xref: /netbsd-src/sys/kern/vfs_bio.c (revision e94a5d02693120d4ad9d909e488894e9fcf0eb76)
1 /*	$NetBSD: vfs_bio.c,v 1.306 2024/12/07 02:27:38 riastradh Exp $	*/
2 
3 /*-
4  * Copyright (c) 2007, 2008, 2009, 2019, 2020 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 1.306 2024/12/07 02:27:38 riastradh Exp $");
127 
128 #ifdef _KERNEL_OPT
129 #include "opt_biohist.h"
130 #include "opt_bufcache.h"
131 #include "opt_dtrace.h"
132 #endif
133 
134 #include <sys/param.h>
135 #include <sys/types.h>
136 
137 #include <sys/bitops.h>
138 #include <sys/buf.h>
139 #include <sys/conf.h>
140 #include <sys/cprng.h>
141 #include <sys/cpu.h>
142 #include <sys/fstrans.h>
143 #include <sys/intr.h>
144 #include <sys/kauth.h>
145 #include <sys/kernel.h>
146 #include <sys/mount.h>
147 #include <sys/proc.h>
148 #include <sys/resourcevar.h>
149 #include <sys/sdt.h>
150 #include <sys/sysctl.h>
151 #include <sys/systm.h>
152 #include <sys/vnode.h>
153 #include <sys/wapbl.h>
154 
155 #include <uvm/uvm.h>	/* extern struct uvm uvm */
156 
157 #include <miscfs/specfs/specdev.h>
158 
159 SDT_PROVIDER_DEFINE(io);
160 
161 SDT_PROBE_DEFINE4(io, kernel, , bbusy__start,
162     "struct buf *"/*bp*/,
163     "bool"/*intr*/, "int"/*timo*/, "kmutex_t *"/*interlock*/);
164 SDT_PROBE_DEFINE5(io, kernel, , bbusy__done,
165     "struct buf *"/*bp*/,
166     "bool"/*intr*/,
167     "int"/*timo*/,
168     "kmutex_t *"/*interlock*/,
169     "int"/*error*/);
170 SDT_PROBE_DEFINE0(io, kernel, , getnewbuf__start);
171 SDT_PROBE_DEFINE1(io, kernel, , getnewbuf__done,  "struct buf *"/*bp*/);
172 SDT_PROBE_DEFINE3(io, kernel, , getblk__start,
173     "struct vnode *"/*vp*/, "daddr_t"/*blkno*/, "int"/*size*/);
174 SDT_PROBE_DEFINE4(io, kernel, , getblk__done,
175     "struct vnode *"/*vp*/, "daddr_t"/*blkno*/, "int"/*size*/,
176     "struct buf *"/*bp*/);
177 SDT_PROBE_DEFINE2(io, kernel, , brelse, "struct buf *"/*bp*/, "int"/*set*/);
178 SDT_PROBE_DEFINE1(io, kernel, , wait__start, "struct buf *"/*bp*/);
179 SDT_PROBE_DEFINE1(io, kernel, , wait__done, "struct buf *"/*bp*/);
180 
181 #ifndef	BUFPAGES
182 # define BUFPAGES 0
183 #endif
184 
185 #ifdef BUFCACHE
186 # if (BUFCACHE < 5) || (BUFCACHE > 95)
187 #  error BUFCACHE is not between 5 and 95
188 # endif
189 #else
190 # define BUFCACHE 15
191 #endif
192 
193 u_int	nbuf;			/* desired number of buffer headers */
194 u_int	bufpages = BUFPAGES;	/* optional hardwired count */
195 u_int	bufcache = BUFCACHE;	/* max % of RAM to use for buffer cache */
196 
197 /*
198  * Definitions for the buffer free lists.
199  */
200 #define	BQUEUES		3		/* number of free buffer queues */
201 
202 #define	BQ_LOCKED	0		/* super-blocks &c */
203 #define	BQ_LRU		1		/* lru, useful buffers */
204 #define	BQ_AGE		2		/* rubbish */
205 
206 struct bqueue {
207 	TAILQ_HEAD(, buf) bq_queue;
208 	uint64_t bq_bytes;
209 	buf_t *bq_marker;
210 };
211 static struct bqueue bufqueues[BQUEUES] __cacheline_aligned;
212 
213 /* Function prototypes */
214 static void buf_setwm(void);
215 static int buf_trim(void);
216 static void *bufpool_page_alloc(struct pool *, int);
217 static void bufpool_page_free(struct pool *, void *);
218 static buf_t *bio_doread(struct vnode *, daddr_t, int, int);
219 static buf_t *getnewbuf(int, int, int);
220 static int buf_lotsfree(void);
221 static int buf_canrelease(void);
222 static u_long buf_mempoolidx(u_long);
223 static u_long buf_roundsize(u_long);
224 static void *buf_alloc(size_t);
225 static void buf_mrelease(void *, size_t);
226 static void binsheadfree(buf_t *, struct bqueue *);
227 static void binstailfree(buf_t *, struct bqueue *);
228 #ifdef DEBUG
229 static int checkfreelist(buf_t *, struct bqueue *, int);
230 #endif
231 static void biointr(void *);
232 static void biodone2(buf_t *);
233 static void sysctl_kern_buf_setup(void);
234 static void sysctl_vm_buf_setup(void);
235 
236 /* Initialization for biohist */
237 
238 #include <sys/biohist.h>
239 
240 BIOHIST_DEFINE(biohist);
241 
242 void
243 biohist_init(void)
244 {
245 
246 	BIOHIST_INIT(biohist, BIOHIST_SIZE);
247 }
248 
249 /*
250  * Definitions for the buffer hash lists.
251  */
252 #define	BUFHASH(dvp, lbn)	\
253 	(&bufhashtbl[(((long)(dvp) >> 8) + (int)(lbn)) & bufhash])
254 LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
255 u_long	bufhash;
256 
257 static int     bufhash_stats(struct hashstat_sysctl *, bool);
258 
259 static kcondvar_t needbuffer_cv;
260 
261 /*
262  * Buffer queue lock.
263  */
264 kmutex_t bufcache_lock __cacheline_aligned;
265 kmutex_t buffer_lock __cacheline_aligned;
266 
267 /* Software ISR for completed transfers. */
268 static void *biodone_sih;
269 
270 /* Buffer pool for I/O buffers. */
271 static pool_cache_t buf_cache;
272 static pool_cache_t bufio_cache;
273 
274 #define MEMPOOL_INDEX_OFFSET (ilog2(DEV_BSIZE))	/* smallest pool is 512 bytes */
275 #define NMEMPOOLS (ilog2(MAXBSIZE) - MEMPOOL_INDEX_OFFSET + 1)
276 __CTASSERT((1 << (NMEMPOOLS + MEMPOOL_INDEX_OFFSET - 1)) == MAXBSIZE);
277 
278 /* Buffer memory pools */
279 static struct pool bmempools[NMEMPOOLS];
280 
281 static struct vm_map *buf_map;
282 
283 /*
284  * Buffer memory pool allocator.
285  */
286 static void *
287 bufpool_page_alloc(struct pool *pp, int flags)
288 {
289 
290 	return (void *)uvm_km_alloc(buf_map,
291 	    MAXBSIZE, MAXBSIZE,
292 	    ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT|UVM_KMF_TRYLOCK)
293 	    | UVM_KMF_WIRED);
294 }
295 
296 static void
297 bufpool_page_free(struct pool *pp, void *v)
298 {
299 
300 	uvm_km_free(buf_map, (vaddr_t)v, MAXBSIZE, UVM_KMF_WIRED);
301 }
302 
303 static struct pool_allocator bufmempool_allocator = {
304 	.pa_alloc = bufpool_page_alloc,
305 	.pa_free = bufpool_page_free,
306 	.pa_pagesz = MAXBSIZE,
307 };
308 
309 /* Buffer memory management variables */
310 u_long bufmem_valimit;
311 u_long bufmem_hiwater;
312 u_long bufmem_lowater;
313 u_long bufmem;
314 
315 /*
316  * MD code can call this to set a hard limit on the amount
317  * of virtual memory used by the buffer cache.
318  */
319 int
320 buf_setvalimit(vsize_t sz)
321 {
322 
323 	/* We need to accommodate at least NMEMPOOLS of MAXBSIZE each */
324 	if (sz < NMEMPOOLS * MAXBSIZE)
325 		return SET_ERROR(EINVAL);
326 
327 	bufmem_valimit = sz;
328 	return 0;
329 }
330 
331 static void
332 buf_setwm(void)
333 {
334 
335 	bufmem_hiwater = buf_memcalc();
336 	/* lowater is approx. 2% of memory (with bufcache = 15) */
337 #define	BUFMEM_WMSHIFT	3
338 #define	BUFMEM_HIWMMIN	(64 * 1024 << BUFMEM_WMSHIFT)
339 	if (bufmem_hiwater < BUFMEM_HIWMMIN)
340 		/* Ensure a reasonable minimum value */
341 		bufmem_hiwater = BUFMEM_HIWMMIN;
342 	bufmem_lowater = bufmem_hiwater >> BUFMEM_WMSHIFT;
343 }
344 
345 #ifdef DEBUG
346 int debug_verify_freelist = 0;
347 static int
348 checkfreelist(buf_t *bp, struct bqueue *dp, int ison)
349 {
350 	buf_t *b;
351 
352 	if (!debug_verify_freelist)
353 		return 1;
354 
355 	TAILQ_FOREACH(b, &dp->bq_queue, b_freelist) {
356 		if (b == bp)
357 			return ison ? 1 : 0;
358 	}
359 
360 	return ison ? 0 : 1;
361 }
362 #endif
363 
364 /*
365  * Insq/Remq for the buffer hash lists.
366  * Call with buffer queue locked.
