xref: /freebsd/sys/kern/vfs_bio.c (revision e1c167d0190dfde72fa06350ef0049181cf79a4f)
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/ktr.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/malloc.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/proc.h>
69 #include <sys/racct.h>
70 #include <sys/refcount.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/smp.h>
74 #include <sys/sysctl.h>
75 #include <sys/sysproto.h>
76 #include <sys/vmem.h>
77 #include <sys/vmmeter.h>
78 #include <sys/vnode.h>
79 #include <sys/watchdog.h>
80 #include <geom/geom.h>
81 #include <vm/vm.h>
82 #include <vm/vm_param.h>
83 #include <vm/vm_kern.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_pager.h>
88 #include <vm/vm_extern.h>
89 #include <vm/vm_map.h>
90 #include <vm/swap_pager.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_range(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 | CTLFLAG_STATS,
254     &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
255 static int recursiveflushes;
256 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
257     &recursiveflushes, 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_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
265     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
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_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
270     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
271     "When the number of dirty buffers is considered severe");
272 int dirtybufthresh;
273 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
274     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
275     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
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_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
283     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
284    "Target number of free buffers");
285 static int hifreebuffers;
286 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
287     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
288     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
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 | CTLFLAG_STATS,
313     &barrierwrites, 0, "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)(((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)(((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 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
796 	    SHUTDOWN_PRI_LAST + 100);
797 
798 	bd = arg;
799 	for (;;) {
800 		kthread_suspend_check();
801 
802 		/*
803 		 * Free buffers from the clean queue until we meet our
804 		 * targets.
805 		 *
806 		 * Theory of operation:  The buffer cache is most efficient
807 		 * when some free buffer headers and space are always
808 		 * available to getnewbuf().  This daemon attempts to prevent
809 		 * the excessive blocking and synchronization associated
810 		 * with shortfall.  It goes through three phases according
811 		 * demand:
812 		 *
813 		 * 1)	The daemon wakes up voluntarily once per-second
814 		 *	during idle periods when the counters are below
815 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
816 		 *
817 		 * 2)	The daemon wakes up as we cross the thresholds
818 		 *	ahead of any potential blocking.  This may bounce
819 		 *	slightly according to the rate of consumption and
820 		 *	release.
821 		 *
822 		 * 3)	The daemon and consumers are starved for working
823 		 *	clean buffers.  This is the 'bufspace' sleep below
824 		 *	which will inefficiently trade bufs with bqrelse
825 		 *	until we return to condition 2.
826 		 */
827 		while (bd->bd_bufspace > bd->bd_lobufspace ||
828 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
829 			if (buf_recycle(bd, false) != 0) {
830 				if (bd_flushall(bd))
831 					continue;
832 				/*
833 				 * Speedup dirty if we've run out of clean
834 				 * buffers.  This is possible in particular
835 				 * because softdep may held many bufs locked
836 				 * pending writes to other bufs which are
837 				 * marked for delayed write, exhausting
838 				 * clean space until they are written.
839 				 */
840 				bd_speedup();
841 				BD_LOCK(bd);
842 				if (bd->bd_wanted) {
843 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
844 					    PRIBIO|PDROP, "bufspace", hz/10);
845 				} else
846 					BD_UNLOCK(bd);
847 			}
848 			maybe_yield();
849 		}
850 		bufspace_daemon_wait(bd);
851 	}
852 }
853 
854 /*
855  *	bufmallocadjust:
856  *
857  *	Adjust the reported bufspace for a malloc managed buffer, possibly
858  *	waking any waiters.
859  */
860 static void
861 bufmallocadjust(struct buf *bp, int bufsize)
862 {
863 	int diff;
864 
865 	KASSERT((bp->b_flags & B_MALLOC) != 0,
866 	    ("bufmallocadjust: non-malloc buf %p", bp));
867 	diff = bufsize - bp->b_bufsize;
868 	if (diff < 0)
869 		atomic_subtract_long(&bufmallocspace, -diff);
870 	else
871 		atomic_add_long(&bufmallocspace, diff);
872 	bp->b_bufsize = bufsize;
873 }
874 
875 /*
876  *	runningwakeup:
877  *
878  *	Wake up processes that are waiting on asynchronous writes to fall
879  *	below lorunningspace.
880  */
881 static void
882 runningwakeup(void)
883 {
884 
885 	mtx_lock(&rbreqlock);
886 	if (runningbufreq) {
887 		runningbufreq = 0;
888 		wakeup(&runningbufreq);
889 	}
890 	mtx_unlock(&rbreqlock);
891 }
892 
893 /*
894  *	runningbufwakeup:
895  *
896  *	Decrement the outstanding write count according.
897  */
898 void
899 runningbufwakeup(struct buf *bp)
900 {
901 	long space, bspace;
902 
903 	bspace = bp->b_runningbufspace;
904 	if (bspace == 0)
905 		return;
906 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
907 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
908 	    space, bspace));
909 	bp->b_runningbufspace = 0;
910 	/*
911 	 * Only acquire the lock and wakeup on the transition from exceeding
912 	 * the threshold to falling below it.
913 	 */
914 	if (space < lorunningspace)
915 		return;
916 	if (space - bspace > lorunningspace)
917 		return;
918 	runningwakeup();
919 }
920 
921 /*
922  *	waitrunningbufspace()
923  *
924  *	runningbufspace is a measure of the amount of I/O currently
925  *	running.  This routine is used in async-write situations to
926  *	prevent creating huge backups of pending writes to a device.
927  *	Only asynchronous writes are governed by this function.
928  *
929  *	This does NOT turn an async write into a sync write.  It waits
930  *	for earlier writes to complete and generally returns before the
931  *	caller's write has reached the device.
932  */
933 void
934 waitrunningbufspace(void)
935 {
936 
937 	mtx_lock(&rbreqlock);
938 	while (runningbufspace > hirunningspace) {
939 		runningbufreq = 1;
940 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
941 	}
942 	mtx_unlock(&rbreqlock);
943 }
944 
945 
946 /*
947  *	vfs_buf_test_cache:
948  *
949  *	Called when a buffer is extended.  This function clears the B_CACHE
950  *	bit if the newly extended portion of the buffer does not contain
951  *	valid data.
952  */
953 static __inline void
954 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
955     vm_offset_t size, vm_page_t m)
956 {
957 
958 	/*
959 	 * This function and its results are protected by higher level
960 	 * synchronization requiring vnode and buf locks to page in and
961 	 * validate pages.
962 	 */
963 	if (bp->b_flags & B_CACHE) {
964 		int base = (foff + off) & PAGE_MASK;
965 		if (vm_page_is_valid(m, base, size) == 0)
966 			bp->b_flags &= ~B_CACHE;
967 	}
968 }
969 
970 /* Wake up the buffer daemon if necessary */
971 static void
972 bd_wakeup(void)
973 {
974 
975 	mtx_lock(&bdlock);
976 	if (bd_request == 0) {
977 		bd_request = 1;
978 		wakeup(&bd_request);
979 	}
980 	mtx_unlock(&bdlock);
981 }
982 
983 /*
984  * Adjust the maxbcachbuf tunable.
985  */
986 static void
987 maxbcachebuf_adjust(void)
988 {
989 	int i;
990 
991 	/*
992 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
993 	 */
994 	i = 2;
995 	while (i * 2 <= maxbcachebuf)
996 		i *= 2;
997 	maxbcachebuf = i;
998 	if (maxbcachebuf < MAXBSIZE)
999 		maxbcachebuf = MAXBSIZE;
1000 	if (maxbcachebuf > MAXPHYS)
1001 		maxbcachebuf = MAXPHYS;
1002 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1003 		printf("maxbcachebuf=%d\n", maxbcachebuf);
1004 }
1005 
1006 /*
1007  * bd_speedup - speedup the buffer cache flushing code
1008  */
1009 void
1010 bd_speedup(void)
1011 {
1012 	int needwake;
1013 
1014 	mtx_lock(&bdlock);
1015 	needwake = 0;
1016 	if (bd_speedupreq == 0 || bd_request == 0)
1017 		needwake = 1;
1018 	bd_speedupreq = 1;
1019 	bd_request = 1;
1020 	if (needwake)
1021 		wakeup(&bd_request);
1022 	mtx_unlock(&bdlock);
1023 }
1024 
1025 #ifdef __i386__
1026 #define	TRANSIENT_DENOM	5
1027 #else
1028 #define	TRANSIENT_DENOM 10
1029 #endif
1030 
1031 /*
1032  * Calculating buffer cache scaling values and reserve space for buffer
1033  * headers.  This is called during low level kernel initialization and
1034  * may be called more then once.  We CANNOT write to the memory area
1035  * being reserved at this time.
1036  */
1037 caddr_t
1038 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1039 {
1040 	int tuned_nbuf;
1041 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1042 
1043 	/*
1044 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1045 	 * PAGE_SIZE is >= 1K)
1046 	 */
1047 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1048 
1049 	maxbcachebuf_adjust();
1050 	/*
1051 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1052 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1053 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1054 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1055 	 * the buffer cache we limit the eventual kva reservation to
1056 	 * maxbcache bytes.
1057 	 *
1058 	 * factor represents the 1/4 x ram conversion.
1059 	 */
1060 	if (nbuf == 0) {
1061 		int factor = 4 * BKVASIZE / 1024;
1062 
1063 		nbuf = 50;
1064 		if (physmem_est > 4096)
1065 			nbuf += min((physmem_est - 4096) / factor,
1066 			    65536 / factor);
1067 		if (physmem_est > 65536)
1068 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1069 			    32 * 1024 * 1024 / (factor * 5));
1070 
1071 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1072 			nbuf = maxbcache / BKVASIZE;
1073 		tuned_nbuf = 1;
1074 	} else
1075 		tuned_nbuf = 0;
1076 
1077 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1078 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1079 	if (nbuf > maxbuf) {
1080 		if (!tuned_nbuf)
1081 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1082 			    maxbuf);
1083 		nbuf = maxbuf;
1084 	}
1085 
1086 	/*
1087 	 * Ideal allocation size for the transient bio submap is 10%
1088 	 * of the maximal space buffer map.  This roughly corresponds
1089 	 * to the amount of the buffer mapped for typical UFS load.
1090 	 *
1091 	 * Clip the buffer map to reserve space for the transient
1092 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1093 	 * maximum buffer map extent on the platform.
1094 	 *
1095 	 * The fall-back to the maxbuf in case of maxbcache unset,
1096 	 * allows to not trim the buffer KVA for the architectures
1097 	 * with ample KVA space.
1098 	 */
1099 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1100 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1101 		buf_sz = (long)nbuf * BKVASIZE;
1102 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1103 		    (TRANSIENT_DENOM - 1)) {
1104 			/*
1105 			 * There is more KVA than memory.  Do not
1106 			 * adjust buffer map size, and assign the rest
1107 			 * of maxbuf to transient map.
1108 			 */
1109 			biotmap_sz = maxbuf_sz - buf_sz;
1110 		} else {
1111 			/*
1112 			 * Buffer map spans all KVA we could afford on
1113 			 * this platform.  Give 10% (20% on i386) of
1114 			 * the buffer map to the transient bio map.
1115 			 */
1116 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1117 			buf_sz -= biotmap_sz;
1118 		}
1119 		if (biotmap_sz / INT_MAX > MAXPHYS)
1120 			bio_transient_maxcnt = INT_MAX;
1121 		else
1122 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1123 		/*
1124 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1125 		 * using the transient mapping.
