Home | History | Annotate | Line # | Download | only in kern
vfs_bio.c revision 1.265
      1 /*	$NetBSD: vfs_bio.c,v 1.265 2016/12/26 23:15:15 pgoyette 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 1.265 2016/12/26 23:15:15 pgoyette 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 #ifdef BIOHIST
    203 #include <sys/biohist.h>
    204 #endif	/* BIOHIST */
    205 
    206 KERNHIST_DEFINE(biohist);
    207 
    208 void
    209 biohist_init(void)
    210 {
    211 
    212 	KERNHIST_INIT(biohist, BIOHIST_SIZE);
    213 }
    214 
    215 /*
    216  * Definitions for the buffer hash lists.
    217  */
    218 #define	BUFHASH(dvp, lbn)	\
    219 	(&bufhashtbl[(((long)(dvp) >> 8) + (int)(lbn)) & bufhash])
    220 LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
    221 u_long	bufhash;
    222 struct bqueue bufqueues[BQUEUES];
    223 
    224 static kcondvar_t needbuffer_cv;
    225 
    226 /*
    227  * Buffer queue lock.
    228  */
    229 kmutex_t bufcache_lock;
    230 kmutex_t buffer_lock;
    231 
    232 /* Software ISR for completed transfers. */
    233 static void *biodone_sih;
    234 
    235 /* Buffer pool for I/O buffers. */
    236 static pool_cache_t buf_cache;
    237 static pool_cache_t bufio_cache;
    238 
    239 #define MEMPOOL_INDEX_OFFSET (ilog2(DEV_BSIZE))	/* smallest pool is 512 bytes */
    240 #define NMEMPOOLS (ilog2(MAXBSIZE) - MEMPOOL_INDEX_OFFSET + 1)
    241 __CTASSERT((1 << (NMEMPOOLS + MEMPOOL_INDEX_OFFSET - 1)) == MAXBSIZE);
    242 
    243 /* Buffer memory pools */
    244 static struct pool bmempools[NMEMPOOLS];
    245 
    246 static struct vm_map *buf_map;
    247 
    248 /*
    249  * Buffer memory pool allocator.
    250  */
    251 static void *
    252 bufpool_page_alloc(struct pool *pp, int flags)
    253 {
    254 
    255 	return (void *)uvm_km_alloc(buf_map,
    256 	    MAXBSIZE, MAXBSIZE,
    257 	    ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT|UVM_KMF_TRYLOCK)
    258 	    | UVM_KMF_WIRED);
    259 }
    260 
    261 static void
    262 bufpool_page_free(struct pool *pp, void *v)
    263 {
    264 
    265 	uvm_km_free(buf_map, (vaddr_t)v, MAXBSIZE, UVM_KMF_WIRED);
    266 }
    267 
    268 static struct pool_allocator bufmempool_allocator = {
    269 	.pa_alloc = bufpool_page_alloc,
    270 	.pa_free = bufpool_page_free,
    271 	.pa_pagesz = MAXBSIZE,
    272 };
    273 
    274 /* Buffer memory management variables */
    275 u_long bufmem_valimit;
    276 u_long bufmem_hiwater;
    277 u_long bufmem_lowater;
    278 u_long bufmem;
    279 
    280 /*
    281  * MD code can call this to set a hard limit on the amount
    282  * of virtual memory used by the buffer cache.
    283  */
    284 int
    285 buf_setvalimit(vsize_t sz)
    286 {
    287 
    288 	/* We need to accommodate at least NMEMPOOLS of MAXBSIZE each */
    289 	if (sz < NMEMPOOLS * MAXBSIZE)
    290 		return EINVAL;
    291 
    292 	bufmem_valimit = sz;
    293 	return 0;
    294 }
    295 
    296 static void
    297 buf_setwm(void)
    298 {
    299 
    300 	bufmem_hiwater = buf_memcalc();
    301 	/* lowater is approx. 2% of memory (with bufcache = 15) */
    302 #define	BUFMEM_WMSHIFT	3
    303 #define	BUFMEM_HIWMMIN	(64 * 1024 << BUFMEM_WMSHIFT)
    304 	if (bufmem_hiwater < BUFMEM_HIWMMIN)
    305 		/* Ensure a reasonable minimum value */
    306 		bufmem_hiwater = BUFMEM_HIWMMIN;
    307 	bufmem_lowater = bufmem_hiwater >> BUFMEM_WMSHIFT;
    308 }
    309 
    310 #ifdef DEBUG
    311 int debug_verify_freelist = 0;
    312 static int
    313 checkfreelist(buf_t *bp, struct bqueue *dp, int ison)
    314 {
    315 	buf_t *b;
    316 
    317 	if (!debug_verify_freelist)
    318 		return 1;
    319 
    320 	TAILQ_FOREACH(b, &dp->bq_queue, b_freelist) {
    321 		if (b == bp)
    322 			return ison ? 1 : 0;
    323 	}
    324 
    325 	return ison ? 0 : 1;
    326 }
    327 #endif
    328 
    329 /*
    330  * Insq/Remq for the buffer hash lists.
    331  * Call with buffer queue locked.
    332  */
    333 static void
    334 binsheadfree(buf_t *bp, struct bqueue *dp)
    335 {
    336 
    337 	KASSERT(mutex_owned(&bufcache_lock));
    338 	KASSERT(bp->b_freelistindex == -1);
    339 	TAILQ_INSERT_HEAD(&dp->bq_queue, bp, b_freelist);
    340 	dp->bq_bytes += bp->b_bufsize;
    341 	bp->b_freelistindex = dp - bufqueues;
    342 }
    343 
    344 static void
    345 binstailfree(buf_t *bp, struct bqueue *dp)
    346 {
    347 
    348 	KASSERT(mutex_owned(&bufcache_lock));
    349 	KASSERTMSG(bp->b_freelistindex == -1, "double free of buffer? "
    350 	    "bp=%p, b_freelistindex=%d\n", bp, bp->b_freelistindex);
    351 	TAILQ_INSERT_TAIL(&dp->bq_queue, bp, b_freelist);
    352 	dp->bq_bytes += bp->b_bufsize;
    353 	bp->b_freelistindex = dp - bufqueues;
    354 }
    355 
    356 void
    357 bremfree(buf_t *bp)
    358 {
    359 	struct bqueue *dp;
    360 	int bqidx = bp->b_freelistindex;
    361 
    362 	KASSERT(mutex_owned(&bufcache_lock));
    363 
    364 	KASSERT(bqidx != -1);
    365 	dp = &bufqueues[bqidx];
    366 	KDASSERT(checkfreelist(bp, dp, 1));
    367 	KASSERT(dp->bq_bytes >= bp->b_bufsize);
    368 	TAILQ_REMOVE(&dp->bq_queue, bp, b_freelist);
    369 	dp->bq_bytes -= bp->b_bufsize;
    370 
    371 	/* For the sysctl helper. */
    372 	if (bp == dp->bq_marker)
    373 		dp->bq_marker = NULL;
    374 
    375 #if defined(DIAGNOSTIC)
    376 	bp->b_freelistindex = -1;
    377 #endif /* defined(DIAGNOSTIC) */
    378 }
    379 
    380 /*
    381  * Add a reference to an buffer structure that came from buf_cache.
    382  */
    383 static inline void
    384 bref(buf_t *bp)
    385 {
    386 
    387 	KASSERT(mutex_owned(&bufcache_lock));
    388 	KASSERT(bp->b_refcnt > 0);
    389 
    390 	bp->b_refcnt++;
    391 }
    392 
    393 /*
    394  * Free an unused buffer structure that came from buf_cache.
    395  */
    396 static inline void
    397 brele(buf_t *bp)
    398 {
    399 
    400 	KASSERT(mutex_owned(&bufcache_lock));
    401 	KASSERT(bp->b_refcnt > 0);
    402 
    403 	if (bp->b_refcnt-- == 1) {
    404 		buf_destroy(bp);
    405 #ifdef DEBUG
    406 		memset((char *)bp, 0, sizeof(*bp));
    407 #endif
    408 		pool_cache_put(buf_cache, bp);
    409 	}
    410 }
    411 
    412 /*
    413  * note that for some ports this is used by pmap bootstrap code to
    414  * determine kva size.
    415  */
    416 u_long
    417 buf_memcalc(void)
    418 {
    419 	u_long n;
    420 	vsize_t mapsz = 0;
    421 
    422 	/*
    423 	 * Determine the upper bound of memory to use for buffers.
    424 	 *
    425 	 *	- If bufpages is specified, use that as the number
    426 	 *	  pages.
    427 	 *
    428 	 *	- Otherwise, use bufcache as the percentage of
    429 	 *	  physical memory.
    430 	 */
    431 	if (bufpages != 0) {
    432 		n = bufpages;
    433 	} else {
    434 		if (bufcache < 5) {
    435 			printf("forcing bufcache %d -> 5", bufcache);
    436 			bufcache = 5;
    437 		}
    438 		if (bufcache > 95) {
    439 			printf("forcing bufcache %d -> 95", bufcache);
    440 			bufcache = 95;
    441 		}
    442 		if (buf_map != NULL)
    443 			mapsz = vm_map_max(buf_map) - vm_map_min(buf_map);
    444 		n = calc_cache_size(mapsz, bufcache,
    445 		    (buf_map != kernel_map) ? 100 : BUFCACHE_VA_MAXPCT)
    446 		    / PAGE_SIZE;
    447 	}
    448 
    449 	n <<= PAGE_SHIFT;
    450 	if (bufmem_valimit != 0 && n > bufmem_valimit)
    451 		n = bufmem_valimit;
    452 
    453 	return (n);
    454 }
    455 
    456 /*
    457  * Initialize buffers and hash links for buffers.
