341 	KASSERT(mutex_owned(&bufcache_lock));
342 
343 	KASSERT(bqidx != -1);
344 	dp = &bufqueues[bqidx];
345 	KDASSERT(checkfreelist(bp, dp, 1));
346 	KASSERT(dp->bq_bytes >= bp->b_bufsize);
347 	TAILQ_REMOVE(&dp->bq_queue, bp, b_freelist);
348 	dp->bq_bytes -= bp->b_bufsize;
349 
350 	/* For the sysctl helper. */
351 	if (bp == dp->bq_marker)
352 		dp->bq_marker = NULL;
353 
354 #if defined(DIAGNOSTIC)
355 	bp->b_freelistindex = -1;
356 #endif /* defined(DIAGNOSTIC) */
357 }
358 
359 /*
360  * Add a reference to an buffer structure that came from buf_cache.
361  */
362 static inline void
363 bref(buf_t *bp)
364 {
365 
366 	KASSERT(mutex_owned(&bufcache_lock));
367 	KASSERT(bp->b_refcnt > 0);
368 
369 	bp->b_refcnt++;
370 }
371 
372 /*
373  * Free an unused buffer structure that came from buf_cache.
374  */
375 static inline void
376 brele(buf_t *bp)
377 {
378 
379 	KASSERT(mutex_owned(&bufcache_lock));
380 	KASSERT(bp->b_refcnt > 0);
381 
382 	if (bp->b_refcnt-- == 1) {
383 		buf_destroy(bp);
384 #ifdef DEBUG
385 		memset((char *)bp, 0, sizeof(*bp));
386 #endif
387 		pool_cache_put(buf_cache, bp);
388 	}
389 }
390 
391 /*
392  * note that for some ports this is used by pmap bootstrap code to
393  * determine kva size.
394  */
395 u_long
396 buf_memcalc(void)
397 {
398 	u_long n;
399 
400 	/*
401 	 * Determine the upper bound of memory to use for buffers.
402 	 *
403 	 *	- If bufpages is specified, use that as the number
404 	 *	  pages.
405 	 *
406 	 *	- Otherwise, use bufcache as the percentage of
407 	 *	  physical memory.
408 	 */
409 	if (bufpages != 0) {
410 		n = bufpages;
411 	} else {
412 		if (bufcache < 5) {
413 			printf("forcing bufcache %d -> 5", bufcache);
414 			bufcache = 5;
415 		}
416 		if (bufcache > 95) {
417 			printf("forcing bufcache %d -> 95", bufcache);
418 			bufcache = 95;
419 		}
420 		n = calc_cache_size(buf_map, bufcache,
421 		    (buf_map != kernel_map) ? 100 : BUFCACHE_VA_MAXPCT)
422 		    / PAGE_SIZE;
423 	}
424 
425 	n <<= PAGE_SHIFT;
426 	if (bufmem_valimit != 0 && n > bufmem_valimit)
427 		n = bufmem_valimit;
428 
429 	return (n);
430 }
431 
432 /*
433  * Initialize buffers and hash links for buffers.
434  */
435 void
436 bufinit(void)
437 {
438 	struct bqueue *dp;
439 	int use_std;
440 	u_int i;
441 
442 	mutex_init(&bufcache_lock, MUTEX_DEFAULT, IPL_NONE);
443 	mutex_init(&buffer_lock, MUTEX_DEFAULT, IPL_NONE);
444 	cv_init(&needbuffer_cv, "needbuf");
445 
446 	if (bufmem_valimit != 0) {
447 		vaddr_t minaddr = 0, maxaddr;
448 		buf_map = uvm_km_suballoc(kernel_map, &minaddr, &maxaddr,
449 					  bufmem_valimit, 0, false, 0);
450 		if (buf_map == NULL)
451 			panic("bufinit: cannot allocate submap");
452 	} else
453 		buf_map = kernel_map;
454 
455 	/*
456 	 * Initialize buffer cache memory parameters.
457 	 */
458 	bufmem = 0;
459 	buf_setwm();
460 
461 	/* On "small" machines use small pool page sizes where possible */
462 	use_std = (physmem < atop(16*1024*1024));
463 
464 	/*
465 	 * Also use them on systems that can map the pool pages using
466 	 * a direct-mapped segment.
467 	 */
468 #ifdef PMAP_MAP_POOLPAGE
469 	use_std = 1;
470 #endif
471 
472 	buf_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
473 	    "bufpl", NULL, IPL_SOFTBIO, NULL, NULL, NULL);
474 	bufio_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
475 	    "biopl", NULL, IPL_BIO, NULL, NULL, NULL);
476 
477 	for (i = 0; i < NMEMPOOLS; i++) {
478 		struct pool_allocator *pa;
479 		struct pool *pp = &bmempools[i];
480 		u_int size = 1 << (i + MEMPOOL_INDEX_OFFSET);
481 		char *name = kmem_alloc(8, KM_SLEEP); /* XXX: never freed */
482 		if (__predict_false(size >= 1048576))
483 			(void)snprintf(name, 8, "buf%um", size / 1048576);
484 		else if (__predict_true(size >= 1024))
485 			(void)snprintf(name, 8, "buf%uk", size / 1024);
486 		else
487 			(void)snprintf(name, 8, "buf%ub", size);
488 		pa = (size <= PAGE_SIZE && use_std)
489 			? &pool_allocator_nointr
490 			: &bufmempool_allocator;
491 		pool_init(pp, size, 0, 0, 0, name, pa, IPL_NONE);
492 		pool_setlowat(pp, 1);
493 		pool_sethiwat(pp, 1);
494 	}
495 
496 	/* Initialize the buffer queues */
497 	for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) {
498 		TAILQ_INIT(&dp->bq_queue);
499 		dp->bq_bytes = 0;
500 	}
501 
502 	/*
503 	 * Estimate hash table size based on the amount of memory we
504 	 * intend to use for the buffer cache. The average buffer
505 	 * size is dependent on our clients (i.e. filesystems).
506 	 *
507 	 * For now, use an empirical 3K per buffer.
508 	 */
509 	nbuf = (bufmem_hiwater / 1024) / 3;
510 	bufhashtbl = hashinit(nbuf, HASH_LIST, true, &bufhash);
511 
512 	sysctl_kern_buf_setup();
513 	sysctl_vm_buf_setup();
514 }
515 
516 void
517 bufinit2(void)
518 {
519 
520 	biodone_sih = softint_establish(SOFTINT_BIO | SOFTINT_MPSAFE, biointr,
521 	    NULL);
522 	if (biodone_sih == NULL)
523 		panic("bufinit2: can't establish soft interrupt");
524 }
525 
526 static int
527 buf_lotsfree(void)
528 {
529 	int try, thresh;
530 
531 	/* Always allocate if less than the low water mark. */
532 	if (bufmem < bufmem_lowater)
533 		return 1;
534 
535 	/* Never allocate if greater than the high water mark. */
536 	if (bufmem > bufmem_hiwater)
537 		return 0;
538 
539 	/* If there's anything on the AGE list, it should be eaten. */
540 	if (TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue) != NULL)
541 		return 0;
542 
543 	/*
544 	 * The probabily of getting a new allocation is inversely
545 	 * proportional to the current size of the cache, using
546 	 * a granularity of 16 steps.
547 	 */
548 	try = random() & 0x0000000fL;
549 
550 	/* Don't use "16 * bufmem" here to avoid a 32-bit overflow. */
551 	thresh = (bufmem - bufmem_lowater) /
552 	    ((bufmem_hiwater - bufmem_lowater) / 16);
553 
554 	if (try >= thresh)
555 		return 1;
556 
557 	/* Otherwise don't allocate. */
558 	return 0;
559 }
560 
561 /*
562  * Return estimate of bytes we think need to be
563  * released to help resolve low memory conditions.
564  *
565  * => called with bufcache_lock held.
