xref: /netbsd-src/sys/kern/vfs_bio.c (revision d909946ca08dceb44d7d0f22ec9488679695d976)
1 /*	$NetBSD: vfs_bio.c,v 1.260 2016/07/31 04:05:32 dholland 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 1.260 2016/07/31 04:05:32 dholland Exp $");
127 
128 #ifdef _KERNEL_OPT
129 #include "opt_bufcache.h"
130 #include "opt_dtrace.h"
131 #endif
132 
133 #include <sys/param.h>
134 #include <sys/systm.h>
135 #include <sys/kernel.h>
136 #include <sys/proc.h>
137 #include <sys/buf.h>
138 #include <sys/vnode.h>
139 #include <sys/mount.h>
140 #include <sys/resourcevar.h>
141 #include <sys/sysctl.h>
142 #include <sys/conf.h>
143 #include <sys/kauth.h>
144 #include <sys/fstrans.h>
145 #include <sys/intr.h>
146 #include <sys/cpu.h>
147 #include <sys/wapbl.h>
148 #include <sys/bitops.h>
149 #include <sys/cprng.h>
150 #include <sys/sdt.h>
151 
152 #include <uvm/uvm.h>	/* extern struct uvm uvm */
153 
154 #include <miscfs/specfs/specdev.h>
155 
156 #ifndef	BUFPAGES
157 # define BUFPAGES 0
158 #endif
159 
160 #ifdef BUFCACHE
161 # if (BUFCACHE < 5) || (BUFCACHE > 95)
162 #  error BUFCACHE is not between 5 and 95
163 # endif
164 #else
165 # define BUFCACHE 15
166 #endif
167 
168 u_int	nbuf;			/* desired number of buffer headers */
169 u_int	bufpages = BUFPAGES;	/* optional hardwired count */
170 u_int	bufcache = BUFCACHE;	/* max % of RAM to use for buffer cache */
171 
172 /* Function prototypes */
173 struct bqueue;
174 
175 static void buf_setwm(void);
176 static int buf_trim(void);
177 static void *bufpool_page_alloc(struct pool *, int);
178 static void bufpool_page_free(struct pool *, void *);
179 static buf_t *bio_doread(struct vnode *, daddr_t, int, int);
180 static buf_t *getnewbuf(int, int, int);
181 static int buf_lotsfree(void);
182 static int buf_canrelease(void);
183 static u_long buf_mempoolidx(u_long);
184 static u_long buf_roundsize(u_long);
185 static void *buf_alloc(size_t);
186 static void buf_mrelease(void *, size_t);
187 static void binsheadfree(buf_t *, struct bqueue *);
188 static void binstailfree(buf_t *, struct bqueue *);
189 #ifdef DEBUG
190 static int checkfreelist(buf_t *, struct bqueue *, int);
191 #endif
192 static void biointr(void *);
193 static void biodone2(buf_t *);
194 static void bref(buf_t *);
195 static void brele(buf_t *);
196 static void sysctl_kern_buf_setup(void);
197 static void sysctl_vm_buf_setup(void);
198 
199 /*
200  * Definitions for the buffer hash lists.
201  */
202 #define	BUFHASH(dvp, lbn)	\
203 	(&bufhashtbl[(((long)(dvp) >> 8) + (int)(lbn)) & bufhash])
204 LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
205 u_long	bufhash;
206 struct bqueue bufqueues[BQUEUES];
207 
208 static kcondvar_t needbuffer_cv;
209 
210 /*
211  * Buffer queue lock.
212  */
213 kmutex_t bufcache_lock;
214 kmutex_t buffer_lock;
215 
216 /* Software ISR for completed transfers. */
217 static void *biodone_sih;
218 
219 /* Buffer pool for I/O buffers. */
220 static pool_cache_t buf_cache;
221 static pool_cache_t bufio_cache;
222 
223 #define MEMPOOL_INDEX_OFFSET (ilog2(DEV_BSIZE))	/* smallest pool is 512 bytes */
224 #define NMEMPOOLS (ilog2(MAXBSIZE) - MEMPOOL_INDEX_OFFSET + 1)
225 __CTASSERT((1 << (NMEMPOOLS + MEMPOOL_INDEX_OFFSET - 1)) == MAXBSIZE);
226 
227 /* Buffer memory pools */
228 static struct pool bmempools[NMEMPOOLS];
229 
230 static struct vm_map *buf_map;
231 
232 /*
233  * Buffer memory pool allocator.
234  */
235 static void *
236 bufpool_page_alloc(struct pool *pp, int flags)
237 {
238 
239 	return (void *)uvm_km_alloc(buf_map,
240 	    MAXBSIZE, MAXBSIZE,
241 	    ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT|UVM_KMF_TRYLOCK)
242 	    | UVM_KMF_WIRED);
243 }
244 
245 static void
246 bufpool_page_free(struct pool *pp, void *v)
247 {
248 
249 	uvm_km_free(buf_map, (vaddr_t)v, MAXBSIZE, UVM_KMF_WIRED);
250 }
251 
252 static struct pool_allocator bufmempool_allocator = {
253 	.pa_alloc = bufpool_page_alloc,
254 	.pa_free = bufpool_page_free,
255 	.pa_pagesz = MAXBSIZE,
256 };
257 
258 /* Buffer memory management variables */
259 u_long bufmem_valimit;
260 u_long bufmem_hiwater;
261 u_long bufmem_lowater;
262 u_long bufmem;
263 
264 /*
265  * MD code can call this to set a hard limit on the amount
266  * of virtual memory used by the buffer cache.
267  */
268 int
269 buf_setvalimit(vsize_t sz)
270 {
271 
272 	/* We need to accommodate at least NMEMPOOLS of MAXBSIZE each */
273 	if (sz < NMEMPOOLS * MAXBSIZE)
274 		return EINVAL;
275 
276 	bufmem_valimit = sz;
277 	return 0;
278 }
279 
280 static void
281 buf_setwm(void)
282 {
283 
284 	bufmem_hiwater = buf_memcalc();
285 	/* lowater is approx. 2% of memory (with bufcache = 15) */
286 #define	BUFMEM_WMSHIFT	3
287 #define	BUFMEM_HIWMMIN	(64 * 1024 << BUFMEM_WMSHIFT)
288 	if (bufmem_hiwater < BUFMEM_HIWMMIN)
289 		/* Ensure a reasonable minimum value */
290 		bufmem_hiwater = BUFMEM_HIWMMIN;
291 	bufmem_lowater = bufmem_hiwater >> BUFMEM_WMSHIFT;
292 }
293 
294 #ifdef DEBUG
295 int debug_verify_freelist = 0;
296 static int
297 checkfreelist(buf_t *bp, struct bqueue *dp, int ison)
298 {
299 	buf_t *b;
300 
301 	if (!debug_verify_freelist)
302 		return 1;
303 
304 	TAILQ_FOREACH(b, &dp->bq_queue, b_freelist) {
305 		if (b == bp)
306 			return ison ? 1 : 0;
307 	}
308 
309 	return ison ? 0 : 1;
310 }
311 #endif
312 
313 /*
314  * Insq/Remq for the buffer hash lists.
315  * Call with buffer queue locked.
316  */
317 static void
318 binsheadfree(buf_t *bp, struct bqueue *dp)
319 {
320 
321 	KASSERT(mutex_owned(&bufcache_lock));
322 	KASSERT(bp->b_freelistindex == -1);
323 	TAILQ_INSERT_HEAD(&dp->bq_queue, bp, b_freelist);
324 	dp->bq_bytes += bp->b_bufsize;
325 	bp->b_freelistindex = dp - bufqueues;
326 }
327 
328 static void
329 binstailfree(buf_t *bp, struct bqueue *dp)
330 {
331 
332 	KASSERT(mutex_owned(&bufcache_lock));
333 	KASSERTMSG(bp->b_freelistindex == -1, "double free of buffer? "
334 	    "bp=%p, b_freelistindex=%d\n", bp, bp->b_freelistindex);
335 	TAILQ_INSERT_TAIL(&dp->bq_queue, bp, b_freelist);
336 	dp->bq_bytes += bp->b_bufsize;
337 	bp->b_freelistindex = dp - bufqueues;
338 }
339 
340 void
341 bremfree(buf_t *bp)
342 {
343 	struct bqueue *dp;
344 	int bqidx = bp->b_freelistindex;
345 
346 	KASSERT(mutex_owned(&bufcache_lock));
347 
348 	KASSERT(bqidx != -1);
349 	dp = &bufqueues[bqidx];
350 	KDASSERT(checkfreelist(bp, dp, 1));
351 	KASSERT(dp->bq_bytes >= bp->b_bufsize);
352 	TAILQ_REMOVE(&dp->bq_queue, bp, b_freelist);
353 	dp->bq_bytes -= bp->b_bufsize;
354 
355 	/* For the sysctl helper. */
356 	if (bp == dp->bq_marker)
357 		dp->bq_marker = NULL;
358 
359 #if defined(DIAGNOSTIC)
360 	bp->b_freelistindex = -1;
361 #endif /* defined(DIAGNOSTIC) */
362 }
363 
364 /*
365  * Add a reference to an buffer structure that came from buf_cache.