367  */
368 static void
369 binsheadfree(buf_t *bp, struct bqueue *dp)
370 {
371 
372 	KASSERT(mutex_owned(&bufcache_lock));
373 	KASSERT(bp->b_freelistindex == -1);
374 	TAILQ_INSERT_HEAD(&dp->bq_queue, bp, b_freelist);
375 	dp->bq_bytes += bp->b_bufsize;
376 	bp->b_freelistindex = dp - bufqueues;
377 }
378 
379 static void
380 binstailfree(buf_t *bp, struct bqueue *dp)
381 {
382 
383 	KASSERT(mutex_owned(&bufcache_lock));
384 	KASSERTMSG(bp->b_freelistindex == -1, "double free of buffer? "
385 	    "bp=%p, b_freelistindex=%d\n", bp, bp->b_freelistindex);
386 	TAILQ_INSERT_TAIL(&dp->bq_queue, bp, b_freelist);
387 	dp->bq_bytes += bp->b_bufsize;
388 	bp->b_freelistindex = dp - bufqueues;
389 }
390 
391 void
392 bremfree(buf_t *bp)
393 {
394 	struct bqueue *dp;
395 	int bqidx = bp->b_freelistindex;
396 
397 	KASSERT(mutex_owned(&bufcache_lock));
398 
399 	KASSERT(bqidx != -1);
400 	dp = &bufqueues[bqidx];
401 	KDASSERT(checkfreelist(bp, dp, 1));
402 	KASSERT(dp->bq_bytes >= bp->b_bufsize);
403 	TAILQ_REMOVE(&dp->bq_queue, bp, b_freelist);
404 	dp->bq_bytes -= bp->b_bufsize;
405 
406 	/* For the sysctl helper. */
407 	if (bp == dp->bq_marker)
408 		dp->bq_marker = NULL;
409 
410 #if defined(DIAGNOSTIC)
411 	bp->b_freelistindex = -1;
412 #endif /* defined(DIAGNOSTIC) */
413 }
414 
415 /*
416  * note that for some ports this is used by pmap bootstrap code to
417  * determine kva size.
418  */
419 u_long
420 buf_memcalc(void)
421 {
422 	u_long n;
423 	vsize_t mapsz = 0;
424 
425 	/*
426 	 * Determine the upper bound of memory to use for buffers.
427 	 *
428 	 *	- If bufpages is specified, use that as the number
429 	 *	  pages.
430 	 *
431 	 *	- Otherwise, use bufcache as the percentage of
432 	 *	  physical memory.
433 	 */
434 	if (bufpages != 0) {
435 		n = bufpages;
436 	} else {
437 		if (bufcache < 5) {
438 			printf("forcing bufcache %d -> 5", bufcache);
439 			bufcache = 5;
440 		}
441 		if (bufcache > 95) {
442 			printf("forcing bufcache %d -> 95", bufcache);
443 			bufcache = 95;
444 		}
445 		if (buf_map != NULL)
446 			mapsz = vm_map_max(buf_map) - vm_map_min(buf_map);
447 		n = calc_cache_size(mapsz, bufcache,
448 		    (buf_map != kernel_map) ? 100 : BUFCACHE_VA_MAXPCT)
449 		    / PAGE_SIZE;
450 	}
451 
452 	n <<= PAGE_SHIFT;
453 	if (bufmem_valimit != 0 && n > bufmem_valimit)
454 		n = bufmem_valimit;
455 
456 	return n;
457 }
458 
459 /*
460  * Initialize buffers and hash links for buffers.
461  */
462 void
463 bufinit(void)
464 {
465 	struct bqueue *dp;
466 	int use_std;
467 	u_int i;
468 
469 	biodone_vfs = biodone;
470 
471 	mutex_init(&bufcache_lock, MUTEX_DEFAULT, IPL_NONE);
472 	mutex_init(&buffer_lock, MUTEX_DEFAULT, IPL_NONE);
473 	cv_init(&needbuffer_cv, "needbuf");
474 
475 	if (bufmem_valimit != 0) {
476 		vaddr_t minaddr = 0, maxaddr;
477 		buf_map = uvm_km_suballoc(kernel_map, &minaddr, &maxaddr,
478 		    bufmem_valimit, 0, false, 0);
479 		if (buf_map == NULL)
480 			panic("bufinit: cannot allocate submap");
481 	} else
482 		buf_map = kernel_map;
483 
484 	/*
485 	 * Initialize buffer cache memory parameters.
486 	 */
487 	bufmem = 0;
488 	buf_setwm();
489 
490 	/* On "small" machines use small pool page sizes where possible */
491 	use_std = (physmem < atop(16*1024*1024));
492 
493 	/*
494 	 * Also use them on systems that can map the pool pages using
495 	 * a direct-mapped segment.
496 	 */
497 #ifdef PMAP_MAP_POOLPAGE
498 	use_std = 1;
499 #endif
500 
501 	buf_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
502 	    "bufpl", NULL, IPL_SOFTBIO, NULL, NULL, NULL);
503 	bufio_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
504 	    "biopl", NULL, IPL_BIO, NULL, NULL, NULL);
505 
506 	for (i = 0; i < NMEMPOOLS; i++) {
507 		struct pool_allocator *pa;
508 		struct pool *pp = &bmempools[i];
509 		u_int size = 1 << (i + MEMPOOL_INDEX_OFFSET);
510 		char *name = kmem_alloc(8, KM_SLEEP); /* XXX: never freed */
511 
512 		if (__predict_false(size >= 1048576))
513 			(void)snprintf(name, 8, "buf%um", size / 1048576);
514 		else if (__predict_true(size >= 1024))
515 			(void)snprintf(name, 8, "buf%uk", size / 1024);
516 		else
517 			(void)snprintf(name, 8, "buf%ub", size);
518 		pa = (size <= PAGE_SIZE && use_std)
519 		    ? &pool_allocator_nointr
520 		    : &bufmempool_allocator;
521 		pool_init(pp, size, DEV_BSIZE, 0, 0, name, pa, IPL_NONE);
522 		pool_setlowat(pp, 1);
523 		pool_sethiwat(pp, 1);
524 	}
525 
526 	/* Initialize the buffer queues */
527 	for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) {
528 		TAILQ_INIT(&dp->bq_queue);
529 		dp->bq_bytes = 0;
530 	}
531 
532 	/*
533 	 * Estimate hash table size based on the amount of memory we
534 	 * intend to use for the buffer cache. The average buffer
535 	 * size is dependent on our clients (i.e. filesystems).
536 	 *
537 	 * For now, use an empirical 3K per buffer.
538 	 */
539 	nbuf = (bufmem_hiwater / 1024) / 3;
540 	bufhashtbl = hashinit(nbuf, HASH_LIST, true, &bufhash);
541 
542 	sysctl_kern_buf_setup();
543 	sysctl_vm_buf_setup();
544 	hashstat_register("bufhash", bufhash_stats);
545 }
546 
547 void
548 bufinit2(void)
549 {
550 
551 	biodone_sih = softint_establish(SOFTINT_BIO | SOFTINT_MPSAFE, biointr,
552 	    NULL);
553 	if (biodone_sih == NULL)
554 		panic("bufinit2: can't establish soft interrupt");
555 }
556 
557 static int
558 buf_lotsfree(void)
559 {
560 	u_long guess;
561 
562 	/* Always allocate if less than the low water mark. */
563 	if (bufmem < bufmem_lowater)
564 		return 1;
565 
566 	/* Never allocate if greater than the high water mark. */
567 	if (bufmem > bufmem_hiwater)
568 		return 0;
569 
570 	/* If there's anything on the AGE list, it should be eaten. */
571 	if (TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue) != NULL)
572 		return 0;
573 
574 	/*
575 	 * The probabily of getting a new allocation is inversely
576 	 * proportional  to the current size of the cache above
577 	 * the low water mark.  Divide the total first to avoid overflows
578 	 * in the product.
579 	 */
580 	guess = cprng_fast32() % 16;
581 
582 	if ((bufmem_hiwater - bufmem_lowater) / 16 * guess >=
583 	    (bufmem - bufmem_lowater))
584 		return 1;
585 
586 	/* Otherwise don't allocate. */
587 	return 0;
588 }
589 
590 /*
591  * Return estimate of bytes we think need to be
592  * released to help resolve low memory conditions.
593  *
594  * => called with bufcache_lock held.
595  */
596 static int
597 buf_canrelease(void)
598 {
599 	int pagedemand, ninvalid = 0;
600 
601 	KASSERT(mutex_owned(&bufcache_lock));
602 
603 	if (bufmem < bufmem_lowater)
604 		return 0;
605 
606 	if (bufmem > bufmem_hiwater)
607 		return bufmem - bufmem_hiwater;
608 
609 	ninvalid += bufqueues[BQ_AGE].bq_bytes;
610 
611 	pagedemand = uvmexp.freetarg - uvm_availmem(false);
612 	if (pagedemand < 0)
613 		return ninvalid;
614 	return MAX(ninvalid, MIN(2 * MAXBSIZE,
615 	    MIN((bufmem - bufmem_lowater) / 16, pagedemand * PAGE_SIZE)));
616 }
617 
618 /*
619  * Buffer memory allocation helper functions
620  */
621 static u_long
622 buf_mempoolidx(u_long size)
623 {
624 	u_int n = 0;
625 
626 	size -= 1;
627 	size >>= MEMPOOL_INDEX_OFFSET;
628 	while (size) {
629 		size >>= 1;
630 		n += 1;
631 	}
632 	if (n >= NMEMPOOLS)
633 		panic("buf mem pool index %d", n);
634 	return n;
635 }
636 
637 static u_long
638 buf_roundsize(u_long size)
639 {
640 
641 	/* Round up to nearest power of 2 */
642 	return (1 << (buf_mempoolidx(size) + MEMPOOL_INDEX_OFFSET));
643 }
644 
645 static void *
646 buf_alloc(size_t size)
647 {
648 	u_int n = buf_mempoolidx(size);
649 	void *addr;
650 
651 	while (1) {
652 		addr = pool_get(&bmempools[n], PR_NOWAIT);
653 		if (addr != NULL)
654 			break;
655 
656 		/* No memory, see if we can free some. If so, try again */
657 		mutex_enter(&bufcache_lock);
658 		if (buf_drain(1) > 0) {
659 			mutex_exit(&bufcache_lock);
660 			continue;
661 		}
662 
663 		if (curlwp == uvm.pagedaemon_lwp) {
664 			mutex_exit(&bufcache_lock);
665 			return NULL;
666 		}
667 
668 		/* Wait for buffers to arrive on the LRU queue */
669 		cv_timedwait(&needbuffer_cv, &bufcache_lock, hz / 4);
670 		mutex_exit(&bufcache_lock);
671 	}
672 
673 	return addr;
674 }
675 
676 static void
677 buf_mrelease(void *addr, size_t size)
678 {
679 
680 	pool_put(&bmempools[buf_mempoolidx(size)], addr);
681 }
682 
683 /*
684  * bread()/breadn() helper.