    458  */
    459 void
    460 bufinit(void)
    461 {
    462 	struct bqueue *dp;
    463 	int use_std;
    464 	u_int i;
    465 
    466 	biodone_vfs = biodone;
    467 
    468 	mutex_init(&bufcache_lock, MUTEX_DEFAULT, IPL_NONE);
    469 	mutex_init(&buffer_lock, MUTEX_DEFAULT, IPL_NONE);
    470 	cv_init(&needbuffer_cv, "needbuf");
    471 
    472 	if (bufmem_valimit != 0) {
    473 		vaddr_t minaddr = 0, maxaddr;
    474 		buf_map = uvm_km_suballoc(kernel_map, &minaddr, &maxaddr,
    475 					  bufmem_valimit, 0, false, 0);
    476 		if (buf_map == NULL)
    477 			panic("bufinit: cannot allocate submap");
    478 	} else
    479 		buf_map = kernel_map;
    480 
    481 	/*
    482 	 * Initialize buffer cache memory parameters.
    483 	 */
    484 	bufmem = 0;
    485 	buf_setwm();
    486 
    487 	/* On "small" machines use small pool page sizes where possible */
    488 	use_std = (physmem < atop(16*1024*1024));
    489 
    490 	/*
    491 	 * Also use them on systems that can map the pool pages using
    492 	 * a direct-mapped segment.
    493 	 */
    494 #ifdef PMAP_MAP_POOLPAGE
    495 	use_std = 1;
    496 #endif
    497 
    498 	buf_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
    499 	    "bufpl", NULL, IPL_SOFTBIO, NULL, NULL, NULL);
    500 	bufio_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
    501 	    "biopl", NULL, IPL_BIO, NULL, NULL, NULL);
    502 
    503 	for (i = 0; i < NMEMPOOLS; i++) {
    504 		struct pool_allocator *pa;
    505 		struct pool *pp = &bmempools[i];
    506 		u_int size = 1 << (i + MEMPOOL_INDEX_OFFSET);
    507 		char *name = kmem_alloc(8, KM_SLEEP); /* XXX: never freed */
    508 		if (__predict_false(size >= 1048576))
    509 			(void)snprintf(name, 8, "buf%um", size / 1048576);
    510 		else if (__predict_true(size >= 1024))
    511 			(void)snprintf(name, 8, "buf%uk", size / 1024);
    512 		else
    513 			(void)snprintf(name, 8, "buf%ub", size);
    514 		pa = (size <= PAGE_SIZE && use_std)
    515 			? &pool_allocator_nointr
    516 			: &bufmempool_allocator;
    517 		pool_init(pp, size, 0, 0, 0, name, pa, IPL_NONE);
    518 		pool_setlowat(pp, 1);
    519 		pool_sethiwat(pp, 1);
    520 	}
    521 
    522 	/* Initialize the buffer queues */
    523 	for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) {
    524 		TAILQ_INIT(&dp->bq_queue);
    525 		dp->bq_bytes = 0;
    526 	}
    527 
    528 	/*
    529 	 * Estimate hash table size based on the amount of memory we
    530 	 * intend to use for the buffer cache. The average buffer
    531 	 * size is dependent on our clients (i.e. filesystems).
    532 	 *
    533 	 * For now, use an empirical 3K per buffer.
    534 	 */
    535 	nbuf = (bufmem_hiwater / 1024) / 3;
    536 	bufhashtbl = hashinit(nbuf, HASH_LIST, true, &bufhash);
    537 
    538 	sysctl_kern_buf_setup();
    539 	sysctl_vm_buf_setup();
    540 }
    541 
    542 void
    543 bufinit2(void)
    544 {
    545 
    546 	biodone_sih = softint_establish(SOFTINT_BIO | SOFTINT_MPSAFE, biointr,
    547 	    NULL);
    548 	if (biodone_sih == NULL)
    549 		panic("bufinit2: can't establish soft interrupt");
    550 }
    551 
    552 static int
    553 buf_lotsfree(void)
    554 {
    555 	u_long guess;
    556 
    557 	/* Always allocate if less than the low water mark. */
    558 	if (bufmem < bufmem_lowater)
    559 		return 1;
    560 
    561 	/* Never allocate if greater than the high water mark. */
    562 	if (bufmem > bufmem_hiwater)
    563 		return 0;
    564 
    565 	/* If there's anything on the AGE list, it should be eaten. */
    566 	if (TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue) != NULL)
    567 		return 0;
    568 
    569 	/*
    570 	 * The probabily of getting a new allocation is inversely
    571 	 * proportional  to the current size of the cache above
    572 	 * the low water mark.  Divide the total first to avoid overflows
    573 	 * in the product.
    574 	 */
    575 	guess = cprng_fast32() % 16;
    576 
    577 	if ((bufmem_hiwater - bufmem_lowater) / 16 * guess >=
    578 	    (bufmem - bufmem_lowater))
    579 		return 1;
    580 
    581 	/* Otherwise don't allocate. */
    582 	return 0;
    583 }
    584 
    585 /*
    586  * Return estimate of bytes we think need to be
    587  * released to help resolve low memory conditions.
    588  *
    589  * => called with bufcache_lock held.
    590  */
    591 static int
    592 buf_canrelease(void)
    593 {
    594 	int pagedemand, ninvalid = 0;
    595 
    596 	KASSERT(mutex_owned(&bufcache_lock));
    597 
    598 	if (bufmem < bufmem_lowater)
    599 		return 0;
    600 
    601 	if (bufmem > bufmem_hiwater)
    602 		return bufmem - bufmem_hiwater;
    603 
    604 	ninvalid += bufqueues[BQ_AGE].bq_bytes;
    605 
    606 	pagedemand = uvmexp.freetarg - uvmexp.free;
    607 	if (pagedemand < 0)
    608 		return ninvalid;
    609 	return MAX(ninvalid, MIN(2 * MAXBSIZE,
    610 	    MIN((bufmem - bufmem_lowater) / 16, pagedemand * PAGE_SIZE)));
    611 }
    612 
    613 /*
    614  * Buffer memory allocation helper functions
    615  */
    616 static u_long
    617 buf_mempoolidx(u_long size)
    618 {
    619 	u_int n = 0;
    620 
    621 	size -= 1;
    622 	size >>= MEMPOOL_INDEX_OFFSET;
    623 	while (size) {
    624 		size >>= 1;
    625 		n += 1;
    626 	}
    627 	if (n >= NMEMPOOLS)
    628 		panic("buf mem pool index %d", n);
    629 	return n;
    630 }
    631 
    632 static u_long
    633 buf_roundsize(u_long size)
    634 {
    635 	/* Round up to nearest power of 2 */
    636 	return (1 << (buf_mempoolidx(size) + MEMPOOL_INDEX_OFFSET));
    637 }
    638 
    639 static void *
    640 buf_alloc(size_t size)
    641 {
    642 	u_int n = buf_mempoolidx(size);
    643 	void *addr;
    644 
    645 	while (1) {
    646 		addr = pool_get(&bmempools[n], PR_NOWAIT);
    647 		if (addr != NULL)
    648 			break;
    649 
    650 		/* No memory, see if we can free some. If so, try again */
    651 		mutex_enter(&bufcache_lock);
    652 		if (buf_drain(1) > 0) {
    653 			mutex_exit(&bufcache_lock);
    654 			continue;
    655 		}
    656 
    657 		if (curlwp == uvm.pagedaemon_lwp) {
    658 			mutex_exit(&bufcache_lock);
    659 			return NULL;
    660 		}
    661 
    662 		/* Wait for buffers to arrive on the LRU queue */
    663 		cv_timedwait(&needbuffer_cv, &bufcache_lock, hz / 4);
    664 		mutex_exit(&bufcache_lock);
    665 	}
    666 
    667 	return addr;
    668 }
    669 
    670 static void
    671 buf_mrelease(void *addr, size_t size)
    672 {
    673 
    674 	pool_put(&bmempools[buf_mempoolidx(size)], addr);
    675 }
    676 
    677 /*
    678  * bread()/breadn() helper.
    679  */
    680 static buf_t *
    681 bio_doread(struct vnode *vp, daddr_t blkno, int size, int async)
    682 {
    683 	buf_t *bp;
    684 	struct mount *mp;
    685 
    686 	bp = getblk(vp, blkno, size, 0, 0);
    687 
    688 	/*
    689 	 * getblk() may return NULL if we are the pagedaemon.
    690 	 */
    691 	if (bp == NULL) {
    692 		KASSERT(curlwp == uvm.pagedaemon_lwp);
    693 		return NULL;
    694 	}
    695 
    696 	/*
    697 	 * If buffer does not have data valid, start a read.
    698 	 * Note that if buffer is BC_INVAL, getblk() won't return it.
    699 	 * Therefore, it's valid if its I/O has completed or been delayed.
    700 	 */
    701 	if (!ISSET(bp->b_oflags, (BO_DONE | BO_DELWRI))) {
    702 		/* Start I/O for the buffer. */
    703 		SET(bp->b_flags, B_READ | async);
    704 		if (async)
    705 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
    706 		else
    707 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
    708 		VOP_STRATEGY(vp, bp);
    709 
    710 		/* Pay for the read. */
    711 		curlwp->l_ru.ru_inblock++;
    712 	} else if (async)
    713 		brelse(bp, 0);
    714 
    715 	if (vp->v_type == VBLK)
    716 		mp = spec_node_getmountedfs(vp);
    717 	else
    718 		mp = vp->v_mount;
    719 
    720 	/*
    721 	 * Collect statistics on synchronous and asynchronous reads.