366  */
367 static inline void
368 bref(buf_t *bp)
369 {
370 
371 	KASSERT(mutex_owned(&bufcache_lock));
372 	KASSERT(bp->b_refcnt > 0);
373 
374 	bp->b_refcnt++;
375 }
376 
377 /*
378  * Free an unused buffer structure that came from buf_cache.
379  */
380 static inline void
381 brele(buf_t *bp)
382 {
383 
384 	KASSERT(mutex_owned(&bufcache_lock));
385 	KASSERT(bp->b_refcnt > 0);
386 
387 	if (bp->b_refcnt-- == 1) {
388 		buf_destroy(bp);
389 #ifdef DEBUG
390 		memset((char *)bp, 0, sizeof(*bp));
391 #endif
392 		pool_cache_put(buf_cache, bp);
393 	}
394 }
395 
396 /*
397  * note that for some ports this is used by pmap bootstrap code to
398  * determine kva size.
399  */
400 u_long
401 buf_memcalc(void)
402 {
403 	u_long n;
404 	vsize_t mapsz = 0;
405 
406 	/*
407 	 * Determine the upper bound of memory to use for buffers.
408 	 *
409 	 *	- If bufpages is specified, use that as the number
410 	 *	  pages.
411 	 *
412 	 *	- Otherwise, use bufcache as the percentage of
413 	 *	  physical memory.
414 	 */
415 	if (bufpages != 0) {
416 		n = bufpages;
417 	} else {
418 		if (bufcache < 5) {
419 			printf("forcing bufcache %d -> 5", bufcache);
420 			bufcache = 5;
421 		}
422 		if (bufcache > 95) {
423 			printf("forcing bufcache %d -> 95", bufcache);
424 			bufcache = 95;
425 		}
426 		if (buf_map != NULL)
427 			mapsz = vm_map_max(buf_map) - vm_map_min(buf_map);
428 		n = calc_cache_size(mapsz, bufcache,
429 		    (buf_map != kernel_map) ? 100 : BUFCACHE_VA_MAXPCT)
430 		    / PAGE_SIZE;
431 	}
432 
433 	n <<= PAGE_SHIFT;
434 	if (bufmem_valimit != 0 && n > bufmem_valimit)
435 		n = bufmem_valimit;
436 
437 	return (n);
438 }
439 
440 /*
441  * Initialize buffers and hash links for buffers.
442  */
443 void
444 bufinit(void)
445 {
446 	struct bqueue *dp;
447 	int use_std;
448 	u_int i;
449 
450 	biodone_vfs = biodone;
451 
452 	mutex_init(&bufcache_lock, MUTEX_DEFAULT, IPL_NONE);
453 	mutex_init(&buffer_lock, MUTEX_DEFAULT, IPL_NONE);
454 	cv_init(&needbuffer_cv, "needbuf");
455 
456 	if (bufmem_valimit != 0) {
457 		vaddr_t minaddr = 0, maxaddr;
458 		buf_map = uvm_km_suballoc(kernel_map, &minaddr, &maxaddr,
459 					  bufmem_valimit, 0, false, 0);
460 		if (buf_map == NULL)
461 			panic("bufinit: cannot allocate submap");
462 	} else
463 		buf_map = kernel_map;
464 
465 	/*
466 	 * Initialize buffer cache memory parameters.
467 	 */
468 	bufmem = 0;
469 	buf_setwm();
470 
471 	/* On "small" machines use small pool page sizes where possible */
472 	use_std = (physmem < atop(16*1024*1024));
473 
474 	/*
475 	 * Also use them on systems that can map the pool pages using
476 	 * a direct-mapped segment.
477 	 */
478 #ifdef PMAP_MAP_POOLPAGE
479 	use_std = 1;
480 #endif
481 
482 	buf_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
483 	    "bufpl", NULL, IPL_SOFTBIO, NULL, NULL, NULL);
484 	bufio_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
485 	    "biopl", NULL, IPL_BIO, NULL, NULL, NULL);
486 
487 	for (i = 0; i < NMEMPOOLS; i++) {
488 		struct pool_allocator *pa;
489 		struct pool *pp = &bmempools[i];
490 		u_int size = 1 << (i + MEMPOOL_INDEX_OFFSET);
491 		char *name = kmem_alloc(8, KM_SLEEP); /* XXX: never freed */
492 		if (__predict_false(size >= 1048576))
493 			(void)snprintf(name, 8, "buf%um", size / 1048576);
494 		else if (__predict_true(size >= 1024))
495 			(void)snprintf(name, 8, "buf%uk", size / 1024);
496 		else
497 			(void)snprintf(name, 8, "buf%ub", size);
498 		pa = (size <= PAGE_SIZE && use_std)
499 			? &pool_allocator_nointr
500 			: &bufmempool_allocator;
501 		pool_init(pp, size, 0, 0, 0, name, pa, IPL_NONE);
502 		pool_setlowat(pp, 1);
503 		pool_sethiwat(pp, 1);
504 	}
505 
506 	/* Initialize the buffer queues */
507 	for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) {
508 		TAILQ_INIT(&dp->bq_queue);
509 		dp->bq_bytes = 0;
510 	}
511 
512 	/*
513 	 * Estimate hash table size based on the amount of memory we
514 	 * intend to use for the buffer cache. The average buffer
515 	 * size is dependent on our clients (i.e. filesystems).
516 	 *
517 	 * For now, use an empirical 3K per buffer.
518 	 */
519 	nbuf = (bufmem_hiwater / 1024) / 3;
520 	bufhashtbl = hashinit(nbuf, HASH_LIST, true, &bufhash);
521 
522 	sysctl_kern_buf_setup();
523 	sysctl_vm_buf_setup();
524 }
525 
526 void
527 bufinit2(void)
528 {
529 
530 	biodone_sih = softint_establish(SOFTINT_BIO | SOFTINT_MPSAFE, biointr,
531 	    NULL);
532 	if (biodone_sih == NULL)
533 		panic("bufinit2: can't establish soft interrupt");
534 }
535 
536 static int
537 buf_lotsfree(void)
538 {
539 	u_long guess;
540 
541 	/* Always allocate if less than the low water mark. */
542 	if (bufmem < bufmem_lowater)
543 		return 1;
544 
545 	/* Never allocate if greater than the high water mark. */
546 	if (bufmem > bufmem_hiwater)
547 		return 0;
548 
549 	/* If there's anything on the AGE list, it should be eaten. */
550 	if (TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue) != NULL)
551 		return 0;
552 
553 	/*
554 	 * The probabily of getting a new allocation is inversely
555 	 * proportional  to the current size of the cache above
556 	 * the low water mark.  Divide the total first to avoid overflows
557 	 * in the product.
558 	 */
559 	guess = cprng_fast32() % 16;
560 
561 	if ((bufmem_hiwater - bufmem_lowater) / 16 * guess >=
562 	    (bufmem - bufmem_lowater))
563 		return 1;
564 
565 	/* Otherwise don't allocate. */
566 	return 0;
567 }
568 
569 /*
570  * Return estimate of bytes we think need to be
571  * released to help resolve low memory conditions.
572  *
573  * => called with bufcache_lock held.
574  */
575 static int
576 buf_canrelease(void)
577 {
578 	int pagedemand, ninvalid = 0;
579 
580 	KASSERT(mutex_owned(&bufcache_lock));
581 
582 	if (bufmem < bufmem_lowater)
583 		return 0;
584 
585 	if (bufmem > bufmem_hiwater)
586 		return bufmem - bufmem_hiwater;
587 
588 	ninvalid += bufqueues[BQ_AGE].bq_bytes;
589 
590 	pagedemand = uvmexp.freetarg - uvmexp.free;
591 	if (pagedemand < 0)
592 		return ninvalid;
593 	return MAX(ninvalid, MIN(2 * MAXBSIZE,
594 	    MIN((bufmem - bufmem_lowater) / 16, pagedemand * PAGE_SIZE)));
595 }
596 
597 /*
598  * Buffer memory allocation helper functions
599  */
600 static u_long
601 buf_mempoolidx(u_long size)
602 {
603 	u_int n = 0;
604 
605 	size -= 1;
606 	size >>= MEMPOOL_INDEX_OFFSET;
607 	while (size) {
608 		size >>= 1;
609 		n += 1;
610 	}
611 	if (n >= NMEMPOOLS)
612 		panic("buf mem pool index %d", n);
613 	return n;
614 }
615 
616 static u_long
617 buf_roundsize(u_long size)
618 {
619 	/* Round up to nearest power of 2 */
620 	return (1 << (buf_mempoolidx(size) + MEMPOOL_INDEX_OFFSET));
621 }
622 
623 static void *
624 buf_alloc(size_t size)
625 {
626 	u_int n = buf_mempoolidx(size);
627 	void *addr;
628 
629 	while (1) {
630 		addr = pool_get(&bmempools[n], PR_NOWAIT);
631 		if (addr != NULL)
632 			break;
633 
634 		/* No memory, see if we can free some. If so, try again */
635 		mutex_enter(&bufcache_lock);
636 		if (buf_drain(1) > 0) {
637 			mutex_exit(&bufcache_lock);
638 			continue;
639 		}
640 
641 		if (curlwp == uvm.pagedaemon_lwp) {
642 			mutex_exit(&bufcache_lock);
643 			return NULL;
644 		}
645 
646 		/* Wait for buffers to arrive on the LRU queue */
647 		cv_timedwait(&needbuffer_cv, &bufcache_lock, hz / 4);
648 		mutex_exit(&bufcache_lock);
649 	}
650 
651 	return addr;
652 }
653 
654 static void
655 buf_mrelease(void *addr, size_t size)
656 {
657 
658 	pool_put(&bmempools[buf_mempoolidx(size)], addr);
659 }
660 
661 /*
662  * bread()/breadn() helper.
