xref: /netbsd-src/sys/kern/vfs_bio.c (revision e89934bbf778a6d6d6894877c4da59d0c7835b0f)
1 /*	$NetBSD: vfs_bio.c,v 1.269 2017/01/20 09:45:13 skrll 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 1.269 2017/01/20 09:45:13 skrll Exp $");
127 
128 #ifdef _KERNEL_OPT
129 #include "opt_bufcache.h"
130 #include "opt_dtrace.h"
131 #include "opt_biohist.h"
132 #endif
133 
134 #include <sys/param.h>
135 #include <sys/systm.h>
136 #include <sys/kernel.h>
137 #include <sys/proc.h>
138 #include <sys/buf.h>
139 #include <sys/vnode.h>
140 #include <sys/mount.h>
141 #include <sys/resourcevar.h>
142 #include <sys/sysctl.h>
143 #include <sys/conf.h>
144 #include <sys/kauth.h>
145 #include <sys/fstrans.h>
146 #include <sys/intr.h>
147 #include <sys/cpu.h>
148 #include <sys/wapbl.h>
149 #include <sys/bitops.h>
150 #include <sys/cprng.h>
151 #include <sys/sdt.h>
152 
153 #include <uvm/uvm.h>	/* extern struct uvm uvm */
154 
155 #include <miscfs/specfs/specdev.h>
156 
157 #ifndef	BUFPAGES
158 # define BUFPAGES 0
159 #endif
160 
161 #ifdef BUFCACHE
162 # if (BUFCACHE < 5) || (BUFCACHE > 95)
163 #  error BUFCACHE is not between 5 and 95
164 # endif
165 #else
166 # define BUFCACHE 15
167 #endif
168 
169 u_int	nbuf;			/* desired number of buffer headers */
170 u_int	bufpages = BUFPAGES;	/* optional hardwired count */
171 u_int	bufcache = BUFCACHE;	/* max % of RAM to use for buffer cache */
172 
173 /* Function prototypes */
174 struct bqueue;
175 
176 static void buf_setwm(void);
177 static int buf_trim(void);
178 static void *bufpool_page_alloc(struct pool *, int);
179 static void bufpool_page_free(struct pool *, void *);
180 static buf_t *bio_doread(struct vnode *, daddr_t, int, int);
181 static buf_t *getnewbuf(int, int, int);
182 static int buf_lotsfree(void);
183 static int buf_canrelease(void);
184 static u_long buf_mempoolidx(u_long);
185 static u_long buf_roundsize(u_long);
186 static void *buf_alloc(size_t);
187 static void buf_mrelease(void *, size_t);
188 static void binsheadfree(buf_t *, struct bqueue *);
189 static void binstailfree(buf_t *, struct bqueue *);
190 #ifdef DEBUG
191 static int checkfreelist(buf_t *, struct bqueue *, int);
192 #endif
193 static void biointr(void *);
194 static void biodone2(buf_t *);
195 static void bref(buf_t *);
196 static void brele(buf_t *);
197 static void sysctl_kern_buf_setup(void);
198 static void sysctl_vm_buf_setup(void);
199 
200 /* Initialization for biohist */
201 
202 #include <sys/biohist.h>
203 
204 BIOHIST_DEFINE(biohist);
205 
206 void
207 biohist_init(void)
208 {
209 
210 	BIOHIST_INIT(biohist, BIOHIST_SIZE);
211 }
212 
213 /*
214  * Definitions for the buffer hash lists.
215  */
216 #define	BUFHASH(dvp, lbn)	\
217 	(&bufhashtbl[(((long)(dvp) >> 8) + (int)(lbn)) & bufhash])
218 LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
219 u_long	bufhash;
220 struct bqueue bufqueues[BQUEUES];
221 
222 static kcondvar_t needbuffer_cv;
223 
224 /*
225  * Buffer queue lock.
226  */
227 kmutex_t bufcache_lock;
228 kmutex_t buffer_lock;
229 
230 /* Software ISR for completed transfers. */
231 static void *biodone_sih;
232 
233 /* Buffer pool for I/O buffers. */
234 static pool_cache_t buf_cache;
235 static pool_cache_t bufio_cache;
236 
237 #define MEMPOOL_INDEX_OFFSET (ilog2(DEV_BSIZE))	/* smallest pool is 512 bytes */
238 #define NMEMPOOLS (ilog2(MAXBSIZE) - MEMPOOL_INDEX_OFFSET + 1)
239 __CTASSERT((1 << (NMEMPOOLS + MEMPOOL_INDEX_OFFSET - 1)) == MAXBSIZE);
240 
241 /* Buffer memory pools */
242 static struct pool bmempools[NMEMPOOLS];
243 
244 static struct vm_map *buf_map;
245 
246 /*
247  * Buffer memory pool allocator.
248  */
249 static void *
250 bufpool_page_alloc(struct pool *pp, int flags)
251 {
252 
253 	return (void *)uvm_km_alloc(buf_map,
254 	    MAXBSIZE, MAXBSIZE,
255 	    ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT|UVM_KMF_TRYLOCK)
256 	    | UVM_KMF_WIRED);
257 }
258 
259 static void
260 bufpool_page_free(struct pool *pp, void *v)
261 {
262 
263 	uvm_km_free(buf_map, (vaddr_t)v, MAXBSIZE, UVM_KMF_WIRED);
264 }
265 
266 static struct pool_allocator bufmempool_allocator = {
267 	.pa_alloc = bufpool_page_alloc,
268 	.pa_free = bufpool_page_free,
269 	.pa_pagesz = MAXBSIZE,
270 };
271 
272 /* Buffer memory management variables */
273 u_long bufmem_valimit;
274 u_long bufmem_hiwater;
275 u_long bufmem_lowater;
276 u_long bufmem;
277 
278 /*
279  * MD code can call this to set a hard limit on the amount
280  * of virtual memory used by the buffer cache.
281  */
282 int
283 buf_setvalimit(vsize_t sz)
284 {
285 
286 	/* We need to accommodate at least NMEMPOOLS of MAXBSIZE each */
287 	if (sz < NMEMPOOLS * MAXBSIZE)
288 		return EINVAL;
289 
290 	bufmem_valimit = sz;
291 	return 0;
292 }
293 
294 static void
295 buf_setwm(void)
296 {
297 
298 	bufmem_hiwater = buf_memcalc();
299 	/* lowater is approx. 2% of memory (with bufcache = 15) */
300 #define	BUFMEM_WMSHIFT	3
301 #define	BUFMEM_HIWMMIN	(64 * 1024 << BUFMEM_WMSHIFT)
302 	if (bufmem_hiwater < BUFMEM_HIWMMIN)
303 		/* Ensure a reasonable minimum value */
304 		bufmem_hiwater = BUFMEM_HIWMMIN;
305 	bufmem_lowater = bufmem_hiwater >> BUFMEM_WMSHIFT;
306 }
307 
308 #ifdef DEBUG
309 int debug_verify_freelist = 0;
310 static int
311 checkfreelist(buf_t *bp, struct bqueue *dp, int ison)
312 {
313 	buf_t *b;
314 
315 	if (!debug_verify_freelist)
316 		return 1;
317 
318 	TAILQ_FOREACH(b, &dp->bq_queue, b_freelist) {
319 		if (b == bp)
320 			return ison ? 1 : 0;
321 	}
322 
323 	return ison ? 0 : 1;
324 }
325 #endif
326 
327 /*
328  * Insq/Remq for the buffer hash lists.
329  * Call with buffer queue locked.
330  */
331 static void
332 binsheadfree(buf_t *bp, struct bqueue *dp)
333 {
334 
335 	KASSERT(mutex_owned(&bufcache_lock));
336 	KASSERT(bp->b_freelistindex == -1);
337 	TAILQ_INSERT_HEAD(&dp->bq_queue, bp, b_freelist);
338 	dp->bq_bytes += bp->b_bufsize;
339 	bp->b_freelistindex = dp - bufqueues;
340 }
341 
342 static void
343 binstailfree(buf_t *bp, struct bqueue *dp)
344 {
345 
346 	KASSERT(mutex_owned(&bufcache_lock));
347 	KASSERTMSG(bp->b_freelistindex == -1, "double free of buffer? "
348 	    "bp=%p, b_freelistindex=%d\n", bp, bp->b_freelistindex);
349 	TAILQ_INSERT_TAIL(&dp->bq_queue, bp, b_freelist);
350 	dp->bq_bytes += bp->b_bufsize;
351 	bp->b_freelistindex = dp - bufqueues;
352 }
353 
354 void
355 bremfree(buf_t *bp)
356 {
357 	struct bqueue *dp;
358 	int bqidx = bp->b_freelistindex;
359 
360 	KASSERT(mutex_owned(&bufcache_lock));
361 
362 	KASSERT(bqidx != -1);
363 	dp = &bufqueues[bqidx];
364 	KDASSERT(checkfreelist(bp, dp, 1));
365 	KASSERT(dp->bq_bytes >= bp->b_bufsize);
366 	TAILQ_REMOVE(&dp->bq_queue, bp, b_freelist);
367 	dp->bq_bytes -= bp->b_bufsize;
368 
369 	/* For the sysctl helper. */
370 	if (bp == dp->bq_marker)
371 		dp->bq_marker = NULL;
372 
373 #if defined(DIAGNOSTIC)
374 	bp->b_freelistindex = -1;
375 #endif /* defined(DIAGNOSTIC) */
376 }
377 
378 /*
379  * Add a reference to an buffer structure that came from buf_cache.