685  */
686 static buf_t *
687 bio_doread(struct vnode *vp, daddr_t blkno, int size, int async)
688 {
689 	buf_t *bp;
690 	struct mount *mp;
691 
692 	bp = getblk(vp, blkno, size, 0, 0);
693 
694 	/*
695 	 * getblk() may return NULL if we are the pagedaemon.
696 	 */
697 	if (bp == NULL) {
698 		KASSERT(curlwp == uvm.pagedaemon_lwp);
699 		return NULL;
700 	}
701 
702 	/*
703 	 * If buffer does not have data valid, start a read.
704 	 * Note that if buffer is BC_INVAL, getblk() won't return it.
705 	 * Therefore, it's valid if its I/O has completed or been delayed.
706 	 */
707 	if (!ISSET(bp->b_oflags, (BO_DONE | BO_DELWRI))) {
708 		/* Start I/O for the buffer. */
709 		SET(bp->b_flags, B_READ | async);
710 		if (async)
711 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
712 		else
713 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
714 		VOP_STRATEGY(vp, bp);
715 
716 		/* Pay for the read. */
717 		curlwp->l_ru.ru_inblock++;
718 	} else if (async)
719 		brelse(bp, 0);
720 
721 	if (vp->v_type == VBLK)
722 		mp = spec_node_getmountedfs(vp);
723 	else
724 		mp = vp->v_mount;
725 
726 	/*
727 	 * Collect statistics on synchronous and asynchronous reads.
728 	 * Reads from block devices are charged to their associated
729 	 * filesystem (if any).
730 	 */
731 	if (mp != NULL) {
732 		if (async == 0)
733 			mp->mnt_stat.f_syncreads++;
734 		else
735 			mp->mnt_stat.f_asyncreads++;
736 	}
737 
738 	return bp;
739 }
740 
741 /*
742  * Read a disk block.
743  * This algorithm described in Bach (p.54).
744  */
745 int
746 bread(struct vnode *vp, daddr_t blkno, int size, int flags, buf_t **bpp)
747 {
748 	buf_t *bp;
749 	int error;
750 
751 	BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
752 
753 	/* Get buffer for block. */
754 	bp = *bpp = bio_doread(vp, blkno, size, 0);
755 	if (bp == NULL)
756 		return SET_ERROR(ENOMEM);
757 
758 	/* Wait for the read to complete, and return result. */
759 	error = biowait(bp);
760 	if (error == 0 && (flags & B_MODIFY) != 0)
761 		error = fscow_run(bp, true);
762 	if (error) {
763 		brelse(bp, 0);
764 		*bpp = NULL;
765 	}
766 
767 	return error;
768 }
769 
770 /*
771  * Read-ahead multiple disk blocks. The first is sync, the rest async.
772  * Trivial modification to the breada algorithm presented in Bach (p.55).
773  */
774 int
775 breadn(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablks,
776     int *rasizes, int nrablks, int flags, buf_t **bpp)
777 {
778 	buf_t *bp;
779 	int error, i;
780 
781 	BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
782 
783 	bp = *bpp = bio_doread(vp, blkno, size, 0);
784 	if (bp == NULL)
785 		return SET_ERROR(ENOMEM);
786 
787 	/*
788 	 * For each of the read-ahead blocks, start a read, if necessary.
789 	 */
790 	mutex_enter(&bufcache_lock);
791 	for (i = 0; i < nrablks; i++) {
792 		/* If it's in the cache, just go on to next one. */
793 		if (incore(vp, rablks[i]))
794 			continue;
795 
796 		/* Get a buffer for the read-ahead block */
797 		mutex_exit(&bufcache_lock);
798 		(void) bio_doread(vp, rablks[i], rasizes[i], B_ASYNC);
799 		mutex_enter(&bufcache_lock);
800 	}
801 	mutex_exit(&bufcache_lock);
802 
803 	/* Otherwise, we had to start a read for it; wait until it's valid. */
804 	error = biowait(bp);
805 	if (error == 0 && (flags & B_MODIFY) != 0)
806 		error = fscow_run(bp, true);
807 	if (error) {
808 		brelse(bp, 0);
809 		*bpp = NULL;
810 	}
811 
812 	return error;
813 }
814 
815 /*
816  * Block write.  Described in Bach (p.56)
817  */
818 int
819 bwrite(buf_t *bp)
820 {
821 	int rv, sync, wasdelayed;
822 	struct vnode *vp;
823 	struct mount *mp;
824 
825 	BIOHIST_FUNC(__func__); BIOHIST_CALLARGS(biohist, "bp=%#jx",
826 	    (uintptr_t)bp, 0, 0, 0);
827 
828 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
829 	KASSERT(!cv_has_waiters(&bp->b_done));
830 
831 	vp = bp->b_vp;
832 
833 	/*
834 	 * dholland 20160728 AFAICT vp==NULL must be impossible as it
835 	 * will crash upon reaching VOP_STRATEGY below... see further
836 	 * analysis on tech-kern.
837 	 */
838 	KASSERTMSG(vp != NULL, "bwrite given buffer with null vnode");
839 
840 	if (vp != NULL) {
841 		KASSERT(bp->b_objlock == vp->v_interlock);
842 		if (vp->v_type == VBLK)
843 			mp = spec_node_getmountedfs(vp);
844 		else
845 			mp = vp->v_mount;
846 	} else {
847 		mp = NULL;
848 	}
849 
850 	if (mp && mp->mnt_wapbl) {
851 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
852 			bdwrite(bp);
853 			return 0;
854 		}
855 	}
856 
857 	/*
858 	 * Remember buffer type, to switch on it later.  If the write was
859 	 * synchronous, but the file system was mounted with MNT_ASYNC,
860 	 * convert it to a delayed write.
861 	 * XXX note that this relies on delayed tape writes being converted
862 	 * to async, not sync writes (which is safe, but ugly).
863 	 */
864 	sync = !ISSET(bp->b_flags, B_ASYNC);
865 	if (sync && mp != NULL && ISSET(mp->mnt_flag, MNT_ASYNC)) {
866 		bdwrite(bp);
867 		return 0;
868 	}
869 
870 	/*
871 	 * Collect statistics on synchronous and asynchronous writes.
872 	 * Writes to block devices are charged to their associated
873 	 * filesystem (if any).
874 	 */
875 	if (mp != NULL) {
876 		if (sync)
877 			mp->mnt_stat.f_syncwrites++;
878 		else
879 			mp->mnt_stat.f_asyncwrites++;
880 	}
881 
882 	/*
883 	 * Pay for the I/O operation and make sure the buf is on the correct
884 	 * vnode queue.
885 	 */
886 	bp->b_error = 0;
887 	wasdelayed = ISSET(bp->b_oflags, BO_DELWRI);
888 	CLR(bp->b_flags, B_READ);
889 	if (wasdelayed) {
890 		mutex_enter(&bufcache_lock);
891 		mutex_enter(bp->b_objlock);
892 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
893 		reassignbuf(bp, bp->b_vp);
894 		/* Wake anyone trying to busy the buffer via vnode's lists. */
895 		cv_broadcast(&bp->b_busy);
896 		mutex_exit(&bufcache_lock);
897 	} else {
898 		curlwp->l_ru.ru_oublock++;
899 		mutex_enter(bp->b_objlock);
900 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
901 	}
902 	if (vp != NULL)
903 		vp->v_numoutput++;
904 	mutex_exit(bp->b_objlock);
905 
906 	/* Initiate disk write. */
907 	if (sync)
908 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
909 	else
910 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
911 
912 	VOP_STRATEGY(vp, bp);
913 
914 	if (sync) {
915 		/* If I/O was synchronous, wait for it to complete. */
916 		rv = biowait(bp);
917 
918 		/* Release the buffer. */
919 		brelse(bp, 0);
920 
921 		return rv;
922 	} else {
923 		return 0;
924 	}
925 }
926 
927 int
928 vn_bwrite(void *v)
929 {
930 	struct vop_bwrite_args *ap = v;
931 
932 	return bwrite(ap->a_bp);
933 }
934 
935 /*
936  * Delayed write.
937  *
938  * The buffer is marked dirty, but is not queued for I/O.
939  * This routine should be used when the buffer is expected
940  * to be modified again soon, typically a small write that
941  * partially fills a buffer.
942  *
943  * NB: magnetic tapes cannot be delayed; they must be
944  * written in the order that the writes are requested.
945  *
946  * Described in Leffler, et al. (pp. 208-213).
