xref: /freebsd/sys/kern/vfs_bio.c (revision 74fe6c29fb7eef3418d7919dcd41dc1a04a982a1)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2004 Poul-Henning Kamp
5  * Copyright (c) 1994,1997 John S. Dyson
6  * Copyright (c) 2013 The FreeBSD Foundation
7  * All rights reserved.
8  *
9  * Portions of this software were developed by Konstantin Belousov
10  * under sponsorship from the FreeBSD Foundation.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * this file contains a new buffer I/O scheme implementing a coherent
36  * VM object and buffer cache scheme.  Pains have been taken to make
37  * sure that the performance degradation associated with schemes such
38  * as this is not realized.
39  *
40  * Author:  John S. Dyson
41  * Significant help during the development and debugging phases
42  * had been provided by David Greenman, also of the FreeBSD core team.
43  *
44  * see man buf(9) for more info.
45  */
46 
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/bio.h>
53 #include <sys/bitset.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/buf.h>
57 #include <sys/devicestat.h>
58 #include <sys/eventhandler.h>
59 #include <sys/fail.h>
60 #include <sys/limits.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/mount.h>
64 #include <sys/mutex.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/proc.h>
68 #include <sys/racct.h>
69 #include <sys/resourcevar.h>
70 #include <sys/rwlock.h>
71 #include <sys/smp.h>
72 #include <sys/sysctl.h>
73 #include <sys/sysproto.h>
74 #include <sys/vmem.h>
75 #include <sys/vmmeter.h>
76 #include <sys/vnode.h>
77 #include <sys/watchdog.h>
78 #include <geom/geom.h>
79 #include <vm/vm.h>
80 #include <vm/vm_param.h>
81 #include <vm/vm_kern.h>
82 #include <vm/vm_object.h>
83 #include <vm/vm_page.h>
84 #include <vm/vm_pageout.h>
85 #include <vm/vm_pager.h>
86 #include <vm/vm_extern.h>
87 #include <vm/vm_map.h>
88 #include <vm/swap_pager.h>
89 #include "opt_compat.h"
90 #include "opt_swap.h"
91 
92 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
93 
94 struct	bio_ops bioops;		/* I/O operation notification */
95 
96 struct	buf_ops buf_ops_bio = {
97 	.bop_name	=	"buf_ops_bio",
98 	.bop_write	=	bufwrite,
99 	.bop_strategy	=	bufstrategy,
100 	.bop_sync	=	bufsync,
101 	.bop_bdflush	=	bufbdflush,
102 };
103 
104 struct bufqueue {
105 	struct mtx_padalign	bq_lock;
106 	TAILQ_HEAD(, buf)	bq_queue;
107 	uint8_t			bq_index;
108 	uint16_t		bq_subqueue;
109 	int			bq_len;
110 } __aligned(CACHE_LINE_SIZE);
111 
112 #define	BQ_LOCKPTR(bq)		(&(bq)->bq_lock)
113 #define	BQ_LOCK(bq)		mtx_lock(BQ_LOCKPTR((bq)))
114 #define	BQ_UNLOCK(bq)		mtx_unlock(BQ_LOCKPTR((bq)))
115 #define	BQ_ASSERT_LOCKED(bq)	mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
116 
117 struct bufdomain {
118 	struct bufqueue	bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
119 	struct bufqueue bd_dirtyq;
120 	struct bufqueue	*bd_cleanq;
121 	struct mtx_padalign bd_run_lock;
122 	/* Constants */
123 	long		bd_maxbufspace;
124 	long		bd_hibufspace;
125 	long 		bd_lobufspace;
126 	long 		bd_bufspacethresh;
127 	int		bd_hifreebuffers;
128 	int		bd_lofreebuffers;
129 	int		bd_hidirtybuffers;
130 	int		bd_lodirtybuffers;
131 	int		bd_dirtybufthresh;
132 	int		bd_lim;
133 	/* atomics */
134 	int		bd_wanted;
135 	int __aligned(CACHE_LINE_SIZE)	bd_numdirtybuffers;
136 	int __aligned(CACHE_LINE_SIZE)	bd_running;
137 	long __aligned(CACHE_LINE_SIZE) bd_bufspace;
138 	int __aligned(CACHE_LINE_SIZE)	bd_freebuffers;
139 } __aligned(CACHE_LINE_SIZE);
140 
141 #define	BD_LOCKPTR(bd)		(&(bd)->bd_cleanq->bq_lock)
142 #define	BD_LOCK(bd)		mtx_lock(BD_LOCKPTR((bd)))
143 #define	BD_UNLOCK(bd)		mtx_unlock(BD_LOCKPTR((bd)))
144 #define	BD_ASSERT_LOCKED(bd)	mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
145 #define	BD_RUN_LOCKPTR(bd)	(&(bd)->bd_run_lock)
146 #define	BD_RUN_LOCK(bd)		mtx_lock(BD_RUN_LOCKPTR((bd)))
147 #define	BD_RUN_UNLOCK(bd)	mtx_unlock(BD_RUN_LOCKPTR((bd)))
148 #define	BD_DOMAIN(bd)		(bd - bdomain)
149 
150 static struct buf *buf;		/* buffer header pool */
151 extern struct buf *swbuf;	/* Swap buffer header pool. */
152 caddr_t unmapped_buf;
153 
154 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
155 struct proc *bufdaemonproc;
156 
157 static int inmem(struct vnode *vp, daddr_t blkno);
158 static void vm_hold_free_pages(struct buf *bp, int newbsize);
159 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
160 		vm_offset_t to);
161 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
162 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
163 		vm_page_t m);
164 static void vfs_clean_pages_dirty_buf(struct buf *bp);
165 static void vfs_setdirty_locked_object(struct buf *bp);
166 static void vfs_vmio_invalidate(struct buf *bp);
167 static void vfs_vmio_truncate(struct buf *bp, int npages);
168 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
169 static int vfs_bio_clcheck(struct vnode *vp, int size,
170 		daddr_t lblkno, daddr_t blkno);
171 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
172 		void (*)(struct buf *));
173 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
174 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
175 static void buf_daemon(void);
176 static __inline void bd_wakeup(void);
177 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
178 static void bufkva_reclaim(vmem_t *, int);
179 static void bufkva_free(struct buf *);
180 static int buf_import(void *, void **, int, int, int);
181 static void buf_release(void *, void **, int);
182 static void maxbcachebuf_adjust(void);
183 static inline struct bufdomain *bufdomain(struct buf *);
184 static void bq_remove(struct bufqueue *bq, struct buf *bp);
185 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
186 static int buf_recycle(struct bufdomain *, bool kva);
187 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
188 	    const char *lockname);
189 static void bd_init(struct bufdomain *bd);
190 static int bd_flushall(struct bufdomain *bd);
191 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
192 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
193 
194 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
195 int vmiodirenable = TRUE;
196 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
197     "Use the VM system for directory writes");
198 long runningbufspace;
199 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
200     "Amount of presently outstanding async buffer io");
201 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
202     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
203 static counter_u64_t bufkvaspace;
204 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
205     "Kernel virtual memory used for buffers");
206 static long maxbufspace;
207 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
208     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
209     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
210     "Maximum allowed value of bufspace (including metadata)");
211 static long bufmallocspace;
212 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
213     "Amount of malloced memory for buffers");
214 static long maxbufmallocspace;
215 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
216     0, "Maximum amount of malloced memory for buffers");
217 static long lobufspace;
218 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
219     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
220     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
221     "Minimum amount of buffers we want to have");
222 long hibufspace;
223 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
224     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
225     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
226     "Maximum allowed value of bufspace (excluding metadata)");
227 long bufspacethresh;
228 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
229     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
230     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
231     "Bufspace consumed before waking the daemon to free some");
232 static counter_u64_t buffreekvacnt;
233 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
234     "Number of times we have freed the KVA space from some buffer");
235 static counter_u64_t bufdefragcnt;
236 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
237     "Number of times we have had to repeat buffer allocation to defragment");
238 static long lorunningspace;
239 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
240     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
241     "Minimum preferred space used for in-progress I/O");
242 static long hirunningspace;
243 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
244     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
245     "Maximum amount of space to use for in-progress I/O");
246 int dirtybufferflushes;
247 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
248     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
249 int bdwriteskip;
250 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
251     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
252 int altbufferflushes;
253 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
254     0, "Number of fsync flushes to limit dirty buffers");
255 static int recursiveflushes;
256 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
257     0, "Number of flushes skipped due to being recursive");
258 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
259 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
260     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
261     "Number of buffers that are dirty (has unwritten changes) at the moment");
262 static int lodirtybuffers;
263 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
264     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
265     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "L",
266     "How many buffers we want to have free before bufdaemon can sleep");
267 static int hidirtybuffers;
268 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
269     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
270     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "L",
271     "When the number of dirty buffers is considered severe");
272 int dirtybufthresh;
273 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
274     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
275     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "L",
276     "Number of bdwrite to bawrite conversions to clear dirty buffers");
277 static int numfreebuffers;
278 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
279     "Number of free buffers");
280 static int lofreebuffers;
281 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
282     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
283     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "L",
284    "Target number of free buffers");
285 static int hifreebuffers;
286 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
287     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
288     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "L",
289    "Threshold for clean buffer recycling");
290 static counter_u64_t getnewbufcalls;
291 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
292    &getnewbufcalls, "Number of calls to getnewbuf");
293 static counter_u64_t getnewbufrestarts;
294 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
295     &getnewbufrestarts,
296     "Number of times getnewbuf has had to restart a buffer acquisition");
297 static counter_u64_t mappingrestarts;
298 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
299     &mappingrestarts,
300     "Number of times getblk has had to restart a buffer mapping for "
301     "unmapped buffer");
302 static counter_u64_t numbufallocfails;
303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
304     &numbufallocfails, "Number of times buffer allocations failed");
305 static int flushbufqtarget = 100;
306 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
307     "Amount of work to do in flushbufqueues when helping bufdaemon");
308 static counter_u64_t notbufdflushes;
309 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
310     "Number of dirty buffer flushes done by the bufdaemon helpers");
311 static long barrierwrites;
312 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
313     "Number of barrier writes");
314 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
315     &unmapped_buf_allowed, 0,
316     "Permit the use of the unmapped i/o");
317 int maxbcachebuf = MAXBCACHEBUF;
318 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
319     "Maximum size of a buffer cache block");
320 
321 /*
322  * This lock synchronizes access to bd_request.
323  */
324 static struct mtx_padalign __exclusive_cache_line bdlock;
325 
326 /*
327  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
328  * waitrunningbufspace().
329  */
330 static struct mtx_padalign __exclusive_cache_line rbreqlock;
331 
332 /*
333  * Lock that protects bdirtywait.
334  */
335 static struct mtx_padalign __exclusive_cache_line bdirtylock;
336 
337 /*
338  * Wakeup point for bufdaemon, as well as indicator of whether it is already
339  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
340  * is idling.
341  */
342 static int bd_request;
343 
344 /*
345  * Request for the buf daemon to write more buffers than is indicated by
346  * lodirtybuf.  This may be necessary to push out excess dependencies or
347  * defragment the address space where a simple count of the number of dirty
348  * buffers is insufficient to characterize the demand for flushing them.
349  */
350 static int bd_speedupreq;
351 
352 /*
353  * Synchronization (sleep/wakeup) variable for active buffer space requests.
354  * Set when wait starts, cleared prior to wakeup().
355  * Used in runningbufwakeup() and waitrunningbufspace().
356  */
357 static int runningbufreq;
358 
359 /*
360  * Synchronization for bwillwrite() waiters.
361  */
362 static int bdirtywait;
363 
364 /*
365  * Definitions for the buffer free lists.
366  */
367 #define QUEUE_NONE	0	/* on no queue */
368 #define QUEUE_EMPTY	1	/* empty buffer headers */
369 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
370 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
371 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
372 
373 /* Maximum number of buffer domains. */
374 #define	BUF_DOMAINS	8
375 
376 struct bufdomainset bdlodirty;		/* Domains > lodirty */
377 struct bufdomainset bdhidirty;		/* Domains > hidirty */
378 
379 /* Configured number of clean queues. */
380 static int __read_mostly buf_domains;
381 
382 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
383 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
384 struct bufqueue __exclusive_cache_line bqempty;
385 
386 /*
387  * per-cpu empty buffer cache.
388  */
389 uma_zone_t buf_zone;
390 
391 /*
392  * Single global constant for BUF_WMESG, to avoid getting multiple references.
393  * buf_wmesg is referred from macros.
394  */
395 const char *buf_wmesg = BUF_WMESG;
396 
397 static int
398 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
399 {
400 	long value;
401 	int error;
402 
403 	value = *(long *)arg1;
404 	error = sysctl_handle_long(oidp, &value, 0, req);
405 	if (error != 0 || req->newptr == NULL)
406 		return (error);
407 	mtx_lock(&rbreqlock);
408 	if (arg1 == &hirunningspace) {
409 		if (value < lorunningspace)
410 			error = EINVAL;
411 		else
412 			hirunningspace = value;
413 	} else {
414 		KASSERT(arg1 == &lorunningspace,
415 		    ("%s: unknown arg1", __func__));
416 		if (value > hirunningspace)
417 			error = EINVAL;
418 		else
419 			lorunningspace = value;
420 	}
421 	mtx_unlock(&rbreqlock);
422 	return (error);
423 }
424 
425 static int
426 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
427 {
428 	int error;
429 	int value;
430 	int i;
431 
432 	value = *(int *)arg1;
433 	error = sysctl_handle_int(oidp, &value, 0, req);
434 	if (error != 0 || req->newptr == NULL)
435 		return (error);
436 	*(int *)arg1 = value;
437 	for (i = 0; i < buf_domains; i++)
438 		*(int *)(((uintptr_t)&bdomain[i]) + arg2) =
439 		    value / buf_domains;
440 
441 	return (error);
442 }
443 
444 static int
445 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
446 {
447 	long value;
448 	int error;
449 	int i;
450 
451 	value = *(long *)arg1;
452 	error = sysctl_handle_long(oidp, &value, 0, req);
453 	if (error != 0 || req->newptr == NULL)
454 		return (error);
455 	*(long *)arg1 = value;
456 	for (i = 0; i < buf_domains; i++)
457 		*(long *)(((uintptr_t)&bdomain[i]) + arg2) =
458 		    value / buf_domains;
459 
460 	return (error);
461 }
462 
463 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
464     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
465 static int
466 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
467 {
468 	long lvalue;
469 	int ivalue;
470 	int i;
471 
472 	lvalue = 0;
473 	for (i = 0; i < buf_domains; i++)
474 		lvalue += bdomain[i].bd_bufspace;
475 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
476 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
477 	if (lvalue > INT_MAX)
478 		/* On overflow, still write out a long to trigger ENOMEM. */
479 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
480 	ivalue = lvalue;
481 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
482 }
483 #else
484 static int
485 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
486 {
487 	long lvalue;
488 	int i;
489 
490 	lvalue = 0;
491 	for (i = 0; i < buf_domains; i++)
492 		lvalue += bdomain[i].bd_bufspace;
493 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
494 }
495 #endif
496 
497 static int
498 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
499 {
500 	int value;
501 	int i;
502 
503 	value = 0;
504 	for (i = 0; i < buf_domains; i++)
505 		value += bdomain[i].bd_numdirtybuffers;
506 	return (sysctl_handle_int(oidp, &value, 0, req));
507 }
508 
509 /*
510  *	bdirtywakeup:
511  *
512  *	Wakeup any bwillwrite() waiters.
513  */
514 static void
515 bdirtywakeup(void)
516 {
517 	mtx_lock(&bdirtylock);
518 	if (bdirtywait) {
519 		bdirtywait = 0;
520 		wakeup(&bdirtywait);
521 	}
522 	mtx_unlock(&bdirtylock);
523 }
524 
525 /*
526  *	bd_clear:
527  *
528  *	Clear a domain from the appropriate bitsets when dirtybuffers
529  *	is decremented.
530  */
531 static void
532 bd_clear(struct bufdomain *bd)
533 {
534 
535 	mtx_lock(&bdirtylock);
536 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
537 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
538 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
539 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
540 	mtx_unlock(&bdirtylock);
541 }
542 
543 /*
544  *	bd_set:
545  *
546  *	Set a domain in the appropriate bitsets when dirtybuffers
547  *	is incremented.
548  */
549 static void
550 bd_set(struct bufdomain *bd)
551 {
552 
553 	mtx_lock(&bdirtylock);
554 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
555 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
556 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
557 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
558 	mtx_unlock(&bdirtylock);
559 }
560 
561 /*
562  *	bdirtysub:
563  *
564  *	Decrement the numdirtybuffers count by one and wakeup any
565  *	threads blocked in bwillwrite().
566  */
567 static void
568 bdirtysub(struct buf *bp)
569 {
570 	struct bufdomain *bd;
571 	int num;
572 
573 	bd = bufdomain(bp);
574 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
575 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
576 		bdirtywakeup();
577 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
578 		bd_clear(bd);
579 }
580 
581 /*
582  *	bdirtyadd:
583  *
584  *	Increment the numdirtybuffers count by one and wakeup the buf
585  *	daemon if needed.
586  */
587 static void
588 bdirtyadd(struct buf *bp)
589 {
590 	struct bufdomain *bd;
591 	int num;
592 
593 	/*
594 	 * Only do the wakeup once as we cross the boundary.  The
595 	 * buf daemon will keep running until the condition clears.
596 	 */
597 	bd = bufdomain(bp);
598 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
599 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
600 		bd_wakeup();
601 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
602 		bd_set(bd);
603 }
604 
605 /*
606  *	bufspace_daemon_wakeup:
607  *
608  *	Wakeup the daemons responsible for freeing clean bufs.
609  */
610 static void
611 bufspace_daemon_wakeup(struct bufdomain *bd)
612 {
613 
614 	/*
615 	 * avoid the lock if the daemon is running.
616 	 */
617 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
618 		BD_RUN_LOCK(bd);
619 		atomic_store_int(&bd->bd_running, 1);
620 		wakeup(&bd->bd_running);
621 		BD_RUN_UNLOCK(bd);
622 	}
623 }
624 
625 /*
626  *	bufspace_daemon_wait:
627  *
628  *	Sleep until the domain falls below a limit or one second passes.
629  */
630 static void
631 bufspace_daemon_wait(struct bufdomain *bd)
632 {
633 	/*
634 	 * Re-check our limits and sleep.  bd_running must be
635 	 * cleared prior to checking the limits to avoid missed
636 	 * wakeups.  The waker will adjust one of bufspace or
637 	 * freebuffers prior to checking bd_running.
638 	 */
639 	BD_RUN_LOCK(bd);
640 	atomic_store_int(&bd->bd_running, 0);
641 	if (bd->bd_bufspace < bd->bd_bufspacethresh &&
642 	    bd->bd_freebuffers > bd->bd_lofreebuffers) {
643 		msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
644 		    "-", hz);
645 	} else {
646 		/* Avoid spurious wakeups while running. */
647 		atomic_store_int(&bd->bd_running, 1);
648 		BD_RUN_UNLOCK(bd);
649 	}
650 }
651 
652 /*
653  *	bufspace_adjust:
654  *
655  *	Adjust the reported bufspace for a KVA managed buffer, possibly
656  * 	waking any waiters.
657  */
658 static void
659 bufspace_adjust(struct buf *bp, int bufsize)
660 {
661 	struct bufdomain *bd;
662 	long space;
663 	int diff;
664 
665 	KASSERT((bp->b_flags & B_MALLOC) == 0,
666 	    ("bufspace_adjust: malloc buf %p", bp));
667 	bd = bufdomain(bp);
668 	diff = bufsize - bp->b_bufsize;
669 	if (diff < 0) {
670 		atomic_subtract_long(&bd->bd_bufspace, -diff);
671 	} else if (diff > 0) {
672 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
673 		/* Wake up the daemon on the transition. */
674 		if (space < bd->bd_bufspacethresh &&
675 		    space + diff >= bd->bd_bufspacethresh)
676 			bufspace_daemon_wakeup(bd);
677 	}
678 	bp->b_bufsize = bufsize;
679 }
680 
681 /*
682  *	bufspace_reserve:
683  *
684  *	Reserve bufspace before calling allocbuf().  metadata has a
685  *	different space limit than data.
686  */
687 static int
688 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
689 {
690 	long limit, new;
691 	long space;
692 
693 	if (metadata)
694 		limit = bd->bd_maxbufspace;
695 	else
696 		limit = bd->bd_hibufspace;
697 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
698 	new = space + size;
699 	if (new > limit) {
700 		atomic_subtract_long(&bd->bd_bufspace, size);
701 		return (ENOSPC);
702 	}
703 
704 	/* Wake up the daemon on the transition. */
705 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
706 		bufspace_daemon_wakeup(bd);
707 
708 	return (0);
709 }
710 
711 /*
712  *	bufspace_release:
713  *
714  *	Release reserved bufspace after bufspace_adjust() has consumed it.
715  */
716 static void
717 bufspace_release(struct bufdomain *bd, int size)
718 {
719 
720 	atomic_subtract_long(&bd->bd_bufspace, size);
721 }
722 
723 /*
724  *	bufspace_wait:
725  *
726  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
727  *	supplied.  bd_wanted must be set prior to polling for space.  The
728  *	operation must be re-tried on return.