663  */
664 static buf_t *
665 bio_doread(struct vnode *vp, daddr_t blkno, int size, int async)
666 {
667 	buf_t *bp;
668 	struct mount *mp;
669 
670 	bp = getblk(vp, blkno, size, 0, 0);
671 
672 	/*
673 	 * getblk() may return NULL if we are the pagedaemon.
674 	 */
675 	if (bp == NULL) {
676 		KASSERT(curlwp == uvm.pagedaemon_lwp);
677 		return NULL;
678 	}
679 
680 	/*
681 	 * If buffer does not have data valid, start a read.
682 	 * Note that if buffer is BC_INVAL, getblk() won't return it.
683 	 * Therefore, it's valid if its I/O has completed or been delayed.
684 	 */
685 	if (!ISSET(bp->b_oflags, (BO_DONE | BO_DELWRI))) {
686 		/* Start I/O for the buffer. */
687 		SET(bp->b_flags, B_READ | async);
688 		if (async)
689 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
690 		else
691 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
692 		VOP_STRATEGY(vp, bp);
693 
694 		/* Pay for the read. */
695 		curlwp->l_ru.ru_inblock++;
696 	} else if (async)
697 		brelse(bp, 0);
698 
699 	if (vp->v_type == VBLK)
700 		mp = spec_node_getmountedfs(vp);
701 	else
702 		mp = vp->v_mount;
703 
704 	/*
705 	 * Collect statistics on synchronous and asynchronous reads.
706 	 * Reads from block devices are charged to their associated
707 	 * filesystem (if any).
708 	 */
709 	if (mp != NULL) {
710 		if (async == 0)
711 			mp->mnt_stat.f_syncreads++;
712 		else
713 			mp->mnt_stat.f_asyncreads++;
714 	}
715 
716 	return (bp);
717 }
718 
719 /*
720  * Read a disk block.
721  * This algorithm described in Bach (p.54).
722  */
723 int
724 bread(struct vnode *vp, daddr_t blkno, int size, int flags, buf_t **bpp)
725 {
726 	buf_t *bp;
727 	int error;
728 
729 	/* Get buffer for block. */
730 	bp = *bpp = bio_doread(vp, blkno, size, 0);
731 	if (bp == NULL)
732 		return ENOMEM;
733 
734 	/* Wait for the read to complete, and return result. */
735 	error = biowait(bp);
736 	if (error == 0 && (flags & B_MODIFY) != 0)
737 		error = fscow_run(bp, true);
738 	if (error) {
739 		brelse(bp, 0);
740 		*bpp = NULL;
741 	}
742 
743 	return error;
744 }
745 
746 /*
747  * Read-ahead multiple disk blocks. The first is sync, the rest async.
748  * Trivial modification to the breada algorithm presented in Bach (p.55).
749  */
750 int
751 breadn(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablks,
752     int *rasizes, int nrablks, int flags, buf_t **bpp)
753 {
754 	buf_t *bp;
755 	int error, i;
756 
757 	bp = *bpp = bio_doread(vp, blkno, size, 0);
758 	if (bp == NULL)
759 		return ENOMEM;
760 
761 	/*
762 	 * For each of the read-ahead blocks, start a read, if necessary.
763 	 */
764 	mutex_enter(&bufcache_lock);
765 	for (i = 0; i < nrablks; i++) {
766 		/* If it's in the cache, just go on to next one. */
767 		if (incore(vp, rablks[i]))
768 			continue;
769 
770 		/* Get a buffer for the read-ahead block */
771 		mutex_exit(&bufcache_lock);
772 		(void) bio_doread(vp, rablks[i], rasizes[i], B_ASYNC);
773 		mutex_enter(&bufcache_lock);
774 	}
775 	mutex_exit(&bufcache_lock);
776 
777 	/* Otherwise, we had to start a read for it; wait until it's valid. */
778 	error = biowait(bp);
779 	if (error == 0 && (flags & B_MODIFY) != 0)
780 		error = fscow_run(bp, true);
781 	if (error) {
782 		brelse(bp, 0);
783 		*bpp = NULL;
784 	}
785 
786 	return error;
787 }
788 
789 /*
790  * Block write.  Described in Bach (p.56)
791  */
792 int
793 bwrite(buf_t *bp)
794 {
795 	int rv, sync, wasdelayed;
796 	struct vnode *vp;
797 	struct mount *mp;
798 
799 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
800 	KASSERT(!cv_has_waiters(&bp->b_done));
801 
802 	vp = bp->b_vp;
803 
804 	/*
805 	 * dholland 20160728 AFAICT vp==NULL must be impossible as it
806 	 * will crash upon reaching VOP_STRATEGY below... see further
807 	 * analysis on tech-kern.
808 	 */
809 	KASSERTMSG(vp != NULL, "bwrite given buffer with null vnode");
810 
811 	if (vp != NULL) {
812 		KASSERT(bp->b_objlock == vp->v_interlock);
813 		if (vp->v_type == VBLK)
814 			mp = spec_node_getmountedfs(vp);
815 		else
816 			mp = vp->v_mount;
817 	} else {
818 		mp = NULL;
819 	}
820 
821 	if (mp && mp->mnt_wapbl) {
822 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
823 			bdwrite(bp);
824 			return 0;
825 		}
826 	}
827 
828 	/*
829 	 * Remember buffer type, to switch on it later.  If the write was
830 	 * synchronous, but the file system was mounted with MNT_ASYNC,
831 	 * convert it to a delayed write.
832 	 * XXX note that this relies on delayed tape writes being converted
833 	 * to async, not sync writes (which is safe, but ugly).
834 	 */
835 	sync = !ISSET(bp->b_flags, B_ASYNC);
836 	if (sync && mp != NULL && ISSET(mp->mnt_flag, MNT_ASYNC)) {
837 		bdwrite(bp);
838 		return (0);
839 	}
840 
841 	/*
842 	 * Collect statistics on synchronous and asynchronous writes.
843 	 * Writes to block devices are charged to their associated
844 	 * filesystem (if any).
845 	 */
846 	if (mp != NULL) {
847 		if (sync)
848 			mp->mnt_stat.f_syncwrites++;
849 		else
850 			mp->mnt_stat.f_asyncwrites++;
851 	}
852 
853 	/*
854 	 * Pay for the I/O operation and make sure the buf is on the correct
855 	 * vnode queue.
856 	 */
857 	bp->b_error = 0;
858 	wasdelayed = ISSET(bp->b_oflags, BO_DELWRI);
859 	CLR(bp->b_flags, B_READ);
860 	if (wasdelayed) {
861 		mutex_enter(&bufcache_lock);
862 		mutex_enter(bp->b_objlock);
863 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
864 		reassignbuf(bp, bp->b_vp);
865 		mutex_exit(&bufcache_lock);
866 	} else {
867 		curlwp->l_ru.ru_oublock++;
868 		mutex_enter(bp->b_objlock);
869 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
870 	}
871 	if (vp != NULL)
872 		vp->v_numoutput++;
873 	mutex_exit(bp->b_objlock);
874 
875 	/* Initiate disk write. */
876 	if (sync)
877 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
878 	else
879 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
880 
881 	VOP_STRATEGY(vp, bp);
882 
883 	if (sync) {
884 		/* If I/O was synchronous, wait for it to complete. */
885 		rv = biowait(bp);
886 
887 		/* Release the buffer. */
888 		brelse(bp, 0);
889 
890 		return (rv);
891 	} else {
892 		return (0);
893 	}
894 }
895 
896 int
897 vn_bwrite(void *v)
898 {
899 	struct vop_bwrite_args *ap = v;
900 
901 	return (bwrite(ap->a_bp));
902 }
903 
904 /*
905  * Delayed write.
906  *
907  * The buffer is marked dirty, but is not queued for I/O.
908  * This routine should be used when the buffer is expected
909  * to be modified again soon, typically a small write that
910  * partially fills a buffer.