947  */
948 void
949 bdwrite(buf_t *bp)
950 {
951 
952 	BIOHIST_FUNC(__func__); BIOHIST_CALLARGS(biohist, "bp=%#jx",
953 	    (uintptr_t)bp, 0, 0, 0);
954 
955 	KASSERT(bp->b_vp == NULL || bp->b_vp->v_tag != VT_UFS ||
956 	    bp->b_vp->v_type == VBLK || ISSET(bp->b_flags, B_COWDONE));
957 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
958 	KASSERT(!cv_has_waiters(&bp->b_done));
959 
960 	/* If this is a tape block, write the block now. */
961 	if (bdev_type(bp->b_dev) == D_TAPE) {
962 		bawrite(bp);
963 		return;
964 	}
965 
966 	if (wapbl_vphaswapbl(bp->b_vp)) {
967 		struct mount *mp = wapbl_vptomp(bp->b_vp);
968 
969 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
970 			WAPBL_ADD_BUF(mp, bp);
971 		}
972 	}
973 
974 	/*
975 	 * If the block hasn't been seen before:
976 	 *	(1) Mark it as having been seen,
977 	 *	(2) Charge for the write,
978 	 *	(3) Make sure it's on its vnode's correct block list.
979 	 */
980 	KASSERT(bp->b_vp == NULL || bp->b_objlock == bp->b_vp->v_interlock);
981 
982 	if (!ISSET(bp->b_oflags, BO_DELWRI)) {
983 		mutex_enter(&bufcache_lock);
984 		mutex_enter(bp->b_objlock);
985 		SET(bp->b_oflags, BO_DELWRI);
986 		curlwp->l_ru.ru_oublock++;
987 		reassignbuf(bp, bp->b_vp);
988 		/* Wake anyone trying to busy the buffer via vnode's lists. */
989 		cv_broadcast(&bp->b_busy);
990 		mutex_exit(&bufcache_lock);
991 	} else {
992 		mutex_enter(bp->b_objlock);
993 	}
994 	/* Otherwise, the "write" is done, so mark and release the buffer. */
995 	CLR(bp->b_oflags, BO_DONE);
996 	mutex_exit(bp->b_objlock);
997 
998 	brelse(bp, 0);
999 }
1000 
1001 /*
1002  * Asynchronous block write; just an asynchronous bwrite().
1003  */
1004 void
1005 bawrite(buf_t *bp)
1006 {
1007 
1008 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1009 	KASSERT(bp->b_vp != NULL);
1010 
1011 	SET(bp->b_flags, B_ASYNC);
1012 	VOP_BWRITE(bp->b_vp, bp);
1013 }
1014 
1015 /*
1016  * Release a buffer on to the free lists.
1017  * Described in Bach (p. 46).
1018  */
1019 void
1020 brelsel(buf_t *bp, int set)
1021 {
1022 	struct bqueue *bufq;
1023 	struct vnode *vp;
1024 
1025 	SDT_PROBE2(io, kernel, , brelse,  bp, set);
1026 
1027 	KASSERT(bp != NULL);
1028 	KASSERT(mutex_owned(&bufcache_lock));
1029 	KASSERT(!cv_has_waiters(&bp->b_done));
1030 
1031 	SET(bp->b_cflags, set);
1032 
1033 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1034 	KASSERT(bp->b_iodone == NULL);
1035 
1036 	/* Wake up any processes waiting for any buffer to become free. */
1037 	cv_signal(&needbuffer_cv);
1038 
1039 	/* Wake up any proceeses waiting for _this_ buffer to become free */
1040 	if (ISSET(bp->b_cflags, BC_WANTED))
1041 		CLR(bp->b_cflags, BC_WANTED|BC_AGE);
1042 
1043 	/* If it's clean clear the copy-on-write flag. */
1044 	if (ISSET(bp->b_flags, B_COWDONE)) {
1045 		mutex_enter(bp->b_objlock);
1046 		if (!ISSET(bp->b_oflags, BO_DELWRI))
1047 			CLR(bp->b_flags, B_COWDONE);
1048 		mutex_exit(bp->b_objlock);
1049 	}
1050 
1051 	/*
1052 	 * Determine which queue the buffer should be on, then put it there.
1053 	 */
1054 
1055 	/* If it's locked, don't report an error; try again later. */
1056 	if (ISSET(bp->b_flags, B_LOCKED))
1057 		bp->b_error = 0;
1058 
1059 	/* If it's not cacheable, or an error, mark it invalid. */
1060 	if (ISSET(bp->b_cflags, BC_NOCACHE) || bp->b_error != 0)
1061 		SET(bp->b_cflags, BC_INVAL);
1062 
1063 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1064 		/*
1065 		 * This is a delayed write buffer that was just flushed to
1066 		 * disk.  It is still on the LRU queue.  If it's become
1067 		 * invalid, then we need to move it to a different queue;
1068 		 * otherwise leave it in its current position.
1069 		 */
1070 		CLR(bp->b_cflags, BC_VFLUSH);
1071 		if (!ISSET(bp->b_cflags, BC_INVAL|BC_AGE) &&
1072 		    !ISSET(bp->b_flags, B_LOCKED) && bp->b_error == 0) {
1073 			KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 1));
1074 			goto already_queued;
1075 		} else {
1076 			bremfree(bp);
1077 		}
1078 	}
1079 
1080 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_AGE], 0));
1081 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 0));
1082 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LOCKED], 0));
1083 
1084 	if ((bp->b_bufsize <= 0) || ISSET(bp->b_cflags, BC_INVAL)) {
1085 		/*
1086 		 * If it's invalid or empty, dissociate it from its vnode
1087 		 * and put on the head of the appropriate queue.
1088 		 */
1089 		if (ISSET(bp->b_flags, B_LOCKED)) {
1090 			if (wapbl_vphaswapbl(vp = bp->b_vp)) {
1091 				struct mount *mp = wapbl_vptomp(vp);
1092 
1093 				KASSERT(bp->b_iodone !=
1094 				    mp->mnt_wapbl_op->wo_wapbl_biodone);
1095 				WAPBL_REMOVE_BUF(mp, bp);
1096 			}
1097 		}
1098 
1099 		mutex_enter(bp->b_objlock);
1100 		CLR(bp->b_oflags, BO_DONE|BO_DELWRI);
1101 		if ((vp = bp->b_vp) != NULL) {
1102 			KASSERT(bp->b_objlock == vp->v_interlock);
1103 			reassignbuf(bp, bp->b_vp);
1104 			brelvp(bp);
1105 			mutex_exit(vp->v_interlock);
1106 		} else {
1107 			KASSERT(bp->b_objlock == &buffer_lock);
1108 			mutex_exit(bp->b_objlock);
1109 		}
1110 		/* We want to dispose of the buffer, so wake everybody. */
1111 		cv_broadcast(&bp->b_busy);
1112 		if (bp->b_bufsize <= 0)
1113 			/* no data */
1114 			goto already_queued;
1115 		else
1116 			/* invalid data */
1117 			bufq = &bufqueues[BQ_AGE];
1118 		binsheadfree(bp, bufq);
1119 	} else  {
1120 		/*
1121 		 * It has valid data.  Put it on the end of the appropriate
1122 		 * queue, so that it'll stick around for as long as possible.
1123 		 * If buf is AGE, but has dependencies, must put it on last
1124 		 * bufqueue to be scanned, ie LRU. This protects against the
1125 		 * livelock where BQ_AGE only has buffers with dependencies,
1126 		 * and we thus never get to the dependent buffers in BQ_LRU.
1127 		 */
1128 		if (ISSET(bp->b_flags, B_LOCKED)) {
1129 			/* locked in core */
1130 			bufq = &bufqueues[BQ_LOCKED];
1131 		} else if (!ISSET(bp->b_cflags, BC_AGE)) {
1132 			/* valid data */
1133 			bufq = &bufqueues[BQ_LRU];
1134 		} else {
1135 			/* stale but valid data */
1136 			bufq = &bufqueues[BQ_AGE];
1137 		}
1138 		binstailfree(bp, bufq);
1139 	}
1140 already_queued:
1141 	/* Unlock the buffer. */
1142 	CLR(bp->b_cflags, BC_AGE|BC_BUSY|BC_NOCACHE);
1143 	CLR(bp->b_flags, B_ASYNC);
1144 
1145 	/*
1146 	 * Wake only the highest priority waiter on the lock, in order to
1147 	 * prevent a thundering herd: many LWPs simultaneously awakening and
1148 	 * competing for the buffer's lock.  Testing in 2019 revealed this
1149 	 * to reduce contention on bufcache_lock tenfold during a kernel
1150 	 * compile.  Here and elsewhere, when the buffer is changing
1151 	 * identity, being disposed of, or moving from one list to another,
1152 	 * we wake all lock requestors.
1153 	 */
1154 	if (bp->b_bufsize <= 0) {
1155 		cv_broadcast(&bp->b_busy);
1156 		buf_destroy(bp);
1157 #ifdef DEBUG
1158 		memset((char *)bp, 0, sizeof(*bp));
1159 #endif
1160 		pool_cache_put(buf_cache, bp);
1161 	} else
1162 		cv_signal(&bp->b_busy);
1163 }
1164 
1165 void
1166 brelse(buf_t *bp, int set)
1167 {
1168 
1169 	mutex_enter(&bufcache_lock);
1170 	brelsel(bp, set);
1171 	mutex_exit(&bufcache_lock);
1172 }
1173 
1174 /*
1175  * Determine if a block is in the cache.
1176  * Just look on what would be its hash chain.  If it's there, return
1177  * a pointer to it, unless it's marked invalid.  If it's marked invalid,
1178  * we normally don't return the buffer, unless the caller explicitly
1179  * wants us to.
1180  */
1181 buf_t *
1182 incore(struct vnode *vp, daddr_t blkno)
1183 {
1184 	buf_t *bp;
1185 
1186 	KASSERT(mutex_owned(&bufcache_lock));
1187 
1188 	/* Search hash chain */
1189 	LIST_FOREACH(bp, BUFHASH(vp, blkno), b_hash) {
1190 		if (bp->b_lblkno == blkno && bp->b_vp == vp &&
1191 		    !ISSET(bp->b_cflags, BC_INVAL)) {
1192 			KASSERT(bp->b_objlock == vp->v_interlock);
1193 			return (bp);
1194 		}
1195 	}
1196 
1197 	return NULL;
1198 }
1199 
1200 /*
1201  * Get a block of requested size that is associated with
1202  * a given vnode and block offset. If it is found in the
1203  * block cache, mark it as having been found, make it busy
1204  * and return it. Otherwise, return an empty block of the
1205  * correct size. It is up to the caller to insure that the
1206  * cached blocks be of the correct size.