380  */
381 static inline void
382 bref(buf_t *bp)
383 {
384 
385 	KASSERT(mutex_owned(&bufcache_lock));
386 	KASSERT(bp->b_refcnt > 0);
387 
388 	bp->b_refcnt++;
389 }
390 
391 /*
392  * Free an unused buffer structure that came from buf_cache.
393  */
394 static inline void
395 brele(buf_t *bp)
396 {
397 
398 	KASSERT(mutex_owned(&bufcache_lock));
399 	KASSERT(bp->b_refcnt > 0);
400 
401 	if (bp->b_refcnt-- == 1) {
402 		buf_destroy(bp);
403 #ifdef DEBUG
404 		memset((char *)bp, 0, sizeof(*bp));
405 #endif
406 		pool_cache_put(buf_cache, bp);
407 	}
408 }
409 
410 /*
411  * note that for some ports this is used by pmap bootstrap code to
412  * determine kva size.
413  */
414 u_long
415 buf_memcalc(void)
416 {
417 	u_long n;
418 	vsize_t mapsz = 0;
419 
420 	/*
421 	 * Determine the upper bound of memory to use for buffers.
422 	 *
423 	 *	- If bufpages is specified, use that as the number
424 	 *	  pages.
425 	 *
426 	 *	- Otherwise, use bufcache as the percentage of
427 	 *	  physical memory.
428 	 */
429 	if (bufpages != 0) {
430 		n = bufpages;
431 	} else {
432 		if (bufcache < 5) {
433 			printf("forcing bufcache %d -> 5", bufcache);
434 			bufcache = 5;
435 		}
436 		if (bufcache > 95) {
437 			printf("forcing bufcache %d -> 95", bufcache);
438 			bufcache = 95;
439 		}
440 		if (buf_map != NULL)
441 			mapsz = vm_map_max(buf_map) - vm_map_min(buf_map);
442 		n = calc_cache_size(mapsz, bufcache,
443 		    (buf_map != kernel_map) ? 100 : BUFCACHE_VA_MAXPCT)
444 		    / PAGE_SIZE;
445 	}
446 
447 	n <<= PAGE_SHIFT;
448 	if (bufmem_valimit != 0 && n > bufmem_valimit)
449 		n = bufmem_valimit;
450 
451 	return (n);
452 }
453 
454 /*
455  * Initialize buffers and hash links for buffers.
456  */
457 void
458 bufinit(void)
459 {
460 	struct bqueue *dp;
461 	int use_std;
462 	u_int i;
463 
464 	biodone_vfs = biodone;
465 
466 	mutex_init(&bufcache_lock, MUTEX_DEFAULT, IPL_NONE);
467 	mutex_init(&buffer_lock, MUTEX_DEFAULT, IPL_NONE);
468 	cv_init(&needbuffer_cv, "needbuf");
469 
470 	if (bufmem_valimit != 0) {
471 		vaddr_t minaddr = 0, maxaddr;
472 		buf_map = uvm_km_suballoc(kernel_map, &minaddr, &maxaddr,
473 					  bufmem_valimit, 0, false, 0);
474 		if (buf_map == NULL)
475 			panic("bufinit: cannot allocate submap");
476 	} else
477 		buf_map = kernel_map;
478 
479 	/*
480 	 * Initialize buffer cache memory parameters.
481 	 */
482 	bufmem = 0;
483 	buf_setwm();
484 
485 	/* On "small" machines use small pool page sizes where possible */
486 	use_std = (physmem < atop(16*1024*1024));
487 
488 	/*
489 	 * Also use them on systems that can map the pool pages using
490 	 * a direct-mapped segment.
491 	 */
492 #ifdef PMAP_MAP_POOLPAGE
493 	use_std = 1;
494 #endif
495 
496 	buf_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
497 	    "bufpl", NULL, IPL_SOFTBIO, NULL, NULL, NULL);
498 	bufio_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
499 	    "biopl", NULL, IPL_BIO, NULL, NULL, NULL);
500 
501 	for (i = 0; i < NMEMPOOLS; i++) {
502 		struct pool_allocator *pa;
503 		struct pool *pp = &bmempools[i];
504 		u_int size = 1 << (i + MEMPOOL_INDEX_OFFSET);
505 		char *name = kmem_alloc(8, KM_SLEEP); /* XXX: never freed */
506 		if (__predict_false(size >= 1048576))
507 			(void)snprintf(name, 8, "buf%um", size / 1048576);
508 		else if (__predict_true(size >= 1024))
509 			(void)snprintf(name, 8, "buf%uk", size / 1024);
510 		else
511 			(void)snprintf(name, 8, "buf%ub", size);
512 		pa = (size <= PAGE_SIZE && use_std)
513 			? &pool_allocator_nointr
514 			: &bufmempool_allocator;
515 		pool_init(pp, size, 0, 0, 0, name, pa, IPL_NONE);
516 		pool_setlowat(pp, 1);
517 		pool_sethiwat(pp, 1);
518 	}
519 
520 	/* Initialize the buffer queues */
521 	for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) {
522 		TAILQ_INIT(&dp->bq_queue);
523 		dp->bq_bytes = 0;
524 	}
525 
526 	/*
527 	 * Estimate hash table size based on the amount of memory we
528 	 * intend to use for the buffer cache. The average buffer
529 	 * size is dependent on our clients (i.e. filesystems).
530 	 *
531 	 * For now, use an empirical 3K per buffer.
532 	 */
533 	nbuf = (bufmem_hiwater / 1024) / 3;
534 	bufhashtbl = hashinit(nbuf, HASH_LIST, true, &bufhash);
535 
536 	sysctl_kern_buf_setup();
537 	sysctl_vm_buf_setup();
538 }
539 
540 void
541 bufinit2(void)
542 {
543 
544 	biodone_sih = softint_establish(SOFTINT_BIO | SOFTINT_MPSAFE, biointr,
545 	    NULL);
546 	if (biodone_sih == NULL)
547 		panic("bufinit2: can't establish soft interrupt");
548 }
549 
550 static int
551 buf_lotsfree(void)
552 {
553 	u_long guess;
554 
555 	/* Always allocate if less than the low water mark. */
556 	if (bufmem < bufmem_lowater)
557 		return 1;
558 
559 	/* Never allocate if greater than the high water mark. */
560 	if (bufmem > bufmem_hiwater)
561 		return 0;
562 
563 	/* If there's anything on the AGE list, it should be eaten. */
564 	if (TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue) != NULL)
565 		return 0;
566 
567 	/*
568 	 * The probabily of getting a new allocation is inversely
569 	 * proportional  to the current size of the cache above
570 	 * the low water mark.  Divide the total first to avoid overflows
571 	 * in the product.
572 	 */
573 	guess = cprng_fast32() % 16;
574 
575 	if ((bufmem_hiwater - bufmem_lowater) / 16 * guess >=
576 	    (bufmem - bufmem_lowater))
577 		return 1;
578 
579 	/* Otherwise don't allocate. */
580 	return 0;
581 }
582 
583 /*
584  * Return estimate of bytes we think need to be
585  * released to help resolve low memory conditions.
586  *
587  * => called with bufcache_lock held.
588  */
589 static int
590 buf_canrelease(void)
591 {
592 	int pagedemand, ninvalid = 0;
593 
594 	KASSERT(mutex_owned(&bufcache_lock));
595 
596 	if (bufmem < bufmem_lowater)
597 		return 0;
598 
599 	if (bufmem > bufmem_hiwater)
600 		return bufmem - bufmem_hiwater;
601 
602 	ninvalid += bufqueues[BQ_AGE].bq_bytes;
603 
604 	pagedemand = uvmexp.freetarg - uvmexp.free;
605 	if (pagedemand < 0)
606 		return ninvalid;
607 	return MAX(ninvalid, MIN(2 * MAXBSIZE,
608 	    MIN((bufmem - bufmem_lowater) / 16, pagedemand * PAGE_SIZE)));
609 }
610 
611 /*
612  * Buffer memory allocation helper functions
613  */
614 static u_long
615 buf_mempoolidx(u_long size)
616 {
617 	u_int n = 0;
618 
619 	size -= 1;
620 	size >>= MEMPOOL_INDEX_OFFSET;
621 	while (size) {
622 		size >>= 1;
623 		n += 1;
624 	}
625 	if (n >= NMEMPOOLS)
626 		panic("buf mem pool index %d", n);
627 	return n;
628 }
629 
630 static u_long
631 buf_roundsize(u_long size)
632 {
633 	/* Round up to nearest power of 2 */
634 	return (1 << (buf_mempoolidx(size) + MEMPOOL_INDEX_OFFSET));
635 }
636 
637 static void *
638 buf_alloc(size_t size)
639 {
640 	u_int n = buf_mempoolidx(size);
641 	void *addr;
642 
643 	while (1) {
644 		addr = pool_get(&bmempools[n], PR_NOWAIT);
645 		if (addr != NULL)
646 			break;
647 
648 		/* No memory, see if we can free some. If so, try again */
649 		mutex_enter(&bufcache_lock);
650 		if (buf_drain(1) > 0) {
651 			mutex_exit(&bufcache_lock);
652 			continue;
653 		}
654 
655 		if (curlwp == uvm.pagedaemon_lwp) {
656 			mutex_exit(&bufcache_lock);
657 			return NULL;
658 		}
659 
660 		/* Wait for buffers to arrive on the LRU queue */
661 		cv_timedwait(&needbuffer_cv, &bufcache_lock, hz / 4);
662 		mutex_exit(&bufcache_lock);
663 	}
664 
665 	return addr;
666 }
667 
668 static void
669 buf_mrelease(void *addr, size_t size)
670 {
671 
672 	pool_put(&bmempools[buf_mempoolidx(size)], addr);
673 }
674 
675 /*
676  * bread()/breadn() helper.