    722 	 * Reads from block devices are charged to their associated
    723 	 * filesystem (if any).
    724 	 */
    725 	if (mp != NULL) {
    726 		if (async == 0)
    727 			mp->mnt_stat.f_syncreads++;
    728 		else
    729 			mp->mnt_stat.f_asyncreads++;
    730 	}
    731 
    732 	return (bp);
    733 }
    734 
    735 /*
    736  * Read a disk block.
    737  * This algorithm described in Bach (p.54).
    738  */
    739 int
    740 bread(struct vnode *vp, daddr_t blkno, int size, int flags, buf_t **bpp)
    741 {
    742 	buf_t *bp;
    743 	int error;
    744 
    745 	KERNHIST_FUNC(__func__); KERNHIST_CALLED(biohist);
    746 
    747 	/* Get buffer for block. */
    748 	bp = *bpp = bio_doread(vp, blkno, size, 0);
    749 	if (bp == NULL)
    750 		return ENOMEM;
    751 
    752 	/* Wait for the read to complete, and return result. */
    753 	error = biowait(bp);
    754 	if (error == 0 && (flags & B_MODIFY) != 0)
    755 		error = fscow_run(bp, true);
    756 	if (error) {
    757 		brelse(bp, 0);
    758 		*bpp = NULL;
    759 	}
    760 
    761 	return error;
    762 }
    763 
    764 /*
    765  * Read-ahead multiple disk blocks. The first is sync, the rest async.
    766  * Trivial modification to the breada algorithm presented in Bach (p.55).
    767  */
    768 int
    769 breadn(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablks,
    770     int *rasizes, int nrablks, int flags, buf_t **bpp)
    771 {
    772 	buf_t *bp;
    773 	int error, i;
    774 
    775 	KERNHIST_FUNC(__func__); KERNHIST_CALLED(biohist);
    776 
    777 	bp = *bpp = bio_doread(vp, blkno, size, 0);
    778 	if (bp == NULL)
    779 		return ENOMEM;
    780 
    781 	/*
    782 	 * For each of the read-ahead blocks, start a read, if necessary.
    783 	 */
    784 	mutex_enter(&bufcache_lock);
    785 	for (i = 0; i < nrablks; i++) {
    786 		/* If it's in the cache, just go on to next one. */
    787 		if (incore(vp, rablks[i]))
    788 			continue;
    789 
    790 		/* Get a buffer for the read-ahead block */
    791 		mutex_exit(&bufcache_lock);
    792 		(void) bio_doread(vp, rablks[i], rasizes[i], B_ASYNC);
    793 		mutex_enter(&bufcache_lock);
    794 	}
    795 	mutex_exit(&bufcache_lock);
    796 
    797 	/* Otherwise, we had to start a read for it; wait until it's valid. */
    798 	error = biowait(bp);
    799 	if (error == 0 && (flags & B_MODIFY) != 0)
    800 		error = fscow_run(bp, true);
    801 	if (error) {
    802 		brelse(bp, 0);
    803 		*bpp = NULL;
    804 	}
    805 
    806 	return error;
    807 }
    808 
    809 /*
    810  * Block write.  Described in Bach (p.56)
    811  */
    812 int
    813 bwrite(buf_t *bp)
    814 {
    815 	int rv, sync, wasdelayed;
    816 	struct vnode *vp;
    817 	struct mount *mp;
    818 
    819 	KERNHIST_FUNC(__func__); KERNHIST_CALLARGS(biohist, "bp=%p",
    820 	    bp, 0, 0, 0);
    821 
    822 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
    823 	KASSERT(!cv_has_waiters(&bp->b_done));
    824 
    825 	vp = bp->b_vp;
    826 
    827 	/*
    828 	 * dholland 20160728 AFAICT vp==NULL must be impossible as it
    829 	 * will crash upon reaching VOP_STRATEGY below... see further
    830 	 * analysis on tech-kern.
    831 	 */
    832 	KASSERTMSG(vp != NULL, "bwrite given buffer with null vnode");
    833 
    834 	if (vp != NULL) {
    835 		KASSERT(bp->b_objlock == vp->v_interlock);
    836 		if (vp->v_type == VBLK)
    837 			mp = spec_node_getmountedfs(vp);
    838 		else
    839 			mp = vp->v_mount;
    840 	} else {
    841 		mp = NULL;
    842 	}
    843 
    844 	if (mp && mp->mnt_wapbl) {
    845 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
    846 			bdwrite(bp);
    847 			return 0;
    848 		}
    849 	}
    850 
    851 	/*
    852 	 * Remember buffer type, to switch on it later.  If the write was
    853 	 * synchronous, but the file system was mounted with MNT_ASYNC,
    854 	 * convert it to a delayed write.
    855 	 * XXX note that this relies on delayed tape writes being converted
    856 	 * to async, not sync writes (which is safe, but ugly).
    857 	 */
    858 	sync = !ISSET(bp->b_flags, B_ASYNC);
    859 	if (sync && mp != NULL && ISSET(mp->mnt_flag, MNT_ASYNC)) {
    860 		bdwrite(bp);
    861 		return (0);
    862 	}
    863 
    864 	/*
    865 	 * Collect statistics on synchronous and asynchronous writes.
    866 	 * Writes to block devices are charged to their associated
    867 	 * filesystem (if any).
    868 	 */
    869 	if (mp != NULL) {
    870 		if (sync)
    871 			mp->mnt_stat.f_syncwrites++;
    872 		else
    873 			mp->mnt_stat.f_asyncwrites++;
    874 	}
    875 
    876 	/*
    877 	 * Pay for the I/O operation and make sure the buf is on the correct
    878 	 * vnode queue.
    879 	 */
    880 	bp->b_error = 0;
    881 	wasdelayed = ISSET(bp->b_oflags, BO_DELWRI);
    882 	CLR(bp->b_flags, B_READ);
    883 	if (wasdelayed) {
    884 		mutex_enter(&bufcache_lock);
    885 		mutex_enter(bp->b_objlock);
    886 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
    887 		reassignbuf(bp, bp->b_vp);
    888 		mutex_exit(&bufcache_lock);
    889 	} else {
    890 		curlwp->l_ru.ru_oublock++;
    891 		mutex_enter(bp->b_objlock);
    892 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
    893 	}
    894 	if (vp != NULL)
    895 		vp->v_numoutput++;
    896 	mutex_exit(bp->b_objlock);
    897 
    898 	/* Initiate disk write. */
    899 	if (sync)
    900 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
    901 	else
    902 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
    903 
    904 	VOP_STRATEGY(vp, bp);
    905 
    906 	if (sync) {
    907 		/* If I/O was synchronous, wait for it to complete. */
    908 		rv = biowait(bp);
    909 
    910 		/* Release the buffer. */
    911 		brelse(bp, 0);
    912 
    913 		return (rv);
    914 	} else {
    915 		return (0);
    916 	}
    917 }
    918 
    919 int
    920 vn_bwrite(void *v)
    921 {
    922 	struct vop_bwrite_args *ap = v;
    923 
    924 	return (bwrite(ap->a_bp));
    925 }
    926 
    927 /*
    928  * Delayed write.
    929  *
    930  * The buffer is marked dirty, but is not queued for I/O.
    931  * This routine should be used when the buffer is expected
    932  * to be modified again soon, typically a small write that
    933  * partially fills a buffer.
    934  *
    935  * NB: magnetic tapes cannot be delayed; they must be
    936  * written in the order that the writes are requested.
    937  *
    938  * Described in Leffler, et al. (pp. 208-213).
    939  */
    940 void
    941 bdwrite(buf_t *bp)
    942 {
    943 
    944 	KERNHIST_FUNC(__func__); KERNHIST_CALLARGS(biohist, "bp=%p",
    945 	    bp, 0, 0, 0);
    946 
    947 	KASSERT(bp->b_vp == NULL || bp->b_vp->v_tag != VT_UFS ||
    948 	    bp->b_vp->v_type == VBLK || ISSET(bp->b_flags, B_COWDONE));
    949 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
    950 	KASSERT(!cv_has_waiters(&bp->b_done));
    951 
    952 	/* If this is a tape block, write the block now. */
    953 	if (bdev_type(bp->b_dev) == D_TAPE) {
    954 		bawrite(bp);
    955 		return;
    956 	}
    957 
    958 	if (wapbl_vphaswapbl(bp->b_vp)) {
    959 		struct mount *mp = wapbl_vptomp(bp->b_vp);
    960 
    961 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
    962 			WAPBL_ADD_BUF(mp, bp);
    963 		}
    964 	}
    965 
    966 	/*
    967 	 * If the block hasn't been seen before:
    968 	 *	(1) Mark it as having been seen,
    969 	 *	(2) Charge for the write,
    970 	 *	(3) Make sure it's on its vnode's correct block list.
    971 	 */
    972 	KASSERT(bp->b_vp == NULL || bp->b_objlock == bp->b_vp->v_interlock);
    973 
    974 	if (!ISSET(bp->b_oflags, BO_DELWRI)) {
    975 		mutex_enter(&bufcache_lock);
    976 		mutex_enter(bp->b_objlock);
    977 		SET(bp->b_oflags, BO_DELWRI);
    978 		curlwp->l_ru.ru_oublock++;
    979 		reassignbuf(bp, bp->b_vp);
    980 		mutex_exit(&bufcache_lock);
    981 	} else {
    982 		mutex_enter(bp->b_objlock);
    983 	}
    984 	/* Otherwise, the "write" is done, so mark and release the buffer. */
    985 	CLR(bp->b_oflags, BO_DONE);
    986 	mutex_exit(bp->b_objlock);
    987 
    988 	brelse(bp, 0);
    989 }
    990 
    991 /*
    992  * Asynchronous block write; just an asynchronous bwrite().