1126 		 */
1127 		if (bio_transient_maxcnt > 1024)
1128 			bio_transient_maxcnt = 1024;
1129 		if (tuned_nbuf)
1130 			nbuf = buf_sz / BKVASIZE;
1131 	}
1132 
1133 	if (nswbuf == 0) {
1134 		nswbuf = min(nbuf / 4, 256);
1135 		if (nswbuf < NSWBUF_MIN)
1136 			nswbuf = NSWBUF_MIN;
1137 	}
1138 
1139 	/*
1140 	 * Reserve space for the buffer cache buffers
1141 	 */
1142 	buf = (void *)v;
1143 	v = (caddr_t)(buf + nbuf);
1144 
1145 	return(v);
1146 }
1147 
1148 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1149 void
1150 bufinit(void)
1151 {
1152 	struct buf *bp;
1153 	int i;
1154 
1155 	KASSERT(maxbcachebuf >= MAXBSIZE,
1156 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1157 	    MAXBSIZE));
1158 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1159 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1160 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1161 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1162 
1163 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1164 
1165 	/* finally, initialize each buffer header and stick on empty q */
1166 	for (i = 0; i < nbuf; i++) {
1167 		bp = &buf[i];
1168 		bzero(bp, sizeof *bp);
1169 		bp->b_flags = B_INVAL;
1170 		bp->b_rcred = NOCRED;
1171 		bp->b_wcred = NOCRED;
1172 		bp->b_qindex = QUEUE_NONE;
1173 		bp->b_domain = -1;
1174 		bp->b_subqueue = mp_maxid + 1;
1175 		bp->b_xflags = 0;
1176 		bp->b_data = bp->b_kvabase = unmapped_buf;
1177 		LIST_INIT(&bp->b_dep);
1178 		BUF_LOCKINIT(bp);
1179 		bq_insert(&bqempty, bp, false);
1180 	}
1181 
1182 	/*
1183 	 * maxbufspace is the absolute maximum amount of buffer space we are
1184 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1185 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1186 	 * used by most other requests.  The differential is required to
1187 	 * ensure that metadata deadlocks don't occur.
1188 	 *
1189 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1190 	 * this may result in KVM fragmentation which is not handled optimally
1191 	 * by the system. XXX This is less true with vmem.  We could use
1192 	 * PAGE_SIZE.
1193 	 */
1194 	maxbufspace = (long)nbuf * BKVASIZE;
1195 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1196 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1197 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1198 
1199 	/*
1200 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1201 	 * arbitrarily and may need further tuning. It corresponds to
1202 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1203 	 * which fits with many RAID controllers' tagged queuing limits.
1204 	 * The lower 1 MiB limit is the historical upper limit for
1205 	 * hirunningspace.
1206 	 */
1207 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1208 	    16 * 1024 * 1024), 1024 * 1024);
1209 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1210 
1211 	/*
1212 	 * Limit the amount of malloc memory since it is wired permanently into
1213 	 * the kernel space.  Even though this is accounted for in the buffer
1214 	 * allocation, we don't want the malloced region to grow uncontrolled.
1215 	 * The malloc scheme improves memory utilization significantly on
1216 	 * average (small) directories.
1217 	 */
1218 	maxbufmallocspace = hibufspace / 20;
1219 
1220 	/*
1221 	 * Reduce the chance of a deadlock occurring by limiting the number
1222 	 * of delayed-write dirty buffers we allow to stack up.
1223 	 */
1224 	hidirtybuffers = nbuf / 4 + 20;
1225 	dirtybufthresh = hidirtybuffers * 9 / 10;
1226 	/*
1227 	 * To support extreme low-memory systems, make sure hidirtybuffers
1228 	 * cannot eat up all available buffer space.  This occurs when our
1229 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1230 	 * buffer space assuming BKVASIZE'd buffers.
1231 	 */
1232 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1233 		hidirtybuffers >>= 1;
1234 	}
1235 	lodirtybuffers = hidirtybuffers / 2;
1236 
1237 	/*
1238 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1239 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1240 	 * are counted as free but will be unavailable to threads executing
1241 	 * on other cpus.
1242 	 *
1243 	 * hifreebuffers is the free target for the bufspace daemon.  This
1244 	 * should be set appropriately to limit work per-iteration.
1245 	 */
1246 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1247 	hifreebuffers = (3 * lofreebuffers) / 2;
1248 	numfreebuffers = nbuf;
1249 
1250 	/* Setup the kva and free list allocators. */
1251 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1252 	buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1253 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1254 
1255 	/*
1256 	 * Size the clean queue according to the amount of buffer space.
1257 	 * One queue per-256mb up to the max.  More queues gives better
1258 	 * concurrency but less accurate LRU.
1259 	 */
1260 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1261 	for (i = 0 ; i < buf_domains; i++) {
1262 		struct bufdomain *bd;
1263 
1264 		bd = &bdomain[i];
1265 		bd_init(bd);
1266 		bd->bd_freebuffers = nbuf / buf_domains;
1267 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1268 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1269 		bd->bd_bufspace = 0;
1270 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1271 		bd->bd_hibufspace = hibufspace / buf_domains;
1272 		bd->bd_lobufspace = lobufspace / buf_domains;
1273 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1274 		bd->bd_numdirtybuffers = 0;
1275 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1276 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1277 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1278 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1279 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1280 	}
1281 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1282 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1283 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1284 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1285 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1286 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1287 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1288 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1289 }
1290 
1291 #ifdef INVARIANTS
1292 static inline void
1293 vfs_buf_check_mapped(struct buf *bp)
1294 {
1295 
1296 	KASSERT(bp->b_kvabase != unmapped_buf,
1297 	    ("mapped buf: b_kvabase was not updated %p", bp));
1298 	KASSERT(bp->b_data != unmapped_buf,
1299 	    ("mapped buf: b_data was not updated %p", bp));
1300 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1301 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1302 }
1303 
1304 static inline void
1305 vfs_buf_check_unmapped(struct buf *bp)
1306 {
1307 
1308 	KASSERT(bp->b_data == unmapped_buf,
1309 	    ("unmapped buf: corrupted b_data %p", bp));
1310 }
1311 
1312 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1313 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1314 #else
1315 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1316 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1317 #endif
1318 
1319 static int
1320 isbufbusy(struct buf *bp)
1321 {
1322 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1323 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1324 		return (1);
1325 	return (0);
1326 }
1327 
1328 /*
1329  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1330  */
1331 void
1332 bufshutdown(int show_busybufs)
1333 {
1334 	static int first_buf_printf = 1;
1335 	struct buf *bp;
1336 	int iter, nbusy, pbusy;
1337 #ifndef PREEMPTION
1338 	int subiter;
1339 #endif
1340 
1341 	/*
1342 	 * Sync filesystems for shutdown
1343 	 */
1344 	wdog_kern_pat(WD_LASTVAL);
1345 	sys_sync(curthread, NULL);
1346 
1347 	/*
1348 	 * With soft updates, some buffers that are
1349 	 * written will be remarked as dirty until other
1350 	 * buffers are written.
1351 	 */
1352 	for (iter = pbusy = 0; iter < 20; iter++) {
1353 		nbusy = 0;
1354 		for (bp = &buf[nbuf]; --bp >= buf; )
1355 			if (isbufbusy(bp))
1356 				nbusy++;
1357 		if (nbusy == 0) {
1358 			if (first_buf_printf)
1359 				printf("All buffers synced.");
1360 			break;
1361 		}
1362 		if (first_buf_printf) {
1363 			printf("Syncing disks, buffers remaining... ");
1364 			first_buf_printf = 0;
1365 		}
1366 		printf("%d ", nbusy);
1367 		if (nbusy < pbusy)
1368 			iter = 0;
1369 		pbusy = nbusy;
1370 
1371 		wdog_kern_pat(WD_LASTVAL);
1372 		sys_sync(curthread, NULL);
1373 
1374 #ifdef PREEMPTION
1375 		/*
1376 		 * Spin for a while to allow interrupt threads to run.
1377 		 */
1378 		DELAY(50000 * iter);
1379 #else
1380 		/*
1381 		 * Context switch several times to allow interrupt
1382 		 * threads to run.
1383 		 */
1384 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1385 			thread_lock(curthread);
1386 			mi_switch(SW_VOL, NULL);
1387 			thread_unlock(curthread);
1388 			DELAY(1000);
1389 		}
1390 #endif
1391 	}
1392 	printf("\n");
1393 	/*
1394 	 * Count only busy local buffers to prevent forcing
1395 	 * a fsck if we're just a client of a wedged NFS server
1396 	 */
1397 	nbusy = 0;
1398 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1399 		if (isbufbusy(bp)) {
1400 #if 0
1401 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1402 			if (bp->b_dev == NULL) {
1403 				TAILQ_REMOVE(&mountlist,
1404 				    bp->b_vp->v_mount, mnt_list);
1405 				continue;
1406 			}
1407 #endif
1408 			nbusy++;
1409 			if (show_busybufs > 0) {
1410 				printf(
1411 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1412 				    nbusy, bp, bp->b_vp, bp->b_flags,
1413 				    (intmax_t)bp->b_blkno,
1414 				    (intmax_t)bp->b_lblkno);
1415 				BUF_LOCKPRINTINFO(bp);
1416 				if (show_busybufs > 1)
1417 					vn_printf(bp->b_vp,
1418 					    "vnode content: ");
1419 			}
1420 		}
1421 	}
1422 	if (nbusy) {
1423 		/*
1424 		 * Failed to sync all blocks. Indicate this and don't
1425 		 * unmount filesystems (thus forcing an fsck on reboot).
1426 		 */
1427 		printf("Giving up on %d buffers\n", nbusy);
1428 		DELAY(5000000);	/* 5 seconds */
1429 	} else {
1430 		if (!first_buf_printf)
1431 			printf("Final sync complete\n");
1432 		/*
1433 		 * Unmount filesystems
1434 		 */
1435 		if (panicstr == NULL)
1436 			vfs_unmountall();
1437 	}
1438 	swapoff_all();
1439 	DELAY(100000);		/* wait for console output to finish */
1440 }
1441 
1442 static void
1443 bpmap_qenter(struct buf *bp)
1444 {
1445 
1446 	BUF_CHECK_MAPPED(bp);
1447 
1448 	/*
1449 	 * bp->b_data is relative to bp->b_offset, but
1450 	 * bp->b_offset may be offset into the first page.
1451 	 */
1452 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1453 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1454 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1455 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1456 }
1457 
1458 static inline struct bufdomain *
1459 bufdomain(struct buf *bp)
1460 {
1461 
1462 	return (&bdomain[bp->b_domain]);
1463 }
1464 
1465 static struct bufqueue *
1466 bufqueue(struct buf *bp)
1467 {
1468 
1469 	switch (bp->b_qindex) {
1470 	case QUEUE_NONE:
1471 		/* FALLTHROUGH */
1472 	case QUEUE_SENTINEL:
1473 		return (NULL);
1474 	case QUEUE_EMPTY:
1475 		return (&bqempty);
1476 	case QUEUE_DIRTY:
1477 		return (&bufdomain(bp)->bd_dirtyq);
1478 	case QUEUE_CLEAN:
1479 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1480 	default:
1481 		break;
1482 	}
1483 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1484 }
1485 
1486 /*
1487  * Return the locked bufqueue that bp is a member of.