677  */
678 static buf_t *
679 bio_doread(struct vnode *vp, daddr_t blkno, int size, int async)
680 {
681 	buf_t *bp;
682 	struct mount *mp;
683 
684 	bp = getblk(vp, blkno, size, 0, 0);
685 
686 	/*
687 	 * getblk() may return NULL if we are the pagedaemon.
688 	 */
689 	if (bp == NULL) {
690 		KASSERT(curlwp == uvm.pagedaemon_lwp);
691 		return NULL;
692 	}
693 
694 	/*
695 	 * If buffer does not have data valid, start a read.
696 	 * Note that if buffer is BC_INVAL, getblk() won't return it.
697 	 * Therefore, it's valid if its I/O has completed or been delayed.
698 	 */
699 	if (!ISSET(bp->b_oflags, (BO_DONE | BO_DELWRI))) {
700 		/* Start I/O for the buffer. */
701 		SET(bp->b_flags, B_READ | async);
702 		if (async)
703 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
704 		else
705 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
706 		VOP_STRATEGY(vp, bp);
707 
708 		/* Pay for the read. */
709 		curlwp->l_ru.ru_inblock++;
710 	} else if (async)
711 		brelse(bp, 0);
712 
713 	if (vp->v_type == VBLK)
714 		mp = spec_node_getmountedfs(vp);
715 	else
716 		mp = vp->v_mount;
717 
718 	/*
719 	 * Collect statistics on synchronous and asynchronous reads.
720 	 * Reads from block devices are charged to their associated
721 	 * filesystem (if any).
722 	 */
723 	if (mp != NULL) {
724 		if (async == 0)
725 			mp->mnt_stat.f_syncreads++;
726 		else
727 			mp->mnt_stat.f_asyncreads++;
728 	}
729 
730 	return (bp);
731 }
732 
733 /*
734  * Read a disk block.
735  * This algorithm described in Bach (p.54).
736  */
737 int
738 bread(struct vnode *vp, daddr_t blkno, int size, int flags, buf_t **bpp)
739 {
740 	buf_t *bp;
741 	int error;
742 
743 	BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
744 
745 	/* Get buffer for block. */
746 	bp = *bpp = bio_doread(vp, blkno, size, 0);
747 	if (bp == NULL)
748 		return ENOMEM;
749 
750 	/* Wait for the read to complete, and return result. */
751 	error = biowait(bp);
752 	if (error == 0 && (flags & B_MODIFY) != 0)
753 		error = fscow_run(bp, true);
754 	if (error) {
755 		brelse(bp, 0);
756 		*bpp = NULL;
757 	}
758 
759 	return error;
760 }
761 
762 /*
763  * Read-ahead multiple disk blocks. The first is sync, the rest async.
764  * Trivial modification to the breada algorithm presented in Bach (p.55).
765  */
766 int
767 breadn(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablks,
768     int *rasizes, int nrablks, int flags, buf_t **bpp)
769 {
770 	buf_t *bp;
771 	int error, i;
772 
773 	BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
774 
775 	bp = *bpp = bio_doread(vp, blkno, size, 0);
776 	if (bp == NULL)
777 		return ENOMEM;
778 
779 	/*
780 	 * For each of the read-ahead blocks, start a read, if necessary.
781 	 */
782 	mutex_enter(&bufcache_lock);
783 	for (i = 0; i < nrablks; i++) {
784 		/* If it's in the cache, just go on to next one. */
785 		if (incore(vp, rablks[i]))
786 			continue;
787 
788 		/* Get a buffer for the read-ahead block */
789 		mutex_exit(&bufcache_lock);
790 		(void) bio_doread(vp, rablks[i], rasizes[i], B_ASYNC);
791 		mutex_enter(&bufcache_lock);
792 	}
793 	mutex_exit(&bufcache_lock);
794 
795 	/* Otherwise, we had to start a read for it; wait until it's valid. */
796 	error = biowait(bp);
797 	if (error == 0 && (flags & B_MODIFY) != 0)
798 		error = fscow_run(bp, true);
799 	if (error) {
800 		brelse(bp, 0);
801 		*bpp = NULL;
802 	}
803 
804 	return error;
805 }
806 
807 /*
808  * Block write.  Described in Bach (p.56)
809  */
810 int
811 bwrite(buf_t *bp)
812 {
813 	int rv, sync, wasdelayed;
814 	struct vnode *vp;
815 	struct mount *mp;
816 
817 	BIOHIST_FUNC(__func__); BIOHIST_CALLARGS(biohist, "bp=%p",
818 	    bp, 0, 0, 0);
819 
820 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
821 	KASSERT(!cv_has_waiters(&bp->b_done));
822 
823 	vp = bp->b_vp;
824 
825 	/*
826 	 * dholland 20160728 AFAICT vp==NULL must be impossible as it
827 	 * will crash upon reaching VOP_STRATEGY below... see further
828 	 * analysis on tech-kern.
829 	 */
830 	KASSERTMSG(vp != NULL, "bwrite given buffer with null vnode");
831 
832 	if (vp != NULL) {
833 		KASSERT(bp->b_objlock == vp->v_interlock);
834 		if (vp->v_type == VBLK)
835 			mp = spec_node_getmountedfs(vp);
836 		else
837 			mp = vp->v_mount;
838 	} else {
839 		mp = NULL;
840 	}
841 
842 	if (mp && mp->mnt_wapbl) {
843 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
844 			bdwrite(bp);
845 			return 0;
846 		}
847 	}
848 
849 	/*
850 	 * Remember buffer type, to switch on it later.  If the write was
851 	 * synchronous, but the file system was mounted with MNT_ASYNC,
852 	 * convert it to a delayed write.
853 	 * XXX note that this relies on delayed tape writes being converted
854 	 * to async, not sync writes (which is safe, but ugly).
855 	 */
856 	sync = !ISSET(bp->b_flags, B_ASYNC);
857 	if (sync && mp != NULL && ISSET(mp->mnt_flag, MNT_ASYNC)) {
858 		bdwrite(bp);
859 		return (0);
860 	}
861 
862 	/*
863 	 * Collect statistics on synchronous and asynchronous writes.
864 	 * Writes to block devices are charged to their associated
865 	 * filesystem (if any).
866 	 */
867 	if (mp != NULL) {
868 		if (sync)
869 			mp->mnt_stat.f_syncwrites++;
870 		else
871 			mp->mnt_stat.f_asyncwrites++;
872 	}
873 
874 	/*
875 	 * Pay for the I/O operation and make sure the buf is on the correct
876 	 * vnode queue.
877 	 */
878 	bp->b_error = 0;
879 	wasdelayed = ISSET(bp->b_oflags, BO_DELWRI);
880 	CLR(bp->b_flags, B_READ);
881 	if (wasdelayed) {
882 		mutex_enter(&bufcache_lock);
883 		mutex_enter(bp->b_objlock);
884 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
885 		reassignbuf(bp, bp->b_vp);
886 		mutex_exit(&bufcache_lock);
887 	} else {
888 		curlwp->l_ru.ru_oublock++;
889 		mutex_enter(bp->b_objlock);
890 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
891 	}
892 	if (vp != NULL)
893 		vp->v_numoutput++;
894 	mutex_exit(bp->b_objlock);
895 
896 	/* Initiate disk write. */
897 	if (sync)
898 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
899 	else
900 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
901 
902 	VOP_STRATEGY(vp, bp);
903 
904 	if (sync) {
905 		/* If I/O was synchronous, wait for it to complete. */
906 		rv = biowait(bp);
907 
908 		/* Release the buffer. */
909 		brelse(bp, 0);
910 
911 		return (rv);
912 	} else {
913 		return (0);
914 	}
915 }
916 
917 int
918 vn_bwrite(void *v)
919 {
920 	struct vop_bwrite_args *ap = v;
921 
922 	return (bwrite(ap->a_bp));
923 }
924 
925 /*
926  * Delayed write.
927  *
928  * The buffer is marked dirty, but is not queued for I/O.
929  * This routine should be used when the buffer is expected
930  * to be modified again soon, typically a small write that
931  * partially fills a buffer.
932  *
933  * NB: magnetic tapes cannot be delayed; they must be
934  * written in the order that the writes are requested.
935  *
936  * Described in Leffler, et al. (pp. 208-213).