1207  */
1208 buf_t *
1209 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo)
1210 {
1211 	int err, preserve;
1212 	buf_t *bp;
1213 
1214 	mutex_enter(&bufcache_lock);
1215 	SDT_PROBE3(io, kernel, , getblk__start,  vp, blkno, size);
1216 loop:
1217 	bp = incore(vp, blkno);
1218 	if (bp != NULL) {
1219 		err = bbusy(bp, ((slpflag & PCATCH) != 0), slptimeo, NULL);
1220 		if (err != 0) {
1221 			if (err == EPASSTHROUGH)
1222 				goto loop;
1223 			mutex_exit(&bufcache_lock);
1224 			SDT_PROBE4(io, kernel, , getblk__done,
1225 			    vp, blkno, size, NULL);
1226 			return NULL;
1227 		}
1228 		KASSERT(!cv_has_waiters(&bp->b_done));
1229 #ifdef DIAGNOSTIC
1230 		if (ISSET(bp->b_oflags, BO_DONE|BO_DELWRI) &&
1231 		    bp->b_bcount < size && vp->v_type != VBLK)
1232 			panic("getblk: block size invariant failed");
1233 #endif
1234 		bremfree(bp);
1235 		preserve = 1;
1236 	} else {
1237 		if ((bp = getnewbuf(slpflag, slptimeo, 0)) == NULL)
1238 			goto loop;
1239 
1240 		if (incore(vp, blkno) != NULL) {
1241 			/* The block has come into memory in the meantime. */
1242 			brelsel(bp, 0);
1243 			goto loop;
1244 		}
1245 
1246 		LIST_INSERT_HEAD(BUFHASH(vp, blkno), bp, b_hash);
1247 		bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno;
1248 		mutex_enter(vp->v_interlock);
1249 		bgetvp(vp, bp);
1250 		mutex_exit(vp->v_interlock);
1251 		preserve = 0;
1252 	}
1253 	mutex_exit(&bufcache_lock);
1254 
1255 	/*
1256 	 * LFS can't track total size of B_LOCKED buffer (locked_queue_bytes)
1257 	 * if we re-size buffers here.
1258 	 */
1259 	if (ISSET(bp->b_flags, B_LOCKED)) {
1260 		KASSERT(bp->b_bufsize >= size);
1261 	} else {
1262 		if (allocbuf(bp, size, preserve)) {
1263 			mutex_enter(&bufcache_lock);
1264 			LIST_REMOVE(bp, b_hash);
1265 			brelsel(bp, BC_INVAL);
1266 			mutex_exit(&bufcache_lock);
1267 			SDT_PROBE4(io, kernel, , getblk__done,
1268 			    vp, blkno, size, NULL);
1269 			return NULL;
1270 		}
1271 	}
1272 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1273 	SDT_PROBE4(io, kernel, , getblk__done,  vp, blkno, size, bp);
1274 	return bp;
1275 }
1276 
1277 /*
1278  * Get an empty, disassociated buffer of given size.
1279  */
1280 buf_t *
1281 geteblk(int size)
1282 {
1283 	buf_t *bp;
1284 	int error __diagused;
1285 
1286 	mutex_enter(&bufcache_lock);
1287 	while ((bp = getnewbuf(0, 0, 0)) == NULL)
1288 		continue;
1289 
1290 	SET(bp->b_cflags, BC_INVAL);
1291 	LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1292 	mutex_exit(&bufcache_lock);
1293 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1294 	error = allocbuf(bp, size, 0);
1295 	KASSERT(error == 0);
1296 	return bp;
1297 }
1298 
1299 /*
1300  * Expand or contract the actual memory allocated to a buffer.
1301  *
1302  * If the buffer shrinks, data is lost, so it's up to the
1303  * caller to have written it out *first*; this routine will not
1304  * start a write.  If the buffer grows, it's the callers
1305  * responsibility to fill out the buffer's additional contents.
1306  */
1307 int
1308 allocbuf(buf_t *bp, int size, int preserve)
1309 {
1310 	void *addr;
1311 	vsize_t oldsize, desired_size;
1312 	int oldcount;
1313 	int delta;
1314 
1315 	desired_size = buf_roundsize(size);
1316 	if (desired_size > MAXBSIZE)
1317 		printf("allocbuf: buffer larger than MAXBSIZE requested");
1318 
1319 	oldcount = bp->b_bcount;
1320 
1321 	bp->b_bcount = size;
1322 
1323 	oldsize = bp->b_bufsize;
1324 	if (oldsize == desired_size) {
1325 		/*
1326 		 * Do not short cut the WAPBL resize, as the buffer length
1327 		 * could still have changed and this would corrupt the
1328 		 * tracking of the transaction length.
1329 		 */
1330 		goto out;
1331 	}
1332 
1333 	/*
1334 	 * If we want a buffer of a different size, re-allocate the
1335 	 * buffer's memory; copy old content only if needed.
1336 	 */
1337 	addr = buf_alloc(desired_size);
1338 	if (addr == NULL)
1339 		return SET_ERROR(ENOMEM);
1340 	if (preserve)
1341 		memcpy(addr, bp->b_data, MIN(oldsize,desired_size));
1342 	if (bp->b_data != NULL)
1343 		buf_mrelease(bp->b_data, oldsize);
1344 	bp->b_data = addr;
1345 	bp->b_bufsize = desired_size;
1346 
1347 	/*
1348 	 * Update overall buffer memory counter (protected by bufcache_lock)
1349 	 */
1350 	delta = (long)desired_size - (long)oldsize;
1351 
1352 	mutex_enter(&bufcache_lock);
1353 	if ((bufmem += delta) > bufmem_hiwater) {
1354 		/*
1355 		 * Need to trim overall memory usage.
1356 		 */
1357 		while (buf_canrelease()) {
1358 			if (preempt_needed()) {
1359 				mutex_exit(&bufcache_lock);
1360 				preempt();
1361 				mutex_enter(&bufcache_lock);
1362 			}
1363 			if (buf_trim() == 0)
1364 				break;
1365 		}
1366 	}
1367 	mutex_exit(&bufcache_lock);
1368 
1369 out:
1370 	if (wapbl_vphaswapbl(bp->b_vp)) {
1371 		WAPBL_RESIZE_BUF(wapbl_vptomp(bp->b_vp), bp,
1372 		    oldsize, oldcount);
1373 	}
1374 
1375 	return 0;
1376 }
1377 
1378 /*
1379  * Find a buffer which is available for use.
1380  * Select something from a free list.
1381  * Preference is to AGE list, then LRU list.
1382  *
1383  * Called with the buffer queues locked.
1384  * Return buffer locked.
1385  */
1386 static buf_t *
1387 getnewbuf(int slpflag, int slptimeo, int from_bufq)
1388 {
1389 	buf_t *bp;
1390 	struct vnode *vp;
1391 	struct mount *transmp = NULL;
1392 
1393 	SDT_PROBE0(io, kernel, , getnewbuf__start);
1394 
1395 start:
1396 	KASSERT(mutex_owned(&bufcache_lock));
1397 
1398 	/*
1399 	 * Get a new buffer from the pool.
1400 	 */
1401 	if (!from_bufq && buf_lotsfree()) {
1402 		mutex_exit(&bufcache_lock);
1403 		bp = pool_cache_get(buf_cache, PR_NOWAIT);
1404 		if (bp != NULL) {
1405 			memset((char *)bp, 0, sizeof(*bp));
1406 			buf_init(bp);
1407 			SET(bp->b_cflags, BC_BUSY);	/* mark buffer busy */
1408 			mutex_enter(&bufcache_lock);
1409 #if defined(DIAGNOSTIC)
1410 			bp->b_freelistindex = -1;
1411 #endif /* defined(DIAGNOSTIC) */
1412 			SDT_PROBE1(io, kernel, , getnewbuf__done,  bp);
1413 			return bp;
1414 		}
1415 		mutex_enter(&bufcache_lock);
1416 	}
1417 
1418 	KASSERT(mutex_owned(&bufcache_lock));
1419 	if ((bp = TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue)) != NULL) {
1420 		KASSERT(!ISSET(bp->b_oflags, BO_DELWRI));
1421 	} else {
1422 		TAILQ_FOREACH(bp, &bufqueues[BQ_LRU].bq_queue, b_freelist) {
1423 			if (ISSET(bp->b_cflags, BC_VFLUSH) ||
1424 			    !ISSET(bp->b_oflags, BO_DELWRI))
1425 				break;
1426 			if (fstrans_start_nowait(bp->b_vp->v_mount) == 0) {
1427 				KASSERT(transmp == NULL);
1428 				transmp = bp->b_vp->v_mount;
1429 				break;
1430 			}
1431 		}
1432 	}
1433 	if (bp != NULL) {
1434 		KASSERT(!ISSET(bp->b_cflags, BC_BUSY) ||
1435 		    ISSET(bp->b_cflags, BC_VFLUSH));
1436 		bremfree(bp);
1437 
1438 		/* Buffer is no longer on free lists. */
1439 		SET(bp->b_cflags, BC_BUSY);
1440 
1441 		/* Wake anyone trying to lock the old identity. */
1442 		cv_broadcast(&bp->b_busy);
1443 	} else {
1444 		/*
1445 		 * XXX: !from_bufq should be removed.