1488  */
1489 static struct bufqueue *
1490 bufqueue_acquire(struct buf *bp)
1491 {
1492 	struct bufqueue *bq, *nbq;
1493 
1494 	/*
1495 	 * bp can be pushed from a per-cpu queue to the
1496 	 * cleanq while we're waiting on the lock.  Retry
1497 	 * if the queues don't match.
1498 	 */
1499 	bq = bufqueue(bp);
1500 	BQ_LOCK(bq);
1501 	for (;;) {
1502 		nbq = bufqueue(bp);
1503 		if (bq == nbq)
1504 			break;
1505 		BQ_UNLOCK(bq);
1506 		BQ_LOCK(nbq);
1507 		bq = nbq;
1508 	}
1509 	return (bq);
1510 }
1511 
1512 /*
1513  *	binsfree:
1514  *
1515  *	Insert the buffer into the appropriate free list.  Requires a
1516  *	locked buffer on entry and buffer is unlocked before return.
1517  */
1518 static void
1519 binsfree(struct buf *bp, int qindex)
1520 {
1521 	struct bufdomain *bd;
1522 	struct bufqueue *bq;
1523 
1524 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1525 	    ("binsfree: Invalid qindex %d", qindex));
1526 	BUF_ASSERT_XLOCKED(bp);
1527 
1528 	/*
1529 	 * Handle delayed bremfree() processing.
1530 	 */
1531 	if (bp->b_flags & B_REMFREE) {
1532 		if (bp->b_qindex == qindex) {
1533 			bp->b_flags |= B_REUSE;
1534 			bp->b_flags &= ~B_REMFREE;
1535 			BUF_UNLOCK(bp);
1536 			return;
1537 		}
1538 		bq = bufqueue_acquire(bp);
1539 		bq_remove(bq, bp);
1540 		BQ_UNLOCK(bq);
1541 	}
1542 	bd = bufdomain(bp);
1543 	if (qindex == QUEUE_CLEAN) {
1544 		if (bd->bd_lim != 0)
1545 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1546 		else
1547 			bq = bd->bd_cleanq;
1548 	} else
1549 		bq = &bd->bd_dirtyq;
1550 	bq_insert(bq, bp, true);
1551 }
1552 
1553 /*
1554  * buf_free:
1555  *
1556  *	Free a buffer to the buf zone once it no longer has valid contents.
1557  */
1558 static void
1559 buf_free(struct buf *bp)
1560 {
1561 
1562 	if (bp->b_flags & B_REMFREE)
1563 		bremfreef(bp);
1564 	if (bp->b_vflags & BV_BKGRDINPROG)
1565 		panic("losing buffer 1");
1566 	if (bp->b_rcred != NOCRED) {
1567 		crfree(bp->b_rcred);
1568 		bp->b_rcred = NOCRED;
1569 	}
1570 	if (bp->b_wcred != NOCRED) {
1571 		crfree(bp->b_wcred);
1572 		bp->b_wcred = NOCRED;
1573 	}
1574 	if (!LIST_EMPTY(&bp->b_dep))
1575 		buf_deallocate(bp);
1576 	bufkva_free(bp);
1577 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1578 	BUF_UNLOCK(bp);
1579 	uma_zfree(buf_zone, bp);
1580 }
1581 
1582 /*
1583  * buf_import:
1584  *
1585  *	Import bufs into the uma cache from the buf list.  The system still
1586  *	expects a static array of bufs and much of the synchronization
1587  *	around bufs assumes type stable storage.  As a result, UMA is used
1588  *	only as a per-cpu cache of bufs still maintained on a global list.
1589  */
1590 static int
1591 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1592 {
1593 	struct buf *bp;
1594 	int i;
1595 
1596 	BQ_LOCK(&bqempty);
1597 	for (i = 0; i < cnt; i++) {
1598 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1599 		if (bp == NULL)
1600 			break;
1601 		bq_remove(&bqempty, bp);
1602 		store[i] = bp;
1603 	}
1604 	BQ_UNLOCK(&bqempty);
1605 
1606 	return (i);
1607 }
1608 
1609 /*
1610  * buf_release:
1611  *
1612  *	Release bufs from the uma cache back to the buffer queues.
1613  */
1614 static void
1615 buf_release(void *arg, void **store, int cnt)
1616 {
1617 	struct bufqueue *bq;
1618 	struct buf *bp;
1619         int i;
1620 
1621 	bq = &bqempty;
1622 	BQ_LOCK(bq);
1623         for (i = 0; i < cnt; i++) {
1624 		bp = store[i];
1625 		/* Inline bq_insert() to batch locking. */
1626 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1627 		bp->b_flags &= ~(B_AGE | B_REUSE);
1628 		bq->bq_len++;
1629 		bp->b_qindex = bq->bq_index;
1630 	}
1631 	BQ_UNLOCK(bq);
1632 }
1633 
1634 /*
1635  * buf_alloc:
1636  *
1637  *	Allocate an empty buffer header.
1638  */
1639 static struct buf *
1640 buf_alloc(struct bufdomain *bd)
1641 {
1642 	struct buf *bp;
1643 	int freebufs;
1644 
1645 	/*
1646 	 * We can only run out of bufs in the buf zone if the average buf
1647 	 * is less than BKVASIZE.  In this case the actual wait/block will
1648 	 * come from buf_reycle() failing to flush one of these small bufs.
1649 	 */
1650 	bp = NULL;
1651 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1652 	if (freebufs > 0)
1653 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1654 	if (bp == NULL) {
1655 		atomic_add_int(&bd->bd_freebuffers, 1);
1656 		bufspace_daemon_wakeup(bd);
1657 		counter_u64_add(numbufallocfails, 1);
1658 		return (NULL);
1659 	}
1660 	/*
1661 	 * Wake-up the bufspace daemon on transition below threshold.
1662 	 */
1663 	if (freebufs == bd->bd_lofreebuffers)
1664 		bufspace_daemon_wakeup(bd);
1665 
1666 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1667 		panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1668 
1669 	KASSERT(bp->b_vp == NULL,
1670 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1671 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1672 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1673 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1674 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1675 	KASSERT(bp->b_npages == 0,
1676 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1677 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1678 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1679 
1680 	bp->b_domain = BD_DOMAIN(bd);
1681 	bp->b_flags = 0;
1682 	bp->b_ioflags = 0;
1683 	bp->b_xflags = 0;
1684 	bp->b_vflags = 0;
1685 	bp->b_vp = NULL;
1686 	bp->b_blkno = bp->b_lblkno = 0;
1687 	bp->b_offset = NOOFFSET;
1688 	bp->b_iodone = 0;
1689 	bp->b_error = 0;
1690 	bp->b_resid = 0;
1691 	bp->b_bcount = 0;
1692 	bp->b_npages = 0;
1693 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1694 	bp->b_bufobj = NULL;
1695 	bp->b_data = bp->b_kvabase = unmapped_buf;
1696 	bp->b_fsprivate1 = NULL;
1697 	bp->b_fsprivate2 = NULL;
1698 	bp->b_fsprivate3 = NULL;
1699 	LIST_INIT(&bp->b_dep);
1700 
1701 	return (bp);
1702 }
1703 
1704 /*
1705  *	buf_recycle:
1706  *
1707  *	Free a buffer from the given bufqueue.  kva controls whether the
1708  *	freed buf must own some kva resources.  This is used for
1709  *	defragmenting.
1710  */
1711 static int
1712 buf_recycle(struct bufdomain *bd, bool kva)
1713 {
1714 	struct bufqueue *bq;
1715 	struct buf *bp, *nbp;
1716 
1717 	if (kva)
1718 		counter_u64_add(bufdefragcnt, 1);
1719 	nbp = NULL;
1720 	bq = bd->bd_cleanq;
1721 	BQ_LOCK(bq);
1722 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1723 	    ("buf_recycle: Locks don't match"));
1724 	nbp = TAILQ_FIRST(&bq->bq_queue);
1725 
1726 	/*
1727 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1728 	 * depending.
1729 	 */
1730 	while ((bp = nbp) != NULL) {
1731 		/*
1732 		 * Calculate next bp (we can only use it if we do not
1733 		 * release the bqlock).
1734 		 */
1735 		nbp = TAILQ_NEXT(bp, b_freelist);
1736 
1737 		/*
1738 		 * If we are defragging then we need a buffer with
1739 		 * some kva to reclaim.
1740 		 */
1741 		if (kva && bp->b_kvasize == 0)
1742 			continue;
1743 
1744 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1745 			continue;
1746 
1747 		/*
1748 		 * Implement a second chance algorithm for frequently
1749 		 * accessed buffers.
1750 		 */
1751 		if ((bp->b_flags & B_REUSE) != 0) {
1752 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1753 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1754 			bp->b_flags &= ~B_REUSE;
1755 			BUF_UNLOCK(bp);
1756 			continue;
1757 		}
1758 
1759 		/*
1760 		 * Skip buffers with background writes in progress.
1761 		 */
1762 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1763 			BUF_UNLOCK(bp);
1764 			continue;
1765 		}
1766 
1767 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1768 		    ("buf_recycle: inconsistent queue %d bp %p",
1769 		    bp->b_qindex, bp));
1770 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1771 		    ("getnewbuf: queue domain %d doesn't match request %d",
1772 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1773 		/*
1774 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1775 		 * the scan from this point on.
1776 		 */
1777 		bq_remove(bq, bp);
1778 		BQ_UNLOCK(bq);
1779 
1780 		/*
1781 		 * Requeue the background write buffer with error and
1782 		 * restart the scan.
1783 		 */
1784 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1785 			bqrelse(bp);
1786 			BQ_LOCK(bq);
1787 			nbp = TAILQ_FIRST(&bq->bq_queue);
1788 			continue;
1789 		}
1790 		bp->b_flags |= B_INVAL;
1791 		brelse(bp);
1792 		return (0);
1793 	}
1794 	bd->bd_wanted = 1;
1795 	BQ_UNLOCK(bq);
1796 
1797 	return (ENOBUFS);
1798 }
1799 
1800 /*
1801  *	bremfree:
1802  *
1803  *	Mark the buffer for removal from the appropriate free list.
1804  *
1805  */
1806 void
1807 bremfree(struct buf *bp)
1808 {
1809 
1810 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1811 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1812 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1813 	KASSERT(bp->b_qindex != QUEUE_NONE,
1814 	    ("bremfree: buffer %p not on a queue.", bp));
1815 	BUF_ASSERT_XLOCKED(bp);
1816 
1817 	bp->b_flags |= B_REMFREE;
1818 }
1819 
1820 /*
1821  *	bremfreef:
1822  *
1823  *	Force an immediate removal from a free list.  Used only in nfs when
1824  *	it abuses the b_freelist pointer.
1825  */
1826 void
1827 bremfreef(struct buf *bp)
1828 {
1829 	struct bufqueue *bq;
1830 
1831 	bq = bufqueue_acquire(bp);
1832 	bq_remove(bq, bp);
1833 	BQ_UNLOCK(bq);
1834 }
1835 
1836 static void
1837 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1838 {
1839 
1840 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1841 	TAILQ_INIT(&bq->bq_queue);
1842 	bq->bq_len = 0;
1843 	bq->bq_index = qindex;
1844 	bq->bq_subqueue = subqueue;
1845 }
1846 
1847 static void
1848 bd_init(struct bufdomain *bd)
1849 {
1850 	int i;
1851 
1852 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1853 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1854 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1855 	for (i = 0; i <= mp_maxid; i++)
1856 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1857 		    "bufq clean subqueue lock");
1858 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1859 }
1860 
1861 /*
1862  *	bq_remove:
1863  *
1864  *	Removes a buffer from the free list, must be called with the
1865  *	correct qlock held.