937  */
938 void
939 bdwrite(buf_t *bp)
940 {
941 
942 	BIOHIST_FUNC(__func__); BIOHIST_CALLARGS(biohist, "bp=%p",
943 	    bp, 0, 0, 0);
944 
945 	KASSERT(bp->b_vp == NULL || bp->b_vp->v_tag != VT_UFS ||
946 	    bp->b_vp->v_type == VBLK || ISSET(bp->b_flags, B_COWDONE));
947 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
948 	KASSERT(!cv_has_waiters(&bp->b_done));
949 
950 	/* If this is a tape block, write the block now. */
951 	if (bdev_type(bp->b_dev) == D_TAPE) {
952 		bawrite(bp);
953 		return;
954 	}
955 
956 	if (wapbl_vphaswapbl(bp->b_vp)) {
957 		struct mount *mp = wapbl_vptomp(bp->b_vp);
958 
959 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
960 			WAPBL_ADD_BUF(mp, bp);
961 		}
962 	}
963 
964 	/*
965 	 * If the block hasn't been seen before:
966 	 *	(1) Mark it as having been seen,
967 	 *	(2) Charge for the write,
968 	 *	(3) Make sure it's on its vnode's correct block list.
969 	 */
970 	KASSERT(bp->b_vp == NULL || bp->b_objlock == bp->b_vp->v_interlock);
971 
972 	if (!ISSET(bp->b_oflags, BO_DELWRI)) {
973 		mutex_enter(&bufcache_lock);
974 		mutex_enter(bp->b_objlock);
975 		SET(bp->b_oflags, BO_DELWRI);
976 		curlwp->l_ru.ru_oublock++;
977 		reassignbuf(bp, bp->b_vp);
978 		mutex_exit(&bufcache_lock);
979 	} else {
980 		mutex_enter(bp->b_objlock);
981 	}
982 	/* Otherwise, the "write" is done, so mark and release the buffer. */
983 	CLR(bp->b_oflags, BO_DONE);
984 	mutex_exit(bp->b_objlock);
985 
986 	brelse(bp, 0);
987 }
988 
989 /*
990  * Asynchronous block write; just an asynchronous bwrite().
991  */
992 void
993 bawrite(buf_t *bp)
994 {
995 
996 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
997 	KASSERT(bp->b_vp != NULL);
998 
999 	SET(bp->b_flags, B_ASYNC);
1000 	VOP_BWRITE(bp->b_vp, bp);
1001 }
1002 
1003 /*
1004  * Release a buffer on to the free lists.
1005  * Described in Bach (p. 46).
1006  */
1007 void
1008 brelsel(buf_t *bp, int set)
1009 {
1010 	struct bqueue *bufq;
1011 	struct vnode *vp;
1012 
1013 	KASSERT(bp != NULL);
1014 	KASSERT(mutex_owned(&bufcache_lock));
1015 	KASSERT(!cv_has_waiters(&bp->b_done));
1016 	KASSERT(bp->b_refcnt > 0);
1017 
1018 	SET(bp->b_cflags, set);
1019 
1020 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1021 	KASSERT(bp->b_iodone == NULL);
1022 
1023 	/* Wake up any processes waiting for any buffer to become free. */
1024 	cv_signal(&needbuffer_cv);
1025 
1026 	/* Wake up any proceeses waiting for _this_ buffer to become free */
1027 	if (ISSET(bp->b_cflags, BC_WANTED))
1028 		CLR(bp->b_cflags, BC_WANTED|BC_AGE);
1029 
1030 	/* If it's clean clear the copy-on-write flag. */
1031 	if (ISSET(bp->b_flags, B_COWDONE)) {
1032 		mutex_enter(bp->b_objlock);
1033 		if (!ISSET(bp->b_oflags, BO_DELWRI))
1034 			CLR(bp->b_flags, B_COWDONE);
1035 		mutex_exit(bp->b_objlock);
1036 	}
1037 
1038 	/*
1039 	 * Determine which queue the buffer should be on, then put it there.
1040 	 */
1041 
1042 	/* If it's locked, don't report an error; try again later. */
1043 	if (ISSET(bp->b_flags, B_LOCKED))
1044 		bp->b_error = 0;
1045 
1046 	/* If it's not cacheable, or an error, mark it invalid. */
1047 	if (ISSET(bp->b_cflags, BC_NOCACHE) || bp->b_error != 0)
1048 		SET(bp->b_cflags, BC_INVAL);
1049 
1050 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1051 		/*
1052 		 * This is a delayed write buffer that was just flushed to
1053 		 * disk.  It is still on the LRU queue.  If it's become
1054 		 * invalid, then we need to move it to a different queue;
1055 		 * otherwise leave it in its current position.
1056 		 */
1057 		CLR(bp->b_cflags, BC_VFLUSH);
1058 		if (!ISSET(bp->b_cflags, BC_INVAL|BC_AGE) &&
1059 		    !ISSET(bp->b_flags, B_LOCKED) && bp->b_error == 0) {
1060 			KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 1));
1061 			goto already_queued;
1062 		} else {
1063 			bremfree(bp);
1064 		}
1065 	}
1066 
1067 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_AGE], 0));
1068 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 0));
1069 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LOCKED], 0));
1070 
1071 	if ((bp->b_bufsize <= 0) || ISSET(bp->b_cflags, BC_INVAL)) {
1072 		/*
1073 		 * If it's invalid or empty, dissociate it from its vnode
1074 		 * and put on the head of the appropriate queue.
1075 		 */
1076 		if (ISSET(bp->b_flags, B_LOCKED)) {
1077 			if (wapbl_vphaswapbl(vp = bp->b_vp)) {
1078 				struct mount *mp = wapbl_vptomp(vp);
1079 
1080 				KASSERT(bp->b_iodone
1081 				    != mp->mnt_wapbl_op->wo_wapbl_biodone);
1082 				WAPBL_REMOVE_BUF(mp, bp);
1083 			}
1084 		}
1085 
1086 		mutex_enter(bp->b_objlock);
1087 		CLR(bp->b_oflags, BO_DONE|BO_DELWRI);
1088 		if ((vp = bp->b_vp) != NULL) {
1089 			KASSERT(bp->b_objlock == vp->v_interlock);
1090 			reassignbuf(bp, bp->b_vp);
1091 			brelvp(bp);
1092 			mutex_exit(vp->v_interlock);
1093 		} else {
1094 			KASSERT(bp->b_objlock == &buffer_lock);
1095 			mutex_exit(bp->b_objlock);
1096 		}
1097 
1098 		if (bp->b_bufsize <= 0)
1099 			/* no data */
1100 			goto already_queued;
1101 		else
1102 			/* invalid data */
1103 			bufq = &bufqueues[BQ_AGE];
1104 		binsheadfree(bp, bufq);
1105 	} else  {
1106 		/*
1107 		 * It has valid data.  Put it on the end of the appropriate
1108 		 * queue, so that it'll stick around for as long as possible.
1109 		 * If buf is AGE, but has dependencies, must put it on last
1110 		 * bufqueue to be scanned, ie LRU. This protects against the
1111 		 * livelock where BQ_AGE only has buffers with dependencies,
1112 		 * and we thus never get to the dependent buffers in BQ_LRU.
1113 		 */
1114 		if (ISSET(bp->b_flags, B_LOCKED)) {
1115 			/* locked in core */
1116 			bufq = &bufqueues[BQ_LOCKED];
1117 		} else if (!ISSET(bp->b_cflags, BC_AGE)) {
1118 			/* valid data */
1119 			bufq = &bufqueues[BQ_LRU];
1120 		} else {
1121 			/* stale but valid data */
1122 			bufq = &bufqueues[BQ_AGE];
1123 		}
1124 		binstailfree(bp, bufq);
1125 	}
1126 already_queued:
1127 	/* Unlock the buffer. */
1128 	CLR(bp->b_cflags, BC_AGE|BC_BUSY|BC_NOCACHE);
1129 	CLR(bp->b_flags, B_ASYNC);
1130 	cv_broadcast(&bp->b_busy);
1131 
1132 	if (bp->b_bufsize <= 0)
1133 		brele(bp);
1134 }
1135 
1136 void
1137 brelse(buf_t *bp, int set)
1138 {
1139 
1140 	mutex_enter(&bufcache_lock);
1141 	brelsel(bp, set);
1142 	mutex_exit(&bufcache_lock);
1143 }
1144 
1145 /*
1146  * Determine if a block is in the cache.
1147  * Just look on what would be its hash chain.  If it's there, return
1148  * a pointer to it, unless it's marked invalid.  If it's marked invalid,
1149  * we normally don't return the buffer, unless the caller explicitly
1150  * wants us to.
1151  */
1152 buf_t *
1153 incore(struct vnode *vp, daddr_t blkno)
1154 {
1155 	buf_t *bp;
1156 
1157 	KASSERT(mutex_owned(&bufcache_lock));
1158 
1159 	/* Search hash chain */
1160 	LIST_FOREACH(bp, BUFHASH(vp, blkno), b_hash) {
1161 		if (bp->b_lblkno == blkno && bp->b_vp == vp &&
1162 		    !ISSET(bp->b_cflags, BC_INVAL)) {
1163 		    	KASSERT(bp->b_objlock == vp->v_interlock);
1164 		    	return (bp);
1165 		}
1166 	}
1167 
1168 	return (NULL);
1169 }
1170 
1171 /*
1172  * Get a block of requested size that is associated with
1173  * a given vnode and block offset. If it is found in the
1174  * block cache, mark it as having been found, make it busy
1175  * and return it. Otherwise, return an empty block of the
1176  * correct size. It is up to the caller to insure that the
1177  * cached blocks be of the correct size.