566  */
567 static int
568 buf_canrelease(void)
569 {
570 	int pagedemand, ninvalid = 0;
571 
572 	KASSERT(mutex_owned(&bufcache_lock));
573 
574 	if (bufmem < bufmem_lowater)
575 		return 0;
576 
577 	if (bufmem > bufmem_hiwater)
578 		return bufmem - bufmem_hiwater;
579 
580 	ninvalid += bufqueues[BQ_AGE].bq_bytes;
581 
582 	pagedemand = uvmexp.freetarg - uvmexp.free;
583 	if (pagedemand < 0)
584 		return ninvalid;
585 	return MAX(ninvalid, MIN(2 * MAXBSIZE,
586 	    MIN((bufmem - bufmem_lowater) / 16, pagedemand * PAGE_SIZE)));
587 }
588 
589 /*
590  * Buffer memory allocation helper functions
591  */
592 static u_long
593 buf_mempoolidx(u_long size)
594 {
595 	u_int n = 0;
596 
597 	size -= 1;
598 	size >>= MEMPOOL_INDEX_OFFSET;
599 	while (size) {
600 		size >>= 1;
601 		n += 1;
602 	}
603 	if (n >= NMEMPOOLS)
604 		panic("buf mem pool index %d", n);
605 	return n;
606 }
607 
608 static u_long
609 buf_roundsize(u_long size)
610 {
611 	/* Round up to nearest power of 2 */
612 	return (1 << (buf_mempoolidx(size) + MEMPOOL_INDEX_OFFSET));
613 }
614 
615 static void *
616 buf_alloc(size_t size)
617 {
618 	u_int n = buf_mempoolidx(size);
619 	void *addr;
620 
621 	while (1) {
622 		addr = pool_get(&bmempools[n], PR_NOWAIT);
623 		if (addr != NULL)
624 			break;
625 
626 		/* No memory, see if we can free some. If so, try again */
627 		mutex_enter(&bufcache_lock);
628 		if (buf_drain(1) > 0) {
629 			mutex_exit(&bufcache_lock);
630 			continue;
631 		}
632 
633 		if (curlwp == uvm.pagedaemon_lwp) {
634 			mutex_exit(&bufcache_lock);
635 			return NULL;
636 		}
637 
638 		/* Wait for buffers to arrive on the LRU queue */
639 		cv_timedwait(&needbuffer_cv, &bufcache_lock, hz / 4);
640 		mutex_exit(&bufcache_lock);
641 	}
642 
643 	return addr;
644 }
645 
646 static void
647 buf_mrelease(void *addr, size_t size)
648 {
649 
650 	pool_put(&bmempools[buf_mempoolidx(size)], addr);
651 }
652 
653 /*
654  * bread()/breadn() helper.
655  */
656 static buf_t *
657 bio_doread(struct vnode *vp, daddr_t blkno, int size, kauth_cred_t cred,
658     int async)
659 {
660 	buf_t *bp;
661 	struct mount *mp;
662 
663 	bp = getblk(vp, blkno, size, 0, 0);
664 
665 	/*
666 	 * getblk() may return NULL if we are the pagedaemon.
667 	 */
668 	if (bp == NULL) {
669 		KASSERT(curlwp == uvm.pagedaemon_lwp);
670 		return NULL;
671 	}
672 
673 	/*
674 	 * If buffer does not have data valid, start a read.
675 	 * Note that if buffer is BC_INVAL, getblk() won't return it.
676 	 * Therefore, it's valid if its I/O has completed or been delayed.
677 	 */
678 	if (!ISSET(bp->b_oflags, (BO_DONE | BO_DELWRI))) {
679 		/* Start I/O for the buffer. */
680 		SET(bp->b_flags, B_READ | async);
681 		if (async)
682 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
683 		else
684 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
685 		VOP_STRATEGY(vp, bp);
686 
687 		/* Pay for the read. */
688 		curlwp->l_ru.ru_inblock++;
689 	} else if (async)
690 		brelse(bp, 0);
691 
692 	if (vp->v_type == VBLK)
693 		mp = vp->v_specmountpoint;
694 	else
695 		mp = vp->v_mount;
696 
697 	/*
698 	 * Collect statistics on synchronous and asynchronous reads.
699 	 * Reads from block devices are charged to their associated
700 	 * filesystem (if any).
701 	 */
702 	if (mp != NULL) {
703 		if (async == 0)
704 			mp->mnt_stat.f_syncreads++;
705 		else
706 			mp->mnt_stat.f_asyncreads++;
707 	}
708 
709 	return (bp);
710 }
711 
712 /*
713  * Read a disk block.
714  * This algorithm described in Bach (p.54).
715  */
716 int
717 bread(struct vnode *vp, daddr_t blkno, int size, kauth_cred_t cred,
718     int flags, buf_t **bpp)
719 {
720 	buf_t *bp;
721 	int error;
722 
723 	/* Get buffer for block. */
724 	bp = *bpp = bio_doread(vp, blkno, size, cred, 0);
725 	if (bp == NULL)
726 		return ENOMEM;
727 
728 	/* Wait for the read to complete, and return result. */
729 	error = biowait(bp);
730 	if (error == 0 && (flags & B_MODIFY) != 0)
731 		error = fscow_run(bp, true);
732 	if (error) {
733 		brelse(bp, 0);
734 		*bpp = NULL;
735 	}
736 
737 	return error;
738 }
739 
740 /*
741  * Read-ahead multiple disk blocks. The first is sync, the rest async.
742  * Trivial modification to the breada algorithm presented in Bach (p.55).
743  */
744 int
745 breadn(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablks,
746     int *rasizes, int nrablks, kauth_cred_t cred, int flags, buf_t **bpp)
747 {
748 	buf_t *bp;
749 	int error, i;
750 
751 	bp = *bpp = bio_doread(vp, blkno, size, cred, 0);
752 	if (bp == NULL)
753 		return ENOMEM;
754 
755 	/*
756 	 * For each of the read-ahead blocks, start a read, if necessary.
757 	 */
758 	mutex_enter(&bufcache_lock);
759 	for (i = 0; i < nrablks; i++) {
760 		/* If it's in the cache, just go on to next one. */
761 		if (incore(vp, rablks[i]))
762 			continue;
763 
764 		/* Get a buffer for the read-ahead block */
765 		mutex_exit(&bufcache_lock);
766 		(void) bio_doread(vp, rablks[i], rasizes[i], cred, B_ASYNC);
767 		mutex_enter(&bufcache_lock);
768 	}
769 	mutex_exit(&bufcache_lock);
770 
771 	/* Otherwise, we had to start a read for it; wait until it's valid. */
772 	error = biowait(bp);
773 	if (error == 0 && (flags & B_MODIFY) != 0)
774 		error = fscow_run(bp, true);
775 	if (error) {
776 		brelse(bp, 0);
777 		*bpp = NULL;
778 	}
779 
780 	return error;
781 }
782 
783 /*
784  * Block write.  Described in Bach (p.56)
785  */
786 int
787 bwrite(buf_t *bp)
788 {
789 	int rv, sync, wasdelayed;
790 	struct vnode *vp;
791 	struct mount *mp;
792 
793 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
794 	KASSERT(!cv_has_waiters(&bp->b_done));
795 
796 	vp = bp->b_vp;
797 	if (vp != NULL) {
798 		KASSERT(bp->b_objlock == vp->v_interlock);
799 		if (vp->v_type == VBLK)
800 			mp = vp->v_specmountpoint;
801 		else
802 			mp = vp->v_mount;
803 	} else {
804 		mp = NULL;
805 	}
806 
807 	if (mp && mp->mnt_wapbl) {
808 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
809 			bdwrite(bp);
810 			return 0;
811 		}
812 	}
813 
814 	/*
815 	 * Remember buffer type, to switch on it later.  If the write was
816 	 * synchronous, but the file system was mounted with MNT_ASYNC,
817 	 * convert it to a delayed write.