729  */
730 static void
731 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
732     int slpflag, int slptimeo)
733 {
734 	struct thread *td;
735 	int error, fl, norunbuf;
736 
737 	if ((gbflags & GB_NOWAIT_BD) != 0)
738 		return;
739 
740 	td = curthread;
741 	BD_LOCK(bd);
742 	while (bd->bd_wanted) {
743 		if (vp != NULL && vp->v_type != VCHR &&
744 		    (td->td_pflags & TDP_BUFNEED) == 0) {
745 			BD_UNLOCK(bd);
746 			/*
747 			 * getblk() is called with a vnode locked, and
748 			 * some majority of the dirty buffers may as
749 			 * well belong to the vnode.  Flushing the
750 			 * buffers there would make a progress that
751 			 * cannot be achieved by the buf_daemon, that
752 			 * cannot lock the vnode.
753 			 */
754 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
755 			    (td->td_pflags & TDP_NORUNNINGBUF);
756 
757 			/*
758 			 * Play bufdaemon.  The getnewbuf() function
759 			 * may be called while the thread owns lock
760 			 * for another dirty buffer for the same
761 			 * vnode, which makes it impossible to use
762 			 * VOP_FSYNC() there, due to the buffer lock
763 			 * recursion.
764 			 */
765 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
766 			fl = buf_flush(vp, bd, flushbufqtarget);
767 			td->td_pflags &= norunbuf;
768 			BD_LOCK(bd);
769 			if (fl != 0)
770 				continue;
771 			if (bd->bd_wanted == 0)
772 				break;
773 		}
774 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
775 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
776 		if (error != 0)
777 			break;
778 	}
779 	BD_UNLOCK(bd);
780 }
781 
782 
783 /*
784  *	bufspace_daemon:
785  *
786  *	buffer space management daemon.  Tries to maintain some marginal
787  *	amount of free buffer space so that requesting processes neither
788  *	block nor work to reclaim buffers.
789  */
790 static void
791 bufspace_daemon(void *arg)
792 {
793 	struct bufdomain *bd;
794 
795 	bd = arg;
796 	for (;;) {
797 		kproc_suspend_check(curproc);
798 
799 		/*
800 		 * Free buffers from the clean queue until we meet our
801 		 * targets.
802 		 *
803 		 * Theory of operation:  The buffer cache is most efficient
804 		 * when some free buffer headers and space are always
805 		 * available to getnewbuf().  This daemon attempts to prevent
806 		 * the excessive blocking and synchronization associated
807 		 * with shortfall.  It goes through three phases according
808 		 * demand:
809 		 *
810 		 * 1)	The daemon wakes up voluntarily once per-second
811 		 *	during idle periods when the counters are below
812 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
813 		 *
814 		 * 2)	The daemon wakes up as we cross the thresholds
815 		 *	ahead of any potential blocking.  This may bounce
816 		 *	slightly according to the rate of consumption and
817 		 *	release.
818 		 *
819 		 * 3)	The daemon and consumers are starved for working
820 		 *	clean buffers.  This is the 'bufspace' sleep below
821 		 *	which will inefficiently trade bufs with bqrelse
822 		 *	until we return to condition 2.
823 		 */
824 		do {
825 			if (buf_recycle(bd, false) != 0) {
826 				if (bd_flushall(bd))
827 					continue;
828 				/*
829 				 * Speedup dirty if we've run out of clean
830 				 * buffers.  This is possible in particular
831 				 * because softdep may held many bufs locked
832 				 * pending writes to other bufs which are
833 				 * marked for delayed write, exhausting
834 				 * clean space until they are written.
835 				 */
836 				bd_speedup();
837 				BD_LOCK(bd);
838 				if (bd->bd_wanted) {
839 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
840 					    PRIBIO|PDROP, "bufspace", hz/10);
841 				} else
842 					BD_UNLOCK(bd);
843 			}
844 			maybe_yield();
845 		} while (bd->bd_bufspace > bd->bd_lobufspace ||
846 		    bd->bd_freebuffers < bd->bd_hifreebuffers);
847 
848 		bufspace_daemon_wait(bd);
849 	}
850 }
851 
852 /*
853  *	bufmallocadjust:
854  *
855  *	Adjust the reported bufspace for a malloc managed buffer, possibly
856  *	waking any waiters.
857  */
858 static void
859 bufmallocadjust(struct buf *bp, int bufsize)
860 {
861 	int diff;
862 
863 	KASSERT((bp->b_flags & B_MALLOC) != 0,
864 	    ("bufmallocadjust: non-malloc buf %p", bp));
865 	diff = bufsize - bp->b_bufsize;
866 	if (diff < 0)
867 		atomic_subtract_long(&bufmallocspace, -diff);
868 	else
869 		atomic_add_long(&bufmallocspace, diff);
870 	bp->b_bufsize = bufsize;
871 }
872 
873 /*
874  *	runningwakeup:
875  *
876  *	Wake up processes that are waiting on asynchronous writes to fall
877  *	below lorunningspace.
878  */
879 static void
880 runningwakeup(void)
881 {
882 
883 	mtx_lock(&rbreqlock);
884 	if (runningbufreq) {
885 		runningbufreq = 0;
886 		wakeup(&runningbufreq);
887 	}
888 	mtx_unlock(&rbreqlock);
889 }
890 
891 /*
892  *	runningbufwakeup:
893  *
894  *	Decrement the outstanding write count according.
895  */
896 void
897 runningbufwakeup(struct buf *bp)
898 {
899 	long space, bspace;
900 
901 	bspace = bp->b_runningbufspace;
902 	if (bspace == 0)
903 		return;
904 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
905 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
906 	    space, bspace));
907 	bp->b_runningbufspace = 0;
908 	/*
909 	 * Only acquire the lock and wakeup on the transition from exceeding
910 	 * the threshold to falling below it.
911 	 */
912 	if (space < lorunningspace)
913 		return;
914 	if (space - bspace > lorunningspace)
915 		return;
916 	runningwakeup();
917 }
918 
919 /*
920  *	waitrunningbufspace()
921  *
922  *	runningbufspace is a measure of the amount of I/O currently
923  *	running.  This routine is used in async-write situations to
924  *	prevent creating huge backups of pending writes to a device.
925  *	Only asynchronous writes are governed by this function.
926  *
927  *	This does NOT turn an async write into a sync write.  It waits
928  *	for earlier writes to complete and generally returns before the
929  *	caller's write has reached the device.
930  */
931 void
932 waitrunningbufspace(void)
933 {
934 
935 	mtx_lock(&rbreqlock);
936 	while (runningbufspace > hirunningspace) {
937 		runningbufreq = 1;
938 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
939 	}
940 	mtx_unlock(&rbreqlock);
941 }
942 
943 
944 /*
945  *	vfs_buf_test_cache:
946  *
947  *	Called when a buffer is extended.  This function clears the B_CACHE
948  *	bit if the newly extended portion of the buffer does not contain
949  *	valid data.
950  */
951 static __inline void
952 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
953     vm_offset_t size, vm_page_t m)
954 {
955 
956 	VM_OBJECT_ASSERT_LOCKED(m->object);
957 	if (bp->b_flags & B_CACHE) {
958 		int base = (foff + off) & PAGE_MASK;
959 		if (vm_page_is_valid(m, base, size) == 0)
960 			bp->b_flags &= ~B_CACHE;
961 	}
962 }
963 
964 /* Wake up the buffer daemon if necessary */
965 static void
966 bd_wakeup(void)
967 {
968 
969 	mtx_lock(&bdlock);
970 	if (bd_request == 0) {
971 		bd_request = 1;
972 		wakeup(&bd_request);
973 	}
974 	mtx_unlock(&bdlock);
975 }
976 
977 /*
978  * Adjust the maxbcachbuf tunable.
979  */
980 static void
981 maxbcachebuf_adjust(void)
982 {
983 	int i;
984 
985 	/*
986 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
987 	 */
988 	i = 2;
989 	while (i * 2 <= maxbcachebuf)
990 		i *= 2;
991 	maxbcachebuf = i;
992 	if (maxbcachebuf < MAXBSIZE)
993 		maxbcachebuf = MAXBSIZE;
994 	if (maxbcachebuf > MAXPHYS)
995 		maxbcachebuf = MAXPHYS;
996 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
997 		printf("maxbcachebuf=%d\n", maxbcachebuf);
998 }
999 
1000 /*
1001  * bd_speedup - speedup the buffer cache flushing code
1002  */
1003 void
1004 bd_speedup(void)
1005 {
1006 	int needwake;
1007 
1008 	mtx_lock(&bdlock);
1009 	needwake = 0;
1010 	if (bd_speedupreq == 0 || bd_request == 0)
1011 		needwake = 1;
1012 	bd_speedupreq = 1;
1013 	bd_request = 1;
1014 	if (needwake)
1015 		wakeup(&bd_request);
1016 	mtx_unlock(&bdlock);
1017 }
1018 
1019 #ifndef NSWBUF_MIN
1020 #define	NSWBUF_MIN	16
1021 #endif
1022 
1023 #ifdef __i386__
1024 #define	TRANSIENT_DENOM	5
1025 #else
1026 #define	TRANSIENT_DENOM 10
1027 #endif
1028 
1029 /*
1030  * Calculating buffer cache scaling values and reserve space for buffer
1031  * headers.  This is called during low level kernel initialization and
1032  * may be called more then once.  We CANNOT write to the memory area
1033  * being reserved at this time.
1034  */
1035 caddr_t
1036 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1037 {
1038 	int tuned_nbuf;
1039 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1040 
1041 	/*
1042 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1043 	 * PAGE_SIZE is >= 1K)
1044 	 */
1045 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1046 
1047 	maxbcachebuf_adjust();
1048 	/*
1049 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1050 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1051 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1052 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1053 	 * the buffer cache we limit the eventual kva reservation to
1054 	 * maxbcache bytes.
1055 	 *
1056 	 * factor represents the 1/4 x ram conversion.
1057 	 */
1058 	if (nbuf == 0) {
1059 		int factor = 4 * BKVASIZE / 1024;
1060 
1061 		nbuf = 50;
1062 		if (physmem_est > 4096)
1063 			nbuf += min((physmem_est - 4096) / factor,
1064 			    65536 / factor);
1065 		if (physmem_est > 65536)
1066 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1067 			    32 * 1024 * 1024 / (factor * 5));
1068 
1069 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1070 			nbuf = maxbcache / BKVASIZE;
1071 		tuned_nbuf = 1;
1072 	} else
1073 		tuned_nbuf = 0;
1074 
1075 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1076 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1077 	if (nbuf > maxbuf) {
1078 		if (!tuned_nbuf)
1079 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1080 			    maxbuf);
1081 		nbuf = maxbuf;
1082 	}
1083 
1084 	/*
1085 	 * Ideal allocation size for the transient bio submap is 10%
1086 	 * of the maximal space buffer map.  This roughly corresponds
1087 	 * to the amount of the buffer mapped for typical UFS load.
1088 	 *
1089 	 * Clip the buffer map to reserve space for the transient
1090 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1091 	 * maximum buffer map extent on the platform.
1092 	 *
1093 	 * The fall-back to the maxbuf in case of maxbcache unset,
1094 	 * allows to not trim the buffer KVA for the architectures
1095 	 * with ample KVA space.
1096 	 */
1097 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1098 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1099 		buf_sz = (long)nbuf * BKVASIZE;
1100 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1101 		    (TRANSIENT_DENOM - 1)) {
1102 			/*
1103 			 * There is more KVA than memory.  Do not
1104 			 * adjust buffer map size, and assign the rest
1105 			 * of maxbuf to transient map.
1106 			 */
1107 			biotmap_sz = maxbuf_sz - buf_sz;
1108 		} else {
1109 			/*
1110 			 * Buffer map spans all KVA we could afford on
1111 			 * this platform.  Give 10% (20% on i386) of
1112 			 * the buffer map to the transient bio map.
1113 			 */
1114 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1115 			buf_sz -= biotmap_sz;
1116 		}
1117 		if (biotmap_sz / INT_MAX > MAXPHYS)
1118 			bio_transient_maxcnt = INT_MAX;
1119 		else
1120 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1121 		/*
1122 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1123 		 * using the transient mapping.
1124 		 */
1125 		if (bio_transient_maxcnt > 1024)
1126 			bio_transient_maxcnt = 1024;
1127 		if (tuned_nbuf)
1128 			nbuf = buf_sz / BKVASIZE;
1129 	}
1130 
1131 	/*
1132 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
1133 	 * We have no less then 16 and no more then 256.
1134 	 */
1135 	nswbuf = min(nbuf / 4, 256);
1136 	TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf);
1137 	if (nswbuf < NSWBUF_MIN)
1138 		nswbuf = NSWBUF_MIN;
1139 
1140 	/*
1141 	 * Reserve space for the buffer cache buffers
1142 	 */
1143 	swbuf = (void *)v;
1144 	v = (caddr_t)(swbuf + nswbuf);
1145 	buf = (void *)v;
1146 	v = (caddr_t)(buf + nbuf);
1147 
1148 	return(v);
1149 }
1150 
1151 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1152 void
1153 bufinit(void)
1154 {
1155 	struct buf *bp;
1156 	int i;
1157 
1158 	KASSERT(maxbcachebuf >= MAXBSIZE,
1159 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1160 	    MAXBSIZE));
1161 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1162 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1163 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1164 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1165 
1166 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1167 
1168 	/* finally, initialize each buffer header and stick on empty q */
1169 	for (i = 0; i < nbuf; i++) {
1170 		bp = &buf[i];
1171 		bzero(bp, sizeof *bp);
1172 		bp->b_flags = B_INVAL;
1173 		bp->b_rcred = NOCRED;
1174 		bp->b_wcred = NOCRED;
1175 		bp->b_qindex = QUEUE_NONE;
1176 		bp->b_domain = -1;
1177 		bp->b_subqueue = mp_maxid + 1;
1178 		bp->b_xflags = 0;
1179 		bp->b_data = bp->b_kvabase = unmapped_buf;
1180 		LIST_INIT(&bp->b_dep);
1181 		BUF_LOCKINIT(bp);
1182 		bq_insert(&bqempty, bp, false);
1183 	}
1184 
1185 	/*
1186 	 * maxbufspace is the absolute maximum amount of buffer space we are
1187 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1188 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1189 	 * used by most other requests.  The differential is required to
1190 	 * ensure that metadata deadlocks don't occur.
1191 	 *
1192 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1193 	 * this may result in KVM fragmentation which is not handled optimally
1194 	 * by the system. XXX This is less true with vmem.  We could use
1195 	 * PAGE_SIZE.
1196 	 */
1197 	maxbufspace = (long)nbuf * BKVASIZE;
1198 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1199 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1200 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1201 
1202 	/*
1203 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1204 	 * arbitrarily and may need further tuning. It corresponds to
1205 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1206 	 * which fits with many RAID controllers' tagged queuing limits.
1207 	 * The lower 1 MiB limit is the historical upper limit for
1208 	 * hirunningspace.
1209 	 */
1210 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1211 	    16 * 1024 * 1024), 1024 * 1024);
1212 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1213 
1214 	/*
1215 	 * Limit the amount of malloc memory since it is wired permanently into
1216 	 * the kernel space.  Even though this is accounted for in the buffer
1217 	 * allocation, we don't want the malloced region to grow uncontrolled.
1218 	 * The malloc scheme improves memory utilization significantly on
1219 	 * average (small) directories.
1220 	 */
1221 	maxbufmallocspace = hibufspace / 20;
1222 
1223 	/*
1224 	 * Reduce the chance of a deadlock occurring by limiting the number
1225 	 * of delayed-write dirty buffers we allow to stack up.
1226 	 */
1227 	hidirtybuffers = nbuf / 4 + 20;
1228 	dirtybufthresh = hidirtybuffers * 9 / 10;
1229 	/*
1230 	 * To support extreme low-memory systems, make sure hidirtybuffers
1231 	 * cannot eat up all available buffer space.  This occurs when our
1232 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1233 	 * buffer space assuming BKVASIZE'd buffers.
1234 	 */
1235 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1236 		hidirtybuffers >>= 1;
1237 	}
1238 	lodirtybuffers = hidirtybuffers / 2;
1239 
1240 	/*
1241 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1242 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1243 	 * are counted as free but will be unavailable to threads executing
1244 	 * on other cpus.
1245 	 *
1246 	 * hifreebuffers is the free target for the bufspace daemon.  This
1247 	 * should be set appropriately to limit work per-iteration.
1248 	 */
1249 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1250 	hifreebuffers = (3 * lofreebuffers) / 2;
1251 	numfreebuffers = nbuf;
1252 
1253 	/* Setup the kva and free list allocators. */
1254 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1255 	buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1256 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1257 
1258 	/*
1259 	 * Size the clean queue according to the amount of buffer space.
1260 	 * One queue per-256mb up to the max.  More queues gives better
1261 	 * concurrency but less accurate LRU.
1262 	 */
1263 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1264 	for (i = 0 ; i < buf_domains; i++) {
1265 		struct bufdomain *bd;
1266 
1267 		bd = &bdomain[i];
1268 		bd_init(bd);
1269 		bd->bd_freebuffers = nbuf / buf_domains;
1270 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1271 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1272 		bd->bd_bufspace = 0;
1273 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1274 		bd->bd_hibufspace = hibufspace / buf_domains;
1275 		bd->bd_lobufspace = lobufspace / buf_domains;
1276 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1277 		bd->bd_numdirtybuffers = 0;
1278 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1279 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1280 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1281 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1282 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1283 	}
1284 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1285 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1286 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1287 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1288 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1289 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1290 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1291 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1292 }
1293 
1294 #ifdef INVARIANTS
1295 static inline void
1296 vfs_buf_check_mapped(struct buf *bp)
1297 {
1298 
1299 	KASSERT(bp->b_kvabase != unmapped_buf,
1300 	    ("mapped buf: b_kvabase was not updated %p", bp));
1301 	KASSERT(bp->b_data != unmapped_buf,
1302 	    ("mapped buf: b_data was not updated %p", bp));
1303 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1304 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1305 }
1306 
1307 static inline void
1308 vfs_buf_check_unmapped(struct buf *bp)
1309 {
1310 
1311 	KASSERT(bp->b_data == unmapped_buf,
1312 	    ("unmapped buf: corrupted b_data %p", bp));
1313 }
1314 
1315 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1316 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1317 #else
1318 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1319 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1320 #endif
1321 
1322 static int
1323 isbufbusy(struct buf *bp)
1324 {
1325 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1326 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1327 		return (1);
1328 	return (0);
1329 }
1330 
1331 /*
1332  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1333  */
1334 void
1335 bufshutdown(int show_busybufs)
1336 {
1337 	static int first_buf_printf = 1;
1338 	struct buf *bp;
1339 	int iter, nbusy, pbusy;
1340 #ifndef PREEMPTION
1341 	int subiter;
1342 #endif
1343 
1344 	/*
1345 	 * Sync filesystems for shutdown
1346 	 */
1347 	wdog_kern_pat(WD_LASTVAL);
1348 	sys_sync(curthread, NULL);
1349 
1350 	/*
1351 	 * With soft updates, some buffers that are
1352 	 * written will be remarked as dirty until other
1353 	 * buffers are written.
1354 	 */
1355 	for (iter = pbusy = 0; iter < 20; iter++) {
1356 		nbusy = 0;
1357 		for (bp = &buf[nbuf]; --bp >= buf; )
1358 			if (isbufbusy(bp))
1359 				nbusy++;
1360 		if (nbusy == 0) {
1361 			if (first_buf_printf)
1362 				printf("All buffers synced.");
1363 			break;
1364 		}
1365 		if (first_buf_printf) {
1366 			printf("Syncing disks, buffers remaining... ");
1367 			first_buf_printf = 0;
1368 		}
1369 		printf("%d ", nbusy);
1370 		if (nbusy < pbusy)
1371 			iter = 0;
1372 		pbusy = nbusy;
1373 
1374 		wdog_kern_pat(WD_LASTVAL);
1375 		sys_sync(curthread, NULL);
1376 
1377 #ifdef PREEMPTION
1378 		/*
1379 		 * Drop Giant and spin for a while to allow
1380 		 * interrupt threads to run.
1381 		 */
1382 		DROP_GIANT();
1383 		DELAY(50000 * iter);
1384 		PICKUP_GIANT();
1385 #else
1386 		/*
1387 		 * Drop Giant and context switch several times to
1388 		 * allow interrupt threads to run.
1389 		 */
1390 		DROP_GIANT();
1391 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1392 			thread_lock(curthread);
1393 			mi_switch(SW_VOL, NULL);
1394 			thread_unlock(curthread);
1395 			DELAY(1000);
1396 		}
1397 		PICKUP_GIANT();
1398 #endif
1399 	}
1400 	printf("\n");
1401 	/*
1402 	 * Count only busy local buffers to prevent forcing
1403 	 * a fsck if we're just a client of a wedged NFS server
1404 	 */
1405 	nbusy = 0;
1406 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1407 		if (isbufbusy(bp)) {
1408 #if 0
1409 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1410 			if (bp->b_dev == NULL) {
1411 				TAILQ_REMOVE(&mountlist,
1412 				    bp->b_vp->v_mount, mnt_list);
1413 				continue;
1414 			}
1415 #endif
1416 			nbusy++;
1417 			if (show_busybufs > 0) {
1418 				printf(
1419 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1420 				    nbusy, bp, bp->b_vp, bp->b_flags,
1421 				    (intmax_t)bp->b_blkno,
1422 				    (intmax_t)bp->b_lblkno);
1423 				BUF_LOCKPRINTINFO(bp);
1424 				if (show_busybufs > 1)
1425 					vn_printf(bp->b_vp,
1426 					    "vnode content: ");
1427 			}
1428 		}
1429 	}
1430 	if (nbusy) {
1431 		/*
1432 		 * Failed to sync all blocks. Indicate this and don't
1433 		 * unmount filesystems (thus forcing an fsck on reboot).