1178  */
1179 buf_t *
1180 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo)
1181 {
1182 	int err, preserve;
1183 	buf_t *bp;
1184 
1185 	mutex_enter(&bufcache_lock);
1186  loop:
1187 	bp = incore(vp, blkno);
1188 	if (bp != NULL) {
1189 		err = bbusy(bp, ((slpflag & PCATCH) != 0), slptimeo, NULL);
1190 		if (err != 0) {
1191 			if (err == EPASSTHROUGH)
1192 				goto loop;
1193 			mutex_exit(&bufcache_lock);
1194 			return (NULL);
1195 		}
1196 		KASSERT(!cv_has_waiters(&bp->b_done));
1197 #ifdef DIAGNOSTIC
1198 		if (ISSET(bp->b_oflags, BO_DONE|BO_DELWRI) &&
1199 		    bp->b_bcount < size && vp->v_type != VBLK)
1200 			panic("getblk: block size invariant failed");
1201 #endif
1202 		bremfree(bp);
1203 		preserve = 1;
1204 	} else {
1205 		if ((bp = getnewbuf(slpflag, slptimeo, 0)) == NULL)
1206 			goto loop;
1207 
1208 		if (incore(vp, blkno) != NULL) {
1209 			/* The block has come into memory in the meantime. */
1210 			brelsel(bp, 0);
1211 			goto loop;
1212 		}
1213 
1214 		LIST_INSERT_HEAD(BUFHASH(vp, blkno), bp, b_hash);
1215 		bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno;
1216 		mutex_enter(vp->v_interlock);
1217 		bgetvp(vp, bp);
1218 		mutex_exit(vp->v_interlock);
1219 		preserve = 0;
1220 	}
1221 	mutex_exit(&bufcache_lock);
1222 
1223 	/*
1224 	 * LFS can't track total size of B_LOCKED buffer (locked_queue_bytes)
1225 	 * if we re-size buffers here.
1226 	 */
1227 	if (ISSET(bp->b_flags, B_LOCKED)) {
1228 		KASSERT(bp->b_bufsize >= size);
1229 	} else {
1230 		if (allocbuf(bp, size, preserve)) {
1231 			mutex_enter(&bufcache_lock);
1232 			LIST_REMOVE(bp, b_hash);
1233 			mutex_exit(&bufcache_lock);
1234 			brelse(bp, BC_INVAL);
1235 			return NULL;
1236 		}
1237 	}
1238 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1239 	return (bp);
1240 }
1241 
1242 /*
1243  * Get an empty, disassociated buffer of given size.
1244  */
1245 buf_t *
1246 geteblk(int size)
1247 {
1248 	buf_t *bp;
1249 	int error __diagused;
1250 
1251 	mutex_enter(&bufcache_lock);
1252 	while ((bp = getnewbuf(0, 0, 0)) == NULL)
1253 		;
1254 
1255 	SET(bp->b_cflags, BC_INVAL);
1256 	LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1257 	mutex_exit(&bufcache_lock);
1258 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1259 	error = allocbuf(bp, size, 0);
1260 	KASSERT(error == 0);
1261 	return (bp);
1262 }
1263 
1264 /*
1265  * Expand or contract the actual memory allocated to a buffer.
1266  *
1267  * If the buffer shrinks, data is lost, so it's up to the
1268  * caller to have written it out *first*; this routine will not
1269  * start a write.  If the buffer grows, it's the callers
1270  * responsibility to fill out the buffer's additional contents.
1271  */
1272 int
1273 allocbuf(buf_t *bp, int size, int preserve)
1274 {
1275 	void *addr;
1276 	vsize_t oldsize, desired_size;
1277 	int oldcount;
1278 	int delta;
1279 
1280 	desired_size = buf_roundsize(size);
1281 	if (desired_size > MAXBSIZE)
1282 		printf("allocbuf: buffer larger than MAXBSIZE requested");
1283 
1284 	oldcount = bp->b_bcount;
1285 
1286 	bp->b_bcount = size;
1287 
1288 	oldsize = bp->b_bufsize;
1289 	if (oldsize == desired_size) {
1290 		/*
1291 		 * Do not short cut the WAPBL resize, as the buffer length
1292 		 * could still have changed and this would corrupt the
1293 		 * tracking of the transaction length.
1294 		 */
1295 		goto out;
1296 	}
1297 
1298 	/*
1299 	 * If we want a buffer of a different size, re-allocate the
1300 	 * buffer's memory; copy old content only if needed.
1301 	 */
1302 	addr = buf_alloc(desired_size);
1303 	if (addr == NULL)
1304 		return ENOMEM;
1305 	if (preserve)
1306 		memcpy(addr, bp->b_data, MIN(oldsize,desired_size));
1307 	if (bp->b_data != NULL)
1308 		buf_mrelease(bp->b_data, oldsize);
1309 	bp->b_data = addr;
1310 	bp->b_bufsize = desired_size;
1311 
1312 	/*
1313 	 * Update overall buffer memory counter (protected by bufcache_lock)
1314 	 */
1315 	delta = (long)desired_size - (long)oldsize;
1316 
1317 	mutex_enter(&bufcache_lock);
1318 	if ((bufmem += delta) > bufmem_hiwater) {
1319 		/*
1320 		 * Need to trim overall memory usage.
1321 		 */
1322 		while (buf_canrelease()) {
1323 			if (curcpu()->ci_schedstate.spc_flags &
1324 			    SPCF_SHOULDYIELD) {
1325 				mutex_exit(&bufcache_lock);
1326 				preempt();
1327 				mutex_enter(&bufcache_lock);
1328 			}
1329 			if (buf_trim() == 0)
1330 				break;
1331 		}
1332 	}
1333 	mutex_exit(&bufcache_lock);
1334 
1335  out:
1336 	if (wapbl_vphaswapbl(bp->b_vp))
1337 		WAPBL_RESIZE_BUF(wapbl_vptomp(bp->b_vp), bp, oldsize, oldcount);
1338 
1339 	return 0;
1340 }
1341 
1342 /*
1343  * Find a buffer which is available for use.
1344  * Select something from a free list.
1345  * Preference is to AGE list, then LRU list.
1346  *
1347  * Called with the buffer queues locked.
1348  * Return buffer locked.
1349  */
1350 buf_t *
1351 getnewbuf(int slpflag, int slptimeo, int from_bufq)
1352 {
1353 	buf_t *bp;
1354 	struct vnode *vp;
1355 
1356  start:
1357 	KASSERT(mutex_owned(&bufcache_lock));
1358 
1359 	/*
1360 	 * Get a new buffer from the pool.
1361 	 */
1362 	if (!from_bufq && buf_lotsfree()) {
1363 		mutex_exit(&bufcache_lock);
1364 		bp = pool_cache_get(buf_cache, PR_NOWAIT);
1365 		if (bp != NULL) {
1366 			memset((char *)bp, 0, sizeof(*bp));
1367 			buf_init(bp);
1368 			SET(bp->b_cflags, BC_BUSY);	/* mark buffer busy */
1369 			mutex_enter(&bufcache_lock);
1370 #if defined(DIAGNOSTIC)
1371 			bp->b_freelistindex = -1;
1372 #endif /* defined(DIAGNOSTIC) */
1373 			return (bp);
1374 		}
1375 		mutex_enter(&bufcache_lock);
1376 	}
1377 
1378 	KASSERT(mutex_owned(&bufcache_lock));
1379 	if ((bp = TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue)) != NULL ||
1380 	    (bp = TAILQ_FIRST(&bufqueues[BQ_LRU].bq_queue)) != NULL) {
1381 	    	KASSERT(!ISSET(bp->b_cflags, BC_BUSY) || ISSET(bp->b_cflags, BC_VFLUSH));
1382 		bremfree(bp);
1383 
1384 		/* Buffer is no longer on free lists. */
1385 		SET(bp->b_cflags, BC_BUSY);
1386 	} else {
1387 		/*
1388 		 * XXX: !from_bufq should be removed.
1389 		 */
1390 		if (!from_bufq || curlwp != uvm.pagedaemon_lwp) {
1391 			/* wait for a free buffer of any kind */
1392 			if ((slpflag & PCATCH) != 0)
1393 				(void)cv_timedwait_sig(&needbuffer_cv,
1394 				    &bufcache_lock, slptimeo);
1395 			else
1396 				(void)cv_timedwait(&needbuffer_cv,
1397 				    &bufcache_lock, slptimeo);
1398 		}
1399 		return (NULL);
1400 	}
1401 
1402 #ifdef DIAGNOSTIC
1403 	if (bp->b_bufsize <= 0)
1404 		panic("buffer %p: on queue but empty", bp);
1405 #endif
1406 
1407 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1408 		/*
1409 		 * This is a delayed write buffer being flushed to disk.  Make
1410 		 * sure it gets aged out of the queue when it's finished, and
1411 		 * leave it off the LRU queue.