911  *
912  * NB: magnetic tapes cannot be delayed; they must be
913  * written in the order that the writes are requested.
914  *
915  * Described in Leffler, et al. (pp. 208-213).
916  */
917 void
918 bdwrite(buf_t *bp)
919 {
920 
921 	KASSERT(bp->b_vp == NULL || bp->b_vp->v_tag != VT_UFS ||
922 	    bp->b_vp->v_type == VBLK || ISSET(bp->b_flags, B_COWDONE));
923 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
924 	KASSERT(!cv_has_waiters(&bp->b_done));
925 
926 	/* If this is a tape block, write the block now. */
927 	if (bdev_type(bp->b_dev) == D_TAPE) {
928 		bawrite(bp);
929 		return;
930 	}
931 
932 	if (wapbl_vphaswapbl(bp->b_vp)) {
933 		struct mount *mp = wapbl_vptomp(bp->b_vp);
934 
935 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
936 			WAPBL_ADD_BUF(mp, bp);
937 		}
938 	}
939 
940 	/*
941 	 * If the block hasn't been seen before:
942 	 *	(1) Mark it as having been seen,
943 	 *	(2) Charge for the write,
944 	 *	(3) Make sure it's on its vnode's correct block list.
945 	 */
946 	KASSERT(bp->b_vp == NULL || bp->b_objlock == bp->b_vp->v_interlock);
947 
948 	if (!ISSET(bp->b_oflags, BO_DELWRI)) {
949 		mutex_enter(&bufcache_lock);
950 		mutex_enter(bp->b_objlock);
951 		SET(bp->b_oflags, BO_DELWRI);
952 		curlwp->l_ru.ru_oublock++;
953 		reassignbuf(bp, bp->b_vp);
954 		mutex_exit(&bufcache_lock);
955 	} else {
956 		mutex_enter(bp->b_objlock);
957 	}
958 	/* Otherwise, the "write" is done, so mark and release the buffer. */
959 	CLR(bp->b_oflags, BO_DONE);
960 	mutex_exit(bp->b_objlock);
961 
962 	brelse(bp, 0);
963 }
964 
965 /*
966  * Asynchronous block write; just an asynchronous bwrite().
967  */
968 void
969 bawrite(buf_t *bp)
970 {
971 
972 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
973 	KASSERT(bp->b_vp != NULL);
974 
975 	SET(bp->b_flags, B_ASYNC);
976 	VOP_BWRITE(bp->b_vp, bp);
977 }
978 
979 /*
980  * Release a buffer on to the free lists.
981  * Described in Bach (p. 46).
982  */
983 void
984 brelsel(buf_t *bp, int set)
985 {
986 	struct bqueue *bufq;
987 	struct vnode *vp;
988 
989 	KASSERT(bp != NULL);
990 	KASSERT(mutex_owned(&bufcache_lock));
991 	KASSERT(!cv_has_waiters(&bp->b_done));
992 	KASSERT(bp->b_refcnt > 0);
993 
994 	SET(bp->b_cflags, set);
995 
996 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
997 	KASSERT(bp->b_iodone == NULL);
998 
999 	/* Wake up any processes waiting for any buffer to become free. */
1000 	cv_signal(&needbuffer_cv);
1001 
1002 	/* Wake up any proceeses waiting for _this_ buffer to become */
1003 	if (ISSET(bp->b_cflags, BC_WANTED))
1004 		CLR(bp->b_cflags, BC_WANTED|BC_AGE);
1005 
1006 	/* If it's clean clear the copy-on-write flag. */
1007 	if (ISSET(bp->b_flags, B_COWDONE)) {
1008 		mutex_enter(bp->b_objlock);
1009 		if (!ISSET(bp->b_oflags, BO_DELWRI))
1010 			CLR(bp->b_flags, B_COWDONE);
1011 		mutex_exit(bp->b_objlock);
1012 	}
1013 
1014 	/*
1015 	 * Determine which queue the buffer should be on, then put it there.
1016 	 */
1017 
1018 	/* If it's locked, don't report an error; try again later. */
1019 	if (ISSET(bp->b_flags, B_LOCKED))
1020 		bp->b_error = 0;
1021 
1022 	/* If it's not cacheable, or an error, mark it invalid. */
1023 	if (ISSET(bp->b_cflags, BC_NOCACHE) || bp->b_error != 0)
1024 		SET(bp->b_cflags, BC_INVAL);
1025 
1026 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1027 		/*
1028 		 * This is a delayed write buffer that was just flushed to
1029 		 * disk.  It is still on the LRU queue.  If it's become
1030 		 * invalid, then we need to move it to a different queue;
1031 		 * otherwise leave it in its current position.
1032 		 */
1033 		CLR(bp->b_cflags, BC_VFLUSH);
1034 		if (!ISSET(bp->b_cflags, BC_INVAL|BC_AGE) &&
1035 		    !ISSET(bp->b_flags, B_LOCKED) && bp->b_error == 0) {
1036 			KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 1));
1037 			goto already_queued;
1038 		} else {
1039 			bremfree(bp);
1040 		}
1041 	}
1042 
1043 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_AGE], 0));
1044 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 0));
1045 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LOCKED], 0));
1046 
1047 	if ((bp->b_bufsize <= 0) || ISSET(bp->b_cflags, BC_INVAL)) {
1048 		/*
1049 		 * If it's invalid or empty, dissociate it from its vnode
1050 		 * and put on the head of the appropriate queue.
1051 		 */
1052 		if (ISSET(bp->b_flags, B_LOCKED)) {
1053 			if (wapbl_vphaswapbl(vp = bp->b_vp)) {
1054 				struct mount *mp = wapbl_vptomp(vp);
1055 
1056 				KASSERT(bp->b_iodone
1057 				    != mp->mnt_wapbl_op->wo_wapbl_biodone);
1058 				WAPBL_REMOVE_BUF(mp, bp);
1059 			}
1060 		}
1061 
1062 		mutex_enter(bp->b_objlock);
1063 		CLR(bp->b_oflags, BO_DONE|BO_DELWRI);
1064 		if ((vp = bp->b_vp) != NULL) {
1065 			KASSERT(bp->b_objlock == vp->v_interlock);
1066 			reassignbuf(bp, bp->b_vp);
1067 			brelvp(bp);
1068 			mutex_exit(vp->v_interlock);
1069 		} else {
1070 			KASSERT(bp->b_objlock == &buffer_lock);
1071 			mutex_exit(bp->b_objlock);
1072 		}
1073 
1074 		if (bp->b_bufsize <= 0)
1075 			/* no data */
1076 			goto already_queued;
1077 		else
1078 			/* invalid data */
1079 			bufq = &bufqueues[BQ_AGE];
1080 		binsheadfree(bp, bufq);
1081 	} else  {
1082 		/*
1083 		 * It has valid data.  Put it on the end of the appropriate
1084 		 * queue, so that it'll stick around for as long as possible.
1085 		 * If buf is AGE, but has dependencies, must put it on last
1086 		 * bufqueue to be scanned, ie LRU. This protects against the
1087 		 * livelock where BQ_AGE only has buffers with dependencies,
1088 		 * and we thus never get to the dependent buffers in BQ_LRU.
1089 		 */
1090 		if (ISSET(bp->b_flags, B_LOCKED)) {
1091 			/* locked in core */
1092 			bufq = &bufqueues[BQ_LOCKED];
1093 		} else if (!ISSET(bp->b_cflags, BC_AGE)) {
1094 			/* valid data */
1095 			bufq = &bufqueues[BQ_LRU];
1096 		} else {
1097 			/* stale but valid data */
1098 			bufq = &bufqueues[BQ_AGE];
1099 		}
1100 		binstailfree(bp, bufq);
1101 	}
1102 already_queued:
1103 	/* Unlock the buffer. */
1104 	CLR(bp->b_cflags, BC_AGE|BC_BUSY|BC_NOCACHE);
1105 	CLR(bp->b_flags, B_ASYNC);
1106 	cv_broadcast(&bp->b_busy);
1107 
1108 	if (bp->b_bufsize <= 0)
1109 		brele(bp);
1110 }
1111 
1112 void
1113 brelse(buf_t *bp, int set)
1114 {
1115 
1116 	mutex_enter(&bufcache_lock);
1117 	brelsel(bp, set);
1118 	mutex_exit(&bufcache_lock);
1119 }
1120 
1121 /*
1122  * Determine if a block is in the cache.
1123  * Just look on what would be its hash chain.  If it's there, return
1124  * a pointer to it, unless it's marked invalid.  If it's marked invalid,
1125  * we normally don't return the buffer, unless the caller explicitly
1126  * wants us to.