    993  */
    994 void
    995 bawrite(buf_t *bp)
    996 {
    997 
    998 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
    999 	KASSERT(bp->b_vp != NULL);
   1000 
   1001 	SET(bp->b_flags, B_ASYNC);
   1002 	VOP_BWRITE(bp->b_vp, bp);
   1003 }
   1004 
   1005 /*
   1006  * Release a buffer on to the free lists.
   1007  * Described in Bach (p. 46).
   1008  */
   1009 void
   1010 brelsel(buf_t *bp, int set)
   1011 {
   1012 	struct bqueue *bufq;
   1013 	struct vnode *vp;
   1014 
   1015 	KASSERT(bp != NULL);
   1016 	KASSERT(mutex_owned(&bufcache_lock));
   1017 	KASSERT(!cv_has_waiters(&bp->b_done));
   1018 	KASSERT(bp->b_refcnt > 0);
   1019 
   1020 	SET(bp->b_cflags, set);
   1021 
   1022 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
   1023 	KASSERT(bp->b_iodone == NULL);
   1024 
   1025 	/* Wake up any processes waiting for any buffer to become free. */
   1026 	cv_signal(&needbuffer_cv);
   1027 
   1028 	/* Wake up any proceeses waiting for _this_ buffer to become free */
   1029 	if (ISSET(bp->b_cflags, BC_WANTED))
   1030 		CLR(bp->b_cflags, BC_WANTED|BC_AGE);
   1031 
   1032 	/* If it's clean clear the copy-on-write flag. */
   1033 	if (ISSET(bp->b_flags, B_COWDONE)) {
   1034 		mutex_enter(bp->b_objlock);
   1035 		if (!ISSET(bp->b_oflags, BO_DELWRI))
   1036 			CLR(bp->b_flags, B_COWDONE);
   1037 		mutex_exit(bp->b_objlock);
   1038 	}
   1039 
   1040 	/*
   1041 	 * Determine which queue the buffer should be on, then put it there.
   1042 	 */
   1043 
   1044 	/* If it's locked, don't report an error; try again later. */
   1045 	if (ISSET(bp->b_flags, B_LOCKED))
   1046 		bp->b_error = 0;
   1047 
   1048 	/* If it's not cacheable, or an error, mark it invalid. */
   1049 	if (ISSET(bp->b_cflags, BC_NOCACHE) || bp->b_error != 0)
   1050 		SET(bp->b_cflags, BC_INVAL);
   1051 
   1052 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
   1053 		/*
   1054 		 * This is a delayed write buffer that was just flushed to
   1055 		 * disk.  It is still on the LRU queue.  If it's become
   1056 		 * invalid, then we need to move it to a different queue;
   1057 		 * otherwise leave it in its current position.
   1058 		 */
   1059 		CLR(bp->b_cflags, BC_VFLUSH);
   1060 		if (!ISSET(bp->b_cflags, BC_INVAL|BC_AGE) &&
   1061 		    !ISSET(bp->b_flags, B_LOCKED) && bp->b_error == 0) {
   1062 			KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 1));
   1063 			goto already_queued;
   1064 		} else {
   1065 			bremfree(bp);
   1066 		}
   1067 	}
   1068 
   1069 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_AGE], 0));
   1070 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 0));
   1071 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LOCKED], 0));
   1072 
   1073 	if ((bp->b_bufsize <= 0) || ISSET(bp->b_cflags, BC_INVAL)) {
   1074 		/*
   1075 		 * If it's invalid or empty, dissociate it from its vnode
   1076 		 * and put on the head of the appropriate queue.
   1077 		 */
   1078 		if (ISSET(bp->b_flags, B_LOCKED)) {
   1079 			if (wapbl_vphaswapbl(vp = bp->b_vp)) {
   1080 				struct mount *mp = wapbl_vptomp(vp);
   1081 
   1082 				KASSERT(bp->b_iodone
   1083 				    != mp->mnt_wapbl_op->wo_wapbl_biodone);
   1084 				WAPBL_REMOVE_BUF(mp, bp);
   1085 			}
   1086 		}
   1087 
   1088 		mutex_enter(bp->b_objlock);
   1089 		CLR(bp->b_oflags, BO_DONE|BO_DELWRI);
   1090 		if ((vp = bp->b_vp) != NULL) {
   1091 			KASSERT(bp->b_objlock == vp->v_interlock);
   1092 			reassignbuf(bp, bp->b_vp);
   1093 			brelvp(bp);
   1094 			mutex_exit(vp->v_interlock);
   1095 		} else {
   1096 			KASSERT(bp->b_objlock == &buffer_lock);
   1097 			mutex_exit(bp->b_objlock);
   1098 		}
   1099 
   1100 		if (bp->b_bufsize <= 0)
   1101 			/* no data */
   1102 			goto already_queued;
   1103 		else
   1104 			/* invalid data */
   1105 			bufq = &bufqueues[BQ_AGE];
   1106 		binsheadfree(bp, bufq);
   1107 	} else  {
   1108 		/*
   1109 		 * It has valid data.  Put it on the end of the appropriate
   1110 		 * queue, so that it'll stick around for as long as possible.
   1111 		 * If buf is AGE, but has dependencies, must put it on last
   1112 		 * bufqueue to be scanned, ie LRU. This protects against the
   1113 		 * livelock where BQ_AGE only has buffers with dependencies,
   1114 		 * and we thus never get to the dependent buffers in BQ_LRU.
   1115 		 */
   1116 		if (ISSET(bp->b_flags, B_LOCKED)) {
   1117 			/* locked in core */
   1118 			bufq = &bufqueues[BQ_LOCKED];
   1119 		} else if (!ISSET(bp->b_cflags, BC_AGE)) {
   1120 			/* valid data */
   1121 			bufq = &bufqueues[BQ_LRU];
   1122 		} else {
   1123 			/* stale but valid data */
   1124 			bufq = &bufqueues[BQ_AGE];
   1125 		}
   1126 		binstailfree(bp, bufq);
   1127 	}
   1128 already_queued:
   1129 	/* Unlock the buffer. */
   1130 	CLR(bp->b_cflags, BC_AGE|BC_BUSY|BC_NOCACHE);
   1131 	CLR(bp->b_flags, B_ASYNC);
   1132 	cv_broadcast(&bp->b_busy);
   1133 
   1134 	if (bp->b_bufsize <= 0)
   1135 		brele(bp);
   1136 }
   1137 
   1138 void
   1139 brelse(buf_t *bp, int set)
   1140 {
   1141 
   1142 	mutex_enter(&bufcache_lock);
   1143 	brelsel(bp, set);
   1144 	mutex_exit(&bufcache_lock);
   1145 }
   1146 
   1147 /*
   1148  * Determine if a block is in the cache.
   1149  * Just look on what would be its hash chain.  If it's there, return
   1150  * a pointer to it, unless it's marked invalid.  If it's marked invalid,
   1151  * we normally don't return the buffer, unless the caller explicitly
   1152  * wants us to.
   1153  */
   1154 buf_t *
   1155 incore(struct vnode *vp, daddr_t blkno)
   1156 {
   1157 	buf_t *bp;
   1158 
   1159 	KASSERT(mutex_owned(&bufcache_lock));
   1160 
   1161 	/* Search hash chain */
   1162 	LIST_FOREACH(bp, BUFHASH(vp, blkno), b_hash) {
   1163 		if (bp->b_lblkno == blkno && bp->b_vp == vp &&
   1164 		    !ISSET(bp->b_cflags, BC_INVAL)) {
   1165 		    	KASSERT(bp->b_objlock == vp->v_interlock);
   1166 		    	return (bp);
   1167 		}
   1168 	}
   1169 
   1170 	return (NULL);
   1171 }
   1172 
   1173 /*
   1174  * Get a block of requested size that is associated with
   1175  * a given vnode and block offset. If it is found in the
   1176  * block cache, mark it as having been found, make it busy
   1177  * and return it. Otherwise, return an empty block of the
   1178  * correct size. It is up to the caller to insure that the
   1179  * cached blocks be of the correct size.
   1180  */
   1181 buf_t *
   1182 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo)
   1183 {
   1184 	int err, preserve;
   1185 	buf_t *bp;
   1186 
   1187 	mutex_enter(&bufcache_lock);
   1188  loop:
   1189 	bp = incore(vp, blkno);
   1190 	if (bp != NULL) {
   1191 		err = bbusy(bp, ((slpflag & PCATCH) != 0), slptimeo, NULL);
   1192 		if (err != 0) {
   1193 			if (err == EPASSTHROUGH)
   1194 				goto loop;
   1195 			mutex_exit(&bufcache_lock);
   1196 			return (NULL);
   1197 		}
   1198 		KASSERT(!cv_has_waiters(&bp->b_done));
   1199 #ifdef DIAGNOSTIC
   1200 		if (ISSET(bp->b_oflags, BO_DONE|BO_DELWRI) &&
   1201 		    bp->b_bcount < size && vp->v_type != VBLK)
   1202 			panic("getblk: block size invariant failed");
   1203 #endif
   1204 		bremfree(bp);
   1205 		preserve = 1;
   1206 	} else {
   1207 		if ((bp = getnewbuf(slpflag, slptimeo, 0)) == NULL)
   1208 			goto loop;
   1209 
   1210 		if (incore(vp, blkno) != NULL) {
   1211 			/* The block has come into memory in the meantime. */
   1212 			brelsel(bp, 0);
   1213 			goto loop;
   1214 		}
   1215 
   1216 		LIST_INSERT_HEAD(BUFHASH(vp, blkno), bp, b_hash);
   1217 		bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno;
   1218 		mutex_enter(vp->v_interlock);
   1219 		bgetvp(vp, bp);
   1220 		mutex_exit(vp->v_interlock);
   1221 		preserve = 0;
   1222 	}
   1223 	mutex_exit(&bufcache_lock);
   1224 
   1225 	/*
   1226 	 * LFS can't track total size of B_LOCKED buffer (locked_queue_bytes)
   1227 	 * if we re-size buffers here.