1866  */
1867 static void
1868 bq_remove(struct bufqueue *bq, struct buf *bp)
1869 {
1870 
1871 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1872 	    bp, bp->b_vp, bp->b_flags);
1873 	KASSERT(bp->b_qindex != QUEUE_NONE,
1874 	    ("bq_remove: buffer %p not on a queue.", bp));
1875 	KASSERT(bufqueue(bp) == bq,
1876 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1877 
1878 	BQ_ASSERT_LOCKED(bq);
1879 	if (bp->b_qindex != QUEUE_EMPTY) {
1880 		BUF_ASSERT_XLOCKED(bp);
1881 	}
1882 	KASSERT(bq->bq_len >= 1,
1883 	    ("queue %d underflow", bp->b_qindex));
1884 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1885 	bq->bq_len--;
1886 	bp->b_qindex = QUEUE_NONE;
1887 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1888 }
1889 
1890 static void
1891 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1892 {
1893 	struct buf *bp;
1894 
1895 	BQ_ASSERT_LOCKED(bq);
1896 	if (bq != bd->bd_cleanq) {
1897 		BD_LOCK(bd);
1898 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1899 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1900 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1901 			    b_freelist);
1902 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1903 		}
1904 		bd->bd_cleanq->bq_len += bq->bq_len;
1905 		bq->bq_len = 0;
1906 	}
1907 	if (bd->bd_wanted) {
1908 		bd->bd_wanted = 0;
1909 		wakeup(&bd->bd_wanted);
1910 	}
1911 	if (bq != bd->bd_cleanq)
1912 		BD_UNLOCK(bd);
1913 }
1914 
1915 static int
1916 bd_flushall(struct bufdomain *bd)
1917 {
1918 	struct bufqueue *bq;
1919 	int flushed;
1920 	int i;
1921 
1922 	if (bd->bd_lim == 0)
1923 		return (0);
1924 	flushed = 0;
1925 	for (i = 0; i <= mp_maxid; i++) {
1926 		bq = &bd->bd_subq[i];
1927 		if (bq->bq_len == 0)
1928 			continue;
1929 		BQ_LOCK(bq);
1930 		bd_flush(bd, bq);
1931 		BQ_UNLOCK(bq);
1932 		flushed++;
1933 	}
1934 
1935 	return (flushed);
1936 }
1937 
1938 static void
1939 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1940 {
1941 	struct bufdomain *bd;
1942 
1943 	if (bp->b_qindex != QUEUE_NONE)
1944 		panic("bq_insert: free buffer %p onto another queue?", bp);
1945 
1946 	bd = bufdomain(bp);
1947 	if (bp->b_flags & B_AGE) {
1948 		/* Place this buf directly on the real queue. */
1949 		if (bq->bq_index == QUEUE_CLEAN)
1950 			bq = bd->bd_cleanq;
1951 		BQ_LOCK(bq);
1952 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1953 	} else {
1954 		BQ_LOCK(bq);
1955 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1956 	}
1957 	bp->b_flags &= ~(B_AGE | B_REUSE);
1958 	bq->bq_len++;
1959 	bp->b_qindex = bq->bq_index;
1960 	bp->b_subqueue = bq->bq_subqueue;
1961 
1962 	/*
1963 	 * Unlock before we notify so that we don't wakeup a waiter that
1964 	 * fails a trylock on the buf and sleeps again.
1965 	 */
1966 	if (unlock)
1967 		BUF_UNLOCK(bp);
1968 
1969 	if (bp->b_qindex == QUEUE_CLEAN) {
1970 		/*
1971 		 * Flush the per-cpu queue and notify any waiters.
1972 		 */
1973 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1974 		    bq->bq_len >= bd->bd_lim))
1975 			bd_flush(bd, bq);
1976 	}
1977 	BQ_UNLOCK(bq);
1978 }
1979 
1980 /*
1981  *	bufkva_free:
1982  *
1983  *	Free the kva allocation for a buffer.
1984  *
1985  */
1986 static void
1987 bufkva_free(struct buf *bp)
1988 {
1989 
1990 #ifdef INVARIANTS
1991 	if (bp->b_kvasize == 0) {
1992 		KASSERT(bp->b_kvabase == unmapped_buf &&
1993 		    bp->b_data == unmapped_buf,
1994 		    ("Leaked KVA space on %p", bp));
1995 	} else if (buf_mapped(bp))
1996 		BUF_CHECK_MAPPED(bp);
1997 	else
1998 		BUF_CHECK_UNMAPPED(bp);
1999 #endif
2000 	if (bp->b_kvasize == 0)
2001 		return;
2002 
2003 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2004 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2005 	counter_u64_add(buffreekvacnt, 1);
2006 	bp->b_data = bp->b_kvabase = unmapped_buf;
2007 	bp->b_kvasize = 0;
2008 }
2009 
2010 /*
2011  *	bufkva_alloc:
2012  *
2013  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2014  */
2015 static int
2016 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2017 {
2018 	vm_offset_t addr;
2019 	int error;
2020 
2021 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2022 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2023 
2024 	bufkva_free(bp);
2025 
2026 	addr = 0;
2027 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2028 	if (error != 0) {
2029 		/*
2030 		 * Buffer map is too fragmented.  Request the caller
2031 		 * to defragment the map.
2032 		 */
2033 		return (error);
2034 	}
2035 	bp->b_kvabase = (caddr_t)addr;
2036 	bp->b_kvasize = maxsize;
2037 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2038 	if ((gbflags & GB_UNMAPPED) != 0) {
2039 		bp->b_data = unmapped_buf;
2040 		BUF_CHECK_UNMAPPED(bp);
2041 	} else {
2042 		bp->b_data = bp->b_kvabase;
2043 		BUF_CHECK_MAPPED(bp);
2044 	}
2045 	return (0);
2046 }
2047 
2048 /*
2049  *	bufkva_reclaim:
2050  *
2051  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2052  *	callback that fires to avoid returning failure.
2053  */
2054 static void
2055 bufkva_reclaim(vmem_t *vmem, int flags)
2056 {
2057 	bool done;
2058 	int q;
2059 	int i;
2060 
2061 	done = false;
2062 	for (i = 0; i < 5; i++) {
2063 		for (q = 0; q < buf_domains; q++)
2064 			if (buf_recycle(&bdomain[q], true) != 0)
2065 				done = true;
2066 		if (done)
2067 			break;
2068 	}
2069 	return;
2070 }
2071 
2072 /*
2073  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2074  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2075  * the buffer is valid and we do not have to do anything.
2076  */
2077 static void
2078 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2079     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2080 {
2081 	struct buf *rabp;
2082 	struct thread *td;
2083 	int i;
2084 
2085 	td = curthread;
2086 
2087 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2088 		if (inmem(vp, *rablkno))
2089 			continue;
2090 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2091 		if ((rabp->b_flags & B_CACHE) != 0) {
2092 			brelse(rabp);
2093 			continue;
2094 		}
2095 #ifdef RACCT
2096 		if (racct_enable) {
2097 			PROC_LOCK(curproc);
2098 			racct_add_buf(curproc, rabp, 0);
2099 			PROC_UNLOCK(curproc);
2100 		}
2101 #endif /* RACCT */
2102 		td->td_ru.ru_inblock++;
2103 		rabp->b_flags |= B_ASYNC;
2104 		rabp->b_flags &= ~B_INVAL;
2105 		if ((flags & GB_CKHASH) != 0) {
2106 			rabp->b_flags |= B_CKHASH;
2107 			rabp->b_ckhashcalc = ckhashfunc;
2108 		}
2109 		rabp->b_ioflags &= ~BIO_ERROR;
2110 		rabp->b_iocmd = BIO_READ;
2111 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2112 			rabp->b_rcred = crhold(cred);
2113 		vfs_busy_pages(rabp, 0);
2114 		BUF_KERNPROC(rabp);
2115 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2116 		bstrategy(rabp);
2117 	}
2118 }
2119 
2120 /*
2121  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2122  *
2123  * Get a buffer with the specified data.  Look in the cache first.  We
2124  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2125  * is set, the buffer is valid and we do not have to do anything, see
2126  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2127  *
2128  * Always return a NULL buffer pointer (in bpp) when returning an error.
2129  */
2130 int
2131 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
2132     int *rabsize, int cnt, struct ucred *cred, int flags,
2133     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2134 {
2135 	struct buf *bp;
2136 	struct thread *td;
2137 	int error, readwait, rv;
2138 
2139 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2140 	td = curthread;
2141 	/*
2142 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2143 	 * are specified.
2144 	 */
2145 	error = getblkx(vp, blkno, size, 0, 0, flags, &bp);
2146 	if (error != 0) {
2147 		*bpp = NULL;
2148 		return (error);
2149 	}
2150 	flags &= ~GB_NOSPARSE;
2151 	*bpp = bp;
2152 
2153 	/*
2154 	 * If not found in cache, do some I/O
2155 	 */
2156 	readwait = 0;
2157 	if ((bp->b_flags & B_CACHE) == 0) {
2158 #ifdef RACCT
2159 		if (racct_enable) {
2160 			PROC_LOCK(td->td_proc);
2161 			racct_add_buf(td->td_proc, bp, 0);
2162 			PROC_UNLOCK(td->td_proc);
2163 		}
2164 #endif /* RACCT */
2165 		td->td_ru.ru_inblock++;
2166 		bp->b_iocmd = BIO_READ;
2167 		bp->b_flags &= ~B_INVAL;
2168 		if ((flags & GB_CKHASH) != 0) {
2169 			bp->b_flags |= B_CKHASH;
2170 			bp->b_ckhashcalc = ckhashfunc;
2171 		}
2172 		bp->b_ioflags &= ~BIO_ERROR;
2173 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2174 			bp->b_rcred = crhold(cred);
2175 		vfs_busy_pages(bp, 0);
2176 		bp->b_iooffset = dbtob(bp->b_blkno);
2177 		bstrategy(bp);
2178 		++readwait;
2179 	}
2180 
2181 	/*
2182 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2183 	 */
2184 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2185 
2186 	rv = 0;
2187 	if (readwait) {
2188 		rv = bufwait(bp);
2189 		if (rv != 0) {
2190 			brelse(bp);
2191 			*bpp = NULL;
2192 		}
2193 	}
2194 	return (rv);
2195 }
2196 
2197 /*
2198  * Write, release buffer on completion.  (Done by iodone
2199  * if async).  Do not bother writing anything if the buffer
2200  * is invalid.
2201  *
2202  * Note that we set B_CACHE here, indicating that buffer is
2203  * fully valid and thus cacheable.  This is true even of NFS
2204  * now so we set it generally.  This could be set either here
2205  * or in biodone() since the I/O is synchronous.  We put it
2206  * here.