1127  */
1128 buf_t *
1129 incore(struct vnode *vp, daddr_t blkno)
1130 {
1131 	buf_t *bp;
1132 
1133 	KASSERT(mutex_owned(&bufcache_lock));
1134 
1135 	/* Search hash chain */
1136 	LIST_FOREACH(bp, BUFHASH(vp, blkno), b_hash) {
1137 		if (bp->b_lblkno == blkno && bp->b_vp == vp &&
1138 		    !ISSET(bp->b_cflags, BC_INVAL)) {
1139 		    	KASSERT(bp->b_objlock == vp->v_interlock);
1140 		    	return (bp);
1141 		}
1142 	}
1143 
1144 	return (NULL);
1145 }
1146 
1147 /*
1148  * Get a block of requested size that is associated with
1149  * a given vnode and block offset. If it is found in the
1150  * block cache, mark it as having been found, make it busy
1151  * and return it. Otherwise, return an empty block of the
1152  * correct size. It is up to the caller to insure that the
1153  * cached blocks be of the correct size.
1154  */
1155 buf_t *
1156 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo)
1157 {
1158 	int err, preserve;
1159 	buf_t *bp;
1160 
1161 	mutex_enter(&bufcache_lock);
1162  loop:
1163 	bp = incore(vp, blkno);
1164 	if (bp != NULL) {
1165 		err = bbusy(bp, ((slpflag & PCATCH) != 0), slptimeo, NULL);
1166 		if (err != 0) {
1167 			if (err == EPASSTHROUGH)
1168 				goto loop;
1169 			mutex_exit(&bufcache_lock);
1170 			return (NULL);
1171 		}
1172 		KASSERT(!cv_has_waiters(&bp->b_done));
1173 #ifdef DIAGNOSTIC
1174 		if (ISSET(bp->b_oflags, BO_DONE|BO_DELWRI) &&
1175 		    bp->b_bcount < size && vp->v_type != VBLK)
1176 			panic("getblk: block size invariant failed");
1177 #endif
1178 		bremfree(bp);
1179 		preserve = 1;
1180 	} else {
1181 		if ((bp = getnewbuf(slpflag, slptimeo, 0)) == NULL)
1182 			goto loop;
1183 
1184 		if (incore(vp, blkno) != NULL) {
1185 			/* The block has come into memory in the meantime. */
1186 			brelsel(bp, 0);
1187 			goto loop;
1188 		}
1189 
1190 		LIST_INSERT_HEAD(BUFHASH(vp, blkno), bp, b_hash);
1191 		bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno;
1192 		mutex_enter(vp->v_interlock);
1193 		bgetvp(vp, bp);
1194 		mutex_exit(vp->v_interlock);
1195 		preserve = 0;
1196 	}
1197 	mutex_exit(&bufcache_lock);
1198 
1199 	/*
1200 	 * LFS can't track total size of B_LOCKED buffer (locked_queue_bytes)
1201 	 * if we re-size buffers here.
1202 	 */
1203 	if (ISSET(bp->b_flags, B_LOCKED)) {
1204 		KASSERT(bp->b_bufsize >= size);
1205 	} else {
1206 		if (allocbuf(bp, size, preserve)) {
1207 			mutex_enter(&bufcache_lock);
1208 			LIST_REMOVE(bp, b_hash);
1209 			mutex_exit(&bufcache_lock);
1210 			brelse(bp, BC_INVAL);
1211 			return NULL;
1212 		}
1213 	}
1214 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1215 	return (bp);
1216 }
1217 
1218 /*
1219  * Get an empty, disassociated buffer of given size.
1220  */
1221 buf_t *
1222 geteblk(int size)
1223 {
1224 	buf_t *bp;
1225 	int error __diagused;
1226 
1227 	mutex_enter(&bufcache_lock);
1228 	while ((bp = getnewbuf(0, 0, 0)) == NULL)
1229 		;
1230 
1231 	SET(bp->b_cflags, BC_INVAL);
1232 	LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1233 	mutex_exit(&bufcache_lock);
1234 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1235 	error = allocbuf(bp, size, 0);
1236 	KASSERT(error == 0);
1237 	return (bp);
1238 }
1239 
1240 /*
1241  * Expand or contract the actual memory allocated to a buffer.
1242  *
1243  * If the buffer shrinks, data is lost, so it's up to the
1244  * caller to have written it out *first*; this routine will not
1245  * start a write.  If the buffer grows, it's the callers
1246  * responsibility to fill out the buffer's additional contents.
1247  */
1248 int
1249 allocbuf(buf_t *bp, int size, int preserve)
1250 {
1251 	void *addr;
1252 	vsize_t oldsize, desired_size;
1253 	int oldcount;
1254 	int delta;
1255 
1256 	desired_size = buf_roundsize(size);
1257 	if (desired_size > MAXBSIZE)
1258 		printf("allocbuf: buffer larger than MAXBSIZE requested");
1259 
1260 	oldcount = bp->b_bcount;
1261 
1262 	bp->b_bcount = size;
1263 
1264 	oldsize = bp->b_bufsize;
1265 	if (oldsize == desired_size) {
1266 		/*
1267 		 * Do not short cut the WAPBL resize, as the buffer length
1268 		 * could still have changed and this would corrupt the
1269 		 * tracking of the transaction length.
1270 		 */
1271 		goto out;
1272 	}
1273 
1274 	/*
1275 	 * If we want a buffer of a different size, re-allocate the
1276 	 * buffer's memory; copy old content only if needed.
1277 	 */
1278 	addr = buf_alloc(desired_size);
1279 	if (addr == NULL)
1280 		return ENOMEM;
1281 	if (preserve)
1282 		memcpy(addr, bp->b_data, MIN(oldsize,desired_size));
1283 	if (bp->b_data != NULL)
1284 		buf_mrelease(bp->b_data, oldsize);
1285 	bp->b_data = addr;
1286 	bp->b_bufsize = desired_size;
1287 
1288 	/*
1289 	 * Update overall buffer memory counter (protected by bufcache_lock)
1290 	 */
1291 	delta = (long)desired_size - (long)oldsize;
1292 
1293 	mutex_enter(&bufcache_lock);
1294 	if ((bufmem += delta) > bufmem_hiwater) {
1295 		/*
1296 		 * Need to trim overall memory usage.
1297 		 */
1298 		while (buf_canrelease()) {
1299 			if (curcpu()->ci_schedstate.spc_flags &
1300 			    SPCF_SHOULDYIELD) {
1301 				mutex_exit(&bufcache_lock);
1302 				preempt();
1303 				mutex_enter(&bufcache_lock);
1304 			}
1305 			if (buf_trim() == 0)
1306 				break;
1307 		}
1308 	}
1309 	mutex_exit(&bufcache_lock);
1310 
1311  out:
1312 	if (wapbl_vphaswapbl(bp->b_vp))
1313 		WAPBL_RESIZE_BUF(wapbl_vptomp(bp->b_vp), bp, oldsize, oldcount);
1314 
1315 	return 0;
1316 }
1317 
1318 /*
1319  * Find a buffer which is available for use.
1320  * Select something from a free list.
1321  * Preference is to AGE list, then LRU list.
1322  *
1323  * Called with the buffer queues locked.
1324  * Return buffer locked.
1325  */
1326 buf_t *
1327 getnewbuf(int slpflag, int slptimeo, int from_bufq)
1328 {
1329 	buf_t *bp;
1330 	struct vnode *vp;
1331 
1332  start:
1333 	KASSERT(mutex_owned(&bufcache_lock));
1334 
1335 	/*
1336 	 * Get a new buffer from the pool.
1337 	 */
1338 	if (!from_bufq && buf_lotsfree()) {
1339 		mutex_exit(&bufcache_lock);
1340 		bp = pool_cache_get(buf_cache, PR_NOWAIT);
1341 		if (bp != NULL) {
1342 			memset((char *)bp, 0, sizeof(*bp));
1343 			buf_init(bp);
1344 			SET(bp->b_cflags, BC_BUSY);	/* mark buffer busy */
1345 			mutex_enter(&bufcache_lock);
1346 #if defined(DIAGNOSTIC)
1347 			bp->b_freelistindex = -1;
1348 #endif /* defined(DIAGNOSTIC) */
1349 			return (bp);
1350 		}
1351 		mutex_enter(&bufcache_lock);
1352 	}
1353 
1354 	KASSERT(mutex_owned(&bufcache_lock));
1355 	if ((bp = TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue)) != NULL ||
1356 	    (bp = TAILQ_FIRST(&bufqueues[BQ_LRU].bq_queue)) != NULL) {
1357 	    	KASSERT(!ISSET(bp->b_cflags, BC_BUSY) || ISSET(bp->b_cflags, BC_VFLUSH));
1358 		bremfree(bp);
1359 
1360 		/* Buffer is no longer on free lists. */
1361 		SET(bp->b_cflags, BC_BUSY);
1362 	} else {
1363 		/*
1364 		 * XXX: !from_bufq should be removed.