818 	 * XXX note that this relies on delayed tape writes being converted
819 	 * to async, not sync writes (which is safe, but ugly).
820 	 */
821 	sync = !ISSET(bp->b_flags, B_ASYNC);
822 	if (sync && mp != NULL && ISSET(mp->mnt_flag, MNT_ASYNC)) {
823 		bdwrite(bp);
824 		return (0);
825 	}
826 
827 	/*
828 	 * Collect statistics on synchronous and asynchronous writes.
829 	 * Writes to block devices are charged to their associated
830 	 * filesystem (if any).
831 	 */
832 	if (mp != NULL) {
833 		if (sync)
834 			mp->mnt_stat.f_syncwrites++;
835 		else
836 			mp->mnt_stat.f_asyncwrites++;
837 	}
838 
839 	/*
840 	 * Pay for the I/O operation and make sure the buf is on the correct
841 	 * vnode queue.
842 	 */
843 	bp->b_error = 0;
844 	wasdelayed = ISSET(bp->b_oflags, BO_DELWRI);
845 	CLR(bp->b_flags, B_READ);
846 	if (wasdelayed) {
847 		mutex_enter(&bufcache_lock);
848 		mutex_enter(bp->b_objlock);
849 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
850 		reassignbuf(bp, bp->b_vp);
851 		mutex_exit(&bufcache_lock);
852 	} else {
853 		curlwp->l_ru.ru_oublock++;
854 		mutex_enter(bp->b_objlock);
855 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
856 	}
857 	if (vp != NULL)
858 		vp->v_numoutput++;
859 	mutex_exit(bp->b_objlock);
860 
861 	/* Initiate disk write. */
862 	if (sync)
863 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
864 	else
865 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
866 
867 	VOP_STRATEGY(vp, bp);
868 
869 	if (sync) {
870 		/* If I/O was synchronous, wait for it to complete. */
871 		rv = biowait(bp);
872 
873 		/* Release the buffer. */
874 		brelse(bp, 0);
875 
876 		return (rv);
877 	} else {
878 		return (0);
879 	}
880 }
881 
882 int
883 vn_bwrite(void *v)
884 {
885 	struct vop_bwrite_args *ap = v;
886 
887 	return (bwrite(ap->a_bp));
888 }
889 
890 /*
891  * Delayed write.
892  *
893  * The buffer is marked dirty, but is not queued for I/O.
894  * This routine should be used when the buffer is expected
895  * to be modified again soon, typically a small write that
896  * partially fills a buffer.
897  *
898  * NB: magnetic tapes cannot be delayed; they must be
899  * written in the order that the writes are requested.
900  *
901  * Described in Leffler, et al. (pp. 208-213).
902  */
903 void
904 bdwrite(buf_t *bp)
905 {
906 
907 	KASSERT(bp->b_vp == NULL || bp->b_vp->v_tag != VT_UFS ||
908 	    bp->b_vp->v_type == VBLK || ISSET(bp->b_flags, B_COWDONE));
909 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
910 	KASSERT(!cv_has_waiters(&bp->b_done));
911 
912 	/* If this is a tape block, write the block now. */
913 	if (bdev_type(bp->b_dev) == D_TAPE) {
914 		bawrite(bp);
915 		return;
916 	}
917 
918 	if (wapbl_vphaswapbl(bp->b_vp)) {
919 		struct mount *mp = wapbl_vptomp(bp->b_vp);
920 
921 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
922 			WAPBL_ADD_BUF(mp, bp);
923 		}
924 	}
925 
926 	/*
927 	 * If the block hasn't been seen before:
928 	 *	(1) Mark it as having been seen,
929 	 *	(2) Charge for the write,
930 	 *	(3) Make sure it's on its vnode's correct block list.
931 	 */
932 	KASSERT(bp->b_vp == NULL || bp->b_objlock == bp->b_vp->v_interlock);
933 
934 	if (!ISSET(bp->b_oflags, BO_DELWRI)) {
935 		mutex_enter(&bufcache_lock);
936 		mutex_enter(bp->b_objlock);
937 		SET(bp->b_oflags, BO_DELWRI);
938 		curlwp->l_ru.ru_oublock++;
939 		reassignbuf(bp, bp->b_vp);
940 		mutex_exit(&bufcache_lock);
941 	} else {
942 		mutex_enter(bp->b_objlock);
943 	}
944 	/* Otherwise, the "write" is done, so mark and release the buffer. */
945 	CLR(bp->b_oflags, BO_DONE);
946 	mutex_exit(bp->b_objlock);
947 
948 	brelse(bp, 0);
949 }
950 
951 /*
952  * Asynchronous block write; just an asynchronous bwrite().
953  */
954 void
955 bawrite(buf_t *bp)
956 {
957 
958 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
959 	KASSERT(bp->b_vp != NULL);
960 
961 	SET(bp->b_flags, B_ASYNC);
962 	VOP_BWRITE(bp->b_vp, bp);
963 }
964 
965 /*
966  * Release a buffer on to the free lists.
967  * Described in Bach (p. 46).
968  */
969 void
970 brelsel(buf_t *bp, int set)
971 {
972 	struct bqueue *bufq;
973 	struct vnode *vp;
974 
975 	KASSERT(bp != NULL);
976 	KASSERT(mutex_owned(&bufcache_lock));
977 	KASSERT(!cv_has_waiters(&bp->b_done));
978 	KASSERT(bp->b_refcnt > 0);
979 
980 	SET(bp->b_cflags, set);
981 
982 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
983 	KASSERT(bp->b_iodone == NULL);
984 
985 	/* Wake up any processes waiting for any buffer to become free. */
986 	cv_signal(&needbuffer_cv);
987 
988 	/* Wake up any proceeses waiting for _this_ buffer to become */
989 	if (ISSET(bp->b_cflags, BC_WANTED))
990 		CLR(bp->b_cflags, BC_WANTED|BC_AGE);
991 
992 	/* If it's clean clear the copy-on-write flag. */
993 	if (ISSET(bp->b_flags, B_COWDONE)) {
994 		mutex_enter(bp->b_objlock);
995 		if (!ISSET(bp->b_oflags, BO_DELWRI))
996 			CLR(bp->b_flags, B_COWDONE);
997 		mutex_exit(bp->b_objlock);
998 	}
999 
1000 	/*
1001 	 * Determine which queue the buffer should be on, then put it there.
1002 	 */
1003 
1004 	/* If it's locked, don't report an error; try again later. */
1005 	if (ISSET(bp->b_flags, B_LOCKED))
1006 		bp->b_error = 0;
1007 
1008 	/* If it's not cacheable, or an error, mark it invalid. */
1009 	if (ISSET(bp->b_cflags, BC_NOCACHE) || bp->b_error != 0)
1010 		SET(bp->b_cflags, BC_INVAL);
1011 
1012 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1013 		/*
1014 		 * This is a delayed write buffer that was just flushed to
1015 		 * disk.  It is still on the LRU queue.  If it's become
1016 		 * invalid, then we need to move it to a different queue;
1017 		 * otherwise leave it in its current position.
1018 		 */
1019 		CLR(bp->b_cflags, BC_VFLUSH);
1020 		if (!ISSET(bp->b_cflags, BC_INVAL|BC_AGE) &&
1021 		    !ISSET(bp->b_flags, B_LOCKED) && bp->b_error == 0) {
1022 			KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 1));
1023 			goto already_queued;
1024 		} else {
1025 			bremfree(bp);
1026 		}
1027 	}
1028 
1029 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_AGE], 0));
1030 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 0));
1031 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LOCKED], 0));
1032 
1033 	if ((bp->b_bufsize <= 0) || ISSET(bp->b_cflags, BC_INVAL)) {
1034 		/*
1035 		 * If it's invalid or empty, dissociate it from its vnode
1036 		 * and put on the head of the appropriate queue.