2207  */
2208 int
2209 bufwrite(struct buf *bp)
2210 {
2211 	int oldflags;
2212 	struct vnode *vp;
2213 	long space;
2214 	int vp_md;
2215 
2216 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2217 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2218 		bp->b_flags |= B_INVAL | B_RELBUF;
2219 		bp->b_flags &= ~B_CACHE;
2220 		brelse(bp);
2221 		return (ENXIO);
2222 	}
2223 	if (bp->b_flags & B_INVAL) {
2224 		brelse(bp);
2225 		return (0);
2226 	}
2227 
2228 	if (bp->b_flags & B_BARRIER)
2229 		atomic_add_long(&barrierwrites, 1);
2230 
2231 	oldflags = bp->b_flags;
2232 
2233 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2234 	    ("FFS background buffer should not get here %p", bp));
2235 
2236 	vp = bp->b_vp;
2237 	if (vp)
2238 		vp_md = vp->v_vflag & VV_MD;
2239 	else
2240 		vp_md = 0;
2241 
2242 	/*
2243 	 * Mark the buffer clean.  Increment the bufobj write count
2244 	 * before bundirty() call, to prevent other thread from seeing
2245 	 * empty dirty list and zero counter for writes in progress,
2246 	 * falsely indicating that the bufobj is clean.
2247 	 */
2248 	bufobj_wref(bp->b_bufobj);
2249 	bundirty(bp);
2250 
2251 	bp->b_flags &= ~B_DONE;
2252 	bp->b_ioflags &= ~BIO_ERROR;
2253 	bp->b_flags |= B_CACHE;
2254 	bp->b_iocmd = BIO_WRITE;
2255 
2256 	vfs_busy_pages(bp, 1);
2257 
2258 	/*
2259 	 * Normal bwrites pipeline writes
2260 	 */
2261 	bp->b_runningbufspace = bp->b_bufsize;
2262 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2263 
2264 #ifdef RACCT
2265 	if (racct_enable) {
2266 		PROC_LOCK(curproc);
2267 		racct_add_buf(curproc, bp, 1);
2268 		PROC_UNLOCK(curproc);
2269 	}
2270 #endif /* RACCT */
2271 	curthread->td_ru.ru_oublock++;
2272 	if (oldflags & B_ASYNC)
2273 		BUF_KERNPROC(bp);
2274 	bp->b_iooffset = dbtob(bp->b_blkno);
2275 	buf_track(bp, __func__);
2276 	bstrategy(bp);
2277 
2278 	if ((oldflags & B_ASYNC) == 0) {
2279 		int rtval = bufwait(bp);
2280 		brelse(bp);
2281 		return (rtval);
2282 	} else if (space > hirunningspace) {
2283 		/*
2284 		 * don't allow the async write to saturate the I/O
2285 		 * system.  We will not deadlock here because
2286 		 * we are blocking waiting for I/O that is already in-progress
2287 		 * to complete. We do not block here if it is the update
2288 		 * or syncer daemon trying to clean up as that can lead
2289 		 * to deadlock.
2290 		 */
2291 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2292 			waitrunningbufspace();
2293 	}
2294 
2295 	return (0);
2296 }
2297 
2298 void
2299 bufbdflush(struct bufobj *bo, struct buf *bp)
2300 {
2301 	struct buf *nbp;
2302 
2303 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2304 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2305 		altbufferflushes++;
2306 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2307 		BO_LOCK(bo);
2308 		/*
2309 		 * Try to find a buffer to flush.
2310 		 */
2311 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2312 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2313 			    BUF_LOCK(nbp,
2314 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2315 				continue;
2316 			if (bp == nbp)
2317 				panic("bdwrite: found ourselves");
2318 			BO_UNLOCK(bo);
2319 			/* Don't countdeps with the bo lock held. */
2320 			if (buf_countdeps(nbp, 0)) {
2321 				BO_LOCK(bo);
2322 				BUF_UNLOCK(nbp);
2323 				continue;
2324 			}
2325 			if (nbp->b_flags & B_CLUSTEROK) {
2326 				vfs_bio_awrite(nbp);
2327 			} else {
2328 				bremfree(nbp);
2329 				bawrite(nbp);
2330 			}
2331 			dirtybufferflushes++;
2332 			break;
2333 		}
2334 		if (nbp == NULL)
2335 			BO_UNLOCK(bo);
2336 	}
2337 }
2338 
2339 /*
2340  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2341  * anything if the buffer is marked invalid.
2342  *
2343  * Note that since the buffer must be completely valid, we can safely
2344  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2345  * biodone() in order to prevent getblk from writing the buffer
2346  * out synchronously.
2347  */
2348 void
2349 bdwrite(struct buf *bp)
2350 {
2351 	struct thread *td = curthread;
2352 	struct vnode *vp;
2353 	struct bufobj *bo;
2354 
2355 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2356 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2357 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2358 	    ("Barrier request in delayed write %p", bp));
2359 
2360 	if (bp->b_flags & B_INVAL) {
2361 		brelse(bp);
2362 		return;
2363 	}
2364 
2365 	/*
2366 	 * If we have too many dirty buffers, don't create any more.
2367 	 * If we are wildly over our limit, then force a complete
2368 	 * cleanup. Otherwise, just keep the situation from getting
2369 	 * out of control. Note that we have to avoid a recursive
2370 	 * disaster and not try to clean up after our own cleanup!
2371 	 */
2372 	vp = bp->b_vp;
2373 	bo = bp->b_bufobj;
2374 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2375 		td->td_pflags |= TDP_INBDFLUSH;
2376 		BO_BDFLUSH(bo, bp);
2377 		td->td_pflags &= ~TDP_INBDFLUSH;
2378 	} else
2379 		recursiveflushes++;
2380 
2381 	bdirty(bp);
2382 	/*
2383 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2384 	 * true even of NFS now.
2385 	 */
2386 	bp->b_flags |= B_CACHE;
2387 
2388 	/*
2389 	 * This bmap keeps the system from needing to do the bmap later,
2390 	 * perhaps when the system is attempting to do a sync.  Since it
2391 	 * is likely that the indirect block -- or whatever other datastructure
2392 	 * that the filesystem needs is still in memory now, it is a good
2393 	 * thing to do this.  Note also, that if the pageout daemon is
2394 	 * requesting a sync -- there might not be enough memory to do
2395 	 * the bmap then...  So, this is important to do.
2396 	 */
2397 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2398 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2399 	}
2400 
2401 	buf_track(bp, __func__);
2402 
2403 	/*
2404 	 * Set the *dirty* buffer range based upon the VM system dirty
2405 	 * pages.
2406 	 *
2407 	 * Mark the buffer pages as clean.  We need to do this here to
2408 	 * satisfy the vnode_pager and the pageout daemon, so that it
2409 	 * thinks that the pages have been "cleaned".  Note that since
2410 	 * the pages are in a delayed write buffer -- the VFS layer
2411 	 * "will" see that the pages get written out on the next sync,
2412 	 * or perhaps the cluster will be completed.
2413 	 */
2414 	vfs_clean_pages_dirty_buf(bp);
2415 	bqrelse(bp);
2416 
2417 	/*
2418 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2419 	 * due to the softdep code.
2420 	 */
2421 }
2422 
2423 /*
2424  *	bdirty:
2425  *
2426  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2427  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2428  *	itself to properly update it in the dirty/clean lists.  We mark it
2429  *	B_DONE to ensure that any asynchronization of the buffer properly
2430  *	clears B_DONE ( else a panic will occur later ).
2431  *
2432  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2433  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2434  *	should only be called if the buffer is known-good.
2435  *
2436  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2437  *	count.
2438  *
2439  *	The buffer must be on QUEUE_NONE.
2440  */
2441 void
2442 bdirty(struct buf *bp)
2443 {
2444 
2445 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2446 	    bp, bp->b_vp, bp->b_flags);
2447 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2448 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2449 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2450 	bp->b_flags &= ~(B_RELBUF);
2451 	bp->b_iocmd = BIO_WRITE;
2452 
2453 	if ((bp->b_flags & B_DELWRI) == 0) {
2454 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2455 		reassignbuf(bp);
2456 		bdirtyadd(bp);
2457 	}
2458 }
2459 
2460 /*
2461  *	bundirty:
2462  *
2463  *	Clear B_DELWRI for buffer.
2464  *
2465  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2466  *	count.
2467  *
2468  *	The buffer must be on QUEUE_NONE.
2469  */
2470 
2471 void
2472 bundirty(struct buf *bp)
2473 {
2474 
2475 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2476 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2477 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2478 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2479 
2480 	if (bp->b_flags & B_DELWRI) {
2481 		bp->b_flags &= ~B_DELWRI;
2482 		reassignbuf(bp);
2483 		bdirtysub(bp);
2484 	}
2485 	/*
2486 	 * Since it is now being written, we can clear its deferred write flag.
2487 	 */
2488 	bp->b_flags &= ~B_DEFERRED;
2489 }
2490 
2491 /*
2492  *	bawrite:
2493  *
2494  *	Asynchronous write.  Start output on a buffer, but do not wait for
2495  *	it to complete.  The buffer is released when the output completes.
2496  *
2497  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2498  *	B_INVAL buffers.  Not us.
2499  */
2500 void
2501 bawrite(struct buf *bp)
2502 {
2503 
2504 	bp->b_flags |= B_ASYNC;
2505 	(void) bwrite(bp);
2506 }
2507 
2508 /*
2509  *	babarrierwrite:
2510  *
2511  *	Asynchronous barrier write.  Start output on a buffer, but do not
2512  *	wait for it to complete.  Place a write barrier after this write so
2513  *	that this buffer and all buffers written before it are committed to
2514  *	the disk before any buffers written after this write are committed
2515  *	to the disk.  The buffer is released when the output completes.
2516  */
2517 void
2518 babarrierwrite(struct buf *bp)
2519 {
2520 
2521 	bp->b_flags |= B_ASYNC | B_BARRIER;
2522 	(void) bwrite(bp);
2523 }
2524 
2525 /*
2526  *	bbarrierwrite:
2527  *
2528  *	Synchronous barrier write.  Start output on a buffer and wait for
2529  *	it to complete.  Place a write barrier after this write so that
2530  *	this buffer and all buffers written before it are committed to
2531  *	the disk before any buffers written after this write are committed
2532  *	to the disk.  The buffer is released when the output completes.
2533  */
2534 int
2535 bbarrierwrite(struct buf *bp)
2536 {
2537 
2538 	bp->b_flags |= B_BARRIER;
2539 	return (bwrite(bp));
2540 }
2541 
2542 /*
2543  *	bwillwrite:
2544  *
2545  *	Called prior to the locking of any vnodes when we are expecting to
2546  *	write.  We do not want to starve the buffer cache with too many
2547  *	dirty buffers so we block here.  By blocking prior to the locking
2548  *	of any vnodes we attempt to avoid the situation where a locked vnode
2549  *	prevents the various system daemons from flushing related buffers.
2550  */
2551 void
2552 bwillwrite(void)
2553 {
2554 
2555 	if (buf_dirty_count_severe()) {
2556 		mtx_lock(&bdirtylock);
2557 		while (buf_dirty_count_severe()) {
2558 			bdirtywait = 1;
2559 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2560 			    "flswai", 0);
2561 		}
2562 		mtx_unlock(&bdirtylock);
2563 	}
2564 }
2565 
2566 /*
2567  * Return true if we have too many dirty buffers.
2568  */
2569 int
2570 buf_dirty_count_severe(void)
2571 {
2572 
2573 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2574 }
2575 
2576 /*
2577  *	brelse:
2578  *
2579  *	Release a busy buffer and, if requested, free its resources.  The
2580  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2581  *	to be accessed later as a cache entity or reused for other purposes.
2582  */
2583 void
2584 brelse(struct buf *bp)
2585 {
2586 	struct mount *v_mnt;
2587 	int qindex;
2588 
2589 	/*
2590 	 * Many functions erroneously call brelse with a NULL bp under rare
2591 	 * error conditions. Simply return when called with a NULL bp.