1446 		 */
1447 		if (!from_bufq || curlwp != uvm.pagedaemon_lwp) {
1448 			/* wait for a free buffer of any kind */
1449 			if ((slpflag & PCATCH) != 0)
1450 				(void)cv_timedwait_sig(&needbuffer_cv,
1451 				    &bufcache_lock, slptimeo);
1452 			else
1453 				(void)cv_timedwait(&needbuffer_cv,
1454 				    &bufcache_lock, slptimeo);
1455 		}
1456 		SDT_PROBE1(io, kernel, , getnewbuf__done,  NULL);
1457 		return NULL;
1458 	}
1459 
1460 #ifdef DIAGNOSTIC
1461 	if (bp->b_bufsize <= 0)
1462 		panic("buffer %p: on queue but empty", bp);
1463 #endif
1464 
1465 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1466 		/*
1467 		 * This is a delayed write buffer being flushed to disk.  Make
1468 		 * sure it gets aged out of the queue when it's finished, and
1469 		 * leave it off the LRU queue.
1470 		 */
1471 		CLR(bp->b_cflags, BC_VFLUSH);
1472 		SET(bp->b_cflags, BC_AGE);
1473 		goto start;
1474 	}
1475 
1476 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1477 	KASSERT(!cv_has_waiters(&bp->b_done));
1478 
1479 	/*
1480 	 * If buffer was a delayed write, start it and return NULL
1481 	 * (since we might sleep while starting the write).
1482 	 */
1483 	if (ISSET(bp->b_oflags, BO_DELWRI)) {
1484 		/*
1485 		 * This buffer has gone through the LRU, so make sure it gets
1486 		 * reused ASAP.
1487 		 */
1488 		SET(bp->b_cflags, BC_AGE);
1489 		mutex_exit(&bufcache_lock);
1490 		bawrite(bp);
1491 		KASSERT(transmp != NULL);
1492 		fstrans_done(transmp);
1493 		mutex_enter(&bufcache_lock);
1494 		SDT_PROBE1(io, kernel, , getnewbuf__done,  NULL);
1495 		return NULL;
1496 	}
1497 
1498 	KASSERT(transmp == NULL);
1499 
1500 	vp = bp->b_vp;
1501 
1502 	/* clear out various other fields */
1503 	bp->b_cflags = BC_BUSY;
1504 	bp->b_oflags = 0;
1505 	bp->b_flags = 0;
1506 	bp->b_dev = NODEV;
1507 	bp->b_blkno = 0;
1508 	bp->b_lblkno = 0;
1509 	bp->b_rawblkno = 0;
1510 	bp->b_iodone = 0;
1511 	bp->b_error = 0;
1512 	bp->b_resid = 0;
1513 	bp->b_bcount = 0;
1514 
1515 	LIST_REMOVE(bp, b_hash);
1516 
1517 	/* Disassociate us from our vnode, if we had one... */
1518 	if (vp != NULL) {
1519 		mutex_enter(vp->v_interlock);
1520 		brelvp(bp);
1521 		mutex_exit(vp->v_interlock);
1522 	}
1523 
1524 	SDT_PROBE1(io, kernel, , getnewbuf__done,  bp);
1525 	return bp;
1526 }
1527 
1528 /*
1529  * Invalidate the specified buffer if it exists.
1530  */
1531 void
1532 binvalbuf(struct vnode *vp, daddr_t blkno)
1533 {
1534 	buf_t *bp;
1535 	int err;
1536 
1537 	mutex_enter(&bufcache_lock);
1538 
1539 loop:
1540 	bp = incore(vp, blkno);
1541 	if (bp != NULL) {
1542 		err = bbusy(bp, 0, 0, NULL);
1543 		if (err == EPASSTHROUGH)
1544 			goto loop;
1545 		bremfree(bp);
1546 		if (ISSET(bp->b_oflags, BO_DELWRI)) {
1547 			SET(bp->b_cflags, BC_NOCACHE);
1548 			mutex_exit(&bufcache_lock);
1549 			bwrite(bp);
1550 		} else {
1551 			brelsel(bp, BC_INVAL);
1552 			mutex_exit(&bufcache_lock);
1553 		}
1554 	} else
1555 		mutex_exit(&bufcache_lock);
1556 }
1557 
1558 /*
1559  * Attempt to free an aged buffer off the queues.
1560  * Called with queue lock held.
1561  * Returns the amount of buffer memory freed.
1562  */
1563 static int
1564 buf_trim(void)
1565 {
1566 	buf_t *bp;
1567 	long size;
1568 
1569 	KASSERT(mutex_owned(&bufcache_lock));
1570 
1571 	/* Instruct getnewbuf() to get buffers off the queues */
1572 	if ((bp = getnewbuf(PCATCH, 1, 1)) == NULL)
1573 		return 0;
1574 
1575 	KASSERT((bp->b_cflags & BC_WANTED) == 0);
1576 	size = bp->b_bufsize;
1577 	bufmem -= size;
1578 	if (size > 0) {
1579 		buf_mrelease(bp->b_data, size);
1580 		bp->b_bcount = bp->b_bufsize = 0;
1581 	}
1582 	/* brelse() will return the buffer to the global buffer pool */
1583 	brelsel(bp, 0);
1584 	return size;
1585 }
1586 
1587 int
1588 buf_drain(int n)
1589 {
1590 	int size = 0, sz;
1591 
1592 	KASSERT(mutex_owned(&bufcache_lock));
1593 
1594 	while (size < n && bufmem > bufmem_lowater) {
1595 		sz = buf_trim();
1596 		if (sz <= 0)
1597 			break;
1598 		size += sz;
1599 	}
1600 
1601 	return size;
1602 }
1603 
1604 /*
1605  * Wait for operations on the buffer to complete.
1606  * When they do, extract and return the I/O's error value.
1607  */
1608 int
1609 biowait(buf_t *bp)
1610 {
1611 
1612 	BIOHIST_FUNC(__func__);
1613 
1614 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1615 
1616 	SDT_PROBE1(io, kernel, , wait__start, bp);
1617 
1618 	mutex_enter(bp->b_objlock);
1619 
1620 	BIOHIST_CALLARGS(biohist, "bp=%#jx, oflags=0x%jx, ret_addr=%#jx",
1621 	    (uintptr_t)bp, bp->b_oflags,
1622 	    (uintptr_t)__builtin_return_address(0), 0);
1623 
1624 	while (!ISSET(bp->b_oflags, BO_DONE | BO_DELWRI)) {
1625 		BIOHIST_LOG(biohist, "waiting bp=%#jx",
1626 		    (uintptr_t)bp, 0, 0, 0);
1627 		cv_wait(&bp->b_done, bp->b_objlock);
1628 	}
1629 	mutex_exit(bp->b_objlock);
1630 
1631 	SDT_PROBE1(io, kernel, , wait__done, bp);
1632 
1633 	BIOHIST_LOG(biohist, "return %jd", bp->b_error, 0, 0, 0);
1634 
1635 	return bp->b_error;
1636 }
1637 
1638 /*
1639  * Mark I/O complete on a buffer.
1640  *
1641  * If a callback has been requested, e.g. the pageout
1642  * daemon, do so. Otherwise, awaken waiting processes.
1643  *
1644  * [ Leffler, et al., says on p.247:
1645  *	"This routine wakes up the blocked process, frees the buffer
1646  *	for an asynchronous write, or, for a request by the pagedaemon
1647  *	process, invokes a procedure specified in the buffer structure" ]
1648  *
1649  * In real life, the pagedaemon (or other system processes) wants
1650  * to do async stuff too, and doesn't want the buffer brelse()'d.
1651  * (for swap pager, that puts swap buffers on the free lists (!!!),
1652  * for the vn device, that puts allocated buffers on the free lists!)