1434 		 */
1435 		printf("Giving up on %d buffers\n", nbusy);
1436 		DELAY(5000000);	/* 5 seconds */
1437 	} else {
1438 		if (!first_buf_printf)
1439 			printf("Final sync complete\n");
1440 		/*
1441 		 * Unmount filesystems
1442 		 */
1443 		if (panicstr == NULL)
1444 			vfs_unmountall();
1445 	}
1446 	swapoff_all();
1447 	DELAY(100000);		/* wait for console output to finish */
1448 }
1449 
1450 static void
1451 bpmap_qenter(struct buf *bp)
1452 {
1453 
1454 	BUF_CHECK_MAPPED(bp);
1455 
1456 	/*
1457 	 * bp->b_data is relative to bp->b_offset, but
1458 	 * bp->b_offset may be offset into the first page.
1459 	 */
1460 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1461 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1462 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1463 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1464 }
1465 
1466 static inline struct bufdomain *
1467 bufdomain(struct buf *bp)
1468 {
1469 
1470 	return (&bdomain[bp->b_domain]);
1471 }
1472 
1473 static struct bufqueue *
1474 bufqueue(struct buf *bp)
1475 {
1476 
1477 	switch (bp->b_qindex) {
1478 	case QUEUE_NONE:
1479 		/* FALLTHROUGH */
1480 	case QUEUE_SENTINEL:
1481 		return (NULL);
1482 	case QUEUE_EMPTY:
1483 		return (&bqempty);
1484 	case QUEUE_DIRTY:
1485 		return (&bufdomain(bp)->bd_dirtyq);
1486 	case QUEUE_CLEAN:
1487 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1488 	default:
1489 		break;
1490 	}
1491 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1492 }
1493 
1494 /*
1495  * Return the locked bufqueue that bp is a member of.
1496  */
1497 static struct bufqueue *
1498 bufqueue_acquire(struct buf *bp)
1499 {
1500 	struct bufqueue *bq, *nbq;
1501 
1502 	/*
1503 	 * bp can be pushed from a per-cpu queue to the
1504 	 * cleanq while we're waiting on the lock.  Retry
1505 	 * if the queues don't match.
1506 	 */
1507 	bq = bufqueue(bp);
1508 	BQ_LOCK(bq);
1509 	for (;;) {
1510 		nbq = bufqueue(bp);
1511 		if (bq == nbq)
1512 			break;
1513 		BQ_UNLOCK(bq);
1514 		BQ_LOCK(nbq);
1515 		bq = nbq;
1516 	}
1517 	return (bq);
1518 }
1519 
1520 /*
1521  *	binsfree:
1522  *
1523  *	Insert the buffer into the appropriate free list.  Requires a
1524  *	locked buffer on entry and buffer is unlocked before return.
1525  */
1526 static void
1527 binsfree(struct buf *bp, int qindex)
1528 {
1529 	struct bufdomain *bd;
1530 	struct bufqueue *bq;
1531 
1532 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1533 	    ("binsfree: Invalid qindex %d", qindex));
1534 	BUF_ASSERT_XLOCKED(bp);
1535 
1536 	/*
1537 	 * Handle delayed bremfree() processing.
1538 	 */
1539 	if (bp->b_flags & B_REMFREE) {
1540 		if (bp->b_qindex == qindex) {
1541 			bp->b_flags |= B_REUSE;
1542 			bp->b_flags &= ~B_REMFREE;
1543 			BUF_UNLOCK(bp);
1544 			return;
1545 		}
1546 		bq = bufqueue_acquire(bp);
1547 		bq_remove(bq, bp);
1548 		BQ_UNLOCK(bq);
1549 	}
1550 	bd = bufdomain(bp);
1551 	if (qindex == QUEUE_CLEAN) {
1552 		if (bd->bd_lim != 0)
1553 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1554 		else
1555 			bq = bd->bd_cleanq;
1556 	} else
1557 		bq = &bd->bd_dirtyq;
1558 	bq_insert(bq, bp, true);
1559 }
1560 
1561 /*
1562  * buf_free:
1563  *
1564  *	Free a buffer to the buf zone once it no longer has valid contents.
1565  */
1566 static void
1567 buf_free(struct buf *bp)
1568 {
1569 
1570 	if (bp->b_flags & B_REMFREE)
1571 		bremfreef(bp);
1572 	if (bp->b_vflags & BV_BKGRDINPROG)
1573 		panic("losing buffer 1");
1574 	if (bp->b_rcred != NOCRED) {
1575 		crfree(bp->b_rcred);
1576 		bp->b_rcred = NOCRED;
1577 	}
1578 	if (bp->b_wcred != NOCRED) {
1579 		crfree(bp->b_wcred);
1580 		bp->b_wcred = NOCRED;
1581 	}
1582 	if (!LIST_EMPTY(&bp->b_dep))
1583 		buf_deallocate(bp);
1584 	bufkva_free(bp);
1585 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1586 	BUF_UNLOCK(bp);
1587 	uma_zfree(buf_zone, bp);
1588 }
1589 
1590 /*
1591  * buf_import:
1592  *
1593  *	Import bufs into the uma cache from the buf list.  The system still
1594  *	expects a static array of bufs and much of the synchronization
1595  *	around bufs assumes type stable storage.  As a result, UMA is used
1596  *	only as a per-cpu cache of bufs still maintained on a global list.
1597  */
1598 static int
1599 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1600 {
1601 	struct buf *bp;
1602 	int i;
1603 
1604 	BQ_LOCK(&bqempty);
1605 	for (i = 0; i < cnt; i++) {
1606 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1607 		if (bp == NULL)
1608 			break;
1609 		bq_remove(&bqempty, bp);
1610 		store[i] = bp;
1611 	}
1612 	BQ_UNLOCK(&bqempty);
1613 
1614 	return (i);
1615 }
1616 
1617 /*
1618  * buf_release:
1619  *
1620  *	Release bufs from the uma cache back to the buffer queues.
1621  */
1622 static void
1623 buf_release(void *arg, void **store, int cnt)
1624 {
1625 	struct bufqueue *bq;
1626 	struct buf *bp;
1627         int i;
1628 
1629 	bq = &bqempty;
1630 	BQ_LOCK(bq);
1631         for (i = 0; i < cnt; i++) {
1632 		bp = store[i];
1633 		/* Inline bq_insert() to batch locking. */
1634 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1635 		bp->b_flags &= ~(B_AGE | B_REUSE);
1636 		bq->bq_len++;
1637 		bp->b_qindex = bq->bq_index;
1638 	}
1639 	BQ_UNLOCK(bq);
1640 }
1641 
1642 /*
1643  * buf_alloc:
1644  *
1645  *	Allocate an empty buffer header.
1646  */
1647 static struct buf *
1648 buf_alloc(struct bufdomain *bd)
1649 {
1650 	struct buf *bp;
1651 	int freebufs;
1652 
1653 	/*
1654 	 * We can only run out of bufs in the buf zone if the average buf
1655 	 * is less than BKVASIZE.  In this case the actual wait/block will
1656 	 * come from buf_reycle() failing to flush one of these small bufs.
1657 	 */
1658 	bp = NULL;
1659 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1660 	if (freebufs > 0)
1661 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1662 	if (bp == NULL) {
1663 		atomic_fetchadd_int(&bd->bd_freebuffers, 1);
1664 		bufspace_daemon_wakeup(bd);
1665 		counter_u64_add(numbufallocfails, 1);
1666 		return (NULL);
1667 	}
1668 	/*
1669 	 * Wake-up the bufspace daemon on transition below threshold.
1670 	 */
1671 	if (freebufs == bd->bd_lofreebuffers)
1672 		bufspace_daemon_wakeup(bd);
1673 
1674 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1675 		panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1676 
1677 	KASSERT(bp->b_vp == NULL,
1678 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1679 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1680 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1681 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1682 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1683 	KASSERT(bp->b_npages == 0,
1684 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1685 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1686 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1687 
1688 	bp->b_domain = BD_DOMAIN(bd);
1689 	bp->b_flags = 0;
1690 	bp->b_ioflags = 0;
1691 	bp->b_xflags = 0;
1692 	bp->b_vflags = 0;
1693 	bp->b_vp = NULL;
1694 	bp->b_blkno = bp->b_lblkno = 0;
1695 	bp->b_offset = NOOFFSET;
1696 	bp->b_iodone = 0;
1697 	bp->b_error = 0;
1698 	bp->b_resid = 0;
1699 	bp->b_bcount = 0;
1700 	bp->b_npages = 0;
1701 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1702 	bp->b_bufobj = NULL;
1703 	bp->b_data = bp->b_kvabase = unmapped_buf;
1704 	bp->b_fsprivate1 = NULL;
1705 	bp->b_fsprivate2 = NULL;
1706 	bp->b_fsprivate3 = NULL;
1707 	LIST_INIT(&bp->b_dep);
1708 
1709 	return (bp);
1710 }
1711 
1712 /*
1713  *	buf_recycle:
1714  *
1715  *	Free a buffer from the given bufqueue.  kva controls whether the
1716  *	freed buf must own some kva resources.  This is used for
1717  *	defragmenting.
1718  */
1719 static int
1720 buf_recycle(struct bufdomain *bd, bool kva)
1721 {
1722 	struct bufqueue *bq;
1723 	struct buf *bp, *nbp;
1724 
1725 	if (kva)
1726 		counter_u64_add(bufdefragcnt, 1);
1727 	nbp = NULL;
1728 	bq = bd->bd_cleanq;
1729 	BQ_LOCK(bq);
1730 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1731 	    ("buf_recycle: Locks don't match"));
1732 	nbp = TAILQ_FIRST(&bq->bq_queue);
1733 
1734 	/*
1735 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1736 	 * depending.
1737 	 */
1738 	while ((bp = nbp) != NULL) {
1739 		/*
1740 		 * Calculate next bp (we can only use it if we do not
1741 		 * release the bqlock).
1742 		 */
1743 		nbp = TAILQ_NEXT(bp, b_freelist);
1744 
1745 		/*
1746 		 * If we are defragging then we need a buffer with
1747 		 * some kva to reclaim.
1748 		 */
1749 		if (kva && bp->b_kvasize == 0)
1750 			continue;
1751 
1752 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1753 			continue;
1754 
1755 		/*
1756 		 * Implement a second chance algorithm for frequently
1757 		 * accessed buffers.
1758 		 */
1759 		if ((bp->b_flags & B_REUSE) != 0) {
1760 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1761 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1762 			bp->b_flags &= ~B_REUSE;
1763 			BUF_UNLOCK(bp);
1764 			continue;
1765 		}
1766 
1767 		/*
1768 		 * Skip buffers with background writes in progress.
1769 		 */
1770 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1771 			BUF_UNLOCK(bp);
1772 			continue;
1773 		}
1774 
1775 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1776 		    ("buf_recycle: inconsistent queue %d bp %p",
1777 		    bp->b_qindex, bp));
1778 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1779 		    ("getnewbuf: queue domain %d doesn't match request %d",
1780 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1781 		/*
1782 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1783 		 * the scan from this point on.
1784 		 */
1785 		bq_remove(bq, bp);
1786 		BQ_UNLOCK(bq);
1787 
1788 		/*
1789 		 * Requeue the background write buffer with error and
1790 		 * restart the scan.
1791 		 */
1792 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1793 			bqrelse(bp);
1794 			BQ_LOCK(bq);
1795 			nbp = TAILQ_FIRST(&bq->bq_queue);
1796 			continue;
1797 		}
1798 		bp->b_flags |= B_INVAL;
1799 		brelse(bp);
1800 		return (0);
1801 	}
1802 	bd->bd_wanted = 1;
1803 	BQ_UNLOCK(bq);
1804 
1805 	return (ENOBUFS);
1806 }
1807 
1808 /*
1809  *	bremfree:
1810  *
1811  *	Mark the buffer for removal from the appropriate free list.
1812  *
1813  */
1814 void
1815 bremfree(struct buf *bp)
1816 {
1817 
1818 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1819 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1820 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1821 	KASSERT(bp->b_qindex != QUEUE_NONE,
1822 	    ("bremfree: buffer %p not on a queue.", bp));
1823 	BUF_ASSERT_XLOCKED(bp);
1824 
1825 	bp->b_flags |= B_REMFREE;
1826 }
1827 
1828 /*
1829  *	bremfreef:
1830  *
1831  *	Force an immediate removal from a free list.  Used only in nfs when
1832  *	it abuses the b_freelist pointer.
1833  */
1834 void
1835 bremfreef(struct buf *bp)
1836 {
1837 	struct bufqueue *bq;
1838 
1839 	bq = bufqueue_acquire(bp);
1840 	bq_remove(bq, bp);
1841 	BQ_UNLOCK(bq);
1842 }
1843 
1844 static void
1845 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1846 {
1847 
1848 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1849 	TAILQ_INIT(&bq->bq_queue);
1850 	bq->bq_len = 0;
1851 	bq->bq_index = qindex;
1852 	bq->bq_subqueue = subqueue;
1853 }
1854 
1855 static void
1856 bd_init(struct bufdomain *bd)
1857 {
1858 	int domain;
1859 	int i;
1860 
1861 	domain = bd - bdomain;
1862 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1863 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1864 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1865 	for (i = 0; i <= mp_maxid; i++)
1866 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1867 		    "bufq clean subqueue lock");
1868 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1869 }
1870 
1871 /*
1872  *	bq_remove:
1873  *
1874  *	Removes a buffer from the free list, must be called with the
1875  *	correct qlock held.
1876  */
1877 static void
1878 bq_remove(struct bufqueue *bq, struct buf *bp)
1879 {
1880 
1881 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1882 	    bp, bp->b_vp, bp->b_flags);
1883 	KASSERT(bp->b_qindex != QUEUE_NONE,
1884 	    ("bq_remove: buffer %p not on a queue.", bp));
1885 	KASSERT(bufqueue(bp) == bq,
1886 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1887 
1888 	BQ_ASSERT_LOCKED(bq);
1889 	if (bp->b_qindex != QUEUE_EMPTY) {
1890 		BUF_ASSERT_XLOCKED(bp);
1891 	}
1892 	KASSERT(bq->bq_len >= 1,
1893 	    ("queue %d underflow", bp->b_qindex));
1894 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1895 	bq->bq_len--;
1896 	bp->b_qindex = QUEUE_NONE;
1897 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1898 }
1899 
1900 static void
1901 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1902 {
1903 	struct buf *bp;
1904 
1905 	BQ_ASSERT_LOCKED(bq);
1906 	if (bq != bd->bd_cleanq) {
1907 		BD_LOCK(bd);
1908 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1909 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1910 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1911 			    b_freelist);
1912 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1913 		}
1914 		bd->bd_cleanq->bq_len += bq->bq_len;
1915 		bq->bq_len = 0;
1916 	}
1917 	if (bd->bd_wanted) {
1918 		bd->bd_wanted = 0;
1919 		wakeup(&bd->bd_wanted);
1920 	}
1921 	if (bq != bd->bd_cleanq)
1922 		BD_UNLOCK(bd);
1923 }
1924 
1925 static int
1926 bd_flushall(struct bufdomain *bd)
1927 {
1928 	struct bufqueue *bq;
1929 	int flushed;
1930 	int i;
1931 
1932 	if (bd->bd_lim == 0)
1933 		return (0);
1934 	flushed = 0;
1935 	for (i = 0; i <= mp_maxid; i++) {
1936 		bq = &bd->bd_subq[i];
1937 		if (bq->bq_len == 0)
1938 			continue;
1939 		BQ_LOCK(bq);
1940 		bd_flush(bd, bq);
1941 		BQ_UNLOCK(bq);
1942 		flushed++;
1943 	}
1944 
1945 	return (flushed);
1946 }
1947 
1948 static void
1949 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1950 {
1951 	struct bufdomain *bd;
1952 
1953 	if (bp->b_qindex != QUEUE_NONE)
1954 		panic("bq_insert: free buffer %p onto another queue?", bp);
1955 
1956 	bd = bufdomain(bp);
1957 	if (bp->b_flags & B_AGE) {
1958 		/* Place this buf directly on the real queue. */
1959 		if (bq->bq_index == QUEUE_CLEAN)
1960 			bq = bd->bd_cleanq;
1961 		BQ_LOCK(bq);
1962 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1963 	} else {
1964 		BQ_LOCK(bq);
1965 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1966 	}
1967 	bp->b_flags &= ~(B_AGE | B_REUSE);
1968 	bq->bq_len++;
1969 	bp->b_qindex = bq->bq_index;
1970 	bp->b_subqueue = bq->bq_subqueue;
1971 
1972 	/*
1973 	 * Unlock before we notify so that we don't wakeup a waiter that
1974 	 * fails a trylock on the buf and sleeps again.
1975 	 */
1976 	if (unlock)
1977 		BUF_UNLOCK(bp);
1978 
1979 	if (bp->b_qindex == QUEUE_CLEAN) {
1980 		/*
1981 		 * Flush the per-cpu queue and notify any waiters.
1982 		 */
1983 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1984 		    bq->bq_len >= bd->bd_lim))
1985 			bd_flush(bd, bq);
1986 	}
1987 	BQ_UNLOCK(bq);
1988 }
1989 
1990 /*
1991  *	bufkva_free:
1992  *
1993  *	Free the kva allocation for a buffer.
1994  *
1995  */
1996 static void
1997 bufkva_free(struct buf *bp)
1998 {
1999 
2000 #ifdef INVARIANTS
2001 	if (bp->b_kvasize == 0) {
2002 		KASSERT(bp->b_kvabase == unmapped_buf &&
2003 		    bp->b_data == unmapped_buf,
2004 		    ("Leaked KVA space on %p", bp));
2005 	} else if (buf_mapped(bp))
2006 		BUF_CHECK_MAPPED(bp);
2007 	else
2008 		BUF_CHECK_UNMAPPED(bp);
2009 #endif
2010 	if (bp->b_kvasize == 0)
2011 		return;
2012 
2013 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2014 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2015 	counter_u64_add(buffreekvacnt, 1);
2016 	bp->b_data = bp->b_kvabase = unmapped_buf;
2017 	bp->b_kvasize = 0;
2018 }
2019 
2020 /*
2021  *	bufkva_alloc:
2022  *
2023  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2024  */
2025 static int
2026 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2027 {
2028 	vm_offset_t addr;
2029 	int error;
2030 
2031 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2032 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2033 
2034 	bufkva_free(bp);
2035 
2036 	addr = 0;
2037 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2038 	if (error != 0) {
2039 		/*
2040 		 * Buffer map is too fragmented.  Request the caller
2041 		 * to defragment the map.
2042 		 */
2043 		return (error);
2044 	}
2045 	bp->b_kvabase = (caddr_t)addr;
2046 	bp->b_kvasize = maxsize;
2047 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2048 	if ((gbflags & GB_UNMAPPED) != 0) {
2049 		bp->b_data = unmapped_buf;
2050 		BUF_CHECK_UNMAPPED(bp);
2051 	} else {
2052 		bp->b_data = bp->b_kvabase;
2053 		BUF_CHECK_MAPPED(bp);
2054 	}
2055 	return (0);
2056 }
2057 
2058 /*
2059  *	bufkva_reclaim:
2060  *
2061  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2062  *	callback that fires to avoid returning failure.
2063  */
2064 static void
2065 bufkva_reclaim(vmem_t *vmem, int flags)
2066 {
2067 	bool done;
2068 	int q;
2069 	int i;
2070 
2071 	done = false;
2072 	for (i = 0; i < 5; i++) {
2073 		for (q = 0; q < buf_domains; q++)
2074 			if (buf_recycle(&bdomain[q], true) != 0)
2075 				done = true;
2076 		if (done)
2077 			break;
2078 	}
2079 	return;
2080 }
2081 
2082 /*
2083  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2084  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2085  * the buffer is valid and we do not have to do anything.
2086  */
2087 static void
2088 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2089     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2090 {
2091 	struct buf *rabp;
2092 	int i;
2093 
2094 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2095 		if (inmem(vp, *rablkno))
2096 			continue;
2097 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2098 		if ((rabp->b_flags & B_CACHE) != 0) {
2099 			brelse(rabp);
2100 			continue;
2101 		}
2102 		if (!TD_IS_IDLETHREAD(curthread)) {
2103 #ifdef RACCT
2104 			if (racct_enable) {
2105 				PROC_LOCK(curproc);
2106 				racct_add_buf(curproc, rabp, 0);
2107 				PROC_UNLOCK(curproc);
2108 			}
2109 #endif /* RACCT */
2110 			curthread->td_ru.ru_inblock++;
2111 		}
2112 		rabp->b_flags |= B_ASYNC;
2113 		rabp->b_flags &= ~B_INVAL;
2114 		if ((flags & GB_CKHASH) != 0) {
2115 			rabp->b_flags |= B_CKHASH;
2116 			rabp->b_ckhashcalc = ckhashfunc;
2117 		}
2118 		rabp->b_ioflags &= ~BIO_ERROR;
2119 		rabp->b_iocmd = BIO_READ;
2120 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2121 			rabp->b_rcred = crhold(cred);
2122 		vfs_busy_pages(rabp, 0);
2123 		BUF_KERNPROC(rabp);
2124 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2125 		bstrategy(rabp);
2126 	}
2127 }
2128 
2129 /*
2130  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2131  *
2132  * Get a buffer with the specified data.  Look in the cache first.  We
2133  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2134  * is set, the buffer is valid and we do not have to do anything, see
2135  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2136  *
2137  * Always return a NULL buffer pointer (in bpp) when returning an error.