2592 	 */
2593 	if (bp == NULL)
2594 		return;
2595 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2596 	    bp, bp->b_vp, bp->b_flags);
2597 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2598 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2599 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2600 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2601 
2602 	if (BUF_LOCKRECURSED(bp)) {
2603 		/*
2604 		 * Do not process, in particular, do not handle the
2605 		 * B_INVAL/B_RELBUF and do not release to free list.
2606 		 */
2607 		BUF_UNLOCK(bp);
2608 		return;
2609 	}
2610 
2611 	if (bp->b_flags & B_MANAGED) {
2612 		bqrelse(bp);
2613 		return;
2614 	}
2615 
2616 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2617 		BO_LOCK(bp->b_bufobj);
2618 		bp->b_vflags &= ~BV_BKGRDERR;
2619 		BO_UNLOCK(bp->b_bufobj);
2620 		bdirty(bp);
2621 	}
2622 
2623 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2624 	    (bp->b_flags & B_INVALONERR)) {
2625 		/*
2626 		 * Forced invalidation of dirty buffer contents, to be used
2627 		 * after a failed write in the rare case that the loss of the
2628 		 * contents is acceptable.  The buffer is invalidated and
2629 		 * freed.
2630 		 */
2631 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2632 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2633 	}
2634 
2635 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2636 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2637 	    !(bp->b_flags & B_INVAL)) {
2638 		/*
2639 		 * Failed write, redirty.  All errors except ENXIO (which
2640 		 * means the device is gone) are treated as being
2641 		 * transient.
2642 		 *
2643 		 * XXX Treating EIO as transient is not correct; the
2644 		 * contract with the local storage device drivers is that
2645 		 * they will only return EIO once the I/O is no longer
2646 		 * retriable.  Network I/O also respects this through the
2647 		 * guarantees of TCP and/or the internal retries of NFS.
2648 		 * ENOMEM might be transient, but we also have no way of
2649 		 * knowing when its ok to retry/reschedule.  In general,
2650 		 * this entire case should be made obsolete through better
2651 		 * error handling/recovery and resource scheduling.
2652 		 *
2653 		 * Do this also for buffers that failed with ENXIO, but have
2654 		 * non-empty dependencies - the soft updates code might need
2655 		 * to access the buffer to untangle them.
2656 		 *
2657 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2658 		 */
2659 		bp->b_ioflags &= ~BIO_ERROR;
2660 		bdirty(bp);
2661 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2662 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2663 		/*
2664 		 * Either a failed read I/O, or we were asked to free or not
2665 		 * cache the buffer, or we failed to write to a device that's
2666 		 * no longer present.
2667 		 */
2668 		bp->b_flags |= B_INVAL;
2669 		if (!LIST_EMPTY(&bp->b_dep))
2670 			buf_deallocate(bp);
2671 		if (bp->b_flags & B_DELWRI)
2672 			bdirtysub(bp);
2673 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2674 		if ((bp->b_flags & B_VMIO) == 0) {
2675 			allocbuf(bp, 0);
2676 			if (bp->b_vp)
2677 				brelvp(bp);
2678 		}
2679 	}
2680 
2681 	/*
2682 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2683 	 * is called with B_DELWRI set, the underlying pages may wind up
2684 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2685 	 * because pages associated with a B_DELWRI bp are marked clean.
2686 	 *
2687 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2688 	 * if B_DELWRI is set.
2689 	 */
2690 	if (bp->b_flags & B_DELWRI)
2691 		bp->b_flags &= ~B_RELBUF;
2692 
2693 	/*
2694 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2695 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2696 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2697 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2698 	 *
2699 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2700 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2701 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2702 	 *
2703 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2704 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2705 	 * the commit state and we cannot afford to lose the buffer. If the
2706 	 * buffer has a background write in progress, we need to keep it
2707 	 * around to prevent it from being reconstituted and starting a second
2708 	 * background write.
2709 	 */
2710 
2711 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2712 
2713 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2714 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2715 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2716 	    vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2717 		vfs_vmio_invalidate(bp);
2718 		allocbuf(bp, 0);
2719 	}
2720 
2721 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2722 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2723 		allocbuf(bp, 0);
2724 		bp->b_flags &= ~B_NOREUSE;
2725 		if (bp->b_vp != NULL)
2726 			brelvp(bp);
2727 	}
2728 
2729 	/*
2730 	 * If the buffer has junk contents signal it and eventually
2731 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2732 	 * doesn't find it.
2733 	 */
2734 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2735 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2736 		bp->b_flags |= B_INVAL;
2737 	if (bp->b_flags & B_INVAL) {
2738 		if (bp->b_flags & B_DELWRI)
2739 			bundirty(bp);
2740 		if (bp->b_vp)
2741 			brelvp(bp);
2742 	}
2743 
2744 	buf_track(bp, __func__);
2745 
2746 	/* buffers with no memory */
2747 	if (bp->b_bufsize == 0) {
2748 		buf_free(bp);
2749 		return;
2750 	}
2751 	/* buffers with junk contents */
2752 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2753 	    (bp->b_ioflags & BIO_ERROR)) {
2754 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2755 		if (bp->b_vflags & BV_BKGRDINPROG)
2756 			panic("losing buffer 2");
2757 		qindex = QUEUE_CLEAN;
2758 		bp->b_flags |= B_AGE;
2759 	/* remaining buffers */
2760 	} else if (bp->b_flags & B_DELWRI)
2761 		qindex = QUEUE_DIRTY;
2762 	else
2763 		qindex = QUEUE_CLEAN;
2764 
2765 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2766 		panic("brelse: not dirty");
2767 
2768 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2769 	/* binsfree unlocks bp. */
2770 	binsfree(bp, qindex);
2771 }
2772 
2773 /*
2774  * Release a buffer back to the appropriate queue but do not try to free
2775  * it.  The buffer is expected to be used again soon.
2776  *
2777  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2778  * biodone() to requeue an async I/O on completion.  It is also used when
2779  * known good buffers need to be requeued but we think we may need the data
2780  * again soon.
2781  *
2782  * XXX we should be able to leave the B_RELBUF hint set on completion.
2783  */
2784 void
2785 bqrelse(struct buf *bp)
2786 {
2787 	int qindex;
2788 
2789 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2790 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2791 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2792 
2793 	qindex = QUEUE_NONE;
2794 	if (BUF_LOCKRECURSED(bp)) {
2795 		/* do not release to free list */
2796 		BUF_UNLOCK(bp);
2797 		return;
2798 	}
2799 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2800 
2801 	if (bp->b_flags & B_MANAGED) {
2802 		if (bp->b_flags & B_REMFREE)
2803 			bremfreef(bp);
2804 		goto out;
2805 	}
2806 
2807 	/* buffers with stale but valid contents */
2808 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2809 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2810 		BO_LOCK(bp->b_bufobj);
2811 		bp->b_vflags &= ~BV_BKGRDERR;
2812 		BO_UNLOCK(bp->b_bufobj);
2813 		qindex = QUEUE_DIRTY;
2814 	} else {
2815 		if ((bp->b_flags & B_DELWRI) == 0 &&
2816 		    (bp->b_xflags & BX_VNDIRTY))
2817 			panic("bqrelse: not dirty");
2818 		if ((bp->b_flags & B_NOREUSE) != 0) {
2819 			brelse(bp);
2820 			return;
2821 		}
2822 		qindex = QUEUE_CLEAN;
2823 	}
2824 	buf_track(bp, __func__);
2825 	/* binsfree unlocks bp. */
2826 	binsfree(bp, qindex);
2827 	return;
2828 
2829 out:
2830 	buf_track(bp, __func__);
2831 	/* unlock */
2832 	BUF_UNLOCK(bp);
2833 }
2834 
2835 /*
2836  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2837  * restore bogus pages.
2838  */
2839 static void
2840 vfs_vmio_iodone(struct buf *bp)
2841 {
2842 	vm_ooffset_t foff;
2843 	vm_page_t m;
2844 	vm_object_t obj;
2845 	struct vnode *vp __unused;
2846 	int i, iosize, resid;
2847 	bool bogus;
2848 
2849 	obj = bp->b_bufobj->bo_object;
2850 	KASSERT(REFCOUNT_COUNT(obj->paging_in_progress) >= bp->b_npages,
2851 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2852 	    REFCOUNT_COUNT(obj->paging_in_progress), bp->b_npages));
2853 
2854 	vp = bp->b_vp;
2855 	KASSERT(vp->v_holdcnt > 0,
2856 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2857 	KASSERT(vp->v_object != NULL,
2858 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2859 
2860 	foff = bp->b_offset;
2861 	KASSERT(bp->b_offset != NOOFFSET,
2862 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2863 
2864 	bogus = false;
2865 	iosize = bp->b_bcount - bp->b_resid;
2866 	for (i = 0; i < bp->b_npages; i++) {
2867 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2868 		if (resid > iosize)
2869 			resid = iosize;
2870 
2871 		/*
2872 		 * cleanup bogus pages, restoring the originals
2873 		 */
2874 		m = bp->b_pages[i];
2875 		if (m == bogus_page) {
2876 			if (bogus == false) {
2877 				bogus = true;
2878 				VM_OBJECT_RLOCK(obj);
2879 			}
2880 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2881 			if (m == NULL)
2882 				panic("biodone: page disappeared!");
2883 			bp->b_pages[i] = m;
2884 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2885 			/*
2886 			 * In the write case, the valid and clean bits are
2887 			 * already changed correctly ( see bdwrite() ), so we
2888 			 * only need to do this here in the read case.
2889 			 */
2890 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2891 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2892 			    "has unexpected dirty bits", m));
2893 			vfs_page_set_valid(bp, foff, m);
2894 		}
2895 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2896 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2897 		    (intmax_t)foff, (uintmax_t)m->pindex));
2898 
2899 		vm_page_sunbusy(m);
2900 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2901 		iosize -= resid;
2902 	}
2903 	if (bogus)
2904 		VM_OBJECT_RUNLOCK(obj);
2905 	vm_object_pip_wakeupn(obj, bp->b_npages);
2906 	if (bogus && buf_mapped(bp)) {
2907 		BUF_CHECK_MAPPED(bp);
2908 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2909 		    bp->b_pages, bp->b_npages);
2910 	}
2911 }
2912 
2913 /*
2914  * Perform page invalidation when a buffer is released.  The fully invalid
2915  * pages will be reclaimed later in vfs_vmio_truncate().
2916  */
2917 static void
2918 vfs_vmio_invalidate(struct buf *bp)
2919 {
2920 	vm_object_t obj;
2921 	vm_page_t m;
2922 	int flags, i, resid, poffset, presid;
2923 
2924 	if (buf_mapped(bp)) {
2925 		BUF_CHECK_MAPPED(bp);
2926 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2927 	} else
2928 		BUF_CHECK_UNMAPPED(bp);
2929 	/*
2930 	 * Get the base offset and length of the buffer.  Note that
2931 	 * in the VMIO case if the buffer block size is not
2932 	 * page-aligned then b_data pointer may not be page-aligned.
2933 	 * But our b_pages[] array *IS* page aligned.
2934 	 *
2935 	 * block sizes less then DEV_BSIZE (usually 512) are not
2936 	 * supported due to the page granularity bits (m->valid,
2937 	 * m->dirty, etc...).