1037 		 */
1038 		if (ISSET(bp->b_flags, B_LOCKED)) {
1039 			if (wapbl_vphaswapbl(vp = bp->b_vp)) {
1040 				struct mount *mp = wapbl_vptomp(vp);
1041 
1042 				KASSERT(bp->b_iodone
1043 				    != mp->mnt_wapbl_op->wo_wapbl_biodone);
1044 				WAPBL_REMOVE_BUF(mp, bp);
1045 			}
1046 		}
1047 
1048 		mutex_enter(bp->b_objlock);
1049 		CLR(bp->b_oflags, BO_DONE|BO_DELWRI);
1050 		if ((vp = bp->b_vp) != NULL) {
1051 			KASSERT(bp->b_objlock == vp->v_interlock);
1052 			reassignbuf(bp, bp->b_vp);
1053 			brelvp(bp);
1054 			mutex_exit(vp->v_interlock);
1055 		} else {
1056 			KASSERT(bp->b_objlock == &buffer_lock);
1057 			mutex_exit(bp->b_objlock);
1058 		}
1059 
1060 		if (bp->b_bufsize <= 0)
1061 			/* no data */
1062 			goto already_queued;
1063 		else
1064 			/* invalid data */
1065 			bufq = &bufqueues[BQ_AGE];
1066 		binsheadfree(bp, bufq);
1067 	} else  {
1068 		/*
1069 		 * It has valid data.  Put it on the end of the appropriate
1070 		 * queue, so that it'll stick around for as long as possible.
1071 		 * If buf is AGE, but has dependencies, must put it on last
1072 		 * bufqueue to be scanned, ie LRU. This protects against the
1073 		 * livelock where BQ_AGE only has buffers with dependencies,
1074 		 * and we thus never get to the dependent buffers in BQ_LRU.
1075 		 */
1076 		if (ISSET(bp->b_flags, B_LOCKED)) {
1077 			/* locked in core */
1078 			bufq = &bufqueues[BQ_LOCKED];
1079 		} else if (!ISSET(bp->b_cflags, BC_AGE)) {
1080 			/* valid data */
1081 			bufq = &bufqueues[BQ_LRU];
1082 		} else {
1083 			/* stale but valid data */
1084 			bufq = &bufqueues[BQ_AGE];
1085 		}
1086 		binstailfree(bp, bufq);
1087 	}
1088 already_queued:
1089 	/* Unlock the buffer. */
1090 	CLR(bp->b_cflags, BC_AGE|BC_BUSY|BC_NOCACHE);
1091 	CLR(bp->b_flags, B_ASYNC);
1092 	cv_broadcast(&bp->b_busy);
1093 
1094 	if (bp->b_bufsize <= 0)
1095 		brele(bp);
1096 }
1097 
1098 void
1099 brelse(buf_t *bp, int set)
1100 {
1101 
1102 	mutex_enter(&bufcache_lock);
1103 	brelsel(bp, set);
1104 	mutex_exit(&bufcache_lock);
1105 }
1106 
1107 /*
1108  * Determine if a block is in the cache.
1109  * Just look on what would be its hash chain.  If it's there, return
1110  * a pointer to it, unless it's marked invalid.  If it's marked invalid,
1111  * we normally don't return the buffer, unless the caller explicitly
1112  * wants us to.
1113  */
1114 buf_t *
1115 incore(struct vnode *vp, daddr_t blkno)
1116 {
1117 	buf_t *bp;
1118 
1119 	KASSERT(mutex_owned(&bufcache_lock));
1120 
1121 	/* Search hash chain */
1122 	LIST_FOREACH(bp, BUFHASH(vp, blkno), b_hash) {
1123 		if (bp->b_lblkno == blkno && bp->b_vp == vp &&
1124 		    !ISSET(bp->b_cflags, BC_INVAL)) {
1125 		    	KASSERT(bp->b_objlock == vp->v_interlock);
1126 		    	return (bp);
1127 		}
1128 	}
1129 
1130 	return (NULL);
1131 }
1132 
1133 /*
1134  * Get a block of requested size that is associated with
1135  * a given vnode and block offset. If it is found in the
1136  * block cache, mark it as having been found, make it busy
1137  * and return it. Otherwise, return an empty block of the
1138  * correct size. It is up to the caller to insure that the
1139  * cached blocks be of the correct size.
1140  */
1141 buf_t *
1142 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo)
1143 {
1144 	int err, preserve;
1145 	buf_t *bp;
1146 
1147 	mutex_enter(&bufcache_lock);
1148  loop:
1149 	bp = incore(vp, blkno);
1150 	if (bp != NULL) {
1151 		err = bbusy(bp, ((slpflag & PCATCH) != 0), slptimeo, NULL);
1152 		if (err != 0) {
1153 			if (err == EPASSTHROUGH)
1154 				goto loop;
1155 			mutex_exit(&bufcache_lock);
1156 			return (NULL);
1157 		}
1158 		KASSERT(!cv_has_waiters(&bp->b_done));
1159 #ifdef DIAGNOSTIC
1160 		if (ISSET(bp->b_oflags, BO_DONE|BO_DELWRI) &&
1161 		    bp->b_bcount < size && vp->v_type != VBLK)
1162 			panic("getblk: block size invariant failed");
1163 #endif
1164 		bremfree(bp);
1165 		preserve = 1;
1166 	} else {
1167 		if ((bp = getnewbuf(slpflag, slptimeo, 0)) == NULL)
1168 			goto loop;
1169 
1170 		if (incore(vp, blkno) != NULL) {
1171 			/* The block has come into memory in the meantime. */
1172 			brelsel(bp, 0);
1173 			goto loop;
1174 		}
1175 
1176 		LIST_INSERT_HEAD(BUFHASH(vp, blkno), bp, b_hash);
1177 		bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno;
1178 		mutex_enter(vp->v_interlock);
1179 		bgetvp(vp, bp);
1180 		mutex_exit(vp->v_interlock);
1181 		preserve = 0;
1182 	}
1183 	mutex_exit(&bufcache_lock);
1184 
1185 	/*
1186 	 * LFS can't track total size of B_LOCKED buffer (locked_queue_bytes)
1187 	 * if we re-size buffers here.
1188 	 */
1189 	if (ISSET(bp->b_flags, B_LOCKED)) {
1190 		KASSERT(bp->b_bufsize >= size);
1191 	} else {
1192 		if (allocbuf(bp, size, preserve)) {
1193 			mutex_enter(&bufcache_lock);
1194 			LIST_REMOVE(bp, b_hash);
1195 			mutex_exit(&bufcache_lock);
1196 			brelse(bp, BC_INVAL);
1197 			return NULL;
1198 		}
1199 	}
1200 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1201 	return (bp);
1202 }
1203 
1204 /*
1205  * Get an empty, disassociated buffer of given size.
1206  */
1207 buf_t *
1208 geteblk(int size)
1209 {
1210 	buf_t *bp;
1211 	int error;
1212 
1213 	mutex_enter(&bufcache_lock);
1214 	while ((bp = getnewbuf(0, 0, 0)) == NULL)
1215 		;
1216 
1217 	SET(bp->b_cflags, BC_INVAL);
1218 	LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1219 	mutex_exit(&bufcache_lock);
1220 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1221 	error = allocbuf(bp, size, 0);
1222 	KASSERT(error == 0);
1223 	return (bp);
1224 }
1225 
1226 /*
1227  * Expand or contract the actual memory allocated to a buffer.
1228  *
1229  * If the buffer shrinks, data is lost, so it's up to the
1230  * caller to have written it out *first*; this routine will not
1231  * start a write.  If the buffer grows, it's the callers
1232  * responsibility to fill out the buffer's additional contents.