2138  */
2139 int
2140 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
2141     int *rabsize, int cnt, struct ucred *cred, int flags,
2142     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2143 {
2144 	struct buf *bp;
2145 	int readwait, rv;
2146 
2147 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2148 	/*
2149 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
2150 	 */
2151 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
2152 	if (bp == NULL)
2153 		return (EBUSY);
2154 
2155 	/*
2156 	 * If not found in cache, do some I/O
2157 	 */
2158 	readwait = 0;
2159 	if ((bp->b_flags & B_CACHE) == 0) {
2160 		if (!TD_IS_IDLETHREAD(curthread)) {
2161 #ifdef RACCT
2162 			if (racct_enable) {
2163 				PROC_LOCK(curproc);
2164 				racct_add_buf(curproc, bp, 0);
2165 				PROC_UNLOCK(curproc);
2166 			}
2167 #endif /* RACCT */
2168 			curthread->td_ru.ru_inblock++;
2169 		}
2170 		bp->b_iocmd = BIO_READ;
2171 		bp->b_flags &= ~B_INVAL;
2172 		if ((flags & GB_CKHASH) != 0) {
2173 			bp->b_flags |= B_CKHASH;
2174 			bp->b_ckhashcalc = ckhashfunc;
2175 		}
2176 		bp->b_ioflags &= ~BIO_ERROR;
2177 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2178 			bp->b_rcred = crhold(cred);
2179 		vfs_busy_pages(bp, 0);
2180 		bp->b_iooffset = dbtob(bp->b_blkno);
2181 		bstrategy(bp);
2182 		++readwait;
2183 	}
2184 
2185 	/*
2186 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2187 	 */
2188 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2189 
2190 	rv = 0;
2191 	if (readwait) {
2192 		rv = bufwait(bp);
2193 		if (rv != 0) {
2194 			brelse(bp);
2195 			*bpp = NULL;
2196 		}
2197 	}
2198 	return (rv);
2199 }
2200 
2201 /*
2202  * Write, release buffer on completion.  (Done by iodone
2203  * if async).  Do not bother writing anything if the buffer
2204  * is invalid.
2205  *
2206  * Note that we set B_CACHE here, indicating that buffer is
2207  * fully valid and thus cacheable.  This is true even of NFS
2208  * now so we set it generally.  This could be set either here
2209  * or in biodone() since the I/O is synchronous.  We put it
2210  * here.
2211  */
2212 int
2213 bufwrite(struct buf *bp)
2214 {
2215 	int oldflags;
2216 	struct vnode *vp;
2217 	long space;
2218 	int vp_md;
2219 
2220 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2221 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2222 		bp->b_flags |= B_INVAL | B_RELBUF;
2223 		bp->b_flags &= ~B_CACHE;
2224 		brelse(bp);
2225 		return (ENXIO);
2226 	}
2227 	if (bp->b_flags & B_INVAL) {
2228 		brelse(bp);
2229 		return (0);
2230 	}
2231 
2232 	if (bp->b_flags & B_BARRIER)
2233 		barrierwrites++;
2234 
2235 	oldflags = bp->b_flags;
2236 
2237 	BUF_ASSERT_HELD(bp);
2238 
2239 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2240 	    ("FFS background buffer should not get here %p", bp));
2241 
2242 	vp = bp->b_vp;
2243 	if (vp)
2244 		vp_md = vp->v_vflag & VV_MD;
2245 	else
2246 		vp_md = 0;
2247 
2248 	/*
2249 	 * Mark the buffer clean.  Increment the bufobj write count
2250 	 * before bundirty() call, to prevent other thread from seeing
2251 	 * empty dirty list and zero counter for writes in progress,
2252 	 * falsely indicating that the bufobj is clean.
2253 	 */
2254 	bufobj_wref(bp->b_bufobj);
2255 	bundirty(bp);
2256 
2257 	bp->b_flags &= ~B_DONE;
2258 	bp->b_ioflags &= ~BIO_ERROR;
2259 	bp->b_flags |= B_CACHE;
2260 	bp->b_iocmd = BIO_WRITE;
2261 
2262 	vfs_busy_pages(bp, 1);
2263 
2264 	/*
2265 	 * Normal bwrites pipeline writes
2266 	 */
2267 	bp->b_runningbufspace = bp->b_bufsize;
2268 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2269 
2270 	if (!TD_IS_IDLETHREAD(curthread)) {
2271 #ifdef RACCT
2272 		if (racct_enable) {
2273 			PROC_LOCK(curproc);
2274 			racct_add_buf(curproc, bp, 1);
2275 			PROC_UNLOCK(curproc);
2276 		}
2277 #endif /* RACCT */
2278 		curthread->td_ru.ru_oublock++;
2279 	}
2280 	if (oldflags & B_ASYNC)
2281 		BUF_KERNPROC(bp);
2282 	bp->b_iooffset = dbtob(bp->b_blkno);
2283 	buf_track(bp, __func__);
2284 	bstrategy(bp);
2285 
2286 	if ((oldflags & B_ASYNC) == 0) {
2287 		int rtval = bufwait(bp);
2288 		brelse(bp);
2289 		return (rtval);
2290 	} else if (space > hirunningspace) {
2291 		/*
2292 		 * don't allow the async write to saturate the I/O
2293 		 * system.  We will not deadlock here because
2294 		 * we are blocking waiting for I/O that is already in-progress
2295 		 * to complete. We do not block here if it is the update
2296 		 * or syncer daemon trying to clean up as that can lead
2297 		 * to deadlock.
2298 		 */
2299 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2300 			waitrunningbufspace();
2301 	}
2302 
2303 	return (0);
2304 }
2305 
2306 void
2307 bufbdflush(struct bufobj *bo, struct buf *bp)
2308 {
2309 	struct buf *nbp;
2310 
2311 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2312 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2313 		altbufferflushes++;
2314 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2315 		BO_LOCK(bo);
2316 		/*
2317 		 * Try to find a buffer to flush.
2318 		 */
2319 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2320 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2321 			    BUF_LOCK(nbp,
2322 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2323 				continue;
2324 			if (bp == nbp)
2325 				panic("bdwrite: found ourselves");
2326 			BO_UNLOCK(bo);
2327 			/* Don't countdeps with the bo lock held. */
2328 			if (buf_countdeps(nbp, 0)) {
2329 				BO_LOCK(bo);
2330 				BUF_UNLOCK(nbp);
2331 				continue;
2332 			}
2333 			if (nbp->b_flags & B_CLUSTEROK) {
2334 				vfs_bio_awrite(nbp);
2335 			} else {
2336 				bremfree(nbp);
2337 				bawrite(nbp);
2338 			}
2339 			dirtybufferflushes++;
2340 			break;
2341 		}
2342 		if (nbp == NULL)
2343 			BO_UNLOCK(bo);
2344 	}
2345 }
2346 
2347 /*
2348  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2349  * anything if the buffer is marked invalid.
2350  *
2351  * Note that since the buffer must be completely valid, we can safely
2352  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2353  * biodone() in order to prevent getblk from writing the buffer
2354  * out synchronously.
2355  */
2356 void
2357 bdwrite(struct buf *bp)
2358 {
2359 	struct thread *td = curthread;
2360 	struct vnode *vp;
2361 	struct bufobj *bo;
2362 
2363 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2364 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2365 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2366 	    ("Barrier request in delayed write %p", bp));
2367 	BUF_ASSERT_HELD(bp);
2368 
2369 	if (bp->b_flags & B_INVAL) {
2370 		brelse(bp);
2371 		return;
2372 	}
2373 
2374 	/*
2375 	 * If we have too many dirty buffers, don't create any more.
2376 	 * If we are wildly over our limit, then force a complete
2377 	 * cleanup. Otherwise, just keep the situation from getting
2378 	 * out of control. Note that we have to avoid a recursive
2379 	 * disaster and not try to clean up after our own cleanup!
2380 	 */
2381 	vp = bp->b_vp;
2382 	bo = bp->b_bufobj;
2383 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2384 		td->td_pflags |= TDP_INBDFLUSH;
2385 		BO_BDFLUSH(bo, bp);
2386 		td->td_pflags &= ~TDP_INBDFLUSH;
2387 	} else
2388 		recursiveflushes++;
2389 
2390 	bdirty(bp);
2391 	/*
2392 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2393 	 * true even of NFS now.
2394 	 */
2395 	bp->b_flags |= B_CACHE;
2396 
2397 	/*
2398 	 * This bmap keeps the system from needing to do the bmap later,
2399 	 * perhaps when the system is attempting to do a sync.  Since it
2400 	 * is likely that the indirect block -- or whatever other datastructure
2401 	 * that the filesystem needs is still in memory now, it is a good
2402 	 * thing to do this.  Note also, that if the pageout daemon is
2403 	 * requesting a sync -- there might not be enough memory to do
2404 	 * the bmap then...  So, this is important to do.
2405 	 */
2406 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2407 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2408 	}
2409 
2410 	buf_track(bp, __func__);
2411 
2412 	/*
2413 	 * Set the *dirty* buffer range based upon the VM system dirty
2414 	 * pages.
2415 	 *
2416 	 * Mark the buffer pages as clean.  We need to do this here to
2417 	 * satisfy the vnode_pager and the pageout daemon, so that it
2418 	 * thinks that the pages have been "cleaned".  Note that since
2419 	 * the pages are in a delayed write buffer -- the VFS layer
2420 	 * "will" see that the pages get written out on the next sync,
2421 	 * or perhaps the cluster will be completed.
2422 	 */
2423 	vfs_clean_pages_dirty_buf(bp);
2424 	bqrelse(bp);
2425 
2426 	/*
2427 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2428 	 * due to the softdep code.
2429 	 */
2430 }
2431 
2432 /*
2433  *	bdirty:
2434  *
2435  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2436  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2437  *	itself to properly update it in the dirty/clean lists.  We mark it
2438  *	B_DONE to ensure that any asynchronization of the buffer properly
2439  *	clears B_DONE ( else a panic will occur later ).
2440  *
2441  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2442  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2443  *	should only be called if the buffer is known-good.
2444  *
2445  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2446  *	count.
2447  *
2448  *	The buffer must be on QUEUE_NONE.
2449  */
2450 void
2451 bdirty(struct buf *bp)
2452 {
2453 
2454 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2455 	    bp, bp->b_vp, bp->b_flags);
2456 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2457 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2458 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2459 	BUF_ASSERT_HELD(bp);
2460 	bp->b_flags &= ~(B_RELBUF);
2461 	bp->b_iocmd = BIO_WRITE;
2462 
2463 	if ((bp->b_flags & B_DELWRI) == 0) {
2464 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2465 		reassignbuf(bp);
2466 		bdirtyadd(bp);
2467 	}
2468 }
2469 
2470 /*
2471  *	bundirty:
2472  *
2473  *	Clear B_DELWRI for buffer.
2474  *
2475  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2476  *	count.
2477  *
2478  *	The buffer must be on QUEUE_NONE.
2479  */
2480 
2481 void
2482 bundirty(struct buf *bp)
2483 {
2484 
2485 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2486 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2487 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2488 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2489 	BUF_ASSERT_HELD(bp);
2490 
2491 	if (bp->b_flags & B_DELWRI) {
2492 		bp->b_flags &= ~B_DELWRI;
2493 		reassignbuf(bp);
2494 		bdirtysub(bp);
2495 	}
2496 	/*
2497 	 * Since it is now being written, we can clear its deferred write flag.
2498 	 */
2499 	bp->b_flags &= ~B_DEFERRED;
2500 }
2501 
2502 /*
2503  *	bawrite:
2504  *
2505  *	Asynchronous write.  Start output on a buffer, but do not wait for
2506  *	it to complete.  The buffer is released when the output completes.
2507  *
2508  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2509  *	B_INVAL buffers.  Not us.
2510  */
2511 void
2512 bawrite(struct buf *bp)
2513 {
2514 
2515 	bp->b_flags |= B_ASYNC;
2516 	(void) bwrite(bp);
2517 }
2518 
2519 /*
2520  *	babarrierwrite:
2521  *
2522  *	Asynchronous barrier write.  Start output on a buffer, but do not
2523  *	wait for it to complete.  Place a write barrier after this write so
2524  *	that this buffer and all buffers written before it are committed to
2525  *	the disk before any buffers written after this write are committed
2526  *	to the disk.  The buffer is released when the output completes.
2527  */
2528 void
2529 babarrierwrite(struct buf *bp)
2530 {
2531 
2532 	bp->b_flags |= B_ASYNC | B_BARRIER;
2533 	(void) bwrite(bp);
2534 }
2535 
2536 /*
2537  *	bbarrierwrite:
2538  *
2539  *	Synchronous barrier write.  Start output on a buffer and wait for
2540  *	it to complete.  Place a write barrier after this write so that
2541  *	this buffer and all buffers written before it are committed to
2542  *	the disk before any buffers written after this write are committed
2543  *	to the disk.  The buffer is released when the output completes.
2544  */
2545 int
2546 bbarrierwrite(struct buf *bp)
2547 {
2548 
2549 	bp->b_flags |= B_BARRIER;
2550 	return (bwrite(bp));
2551 }
2552 
2553 /*
2554  *	bwillwrite:
2555  *
2556  *	Called prior to the locking of any vnodes when we are expecting to
2557  *	write.  We do not want to starve the buffer cache with too many
2558  *	dirty buffers so we block here.  By blocking prior to the locking
2559  *	of any vnodes we attempt to avoid the situation where a locked vnode
2560  *	prevents the various system daemons from flushing related buffers.
2561  */
2562 void
2563 bwillwrite(void)
2564 {
2565 
2566 	if (buf_dirty_count_severe()) {
2567 		mtx_lock(&bdirtylock);
2568 		while (buf_dirty_count_severe()) {
2569 			bdirtywait = 1;
2570 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2571 			    "flswai", 0);
2572 		}
2573 		mtx_unlock(&bdirtylock);
2574 	}
2575 }
2576 
2577 /*
2578  * Return true if we have too many dirty buffers.
2579  */
2580 int
2581 buf_dirty_count_severe(void)
2582 {
2583 
2584 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2585 }
2586 
2587 /*
2588  *	brelse:
2589  *
2590  *	Release a busy buffer and, if requested, free its resources.  The
2591  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2592  *	to be accessed later as a cache entity or reused for other purposes.
2593  */
2594 void
2595 brelse(struct buf *bp)
2596 {
2597 	struct mount *v_mnt;
2598 	int qindex;
2599 
2600 	/*
2601 	 * Many functions erroneously call brelse with a NULL bp under rare
2602 	 * error conditions. Simply return when called with a NULL bp.
2603 	 */
2604 	if (bp == NULL)
2605 		return;
2606 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2607 	    bp, bp->b_vp, bp->b_flags);
2608 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2609 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2610 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2611 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2612 
2613 	if (BUF_LOCKRECURSED(bp)) {
2614 		/*
2615 		 * Do not process, in particular, do not handle the
2616 		 * B_INVAL/B_RELBUF and do not release to free list.
2617 		 */
2618 		BUF_UNLOCK(bp);
2619 		return;
2620 	}
2621 
2622 	if (bp->b_flags & B_MANAGED) {
2623 		bqrelse(bp);
2624 		return;
2625 	}
2626 
2627 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2628 		BO_LOCK(bp->b_bufobj);
2629 		bp->b_vflags &= ~BV_BKGRDERR;
2630 		BO_UNLOCK(bp->b_bufobj);
2631 		bdirty(bp);
2632 	}
2633 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2634 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2635 	    !(bp->b_flags & B_INVAL)) {
2636 		/*
2637 		 * Failed write, redirty.  All errors except ENXIO (which
2638 		 * means the device is gone) are treated as being
2639 		 * transient.
2640 		 *
2641 		 * XXX Treating EIO as transient is not correct; the
2642 		 * contract with the local storage device drivers is that
2643 		 * they will only return EIO once the I/O is no longer
2644 		 * retriable.  Network I/O also respects this through the
2645 		 * guarantees of TCP and/or the internal retries of NFS.
2646 		 * ENOMEM might be transient, but we also have no way of
2647 		 * knowing when its ok to retry/reschedule.  In general,
2648 		 * this entire case should be made obsolete through better
2649 		 * error handling/recovery and resource scheduling.
2650 		 *
2651 		 * Do this also for buffers that failed with ENXIO, but have
2652 		 * non-empty dependencies - the soft updates code might need
2653 		 * to access the buffer to untangle them.
2654 		 *
2655 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2656 		 */
2657 		bp->b_ioflags &= ~BIO_ERROR;
2658 		bdirty(bp);
2659 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2660 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2661 		/*
2662 		 * Either a failed read I/O, or we were asked to free or not
2663 		 * cache the buffer, or we failed to write to a device that's
2664 		 * no longer present.
2665 		 */
2666 		bp->b_flags |= B_INVAL;
2667 		if (!LIST_EMPTY(&bp->b_dep))
2668 			buf_deallocate(bp);
2669 		if (bp->b_flags & B_DELWRI)
2670 			bdirtysub(bp);
2671 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2672 		if ((bp->b_flags & B_VMIO) == 0) {
2673 			allocbuf(bp, 0);
2674 			if (bp->b_vp)
2675 				brelvp(bp);
2676 		}
2677 	}
2678 
2679 	/*
2680 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2681 	 * is called with B_DELWRI set, the underlying pages may wind up
2682 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2683 	 * because pages associated with a B_DELWRI bp are marked clean.
2684 	 *
2685 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2686 	 * if B_DELWRI is set.
2687 	 */
2688 	if (bp->b_flags & B_DELWRI)
2689 		bp->b_flags &= ~B_RELBUF;
2690 
2691 	/*
2692 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2693 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2694 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2695 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2696 	 *
2697 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2698 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2699 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2700 	 *
2701 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2702 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2703 	 * the commit state and we cannot afford to lose the buffer. If the
2704 	 * buffer has a background write in progress, we need to keep it
2705 	 * around to prevent it from being reconstituted and starting a second
2706 	 * background write.
2707 	 */
2708 
2709 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2710 
2711 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2712 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2713 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2714 	    vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2715 		vfs_vmio_invalidate(bp);
2716 		allocbuf(bp, 0);
2717 	}
2718 
2719 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2720 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2721 		allocbuf(bp, 0);
2722 		bp->b_flags &= ~B_NOREUSE;
2723 		if (bp->b_vp != NULL)
2724 			brelvp(bp);
2725 	}
2726 
2727 	/*
2728 	 * If the buffer has junk contents signal it and eventually
2729 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2730 	 * doesn't find it.
2731 	 */
2732 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2733 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2734 		bp->b_flags |= B_INVAL;
2735 	if (bp->b_flags & B_INVAL) {
2736 		if (bp->b_flags & B_DELWRI)
2737 			bundirty(bp);
2738 		if (bp->b_vp)
2739 			brelvp(bp);
2740 	}
2741 
2742 	buf_track(bp, __func__);
2743 
2744 	/* buffers with no memory */
2745 	if (bp->b_bufsize == 0) {
2746 		buf_free(bp);
2747 		return;
2748 	}
2749 	/* buffers with junk contents */
2750 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2751 	    (bp->b_ioflags & BIO_ERROR)) {
2752 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2753 		if (bp->b_vflags & BV_BKGRDINPROG)
2754 			panic("losing buffer 2");
2755 		qindex = QUEUE_CLEAN;
2756 		bp->b_flags |= B_AGE;
2757 	/* remaining buffers */
2758 	} else if (bp->b_flags & B_DELWRI)
2759 		qindex = QUEUE_DIRTY;
2760 	else
2761 		qindex = QUEUE_CLEAN;
2762 
2763 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2764 		panic("brelse: not dirty");
2765 
2766 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2767 	/* binsfree unlocks bp. */
2768 	binsfree(bp, qindex);
2769 }
2770 
2771 /*
2772  * Release a buffer back to the appropriate queue but do not try to free
2773  * it.  The buffer is expected to be used again soon.
2774  *
2775  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2776  * biodone() to requeue an async I/O on completion.  It is also used when
2777  * known good buffers need to be requeued but we think we may need the data
2778  * again soon.
2779  *
2780  * XXX we should be able to leave the B_RELBUF hint set on completion.