1412 		 */
1413 		CLR(bp->b_cflags, BC_VFLUSH);
1414 		SET(bp->b_cflags, BC_AGE);
1415 		goto start;
1416 	}
1417 
1418 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1419 	KASSERT(bp->b_refcnt > 0);
1420     	KASSERT(!cv_has_waiters(&bp->b_done));
1421 
1422 	/*
1423 	 * If buffer was a delayed write, start it and return NULL
1424 	 * (since we might sleep while starting the write).
1425 	 */
1426 	if (ISSET(bp->b_oflags, BO_DELWRI)) {
1427 		/*
1428 		 * This buffer has gone through the LRU, so make sure it gets
1429 		 * reused ASAP.
1430 		 */
1431 		SET(bp->b_cflags, BC_AGE);
1432 		mutex_exit(&bufcache_lock);
1433 		bawrite(bp);
1434 		mutex_enter(&bufcache_lock);
1435 		return (NULL);
1436 	}
1437 
1438 	vp = bp->b_vp;
1439 
1440 	/* clear out various other fields */
1441 	bp->b_cflags = BC_BUSY;
1442 	bp->b_oflags = 0;
1443 	bp->b_flags = 0;
1444 	bp->b_dev = NODEV;
1445 	bp->b_blkno = 0;
1446 	bp->b_lblkno = 0;
1447 	bp->b_rawblkno = 0;
1448 	bp->b_iodone = 0;
1449 	bp->b_error = 0;
1450 	bp->b_resid = 0;
1451 	bp->b_bcount = 0;
1452 
1453 	LIST_REMOVE(bp, b_hash);
1454 
1455 	/* Disassociate us from our vnode, if we had one... */
1456 	if (vp != NULL) {
1457 		mutex_enter(vp->v_interlock);
1458 		brelvp(bp);
1459 		mutex_exit(vp->v_interlock);
1460 	}
1461 
1462 	return (bp);
1463 }
1464 
1465 /*
1466  * Attempt to free an aged buffer off the queues.
1467  * Called with queue lock held.
1468  * Returns the amount of buffer memory freed.
1469  */
1470 static int
1471 buf_trim(void)
1472 {
1473 	buf_t *bp;
1474 	long size;
1475 
1476 	KASSERT(mutex_owned(&bufcache_lock));
1477 
1478 	/* Instruct getnewbuf() to get buffers off the queues */
1479 	if ((bp = getnewbuf(PCATCH, 1, 1)) == NULL)
1480 		return 0;
1481 
1482 	KASSERT((bp->b_cflags & BC_WANTED) == 0);
1483 	size = bp->b_bufsize;
1484 	bufmem -= size;
1485 	if (size > 0) {
1486 		buf_mrelease(bp->b_data, size);
1487 		bp->b_bcount = bp->b_bufsize = 0;
1488 	}
1489 	/* brelse() will return the buffer to the global buffer pool */
1490 	brelsel(bp, 0);
1491 	return size;
1492 }
1493 
1494 int
1495 buf_drain(int n)
1496 {
1497 	int size = 0, sz;
1498 
1499 	KASSERT(mutex_owned(&bufcache_lock));
1500 
1501 	while (size < n && bufmem > bufmem_lowater) {
1502 		sz = buf_trim();
1503 		if (sz <= 0)
1504 			break;
1505 		size += sz;
1506 	}
1507 
1508 	return size;
1509 }
1510 
1511 SDT_PROVIDER_DEFINE(io);
1512 
1513 SDT_PROBE_DEFINE1(io, kernel, , wait__start, "struct buf *"/*bp*/);
1514 SDT_PROBE_DEFINE1(io, kernel, , wait__done, "struct buf *"/*bp*/);
1515 
1516 /*
1517  * Wait for operations on the buffer to complete.
1518  * When they do, extract and return the I/O's error value.
1519  */
1520 int
1521 biowait(buf_t *bp)
1522 {
1523 
1524 	BIOHIST_FUNC(__func__);
1525 
1526 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1527 	KASSERT(bp->b_refcnt > 0);
1528 
1529 	SDT_PROBE1(io, kernel, , wait__start, bp);
1530 
1531 	mutex_enter(bp->b_objlock);
1532 
1533 	BIOHIST_CALLARGS(biohist, "bp=%p, oflags=0x%x",
1534 	    bp, bp->b_oflags, 0, 0);
1535 
1536 	while (!ISSET(bp->b_oflags, BO_DONE | BO_DELWRI)) {
1537 		BIOHIST_LOG(biohist, "waiting bp=%p", bp, 0, 0, 0);
1538 		cv_wait(&bp->b_done, bp->b_objlock);
1539 	}
1540 	mutex_exit(bp->b_objlock);
1541 
1542 	SDT_PROBE1(io, kernel, , wait__done, bp);
1543 
1544 	BIOHIST_LOG(biohist, "return %d", bp->b_error, 0, 0, 0);
1545 
1546 	return bp->b_error;
1547 }
1548 
1549 /*
1550  * Mark I/O complete on a buffer.
1551  *
1552  * If a callback has been requested, e.g. the pageout
1553  * daemon, do so. Otherwise, awaken waiting processes.
1554  *
1555  * [ Leffler, et al., says on p.247:
1556  *	"This routine wakes up the blocked process, frees the buffer
1557  *	for an asynchronous write, or, for a request by the pagedaemon
1558  *	process, invokes a procedure specified in the buffer structure" ]
1559  *
1560  * In real life, the pagedaemon (or other system processes) wants
1561  * to do async stuff too, and doesn't want the buffer brelse()'d.
1562  * (for swap pager, that puts swap buffers on the free lists (!!!),
1563  * for the vn device, that puts allocated buffers on the free lists!)
1564  */
1565 void
1566 biodone(buf_t *bp)
1567 {
1568 	int s;
1569 
1570 	BIOHIST_FUNC(__func__);
1571 
1572 	KASSERT(!ISSET(bp->b_oflags, BO_DONE));
1573 
1574 	if (cpu_intr_p()) {
1575 		/* From interrupt mode: defer to a soft interrupt. */
1576 		s = splvm();
1577 		TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_biodone, bp, b_actq);
1578 
1579 		BIOHIST_CALLARGS(biohist, "bp=%p, softint scheduled",
1580 		    bp, 0, 0, 0);
1581 		softint_schedule(biodone_sih);
1582 		splx(s);
1583 	} else {
1584 		/* Process now - the buffer may be freed soon. */
1585 		biodone2(bp);
1586 	}
1587 }
1588 
1589 SDT_PROBE_DEFINE1(io, kernel, , done, "struct buf *"/*bp*/);
1590 
1591 static void
1592 biodone2(buf_t *bp)
1593 {
1594 	void (*callout)(buf_t *);
1595 
1596 	SDT_PROBE1(io, kernel, ,done, bp);
1597 
1598 	BIOHIST_FUNC(__func__);
1599 	BIOHIST_CALLARGS(biohist, "bp=%p", bp, 0, 0, 0);
1600 
1601 	mutex_enter(bp->b_objlock);
1602 	/* Note that the transfer is done. */
1603 	if (ISSET(bp->b_oflags, BO_DONE))
1604 		panic("biodone2 already");
1605 	CLR(bp->b_flags, B_COWDONE);
1606 	SET(bp->b_oflags, BO_DONE);
1607 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1608 
1609 	/* Wake up waiting writers. */
1610 	if (!ISSET(bp->b_flags, B_READ))
1611 		vwakeup(bp);
1612 
1613 	if ((callout = bp->b_iodone) != NULL) {
1614 		BIOHIST_LOG(biohist, "callout %p", callout, 0, 0, 0);
1615 
1616 		/* Note callout done, then call out. */
1617 		KASSERT(!cv_has_waiters(&bp->b_done));
1618 		KERNEL_LOCK(1, NULL);		/* XXXSMP */
1619 		bp->b_iodone = NULL;
1620 		mutex_exit(bp->b_objlock);
1621 		(*callout)(bp);
1622 		KERNEL_UNLOCK_ONE(NULL);	/* XXXSMP */
1623 	} else if (ISSET(bp->b_flags, B_ASYNC)) {
1624 		/* If async, release. */
1625 		BIOHIST_LOG(biohist, "async", 0, 0, 0, 0);
1626 		KASSERT(!cv_has_waiters(&bp->b_done));
1627 		mutex_exit(bp->b_objlock);
1628 		brelse(bp, 0);
1629 	} else {
1630 		/* Otherwise just wake up waiters in biowait(). */
1631 		BIOHIST_LOG(biohist, "wake-up", 0, 0, 0, 0);
1632 		cv_broadcast(&bp->b_done);
1633 		mutex_exit(bp->b_objlock);
1634 	}
1635 }
1636 
1637 static void
1638 biointr(void *cookie)
1639 {
1640 	struct cpu_info *ci;
1641 	buf_t *bp;
1642 	int s;
1643 
1644 	BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
1645 
1646 	ci = curcpu();
1647 
1648 	s = splvm();
1649 	while (!TAILQ_EMPTY(&ci->ci_data.cpu_biodone)) {
1650 		KASSERT(curcpu() == ci);
1651 
1652 		bp = TAILQ_FIRST(&ci->ci_data.cpu_biodone);
1653 		TAILQ_REMOVE(&ci->ci_data.cpu_biodone, bp, b_actq);
1654 		splx(s);
1655 
1656 		BIOHIST_LOG(biohist, "bp=%p", bp, 0, 0, 0);
1657 		biodone2(bp);
1658 
1659 		s = splvm();
1660 	}
1661 	splx(s);
1662 }
1663 
1664 /*
1665  * Wait for all buffers to complete I/O
1666  * Return the number of "stuck" buffers.