1233  */
1234 int
1235 allocbuf(buf_t *bp, int size, int preserve)
1236 {
1237 	void *addr;
1238 	vsize_t oldsize, desired_size;
1239 	int oldcount;
1240 	int delta;
1241 
1242 	desired_size = buf_roundsize(size);
1243 	if (desired_size > MAXBSIZE)
1244 		printf("allocbuf: buffer larger than MAXBSIZE requested");
1245 
1246 	oldcount = bp->b_bcount;
1247 
1248 	bp->b_bcount = size;
1249 
1250 	oldsize = bp->b_bufsize;
1251 	if (oldsize == desired_size) {
1252 		/*
1253 		 * Do not short cut the WAPBL resize, as the buffer length
1254 		 * could still have changed and this would corrupt the
1255 		 * tracking of the transaction length.
1256 		 */
1257 		goto out;
1258 	}
1259 
1260 	/*
1261 	 * If we want a buffer of a different size, re-allocate the
1262 	 * buffer's memory; copy old content only if needed.
1263 	 */
1264 	addr = buf_alloc(desired_size);
1265 	if (addr == NULL)
1266 		return ENOMEM;
1267 	if (preserve)
1268 		memcpy(addr, bp->b_data, MIN(oldsize,desired_size));
1269 	if (bp->b_data != NULL)
1270 		buf_mrelease(bp->b_data, oldsize);
1271 	bp->b_data = addr;
1272 	bp->b_bufsize = desired_size;
1273 
1274 	/*
1275 	 * Update overall buffer memory counter (protected by bufcache_lock)
1276 	 */
1277 	delta = (long)desired_size - (long)oldsize;
1278 
1279 	mutex_enter(&bufcache_lock);
1280 	if ((bufmem += delta) > bufmem_hiwater) {
1281 		/*
1282 		 * Need to trim overall memory usage.
1283 		 */
1284 		while (buf_canrelease()) {
1285 			if (curcpu()->ci_schedstate.spc_flags &
1286 			    SPCF_SHOULDYIELD) {
1287 				mutex_exit(&bufcache_lock);
1288 				preempt();
1289 				mutex_enter(&bufcache_lock);
1290 			}
1291 			if (buf_trim() == 0)
1292 				break;
1293 		}
1294 	}
1295 	mutex_exit(&bufcache_lock);
1296 
1297  out:
1298 	if (wapbl_vphaswapbl(bp->b_vp))
1299 		WAPBL_RESIZE_BUF(wapbl_vptomp(bp->b_vp), bp, oldsize, oldcount);
1300 
1301 	return 0;
1302 }
1303 
1304 /*
1305  * Find a buffer which is available for use.
1306  * Select something from a free list.
1307  * Preference is to AGE list, then LRU list.
1308  *
1309  * Called with the buffer queues locked.
1310  * Return buffer locked.
1311  */
1312 buf_t *
1313 getnewbuf(int slpflag, int slptimeo, int from_bufq)
1314 {
1315 	buf_t *bp;
1316 	struct vnode *vp;
1317 
1318  start:
1319 	KASSERT(mutex_owned(&bufcache_lock));
1320 
1321 	/*
1322 	 * Get a new buffer from the pool.
1323 	 */
1324 	if (!from_bufq && buf_lotsfree()) {
1325 		mutex_exit(&bufcache_lock);
1326 		bp = pool_cache_get(buf_cache, PR_NOWAIT);
1327 		if (bp != NULL) {
1328 			memset((char *)bp, 0, sizeof(*bp));
1329 			buf_init(bp);
1330 			SET(bp->b_cflags, BC_BUSY);	/* mark buffer busy */
1331 			mutex_enter(&bufcache_lock);
1332 #if defined(DIAGNOSTIC)
1333 			bp->b_freelistindex = -1;
1334 #endif /* defined(DIAGNOSTIC) */
1335 			return (bp);
1336 		}
1337 		mutex_enter(&bufcache_lock);
1338 	}
1339 
1340 	KASSERT(mutex_owned(&bufcache_lock));
1341 	if ((bp = TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue)) != NULL ||
1342 	    (bp = TAILQ_FIRST(&bufqueues[BQ_LRU].bq_queue)) != NULL) {
1343 	    	KASSERT(!ISSET(bp->b_cflags, BC_BUSY) || ISSET(bp->b_cflags, BC_VFLUSH));
1344 		bremfree(bp);
1345 
1346 		/* Buffer is no longer on free lists. */
1347 		SET(bp->b_cflags, BC_BUSY);
1348 	} else {
1349 		/*
1350 		 * XXX: !from_bufq should be removed.
1351 		 */
1352 		if (!from_bufq || curlwp != uvm.pagedaemon_lwp) {
1353 			/* wait for a free buffer of any kind */
1354 			if ((slpflag & PCATCH) != 0)
1355 				(void)cv_timedwait_sig(&needbuffer_cv,
1356 				    &bufcache_lock, slptimeo);
1357 			else
1358 				(void)cv_timedwait(&needbuffer_cv,
1359 				    &bufcache_lock, slptimeo);
1360 		}
1361 		return (NULL);
1362 	}
1363 
1364 #ifdef DIAGNOSTIC
1365 	if (bp->b_bufsize <= 0)
1366 		panic("buffer %p: on queue but empty", bp);
1367 #endif
1368 
1369 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1370 		/*
1371 		 * This is a delayed write buffer being flushed to disk.  Make
1372 		 * sure it gets aged out of the queue when it's finished, and
1373 		 * leave it off the LRU queue.
1374 		 */
1375 		CLR(bp->b_cflags, BC_VFLUSH);
1376 		SET(bp->b_cflags, BC_AGE);
1377 		goto start;
1378 	}
1379 
1380 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1381 	KASSERT(bp->b_refcnt > 0);
1382     	KASSERT(!cv_has_waiters(&bp->b_done));
1383 
1384 	/*
1385 	 * If buffer was a delayed write, start it and return NULL
1386 	 * (since we might sleep while starting the write).
1387 	 */
1388 	if (ISSET(bp->b_oflags, BO_DELWRI)) {
1389 		/*
1390 		 * This buffer has gone through the LRU, so make sure it gets
1391 		 * reused ASAP.
1392 		 */
1393 		SET(bp->b_cflags, BC_AGE);
1394 		mutex_exit(&bufcache_lock);
1395 		bawrite(bp);
1396 		mutex_enter(&bufcache_lock);
1397 		return (NULL);
1398 	}
1399 
1400 	vp = bp->b_vp;
1401 
1402 	/* clear out various other fields */
1403 	bp->b_cflags = BC_BUSY;
1404 	bp->b_oflags = 0;
1405 	bp->b_flags = 0;
1406 	bp->b_dev = NODEV;
1407 	bp->b_blkno = 0;
1408 	bp->b_lblkno = 0;
1409 	bp->b_rawblkno = 0;
1410 	bp->b_iodone = 0;
1411 	bp->b_error = 0;
1412 	bp->b_resid = 0;
1413 	bp->b_bcount = 0;
1414 
1415 	LIST_REMOVE(bp, b_hash);
1416 
1417 	/* Disassociate us from our vnode, if we had one... */
1418 	if (vp != NULL) {
1419 		mutex_enter(vp->v_interlock);
1420 		brelvp(bp);
1421 		mutex_exit(vp->v_interlock);
1422 	}
1423 
1424 	return (bp);
1425 }
1426 
1427 /*
1428  * Attempt to free an aged buffer off the queues.
1429  * Called with queue lock held.
1430  * Returns the amount of buffer memory freed.