1365 		 */
1366 		if (!from_bufq || curlwp != uvm.pagedaemon_lwp) {
1367 			/* wait for a free buffer of any kind */
1368 			if ((slpflag & PCATCH) != 0)
1369 				(void)cv_timedwait_sig(&needbuffer_cv,
1370 				    &bufcache_lock, slptimeo);
1371 			else
1372 				(void)cv_timedwait(&needbuffer_cv,
1373 				    &bufcache_lock, slptimeo);
1374 		}
1375 		return (NULL);
1376 	}
1377 
1378 #ifdef DIAGNOSTIC
1379 	if (bp->b_bufsize <= 0)
1380 		panic("buffer %p: on queue but empty", bp);
1381 #endif
1382 
1383 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1384 		/*
1385 		 * This is a delayed write buffer being flushed to disk.  Make
1386 		 * sure it gets aged out of the queue when it's finished, and
1387 		 * leave it off the LRU queue.
1388 		 */
1389 		CLR(bp->b_cflags, BC_VFLUSH);
1390 		SET(bp->b_cflags, BC_AGE);
1391 		goto start;
1392 	}
1393 
1394 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1395 	KASSERT(bp->b_refcnt > 0);
1396     	KASSERT(!cv_has_waiters(&bp->b_done));
1397 
1398 	/*
1399 	 * If buffer was a delayed write, start it and return NULL
1400 	 * (since we might sleep while starting the write).
1401 	 */
1402 	if (ISSET(bp->b_oflags, BO_DELWRI)) {
1403 		/*
1404 		 * This buffer has gone through the LRU, so make sure it gets
1405 		 * reused ASAP.
1406 		 */
1407 		SET(bp->b_cflags, BC_AGE);
1408 		mutex_exit(&bufcache_lock);
1409 		bawrite(bp);
1410 		mutex_enter(&bufcache_lock);
1411 		return (NULL);
1412 	}
1413 
1414 	vp = bp->b_vp;
1415 
1416 	/* clear out various other fields */
1417 	bp->b_cflags = BC_BUSY;
1418 	bp->b_oflags = 0;
1419 	bp->b_flags = 0;
1420 	bp->b_dev = NODEV;
1421 	bp->b_blkno = 0;
1422 	bp->b_lblkno = 0;
1423 	bp->b_rawblkno = 0;
1424 	bp->b_iodone = 0;
1425 	bp->b_error = 0;
1426 	bp->b_resid = 0;
1427 	bp->b_bcount = 0;
1428 
1429 	LIST_REMOVE(bp, b_hash);
1430 
1431 	/* Disassociate us from our vnode, if we had one... */
1432 	if (vp != NULL) {
1433 		mutex_enter(vp->v_interlock);
1434 		brelvp(bp);
1435 		mutex_exit(vp->v_interlock);
1436 	}
1437 
1438 	return (bp);
1439 }
1440 
1441 /*
1442  * Attempt to free an aged buffer off the queues.
1443  * Called with queue lock held.
1444  * Returns the amount of buffer memory freed.
1445  */
1446 static int
1447 buf_trim(void)
1448 {
1449 	buf_t *bp;
1450 	long size;
1451 
1452 	KASSERT(mutex_owned(&bufcache_lock));
1453 
1454 	/* Instruct getnewbuf() to get buffers off the queues */
1455 	if ((bp = getnewbuf(PCATCH, 1, 1)) == NULL)
1456 		return 0;
1457 
1458 	KASSERT((bp->b_cflags & BC_WANTED) == 0);
1459 	size = bp->b_bufsize;
1460 	bufmem -= size;
1461 	if (size > 0) {
1462 		buf_mrelease(bp->b_data, size);
1463 		bp->b_bcount = bp->b_bufsize = 0;
1464 	}
1465 	/* brelse() will return the buffer to the global buffer pool */
1466 	brelsel(bp, 0);
1467 	return size;
1468 }
1469 
1470 int
1471 buf_drain(int n)
1472 {
1473 	int size = 0, sz;
1474 
1475 	KASSERT(mutex_owned(&bufcache_lock));
1476 
1477 	while (size < n && bufmem > bufmem_lowater) {
1478 		sz = buf_trim();
1479 		if (sz <= 0)
1480 			break;
1481 		size += sz;
1482 	}
1483 
1484 	return size;
1485 }
1486 
1487 SDT_PROVIDER_DEFINE(io);
1488 
1489 SDT_PROBE_DEFINE1(io, kernel, , wait__start, "struct buf *"/*bp*/);
1490 SDT_PROBE_DEFINE1(io, kernel, , wait__done, "struct buf *"/*bp*/);
1491 
1492 /*
1493  * Wait for operations on the buffer to complete.
1494  * When they do, extract and return the I/O's error value.
1495  */
1496 int
1497 biowait(buf_t *bp)
1498 {
1499 
1500 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1501 	KASSERT(bp->b_refcnt > 0);
1502 
1503 	SDT_PROBE1(io, kernel, , wait__start, bp);
1504 
1505 	mutex_enter(bp->b_objlock);
1506 	while (!ISSET(bp->b_oflags, BO_DONE | BO_DELWRI))
1507 		cv_wait(&bp->b_done, bp->b_objlock);
1508 	mutex_exit(bp->b_objlock);
1509 
1510 	SDT_PROBE1(io, kernel, , wait__done, bp);
1511 
1512 	return bp->b_error;
1513 }
1514 
1515 /*
1516  * Mark I/O complete on a buffer.
1517  *
1518  * If a callback has been requested, e.g. the pageout
1519  * daemon, do so. Otherwise, awaken waiting processes.
1520  *
1521  * [ Leffler, et al., says on p.247:
1522  *	"This routine wakes up the blocked process, frees the buffer
1523  *	for an asynchronous write, or, for a request by the pagedaemon
1524  *	process, invokes a procedure specified in the buffer structure" ]
1525  *
1526  * In real life, the pagedaemon (or other system processes) wants
1527  * to do async stuff to, and doesn't want the buffer brelse()'d.
1528  * (for swap pager, that puts swap buffers on the free lists (!!!),
1529  * for the vn device, that puts allocated buffers on the free lists!)
1530  */
1531 void
1532 biodone(buf_t *bp)
1533 {
1534 	int s;
1535 
1536 	KASSERT(!ISSET(bp->b_oflags, BO_DONE));
1537 
1538 	if (cpu_intr_p()) {
1539 		/* From interrupt mode: defer to a soft interrupt. */
1540 		s = splvm();
1541 		TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_biodone, bp, b_actq);
1542 		softint_schedule(biodone_sih);
1543 		splx(s);
1544 	} else {
1545 		/* Process now - the buffer may be freed soon. */
1546 		biodone2(bp);
1547 	}
1548 }
1549 
1550 SDT_PROBE_DEFINE1(io, kernel, , done, "struct buf *"/*bp*/);
1551 
1552 static void
1553 biodone2(buf_t *bp)
1554 {
1555 	void (*callout)(buf_t *);
1556 
1557 	SDT_PROBE1(io, kernel, ,done, bp);
1558 
1559 	mutex_enter(bp->b_objlock);
1560 	/* Note that the transfer is done. */
1561 	if (ISSET(bp->b_oflags, BO_DONE))
1562 		panic("biodone2 already");
1563 	CLR(bp->b_flags, B_COWDONE);
1564 	SET(bp->b_oflags, BO_DONE);
1565 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1566 
1567 	/* Wake up waiting writers. */
1568 	if (!ISSET(bp->b_flags, B_READ))
1569 		vwakeup(bp);
1570 
1571 	if ((callout = bp->b_iodone) != NULL) {
1572 		/* Note callout done, then call out. */
1573 		KASSERT(!cv_has_waiters(&bp->b_done));
1574 		KERNEL_LOCK(1, NULL);		/* XXXSMP */
1575 		bp->b_iodone = NULL;
1576 		mutex_exit(bp->b_objlock);
1577 		(*callout)(bp);
1578 		KERNEL_UNLOCK_ONE(NULL);	/* XXXSMP */
1579 	} else if (ISSET(bp->b_flags, B_ASYNC)) {
1580 		/* If async, release. */
1581 		KASSERT(!cv_has_waiters(&bp->b_done));
1582 		mutex_exit(bp->b_objlock);
1583 		brelse(bp, 0);
1584 	} else {
1585 		/* Otherwise just wake up waiters in biowait(). */
1586 		cv_broadcast(&bp->b_done);
1587 		mutex_exit(bp->b_objlock);
1588 	}
1589 }
1590 
1591 static void
1592 biointr(void *cookie)
1593 {
1594 	struct cpu_info *ci;
1595 	buf_t *bp;
1596 	int s;
1597 
1598 	ci = curcpu();
1599 
1600 	while (!TAILQ_EMPTY(&ci->ci_data.cpu_biodone)) {
1601 		KASSERT(curcpu() == ci);
1602 
1603 		s = splvm();
1604 		bp = TAILQ_FIRST(&ci->ci_data.cpu_biodone);
1605 		TAILQ_REMOVE(&ci->ci_data.cpu_biodone, bp, b_actq);
1606 		splx(s);
1607 
1608 		biodone2(bp);
1609 	}
1610 }
1611 
1612 /*
1613  * Wait for all buffers to complete I/O
1614  * Return the number of "stuck" buffers.