1653  */
1654 void
1655 biodone(buf_t *bp)
1656 {
1657 	int s;
1658 
1659 	BIOHIST_FUNC(__func__);
1660 
1661 	KASSERT(!ISSET(bp->b_oflags, BO_DONE));
1662 
1663 	if (cpu_intr_p()) {
1664 		/* From interrupt mode: defer to a soft interrupt. */
1665 		s = splvm();
1666 		TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_biodone, bp, b_actq);
1667 
1668 		BIOHIST_CALLARGS(biohist, "bp=%#jx, softint scheduled",
1669 		    (uintptr_t)bp, 0, 0, 0);
1670 		softint_schedule(biodone_sih);
1671 		splx(s);
1672 	} else {
1673 		/* Process now - the buffer may be freed soon. */
1674 		biodone2(bp);
1675 	}
1676 }
1677 
1678 SDT_PROBE_DEFINE1(io, kernel, , done, "struct buf *"/*bp*/);
1679 
1680 static void
1681 biodone2(buf_t *bp)
1682 {
1683 	void (*callout)(buf_t *);
1684 
1685 	SDT_PROBE1(io, kernel, ,done, bp);
1686 
1687 	BIOHIST_FUNC(__func__);
1688 	BIOHIST_CALLARGS(biohist, "bp=%#jx", (uintptr_t)bp, 0, 0, 0);
1689 
1690 	mutex_enter(bp->b_objlock);
1691 	/* Note that the transfer is done. */
1692 	if (ISSET(bp->b_oflags, BO_DONE))
1693 		panic("biodone2 already");
1694 	CLR(bp->b_flags, B_COWDONE);
1695 	SET(bp->b_oflags, BO_DONE);
1696 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1697 
1698 	/* Wake up waiting writers. */
1699 	if (!ISSET(bp->b_flags, B_READ))
1700 		vwakeup(bp);
1701 
1702 	if ((callout = bp->b_iodone) != NULL) {
1703 		BIOHIST_LOG(biohist, "callout %#jx", (uintptr_t)callout,
1704 		    0, 0, 0);
1705 
1706 		/* Note callout done, then call out. */
1707 		KASSERT(!cv_has_waiters(&bp->b_done));
1708 		bp->b_iodone = NULL;
1709 		mutex_exit(bp->b_objlock);
1710 		(*callout)(bp);
1711 	} else if (ISSET(bp->b_flags, B_ASYNC)) {
1712 		/* If async, release. */
1713 		BIOHIST_LOG(biohist, "async", 0, 0, 0, 0);
1714 		KASSERT(!cv_has_waiters(&bp->b_done));
1715 		mutex_exit(bp->b_objlock);
1716 		brelse(bp, 0);
1717 	} else {
1718 		/* Otherwise just wake up waiters in biowait(). */
1719 		BIOHIST_LOG(biohist, "wake-up", 0, 0, 0, 0);
1720 		cv_broadcast(&bp->b_done);
1721 		mutex_exit(bp->b_objlock);
1722 	}
1723 }
1724 
1725 static void
1726 biointr(void *cookie)
1727 {
1728 	struct cpu_info *ci;
1729 	buf_t *bp;
1730 	int s;
1731 
1732 	BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
1733 
1734 	ci = curcpu();
1735 
1736 	s = splvm();
1737 	while (!TAILQ_EMPTY(&ci->ci_data.cpu_biodone)) {
1738 		KASSERT(curcpu() == ci);
1739 
1740 		bp = TAILQ_FIRST(&ci->ci_data.cpu_biodone);
1741 		TAILQ_REMOVE(&ci->ci_data.cpu_biodone, bp, b_actq);
1742 		splx(s);
1743 
1744 		BIOHIST_LOG(biohist, "bp=%#jx", (uintptr_t)bp, 0, 0, 0);
1745 		biodone2(bp);
1746 
1747 		s = splvm();
1748 	}
1749 	splx(s);
1750 }
1751 
1752 static void
1753 sysctl_fillbuf(const buf_t *i, struct buf_sysctl *o)
1754 {
1755 	const bool allowaddr = get_expose_address(curproc);
1756 
1757 	memset(o, 0, sizeof(*o));
1758 
1759 	o->b_flags = i->b_flags | i->b_cflags | i->b_oflags;
1760 	o->b_error = i->b_error;
1761 	o->b_prio = i->b_prio;
1762 	o->b_dev = i->b_dev;
1763 	o->b_bufsize = i->b_bufsize;
1764 	o->b_bcount = i->b_bcount;
1765 	o->b_resid = i->b_resid;
1766 	COND_SET_VALUE(o->b_addr, PTRTOUINT64(i->b_data), allowaddr);
1767 	o->b_blkno = i->b_blkno;
1768 	o->b_rawblkno = i->b_rawblkno;
1769 	COND_SET_VALUE(o->b_iodone, PTRTOUINT64(i->b_iodone), allowaddr);
1770 	COND_SET_VALUE(o->b_proc, PTRTOUINT64(i->b_proc), allowaddr);
1771 	COND_SET_VALUE(o->b_vp, PTRTOUINT64(i->b_vp), allowaddr);
1772 	COND_SET_VALUE(o->b_saveaddr, PTRTOUINT64(i->b_saveaddr), allowaddr);
1773 	o->b_lblkno = i->b_lblkno;
1774 }
1775 
1776 static int
1777 sysctl_dobuf(SYSCTLFN_ARGS)
1778 {
1779 	buf_t *bp;
1780 	struct buf_sysctl bs;
1781 	struct bqueue *bq;
1782 	char *dp;
1783 	u_int i, op, arg;
1784 	size_t len, needed, elem_size, out_size;
1785 	int error, elem_count, retries;
1786 
1787 	if (namelen == 1 && name[0] == CTL_QUERY)
1788 		return sysctl_query(SYSCTLFN_CALL(rnode));
1789 
1790 	if (namelen != 4)
1791 		return SET_ERROR(EINVAL);
1792 
1793 	retries = 100;
1794 retry:
1795 	dp = oldp;
1796 	len = (oldp != NULL) ? *oldlenp : 0;
1797 	op = name[0];
1798 	arg = name[1];
1799 	elem_size = name[2];
1800 	elem_count = name[3];
1801 	out_size = MIN(sizeof(bs), elem_size);
1802 
1803 	/*
1804 	 * at the moment, these are just "placeholders" to make the
1805 	 * API for retrieving kern.buf data more extensible in the
1806 	 * future.
1807 	 *
1808 	 * XXX kern.buf currently has "netbsd32" issues.  hopefully
1809 	 * these will be resolved at a later point.
1810 	 */
1811 	if (op != KERN_BUF_ALL || arg != KERN_BUF_ALL ||
1812 	    elem_size < 1 || elem_count < 0)
1813 		return SET_ERROR(EINVAL);
1814 
1815 	if (oldp == NULL) {
1816 		/* count only, don't run through the buffer queues */
1817 		needed = pool_cache_nget(buf_cache) -
1818 		    pool_cache_nput(buf_cache);
1819 		*oldlenp = (needed + KERN_BUFSLOP) * elem_size;
1820 
1821 		return 0;
1822 	}
1823 
1824 	error = 0;
1825 	needed = 0;
1826 	sysctl_unlock();
1827 	mutex_enter(&bufcache_lock);
1828 	for (i = 0; i < BQUEUES; i++) {
1829 		bq = &bufqueues[i];
1830 		TAILQ_FOREACH(bp, &bq->bq_queue, b_freelist) {
1831 			bq->bq_marker = bp;
1832 			if (len >= elem_size && elem_count > 0) {
1833 				sysctl_fillbuf(bp, &bs);
1834 				mutex_exit(&bufcache_lock);
1835 				error = copyout(&bs, dp, out_size);
1836 				mutex_enter(&bufcache_lock);
1837 				if (error)
1838 					break;
1839 				if (bq->bq_marker != bp) {
1840 					/*
1841 					 * This sysctl node is only for
1842 					 * statistics.  Retry; if the
1843 					 * queue keeps changing, then
1844 					 * bail out.
1845 					 */
1846 					if (retries-- == 0) {
1847 						error = SET_ERROR(EAGAIN);
1848 						break;
1849 					}
1850 					mutex_exit(&bufcache_lock);
1851 					sysctl_relock();
1852 					goto retry;
1853 				}
1854 				dp += elem_size;
1855 				len -= elem_size;
1856 			}
1857 			needed += elem_size;
1858 			if (elem_count > 0 && elem_count != INT_MAX)
1859 				elem_count--;
1860 		}
1861 		if (error != 0)
1862 			break;
1863 	}
1864 	mutex_exit(&bufcache_lock);
1865 	sysctl_relock();
1866 
1867 	*oldlenp = needed;
1868 
1869 	return error;
1870 }
1871 
1872 static int
1873 sysctl_bufvm_update(SYSCTLFN_ARGS)
1874 {
1875 	int error, rv;
1876 	struct sysctlnode node;
1877 	unsigned int temp_bufcache;
1878 	unsigned long temp_water;
1879 
1880 	/* Take a copy of the supplied node and its data */
1881 	node = *rnode;
1882 	if (node.sysctl_data == &bufcache) {
1883 		node.sysctl_data = &temp_bufcache;
1884 		temp_bufcache = *(unsigned int *)rnode->sysctl_data;
1885 	} else {
1886 		node.sysctl_data = &temp_water;
1887 		temp_water = *(unsigned long *)rnode->sysctl_data;
1888 	}
1889 
1890 	/* Update the copy */
1891 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
1892 	if (error || newp == NULL)
1893 		return error;
1894 
1895 	if (rnode->sysctl_data == &bufcache) {
1896 		if (temp_bufcache > 100)
1897 			return SET_ERROR(EINVAL);
1898 		bufcache = temp_bufcache;
1899 		buf_setwm();
1900 	} else if (rnode->sysctl_data == &bufmem_lowater) {
1901 		if (bufmem_hiwater - temp_water < 16)
1902 			return SET_ERROR(EINVAL);
1903 		bufmem_lowater = temp_water;
1904 	} else if (rnode->sysctl_data == &bufmem_hiwater) {
1905 		if (temp_water - bufmem_lowater < 16)
1906 			return SET_ERROR(EINVAL);
1907 		bufmem_hiwater = temp_water;
1908 	} else
1909 		return SET_ERROR(EINVAL);
1910 
1911 	/* Drain until below new high water mark */
1912 	sysctl_unlock();
1913 	mutex_enter(&bufcache_lock);
1914 	while (bufmem > bufmem_hiwater) {
1915 		rv = buf_drain((bufmem - bufmem_hiwater) / (2 * 1024));
1916 		if (rv <= 0)
1917 			break;
1918 	}
1919 	mutex_exit(&bufcache_lock);
1920 	sysctl_relock();
1921 
1922 	return 0;
1923 }
1924 
1925 static struct sysctllog *vfsbio_sysctllog;
1926 
1927 static void
1928 sysctl_kern_buf_setup(void)
1929 {
1930 
1931 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1932 	    CTLFLAG_PERMANENT,
1933 	    CTLTYPE_NODE, "buf",
1934 	    SYSCTL_DESCR("Kernel buffer cache information"),
1935 	    sysctl_dobuf, 0, NULL, 0,
1936 	    CTL_KERN, KERN_BUF, CTL_EOL);
1937 }
1938 
1939 static void
1940 sysctl_vm_buf_setup(void)
1941 {
1942 
1943 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1944 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1945 	    CTLTYPE_INT, "bufcache",
1946 	    SYSCTL_DESCR("Percentage of physical memory to use for "
1947 		"buffer cache"),
1948 	    sysctl_bufvm_update, 0, &bufcache, 0,
1949 	    CTL_VM, CTL_CREATE, CTL_EOL);
1950 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1951 	    CTLFLAG_PERMANENT|CTLFLAG_READONLY,
1952 	    CTLTYPE_LONG, "bufmem",
1953 	    SYSCTL_DESCR("Amount of kernel memory used by buffer cache"),
1954 	    NULL, 0, &bufmem, 0,
1955 	    CTL_VM, CTL_CREATE, CTL_EOL);
1956 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1957 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1958 	    CTLTYPE_LONG, "bufmem_lowater",
1959 	    SYSCTL_DESCR("Minimum amount of kernel memory to reserve for "
1960 		"buffer cache"),
1961 	    sysctl_bufvm_update, 0, &bufmem_lowater, 0,
1962 	    CTL_VM, CTL_CREATE, CTL_EOL);
1963 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1964 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1965 	    CTLTYPE_LONG, "bufmem_hiwater",
1966 	    SYSCTL_DESCR("Maximum amount of kernel memory to use for "
1967 		"buffer cache"),
1968 	    sysctl_bufvm_update, 0, &bufmem_hiwater, 0,
1969 	    CTL_VM, CTL_CREATE, CTL_EOL);
1970 }
1971 
1972 static int
1973 bufhash_stats(struct hashstat_sysctl *hs, bool fill)
1974 {
1975 	buf_t *bp;
1976 	uint64_t chain;
1977 
1978 	strlcpy(hs->hash_name, "bufhash", sizeof(hs->hash_name));
1979 	strlcpy(hs->hash_desc, "buffer hash", sizeof(hs->hash_desc));
1980 	if (!fill)
1981 		return 0;
1982 
1983 	hs->hash_size = bufhash + 1;
1984 
1985 	for (size_t i = 0; i < hs->hash_size; i++) {
1986 		chain = 0;
1987 
1988 		mutex_enter(&bufcache_lock);
1989 		LIST_FOREACH(bp, &bufhashtbl[i], b_hash) {
1990 			chain++;
1991 		}
1992 		mutex_exit(&bufcache_lock);
1993 
1994 		if (chain > 0) {
1995 			hs->hash_used++;
1996 			hs->hash_items += chain;
1997 			if (chain > hs->hash_maxchain)
1998 				hs->hash_maxchain = chain;
1999 		}
2000 		preempt_point();
2001 	}
2002 
2003 	return 0;
2004 }
2005 
2006 #ifdef DEBUG
2007 /*
2008  * Print out statistics on the current allocation of the buffer pool.