2781  */
2782 void
2783 bqrelse(struct buf *bp)
2784 {
2785 	int qindex;
2786 
2787 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2788 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2789 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2790 
2791 	qindex = QUEUE_NONE;
2792 	if (BUF_LOCKRECURSED(bp)) {
2793 		/* do not release to free list */
2794 		BUF_UNLOCK(bp);
2795 		return;
2796 	}
2797 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2798 
2799 	if (bp->b_flags & B_MANAGED) {
2800 		if (bp->b_flags & B_REMFREE)
2801 			bremfreef(bp);
2802 		goto out;
2803 	}
2804 
2805 	/* buffers with stale but valid contents */
2806 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2807 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2808 		BO_LOCK(bp->b_bufobj);
2809 		bp->b_vflags &= ~BV_BKGRDERR;
2810 		BO_UNLOCK(bp->b_bufobj);
2811 		qindex = QUEUE_DIRTY;
2812 	} else {
2813 		if ((bp->b_flags & B_DELWRI) == 0 &&
2814 		    (bp->b_xflags & BX_VNDIRTY))
2815 			panic("bqrelse: not dirty");
2816 		if ((bp->b_flags & B_NOREUSE) != 0) {
2817 			brelse(bp);
2818 			return;
2819 		}
2820 		qindex = QUEUE_CLEAN;
2821 	}
2822 	buf_track(bp, __func__);
2823 	/* binsfree unlocks bp. */
2824 	binsfree(bp, qindex);
2825 	return;
2826 
2827 out:
2828 	buf_track(bp, __func__);
2829 	/* unlock */
2830 	BUF_UNLOCK(bp);
2831 }
2832 
2833 /*
2834  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2835  * restore bogus pages.
2836  */
2837 static void
2838 vfs_vmio_iodone(struct buf *bp)
2839 {
2840 	vm_ooffset_t foff;
2841 	vm_page_t m;
2842 	vm_object_t obj;
2843 	struct vnode *vp;
2844 	int i, iosize, resid;
2845 	bool bogus;
2846 
2847 	obj = bp->b_bufobj->bo_object;
2848 	KASSERT(obj->paging_in_progress >= bp->b_npages,
2849 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2850 	    obj->paging_in_progress, bp->b_npages));
2851 
2852 	vp = bp->b_vp;
2853 	KASSERT(vp->v_holdcnt > 0,
2854 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2855 	KASSERT(vp->v_object != NULL,
2856 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2857 
2858 	foff = bp->b_offset;
2859 	KASSERT(bp->b_offset != NOOFFSET,
2860 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2861 
2862 	bogus = false;
2863 	iosize = bp->b_bcount - bp->b_resid;
2864 	VM_OBJECT_WLOCK(obj);
2865 	for (i = 0; i < bp->b_npages; i++) {
2866 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2867 		if (resid > iosize)
2868 			resid = iosize;
2869 
2870 		/*
2871 		 * cleanup bogus pages, restoring the originals
2872 		 */
2873 		m = bp->b_pages[i];
2874 		if (m == bogus_page) {
2875 			bogus = true;
2876 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2877 			if (m == NULL)
2878 				panic("biodone: page disappeared!");
2879 			bp->b_pages[i] = m;
2880 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2881 			/*
2882 			 * In the write case, the valid and clean bits are
2883 			 * already changed correctly ( see bdwrite() ), so we
2884 			 * only need to do this here in the read case.
2885 			 */
2886 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2887 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2888 			    "has unexpected dirty bits", m));
2889 			vfs_page_set_valid(bp, foff, m);
2890 		}
2891 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2892 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2893 		    (intmax_t)foff, (uintmax_t)m->pindex));
2894 
2895 		vm_page_sunbusy(m);
2896 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2897 		iosize -= resid;
2898 	}
2899 	vm_object_pip_wakeupn(obj, bp->b_npages);
2900 	VM_OBJECT_WUNLOCK(obj);
2901 	if (bogus && buf_mapped(bp)) {
2902 		BUF_CHECK_MAPPED(bp);
2903 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2904 		    bp->b_pages, bp->b_npages);
2905 	}
2906 }
2907 
2908 /*
2909  * Unwire a page held by a buf and place it on the appropriate vm queue.
2910  */
2911 static void
2912 vfs_vmio_unwire(struct buf *bp, vm_page_t m)
2913 {
2914 	bool freed;
2915 
2916 	vm_page_lock(m);
2917 	if (vm_page_unwire_noq(m)) {
2918 		/*
2919 		 * Determine if the page should be freed before adding
2920 		 * it to the inactive queue.
2921 		 */
2922 		if (m->valid == 0) {
2923 			freed = !vm_page_busied(m);
2924 			if (freed)
2925 				vm_page_free(m);
2926 		} else if ((bp->b_flags & B_DIRECT) != 0)
2927 			freed = vm_page_try_to_free(m);
2928 		else
2929 			freed = false;
2930 		if (!freed) {
2931 			/*
2932 			 * If the page is unlikely to be reused, let the
2933 			 * VM know.  Otherwise, maintain LRU.
2934 			 */
2935 			if ((bp->b_flags & B_NOREUSE) != 0)
2936 				vm_page_deactivate_noreuse(m);
2937 			else if (m->queue == PQ_ACTIVE)
2938 				vm_page_reference(m);
2939 			else
2940 				vm_page_deactivate(m);
2941 		}
2942 	}
2943 	vm_page_unlock(m);
2944 }
2945 
2946 /*
2947  * Perform page invalidation when a buffer is released.  The fully invalid
2948  * pages will be reclaimed later in vfs_vmio_truncate().
2949  */
2950 static void
2951 vfs_vmio_invalidate(struct buf *bp)
2952 {
2953 	vm_object_t obj;
2954 	vm_page_t m;
2955 	int i, resid, poffset, presid;
2956 
2957 	if (buf_mapped(bp)) {
2958 		BUF_CHECK_MAPPED(bp);
2959 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2960 	} else
2961 		BUF_CHECK_UNMAPPED(bp);
2962 	/*
2963 	 * Get the base offset and length of the buffer.  Note that
2964 	 * in the VMIO case if the buffer block size is not
2965 	 * page-aligned then b_data pointer may not be page-aligned.
2966 	 * But our b_pages[] array *IS* page aligned.
2967 	 *
2968 	 * block sizes less then DEV_BSIZE (usually 512) are not
2969 	 * supported due to the page granularity bits (m->valid,
2970 	 * m->dirty, etc...).
2971 	 *
2972 	 * See man buf(9) for more information
2973 	 */
2974 	obj = bp->b_bufobj->bo_object;
2975 	resid = bp->b_bufsize;
2976 	poffset = bp->b_offset & PAGE_MASK;
2977 	VM_OBJECT_WLOCK(obj);
2978 	for (i = 0; i < bp->b_npages; i++) {
2979 		m = bp->b_pages[i];
2980 		if (m == bogus_page)
2981 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2982 		bp->b_pages[i] = NULL;
2983 
2984 		presid = resid > (PAGE_SIZE - poffset) ?
2985 		    (PAGE_SIZE - poffset) : resid;
2986 		KASSERT(presid >= 0, ("brelse: extra page"));
2987 		while (vm_page_xbusied(m)) {
2988 			vm_page_lock(m);
2989 			VM_OBJECT_WUNLOCK(obj);
2990 			vm_page_busy_sleep(m, "mbncsh", true);
2991 			VM_OBJECT_WLOCK(obj);
2992 		}
2993 		if (pmap_page_wired_mappings(m) == 0)
2994 			vm_page_set_invalid(m, poffset, presid);
2995 		vfs_vmio_unwire(bp, m);
2996 		resid -= presid;
2997 		poffset = 0;
2998 	}
2999 	VM_OBJECT_WUNLOCK(obj);
3000 	bp->b_npages = 0;
3001 }
3002 
3003 /*
3004  * Page-granular truncation of an existing VMIO buffer.
3005  */
3006 static void
3007 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3008 {
3009 	vm_object_t obj;
3010 	vm_page_t m;
3011 	int i;
3012 
3013 	if (bp->b_npages == desiredpages)
3014 		return;
3015 
3016 	if (buf_mapped(bp)) {
3017 		BUF_CHECK_MAPPED(bp);
3018 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3019 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3020 	} else
3021 		BUF_CHECK_UNMAPPED(bp);
3022 	obj = bp->b_bufobj->bo_object;
3023 	if (obj != NULL)
3024 		VM_OBJECT_WLOCK(obj);
3025 	for (i = desiredpages; i < bp->b_npages; i++) {
3026 		m = bp->b_pages[i];
3027 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3028 		bp->b_pages[i] = NULL;
3029 		vfs_vmio_unwire(bp, m);
3030 	}
3031 	if (obj != NULL)
3032 		VM_OBJECT_WUNLOCK(obj);
3033 	bp->b_npages = desiredpages;
3034 }
3035 
3036 /*
3037  * Byte granular extension of VMIO buffers.
3038  */
3039 static void
3040 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3041 {
3042 	/*
3043 	 * We are growing the buffer, possibly in a
3044 	 * byte-granular fashion.
3045 	 */
3046 	vm_object_t obj;
3047 	vm_offset_t toff;
3048 	vm_offset_t tinc;
3049 	vm_page_t m;
3050 
3051 	/*
3052 	 * Step 1, bring in the VM pages from the object, allocating
3053 	 * them if necessary.  We must clear B_CACHE if these pages
3054 	 * are not valid for the range covered by the buffer.
3055 	 */
3056 	obj = bp->b_bufobj->bo_object;
3057 	VM_OBJECT_WLOCK(obj);
3058 	if (bp->b_npages < desiredpages) {
3059 		/*
3060 		 * We must allocate system pages since blocking
3061 		 * here could interfere with paging I/O, no
3062 		 * matter which process we are.
3063 		 *
3064 		 * Only exclusive busy can be tested here.
3065 		 * Blocking on shared busy might lead to
3066 		 * deadlocks once allocbuf() is called after
3067 		 * pages are vfs_busy_pages().
3068 		 */
3069 		(void)vm_page_grab_pages(obj,
3070 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3071 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3072 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3073 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3074 		bp->b_npages = desiredpages;
3075 	}
3076 
3077 	/*
3078 	 * Step 2.  We've loaded the pages into the buffer,
3079 	 * we have to figure out if we can still have B_CACHE
3080 	 * set.  Note that B_CACHE is set according to the
3081 	 * byte-granular range ( bcount and size ), not the
3082 	 * aligned range ( newbsize ).
3083 	 *
3084 	 * The VM test is against m->valid, which is DEV_BSIZE
3085 	 * aligned.  Needless to say, the validity of the data
3086 	 * needs to also be DEV_BSIZE aligned.  Note that this
3087 	 * fails with NFS if the server or some other client
3088 	 * extends the file's EOF.  If our buffer is resized,
3089 	 * B_CACHE may remain set! XXX
3090 	 */
3091 	toff = bp->b_bcount;
3092 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3093 	while ((bp->b_flags & B_CACHE) && toff < size) {
3094 		vm_pindex_t pi;
3095 
3096 		if (tinc > (size - toff))
3097 			tinc = size - toff;
3098 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3099 		m = bp->b_pages[pi];
3100 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3101 		toff += tinc;
3102 		tinc = PAGE_SIZE;
3103 	}
3104 	VM_OBJECT_WUNLOCK(obj);
3105 
3106 	/*
3107 	 * Step 3, fixup the KVA pmap.
3108 	 */
3109 	if (buf_mapped(bp))
3110 		bpmap_qenter(bp);
3111 	else
3112 		BUF_CHECK_UNMAPPED(bp);
3113 }
3114 
3115 /*
3116  * Check to see if a block at a particular lbn is available for a clustered
3117  * write.
3118  */
3119 static int
3120 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3121 {
3122 	struct buf *bpa;
3123 	int match;
3124 
3125 	match = 0;
3126 
3127 	/* If the buf isn't in core skip it */
3128 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3129 		return (0);
3130 
3131 	/* If the buf is busy we don't want to wait for it */
3132 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3133 		return (0);
3134 
3135 	/* Only cluster with valid clusterable delayed write buffers */
3136 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3137 	    (B_DELWRI | B_CLUSTEROK))
3138 		goto done;
3139 
3140 	if (bpa->b_bufsize != size)
3141 		goto done;
3142 
3143 	/*
3144 	 * Check to see if it is in the expected place on disk and that the
3145 	 * block has been mapped.
3146 	 */
3147 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3148 		match = 1;
3149 done:
3150 	BUF_UNLOCK(bpa);
3151 	return (match);
3152 }
3153 
3154 /*
3155  *	vfs_bio_awrite:
3156  *
3157  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3158  *	This is much better then the old way of writing only one buffer at
3159  *	a time.  Note that we may not be presented with the buffers in the
3160  *	correct order, so we search for the cluster in both directions.
3161  */
3162 int
3163 vfs_bio_awrite(struct buf *bp)
3164 {
3165 	struct bufobj *bo;
3166 	int i;
3167 	int j;
3168 	daddr_t lblkno = bp->b_lblkno;
3169 	struct vnode *vp = bp->b_vp;
3170 	int ncl;
3171 	int nwritten;
3172 	int size;
3173 	int maxcl;
3174 	int gbflags;
3175 
3176 	bo = &vp->v_bufobj;
3177 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3178 	/*
3179 	 * right now we support clustered writing only to regular files.  If
3180 	 * we find a clusterable block we could be in the middle of a cluster
3181 	 * rather then at the beginning.
3182 	 */
3183 	if ((vp->v_type == VREG) &&
3184 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3185 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3186 
3187 		size = vp->v_mount->mnt_stat.f_iosize;
3188 		maxcl = MAXPHYS / size;
3189 
3190 		BO_RLOCK(bo);
3191 		for (i = 1; i < maxcl; i++)
3192 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3193 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3194 				break;
3195 
3196 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3197 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3198 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3199 				break;
3200 		BO_RUNLOCK(bo);
3201 		--j;
3202 		ncl = i + j;
3203 		/*
3204 		 * this is a possible cluster write
3205 		 */
3206 		if (ncl != 1) {
3207 			BUF_UNLOCK(bp);
3208 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3209 			    gbflags);
3210 			return (nwritten);
3211 		}
3212 	}
3213 	bremfree(bp);
3214 	bp->b_flags |= B_ASYNC;
3215 	/*
3216 	 * default (old) behavior, writing out only one block
3217 	 *
3218 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3219 	 */
3220 	nwritten = bp->b_bufsize;
3221 	(void) bwrite(bp);
3222 
3223 	return (nwritten);
3224 }
3225 
3226 /*
3227  *	getnewbuf_kva:
3228  *
3229  *	Allocate KVA for an empty buf header according to gbflags.
3230  */
3231 static int
3232 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3233 {
3234 
3235 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3236 		/*
3237 		 * In order to keep fragmentation sane we only allocate kva
3238 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3239 		 */
3240 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3241 
3242 		if (maxsize != bp->b_kvasize &&
3243 		    bufkva_alloc(bp, maxsize, gbflags))
3244 			return (ENOSPC);
3245 	}
3246 	return (0);
3247 }
3248 
3249 /*
3250  *	getnewbuf:
3251  *
3252  *	Find and initialize a new buffer header, freeing up existing buffers
3253  *	in the bufqueues as necessary.  The new buffer is returned locked.
3254  *
3255  *	We block if:
3256  *		We have insufficient buffer headers
3257  *		We have insufficient buffer space
3258  *		buffer_arena is too fragmented ( space reservation fails )
3259  *		If we have to flush dirty buffers ( but we try to avoid this )
3260  *
3261  *	The caller is responsible for releasing the reserved bufspace after
3262  *	allocbuf() is called.
3263  */
3264 static struct buf *
3265 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3266 {
3267 	struct bufdomain *bd;
3268 	struct buf *bp;
3269 	bool metadata, reserved;
3270 
3271 	bp = NULL;
3272 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3273 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3274 	if (!unmapped_buf_allowed)
3275 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3276 
3277 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3278 	    vp->v_type == VCHR)
3279 		metadata = true;
3280 	else
3281 		metadata = false;
3282 	if (vp == NULL)
3283 		bd = &bdomain[0];
3284 	else
3285 		bd = &bdomain[vp->v_bufobj.bo_domain];
3286 
3287 	counter_u64_add(getnewbufcalls, 1);
3288 	reserved = false;
3289 	do {
3290 		if (reserved == false &&
3291 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3292 			counter_u64_add(getnewbufrestarts, 1);
3293 			continue;
3294 		}
3295 		reserved = true;
3296 		if ((bp = buf_alloc(bd)) == NULL) {
3297 			counter_u64_add(getnewbufrestarts, 1);
3298 			continue;
3299 		}
3300 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3301 			return (bp);
3302 		break;
3303 	} while (buf_recycle(bd, false) == 0);
3304 
3305 	if (reserved)
3306 		bufspace_release(bd, maxsize);
3307 	if (bp != NULL) {
3308 		bp->b_flags |= B_INVAL;
3309 		brelse(bp);
3310 	}
3311 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3312 
3313 	return (NULL);
3314 }
3315 
3316 /*
3317  *	buf_daemon:
3318  *
3319  *	buffer flushing daemon.  Buffers are normally flushed by the
3320  *	update daemon but if it cannot keep up this process starts to
3321  *	take the load in an attempt to prevent getnewbuf() from blocking.
3322  */
3323 static struct kproc_desc buf_kp = {
3324 	"bufdaemon",
3325 	buf_daemon,
3326 	&bufdaemonproc
3327 };
3328 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3329 
3330 static int
3331 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3332 {
3333 	int flushed;
3334 
3335 	flushed = flushbufqueues(vp, bd, target, 0);
3336 	if (flushed == 0) {
3337 		/*
3338 		 * Could not find any buffers without rollback
3339 		 * dependencies, so just write the first one
3340 		 * in the hopes of eventually making progress.
3341 		 */
3342 		if (vp != NULL && target > 2)
3343 			target /= 2;
3344 		flushbufqueues(vp, bd, target, 1);
3345 	}
3346 	return (flushed);
3347 }
3348 
3349 static void
3350 buf_daemon()
3351 {
3352 	struct bufdomain *bd;
3353 	int speedupreq;
3354 	int lodirty;
3355 	int i;
3356 
3357 	/*
3358 	 * This process needs to be suspended prior to shutdown sync.
3359 	 */
3360 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
3361 	    SHUTDOWN_PRI_LAST);
3362 
3363 	/*
3364 	 * Start the buf clean daemons as children threads.
3365 	 */
3366 	for (i = 0 ; i < buf_domains; i++) {
3367 		int error;
3368 
3369 		error = kthread_add((void (*)(void *))bufspace_daemon,
3370 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3371 		if (error)
3372 			panic("error %d spawning bufspace daemon", error);
3373 	}
3374 
3375 	/*
3376 	 * This process is allowed to take the buffer cache to the limit
3377 	 */
3378 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3379 	mtx_lock(&bdlock);
3380 	for (;;) {
3381 		bd_request = 0;
3382 		mtx_unlock(&bdlock);
3383 
3384 		kproc_suspend_check(bufdaemonproc);
3385 
3386 		/*
3387 		 * Save speedupreq for this pass and reset to capture new
3388 		 * requests.
3389 		 */
3390 		speedupreq = bd_speedupreq;
3391 		bd_speedupreq = 0;
3392 
3393 		/*
3394 		 * Flush each domain sequentially according to its level and
3395 		 * the speedup request.
3396 		 */
3397 		for (i = 0; i < buf_domains; i++) {
3398 			bd = &bdomain[i];
3399 			if (speedupreq)
3400 				lodirty = bd->bd_numdirtybuffers / 2;
3401 			else
3402 				lodirty = bd->bd_lodirtybuffers;
3403 			while (bd->bd_numdirtybuffers > lodirty) {
3404 				if (buf_flush(NULL, bd,
3405 				    bd->bd_numdirtybuffers - lodirty) == 0)
3406 					break;
3407 				kern_yield(PRI_USER);
3408 			}
3409 		}
3410 
3411 		/*
3412 		 * Only clear bd_request if we have reached our low water
3413 		 * mark.  The buf_daemon normally waits 1 second and
3414 		 * then incrementally flushes any dirty buffers that have
3415 		 * built up, within reason.
3416 		 *
3417 		 * If we were unable to hit our low water mark and couldn't
3418 		 * find any flushable buffers, we sleep for a short period
3419 		 * to avoid endless loops on unlockable buffers.
3420 		 */
3421 		mtx_lock(&bdlock);
3422 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3423 			/*
3424 			 * We reached our low water mark, reset the
3425 			 * request and sleep until we are needed again.
3426 			 * The sleep is just so the suspend code works.
3427 			 */
3428 			bd_request = 0;
3429 			/*
3430 			 * Do an extra wakeup in case dirty threshold
3431 			 * changed via sysctl and the explicit transition
3432 			 * out of shortfall was missed.
3433 			 */
3434 			bdirtywakeup();
3435 			if (runningbufspace <= lorunningspace)
3436 				runningwakeup();
3437 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3438 		} else {
3439 			/*
3440 			 * We couldn't find any flushable dirty buffers but
3441 			 * still have too many dirty buffers, we
3442 			 * have to sleep and try again.  (rare)
3443 			 */
3444 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3445 		}
3446 	}
3447 }
3448 
3449 /*
3450  *	flushbufqueues:
3451  *
3452  *	Try to flush a buffer in the dirty queue.  We must be careful to
3453  *	free up B_INVAL buffers instead of write them, which NFS is
3454  *	particularly sensitive to.