1615  */
1616 int
1617 buf_syncwait(void)
1618 {
1619 	buf_t *bp;
1620 	int iter, nbusy, nbusy_prev = 0, ihash;
1621 
1622 	for (iter = 0; iter < 20;) {
1623 		mutex_enter(&bufcache_lock);
1624 		nbusy = 0;
1625 		for (ihash = 0; ihash < bufhash+1; ihash++) {
1626 		    LIST_FOREACH(bp, &bufhashtbl[ihash], b_hash) {
1627 			if ((bp->b_cflags & (BC_BUSY|BC_INVAL)) == BC_BUSY)
1628 				nbusy += ((bp->b_flags & B_READ) == 0);
1629 		    }
1630 		}
1631 		mutex_exit(&bufcache_lock);
1632 
1633 		if (nbusy == 0)
1634 			break;
1635 		if (nbusy_prev == 0)
1636 			nbusy_prev = nbusy;
1637 		printf("%d ", nbusy);
1638 		kpause("bflush", false, MAX(1, hz / 25 * iter), NULL);
1639 		if (nbusy >= nbusy_prev) /* we didn't flush anything */
1640 			iter++;
1641 		else
1642 			nbusy_prev = nbusy;
1643 	}
1644 
1645 	if (nbusy) {
1646 #if defined(DEBUG) || defined(DEBUG_HALT_BUSY)
1647 		printf("giving up\nPrinting vnodes for busy buffers\n");
1648 		for (ihash = 0; ihash < bufhash+1; ihash++) {
1649 		    LIST_FOREACH(bp, &bufhashtbl[ihash], b_hash) {
1650 			if ((bp->b_cflags & (BC_BUSY|BC_INVAL)) == BC_BUSY &&
1651 			    (bp->b_flags & B_READ) == 0)
1652 				vprint(NULL, bp->b_vp);
1653 		    }
1654 		}
1655 #endif
1656 	}
1657 
1658 	return nbusy;
1659 }
1660 
1661 static void
1662 sysctl_fillbuf(buf_t *i, struct buf_sysctl *o)
1663 {
1664 
1665 	o->b_flags = i->b_flags | i->b_cflags | i->b_oflags;
1666 	o->b_error = i->b_error;
1667 	o->b_prio = i->b_prio;
1668 	o->b_dev = i->b_dev;
1669 	o->b_bufsize = i->b_bufsize;
1670 	o->b_bcount = i->b_bcount;
1671 	o->b_resid = i->b_resid;
1672 	o->b_addr = PTRTOUINT64(i->b_data);
1673 	o->b_blkno = i->b_blkno;
1674 	o->b_rawblkno = i->b_rawblkno;
1675 	o->b_iodone = PTRTOUINT64(i->b_iodone);
1676 	o->b_proc = PTRTOUINT64(i->b_proc);
1677 	o->b_vp = PTRTOUINT64(i->b_vp);
1678 	o->b_saveaddr = PTRTOUINT64(i->b_saveaddr);
1679 	o->b_lblkno = i->b_lblkno;
1680 }
1681 
1682 #define KERN_BUFSLOP 20
1683 static int
1684 sysctl_dobuf(SYSCTLFN_ARGS)
1685 {
1686 	buf_t *bp;
1687 	struct buf_sysctl bs;
1688 	struct bqueue *bq;
1689 	char *dp;
1690 	u_int i, op, arg;
1691 	size_t len, needed, elem_size, out_size;
1692 	int error, elem_count, retries;
1693 
1694 	if (namelen == 1 && name[0] == CTL_QUERY)
1695 		return (sysctl_query(SYSCTLFN_CALL(rnode)));
1696 
1697 	if (namelen != 4)
1698 		return (EINVAL);
1699 
1700 	retries = 100;
1701  retry:
1702 	dp = oldp;
1703 	len = (oldp != NULL) ? *oldlenp : 0;
1704 	op = name[0];
1705 	arg = name[1];
1706 	elem_size = name[2];
1707 	elem_count = name[3];
1708 	out_size = MIN(sizeof(bs), elem_size);
1709 
1710 	/*
1711 	 * at the moment, these are just "placeholders" to make the
1712 	 * API for retrieving kern.buf data more extensible in the
1713 	 * future.
1714 	 *
1715 	 * XXX kern.buf currently has "netbsd32" issues.  hopefully
1716 	 * these will be resolved at a later point.
1717 	 */
1718 	if (op != KERN_BUF_ALL || arg != KERN_BUF_ALL ||
1719 	    elem_size < 1 || elem_count < 0)
1720 		return (EINVAL);
1721 
1722 	error = 0;
1723 	needed = 0;
1724 	sysctl_unlock();
1725 	mutex_enter(&bufcache_lock);
1726 	for (i = 0; i < BQUEUES; i++) {
1727 		bq = &bufqueues[i];
1728 		TAILQ_FOREACH(bp, &bq->bq_queue, b_freelist) {
1729 			bq->bq_marker = bp;
1730 			if (len >= elem_size && elem_count > 0) {
1731 				sysctl_fillbuf(bp, &bs);
1732 				mutex_exit(&bufcache_lock);
1733 				error = copyout(&bs, dp, out_size);
1734 				mutex_enter(&bufcache_lock);
1735 				if (error)
1736 					break;
1737 				if (bq->bq_marker != bp) {
1738 					/*
1739 					 * This sysctl node is only for
1740 					 * statistics.  Retry; if the
1741 					 * queue keeps changing, then
1742 					 * bail out.
1743 					 */
1744 					if (retries-- == 0) {
1745 						error = EAGAIN;
1746 						break;
1747 					}
1748 					mutex_exit(&bufcache_lock);
1749 					sysctl_relock();
1750 					goto retry;
1751 				}
1752 				dp += elem_size;
1753 				len -= elem_size;
1754 			}
1755 			needed += elem_size;
1756 			if (elem_count > 0 && elem_count != INT_MAX)
1757 				elem_count--;
1758 		}
1759 		if (error != 0)
1760 			break;
1761 	}
1762 	mutex_exit(&bufcache_lock);
1763 	sysctl_relock();
1764 
1765 	*oldlenp = needed;
1766 	if (oldp == NULL)
1767 		*oldlenp += KERN_BUFSLOP * sizeof(buf_t);
1768 
1769 	return (error);
1770 }
1771 
1772 static int
1773 sysctl_bufvm_update(SYSCTLFN_ARGS)
1774 {
1775 	int error, rv;
1776 	struct sysctlnode node;
1777 	unsigned int temp_bufcache;
1778 	unsigned long temp_water;
1779 
1780 	/* Take a copy of the supplied node and its data */
1781 	node = *rnode;
1782 	if (node.sysctl_data == &bufcache) {
1783 	    node.sysctl_data = &temp_bufcache;
1784 	    temp_bufcache = *(unsigned int *)rnode->sysctl_data;
1785 	} else {
1786 	    node.sysctl_data = &temp_water;
1787 	    temp_water = *(unsigned long *)rnode->sysctl_data;
1788 	}
1789 
1790 	/* Update the copy */
1791 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
1792 	if (error || newp == NULL)
1793 		return (error);
1794 
1795 	if (rnode->sysctl_data == &bufcache) {
1796 		if (temp_bufcache > 100)
1797 			return (EINVAL);
1798 		bufcache = temp_bufcache;
1799 		buf_setwm();
1800 	} else if (rnode->sysctl_data == &bufmem_lowater) {
1801 		if (bufmem_hiwater - temp_water < 16)
1802 			return (EINVAL);
1803 		bufmem_lowater = temp_water;
1804 	} else if (rnode->sysctl_data == &bufmem_hiwater) {
1805 		if (temp_water - bufmem_lowater < 16)
1806 			return (EINVAL);
1807 		bufmem_hiwater = temp_water;
1808 	} else
1809 		return (EINVAL);
1810 
1811 	/* Drain until below new high water mark */
1812 	sysctl_unlock();
1813 	mutex_enter(&bufcache_lock);
1814 	while (bufmem > bufmem_hiwater) {
1815 		rv = buf_drain((bufmem - bufmem_hiwater) / (2 * 1024));
1816 		if (rv <= 0)
1817 			break;
1818 	}
1819 	mutex_exit(&bufcache_lock);
1820 	sysctl_relock();
1821 
1822 	return 0;
1823 }
1824 
1825 static struct sysctllog *vfsbio_sysctllog;
1826 
1827 static void
1828 sysctl_kern_buf_setup(void)
1829 {
1830 
1831 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1832 		       CTLFLAG_PERMANENT,
1833 		       CTLTYPE_NODE, "buf",
1834 		       SYSCTL_DESCR("Kernel buffer cache information"),
1835 		       sysctl_dobuf, 0, NULL, 0,
1836 		       CTL_KERN, KERN_BUF, CTL_EOL);
1837 }
1838 
1839 static void
1840 sysctl_vm_buf_setup(void)
1841 {
1842 
1843 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1844 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1845 		       CTLTYPE_INT, "bufcache",
1846 		       SYSCTL_DESCR("Percentage of physical memory to use for "
1847 				    "buffer cache"),
1848 		       sysctl_bufvm_update, 0, &bufcache, 0,
1849 		       CTL_VM, CTL_CREATE, CTL_EOL);
1850 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1851 		       CTLFLAG_PERMANENT|CTLFLAG_READONLY,
1852 		       CTLTYPE_LONG, "bufmem",
1853 		       SYSCTL_DESCR("Amount of kernel memory used by buffer "
1854 				    "cache"),
1855 		       NULL, 0, &bufmem, 0,
1856 		       CTL_VM, CTL_CREATE, CTL_EOL);
1857 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1858 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1859 		       CTLTYPE_LONG, "bufmem_lowater",
1860 		       SYSCTL_DESCR("Minimum amount of kernel memory to "
1861 				    "reserve for buffer cache"),
1862 		       sysctl_bufvm_update, 0, &bufmem_lowater, 0,
1863 		       CTL_VM, CTL_CREATE, CTL_EOL);
1864 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1865 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1866 		       CTLTYPE_LONG, "bufmem_hiwater",
1867 		       SYSCTL_DESCR("Maximum amount of kernel memory to use "
1868 				    "for buffer cache"),
1869 		       sysctl_bufvm_update, 0, &bufmem_hiwater, 0,
1870 		       CTL_VM, CTL_CREATE, CTL_EOL);
1871 }
1872 
1873 #ifdef DEBUG
1874 /*
1875  * Print out statistics on the current allocation of the buffer pool.