1667  */
1668 int
1669 buf_syncwait(void)
1670 {
1671 	buf_t *bp;
1672 	int iter, nbusy, nbusy_prev = 0, ihash;
1673 
1674 	BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
1675 
1676 	for (iter = 0; iter < 20;) {
1677 		mutex_enter(&bufcache_lock);
1678 		nbusy = 0;
1679 		for (ihash = 0; ihash < bufhash+1; ihash++) {
1680 		    LIST_FOREACH(bp, &bufhashtbl[ihash], b_hash) {
1681 			if ((bp->b_cflags & (BC_BUSY|BC_INVAL)) == BC_BUSY)
1682 				nbusy += ((bp->b_flags & B_READ) == 0);
1683 		    }
1684 		}
1685 		mutex_exit(&bufcache_lock);
1686 
1687 		if (nbusy == 0)
1688 			break;
1689 		if (nbusy_prev == 0)
1690 			nbusy_prev = nbusy;
1691 		printf("%d ", nbusy);
1692 		kpause("bflush", false, MAX(1, hz / 25 * iter), NULL);
1693 		if (nbusy >= nbusy_prev) /* we didn't flush anything */
1694 			iter++;
1695 		else
1696 			nbusy_prev = nbusy;
1697 	}
1698 
1699 	if (nbusy) {
1700 #if defined(DEBUG) || defined(DEBUG_HALT_BUSY)
1701 		printf("giving up\nPrinting vnodes for busy buffers\n");
1702 		for (ihash = 0; ihash < bufhash+1; ihash++) {
1703 		    LIST_FOREACH(bp, &bufhashtbl[ihash], b_hash) {
1704 			if ((bp->b_cflags & (BC_BUSY|BC_INVAL)) == BC_BUSY &&
1705 			    (bp->b_flags & B_READ) == 0)
1706 				vprint(NULL, bp->b_vp);
1707 		    }
1708 		}
1709 #endif
1710 	}
1711 
1712 	return nbusy;
1713 }
1714 
1715 static void
1716 sysctl_fillbuf(buf_t *i, struct buf_sysctl *o)
1717 {
1718 
1719 	o->b_flags = i->b_flags | i->b_cflags | i->b_oflags;
1720 	o->b_error = i->b_error;
1721 	o->b_prio = i->b_prio;
1722 	o->b_dev = i->b_dev;
1723 	o->b_bufsize = i->b_bufsize;
1724 	o->b_bcount = i->b_bcount;
1725 	o->b_resid = i->b_resid;
1726 	o->b_addr = PTRTOUINT64(i->b_data);
1727 	o->b_blkno = i->b_blkno;
1728 	o->b_rawblkno = i->b_rawblkno;
1729 	o->b_iodone = PTRTOUINT64(i->b_iodone);
1730 	o->b_proc = PTRTOUINT64(i->b_proc);
1731 	o->b_vp = PTRTOUINT64(i->b_vp);
1732 	o->b_saveaddr = PTRTOUINT64(i->b_saveaddr);
1733 	o->b_lblkno = i->b_lblkno;
1734 }
1735 
1736 #define KERN_BUFSLOP 20
1737 static int
1738 sysctl_dobuf(SYSCTLFN_ARGS)
1739 {
1740 	buf_t *bp;
1741 	struct buf_sysctl bs;
1742 	struct bqueue *bq;
1743 	char *dp;
1744 	u_int i, op, arg;
1745 	size_t len, needed, elem_size, out_size;
1746 	int error, elem_count, retries;
1747 
1748 	if (namelen == 1 && name[0] == CTL_QUERY)
1749 		return (sysctl_query(SYSCTLFN_CALL(rnode)));
1750 
1751 	if (namelen != 4)
1752 		return (EINVAL);
1753 
1754 	retries = 100;
1755  retry:
1756 	dp = oldp;
1757 	len = (oldp != NULL) ? *oldlenp : 0;
1758 	op = name[0];
1759 	arg = name[1];
1760 	elem_size = name[2];
1761 	elem_count = name[3];
1762 	out_size = MIN(sizeof(bs), elem_size);
1763 
1764 	/*
1765 	 * at the moment, these are just "placeholders" to make the
1766 	 * API for retrieving kern.buf data more extensible in the
1767 	 * future.
1768 	 *
1769 	 * XXX kern.buf currently has "netbsd32" issues.  hopefully
1770 	 * these will be resolved at a later point.
1771 	 */
1772 	if (op != KERN_BUF_ALL || arg != KERN_BUF_ALL ||
1773 	    elem_size < 1 || elem_count < 0)
1774 		return (EINVAL);
1775 
1776 	error = 0;
1777 	needed = 0;
1778 	sysctl_unlock();
1779 	mutex_enter(&bufcache_lock);
1780 	for (i = 0; i < BQUEUES; i++) {
1781 		bq = &bufqueues[i];
1782 		TAILQ_FOREACH(bp, &bq->bq_queue, b_freelist) {
1783 			bq->bq_marker = bp;
1784 			if (len >= elem_size && elem_count > 0) {
1785 				sysctl_fillbuf(bp, &bs);
1786 				mutex_exit(&bufcache_lock);
1787 				error = copyout(&bs, dp, out_size);
1788 				mutex_enter(&bufcache_lock);
1789 				if (error)
1790 					break;
1791 				if (bq->bq_marker != bp) {
1792 					/*
1793 					 * This sysctl node is only for
1794 					 * statistics.  Retry; if the
1795 					 * queue keeps changing, then
1796 					 * bail out.
1797 					 */
1798 					if (retries-- == 0) {
1799 						error = EAGAIN;
1800 						break;
1801 					}
1802 					mutex_exit(&bufcache_lock);
1803 					sysctl_relock();
1804 					goto retry;
1805 				}
1806 				dp += elem_size;
1807 				len -= elem_size;
1808 			}
1809 			needed += elem_size;
1810 			if (elem_count > 0 && elem_count != INT_MAX)
1811 				elem_count--;
1812 		}
1813 		if (error != 0)
1814 			break;
1815 	}
1816 	mutex_exit(&bufcache_lock);
1817 	sysctl_relock();
1818 
1819 	*oldlenp = needed;
1820 	if (oldp == NULL)
1821 		*oldlenp += KERN_BUFSLOP * sizeof(buf_t);
1822 
1823 	return (error);
1824 }
1825 
1826 static int
1827 sysctl_bufvm_update(SYSCTLFN_ARGS)
1828 {
1829 	int error, rv;
1830 	struct sysctlnode node;
1831 	unsigned int temp_bufcache;
1832 	unsigned long temp_water;
1833 
1834 	/* Take a copy of the supplied node and its data */
1835 	node = *rnode;
1836 	if (node.sysctl_data == &bufcache) {
1837 	    node.sysctl_data = &temp_bufcache;
1838 	    temp_bufcache = *(unsigned int *)rnode->sysctl_data;
1839 	} else {
1840 	    node.sysctl_data = &temp_water;
1841 	    temp_water = *(unsigned long *)rnode->sysctl_data;
1842 	}
1843 
1844 	/* Update the copy */
1845 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
1846 	if (error || newp == NULL)
1847 		return (error);
1848 
1849 	if (rnode->sysctl_data == &bufcache) {
1850 		if (temp_bufcache > 100)
1851 			return (EINVAL);
1852 		bufcache = temp_bufcache;
1853 		buf_setwm();
1854 	} else if (rnode->sysctl_data == &bufmem_lowater) {
1855 		if (bufmem_hiwater - temp_water < 16)
1856 			return (EINVAL);
1857 		bufmem_lowater = temp_water;
1858 	} else if (rnode->sysctl_data == &bufmem_hiwater) {
1859 		if (temp_water - bufmem_lowater < 16)
1860 			return (EINVAL);
1861 		bufmem_hiwater = temp_water;
1862 	} else
1863 		return (EINVAL);
1864 
1865 	/* Drain until below new high water mark */
1866 	sysctl_unlock();
1867 	mutex_enter(&bufcache_lock);
1868 	while (bufmem > bufmem_hiwater) {
1869 		rv = buf_drain((bufmem - bufmem_hiwater) / (2 * 1024));
1870 		if (rv <= 0)
1871 			break;
1872 	}
1873 	mutex_exit(&bufcache_lock);
1874 	sysctl_relock();
1875 
1876 	return 0;
1877 }
1878 
1879 static struct sysctllog *vfsbio_sysctllog;
1880 
1881 static void
1882 sysctl_kern_buf_setup(void)
1883 {
1884 
1885 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1886 		       CTLFLAG_PERMANENT,
1887 		       CTLTYPE_NODE, "buf",
1888 		       SYSCTL_DESCR("Kernel buffer cache information"),
1889 		       sysctl_dobuf, 0, NULL, 0,
1890 		       CTL_KERN, KERN_BUF, CTL_EOL);
1891 }
1892 
1893 static void
1894 sysctl_vm_buf_setup(void)
1895 {
1896 
1897 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1898 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1899 		       CTLTYPE_INT, "bufcache",
1900 		       SYSCTL_DESCR("Percentage of physical memory to use for "
1901 				    "buffer cache"),
1902 		       sysctl_bufvm_update, 0, &bufcache, 0,
1903 		       CTL_VM, CTL_CREATE, CTL_EOL);
1904 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1905 		       CTLFLAG_PERMANENT|CTLFLAG_READONLY,
1906 		       CTLTYPE_LONG, "bufmem",
1907 		       SYSCTL_DESCR("Amount of kernel memory used by buffer "
1908 				    "cache"),
1909 		       NULL, 0, &bufmem, 0,
1910 		       CTL_VM, CTL_CREATE, CTL_EOL);
1911 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1912 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1913 		       CTLTYPE_LONG, "bufmem_lowater",
1914 		       SYSCTL_DESCR("Minimum amount of kernel memory to "
1915 				    "reserve for buffer cache"),
1916 		       sysctl_bufvm_update, 0, &bufmem_lowater, 0,
1917 		       CTL_VM, CTL_CREATE, CTL_EOL);
1918 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1919 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1920 		       CTLTYPE_LONG, "bufmem_hiwater",
1921 		       SYSCTL_DESCR("Maximum amount of kernel memory to use "
1922 				    "for buffer cache"),
1923 		       sysctl_bufvm_update, 0, &bufmem_hiwater, 0,
1924 		       CTL_VM, CTL_CREATE, CTL_EOL);
1925 }
1926 
1927 #ifdef DEBUG
1928 /*
1929  * Print out statistics on the current allocation of the buffer pool.