1431  */
1432 static int
1433 buf_trim(void)
1434 {
1435 	buf_t *bp;
1436 	long size = 0;
1437 
1438 	KASSERT(mutex_owned(&bufcache_lock));
1439 
1440 	/* Instruct getnewbuf() to get buffers off the queues */
1441 	if ((bp = getnewbuf(PCATCH, 1, 1)) == NULL)
1442 		return 0;
1443 
1444 	KASSERT((bp->b_cflags & BC_WANTED) == 0);
1445 	size = bp->b_bufsize;
1446 	bufmem -= size;
1447 	if (size > 0) {
1448 		buf_mrelease(bp->b_data, size);
1449 		bp->b_bcount = bp->b_bufsize = 0;
1450 	}
1451 	/* brelse() will return the buffer to the global buffer pool */
1452 	brelsel(bp, 0);
1453 	return size;
1454 }
1455 
1456 int
1457 buf_drain(int n)
1458 {
1459 	int size = 0, sz;
1460 
1461 	KASSERT(mutex_owned(&bufcache_lock));
1462 
1463 	while (size < n && bufmem > bufmem_lowater) {
1464 		sz = buf_trim();
1465 		if (sz <= 0)
1466 			break;
1467 		size += sz;
1468 	}
1469 
1470 	return size;
1471 }
1472 
1473 /*
1474  * Wait for operations on the buffer to complete.
1475  * When they do, extract and return the I/O's error value.
1476  */
1477 int
1478 biowait(buf_t *bp)
1479 {
1480 
1481 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1482 	KASSERT(bp->b_refcnt > 0);
1483 
1484 	mutex_enter(bp->b_objlock);
1485 	while (!ISSET(bp->b_oflags, BO_DONE | BO_DELWRI))
1486 		cv_wait(&bp->b_done, bp->b_objlock);
1487 	mutex_exit(bp->b_objlock);
1488 
1489 	return bp->b_error;
1490 }
1491 
1492 /*
1493  * Mark I/O complete on a buffer.
1494  *
1495  * If a callback has been requested, e.g. the pageout
1496  * daemon, do so. Otherwise, awaken waiting processes.
1497  *
1498  * [ Leffler, et al., says on p.247:
1499  *	"This routine wakes up the blocked process, frees the buffer
1500  *	for an asynchronous write, or, for a request by the pagedaemon
1501  *	process, invokes a procedure specified in the buffer structure" ]
1502  *
1503  * In real life, the pagedaemon (or other system processes) wants
1504  * to do async stuff to, and doesn't want the buffer brelse()'d.
1505  * (for swap pager, that puts swap buffers on the free lists (!!!),
1506  * for the vn device, that puts allocated buffers on the free lists!)
1507  */
1508 void
1509 biodone(buf_t *bp)
1510 {
1511 	int s;
1512 
1513 	KASSERT(!ISSET(bp->b_oflags, BO_DONE));
1514 
1515 	if (cpu_intr_p()) {
1516 		/* From interrupt mode: defer to a soft interrupt. */
1517 		s = splvm();
1518 		TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_biodone, bp, b_actq);
1519 		softint_schedule(biodone_sih);
1520 		splx(s);
1521 	} else {
1522 		/* Process now - the buffer may be freed soon. */
1523 		biodone2(bp);
1524 	}
1525 }
1526 
1527 static void
1528 biodone2(buf_t *bp)
1529 {
1530 	void (*callout)(buf_t *);
1531 
1532 	mutex_enter(bp->b_objlock);
1533 	/* Note that the transfer is done. */
1534 	if (ISSET(bp->b_oflags, BO_DONE))
1535 		panic("biodone2 already");
1536 	CLR(bp->b_flags, B_COWDONE);
1537 	SET(bp->b_oflags, BO_DONE);
1538 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1539 
1540 	/* Wake up waiting writers. */
1541 	if (!ISSET(bp->b_flags, B_READ))
1542 		vwakeup(bp);
1543 
1544 	if ((callout = bp->b_iodone) != NULL) {
1545 		/* Note callout done, then call out. */
1546 		KASSERT(!cv_has_waiters(&bp->b_done));
1547 		KERNEL_LOCK(1, NULL);		/* XXXSMP */
1548 		bp->b_iodone = NULL;
1549 		mutex_exit(bp->b_objlock);
1550 		(*callout)(bp);
1551 		KERNEL_UNLOCK_ONE(NULL);	/* XXXSMP */
1552 	} else if (ISSET(bp->b_flags, B_ASYNC)) {
1553 		/* If async, release. */
1554 		KASSERT(!cv_has_waiters(&bp->b_done));
1555 		mutex_exit(bp->b_objlock);
1556 		brelse(bp, 0);
1557 	} else {
1558 		/* Otherwise just wake up waiters in biowait(). */
1559 		cv_broadcast(&bp->b_done);
1560 		mutex_exit(bp->b_objlock);
1561 	}
1562 }
1563 
1564 static void
1565 biointr(void *cookie)
1566 {
1567 	struct cpu_info *ci;
1568 	buf_t *bp;
1569 	int s;
1570 
1571 	ci = curcpu();
1572 
1573 	while (!TAILQ_EMPTY(&ci->ci_data.cpu_biodone)) {
1574 		KASSERT(curcpu() == ci);
1575 
1576 		s = splvm();
1577 		bp = TAILQ_FIRST(&ci->ci_data.cpu_biodone);
1578 		TAILQ_REMOVE(&ci->ci_data.cpu_biodone, bp, b_actq);
1579 		splx(s);
1580 
1581 		biodone2(bp);
1582 	}
1583 }
1584 
1585 /*
1586  * Wait for all buffers to complete I/O
1587  * Return the number of "stuck" buffers.
1588  */
1589 int
1590 buf_syncwait(void)
1591 {
1592 	buf_t *bp;
1593 	int iter, nbusy, nbusy_prev = 0, dcount, ihash;
1594 
1595 	dcount = 10000;
1596 	for (iter = 0; iter < 20;) {
1597 		mutex_enter(&bufcache_lock);
1598 		nbusy = 0;
1599 		for (ihash = 0; ihash < bufhash+1; ihash++) {
1600 		    LIST_FOREACH(bp, &bufhashtbl[ihash], b_hash) {
1601 			if ((bp->b_cflags & (BC_BUSY|BC_INVAL)) == BC_BUSY)
1602 				nbusy += ((bp->b_flags & B_READ) == 0);
1603 		    }
1604 		}
1605 		mutex_exit(&bufcache_lock);
1606 
1607 		if (nbusy == 0)
1608 			break;
1609 		if (nbusy_prev == 0)
1610 			nbusy_prev = nbusy;
1611 		printf("%d ", nbusy);
1612 		kpause("bflush", false, MAX(1, hz / 25 * iter), NULL);
1613 		if (nbusy >= nbusy_prev) /* we didn't flush anything */
1614 			iter++;
1615 		else
1616 			nbusy_prev = nbusy;
1617 	}
1618 
1619 	if (nbusy) {
1620 #if defined(DEBUG) || defined(DEBUG_HALT_BUSY)
1621 		printf("giving up\nPrinting vnodes for busy buffers\n");
1622 		for (ihash = 0; ihash < bufhash+1; ihash++) {
1623 		    LIST_FOREACH(bp, &bufhashtbl[ihash], b_hash) {
1624 			if ((bp->b_cflags & (BC_BUSY|BC_INVAL)) == BC_BUSY &&
1625 			    (bp->b_flags & B_READ) == 0)
1626 				vprint(NULL, bp->b_vp);
1627 		    }
1628 		}
1629 #endif
1630 	}
1631 
1632 	return nbusy;
1633 }
1634 
1635 static void
1636 sysctl_fillbuf(buf_t *i, struct buf_sysctl *o)
1637 {
1638 
1639 	o->b_flags = i->b_flags | i->b_cflags | i->b_oflags;
1640 	o->b_error = i->b_error;
1641 	o->b_prio = i->b_prio;
1642 	o->b_dev = i->b_dev;
1643 	o->b_bufsize = i->b_bufsize;
1644 	o->b_bcount = i->b_bcount;
1645 	o->b_resid = i->b_resid;
1646 	o->b_addr = PTRTOUINT64(i->b_data);
1647 	o->b_blkno = i->b_blkno;
1648 	o->b_rawblkno = i->b_rawblkno;
1649 	o->b_iodone = PTRTOUINT64(i->b_iodone);
1650 	o->b_proc = PTRTOUINT64(i->b_proc);
1651 	o->b_vp = PTRTOUINT64(i->b_vp);
1652 	o->b_saveaddr = PTRTOUINT64(i->b_saveaddr);
1653 	o->b_lblkno = i->b_lblkno;
1654 }
1655 
1656 #define KERN_BUFSLOP 20
1657 static int
1658 sysctl_dobuf(SYSCTLFN_ARGS)
1659 {
1660 	buf_t *bp;
1661 	struct buf_sysctl bs;
1662 	struct bqueue *bq;
1663 	char *dp;
1664 	u_int i, op, arg;
1665 	size_t len, needed, elem_size, out_size;
1666 	int error, elem_count, retries;
1667 
1668 	if (namelen == 1 && name[0] == CTL_QUERY)
1669 		return (sysctl_query(SYSCTLFN_CALL(rnode)));
1670 
1671 	if (namelen != 4)
1672 		return (EINVAL);
1673 
1674 	retries = 100;
1675  retry:
1676 	dp = oldp;
1677 	len = (oldp != NULL) ? *oldlenp : 0;
1678 	op = name[0];
1679 	arg = name[1];
1680 	elem_size = name[2];
1681 	elem_count = name[3];
1682 	out_size = MIN(sizeof(bs), elem_size);
1683 
1684 	/*
1685 	 * at the moment, these are just "placeholders" to make the
1686 	 * API for retrieving kern.buf data more extensible in the
1687 	 * future.