   1228 	 */
   1229 	if (ISSET(bp->b_flags, B_LOCKED)) {
   1230 		KASSERT(bp->b_bufsize >= size);
   1231 	} else {
   1232 		if (allocbuf(bp, size, preserve)) {
   1233 			mutex_enter(&bufcache_lock);
   1234 			LIST_REMOVE(bp, b_hash);
   1235 			mutex_exit(&bufcache_lock);
   1236 			brelse(bp, BC_INVAL);
   1237 			return NULL;
   1238 		}
   1239 	}
   1240 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
   1241 	return (bp);
   1242 }
   1243 
   1244 /*
   1245  * Get an empty, disassociated buffer of given size.
   1246  */
   1247 buf_t *
   1248 geteblk(int size)
   1249 {
   1250 	buf_t *bp;
   1251 	int error __diagused;
   1252 
   1253 	mutex_enter(&bufcache_lock);
   1254 	while ((bp = getnewbuf(0, 0, 0)) == NULL)
   1255 		;
   1256 
   1257 	SET(bp->b_cflags, BC_INVAL);
   1258 	LIST_INSERT_HEAD(&invalhash, bp, b_hash);
   1259 	mutex_exit(&bufcache_lock);
   1260 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
   1261 	error = allocbuf(bp, size, 0);
   1262 	KASSERT(error == 0);
   1263 	return (bp);
   1264 }
   1265 
   1266 /*
   1267  * Expand or contract the actual memory allocated to a buffer.
   1268  *
   1269  * If the buffer shrinks, data is lost, so it's up to the
   1270  * caller to have written it out *first*; this routine will not
   1271  * start a write.  If the buffer grows, it's the callers
   1272  * responsibility to fill out the buffer's additional contents.
   1273  */
   1274 int
   1275 allocbuf(buf_t *bp, int size, int preserve)
   1276 {
   1277 	void *addr;
   1278 	vsize_t oldsize, desired_size;
   1279 	int oldcount;
   1280 	int delta;
   1281 
   1282 	desired_size = buf_roundsize(size);
   1283 	if (desired_size > MAXBSIZE)
   1284 		printf("allocbuf: buffer larger than MAXBSIZE requested");
   1285 
   1286 	oldcount = bp->b_bcount;
   1287 
   1288 	bp->b_bcount = size;
   1289 
   1290 	oldsize = bp->b_bufsize;
   1291 	if (oldsize == desired_size) {
   1292 		/*
   1293 		 * Do not short cut the WAPBL resize, as the buffer length
   1294 		 * could still have changed and this would corrupt the
   1295 		 * tracking of the transaction length.
   1296 		 */
   1297 		goto out;
   1298 	}
   1299 
   1300 	/*
   1301 	 * If we want a buffer of a different size, re-allocate the
   1302 	 * buffer's memory; copy old content only if needed.
   1303 	 */
   1304 	addr = buf_alloc(desired_size);
   1305 	if (addr == NULL)
   1306 		return ENOMEM;
   1307 	if (preserve)
   1308 		memcpy(addr, bp->b_data, MIN(oldsize,desired_size));
   1309 	if (bp->b_data != NULL)
   1310 		buf_mrelease(bp->b_data, oldsize);
   1311 	bp->b_data = addr;
   1312 	bp->b_bufsize = desired_size;
   1313 
   1314 	/*
   1315 	 * Update overall buffer memory counter (protected by bufcache_lock)
   1316 	 */
   1317 	delta = (long)desired_size - (long)oldsize;
   1318 
   1319 	mutex_enter(&bufcache_lock);
   1320 	if ((bufmem += delta) > bufmem_hiwater) {
   1321 		/*
   1322 		 * Need to trim overall memory usage.
   1323 		 */
   1324 		while (buf_canrelease()) {
   1325 			if (curcpu()->ci_schedstate.spc_flags &
   1326 			    SPCF_SHOULDYIELD) {
   1327 				mutex_exit(&bufcache_lock);
   1328 				preempt();
   1329 				mutex_enter(&bufcache_lock);
   1330 			}
   1331 			if (buf_trim() == 0)
   1332 				break;
   1333 		}
   1334 	}
   1335 	mutex_exit(&bufcache_lock);
   1336 
   1337  out:
   1338 	if (wapbl_vphaswapbl(bp->b_vp))
   1339 		WAPBL_RESIZE_BUF(wapbl_vptomp(bp->b_vp), bp, oldsize, oldcount);
   1340 
   1341 	return 0;
   1342 }
   1343 
   1344 /*
   1345  * Find a buffer which is available for use.
   1346  * Select something from a free list.
   1347  * Preference is to AGE list, then LRU list.
   1348  *
   1349  * Called with the buffer queues locked.
   1350  * Return buffer locked.
   1351  */
   1352 buf_t *
   1353 getnewbuf(int slpflag, int slptimeo, int from_bufq)
   1354 {
   1355 	buf_t *bp;
   1356 	struct vnode *vp;
   1357 
   1358  start:
   1359 	KASSERT(mutex_owned(&bufcache_lock));
   1360 
   1361 	/*
   1362 	 * Get a new buffer from the pool.
   1363 	 */
   1364 	if (!from_bufq && buf_lotsfree()) {
   1365 		mutex_exit(&bufcache_lock);
   1366 		bp = pool_cache_get(buf_cache, PR_NOWAIT);
   1367 		if (bp != NULL) {
   1368 			memset((char *)bp, 0, sizeof(*bp));
   1369 			buf_init(bp);
   1370 			SET(bp->b_cflags, BC_BUSY);	/* mark buffer busy */
   1371 			mutex_enter(&bufcache_lock);
   1372 #if defined(DIAGNOSTIC)
   1373 			bp->b_freelistindex = -1;
   1374 #endif /* defined(DIAGNOSTIC) */
   1375 			return (bp);
   1376 		}
   1377 		mutex_enter(&bufcache_lock);
   1378 	}
   1379 
   1380 	KASSERT(mutex_owned(&bufcache_lock));
   1381 	if ((bp = TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue)) != NULL ||
   1382 	    (bp = TAILQ_FIRST(&bufqueues[BQ_LRU].bq_queue)) != NULL) {
   1383 	    	KASSERT(!ISSET(bp->b_cflags, BC_BUSY) || ISSET(bp->b_cflags, BC_VFLUSH));
   1384 		bremfree(bp);
   1385 
   1386 		/* Buffer is no longer on free lists. */
   1387 		SET(bp->b_cflags, BC_BUSY);
   1388 	} else {
   1389 		/*
   1390 		 * XXX: !from_bufq should be removed.
   1391 		 */
   1392 		if (!from_bufq || curlwp != uvm.pagedaemon_lwp) {
   1393 			/* wait for a free buffer of any kind */
   1394 			if ((slpflag & PCATCH) != 0)
   1395 				(void)cv_timedwait_sig(&needbuffer_cv,
   1396 				    &bufcache_lock, slptimeo);
   1397 			else
   1398 				(void)cv_timedwait(&needbuffer_cv,
   1399 				    &bufcache_lock, slptimeo);
   1400 		}
   1401 		return (NULL);
   1402 	}
   1403 
   1404 #ifdef DIAGNOSTIC
   1405 	if (bp->b_bufsize <= 0)
   1406 		panic("buffer %p: on queue but empty", bp);
   1407 #endif
   1408 
   1409 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
   1410 		/*
   1411 		 * This is a delayed write buffer being flushed to disk.  Make
   1412 		 * sure it gets aged out of the queue when it's finished, and
   1413 		 * leave it off the LRU queue.
   1414 		 */
   1415 		CLR(bp->b_cflags, BC_VFLUSH);
   1416 		SET(bp->b_cflags, BC_AGE);
   1417 		goto start;
   1418 	}
   1419 
   1420 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
   1421 	KASSERT(bp->b_refcnt > 0);
   1422     	KASSERT(!cv_has_waiters(&bp->b_done));
   1423 
   1424 	/*
   1425 	 * If buffer was a delayed write, start it and return NULL
   1426 	 * (since we might sleep while starting the write).
   1427 	 */
   1428 	if (ISSET(bp->b_oflags, BO_DELWRI)) {
   1429 		/*
   1430 		 * This buffer has gone through the LRU, so make sure it gets
   1431 		 * reused ASAP.