2009  * Can be enabled to print out on every ``sync'' by setting "syncprt"
2010  * in vfs_syscalls.c using sysctl.
2011  */
2012 void
2013 vfs_bufstats(void)
2014 {
2015 	int i, j, count;
2016 	buf_t *bp;
2017 	struct bqueue *dp;
2018 	int counts[MAXBSIZE / MIN_PAGE_SIZE + 1];
2019 	static const char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE" };
2020 
2021 	for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) {
2022 		count = 0;
2023 		memset(counts, 0, sizeof(counts));
2024 		TAILQ_FOREACH(bp, &dp->bq_queue, b_freelist) {
2025 			counts[bp->b_bufsize / PAGE_SIZE]++;
2026 			count++;
2027 		}
2028 		printf("%s: total-%d", bname[i], count);
2029 		for (j = 0; j <= MAXBSIZE / PAGE_SIZE; j++)
2030 			if (counts[j] != 0)
2031 				printf(", %d-%d", j * PAGE_SIZE, counts[j]);
2032 		printf("\n");
2033 	}
2034 }
2035 #endif /* DEBUG */
2036 
2037 /* ------------------------------ */
2038 
2039 buf_t *
2040 getiobuf(struct vnode *vp, bool waitok)
2041 {
2042 	buf_t *bp;
2043 
2044 	bp = pool_cache_get(bufio_cache, (waitok ? PR_WAITOK : PR_NOWAIT));
2045 	if (bp == NULL)
2046 		return bp;
2047 
2048 	buf_init(bp);
2049 
2050 	if ((bp->b_vp = vp) != NULL) {
2051 		bp->b_objlock = vp->v_interlock;
2052 	} else {
2053 		KASSERT(bp->b_objlock == &buffer_lock);
2054 	}
2055 
2056 	return bp;
2057 }
2058 
2059 void
2060 putiobuf(buf_t *bp)
2061 {
2062 
2063 	buf_destroy(bp);
2064 	pool_cache_put(bufio_cache, bp);
2065 }
2066 
2067 /*
2068  * nestiobuf_iodone: b_iodone callback for nested buffers.
2069  */
2070 
2071 void
2072 nestiobuf_iodone(buf_t *bp)
2073 {
2074 	buf_t *mbp = bp->b_private;
2075 	int error;
2076 	int donebytes;
2077 
2078 	KASSERT(bp->b_bcount <= bp->b_bufsize);
2079 	KASSERT(mbp != bp);
2080 
2081 	error = bp->b_error;
2082 	if (bp->b_error == 0 &&
2083 	    (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
2084 		/*
2085 		 * Not all got transferred, raise an error. We have no way to
2086 		 * propagate these conditions to mbp.
2087 		 */
2088 		error = SET_ERROR(EIO);
2089 	}
2090 
2091 	donebytes = bp->b_bufsize;
2092 
2093 	putiobuf(bp);
2094 	nestiobuf_done(mbp, donebytes, error);
2095 }
2096 
2097 /*
2098  * nestiobuf_setup: setup a "nested" buffer.
2099  *
2100  * => 'mbp' is a "master" buffer which is being divided into sub pieces.
2101  * => 'bp' should be a buffer allocated by getiobuf.
2102  * => 'offset' is a byte offset in the master buffer.
2103  * => 'size' is a size in bytes of this nested buffer.
2104  */
2105 
2106 void
2107 nestiobuf_setup(buf_t *mbp, buf_t *bp, int offset, size_t size)
2108 {
2109 	const int b_pass = mbp->b_flags & (B_READ|B_PHYS|B_RAW|B_MEDIA_FLAGS);
2110 	struct vnode *vp = mbp->b_vp;
2111 
2112 	KASSERT(mbp->b_bcount >= offset + size);
2113 	bp->b_vp = vp;
2114 	bp->b_dev = mbp->b_dev;
2115 	bp->b_objlock = mbp->b_objlock;
2116 	bp->b_cflags = BC_BUSY;
2117 	bp->b_flags = B_ASYNC | b_pass;
2118 	bp->b_iodone = nestiobuf_iodone;
2119 	bp->b_data = (char *)mbp->b_data + offset;
2120 	bp->b_resid = bp->b_bcount = size;
2121 	bp->b_bufsize = bp->b_bcount;
2122 	bp->b_private = mbp;
2123 	BIO_COPYPRIO(bp, mbp);
2124 	if (BUF_ISWRITE(bp) && vp != NULL) {
2125 		mutex_enter(vp->v_interlock);
2126 		vp->v_numoutput++;
2127 		mutex_exit(vp->v_interlock);
2128 	}
2129 }
2130 
2131 /*
2132  * nestiobuf_done: propagate completion to the master buffer.
2133  *
2134  * => 'donebytes' specifies how many bytes in the 'mbp' is completed.
2135  * => 'error' is an errno(2) that 'donebytes' has been completed with.
2136  */
2137 
2138 void
2139 nestiobuf_done(buf_t *mbp, int donebytes, int error)
2140 {
2141 
2142 	if (donebytes == 0) {
2143 		return;
2144 	}
2145 	mutex_enter(mbp->b_objlock);
2146 	KASSERT(mbp->b_resid >= donebytes);
2147 	mbp->b_resid -= donebytes;
2148 	if (error)
2149 		mbp->b_error = error;
2150 	if (mbp->b_resid == 0) {
2151 		if (mbp->b_error)
2152 			mbp->b_resid = mbp->b_bcount;
2153 		mutex_exit(mbp->b_objlock);
2154 		biodone(mbp);
2155 	} else
2156 		mutex_exit(mbp->b_objlock);
2157 }
2158 
2159 void
2160 buf_init(buf_t *bp)
2161 {
2162 
2163 	cv_init(&bp->b_busy, "biolock");
2164 	cv_init(&bp->b_done, "biowait");
2165 	bp->b_dev = NODEV;
2166 	bp->b_error = 0;
2167 	bp->b_flags = 0;
2168 	bp->b_cflags = 0;
2169 	bp->b_oflags = 0;
2170 	bp->b_objlock = &buffer_lock;
2171 	bp->b_iodone = NULL;
2172 	bp->b_dev = NODEV;
2173 	bp->b_vnbufs.le_next = NOLIST;
2174 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
2175 }
2176 
2177 void
2178 buf_destroy(buf_t *bp)
2179 {
2180 
2181 	cv_destroy(&bp->b_done);
2182 	cv_destroy(&bp->b_busy);
2183 }
2184 
2185 int
2186 bbusy(buf_t *bp, bool intr, int timo, kmutex_t *interlock)
2187 {
2188 	int error;
2189 
2190 	KASSERT(mutex_owned(&bufcache_lock));
2191 
2192 	SDT_PROBE4(io, kernel, , bbusy__start,  bp, intr, timo, interlock);
2193 
2194 	if ((bp->b_cflags & BC_BUSY) != 0) {
2195 		if (curlwp == uvm.pagedaemon_lwp) {
2196 			error = SET_ERROR(EDEADLK);
2197 			goto out;
2198 		}
2199 		bp->b_cflags |= BC_WANTED;
2200 		if (interlock != NULL)
2201 			mutex_exit(interlock);
2202 		if (intr) {
2203 			error = cv_timedwait_sig(&bp->b_busy, &bufcache_lock,
2204 			    timo);
2205 		} else {
2206 			error = cv_timedwait(&bp->b_busy, &bufcache_lock,
2207 			    timo);
2208 		}
2209 		/*
2210 		 * At this point the buffer may be gone: don't touch it
2211 		 * again.  The caller needs to find it again and retry.
2212 		 */
2213 		if (interlock != NULL)
2214 			mutex_enter(interlock);
2215 		if (error == 0)
2216 			error = SET_ERROR(EPASSTHROUGH);
2217 	} else {
2218 		bp->b_cflags |= BC_BUSY;
2219 		error = 0;
2220 	}
2221 
2222 out:	SDT_PROBE5(io, kernel, , bbusy__done,
2223 	    bp, intr, timo, interlock, error);
2224 	return error;
2225 }
2226 
2227 /*
2228  * Nothing outside this file should really need to know about nbuf,
2229  * but a few things still want to read it, so give them a way to do that.
2230  */
2231 u_int
2232 buf_nbuf(void)
2233 {
2234 
2235 	return nbuf;
2236 }
2237