1688 	 *
1689 	 * XXX kern.buf currently has "netbsd32" issues.  hopefully
1690 	 * these will be resolved at a later point.
1691 	 */
1692 	if (op != KERN_BUF_ALL || arg != KERN_BUF_ALL ||
1693 	    elem_size < 1 || elem_count < 0)
1694 		return (EINVAL);
1695 
1696 	error = 0;
1697 	needed = 0;
1698 	sysctl_unlock();
1699 	mutex_enter(&bufcache_lock);
1700 	for (i = 0; i < BQUEUES; i++) {
1701 		bq = &bufqueues[i];
1702 		TAILQ_FOREACH(bp, &bq->bq_queue, b_freelist) {
1703 			bq->bq_marker = bp;
1704 			if (len >= elem_size && elem_count > 0) {
1705 				sysctl_fillbuf(bp, &bs);
1706 				mutex_exit(&bufcache_lock);
1707 				error = copyout(&bs, dp, out_size);
1708 				mutex_enter(&bufcache_lock);
1709 				if (error)
1710 					break;
1711 				if (bq->bq_marker != bp) {
1712 					/*
1713 					 * This sysctl node is only for
1714 					 * statistics.  Retry; if the
1715 					 * queue keeps changing, then
1716 					 * bail out.
1717 					 */
1718 					if (retries-- == 0) {
1719 						error = EAGAIN;
1720 						break;
1721 					}
1722 					mutex_exit(&bufcache_lock);
1723 					sysctl_relock();
1724 					goto retry;
1725 				}
1726 				dp += elem_size;
1727 				len -= elem_size;
1728 			}
1729 			needed += elem_size;
1730 			if (elem_count > 0 && elem_count != INT_MAX)
1731 				elem_count--;
1732 		}
1733 		if (error != 0)
1734 			break;
1735 	}
1736 	mutex_exit(&bufcache_lock);
1737 	sysctl_relock();
1738 
1739 	*oldlenp = needed;
1740 	if (oldp == NULL)
1741 		*oldlenp += KERN_BUFSLOP * sizeof(buf_t);
1742 
1743 	return (error);
1744 }
1745 
1746 static int
1747 sysctl_bufvm_update(SYSCTLFN_ARGS)
1748 {
1749 	int error, rv;
1750 	struct sysctlnode node;
1751 	unsigned int temp_bufcache;
1752 	unsigned long temp_water;
1753 
1754 	/* Take a copy of the supplied node and its data */
1755 	node = *rnode;
1756 	if (node.sysctl_data == &bufcache) {
1757 	    node.sysctl_data = &temp_bufcache;
1758 	    temp_bufcache = *(unsigned int *)rnode->sysctl_data;
1759 	} else {
1760 	    node.sysctl_data = &temp_water;
1761 	    temp_water = *(unsigned long *)rnode->sysctl_data;
1762 	}
1763 
1764 	/* Update the copy */
1765 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
1766 	if (error || newp == NULL)
1767 		return (error);
1768 
1769 	if (rnode->sysctl_data == &bufcache) {
1770 		if (temp_bufcache > 100)
1771 			return (EINVAL);
1772 		bufcache = temp_bufcache;
1773 		buf_setwm();
1774 	} else if (rnode->sysctl_data == &bufmem_lowater) {
1775 		if (bufmem_hiwater - temp_water < 16)
1776 			return (EINVAL);
1777 		bufmem_lowater = temp_water;
1778 	} else if (rnode->sysctl_data == &bufmem_hiwater) {
1779 		if (temp_water - bufmem_lowater < 16)
1780 			return (EINVAL);
1781 		bufmem_hiwater = temp_water;
1782 	} else
1783 		return (EINVAL);
1784 
1785 	/* Drain until below new high water mark */
1786 	sysctl_unlock();
1787 	mutex_enter(&bufcache_lock);
1788 	while (bufmem > bufmem_hiwater) {
1789 		rv = buf_drain((bufmem - bufmem_hiwater) / (2 * 1024));
1790 		if (rv <= 0)
1791 			break;
1792 	}
1793 	mutex_exit(&bufcache_lock);
1794 	sysctl_relock();
1795 
1796 	return 0;
1797 }
1798 
1799 static struct sysctllog *vfsbio_sysctllog;
1800 
1801 static void
1802 sysctl_kern_buf_setup(void)
1803 {
1804 
1805 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1806 		       CTLFLAG_PERMANENT,
1807 		       CTLTYPE_NODE, "kern", NULL,
1808 		       NULL, 0, NULL, 0,
1809 		       CTL_KERN, CTL_EOL);
1810 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1811 		       CTLFLAG_PERMANENT,
1812 		       CTLTYPE_NODE, "buf",
1813 		       SYSCTL_DESCR("Kernel buffer cache information"),
1814 		       sysctl_dobuf, 0, NULL, 0,
1815 		       CTL_KERN, KERN_BUF, CTL_EOL);
1816 }
1817 
1818 static void
1819 sysctl_vm_buf_setup(void)
1820 {
1821 
1822 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1823 		       CTLFLAG_PERMANENT,
1824 		       CTLTYPE_NODE, "vm", NULL,
1825 		       NULL, 0, NULL, 0,
1826 		       CTL_VM, CTL_EOL);
1827 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1828 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1829 		       CTLTYPE_INT, "bufcache",
1830 		       SYSCTL_DESCR("Percentage of physical memory to use for "
1831 				    "buffer cache"),
1832 		       sysctl_bufvm_update, 0, &bufcache, 0,
1833 		       CTL_VM, CTL_CREATE, CTL_EOL);
1834 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1835 		       CTLFLAG_PERMANENT|CTLFLAG_READONLY,
1836 		       CTLTYPE_LONG, "bufmem",
1837 		       SYSCTL_DESCR("Amount of kernel memory used by buffer "
1838 				    "cache"),
1839 		       NULL, 0, &bufmem, 0,
1840 		       CTL_VM, CTL_CREATE, CTL_EOL);
1841 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1842 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1843 		       CTLTYPE_LONG, "bufmem_lowater",
1844 		       SYSCTL_DESCR("Minimum amount of kernel memory to "
1845 				    "reserve for buffer cache"),
1846 		       sysctl_bufvm_update, 0, &bufmem_lowater, 0,
1847 		       CTL_VM, CTL_CREATE, CTL_EOL);
1848 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1849 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1850 		       CTLTYPE_LONG, "bufmem_hiwater",
1851 		       SYSCTL_DESCR("Maximum amount of kernel memory to use "
1852 				    "for buffer cache"),
1853 		       sysctl_bufvm_update, 0, &bufmem_hiwater, 0,
1854 		       CTL_VM, CTL_CREATE, CTL_EOL);
1855 }
1856 
1857 #ifdef DEBUG
1858 /*
1859  * Print out statistics on the current allocation of the buffer pool.