   1432 		 */
   1433 		SET(bp->b_cflags, BC_AGE);
   1434 		mutex_exit(&bufcache_lock);
   1435 		bawrite(bp);
   1436 		mutex_enter(&bufcache_lock);
   1437 		return (NULL);
   1438 	}
   1439 
   1440 	vp = bp->b_vp;
   1441 
   1442 	/* clear out various other fields */
   1443 	bp->b_cflags = BC_BUSY;
   1444 	bp->b_oflags = 0;
   1445 	bp->b_flags = 0;
   1446 	bp->b_dev = NODEV;
   1447 	bp->b_blkno = 0;
   1448 	bp->b_lblkno = 0;
   1449 	bp->b_rawblkno = 0;
   1450 	bp->b_iodone = 0;
   1451 	bp->b_error = 0;
   1452 	bp->b_resid = 0;
   1453 	bp->b_bcount = 0;
   1454 
   1455 	LIST_REMOVE(bp, b_hash);
   1456 
   1457 	/* Disassociate us from our vnode, if we had one... */
   1458 	if (vp != NULL) {
   1459 		mutex_enter(vp->v_interlock);
   1460 		brelvp(bp);
   1461 		mutex_exit(vp->v_interlock);
   1462 	}
   1463 
   1464 	return (bp);
   1465 }
   1466 
   1467 /*
   1468  * Attempt to free an aged buffer off the queues.
   1469  * Called with queue lock held.
   1470  * Returns the amount of buffer memory freed.
   1471  */
   1472 static int
   1473 buf_trim(void)
   1474 {
   1475 	buf_t *bp;
   1476 	long size;
   1477 
   1478 	KASSERT(mutex_owned(&bufcache_lock));
   1479 
   1480 	/* Instruct getnewbuf() to get buffers off the queues */
   1481 	if ((bp = getnewbuf(PCATCH, 1, 1)) == NULL)
   1482 		return 0;
   1483 
   1484 	KASSERT((bp->b_cflags & BC_WANTED) == 0);
   1485 	size = bp->b_bufsize;
   1486 	bufmem -= size;
   1487 	if (size > 0) {
   1488 		buf_mrelease(bp->b_data, size);
   1489 		bp->b_bcount = bp->b_bufsize = 0;
   1490 	}
   1491 	/* brelse() will return the buffer to the global buffer pool */
   1492 	brelsel(bp, 0);
   1493 	return size;
   1494 }
   1495 
   1496 int
   1497 buf_drain(int n)
   1498 {
   1499 	int size = 0, sz;
   1500 
   1501 	KASSERT(mutex_owned(&bufcache_lock));
   1502 
   1503 	while (size < n && bufmem > bufmem_lowater) {
   1504 		sz = buf_trim();
   1505 		if (sz <= 0)
   1506 			break;
   1507 		size += sz;
   1508 	}
   1509 
   1510 	return size;
   1511 }
   1512 
   1513 SDT_PROVIDER_DEFINE(io);
   1514 
   1515 SDT_PROBE_DEFINE1(io, kernel, , wait__start, "struct buf *"/*bp*/);
   1516 SDT_PROBE_DEFINE1(io, kernel, , wait__done, "struct buf *"/*bp*/);
   1517 
   1518 /*
   1519  * Wait for operations on the buffer to complete.
   1520  * When they do, extract and return the I/O's error value.
   1521  */
   1522 int
   1523 biowait(buf_t *bp)
   1524 {
   1525 
   1526 	KERNHIST_FUNC(__func__);
   1527 
   1528 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
   1529 	KASSERT(bp->b_refcnt > 0);
   1530 
   1531 	SDT_PROBE1(io, kernel, , wait__start, bp);
   1532 
   1533 	mutex_enter(bp->b_objlock);
   1534 
   1535 	KERNHIST_CALLARGS(biohist, "bp=%p, oflags=0x%x",
   1536 	    bp, bp->b_oflags, 0, 0);
   1537 
   1538 	while (!ISSET(bp->b_oflags, BO_DONE | BO_DELWRI)) {
   1539 		KERNHIST_LOG(biohist, "waiting bp=%p", bp, 0, 0, 0);
   1540 		cv_wait(&bp->b_done, bp->b_objlock);
   1541 	}
   1542 	mutex_exit(bp->b_objlock);
   1543 
   1544 	SDT_PROBE1(io, kernel, , wait__done, bp);
   1545 
   1546 	KERNHIST_LOG(biohist, "  return %d", bp->b_error, 0, 0, 0);
   1547 
   1548 	return bp->b_error;
   1549 }
   1550 
   1551 /*
   1552  * Mark I/O complete on a buffer.
   1553  *
   1554  * If a callback has been requested, e.g. the pageout
   1555  * daemon, do so. Otherwise, awaken waiting processes.
   1556  *
   1557  * [ Leffler, et al., says on p.247:
   1558  *	"This routine wakes up the blocked process, frees the buffer
   1559  *	for an asynchronous write, or, for a request by the pagedaemon
   1560  *	process, invokes a procedure specified in the buffer structure" ]
   1561  *
   1562  * In real life, the pagedaemon (or other system processes) wants
   1563  * to do async stuff too, and doesn't want the buffer brelse()'d.
   1564  * (for swap pager, that puts swap buffers on the free lists (!!!),
   1565  * for the vn device, that puts allocated buffers on the free lists!)
   1566  */
   1567 void
   1568 biodone(buf_t *bp)
   1569 {
   1570 	int s;
   1571 
   1572 	KERNHIST_FUNC(__func__);
   1573 
   1574 	KASSERT(!ISSET(bp->b_oflags, BO_DONE));
   1575 
   1576 	if (cpu_intr_p()) {
   1577 		/* From interrupt mode: defer to a soft interrupt. */
   1578 		s = splvm();
   1579 		TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_biodone, bp, b_actq);
   1580 
   1581 		KERNHIST_CALLARGS(biohist, "bp=%p, softint scheduled",
   1582 		    bp, 0, 0, 0);
   1583 		softint_schedule(biodone_sih);
   1584 		splx(s);
   1585 	} else {
   1586 		/* Process now - the buffer may be freed soon. */
   1587 		biodone2(bp);
   1588 	}
   1589 }
   1590 
   1591 SDT_PROBE_DEFINE1(io, kernel, , done, "struct buf *"/*bp*/);
   1592 
   1593 static void
   1594 biodone2(buf_t *bp)
   1595 {
   1596 	void (*callout)(buf_t *);
   1597 
   1598 	SDT_PROBE1(io, kernel, ,done, bp);
   1599 
   1600 	KERNHIST_FUNC(__func__);
   1601 	KERNHIST_CALLARGS(biohist, "bp=%p", bp, 0, 0, 0);
   1602 
   1603 	mutex_enter(bp->b_objlock);
   1604 	/* Note that the transfer is done. */
   1605 	if (ISSET(bp->b_oflags, BO_DONE))
   1606 		panic("biodone2 already");
   1607 	CLR(bp->b_flags, B_COWDONE);
   1608 	SET(bp->b_oflags, BO_DONE);
   1609 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
   1610 
   1611 	/* Wake up waiting writers. */
   1612 	if (!ISSET(bp->b_flags, B_READ))
   1613 		vwakeup(bp);
   1614 
   1615 	if ((callout = bp->b_iodone) != NULL) {
   1616 		KERNHIST_LOG(biohist, "callout %p", callout, 0, 0, 0);
   1617 
   1618 		/* Note callout done, then call out. */
   1619 		KASSERT(!cv_has_waiters(&bp->b_done));
   1620 		KERNEL_LOCK(1, NULL);		/* XXXSMP */
   1621 		bp->b_iodone = NULL;
   1622 		mutex_exit(bp->b_objlock);
   1623 		(*callout)(bp);
   1624 		KERNEL_UNLOCK_ONE(NULL);	/* XXXSMP */
   1625 	} else if (ISSET(bp->b_flags, B_ASYNC)) {
   1626 		/* If async, release. */
   1627 		KERNHIST_LOG(biohist, "async", 0, 0, 0, 0);
   1628 		KASSERT(!cv_has_waiters(&bp->b_done));
   1629 		mutex_exit(bp->b_objlock);
   1630 		brelse(bp, 0);
   1631 	} else {
   1632 		/* Otherwise just wake up waiters in biowait(). */
   1633 		KERNHIST_LOG(biohist, "wake-up", 0, 0, 0, 0);
   1634 		cv_broadcast(&bp->b_done);
   1635 		mutex_exit(bp->b_objlock);
   1636 	}
   1637 }
   1638 
   1639 static void
   1640 biointr(void *cookie)
   1641 {
   1642 	struct cpu_info *ci;
   1643 	buf_t *bp;
   1644 	int s;
   1645 
   1646 	KERNHIST_FUNC(__func__); KERNHIST_CALLED(biohist);
   1647 
   1648 	ci = curcpu();
   1649 
   1650 	s = splvm();
   1651 	while (!TAILQ_EMPTY(&ci->ci_data.cpu_biodone)) {
   1652 		KASSERT(curcpu() == ci);
   1653 
   1654 		bp = TAILQ_FIRST(&ci->ci_data.cpu_biodone);
   1655 		TAILQ_REMOVE(&ci->ci_data.cpu_biodone, bp, b_actq);
   1656 		splx(s);
   1657 
   1658 		KERNHIST_LOG(biohist, "bp=%p", bp, 0, 0, 0);
   1659 		biodone2(bp);
   1660 
   1661 		s = splvm();
   1662 	}
   1663 	splx(s);
   1664 }
   1665 
   1666 /*
   1667  * Wait for all buffers to complete I/O
   1668  * Return the number of "stuck" buffers.