3455  */
3456 static int flushwithdeps = 0;
3457 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
3458     0, "Number of buffers flushed with dependecies that require rollbacks");
3459 
3460 static int
3461 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3462     int flushdeps)
3463 {
3464 	struct bufqueue *bq;
3465 	struct buf *sentinel;
3466 	struct vnode *vp;
3467 	struct mount *mp;
3468 	struct buf *bp;
3469 	int hasdeps;
3470 	int flushed;
3471 	int error;
3472 	bool unlock;
3473 
3474 	flushed = 0;
3475 	bq = &bd->bd_dirtyq;
3476 	bp = NULL;
3477 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3478 	sentinel->b_qindex = QUEUE_SENTINEL;
3479 	BQ_LOCK(bq);
3480 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3481 	BQ_UNLOCK(bq);
3482 	while (flushed != target) {
3483 		maybe_yield();
3484 		BQ_LOCK(bq);
3485 		bp = TAILQ_NEXT(sentinel, b_freelist);
3486 		if (bp != NULL) {
3487 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3488 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3489 			    b_freelist);
3490 		} else {
3491 			BQ_UNLOCK(bq);
3492 			break;
3493 		}
3494 		/*
3495 		 * Skip sentinels inserted by other invocations of the
3496 		 * flushbufqueues(), taking care to not reorder them.
3497 		 *
3498 		 * Only flush the buffers that belong to the
3499 		 * vnode locked by the curthread.
3500 		 */
3501 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3502 		    bp->b_vp != lvp)) {
3503 			BQ_UNLOCK(bq);
3504 			continue;
3505 		}
3506 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3507 		BQ_UNLOCK(bq);
3508 		if (error != 0)
3509 			continue;
3510 
3511 		/*
3512 		 * BKGRDINPROG can only be set with the buf and bufobj
3513 		 * locks both held.  We tolerate a race to clear it here.
3514 		 */
3515 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3516 		    (bp->b_flags & B_DELWRI) == 0) {
3517 			BUF_UNLOCK(bp);
3518 			continue;
3519 		}
3520 		if (bp->b_flags & B_INVAL) {
3521 			bremfreef(bp);
3522 			brelse(bp);
3523 			flushed++;
3524 			continue;
3525 		}
3526 
3527 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3528 			if (flushdeps == 0) {
3529 				BUF_UNLOCK(bp);
3530 				continue;
3531 			}
3532 			hasdeps = 1;
3533 		} else
3534 			hasdeps = 0;
3535 		/*
3536 		 * We must hold the lock on a vnode before writing
3537 		 * one of its buffers. Otherwise we may confuse, or
3538 		 * in the case of a snapshot vnode, deadlock the
3539 		 * system.
3540 		 *
3541 		 * The lock order here is the reverse of the normal
3542 		 * of vnode followed by buf lock.  This is ok because
3543 		 * the NOWAIT will prevent deadlock.
3544 		 */
3545 		vp = bp->b_vp;
3546 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3547 			BUF_UNLOCK(bp);
3548 			continue;
3549 		}
3550 		if (lvp == NULL) {
3551 			unlock = true;
3552 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3553 		} else {
3554 			ASSERT_VOP_LOCKED(vp, "getbuf");
3555 			unlock = false;
3556 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3557 			    vn_lock(vp, LK_TRYUPGRADE);
3558 		}
3559 		if (error == 0) {
3560 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3561 			    bp, bp->b_vp, bp->b_flags);
3562 			if (curproc == bufdaemonproc) {
3563 				vfs_bio_awrite(bp);
3564 			} else {
3565 				bremfree(bp);
3566 				bwrite(bp);
3567 				counter_u64_add(notbufdflushes, 1);
3568 			}
3569 			vn_finished_write(mp);
3570 			if (unlock)
3571 				VOP_UNLOCK(vp, 0);
3572 			flushwithdeps += hasdeps;
3573 			flushed++;
3574 
3575 			/*
3576 			 * Sleeping on runningbufspace while holding
3577 			 * vnode lock leads to deadlock.
3578 			 */
3579 			if (curproc == bufdaemonproc &&
3580 			    runningbufspace > hirunningspace)
3581 				waitrunningbufspace();
3582 			continue;
3583 		}
3584 		vn_finished_write(mp);
3585 		BUF_UNLOCK(bp);
3586 	}
3587 	BQ_LOCK(bq);
3588 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3589 	BQ_UNLOCK(bq);
3590 	free(sentinel, M_TEMP);
3591 	return (flushed);
3592 }
3593 
3594 /*
3595  * Check to see if a block is currently memory resident.
3596  */
3597 struct buf *
3598 incore(struct bufobj *bo, daddr_t blkno)
3599 {
3600 	struct buf *bp;
3601 
3602 	BO_RLOCK(bo);
3603 	bp = gbincore(bo, blkno);
3604 	BO_RUNLOCK(bo);
3605 	return (bp);
3606 }
3607 
3608 /*
3609  * Returns true if no I/O is needed to access the
3610  * associated VM object.  This is like incore except
3611  * it also hunts around in the VM system for the data.
3612  */
3613 
3614 static int
3615 inmem(struct vnode * vp, daddr_t blkno)
3616 {
3617 	vm_object_t obj;
3618 	vm_offset_t toff, tinc, size;
3619 	vm_page_t m;
3620 	vm_ooffset_t off;
3621 
3622 	ASSERT_VOP_LOCKED(vp, "inmem");
3623 
3624 	if (incore(&vp->v_bufobj, blkno))
3625 		return 1;
3626 	if (vp->v_mount == NULL)
3627 		return 0;
3628 	obj = vp->v_object;
3629 	if (obj == NULL)
3630 		return (0);
3631 
3632 	size = PAGE_SIZE;
3633 	if (size > vp->v_mount->mnt_stat.f_iosize)
3634 		size = vp->v_mount->mnt_stat.f_iosize;
3635 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3636 
3637 	VM_OBJECT_RLOCK(obj);
3638 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3639 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3640 		if (!m)
3641 			goto notinmem;
3642 		tinc = size;
3643 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3644 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3645 		if (vm_page_is_valid(m,
3646 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3647 			goto notinmem;
3648 	}
3649 	VM_OBJECT_RUNLOCK(obj);
3650 	return 1;
3651 
3652 notinmem:
3653 	VM_OBJECT_RUNLOCK(obj);
3654 	return (0);
3655 }
3656 
3657 /*
3658  * Set the dirty range for a buffer based on the status of the dirty
3659  * bits in the pages comprising the buffer.  The range is limited
3660  * to the size of the buffer.
3661  *
3662  * Tell the VM system that the pages associated with this buffer
3663  * are clean.  This is used for delayed writes where the data is
3664  * going to go to disk eventually without additional VM intevention.
3665  *
3666  * Note that while we only really need to clean through to b_bcount, we
3667  * just go ahead and clean through to b_bufsize.
3668  */
3669 static void
3670 vfs_clean_pages_dirty_buf(struct buf *bp)
3671 {
3672 	vm_ooffset_t foff, noff, eoff;
3673 	vm_page_t m;
3674 	int i;
3675 
3676 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3677 		return;
3678 
3679 	foff = bp->b_offset;
3680 	KASSERT(bp->b_offset != NOOFFSET,
3681 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3682 
3683 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3684 	vfs_drain_busy_pages(bp);
3685 	vfs_setdirty_locked_object(bp);
3686 	for (i = 0; i < bp->b_npages; i++) {
3687 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3688 		eoff = noff;
3689 		if (eoff > bp->b_offset + bp->b_bufsize)
3690 			eoff = bp->b_offset + bp->b_bufsize;
3691 		m = bp->b_pages[i];
3692 		vfs_page_set_validclean(bp, foff, m);
3693 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3694 		foff = noff;
3695 	}
3696 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3697 }
3698 
3699 static void
3700 vfs_setdirty_locked_object(struct buf *bp)
3701 {
3702 	vm_object_t object;
3703 	int i;
3704 
3705 	object = bp->b_bufobj->bo_object;
3706 	VM_OBJECT_ASSERT_WLOCKED(object);
3707 
3708 	/*
3709 	 * We qualify the scan for modified pages on whether the
3710 	 * object has been flushed yet.
3711 	 */
3712 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3713 		vm_offset_t boffset;
3714 		vm_offset_t eoffset;
3715 
3716 		/*
3717 		 * test the pages to see if they have been modified directly
3718 		 * by users through the VM system.
3719 		 */
3720 		for (i = 0; i < bp->b_npages; i++)
3721 			vm_page_test_dirty(bp->b_pages[i]);
3722 
3723 		/*
3724 		 * Calculate the encompassing dirty range, boffset and eoffset,
3725 		 * (eoffset - boffset) bytes.
3726 		 */
3727 
3728 		for (i = 0; i < bp->b_npages; i++) {
3729 			if (bp->b_pages[i]->dirty)
3730 				break;
3731 		}
3732 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3733 
3734 		for (i = bp->b_npages - 1; i >= 0; --i) {
3735 			if (bp->b_pages[i]->dirty) {
3736 				break;
3737 			}
3738 		}
3739 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3740 
3741 		/*
3742 		 * Fit it to the buffer.
3743 		 */
3744 
3745 		if (eoffset > bp->b_bcount)
3746 			eoffset = bp->b_bcount;
3747 
3748 		/*
3749 		 * If we have a good dirty range, merge with the existing
3750 		 * dirty range.
3751 		 */
3752 
3753 		if (boffset < eoffset) {
3754 			if (bp->b_dirtyoff > boffset)
3755 				bp->b_dirtyoff = boffset;
3756 			if (bp->b_dirtyend < eoffset)
3757 				bp->b_dirtyend = eoffset;
3758 		}
3759 	}
3760 }
3761 
3762 /*
3763  * Allocate the KVA mapping for an existing buffer.
3764  * If an unmapped buffer is provided but a mapped buffer is requested, take
3765  * also care to properly setup mappings between pages and KVA.
3766  */
3767 static void
3768 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3769 {
3770 	int bsize, maxsize, need_mapping, need_kva;
3771 	off_t offset;
3772 
3773 	need_mapping = bp->b_data == unmapped_buf &&
3774 	    (gbflags & GB_UNMAPPED) == 0;
3775 	need_kva = bp->b_kvabase == unmapped_buf &&
3776 	    bp->b_data == unmapped_buf &&
3777 	    (gbflags & GB_KVAALLOC) != 0;
3778 	if (!need_mapping && !need_kva)
3779 		return;
3780 
3781 	BUF_CHECK_UNMAPPED(bp);
3782 
3783 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3784 		/*
3785 		 * Buffer is not mapped, but the KVA was already
3786 		 * reserved at the time of the instantiation.  Use the
3787 		 * allocated space.
3788 		 */
3789 		goto has_addr;
3790 	}
3791 
3792 	/*
3793 	 * Calculate the amount of the address space we would reserve
3794 	 * if the buffer was mapped.
3795 	 */
3796 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3797 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3798 	offset = blkno * bsize;
3799 	maxsize = size + (offset & PAGE_MASK);
3800 	maxsize = imax(maxsize, bsize);
3801 
3802 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3803 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3804 			/*
3805 			 * XXXKIB: defragmentation cannot
3806 			 * succeed, not sure what else to do.
3807 			 */
3808 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3809 		}
3810 		counter_u64_add(mappingrestarts, 1);
3811 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3812 	}
3813 has_addr:
3814 	if (need_mapping) {
3815 		/* b_offset is handled by bpmap_qenter. */
3816 		bp->b_data = bp->b_kvabase;
3817 		BUF_CHECK_MAPPED(bp);
3818 		bpmap_qenter(bp);
3819 	}
3820 }
3821 
3822 /*
3823  *	getblk:
3824  *
3825  *	Get a block given a specified block and offset into a file/device.
3826  *	The buffers B_DONE bit will be cleared on return, making it almost
3827  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3828  *	return.  The caller should clear B_INVAL prior to initiating a
3829  *	READ.
3830  *
3831  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3832  *	an existing buffer.
3833  *
3834  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3835  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3836  *	and then cleared based on the backing VM.  If the previous buffer is
3837  *	non-0-sized but invalid, B_CACHE will be cleared.
3838  *
3839  *	If getblk() must create a new buffer, the new buffer is returned with
3840  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3841  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3842  *	backing VM.
3843  *
3844  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3845  *	B_CACHE bit is clear.
3846  *
3847  *	What this means, basically, is that the caller should use B_CACHE to
3848  *	determine whether the buffer is fully valid or not and should clear
3849  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3850  *	the buffer by loading its data area with something, the caller needs
3851  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3852  *	the caller should set B_CACHE ( as an optimization ), else the caller
3853  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3854  *	a write attempt or if it was a successful read.  If the caller
3855  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3856  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3857  */
3858 struct buf *
3859 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3860     int flags)
3861 {
3862 	struct buf *bp;
3863 	struct bufobj *bo;
3864 	int bsize, error, maxsize, vmio;
3865 	off_t offset;
3866 
3867 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3868 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3869 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3870 	ASSERT_VOP_LOCKED(vp, "getblk");
3871 	if (size > maxbcachebuf)
3872 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3873 		    maxbcachebuf);
3874 	if (!unmapped_buf_allowed)
3875 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3876 
3877 	bo = &vp->v_bufobj;
3878 loop:
3879 	BO_RLOCK(bo);
3880 	bp = gbincore(bo, blkno);
3881 	if (bp != NULL) {
3882 		int lockflags;
3883 		/*
3884 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3885 		 * it must be on a queue.
3886 		 */
3887 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3888 
3889 		if (flags & GB_LOCK_NOWAIT)
3890 			lockflags |= LK_NOWAIT;
3891 
3892 		error = BUF_TIMELOCK(bp, lockflags,
3893 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3894 
3895 		/*
3896 		 * If we slept and got the lock we have to restart in case
3897 		 * the buffer changed identities.
3898 		 */
3899 		if (error == ENOLCK)
3900 			goto loop;
3901 		/* We timed out or were interrupted. */
3902 		else if (error)
3903 			return (NULL);
3904 		/* If recursed, assume caller knows the rules. */
3905 		else if (BUF_LOCKRECURSED(bp))
3906 			goto end;
3907 
3908 		/*
3909 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3910 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3911 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3912 		 * backing VM cache.
3913 		 */
3914 		if (bp->b_flags & B_INVAL)
3915 			bp->b_flags &= ~B_CACHE;
3916 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3917 			bp->b_flags |= B_CACHE;
3918 		if (bp->b_flags & B_MANAGED)
3919 			MPASS(bp->b_qindex == QUEUE_NONE);
3920 		else
3921 			bremfree(bp);
3922 
3923 		/*
3924 		 * check for size inconsistencies for non-VMIO case.
3925 		 */
3926 		if (bp->b_bcount != size) {
3927 			if ((bp->b_flags & B_VMIO) == 0 ||
3928 			    (size > bp->b_kvasize)) {
3929 				if (bp->b_flags & B_DELWRI) {
3930 					bp->b_flags |= B_NOCACHE;
3931 					bwrite(bp);
3932 				} else {
3933 					if (LIST_EMPTY(&bp->b_dep)) {
3934 						bp->b_flags |= B_RELBUF;
3935 						brelse(bp);
3936 					} else {
3937 						bp->b_flags |= B_NOCACHE;
3938 						bwrite(bp);
3939 					}
3940 				}
3941 				goto loop;
3942 			}
3943 		}
3944 
3945 		/*
3946 		 * Handle the case of unmapped buffer which should
3947 		 * become mapped, or the buffer for which KVA
3948 		 * reservation is requested.
3949 		 */
3950 		bp_unmapped_get_kva(bp, blkno, size, flags);
3951 
3952 		/*
3953 		 * If the size is inconsistent in the VMIO case, we can resize
3954 		 * the buffer.  This might lead to B_CACHE getting set or
3955 		 * cleared.  If the size has not changed, B_CACHE remains
3956 		 * unchanged from its previous state.
3957 		 */
3958 		allocbuf(bp, size);
3959 
3960 		KASSERT(bp->b_offset != NOOFFSET,
3961 		    ("getblk: no buffer offset"));
3962 
3963 		/*
3964 		 * A buffer with B_DELWRI set and B_CACHE clear must
3965 		 * be committed before we can return the buffer in
3966 		 * order to prevent the caller from issuing a read
3967 		 * ( due to B_CACHE not being set ) and overwriting
3968 		 * it.
3969 		 *
3970 		 * Most callers, including NFS and FFS, need this to
3971 		 * operate properly either because they assume they
3972 		 * can issue a read if B_CACHE is not set, or because
3973 		 * ( for example ) an uncached B_DELWRI might loop due
3974 		 * to softupdates re-dirtying the buffer.  In the latter
3975 		 * case, B_CACHE is set after the first write completes,
3976 		 * preventing further loops.
3977 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3978 		 * above while extending the buffer, we cannot allow the
3979 		 * buffer to remain with B_CACHE set after the write
3980 		 * completes or it will represent a corrupt state.  To
3981 		 * deal with this we set B_NOCACHE to scrap the buffer
3982 		 * after the write.
3983 		 *
3984 		 * We might be able to do something fancy, like setting
3985 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3986 		 * so the below call doesn't set B_CACHE, but that gets real
3987 		 * confusing.  This is much easier.
3988 		 */
3989 
3990 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3991 			bp->b_flags |= B_NOCACHE;
3992 			bwrite(bp);
3993 			goto loop;
3994 		}
3995 		bp->b_flags &= ~B_DONE;
3996 	} else {
3997 		/*
3998 		 * Buffer is not in-core, create new buffer.  The buffer
3999 		 * returned by getnewbuf() is locked.  Note that the returned
4000 		 * buffer is also considered valid (not marked B_INVAL).
4001 		 */
4002 		BO_RUNLOCK(bo);
4003 		/*
4004 		 * If the user does not want us to create the buffer, bail out
4005 		 * here.
4006 		 */
4007 		if (flags & GB_NOCREAT)
4008 			return NULL;
4009 		if (bdomain[bo->bo_domain].bd_freebuffers == 0 &&
4010 		    TD_IS_IDLETHREAD(curthread))
4011 			return NULL;
4012 
4013 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
4014 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4015 		offset = blkno * bsize;
4016 		vmio = vp->v_object != NULL;
4017 		if (vmio) {
4018 			maxsize = size + (offset & PAGE_MASK);
4019 		} else {
4020 			maxsize = size;
4021 			/* Do not allow non-VMIO notmapped buffers. */
4022 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4023 		}
4024 		maxsize = imax(maxsize, bsize);
4025 
4026 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4027 		if (bp == NULL) {
4028 			if (slpflag || slptimeo)
4029 				return NULL;
4030 			/*
4031 			 * XXX This is here until the sleep path is diagnosed
4032 			 * enough to work under very low memory conditions.
4033 			 *
4034 			 * There's an issue on low memory, 4BSD+non-preempt
4035 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4036 			 * exhaustion occurs without sleeping for buffer
4037 			 * reclaimation.  This just sticks in a loop and
4038 			 * constantly attempts to allocate a buffer, which
4039 			 * hits exhaustion and tries to wakeup bufdaemon.
4040 			 * This never happens because we never yield.
4041 			 *
4042 			 * The real solution is to identify and fix these cases
4043 			 * so we aren't effectively busy-waiting in a loop
4044 			 * until the reclaimation path has cycles to run.
4045 			 */
4046 			kern_yield(PRI_USER);
4047 			goto loop;
4048 		}
4049 
4050 		/*
4051 		 * This code is used to make sure that a buffer is not
4052 		 * created while the getnewbuf routine is blocked.
4053 		 * This can be a problem whether the vnode is locked or not.
4054 		 * If the buffer is created out from under us, we have to
4055 		 * throw away the one we just created.
4056 		 *
4057 		 * Note: this must occur before we associate the buffer
4058 		 * with the vp especially considering limitations in
4059 		 * the splay tree implementation when dealing with duplicate
4060 		 * lblkno's.
4061 		 */
4062 		BO_LOCK(bo);
4063 		if (gbincore(bo, blkno)) {
4064 			BO_UNLOCK(bo);
4065 			bp->b_flags |= B_INVAL;
4066 			bufspace_release(bufdomain(bp), maxsize);
4067 			brelse(bp);
4068 			goto loop;
4069 		}
4070 
4071 		/*
4072 		 * Insert the buffer into the hash, so that it can
4073 		 * be found by incore.
4074 		 */
4075 		bp->b_blkno = bp->b_lblkno = blkno;
4076 		bp->b_offset = offset;
4077 		bgetvp(vp, bp);
4078 		BO_UNLOCK(bo);
4079 
4080 		/*
4081 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4082 		 * buffer size starts out as 0, B_CACHE will be set by
4083 		 * allocbuf() for the VMIO case prior to it testing the
4084 		 * backing store for validity.
4085 		 */
4086 
4087 		if (vmio) {
4088 			bp->b_flags |= B_VMIO;
4089 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4090 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4091 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4092 		} else {
4093 			bp->b_flags &= ~B_VMIO;
4094 			KASSERT(bp->b_bufobj->bo_object == NULL,
4095 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4096 			    bp, bp->b_bufobj->bo_object));
4097 			BUF_CHECK_MAPPED(bp);
4098 		}
4099 
4100 		allocbuf(bp, size);
4101 		bufspace_release(bufdomain(bp), maxsize);
4102 		bp->b_flags &= ~B_DONE;
4103 	}
4104 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4105 	BUF_ASSERT_HELD(bp);
4106 end:
4107 	buf_track(bp, __func__);
4108 	KASSERT(bp->b_bufobj == bo,
4109 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4110 	return (bp);
4111 }
4112 
4113 /*
4114  * Get an empty, disassociated buffer of given size.  The buffer is initially
4115  * set to B_INVAL.