2938 	 *
2939 	 * See man buf(9) for more information
2940 	 */
2941 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2942 	obj = bp->b_bufobj->bo_object;
2943 	resid = bp->b_bufsize;
2944 	poffset = bp->b_offset & PAGE_MASK;
2945 	VM_OBJECT_WLOCK(obj);
2946 	for (i = 0; i < bp->b_npages; i++) {
2947 		m = bp->b_pages[i];
2948 		if (m == bogus_page)
2949 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2950 		bp->b_pages[i] = NULL;
2951 
2952 		presid = resid > (PAGE_SIZE - poffset) ?
2953 		    (PAGE_SIZE - poffset) : resid;
2954 		KASSERT(presid >= 0, ("brelse: extra page"));
2955 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
2956 		if (pmap_page_wired_mappings(m) == 0)
2957 			vm_page_set_invalid(m, poffset, presid);
2958 		vm_page_sunbusy(m);
2959 		vm_page_release_locked(m, flags);
2960 		resid -= presid;
2961 		poffset = 0;
2962 	}
2963 	VM_OBJECT_WUNLOCK(obj);
2964 	bp->b_npages = 0;
2965 }
2966 
2967 /*
2968  * Page-granular truncation of an existing VMIO buffer.
2969  */
2970 static void
2971 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2972 {
2973 	vm_object_t obj;
2974 	vm_page_t m;
2975 	int flags, i;
2976 
2977 	if (bp->b_npages == desiredpages)
2978 		return;
2979 
2980 	if (buf_mapped(bp)) {
2981 		BUF_CHECK_MAPPED(bp);
2982 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2983 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2984 	} else
2985 		BUF_CHECK_UNMAPPED(bp);
2986 
2987 	/*
2988 	 * The object lock is needed only if we will attempt to free pages.
2989 	 */
2990 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2991 	if ((bp->b_flags & B_DIRECT) != 0) {
2992 		flags |= VPR_TRYFREE;
2993 		obj = bp->b_bufobj->bo_object;
2994 		VM_OBJECT_WLOCK(obj);
2995 	} else {
2996 		obj = NULL;
2997 	}
2998 	for (i = desiredpages; i < bp->b_npages; i++) {
2999 		m = bp->b_pages[i];
3000 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3001 		bp->b_pages[i] = NULL;
3002 		if (obj != NULL)
3003 			vm_page_release_locked(m, flags);
3004 		else
3005 			vm_page_release(m, flags);
3006 	}
3007 	if (obj != NULL)
3008 		VM_OBJECT_WUNLOCK(obj);
3009 	bp->b_npages = desiredpages;
3010 }
3011 
3012 /*
3013  * Byte granular extension of VMIO buffers.
3014  */
3015 static void
3016 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3017 {
3018 	/*
3019 	 * We are growing the buffer, possibly in a
3020 	 * byte-granular fashion.
3021 	 */
3022 	vm_object_t obj;
3023 	vm_offset_t toff;
3024 	vm_offset_t tinc;
3025 	vm_page_t m;
3026 
3027 	/*
3028 	 * Step 1, bring in the VM pages from the object, allocating
3029 	 * them if necessary.  We must clear B_CACHE if these pages
3030 	 * are not valid for the range covered by the buffer.
3031 	 */
3032 	obj = bp->b_bufobj->bo_object;
3033 	if (bp->b_npages < desiredpages) {
3034 		/*
3035 		 * We must allocate system pages since blocking
3036 		 * here could interfere with paging I/O, no
3037 		 * matter which process we are.
3038 		 *
3039 		 * Only exclusive busy can be tested here.
3040 		 * Blocking on shared busy might lead to
3041 		 * deadlocks once allocbuf() is called after
3042 		 * pages are vfs_busy_pages().
3043 		 */
3044 		VM_OBJECT_WLOCK(obj);
3045 		(void)vm_page_grab_pages(obj,
3046 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3047 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3048 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3049 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3050 		VM_OBJECT_WUNLOCK(obj);
3051 		bp->b_npages = desiredpages;
3052 	}
3053 
3054 	/*
3055 	 * Step 2.  We've loaded the pages into the buffer,
3056 	 * we have to figure out if we can still have B_CACHE
3057 	 * set.  Note that B_CACHE is set according to the
3058 	 * byte-granular range ( bcount and size ), not the
3059 	 * aligned range ( newbsize ).
3060 	 *
3061 	 * The VM test is against m->valid, which is DEV_BSIZE
3062 	 * aligned.  Needless to say, the validity of the data
3063 	 * needs to also be DEV_BSIZE aligned.  Note that this
3064 	 * fails with NFS if the server or some other client
3065 	 * extends the file's EOF.  If our buffer is resized,
3066 	 * B_CACHE may remain set! XXX
3067 	 */
3068 	toff = bp->b_bcount;
3069 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3070 	while ((bp->b_flags & B_CACHE) && toff < size) {
3071 		vm_pindex_t pi;
3072 
3073 		if (tinc > (size - toff))
3074 			tinc = size - toff;
3075 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3076 		m = bp->b_pages[pi];
3077 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3078 		toff += tinc;
3079 		tinc = PAGE_SIZE;
3080 	}
3081 
3082 	/*
3083 	 * Step 3, fixup the KVA pmap.
3084 	 */
3085 	if (buf_mapped(bp))
3086 		bpmap_qenter(bp);
3087 	else
3088 		BUF_CHECK_UNMAPPED(bp);
3089 }
3090 
3091 /*
3092  * Check to see if a block at a particular lbn is available for a clustered
3093  * write.
3094  */
3095 static int
3096 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3097 {
3098 	struct buf *bpa;
3099 	int match;
3100 
3101 	match = 0;
3102 
3103 	/* If the buf isn't in core skip it */
3104 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3105 		return (0);
3106 
3107 	/* If the buf is busy we don't want to wait for it */
3108 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3109 		return (0);
3110 
3111 	/* Only cluster with valid clusterable delayed write buffers */
3112 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3113 	    (B_DELWRI | B_CLUSTEROK))
3114 		goto done;
3115 
3116 	if (bpa->b_bufsize != size)
3117 		goto done;
3118 
3119 	/*
3120 	 * Check to see if it is in the expected place on disk and that the
3121 	 * block has been mapped.
3122 	 */
3123 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3124 		match = 1;
3125 done:
3126 	BUF_UNLOCK(bpa);
3127 	return (match);
3128 }
3129 
3130 /*
3131  *	vfs_bio_awrite:
3132  *
3133  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3134  *	This is much better then the old way of writing only one buffer at
3135  *	a time.  Note that we may not be presented with the buffers in the
3136  *	correct order, so we search for the cluster in both directions.
3137  */
3138 int
3139 vfs_bio_awrite(struct buf *bp)
3140 {
3141 	struct bufobj *bo;
3142 	int i;
3143 	int j;
3144 	daddr_t lblkno = bp->b_lblkno;
3145 	struct vnode *vp = bp->b_vp;
3146 	int ncl;
3147 	int nwritten;
3148 	int size;
3149 	int maxcl;
3150 	int gbflags;
3151 
3152 	bo = &vp->v_bufobj;
3153 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3154 	/*
3155 	 * right now we support clustered writing only to regular files.  If
3156 	 * we find a clusterable block we could be in the middle of a cluster
3157 	 * rather then at the beginning.
3158 	 */
3159 	if ((vp->v_type == VREG) &&
3160 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3161 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3162 
3163 		size = vp->v_mount->mnt_stat.f_iosize;
3164 		maxcl = MAXPHYS / size;
3165 
3166 		BO_RLOCK(bo);
3167 		for (i = 1; i < maxcl; i++)
3168 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3169 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3170 				break;
3171 
3172 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3173 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3174 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3175 				break;
3176 		BO_RUNLOCK(bo);
3177 		--j;
3178 		ncl = i + j;
3179 		/*
3180 		 * this is a possible cluster write
3181 		 */
3182 		if (ncl != 1) {
3183 			BUF_UNLOCK(bp);
3184 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3185 			    gbflags);
3186 			return (nwritten);
3187 		}
3188 	}
3189 	bremfree(bp);
3190 	bp->b_flags |= B_ASYNC;
3191 	/*
3192 	 * default (old) behavior, writing out only one block
3193 	 *
3194 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3195 	 */
3196 	nwritten = bp->b_bufsize;
3197 	(void) bwrite(bp);
3198 
3199 	return (nwritten);
3200 }
3201 
3202 /*
3203  *	getnewbuf_kva:
3204  *
3205  *	Allocate KVA for an empty buf header according to gbflags.
3206  */
3207 static int
3208 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3209 {
3210 
3211 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3212 		/*
3213 		 * In order to keep fragmentation sane we only allocate kva
3214 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3215 		 */
3216 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3217 
3218 		if (maxsize != bp->b_kvasize &&
3219 		    bufkva_alloc(bp, maxsize, gbflags))
3220 			return (ENOSPC);
3221 	}
3222 	return (0);
3223 }
3224 
3225 /*
3226  *	getnewbuf:
3227  *
3228  *	Find and initialize a new buffer header, freeing up existing buffers
3229  *	in the bufqueues as necessary.  The new buffer is returned locked.
3230  *
3231  *	We block if:
3232  *		We have insufficient buffer headers
3233  *		We have insufficient buffer space
3234  *		buffer_arena is too fragmented ( space reservation fails )
3235  *		If we have to flush dirty buffers ( but we try to avoid this )
3236  *
3237  *	The caller is responsible for releasing the reserved bufspace after
3238  *	allocbuf() is called.
3239  */
3240 static struct buf *
3241 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3242 {
3243 	struct bufdomain *bd;
3244 	struct buf *bp;
3245 	bool metadata, reserved;
3246 
3247 	bp = NULL;
3248 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3249 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3250 	if (!unmapped_buf_allowed)
3251 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3252 
3253 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3254 	    vp->v_type == VCHR)
3255 		metadata = true;
3256 	else
3257 		metadata = false;
3258 	if (vp == NULL)
3259 		bd = &bdomain[0];
3260 	else
3261 		bd = &bdomain[vp->v_bufobj.bo_domain];
3262 
3263 	counter_u64_add(getnewbufcalls, 1);
3264 	reserved = false;
3265 	do {
3266 		if (reserved == false &&
3267 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3268 			counter_u64_add(getnewbufrestarts, 1);
3269 			continue;
3270 		}
3271 		reserved = true;
3272 		if ((bp = buf_alloc(bd)) == NULL) {
3273 			counter_u64_add(getnewbufrestarts, 1);
3274 			continue;
3275 		}
3276 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3277 			return (bp);
3278 		break;
3279 	} while (buf_recycle(bd, false) == 0);
3280 
3281 	if (reserved)
3282 		bufspace_release(bd, maxsize);
3283 	if (bp != NULL) {
3284 		bp->b_flags |= B_INVAL;
3285 		brelse(bp);
3286 	}
3287 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3288 
3289 	return (NULL);
3290 }
3291 
3292 /*
3293  *	buf_daemon:
3294  *
3295  *	buffer flushing daemon.  Buffers are normally flushed by the
3296  *	update daemon but if it cannot keep up this process starts to
3297  *	take the load in an attempt to prevent getnewbuf() from blocking.
3298  */
3299 static struct kproc_desc buf_kp = {
3300 	"bufdaemon",
3301 	buf_daemon,
3302 	&bufdaemonproc
3303 };
3304 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3305 
3306 static int
3307 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3308 {
3309 	int flushed;
3310 
3311 	flushed = flushbufqueues(vp, bd, target, 0);
3312 	if (flushed == 0) {
3313 		/*
3314 		 * Could not find any buffers without rollback
3315 		 * dependencies, so just write the first one
3316 		 * in the hopes of eventually making progress.