1876  * Can be enabled to print out on every ``sync'' by setting "syncprt"
1877  * in vfs_syscalls.c using sysctl.
1878  */
1879 void
1880 vfs_bufstats(void)
1881 {
1882 	int i, j, count;
1883 	buf_t *bp;
1884 	struct bqueue *dp;
1885 	int counts[(MAXBSIZE / PAGE_SIZE) + 1];
1886 	static const char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE" };
1887 
1888 	for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) {
1889 		count = 0;
1890 		for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
1891 			counts[j] = 0;
1892 		TAILQ_FOREACH(bp, &dp->bq_queue, b_freelist) {
1893 			counts[bp->b_bufsize/PAGE_SIZE]++;
1894 			count++;
1895 		}
1896 		printf("%s: total-%d", bname[i], count);
1897 		for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
1898 			if (counts[j] != 0)
1899 				printf(", %d-%d", j * PAGE_SIZE, counts[j]);
1900 		printf("\n");
1901 	}
1902 }
1903 #endif /* DEBUG */
1904 
1905 /* ------------------------------ */
1906 
1907 buf_t *
1908 getiobuf(struct vnode *vp, bool waitok)
1909 {
1910 	buf_t *bp;
1911 
1912 	bp = pool_cache_get(bufio_cache, (waitok ? PR_WAITOK : PR_NOWAIT));
1913 	if (bp == NULL)
1914 		return bp;
1915 
1916 	buf_init(bp);
1917 
1918 	if ((bp->b_vp = vp) == NULL)
1919 		bp->b_objlock = &buffer_lock;
1920 	else
1921 		bp->b_objlock = vp->v_interlock;
1922 
1923 	return bp;
1924 }
1925 
1926 void
1927 putiobuf(buf_t *bp)
1928 {
1929 
1930 	buf_destroy(bp);
1931 	pool_cache_put(bufio_cache, bp);
1932 }
1933 
1934 /*
1935  * nestiobuf_iodone: b_iodone callback for nested buffers.
1936  */
1937 
1938 void
1939 nestiobuf_iodone(buf_t *bp)
1940 {
1941 	buf_t *mbp = bp->b_private;
1942 	int error;
1943 	int donebytes;
1944 
1945 	KASSERT(bp->b_bcount <= bp->b_bufsize);
1946 	KASSERT(mbp != bp);
1947 
1948 	error = bp->b_error;
1949 	if (bp->b_error == 0 &&
1950 	    (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
1951 		/*
1952 		 * Not all got transfered, raise an error. We have no way to
1953 		 * propagate these conditions to mbp.
1954 		 */
1955 		error = EIO;
1956 	}
1957 
1958 	donebytes = bp->b_bufsize;
1959 
1960 	putiobuf(bp);
1961 	nestiobuf_done(mbp, donebytes, error);
1962 }
1963 
1964 /*
1965  * nestiobuf_setup: setup a "nested" buffer.
1966  *
1967  * => 'mbp' is a "master" buffer which is being divided into sub pieces.
1968  * => 'bp' should be a buffer allocated by getiobuf.
1969  * => 'offset' is a byte offset in the master buffer.
1970  * => 'size' is a size in bytes of this nested buffer.
1971  */
1972 
1973 void
1974 nestiobuf_setup(buf_t *mbp, buf_t *bp, int offset, size_t size)
1975 {
1976 	const int b_read = mbp->b_flags & B_READ;
1977 	struct vnode *vp = mbp->b_vp;
1978 
1979 	KASSERT(mbp->b_bcount >= offset + size);
1980 	bp->b_vp = vp;
1981 	bp->b_dev = mbp->b_dev;
1982 	bp->b_objlock = mbp->b_objlock;
1983 	bp->b_cflags = BC_BUSY;
1984 	bp->b_flags = B_ASYNC | b_read;
1985 	bp->b_iodone = nestiobuf_iodone;
1986 	bp->b_data = (char *)mbp->b_data + offset;
1987 	bp->b_resid = bp->b_bcount = size;
1988 	bp->b_bufsize = bp->b_bcount;
1989 	bp->b_private = mbp;
1990 	BIO_COPYPRIO(bp, mbp);
1991 	if (!b_read && vp != NULL) {
1992 		mutex_enter(vp->v_interlock);
1993 		vp->v_numoutput++;
1994 		mutex_exit(vp->v_interlock);
1995 	}
1996 }
1997 
1998 /*
1999  * nestiobuf_done: propagate completion to the master buffer.
2000  *
2001  * => 'donebytes' specifies how many bytes in the 'mbp' is completed.
2002  * => 'error' is an errno(2) that 'donebytes' has been completed with.
2003  */
2004 
2005 void
2006 nestiobuf_done(buf_t *mbp, int donebytes, int error)
2007 {
2008 
2009 	if (donebytes == 0) {
2010 		return;
2011 	}
2012 	mutex_enter(mbp->b_objlock);
2013 	KASSERT(mbp->b_resid >= donebytes);
2014 	mbp->b_resid -= donebytes;
2015 	if (error)
2016 		mbp->b_error = error;
2017 	if (mbp->b_resid == 0) {
2018 		if (mbp->b_error)
2019 			mbp->b_resid = mbp->b_bcount;
2020 		mutex_exit(mbp->b_objlock);
2021 		biodone(mbp);
2022 	} else
2023 		mutex_exit(mbp->b_objlock);
2024 }
2025 
2026 void
2027 buf_init(buf_t *bp)
2028 {
2029 
2030 	cv_init(&bp->b_busy, "biolock");
2031 	cv_init(&bp->b_done, "biowait");
2032 	bp->b_dev = NODEV;
2033 	bp->b_error = 0;
2034 	bp->b_flags = 0;
2035 	bp->b_cflags = 0;
2036 	bp->b_oflags = 0;
2037 	bp->b_objlock = &buffer_lock;
2038 	bp->b_iodone = NULL;
2039 	bp->b_refcnt = 1;
2040 	bp->b_dev = NODEV;
2041 	bp->b_vnbufs.le_next = NOLIST;
2042 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
2043 }
2044 
2045 void
2046 buf_destroy(buf_t *bp)
2047 {
2048 
2049 	cv_destroy(&bp->b_done);
2050 	cv_destroy(&bp->b_busy);
2051 }
2052 
2053 int
2054 bbusy(buf_t *bp, bool intr, int timo, kmutex_t *interlock)
2055 {
2056 	int error;
2057 
2058 	KASSERT(mutex_owned(&bufcache_lock));
2059 
2060 	if ((bp->b_cflags & BC_BUSY) != 0) {
2061 		if (curlwp == uvm.pagedaemon_lwp)
2062 			return EDEADLK;
2063 		bp->b_cflags |= BC_WANTED;
2064 		bref(bp);
2065 		if (interlock != NULL)
2066 			mutex_exit(interlock);
2067 		if (intr) {
2068 			error = cv_timedwait_sig(&bp->b_busy, &bufcache_lock,
2069 			    timo);
2070 		} else {
2071 			error = cv_timedwait(&bp->b_busy, &bufcache_lock,
2072 			    timo);
2073 		}
2074 		brele(bp);
2075 		if (interlock != NULL)
2076 			mutex_enter(interlock);
2077 		if (error != 0)
2078 			return error;
2079 		return EPASSTHROUGH;
2080 	}
2081 	bp->b_cflags |= BC_BUSY;
2082 
2083 	return 0;
2084 }
2085