4116  */
4117 struct buf *
4118 geteblk(int size, int flags)
4119 {
4120 	struct buf *bp;
4121 	int maxsize;
4122 
4123 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4124 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4125 		if ((flags & GB_NOWAIT_BD) &&
4126 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4127 			return (NULL);
4128 	}
4129 	allocbuf(bp, size);
4130 	bufspace_release(bufdomain(bp), maxsize);
4131 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4132 	BUF_ASSERT_HELD(bp);
4133 	return (bp);
4134 }
4135 
4136 /*
4137  * Truncate the backing store for a non-vmio buffer.
4138  */
4139 static void
4140 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4141 {
4142 
4143 	if (bp->b_flags & B_MALLOC) {
4144 		/*
4145 		 * malloced buffers are not shrunk
4146 		 */
4147 		if (newbsize == 0) {
4148 			bufmallocadjust(bp, 0);
4149 			free(bp->b_data, M_BIOBUF);
4150 			bp->b_data = bp->b_kvabase;
4151 			bp->b_flags &= ~B_MALLOC;
4152 		}
4153 		return;
4154 	}
4155 	vm_hold_free_pages(bp, newbsize);
4156 	bufspace_adjust(bp, newbsize);
4157 }
4158 
4159 /*
4160  * Extend the backing for a non-VMIO buffer.
4161  */
4162 static void
4163 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4164 {
4165 	caddr_t origbuf;
4166 	int origbufsize;
4167 
4168 	/*
4169 	 * We only use malloced memory on the first allocation.
4170 	 * and revert to page-allocated memory when the buffer
4171 	 * grows.
4172 	 *
4173 	 * There is a potential smp race here that could lead
4174 	 * to bufmallocspace slightly passing the max.  It
4175 	 * is probably extremely rare and not worth worrying
4176 	 * over.
4177 	 */
4178 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4179 	    bufmallocspace < maxbufmallocspace) {
4180 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4181 		bp->b_flags |= B_MALLOC;
4182 		bufmallocadjust(bp, newbsize);
4183 		return;
4184 	}
4185 
4186 	/*
4187 	 * If the buffer is growing on its other-than-first
4188 	 * allocation then we revert to the page-allocation
4189 	 * scheme.
4190 	 */
4191 	origbuf = NULL;
4192 	origbufsize = 0;
4193 	if (bp->b_flags & B_MALLOC) {
4194 		origbuf = bp->b_data;
4195 		origbufsize = bp->b_bufsize;
4196 		bp->b_data = bp->b_kvabase;
4197 		bufmallocadjust(bp, 0);
4198 		bp->b_flags &= ~B_MALLOC;
4199 		newbsize = round_page(newbsize);
4200 	}
4201 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4202 	    (vm_offset_t) bp->b_data + newbsize);
4203 	if (origbuf != NULL) {
4204 		bcopy(origbuf, bp->b_data, origbufsize);
4205 		free(origbuf, M_BIOBUF);
4206 	}
4207 	bufspace_adjust(bp, newbsize);
4208 }
4209 
4210 /*
4211  * This code constitutes the buffer memory from either anonymous system
4212  * memory (in the case of non-VMIO operations) or from an associated
4213  * VM object (in the case of VMIO operations).  This code is able to
4214  * resize a buffer up or down.
4215  *
4216  * Note that this code is tricky, and has many complications to resolve
4217  * deadlock or inconsistent data situations.  Tread lightly!!!
4218  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4219  * the caller.  Calling this code willy nilly can result in the loss of data.
4220  *
4221  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4222  * B_CACHE for the non-VMIO case.
4223  */
4224 int
4225 allocbuf(struct buf *bp, int size)
4226 {
4227 	int newbsize;
4228 
4229 	BUF_ASSERT_HELD(bp);
4230 
4231 	if (bp->b_bcount == size)
4232 		return (1);
4233 
4234 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4235 		panic("allocbuf: buffer too small");
4236 
4237 	newbsize = roundup2(size, DEV_BSIZE);
4238 	if ((bp->b_flags & B_VMIO) == 0) {
4239 		if ((bp->b_flags & B_MALLOC) == 0)
4240 			newbsize = round_page(newbsize);
4241 		/*
4242 		 * Just get anonymous memory from the kernel.  Don't
4243 		 * mess with B_CACHE.
4244 		 */
4245 		if (newbsize < bp->b_bufsize)
4246 			vfs_nonvmio_truncate(bp, newbsize);
4247 		else if (newbsize > bp->b_bufsize)
4248 			vfs_nonvmio_extend(bp, newbsize);
4249 	} else {
4250 		int desiredpages;
4251 
4252 		desiredpages = (size == 0) ? 0 :
4253 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4254 
4255 		if (bp->b_flags & B_MALLOC)
4256 			panic("allocbuf: VMIO buffer can't be malloced");
4257 		/*
4258 		 * Set B_CACHE initially if buffer is 0 length or will become
4259 		 * 0-length.
4260 		 */
4261 		if (size == 0 || bp->b_bufsize == 0)
4262 			bp->b_flags |= B_CACHE;
4263 
4264 		if (newbsize < bp->b_bufsize)
4265 			vfs_vmio_truncate(bp, desiredpages);
4266 		/* XXX This looks as if it should be newbsize > b_bufsize */
4267 		else if (size > bp->b_bcount)
4268 			vfs_vmio_extend(bp, desiredpages, size);
4269 		bufspace_adjust(bp, newbsize);
4270 	}
4271 	bp->b_bcount = size;		/* requested buffer size. */
4272 	return (1);
4273 }
4274 
4275 extern int inflight_transient_maps;
4276 
4277 void
4278 biodone(struct bio *bp)
4279 {
4280 	struct mtx *mtxp;
4281 	void (*done)(struct bio *);
4282 	vm_offset_t start, end;
4283 
4284 	biotrack(bp, __func__);
4285 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4286 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4287 		bp->bio_flags |= BIO_UNMAPPED;
4288 		start = trunc_page((vm_offset_t)bp->bio_data);
4289 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4290 		bp->bio_data = unmapped_buf;
4291 		pmap_qremove(start, atop(end - start));
4292 		vmem_free(transient_arena, start, end - start);
4293 		atomic_add_int(&inflight_transient_maps, -1);
4294 	}
4295 	done = bp->bio_done;
4296 	if (done == NULL) {
4297 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4298 		mtx_lock(mtxp);
4299 		bp->bio_flags |= BIO_DONE;
4300 		wakeup(bp);
4301 		mtx_unlock(mtxp);
4302 	} else
4303 		done(bp);
4304 }
4305 
4306 /*
4307  * Wait for a BIO to finish.
4308  */
4309 int
4310 biowait(struct bio *bp, const char *wchan)
4311 {
4312 	struct mtx *mtxp;
4313 
4314 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4315 	mtx_lock(mtxp);
4316 	while ((bp->bio_flags & BIO_DONE) == 0)
4317 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4318 	mtx_unlock(mtxp);
4319 	if (bp->bio_error != 0)
4320 		return (bp->bio_error);
4321 	if (!(bp->bio_flags & BIO_ERROR))
4322 		return (0);
4323 	return (EIO);
4324 }
4325 
4326 void
4327 biofinish(struct bio *bp, struct devstat *stat, int error)
4328 {
4329 
4330 	if (error) {
4331 		bp->bio_error = error;
4332 		bp->bio_flags |= BIO_ERROR;
4333 	}
4334 	if (stat != NULL)
4335 		devstat_end_transaction_bio(stat, bp);
4336 	biodone(bp);
4337 }
4338 
4339 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4340 void
4341 biotrack_buf(struct bio *bp, const char *location)
4342 {
4343 
4344 	buf_track(bp->bio_track_bp, location);
4345 }
4346 #endif
4347 
4348 /*
4349  *	bufwait:
4350  *
4351  *	Wait for buffer I/O completion, returning error status.  The buffer
4352  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4353  *	error and cleared.
4354  */
4355 int
4356 bufwait(struct buf *bp)
4357 {
4358 	if (bp->b_iocmd == BIO_READ)
4359 		bwait(bp, PRIBIO, "biord");
4360 	else
4361 		bwait(bp, PRIBIO, "biowr");
4362 	if (bp->b_flags & B_EINTR) {
4363 		bp->b_flags &= ~B_EINTR;
4364 		return (EINTR);
4365 	}
4366 	if (bp->b_ioflags & BIO_ERROR) {
4367 		return (bp->b_error ? bp->b_error : EIO);
4368 	} else {
4369 		return (0);
4370 	}
4371 }
4372 
4373 /*
4374  *	bufdone:
4375  *
4376  *	Finish I/O on a buffer, optionally calling a completion function.
4377  *	This is usually called from an interrupt so process blocking is
4378  *	not allowed.
4379  *
4380  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4381  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4382  *	assuming B_INVAL is clear.
4383  *
4384  *	For the VMIO case, we set B_CACHE if the op was a read and no
4385  *	read error occurred, or if the op was a write.  B_CACHE is never
4386  *	set if the buffer is invalid or otherwise uncacheable.
4387  *
4388  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
4389  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4390  *	in the biodone routine.
4391  */
4392 void
4393 bufdone(struct buf *bp)
4394 {
4395 	struct bufobj *dropobj;
4396 	void    (*biodone)(struct buf *);
4397 
4398 	buf_track(bp, __func__);
4399 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4400 	dropobj = NULL;
4401 
4402 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4403 	BUF_ASSERT_HELD(bp);
4404 
4405 	runningbufwakeup(bp);
4406 	if (bp->b_iocmd == BIO_WRITE)
4407 		dropobj = bp->b_bufobj;
4408 	/* call optional completion function if requested */
4409 	if (bp->b_iodone != NULL) {
4410 		biodone = bp->b_iodone;
4411 		bp->b_iodone = NULL;
4412 		(*biodone) (bp);
4413 		if (dropobj)
4414 			bufobj_wdrop(dropobj);
4415 		return;
4416 	}
4417 	if (bp->b_flags & B_VMIO) {
4418 		/*
4419 		 * Set B_CACHE if the op was a normal read and no error
4420 		 * occurred.  B_CACHE is set for writes in the b*write()
4421 		 * routines.
4422 		 */
4423 		if (bp->b_iocmd == BIO_READ &&
4424 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4425 		    !(bp->b_ioflags & BIO_ERROR))
4426 			bp->b_flags |= B_CACHE;
4427 		vfs_vmio_iodone(bp);
4428 	}
4429 	if (!LIST_EMPTY(&bp->b_dep))
4430 		buf_complete(bp);
4431 	if ((bp->b_flags & B_CKHASH) != 0) {
4432 		KASSERT(bp->b_iocmd == BIO_READ,
4433 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4434 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4435 		(*bp->b_ckhashcalc)(bp);
4436 	}
4437 	/*
4438 	 * For asynchronous completions, release the buffer now. The brelse
4439 	 * will do a wakeup there if necessary - so no need to do a wakeup
4440 	 * here in the async case. The sync case always needs to do a wakeup.
4441 	 */
4442 	if (bp->b_flags & B_ASYNC) {
4443 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4444 		    (bp->b_ioflags & BIO_ERROR))
4445 			brelse(bp);
4446 		else
4447 			bqrelse(bp);
4448 	} else
4449 		bdone(bp);
4450 	if (dropobj)
4451 		bufobj_wdrop(dropobj);
4452 }
4453 
4454 /*
4455  * This routine is called in lieu of iodone in the case of
4456  * incomplete I/O.  This keeps the busy status for pages
4457  * consistent.
4458  */
4459 void
4460 vfs_unbusy_pages(struct buf *bp)
4461 {
4462 	int i;
4463 	vm_object_t obj;
4464 	vm_page_t m;
4465 
4466 	runningbufwakeup(bp);
4467 	if (!(bp->b_flags & B_VMIO))
4468 		return;
4469 
4470 	obj = bp->b_bufobj->bo_object;
4471 	VM_OBJECT_WLOCK(obj);
4472 	for (i = 0; i < bp->b_npages; i++) {
4473 		m = bp->b_pages[i];
4474 		if (m == bogus_page) {
4475 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4476 			if (!m)
4477 				panic("vfs_unbusy_pages: page missing\n");
4478 			bp->b_pages[i] = m;
4479 			if (buf_mapped(bp)) {
4480 				BUF_CHECK_MAPPED(bp);
4481 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4482 				    bp->b_pages, bp->b_npages);
4483 			} else
4484 				BUF_CHECK_UNMAPPED(bp);
4485 		}
4486 		vm_page_sunbusy(m);
4487 	}
4488 	vm_object_pip_wakeupn(obj, bp->b_npages);
4489 	VM_OBJECT_WUNLOCK(obj);
4490 }
4491 
4492 /*
4493  * vfs_page_set_valid:
4494  *
4495  *	Set the valid bits in a page based on the supplied offset.   The
4496  *	range is restricted to the buffer's size.
4497  *
4498  *	This routine is typically called after a read completes.
4499  */
4500 static void
4501 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4502 {
4503 	vm_ooffset_t eoff;
4504 
4505 	/*
4506 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4507 	 * page boundary and eoff is not greater than the end of the buffer.
4508 	 * The end of the buffer, in this case, is our file EOF, not the
4509 	 * allocation size of the buffer.
4510 	 */
4511 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4512 	if (eoff > bp->b_offset + bp->b_bcount)
4513 		eoff = bp->b_offset + bp->b_bcount;
4514 
4515 	/*
4516 	 * Set valid range.  This is typically the entire buffer and thus the
4517 	 * entire page.
4518 	 */
4519 	if (eoff > off)
4520 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4521 }
4522 
4523 /*
4524  * vfs_page_set_validclean:
4525  *
4526  *	Set the valid bits and clear the dirty bits in a page based on the
4527  *	supplied offset.   The range is restricted to the buffer's size.
4528  */
4529 static void
4530 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4531 {
4532 	vm_ooffset_t soff, eoff;
4533 
4534 	/*
4535 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4536 	 * page boundary or cross the end of the buffer.  The end of the
4537 	 * buffer, in this case, is our file EOF, not the allocation size
4538 	 * of the buffer.
4539 	 */
4540 	soff = off;
4541 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4542 	if (eoff > bp->b_offset + bp->b_bcount)
4543 		eoff = bp->b_offset + bp->b_bcount;
4544 
4545 	/*
4546 	 * Set valid range.  This is typically the entire buffer and thus the
4547 	 * entire page.
4548 	 */
4549 	if (eoff > soff) {
4550 		vm_page_set_validclean(
4551 		    m,
4552 		   (vm_offset_t) (soff & PAGE_MASK),
4553 		   (vm_offset_t) (eoff - soff)
4554 		);
4555 	}
4556 }
4557 
4558 /*
4559  * Ensure that all buffer pages are not exclusive busied.  If any page is
4560  * exclusive busy, drain it.
4561  */
4562 void
4563 vfs_drain_busy_pages(struct buf *bp)
4564 {
4565 	vm_page_t m;
4566 	int i, last_busied;
4567 
4568 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4569 	last_busied = 0;
4570 	for (i = 0; i < bp->b_npages; i++) {
4571 		m = bp->b_pages[i];
4572 		if (vm_page_xbusied(m)) {
4573 			for (; last_busied < i; last_busied++)
4574 				vm_page_sbusy(bp->b_pages[last_busied]);
4575 			while (vm_page_xbusied(m)) {
4576 				vm_page_lock(m);
4577 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4578 				vm_page_busy_sleep(m, "vbpage", true);
4579 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4580 			}
4581 		}
4582 	}
4583 	for (i = 0; i < last_busied; i++)
4584 		vm_page_sunbusy(bp->b_pages[i]);
4585 }
4586 
4587 /*
4588  * This routine is called before a device strategy routine.
4589  * It is used to tell the VM system that paging I/O is in
4590  * progress, and treat the pages associated with the buffer
4591  * almost as being exclusive busy.  Also the object paging_in_progress
4592  * flag is handled to make sure that the object doesn't become
4593  * inconsistent.
4594  *
4595  * Since I/O has not been initiated yet, certain buffer flags
4596  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4597  * and should be ignored.
4598  */
4599 void
4600 vfs_busy_pages(struct buf *bp, int clear_modify)
4601 {
4602 	vm_object_t obj;
4603 	vm_ooffset_t foff;
4604 	vm_page_t m;
4605 	int i;
4606 	bool bogus;
4607 
4608 	if (!(bp->b_flags & B_VMIO))
4609 		return;
4610 
4611 	obj = bp->b_bufobj->bo_object;
4612 	foff = bp->b_offset;
4613 	KASSERT(bp->b_offset != NOOFFSET,
4614 	    ("vfs_busy_pages: no buffer offset"));
4615 	VM_OBJECT_WLOCK(obj);
4616 	vfs_drain_busy_pages(bp);
4617 	if (bp->b_bufsize != 0)
4618 		vfs_setdirty_locked_object(bp);
4619 	bogus = false;
4620 	for (i = 0; i < bp->b_npages; i++) {
4621 		m = bp->b_pages[i];
4622 
4623 		if ((bp->b_flags & B_CLUSTER) == 0) {
4624 			vm_object_pip_add(obj, 1);
4625 			vm_page_sbusy(m);
4626 		}
4627 		/*
4628 		 * When readying a buffer for a read ( i.e
4629 		 * clear_modify == 0 ), it is important to do
4630 		 * bogus_page replacement for valid pages in
4631 		 * partially instantiated buffers.  Partially
4632 		 * instantiated buffers can, in turn, occur when
4633 		 * reconstituting a buffer from its VM backing store
4634 		 * base.  We only have to do this if B_CACHE is
4635 		 * clear ( which causes the I/O to occur in the
4636 		 * first place ).  The replacement prevents the read
4637 		 * I/O from overwriting potentially dirty VM-backed
4638 		 * pages.  XXX bogus page replacement is, uh, bogus.
4639 		 * It may not work properly with small-block devices.
4640 		 * We need to find a better way.
4641 		 */
4642 		if (clear_modify) {
4643 			pmap_remove_write(m);
4644 			vfs_page_set_validclean(bp, foff, m);
4645 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4646 		    (bp->b_flags & B_CACHE) == 0) {
4647 			bp->b_pages[i] = bogus_page;
4648 			bogus = true;
4649 		}
4650 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4651 	}
4652 	VM_OBJECT_WUNLOCK(obj);
4653 	if (bogus && buf_mapped(bp)) {
4654 		BUF_CHECK_MAPPED(bp);
4655 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4656 		    bp->b_pages, bp->b_npages);
4657 	}
4658 }
4659 
4660 /*
4661  *	vfs_bio_set_valid:
4662  *
4663  *	Set the range within the buffer to valid.  The range is
4664  *	relative to the beginning of the buffer, b_offset.  Note that
4665  *	b_offset itself may be offset from the beginning of the first
4666  *	page.
4667  */
4668 void
4669 vfs_bio_set_valid(struct buf *bp, int base, int size)
4670 {
4671 	int i, n;
4672 	vm_page_t m;
4673 
4674 	if (!(bp->b_flags & B_VMIO))
4675 		return;
4676 
4677 	/*
4678 	 * Fixup base to be relative to beginning of first page.
4679 	 * Set initial n to be the maximum number of bytes in the
4680 	 * first page that can be validated.
4681 	 */
4682 	base += (bp->b_offset & PAGE_MASK);
4683 	n = PAGE_SIZE - (base & PAGE_MASK);
4684 
4685 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4686 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4687 		m = bp->b_pages[i];
4688 		if (n > size)
4689 			n = size;
4690 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4691 		base += n;
4692 		size -= n;
4693 		n = PAGE_SIZE;
4694 	}
4695 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4696 }
4697 
4698 /*
4699  *	vfs_bio_clrbuf:
4700  *
4701  *	If the specified buffer is a non-VMIO buffer, clear the entire
4702  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4703  *	validate only the previously invalid portions of the buffer.
4704  *	This routine essentially fakes an I/O, so we need to clear
4705  *	BIO_ERROR and B_INVAL.
4706  *
4707  *	Note that while we only theoretically need to clear through b_bcount,
4708  *	we go ahead and clear through b_bufsize.
4709  */
4710 void
4711 vfs_bio_clrbuf(struct buf *bp)
4712 {
4713 	int i, j, mask, sa, ea, slide;
4714 
4715 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4716 		clrbuf(bp);
4717 		return;
4718 	}
4719 	bp->b_flags &= ~B_INVAL;
4720 	bp->b_ioflags &= ~BIO_ERROR;
4721 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4722 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4723 	    (bp->b_offset & PAGE_MASK) == 0) {
4724 		if (bp->b_pages[0] == bogus_page)
4725 			goto unlock;
4726 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4727 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4728 		if ((bp->b_pages[0]->valid & mask) == mask)
4729 			goto unlock;
4730 		if ((bp->b_pages[0]->valid & mask) == 0) {
4731 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4732 			bp->b_pages[0]->valid |= mask;
4733 			goto unlock;
4734 		}
4735 	}
4736 	sa = bp->b_offset & PAGE_MASK;
4737 	slide = 0;
4738 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4739 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4740 		ea = slide & PAGE_MASK;
4741 		if (ea == 0)
4742 			ea = PAGE_SIZE;
4743 		if (bp->b_pages[i] == bogus_page)
4744 			continue;
4745 		j = sa / DEV_BSIZE;
4746 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4747 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4748 		if ((bp->b_pages[i]->valid & mask) == mask)
4749 			continue;
4750 		if ((bp->b_pages[i]->valid & mask) == 0)
4751 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4752 		else {
4753 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4754 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4755 					pmap_zero_page_area(bp->b_pages[i],
4756 					    sa, DEV_BSIZE);
4757 				}
4758 			}
4759 		}
4760 		bp->b_pages[i]->valid |= mask;
4761 	}
4762 unlock:
4763 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4764 	bp->b_resid = 0;
4765 }
4766 
4767 void
4768 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4769 {
4770 	vm_page_t m;
4771 	int i, n;
4772 
4773 	if (buf_mapped(bp)) {
4774 		BUF_CHECK_MAPPED(bp);
4775 		bzero(bp->b_data + base, size);
4776 	} else {
4777 		BUF_CHECK_UNMAPPED(bp);
4778 		n = PAGE_SIZE - (base & PAGE_MASK);
4779 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4780 			m = bp->b_pages[i];
4781 			if (n > size)
4782 				n = size;
4783 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4784 			base += n;
4785 			size -= n;
4786 			n = PAGE_SIZE;
4787 		}
4788 	}
4789 }
4790 
4791 /*
4792  * Update buffer flags based on I/O request parameters, optionally releasing the
4793  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4794  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4795  * I/O).  Otherwise the buffer is released to the cache.