   1669  */
   1670 int
   1671 buf_syncwait(void)
   1672 {
   1673 	buf_t *bp;
   1674 	int iter, nbusy, nbusy_prev = 0, ihash;
   1675 
   1676 	KERNHIST_FUNC(__func__); KERNHIST_CALLED(biohist);
   1677 
   1678 	for (iter = 0; iter < 20;) {
   1679 		mutex_enter(&bufcache_lock);
   1680 		nbusy = 0;
   1681 		for (ihash = 0; ihash < bufhash+1; ihash++) {
   1682 		    LIST_FOREACH(bp, &bufhashtbl[ihash], b_hash) {
   1683 			if ((bp->b_cflags & (BC_BUSY|BC_INVAL)) == BC_BUSY)
   1684 				nbusy += ((bp->b_flags & B_READ) == 0);
   1685 		    }
   1686 		}
   1687 		mutex_exit(&bufcache_lock);
   1688 
   1689 		if (nbusy == 0)
   1690 			break;
   1691 		if (nbusy_prev == 0)
   1692 			nbusy_prev = nbusy;
   1693 		printf("%d ", nbusy);
   1694 		kpause("bflush", false, MAX(1, hz / 25 * iter), NULL);
   1695 		if (nbusy >= nbusy_prev) /* we didn't flush anything */
   1696 			iter++;
   1697 		else
   1698 			nbusy_prev = nbusy;
   1699 	}
   1700 
   1701 	if (nbusy) {
   1702 #if defined(DEBUG) || defined(DEBUG_HALT_BUSY)
   1703 		printf("giving up\nPrinting vnodes for busy buffers\n");
   1704 		for (ihash = 0; ihash < bufhash+1; ihash++) {
   1705 		    LIST_FOREACH(bp, &bufhashtbl[ihash], b_hash) {
   1706 			if ((bp->b_cflags & (BC_BUSY|BC_INVAL)) == BC_BUSY &&
   1707 			    (bp->b_flags & B_READ) == 0)
   1708 				vprint(NULL, bp->b_vp);
   1709 		    }
   1710 		}
   1711 #endif
   1712 	}
   1713 
   1714 	return nbusy;
   1715 }
   1716 
   1717 static void
   1718 sysctl_fillbuf(buf_t *i, struct buf_sysctl *o)
   1719 {
   1720 
   1721 	o->b_flags = i->b_flags | i->b_cflags | i->b_oflags;
   1722 	o->b_error = i->b_error;
   1723 	o->b_prio = i->b_prio;
   1724 	o->b_dev = i->b_dev;
   1725 	o->b_bufsize = i->b_bufsize;
   1726 	o->b_bcount = i->b_bcount;
   1727 	o->b_resid = i->b_resid;
   1728 	o->b_addr = PTRTOUINT64(i->b_data);
   1729 	o->b_blkno = i->b_blkno;
   1730 	o->b_rawblkno = i->b_rawblkno;
   1731 	o->b_iodone = PTRTOUINT64(i->b_iodone);
   1732 	o->b_proc = PTRTOUINT64(i->b_proc);
   1733 	o->b_vp = PTRTOUINT64(i->b_vp);
   1734 	o->b_saveaddr = PTRTOUINT64(i->b_saveaddr);
   1735 	o->b_lblkno = i->b_lblkno;
   1736 }
   1737 
   1738 #define KERN_BUFSLOP 20
   1739 static int
   1740 sysctl_dobuf(SYSCTLFN_ARGS)
   1741 {
   1742 	buf_t *bp;
   1743 	struct buf_sysctl bs;
   1744 	struct bqueue *bq;
   1745 	char *dp;
   1746 	u_int i, op, arg;
   1747 	size_t len, needed, elem_size, out_size;
   1748 	int error, elem_count, retries;
   1749 
   1750 	if (namelen == 1 && name[0] == CTL_QUERY)
   1751 		return (sysctl_query(SYSCTLFN_CALL(rnode)));
   1752 
   1753 	if (namelen != 4)
   1754 		return (EINVAL);
   1755 
   1756 	retries = 100;
   1757  retry:
   1758 	dp = oldp;
   1759 	len = (oldp != NULL) ? *oldlenp : 0;
   1760 	op = name[0];
   1761 	arg = name[1];
   1762 	elem_size = name[2];
   1763 	elem_count = name[3];
   1764 	out_size = MIN(sizeof(bs), elem_size);
   1765 
   1766 	/*
   1767 	 * at the moment, these are just "placeholders" to make the
   1768 	 * API for retrieving kern.buf data more extensible in the
   1769 	 * future.
   1770 	 *
   1771 	 * XXX kern.buf currently has "netbsd32" issues.  hopefully
   1772 	 * these will be resolved at a later point.
   1773 	 */
   1774 	if (op != KERN_BUF_ALL || arg != KERN_BUF_ALL ||
   1775 	    elem_size < 1 || elem_count < 0)
   1776 		return (EINVAL);
   1777 
   1778 	error = 0;
   1779 	needed = 0;
   1780 	sysctl_unlock();
   1781 	mutex_enter(&bufcache_lock);
   1782 	for (i = 0; i < BQUEUES; i++) {
   1783 		bq = &bufqueues[i];
   1784 		TAILQ_FOREACH(bp, &bq->bq_queue, b_freelist) {
   1785 			bq->bq_marker = bp;
   1786 			if (len >= elem_size && elem_count > 0) {
   1787 				sysctl_fillbuf(bp, &bs);
   1788 				mutex_exit(&bufcache_lock);
   1789 				error = copyout(&bs, dp, out_size);
   1790 				mutex_enter(&bufcache_lock);
   1791 				if (error)
   1792 					break;
   1793 				if (bq->bq_marker != bp) {
   1794 					/*
   1795 					 * This sysctl node is only for
   1796 					 * statistics.  Retry; if the
   1797 					 * queue keeps changing, then
   1798 					 * bail out.
   1799 					 */
   1800 					if (retries-- == 0) {
   1801 						error = EAGAIN;
   1802 						break;
   1803 					}
   1804 					mutex_exit(&bufcache_lock);
   1805 					sysctl_relock();
   1806 					goto retry;
   1807 				}
   1808 				dp += elem_size;
   1809 				len -= elem_size;
   1810 			}
   1811 			needed += elem_size;
   1812 			if (elem_count > 0 && elem_count != INT_MAX)
   1813 				elem_count--;
   1814 		}
   1815 		if (error != 0)
   1816 			break;
   1817 	}
   1818 	mutex_exit(&bufcache_lock);
   1819 	sysctl_relock();
   1820 
   1821 	*oldlenp = needed;
   1822 	if (oldp == NULL)
   1823 		*oldlenp += KERN_BUFSLOP * sizeof(buf_t);
   1824 
   1825 	return (error);
   1826 }
   1827 
   1828 static int
   1829 sysctl_bufvm_update(SYSCTLFN_ARGS)
   1830 {
   1831 	int error, rv;
   1832 	struct sysctlnode node;
   1833 	unsigned int temp_bufcache;
   1834 	unsigned long temp_water;
   1835 
   1836 	/* Take a copy of the supplied node and its data */
   1837 	node = *rnode;
   1838 	if (node.sysctl_data == &bufcache) {
   1839 	    node.sysctl_data = &temp_bufcache;
   1840 	    temp_bufcache = *(unsigned int *)rnode->sysctl_data;
   1841 	} else {
   1842 	    node.sysctl_data = &temp_water;
   1843 	    temp_water = *(unsigned long *)rnode->sysctl_data;
   1844 	}
   1845 
   1846 	/* Update the copy */
   1847 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
   1848 	if (error || newp == NULL)
   1849 		return (error);
   1850 
   1851 	if (rnode->sysctl_data == &bufcache) {
   1852 		if (temp_bufcache > 100)
   1853 			return (EINVAL);
   1854 		bufcache = temp_bufcache;
   1855 		buf_setwm();
   1856 	} else if (rnode->sysctl_data == &bufmem_lowater) {
   1857 		if (bufmem_hiwater - temp_water < 16)
   1858 			return (EINVAL);
   1859 		bufmem_lowater = temp_water;
   1860 	} else if (rnode->sysctl_data == &bufmem_hiwater) {
   1861 		if (temp_water - bufmem_lowater < 16)
   1862 			return (EINVAL);
   1863 		bufmem_hiwater = temp_water;
   1864 	} else
   1865 		return (EINVAL);
   1866 
   1867 	/* Drain until below new high water mark */
   1868 	sysctl_unlock();
   1869 	mutex_enter(&bufcache_lock);
   1870 	while (bufmem > bufmem_hiwater) {
   1871 		rv = buf_drain((bufmem - bufmem_hiwater) / (2 * 1024));
   1872 		if (rv <= 0)
   1873 			break;
   1874 	}
   1875 	mutex_exit(&bufcache_lock);
   1876 	sysctl_relock();
   1877 
   1878 	return 0;
   1879 }
   1880 
   1881 static struct sysctllog *vfsbio_sysctllog;
   1882 
   1883 static void
   1884 sysctl_kern_buf_setup(void)
   1885 {
   1886 
   1887 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
   1888 		       CTLFLAG_PERMANENT,
   1889 		       CTLTYPE_NODE, "buf",
   1890 		       SYSCTL_DESCR("Kernel buffer cache information"),
   1891 		       sysctl_dobuf, 0, NULL, 0,
   1892 		       CTL_KERN, KERN_BUF, CTL_EOL);
   1893 }
   1894 
   1895 static void
   1896 sysctl_vm_buf_setup(void)
   1897 {
   1898 
   1899 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
   1900 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
   1901 		       CTLTYPE_INT, "bufcache",
   1902 		       SYSCTL_DESCR("Percentage of physical memory to use for "
   1903 				    "buffer cache"),
   1904 		       sysctl_bufvm_update, 0, &bufcache, 0,
   1905 		       CTL_VM, CTL_CREATE, CTL_EOL);
   1906 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
   1907 		       CTLFLAG_PERMANENT|CTLFLAG_READONLY,
   1908 		       CTLTYPE_LONG, "bufmem",
   1909 		       SYSCTL_DESCR("Amount of kernel memory used by buffer "
   1910 				    "cache"),
   1911 		       NULL, 0, &bufmem, 0,
   1912 		       CTL_VM, CTL_CREATE, CTL_EOL);
   1913 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
   1914 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
   1915 		       CTLTYPE_LONG, "bufmem_lowater",
   1916 		       SYSCTL_DESCR("Minimum amount of kernel memory to "
   1917 				    "reserve for buffer cache"),
   1918 		       sysctl_bufvm_update, 0, &bufmem_lowater, 0,
   1919 		       CTL_VM, CTL_CREATE, CTL_EOL);
   1920 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
   1921 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
   1922 		       CTLTYPE_LONG, "bufmem_hiwater",
   1923 		       SYSCTL_DESCR("Maximum amount of kernel memory to use "
   1924 				    "for buffer cache"),
   1925 		       sysctl_bufvm_update, 0, &bufmem_hiwater, 0,
   1926 		       CTL_VM, CTL_CREATE, CTL_EOL);
   1927 }
   1928 
   1929 #ifdef DEBUG
   1930 /*
   1931  * Print out statistics on the current allocation of the buffer pool.