1930  * Can be enabled to print out on every ``sync'' by setting "syncprt"
1931  * in vfs_syscalls.c using sysctl.
1932  */
1933 void
1934 vfs_bufstats(void)
1935 {
1936 	int i, j, count;
1937 	buf_t *bp;
1938 	struct bqueue *dp;
1939 	int counts[MAXBSIZE / MIN_PAGE_SIZE + 1];
1940 	static const char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE" };
1941 
1942 	for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) {
1943 		count = 0;
1944 		memset(counts, 0, sizeof(counts));
1945 		TAILQ_FOREACH(bp, &dp->bq_queue, b_freelist) {
1946 			counts[bp->b_bufsize / PAGE_SIZE]++;
1947 			count++;
1948 		}
1949 		printf("%s: total-%d", bname[i], count);
1950 		for (j = 0; j <= MAXBSIZE / PAGE_SIZE; j++)
1951 			if (counts[j] != 0)
1952 				printf(", %d-%d", j * PAGE_SIZE, counts[j]);
1953 		printf("\n");
1954 	}
1955 }
1956 #endif /* DEBUG */
1957 
1958 /* ------------------------------ */
1959 
1960 buf_t *
1961 getiobuf(struct vnode *vp, bool waitok)
1962 {
1963 	buf_t *bp;
1964 
1965 	bp = pool_cache_get(bufio_cache, (waitok ? PR_WAITOK : PR_NOWAIT));
1966 	if (bp == NULL)
1967 		return bp;
1968 
1969 	buf_init(bp);
1970 
1971 	if ((bp->b_vp = vp) != NULL) {
1972 		bp->b_objlock = vp->v_interlock;
1973 	} else {
1974 		KASSERT(bp->b_objlock == &buffer_lock);
1975 	}
1976 
1977 	return bp;
1978 }
1979 
1980 void
1981 putiobuf(buf_t *bp)
1982 {
1983 
1984 	buf_destroy(bp);
1985 	pool_cache_put(bufio_cache, bp);
1986 }
1987 
1988 /*
1989  * nestiobuf_iodone: b_iodone callback for nested buffers.
1990  */
1991 
1992 void
1993 nestiobuf_iodone(buf_t *bp)
1994 {
1995 	buf_t *mbp = bp->b_private;
1996 	int error;
1997 	int donebytes;
1998 
1999 	KASSERT(bp->b_bcount <= bp->b_bufsize);
2000 	KASSERT(mbp != bp);
2001 
2002 	error = bp->b_error;
2003 	if (bp->b_error == 0 &&
2004 	    (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
2005 		/*
2006 		 * Not all got transfered, raise an error. We have no way to
2007 		 * propagate these conditions to mbp.
2008 		 */
2009 		error = EIO;
2010 	}
2011 
2012 	donebytes = bp->b_bufsize;
2013 
2014 	putiobuf(bp);
2015 	nestiobuf_done(mbp, donebytes, error);
2016 }
2017 
2018 /*
2019  * nestiobuf_setup: setup a "nested" buffer.
2020  *
2021  * => 'mbp' is a "master" buffer which is being divided into sub pieces.
2022  * => 'bp' should be a buffer allocated by getiobuf.
2023  * => 'offset' is a byte offset in the master buffer.
2024  * => 'size' is a size in bytes of this nested buffer.
2025  */
2026 
2027 void
2028 nestiobuf_setup(buf_t *mbp, buf_t *bp, int offset, size_t size)
2029 {
2030 	const int b_read = mbp->b_flags & B_READ;
2031 	struct vnode *vp = mbp->b_vp;
2032 
2033 	KASSERT(mbp->b_bcount >= offset + size);
2034 	bp->b_vp = vp;
2035 	bp->b_dev = mbp->b_dev;
2036 	bp->b_objlock = mbp->b_objlock;
2037 	bp->b_cflags = BC_BUSY;
2038 	bp->b_flags = B_ASYNC | b_read;
2039 	bp->b_iodone = nestiobuf_iodone;
2040 	bp->b_data = (char *)mbp->b_data + offset;
2041 	bp->b_resid = bp->b_bcount = size;
2042 	bp->b_bufsize = bp->b_bcount;
2043 	bp->b_private = mbp;
2044 	BIO_COPYPRIO(bp, mbp);
2045 	if (!b_read && vp != NULL) {
2046 		mutex_enter(vp->v_interlock);
2047 		vp->v_numoutput++;
2048 		mutex_exit(vp->v_interlock);
2049 	}
2050 }
2051 
2052 /*
2053  * nestiobuf_done: propagate completion to the master buffer.
2054  *
2055  * => 'donebytes' specifies how many bytes in the 'mbp' is completed.
2056  * => 'error' is an errno(2) that 'donebytes' has been completed with.
2057  */
2058 
2059 void
2060 nestiobuf_done(buf_t *mbp, int donebytes, int error)
2061 {
2062 
2063 	if (donebytes == 0) {
2064 		return;
2065 	}
2066 	mutex_enter(mbp->b_objlock);
2067 	KASSERT(mbp->b_resid >= donebytes);
2068 	mbp->b_resid -= donebytes;
2069 	if (error)
2070 		mbp->b_error = error;
2071 	if (mbp->b_resid == 0) {
2072 		if (mbp->b_error)
2073 			mbp->b_resid = mbp->b_bcount;
2074 		mutex_exit(mbp->b_objlock);
2075 		biodone(mbp);
2076 	} else
2077 		mutex_exit(mbp->b_objlock);
2078 }
2079 
2080 void
2081 buf_init(buf_t *bp)
2082 {
2083 
2084 	cv_init(&bp->b_busy, "biolock");
2085 	cv_init(&bp->b_done, "biowait");
2086 	bp->b_dev = NODEV;
2087 	bp->b_error = 0;
2088 	bp->b_flags = 0;
2089 	bp->b_cflags = 0;
2090 	bp->b_oflags = 0;
2091 	bp->b_objlock = &buffer_lock;
2092 	bp->b_iodone = NULL;
2093 	bp->b_refcnt = 1;
2094 	bp->b_dev = NODEV;
2095 	bp->b_vnbufs.le_next = NOLIST;
2096 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
2097 }
2098 
2099 void
2100 buf_destroy(buf_t *bp)
2101 {
2102 
2103 	cv_destroy(&bp->b_done);
2104 	cv_destroy(&bp->b_busy);
2105 }
2106 
2107 int
2108 bbusy(buf_t *bp, bool intr, int timo, kmutex_t *interlock)
2109 {
2110 	int error;
2111 
2112 	KASSERT(mutex_owned(&bufcache_lock));
2113 
2114 	if ((bp->b_cflags & BC_BUSY) != 0) {
2115 		if (curlwp == uvm.pagedaemon_lwp)
2116 			return EDEADLK;
2117 		bp->b_cflags |= BC_WANTED;
2118 		bref(bp);
2119 		if (interlock != NULL)
2120 			mutex_exit(interlock);
2121 		if (intr) {
2122 			error = cv_timedwait_sig(&bp->b_busy, &bufcache_lock,
2123 			    timo);
2124 		} else {
2125 			error = cv_timedwait(&bp->b_busy, &bufcache_lock,
2126 			    timo);
2127 		}
2128 		brele(bp);
2129 		if (interlock != NULL)
2130 			mutex_enter(interlock);
2131 		if (error != 0)
2132 			return error;
2133 		return EPASSTHROUGH;
2134 	}
2135 	bp->b_cflags |= BC_BUSY;
2136 
2137 	return 0;
2138 }
2139