4796  */
4797 static void
4798 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4799 {
4800 
4801 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4802 	    ("buf %p non-VMIO noreuse", bp));
4803 
4804 	if ((ioflag & IO_DIRECT) != 0)
4805 		bp->b_flags |= B_DIRECT;
4806 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4807 		bp->b_flags |= B_RELBUF;
4808 		if ((ioflag & IO_NOREUSE) != 0)
4809 			bp->b_flags |= B_NOREUSE;
4810 		if (release)
4811 			brelse(bp);
4812 	} else if (release)
4813 		bqrelse(bp);
4814 }
4815 
4816 void
4817 vfs_bio_brelse(struct buf *bp, int ioflag)
4818 {
4819 
4820 	b_io_dismiss(bp, ioflag, true);
4821 }
4822 
4823 void
4824 vfs_bio_set_flags(struct buf *bp, int ioflag)
4825 {
4826 
4827 	b_io_dismiss(bp, ioflag, false);
4828 }
4829 
4830 /*
4831  * vm_hold_load_pages and vm_hold_free_pages get pages into
4832  * a buffers address space.  The pages are anonymous and are
4833  * not associated with a file object.
4834  */
4835 static void
4836 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4837 {
4838 	vm_offset_t pg;
4839 	vm_page_t p;
4840 	int index;
4841 
4842 	BUF_CHECK_MAPPED(bp);
4843 
4844 	to = round_page(to);
4845 	from = round_page(from);
4846 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4847 
4848 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4849 		/*
4850 		 * note: must allocate system pages since blocking here
4851 		 * could interfere with paging I/O, no matter which
4852 		 * process we are.
4853 		 */
4854 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4855 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4856 		    VM_ALLOC_WAITOK);
4857 		pmap_qenter(pg, &p, 1);
4858 		bp->b_pages[index] = p;
4859 	}
4860 	bp->b_npages = index;
4861 }
4862 
4863 /* Return pages associated with this buf to the vm system */
4864 static void
4865 vm_hold_free_pages(struct buf *bp, int newbsize)
4866 {
4867 	vm_offset_t from;
4868 	vm_page_t p;
4869 	int index, newnpages;
4870 
4871 	BUF_CHECK_MAPPED(bp);
4872 
4873 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4874 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4875 	if (bp->b_npages > newnpages)
4876 		pmap_qremove(from, bp->b_npages - newnpages);
4877 	for (index = newnpages; index < bp->b_npages; index++) {
4878 		p = bp->b_pages[index];
4879 		bp->b_pages[index] = NULL;
4880 		p->wire_count--;
4881 		vm_page_free(p);
4882 	}
4883 	vm_wire_sub(bp->b_npages - newnpages);
4884 	bp->b_npages = newnpages;
4885 }
4886 
4887 /*
4888  * Map an IO request into kernel virtual address space.
4889  *
4890  * All requests are (re)mapped into kernel VA space.
4891  * Notice that we use b_bufsize for the size of the buffer
4892  * to be mapped.  b_bcount might be modified by the driver.
4893  *
4894  * Note that even if the caller determines that the address space should
4895  * be valid, a race or a smaller-file mapped into a larger space may
4896  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4897  * check the return value.
4898  *
4899  * This function only works with pager buffers.
4900  */
4901 int
4902 vmapbuf(struct buf *bp, int mapbuf)
4903 {
4904 	vm_prot_t prot;
4905 	int pidx;
4906 
4907 	if (bp->b_bufsize < 0)
4908 		return (-1);
4909 	prot = VM_PROT_READ;
4910 	if (bp->b_iocmd == BIO_READ)
4911 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4912 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4913 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4914 	    btoc(MAXPHYS))) < 0)
4915 		return (-1);
4916 	bp->b_npages = pidx;
4917 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4918 	if (mapbuf || !unmapped_buf_allowed) {
4919 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4920 		bp->b_data = bp->b_kvabase + bp->b_offset;
4921 	} else
4922 		bp->b_data = unmapped_buf;
4923 	return(0);
4924 }
4925 
4926 /*
4927  * Free the io map PTEs associated with this IO operation.
4928  * We also invalidate the TLB entries and restore the original b_addr.
4929  *
4930  * This function only works with pager buffers.
4931  */
4932 void
4933 vunmapbuf(struct buf *bp)
4934 {
4935 	int npages;
4936 
4937 	npages = bp->b_npages;
4938 	if (buf_mapped(bp))
4939 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4940 	vm_page_unhold_pages(bp->b_pages, npages);
4941 
4942 	bp->b_data = unmapped_buf;
4943 }
4944 
4945 void
4946 bdone(struct buf *bp)
4947 {
4948 	struct mtx *mtxp;
4949 
4950 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4951 	mtx_lock(mtxp);
4952 	bp->b_flags |= B_DONE;
4953 	wakeup(bp);
4954 	mtx_unlock(mtxp);
4955 }
4956 
4957 void
4958 bwait(struct buf *bp, u_char pri, const char *wchan)
4959 {
4960 	struct mtx *mtxp;
4961 
4962 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4963 	mtx_lock(mtxp);
4964 	while ((bp->b_flags & B_DONE) == 0)
4965 		msleep(bp, mtxp, pri, wchan, 0);
4966 	mtx_unlock(mtxp);
4967 }
4968 
4969 int
4970 bufsync(struct bufobj *bo, int waitfor)
4971 {
4972 
4973 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
4974 }
4975 
4976 void
4977 bufstrategy(struct bufobj *bo, struct buf *bp)
4978 {
4979 	int i = 0;
4980 	struct vnode *vp;
4981 
4982 	vp = bp->b_vp;
4983 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4984 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4985 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4986 	i = VOP_STRATEGY(vp, bp);
4987 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4988 }
4989 
4990 /*
4991  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
4992  */
4993 void
4994 bufobj_init(struct bufobj *bo, void *private)
4995 {
4996 	static volatile int bufobj_cleanq;
4997 
4998         bo->bo_domain =
4999             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5000         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5001         bo->bo_private = private;
5002         TAILQ_INIT(&bo->bo_clean.bv_hd);
5003         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5004 }
5005 
5006 void
5007 bufobj_wrefl(struct bufobj *bo)
5008 {
5009 
5010 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5011 	ASSERT_BO_WLOCKED(bo);
5012 	bo->bo_numoutput++;
5013 }
5014 
5015 void
5016 bufobj_wref(struct bufobj *bo)
5017 {
5018 
5019 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5020 	BO_LOCK(bo);
5021 	bo->bo_numoutput++;
5022 	BO_UNLOCK(bo);
5023 }
5024 
5025 void
5026 bufobj_wdrop(struct bufobj *bo)
5027 {
5028 
5029 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5030 	BO_LOCK(bo);
5031 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5032 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5033 		bo->bo_flag &= ~BO_WWAIT;
5034 		wakeup(&bo->bo_numoutput);
5035 	}
5036 	BO_UNLOCK(bo);
5037 }
5038 
5039 int
5040 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5041 {
5042 	int error;
5043 
5044 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5045 	ASSERT_BO_WLOCKED(bo);
5046 	error = 0;
5047 	while (bo->bo_numoutput) {
5048 		bo->bo_flag |= BO_WWAIT;
5049 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5050 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5051 		if (error)
5052 			break;
5053 	}
5054 	return (error);
5055 }
5056 
5057 /*
5058  * Set bio_data or bio_ma for struct bio from the struct buf.
5059  */
5060 void
5061 bdata2bio(struct buf *bp, struct bio *bip)
5062 {
5063 
5064 	if (!buf_mapped(bp)) {
5065 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5066 		bip->bio_ma = bp->b_pages;
5067 		bip->bio_ma_n = bp->b_npages;
5068 		bip->bio_data = unmapped_buf;
5069 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5070 		bip->bio_flags |= BIO_UNMAPPED;
5071 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5072 		    PAGE_SIZE == bp->b_npages,
5073 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5074 		    (long long)bip->bio_length, bip->bio_ma_n));
5075 	} else {
5076 		bip->bio_data = bp->b_data;
5077 		bip->bio_ma = NULL;
5078 	}
5079 }
5080 
5081 /*
5082  * The MIPS pmap code currently doesn't handle aliased pages.
5083  * The VIPT caches may not handle page aliasing themselves, leading
5084  * to data corruption.
5085  *
5086  * As such, this code makes a system extremely unhappy if said
5087  * system doesn't support unaliasing the above situation in hardware.
5088  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5089  * this feature at build time, so it has to be handled in software.
5090  *
5091  * Once the MIPS pmap/cache code grows to support this function on
5092  * earlier chips, it should be flipped back off.
5093  */
5094 #ifdef	__mips__
5095 static int buf_pager_relbuf = 1;
5096 #else
5097 static int buf_pager_relbuf = 0;
5098 #endif
5099 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5100     &buf_pager_relbuf, 0,
5101     "Make buffer pager release buffers after reading");
5102 
5103 /*
5104  * The buffer pager.  It uses buffer reads to validate pages.
5105  *
5106  * In contrast to the generic local pager from vm/vnode_pager.c, this
5107  * pager correctly and easily handles volumes where the underlying
5108  * device block size is greater than the machine page size.  The
5109  * buffer cache transparently extends the requested page run to be
5110  * aligned at the block boundary, and does the necessary bogus page
5111  * replacements in the addends to avoid obliterating already valid
5112  * pages.
5113  *
5114  * The only non-trivial issue is that the exclusive busy state for
5115  * pages, which is assumed by the vm_pager_getpages() interface, is
5116  * incompatible with the VMIO buffer cache's desire to share-busy the
5117  * pages.  This function performs a trivial downgrade of the pages'
5118  * state before reading buffers, and a less trivial upgrade from the
5119  * shared-busy to excl-busy state after the read.
5120  */
5121 int
5122 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5123     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5124     vbg_get_blksize_t get_blksize)
5125 {
5126 	vm_page_t m;
5127 	vm_object_t object;
5128 	struct buf *bp;
5129 	struct mount *mp;
5130 	daddr_t lbn, lbnp;
5131 	vm_ooffset_t la, lb, poff, poffe;
5132 	long bsize;
5133 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5134 	bool redo, lpart;
5135 
5136 	object = vp->v_object;
5137 	mp = vp->v_mount;
5138 	error = 0;
5139 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5140 	if (la >= object->un_pager.vnp.vnp_size)
5141 		return (VM_PAGER_BAD);
5142 
5143 	/*
5144 	 * Change the meaning of la from where the last requested page starts
5145 	 * to where it ends, because that's the end of the requested region
5146 	 * and the start of the potential read-ahead region.
5147 	 */
5148 	la += PAGE_SIZE;
5149 	lpart = la > object->un_pager.vnp.vnp_size;
5150 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5151 
5152 	/*
5153 	 * Calculate read-ahead, behind and total pages.
5154 	 */
5155 	pgsin = count;
5156 	lb = IDX_TO_OFF(ma[0]->pindex);
5157 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5158 	pgsin += pgsin_b;
5159 	if (rbehind != NULL)
5160 		*rbehind = pgsin_b;
5161 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5162 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5163 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5164 		    PAGE_SIZE) - la);
5165 	pgsin += pgsin_a;
5166 	if (rahead != NULL)
5167 		*rahead = pgsin_a;
5168 	VM_CNT_INC(v_vnodein);
5169 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5170 
5171 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5172 	    != 0) ? GB_UNMAPPED : 0;
5173 	VM_OBJECT_WLOCK(object);
5174 again:
5175 	for (i = 0; i < count; i++)
5176 		vm_page_busy_downgrade(ma[i]);
5177 	VM_OBJECT_WUNLOCK(object);
5178 
5179 	lbnp = -1;
5180 	for (i = 0; i < count; i++) {
5181 		m = ma[i];
5182 
5183 		/*
5184 		 * Pages are shared busy and the object lock is not
5185 		 * owned, which together allow for the pages'
5186 		 * invalidation.  The racy test for validity avoids
5187 		 * useless creation of the buffer for the most typical
5188 		 * case when invalidation is not used in redo or for
5189 		 * parallel read.  The shared->excl upgrade loop at
5190 		 * the end of the function catches the race in a
5191 		 * reliable way (protected by the object lock).
5192 		 */
5193 		if (m->valid == VM_PAGE_BITS_ALL)
5194 			continue;
5195 
5196 		poff = IDX_TO_OFF(m->pindex);
5197 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5198 		for (; poff < poffe; poff += bsize) {
5199 			lbn = get_lblkno(vp, poff);
5200 			if (lbn == lbnp)
5201 				goto next_page;
5202 			lbnp = lbn;
5203 
5204 			bsize = get_blksize(vp, lbn);
5205 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5206 			    br_flags, &bp);
5207 			if (error != 0)
5208 				goto end_pages;
5209 			if (LIST_EMPTY(&bp->b_dep)) {
5210 				/*
5211 				 * Invalidation clears m->valid, but
5212 				 * may leave B_CACHE flag if the
5213 				 * buffer existed at the invalidation
5214 				 * time.  In this case, recycle the
5215 				 * buffer to do real read on next
5216 				 * bread() after redo.
5217 				 *
5218 				 * Otherwise B_RELBUF is not strictly
5219 				 * necessary, enable to reduce buf
5220 				 * cache pressure.
5221 				 */
5222 				if (buf_pager_relbuf ||
5223 				    m->valid != VM_PAGE_BITS_ALL)
5224 					bp->b_flags |= B_RELBUF;
5225 
5226 				bp->b_flags &= ~B_NOCACHE;
5227 				brelse(bp);
5228 			} else {
5229 				bqrelse(bp);
5230 			}
5231 		}
5232 		KASSERT(1 /* racy, enable for debugging */ ||
5233 		    m->valid == VM_PAGE_BITS_ALL || i == count - 1,
5234 		    ("buf %d %p invalid", i, m));
5235 		if (i == count - 1 && lpart) {
5236 			VM_OBJECT_WLOCK(object);
5237 			if (m->valid != 0 &&
5238 			    m->valid != VM_PAGE_BITS_ALL)
5239 				vm_page_zero_invalid(m, TRUE);
5240 			VM_OBJECT_WUNLOCK(object);
5241 		}
5242 next_page:;
5243 	}
5244 end_pages:
5245 
5246 	VM_OBJECT_WLOCK(object);
5247 	redo = false;
5248 	for (i = 0; i < count; i++) {
5249 		vm_page_sunbusy(ma[i]);
5250 		ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
5251 
5252 		/*
5253 		 * Since the pages were only sbusy while neither the
5254 		 * buffer nor the object lock was held by us, or
5255 		 * reallocated while vm_page_grab() slept for busy
5256 		 * relinguish, they could have been invalidated.
5257 		 * Recheck the valid bits and re-read as needed.
5258 		 *
5259 		 * Note that the last page is made fully valid in the
5260 		 * read loop, and partial validity for the page at
5261 		 * index count - 1 could mean that the page was
5262 		 * invalidated or removed, so we must restart for
5263 		 * safety as well.
5264 		 */
5265 		if (ma[i]->valid != VM_PAGE_BITS_ALL)
5266 			redo = true;
5267 	}
5268 	if (redo && error == 0)
5269 		goto again;
5270 	VM_OBJECT_WUNLOCK(object);
5271 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5272 }
5273 
5274 #include "opt_ddb.h"
5275 #ifdef DDB
5276 #include <ddb/ddb.h>
5277 
5278 /* DDB command to show buffer data */
5279 DB_SHOW_COMMAND(buffer, db_show_buffer)
5280 {
5281 	/* get args */
5282 	struct buf *bp = (struct buf *)addr;
5283 #ifdef FULL_BUF_TRACKING
5284 	uint32_t i, j;
5285 #endif
5286 
5287 	if (!have_addr) {
5288 		db_printf("usage: show buffer <addr>\n");
5289 		return;
5290 	}
5291 
5292 	db_printf("buf at %p\n", bp);
5293 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
5294 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
5295 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
5296 	db_printf(
5297 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5298 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
5299 	    "b_dep = %p\n",
5300 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5301 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5302 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
5303 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5304 	    bp->b_kvabase, bp->b_kvasize);
5305 	if (bp->b_npages) {
5306 		int i;
5307 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5308 		for (i = 0; i < bp->b_npages; i++) {
5309 			vm_page_t m;
5310 			m = bp->b_pages[i];
5311 			if (m != NULL)
5312 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5313 				    (u_long)m->pindex,
5314 				    (u_long)VM_PAGE_TO_PHYS(m));
5315 			else
5316 				db_printf("( ??? )");
5317 			if ((i + 1) < bp->b_npages)
5318 				db_printf(",");
5319 		}
5320 		db_printf("\n");
5321 	}
5322 	BUF_LOCKPRINTINFO(bp);
5323 #if defined(FULL_BUF_TRACKING)
5324 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5325 
5326 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5327 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5328 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5329 			continue;
5330 		db_printf(" %2u: %s\n", j,
5331 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5332 	}
5333 #elif defined(BUF_TRACKING)
5334 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5335 #endif
5336 	db_printf(" ");
5337 }
5338 
5339 DB_SHOW_COMMAND(bufqueues, bufqueues)
5340 {
5341 	struct bufdomain *bd;
5342 	struct buf *bp;
5343 	long total;
5344 	int i, j, cnt;
5345 
5346 	db_printf("bqempty: %d\n", bqempty.bq_len);
5347 
5348 	for (i = 0; i < buf_domains; i++) {
5349 		bd = &bdomain[i];
5350 		db_printf("Buf domain %d\n", i);
5351 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5352 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5353 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5354 		db_printf("\n");
5355 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5356 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5357 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5358 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5359 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5360 		db_printf("\n");
5361 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5362 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5363 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5364 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5365 		db_printf("\n");
5366 		total = 0;
5367 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5368 			total += bp->b_bufsize;
5369 		db_printf("\tcleanq count\t%d (%ld)\n",
5370 		    bd->bd_cleanq->bq_len, total);
5371 		total = 0;
5372 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5373 			total += bp->b_bufsize;
5374 		db_printf("\tdirtyq count\t%d (%ld)\n",
5375 		    bd->bd_dirtyq.bq_len, total);
5376 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5377 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5378 		db_printf("\tCPU ");
5379 		for (j = 0; j <= mp_maxid; j++)
5380 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5381 		db_printf("\n");
5382 		cnt = 0;
5383 		total = 0;
5384 		for (j = 0; j < nbuf; j++)
5385 			if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5386 				cnt++;
5387 				total += buf[j].b_bufsize;
5388 			}
5389 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5390 		cnt = 0;
5391 		total = 0;
5392 		for (j = 0; j < nbuf; j++)
5393 			if (buf[j].b_domain == i) {
5394 				cnt++;
5395 				total += buf[j].b_bufsize;
5396 			}
5397 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5398 	}
5399 }
5400 
5401 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5402 {
5403 	struct buf *bp;
5404 	int i;
5405 
5406 	for (i = 0; i < nbuf; i++) {
5407 		bp = &buf[i];
5408 		if (BUF_ISLOCKED(bp)) {
5409 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5410 			db_printf("\n");
5411 			if (db_pager_quit)
5412 				break;
5413 		}
5414 	}
5415 }
5416 
5417 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5418 {
5419 	struct vnode *vp;
5420 	struct buf *bp;
5421 
5422 	if (!have_addr) {
5423 		db_printf("usage: show vnodebufs <addr>\n");
5424 		return;
5425 	}
5426 	vp = (struct vnode *)addr;
5427 	db_printf("Clean buffers:\n");
5428 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5429 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5430 		db_printf("\n");
5431 	}
5432 	db_printf("Dirty buffers:\n");
5433 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5434 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5435 		db_printf("\n");
5436 	}
5437 }
5438 
5439 DB_COMMAND(countfreebufs, db_coundfreebufs)
5440 {
5441 	struct buf *bp;
5442 	int i, used = 0, nfree = 0;
5443 
5444 	if (have_addr) {
5445 		db_printf("usage: countfreebufs\n");
5446 		return;
5447 	}
5448 
5449 	for (i = 0; i < nbuf; i++) {
5450 		bp = &buf[i];
5451 		if (bp->b_qindex == QUEUE_EMPTY)
5452 			nfree++;
5453 		else
5454 			used++;
5455 	}
5456 
5457 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5458 	    nfree + used);
5459 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5460 }
5461 #endif /* DDB */
5462