1860  * Can be enabled to print out on every ``sync'' by setting "syncprt"
1861  * in vfs_syscalls.c using sysctl.
1862  */
1863 void
1864 vfs_bufstats(void)
1865 {
1866 	int i, j, count;
1867 	buf_t *bp;
1868 	struct bqueue *dp;
1869 	int counts[(MAXBSIZE / PAGE_SIZE) + 1];
1870 	static const char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE" };
1871 
1872 	for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) {
1873 		count = 0;
1874 		for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
1875 			counts[j] = 0;
1876 		TAILQ_FOREACH(bp, &dp->bq_queue, b_freelist) {
1877 			counts[bp->b_bufsize/PAGE_SIZE]++;
1878 			count++;
1879 		}
1880 		printf("%s: total-%d", bname[i], count);
1881 		for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
1882 			if (counts[j] != 0)
1883 				printf(", %d-%d", j * PAGE_SIZE, counts[j]);
1884 		printf("\n");
1885 	}
1886 }
1887 #endif /* DEBUG */
1888 
1889 /* ------------------------------ */
1890 
1891 buf_t *
1892 getiobuf(struct vnode *vp, bool waitok)
1893 {
1894 	buf_t *bp;
1895 
1896 	bp = pool_cache_get(bufio_cache, (waitok ? PR_WAITOK : PR_NOWAIT));
1897 	if (bp == NULL)
1898 		return bp;
1899 
1900 	buf_init(bp);
1901 
1902 	if ((bp->b_vp = vp) == NULL)
1903 		bp->b_objlock = &buffer_lock;
1904 	else
1905 		bp->b_objlock = vp->v_interlock;
1906 
1907 	return bp;
1908 }
1909 
1910 void
1911 putiobuf(buf_t *bp)
1912 {
1913 
1914 	buf_destroy(bp);
1915 	pool_cache_put(bufio_cache, bp);
1916 }
1917 
1918 /*
1919  * nestiobuf_iodone: b_iodone callback for nested buffers.
1920  */
1921 
1922 void
1923 nestiobuf_iodone(buf_t *bp)
1924 {
1925 	buf_t *mbp = bp->b_private;
1926 	int error;
1927 	int donebytes;
1928 
1929 	KASSERT(bp->b_bcount <= bp->b_bufsize);
1930 	KASSERT(mbp != bp);
1931 
1932 	error = bp->b_error;
1933 	if (bp->b_error == 0 &&
1934 	    (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
1935 		/*
1936 		 * Not all got transfered, raise an error. We have no way to
1937 		 * propagate these conditions to mbp.
1938 		 */
1939 		error = EIO;
1940 	}
1941 
1942 	donebytes = bp->b_bufsize;
1943 
1944 	putiobuf(bp);
1945 	nestiobuf_done(mbp, donebytes, error);
1946 }
1947 
1948 /*
1949  * nestiobuf_setup: setup a "nested" buffer.
1950  *
1951  * => 'mbp' is a "master" buffer which is being divided into sub pieces.
1952  * => 'bp' should be a buffer allocated by getiobuf.
1953  * => 'offset' is a byte offset in the master buffer.
1954  * => 'size' is a size in bytes of this nested buffer.
1955  */
1956 
1957 void
1958 nestiobuf_setup(buf_t *mbp, buf_t *bp, int offset, size_t size)
1959 {
1960 	const int b_read = mbp->b_flags & B_READ;
1961 	struct vnode *vp = mbp->b_vp;
1962 
1963 	KASSERT(mbp->b_bcount >= offset + size);
1964 	bp->b_vp = vp;
1965 	bp->b_dev = mbp->b_dev;
1966 	bp->b_objlock = mbp->b_objlock;
1967 	bp->b_cflags = BC_BUSY;
1968 	bp->b_flags = B_ASYNC | b_read;
1969 	bp->b_iodone = nestiobuf_iodone;
1970 	bp->b_data = (char *)mbp->b_data + offset;
1971 	bp->b_resid = bp->b_bcount = size;
1972 	bp->b_bufsize = bp->b_bcount;
1973 	bp->b_private = mbp;
1974 	BIO_COPYPRIO(bp, mbp);
1975 	if (!b_read && vp != NULL) {
1976 		mutex_enter(vp->v_interlock);
1977 		vp->v_numoutput++;
1978 		mutex_exit(vp->v_interlock);
1979 	}
1980 }
1981 
1982 /*
1983  * nestiobuf_done: propagate completion to the master buffer.
1984  *
1985  * => 'donebytes' specifies how many bytes in the 'mbp' is completed.
1986  * => 'error' is an errno(2) that 'donebytes' has been completed with.
1987  */
1988 
1989 void
1990 nestiobuf_done(buf_t *mbp, int donebytes, int error)
1991 {
1992 
1993 	if (donebytes == 0) {
1994 		return;
1995 	}
1996 	mutex_enter(mbp->b_objlock);
1997 	KASSERT(mbp->b_resid >= donebytes);
1998 	mbp->b_resid -= donebytes;
1999 	if (error)
2000 		mbp->b_error = error;
2001 	if (mbp->b_resid == 0) {
2002 		if (mbp->b_error)
2003 			mbp->b_resid = mbp->b_bcount;
2004 		mutex_exit(mbp->b_objlock);
2005 		biodone(mbp);
2006 	} else
2007 		mutex_exit(mbp->b_objlock);
2008 }
2009 
2010 void
2011 buf_init(buf_t *bp)
2012 {
2013 
2014 	cv_init(&bp->b_busy, "biolock");
2015 	cv_init(&bp->b_done, "biowait");
2016 	bp->b_dev = NODEV;
2017 	bp->b_error = 0;
2018 	bp->b_flags = 0;
2019 	bp->b_cflags = 0;
2020 	bp->b_oflags = 0;
2021 	bp->b_objlock = &buffer_lock;
2022 	bp->b_iodone = NULL;
2023 	bp->b_refcnt = 1;
2024 	bp->b_dev = NODEV;
2025 	bp->b_vnbufs.le_next = NOLIST;
2026 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
2027 }
2028 
2029 void
2030 buf_destroy(buf_t *bp)
2031 {
2032 
2033 	cv_destroy(&bp->b_done);
2034 	cv_destroy(&bp->b_busy);
2035 }
2036 
2037 int
2038 bbusy(buf_t *bp, bool intr, int timo, kmutex_t *interlock)
2039 {
2040 	int error;
2041 
2042 	KASSERT(mutex_owned(&bufcache_lock));
2043 
2044 	if ((bp->b_cflags & BC_BUSY) != 0) {
2045 		if (curlwp == uvm.pagedaemon_lwp)
2046 			return EDEADLK;
2047 		bp->b_cflags |= BC_WANTED;
2048 		bref(bp);
2049 		if (interlock != NULL)
2050 			mutex_exit(interlock);
2051 		if (intr) {
2052 			error = cv_timedwait_sig(&bp->b_busy, &bufcache_lock,
2053 			    timo);
2054 		} else {
2055 			error = cv_timedwait(&bp->b_busy, &bufcache_lock,
2056 			    timo);
2057 		}
2058 		brele(bp);
2059 		if (interlock != NULL)
2060 			mutex_enter(interlock);
2061 		if (error != 0)
2062 			return error;
2063 		return EPASSTHROUGH;
2064 	}
2065 	bp->b_cflags |= BC_BUSY;
2066 
2067 	return 0;
2068 }
2069