   1932  * Can be enabled to print out on every ``sync'' by setting "syncprt"
   1933  * in vfs_syscalls.c using sysctl.
   1934  */
   1935 void
   1936 vfs_bufstats(void)
   1937 {
   1938 	int i, j, count;
   1939 	buf_t *bp;
   1940 	struct bqueue *dp;
   1941 	int counts[MAXBSIZE / MIN_PAGE_SIZE + 1];
   1942 	static const char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE" };
   1943 
   1944 	for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) {
   1945 		count = 0;
   1946 		memset(counts, 0, sizeof(counts));
   1947 		TAILQ_FOREACH(bp, &dp->bq_queue, b_freelist) {
   1948 			counts[bp->b_bufsize / PAGE_SIZE]++;
   1949 			count++;
   1950 		}
   1951 		printf("%s: total-%d", bname[i], count);
   1952 		for (j = 0; j <= MAXBSIZE / PAGE_SIZE; j++)
   1953 			if (counts[j] != 0)
   1954 				printf(", %d-%d", j * PAGE_SIZE, counts[j]);
   1955 		printf("\n");
   1956 	}
   1957 }
   1958 #endif /* DEBUG */
   1959 
   1960 /* ------------------------------ */
   1961 
   1962 buf_t *
   1963 getiobuf(struct vnode *vp, bool waitok)
   1964 {
   1965 	buf_t *bp;
   1966 
   1967 	bp = pool_cache_get(bufio_cache, (waitok ? PR_WAITOK : PR_NOWAIT));
   1968 	if (bp == NULL)
   1969 		return bp;
   1970 
   1971 	buf_init(bp);
   1972 
   1973 	if ((bp->b_vp = vp) == NULL)
   1974 		bp->b_objlock = &buffer_lock;
   1975 	else
   1976 		bp->b_objlock = vp->v_interlock;
   1977 
   1978 	return bp;
   1979 }
   1980 
   1981 void
   1982 putiobuf(buf_t *bp)
   1983 {
   1984 
   1985 	buf_destroy(bp);
   1986 	pool_cache_put(bufio_cache, bp);
   1987 }
   1988 
   1989 /*
   1990  * nestiobuf_iodone: b_iodone callback for nested buffers.
   1991  */
   1992 
   1993 void
   1994 nestiobuf_iodone(buf_t *bp)
   1995 {
   1996 	buf_t *mbp = bp->b_private;
   1997 	int error;
   1998 	int donebytes;
   1999 
   2000 	KASSERT(bp->b_bcount <= bp->b_bufsize);
   2001 	KASSERT(mbp != bp);
   2002 
   2003 	error = bp->b_error;
   2004 	if (bp->b_error == 0 &&
   2005 	    (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
   2006 		/*
   2007 		 * Not all got transfered, raise an error. We have no way to
   2008 		 * propagate these conditions to mbp.
   2009 		 */
   2010 		error = EIO;
   2011 	}
   2012 
   2013 	donebytes = bp->b_bufsize;
   2014 
   2015 	putiobuf(bp);
   2016 	nestiobuf_done(mbp, donebytes, error);
   2017 }
   2018 
   2019 /*
   2020  * nestiobuf_setup: setup a "nested" buffer.
   2021  *
   2022  * => 'mbp' is a "master" buffer which is being divided into sub pieces.
   2023  * => 'bp' should be a buffer allocated by getiobuf.
   2024  * => 'offset' is a byte offset in the master buffer.
   2025  * => 'size' is a size in bytes of this nested buffer.
   2026  */
   2027 
   2028 void
   2029 nestiobuf_setup(buf_t *mbp, buf_t *bp, int offset, size_t size)
   2030 {
   2031 	const int b_read = mbp->b_flags & B_READ;
   2032 	struct vnode *vp = mbp->b_vp;
   2033 
   2034 	KASSERT(mbp->b_bcount >= offset + size);
   2035 	bp->b_vp = vp;
   2036 	bp->b_dev = mbp->b_dev;
   2037 	bp->b_objlock = mbp->b_objlock;
   2038 	bp->b_cflags = BC_BUSY;
   2039 	bp->b_flags = B_ASYNC | b_read;
   2040 	bp->b_iodone = nestiobuf_iodone;
   2041 	bp->b_data = (char *)mbp->b_data + offset;
   2042 	bp->b_resid = bp->b_bcount = size;
   2043 	bp->b_bufsize = bp->b_bcount;
   2044 	bp->b_private = mbp;
   2045 	BIO_COPYPRIO(bp, mbp);
   2046 	if (!b_read && vp != NULL) {
   2047 		mutex_enter(vp->v_interlock);
   2048 		vp->v_numoutput++;
   2049 		mutex_exit(vp->v_interlock);
   2050 	}
   2051 }
   2052 
   2053 /*
   2054  * nestiobuf_done: propagate completion to the master buffer.
   2055  *
   2056  * => 'donebytes' specifies how many bytes in the 'mbp' is completed.
   2057  * => 'error' is an errno(2) that 'donebytes' has been completed with.
   2058  */
   2059 
   2060 void
   2061 nestiobuf_done(buf_t *mbp, int donebytes, int error)
   2062 {
   2063 
   2064 	if (donebytes == 0) {
   2065 		return;
   2066 	}
   2067 	mutex_enter(mbp->b_objlock);
   2068 	KASSERT(mbp->b_resid >= donebytes);
   2069 	mbp->b_resid -= donebytes;
   2070 	if (error)
   2071 		mbp->b_error = error;
   2072 	if (mbp->b_resid == 0) {
   2073 		if (mbp->b_error)
   2074 			mbp->b_resid = mbp->b_bcount;
   2075 		mutex_exit(mbp->b_objlock);
   2076 		biodone(mbp);
   2077 	} else
   2078 		mutex_exit(mbp->b_objlock);
   2079 }
   2080 
   2081 void
   2082 buf_init(buf_t *bp)
   2083 {
   2084 
   2085 	cv_init(&bp->b_busy, "biolock");
   2086 	cv_init(&bp->b_done, "biowait");
   2087 	bp->b_dev = NODEV;
   2088 	bp->b_error = 0;
   2089 	bp->b_flags = 0;
   2090 	bp->b_cflags = 0;
   2091 	bp->b_oflags = 0;
   2092 	bp->b_objlock = &buffer_lock;
   2093 	bp->b_iodone = NULL;
   2094 	bp->b_refcnt = 1;
   2095 	bp->b_dev = NODEV;
   2096 	bp->b_vnbufs.le_next = NOLIST;
   2097 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
   2098 }
   2099 
   2100 void
   2101 buf_destroy(buf_t *bp)
   2102 {
   2103 
   2104 	cv_destroy(&bp->b_done);
   2105 	cv_destroy(&bp->b_busy);
   2106 }
   2107 
   2108 int
   2109 bbusy(buf_t *bp, bool intr, int timo, kmutex_t *interlock)
   2110 {
   2111 	int error;
   2112 
   2113 	KASSERT(mutex_owned(&bufcache_lock));
   2114 
   2115 	if ((bp->b_cflags & BC_BUSY) != 0) {
   2116 		if (curlwp == uvm.pagedaemon_lwp)
   2117 			return EDEADLK;
   2118 		bp->b_cflags |= BC_WANTED;
   2119 		bref(bp);
   2120 		if (interlock != NULL)
   2121 			mutex_exit(interlock);
   2122 		if (intr) {
   2123 			error = cv_timedwait_sig(&bp->b_busy, &bufcache_lock,
   2124 			    timo);
   2125 		} else {
   2126 			error = cv_timedwait(&bp->b_busy, &bufcache_lock,
   2127 			    timo);
   2128 		}
   2129 		brele(bp);
   2130 		if (interlock != NULL)
   2131 			mutex_enter(interlock);
   2132 		if (error != 0)
   2133 			return error;
   2134 		return EPASSTHROUGH;
   2135 	}
   2136 	bp->b_cflags |= BC_BUSY;
   2137 
   2138 	return 0;
   2139 }
   2140