3317 		 */
3318 		if (vp != NULL && target > 2)
3319 			target /= 2;
3320 		flushbufqueues(vp, bd, target, 1);
3321 	}
3322 	return (flushed);
3323 }
3324 
3325 static void
3326 buf_daemon()
3327 {
3328 	struct bufdomain *bd;
3329 	int speedupreq;
3330 	int lodirty;
3331 	int i;
3332 
3333 	/*
3334 	 * This process needs to be suspended prior to shutdown sync.
3335 	 */
3336 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3337 	    SHUTDOWN_PRI_LAST + 100);
3338 
3339 	/*
3340 	 * Start the buf clean daemons as children threads.
3341 	 */
3342 	for (i = 0 ; i < buf_domains; i++) {
3343 		int error;
3344 
3345 		error = kthread_add((void (*)(void *))bufspace_daemon,
3346 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3347 		if (error)
3348 			panic("error %d spawning bufspace daemon", error);
3349 	}
3350 
3351 	/*
3352 	 * This process is allowed to take the buffer cache to the limit
3353 	 */
3354 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3355 	mtx_lock(&bdlock);
3356 	for (;;) {
3357 		bd_request = 0;
3358 		mtx_unlock(&bdlock);
3359 
3360 		kthread_suspend_check();
3361 
3362 		/*
3363 		 * Save speedupreq for this pass and reset to capture new
3364 		 * requests.
3365 		 */
3366 		speedupreq = bd_speedupreq;
3367 		bd_speedupreq = 0;
3368 
3369 		/*
3370 		 * Flush each domain sequentially according to its level and
3371 		 * the speedup request.
3372 		 */
3373 		for (i = 0; i < buf_domains; i++) {
3374 			bd = &bdomain[i];
3375 			if (speedupreq)
3376 				lodirty = bd->bd_numdirtybuffers / 2;
3377 			else
3378 				lodirty = bd->bd_lodirtybuffers;
3379 			while (bd->bd_numdirtybuffers > lodirty) {
3380 				if (buf_flush(NULL, bd,
3381 				    bd->bd_numdirtybuffers - lodirty) == 0)
3382 					break;
3383 				kern_yield(PRI_USER);
3384 			}
3385 		}
3386 
3387 		/*
3388 		 * Only clear bd_request if we have reached our low water
3389 		 * mark.  The buf_daemon normally waits 1 second and
3390 		 * then incrementally flushes any dirty buffers that have
3391 		 * built up, within reason.
3392 		 *
3393 		 * If we were unable to hit our low water mark and couldn't
3394 		 * find any flushable buffers, we sleep for a short period
3395 		 * to avoid endless loops on unlockable buffers.
3396 		 */
3397 		mtx_lock(&bdlock);
3398 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3399 			/*
3400 			 * We reached our low water mark, reset the
3401 			 * request and sleep until we are needed again.
3402 			 * The sleep is just so the suspend code works.
3403 			 */
3404 			bd_request = 0;
3405 			/*
3406 			 * Do an extra wakeup in case dirty threshold
3407 			 * changed via sysctl and the explicit transition
3408 			 * out of shortfall was missed.
3409 			 */
3410 			bdirtywakeup();
3411 			if (runningbufspace <= lorunningspace)
3412 				runningwakeup();
3413 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3414 		} else {
3415 			/*
3416 			 * We couldn't find any flushable dirty buffers but
3417 			 * still have too many dirty buffers, we
3418 			 * have to sleep and try again.  (rare)
3419 			 */
3420 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3421 		}
3422 	}
3423 }
3424 
3425 /*
3426  *	flushbufqueues:
3427  *
3428  *	Try to flush a buffer in the dirty queue.  We must be careful to
3429  *	free up B_INVAL buffers instead of write them, which NFS is
3430  *	particularly sensitive to.
3431  */
3432 static int flushwithdeps = 0;
3433 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3434     &flushwithdeps, 0,
3435     "Number of buffers flushed with dependecies that require rollbacks");
3436 
3437 static int
3438 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3439     int flushdeps)
3440 {
3441 	struct bufqueue *bq;
3442 	struct buf *sentinel;
3443 	struct vnode *vp;
3444 	struct mount *mp;
3445 	struct buf *bp;
3446 	int hasdeps;
3447 	int flushed;
3448 	int error;
3449 	bool unlock;
3450 
3451 	flushed = 0;
3452 	bq = &bd->bd_dirtyq;
3453 	bp = NULL;
3454 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3455 	sentinel->b_qindex = QUEUE_SENTINEL;
3456 	BQ_LOCK(bq);
3457 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3458 	BQ_UNLOCK(bq);
3459 	while (flushed != target) {
3460 		maybe_yield();
3461 		BQ_LOCK(bq);
3462 		bp = TAILQ_NEXT(sentinel, b_freelist);
3463 		if (bp != NULL) {
3464 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3465 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3466 			    b_freelist);
3467 		} else {
3468 			BQ_UNLOCK(bq);
3469 			break;
3470 		}
3471 		/*
3472 		 * Skip sentinels inserted by other invocations of the
3473 		 * flushbufqueues(), taking care to not reorder them.
3474 		 *
3475 		 * Only flush the buffers that belong to the
3476 		 * vnode locked by the curthread.
3477 		 */
3478 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3479 		    bp->b_vp != lvp)) {
3480 			BQ_UNLOCK(bq);
3481 			continue;
3482 		}
3483 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3484 		BQ_UNLOCK(bq);
3485 		if (error != 0)
3486 			continue;
3487 
3488 		/*
3489 		 * BKGRDINPROG can only be set with the buf and bufobj
3490 		 * locks both held.  We tolerate a race to clear it here.
3491 		 */
3492 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3493 		    (bp->b_flags & B_DELWRI) == 0) {
3494 			BUF_UNLOCK(bp);
3495 			continue;
3496 		}
3497 		if (bp->b_flags & B_INVAL) {
3498 			bremfreef(bp);
3499 			brelse(bp);
3500 			flushed++;
3501 			continue;
3502 		}
3503 
3504 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3505 			if (flushdeps == 0) {
3506 				BUF_UNLOCK(bp);
3507 				continue;
3508 			}
3509 			hasdeps = 1;
3510 		} else
3511 			hasdeps = 0;
3512 		/*
3513 		 * We must hold the lock on a vnode before writing
3514 		 * one of its buffers. Otherwise we may confuse, or
3515 		 * in the case of a snapshot vnode, deadlock the
3516 		 * system.
3517 		 *
3518 		 * The lock order here is the reverse of the normal
3519 		 * of vnode followed by buf lock.  This is ok because
3520 		 * the NOWAIT will prevent deadlock.
3521 		 */
3522 		vp = bp->b_vp;
3523 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3524 			BUF_UNLOCK(bp);
3525 			continue;
3526 		}
3527 		if (lvp == NULL) {
3528 			unlock = true;
3529 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3530 		} else {
3531 			ASSERT_VOP_LOCKED(vp, "getbuf");
3532 			unlock = false;
3533 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3534 			    vn_lock(vp, LK_TRYUPGRADE);
3535 		}
3536 		if (error == 0) {
3537 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3538 			    bp, bp->b_vp, bp->b_flags);
3539 			if (curproc == bufdaemonproc) {
3540 				vfs_bio_awrite(bp);
3541 			} else {
3542 				bremfree(bp);
3543 				bwrite(bp);
3544 				counter_u64_add(notbufdflushes, 1);
3545 			}
3546 			vn_finished_write(mp);
3547 			if (unlock)
3548 				VOP_UNLOCK(vp, 0);
3549 			flushwithdeps += hasdeps;
3550 			flushed++;
3551 
3552 			/*
3553 			 * Sleeping on runningbufspace while holding
3554 			 * vnode lock leads to deadlock.
3555 			 */
3556 			if (curproc == bufdaemonproc &&
3557 			    runningbufspace > hirunningspace)
3558 				waitrunningbufspace();
3559 			continue;
3560 		}
3561 		vn_finished_write(mp);
3562 		BUF_UNLOCK(bp);
3563 	}
3564 	BQ_LOCK(bq);
3565 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3566 	BQ_UNLOCK(bq);
3567 	free(sentinel, M_TEMP);
3568 	return (flushed);
3569 }
3570 
3571 /*
3572  * Check to see if a block is currently memory resident.
3573  */
3574 struct buf *
3575 incore(struct bufobj *bo, daddr_t blkno)
3576 {
3577 	struct buf *bp;
3578 
3579 	BO_RLOCK(bo);
3580 	bp = gbincore(bo, blkno);
3581 	BO_RUNLOCK(bo);
3582 	return (bp);
3583 }
3584 
3585 /*
3586  * Returns true if no I/O is needed to access the
3587  * associated VM object.  This is like incore except
3588  * it also hunts around in the VM system for the data.
3589  */
3590 
3591 static int
3592 inmem(struct vnode * vp, daddr_t blkno)
3593 {
3594 	vm_object_t obj;
3595 	vm_offset_t toff, tinc, size;
3596 	vm_page_t m;
3597 	vm_ooffset_t off;
3598 
3599 	ASSERT_VOP_LOCKED(vp, "inmem");
3600 
3601 	if (incore(&vp->v_bufobj, blkno))
3602 		return 1;
3603 	if (vp->v_mount == NULL)
3604 		return 0;
3605 	obj = vp->v_object;
3606 	if (obj == NULL)
3607 		return (0);
3608 
3609 	size = PAGE_SIZE;
3610 	if (size > vp->v_mount->mnt_stat.f_iosize)
3611 		size = vp->v_mount->mnt_stat.f_iosize;
3612 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3613 
3614 	VM_OBJECT_RLOCK(obj);
3615 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3616 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3617 		if (!m)
3618 			goto notinmem;
3619 		tinc = size;
3620 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3621 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3622 		if (vm_page_is_valid(m,
3623 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3624 			goto notinmem;
3625 	}
3626 	VM_OBJECT_RUNLOCK(obj);
3627 	return 1;
3628 
3629 notinmem:
3630 	VM_OBJECT_RUNLOCK(obj);
3631 	return (0);
3632 }
3633 
3634 /*
3635  * Set the dirty range for a buffer based on the status of the dirty
3636  * bits in the pages comprising the buffer.  The range is limited
3637  * to the size of the buffer.
3638  *
3639  * Tell the VM system that the pages associated with this buffer
3640  * are clean.  This is used for delayed writes where the data is
3641  * going to go to disk eventually without additional VM intevention.
3642  *
3643  * Note that while we only really need to clean through to b_bcount, we
3644  * just go ahead and clean through to b_bufsize.
3645  */
3646 static void
3647 vfs_clean_pages_dirty_buf(struct buf *bp)
3648 {
3649 	vm_ooffset_t foff, noff, eoff;
3650 	vm_page_t m;
3651 	int i;
3652 
3653 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3654 		return;
3655 
3656 	foff = bp->b_offset;
3657 	KASSERT(bp->b_offset != NOOFFSET,
3658 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3659 
3660 	vfs_busy_pages_acquire(bp);
3661 	vfs_setdirty_range(bp);
3662 	for (i = 0; i < bp->b_npages; i++) {
3663 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3664 		eoff = noff;
3665 		if (eoff > bp->b_offset + bp->b_bufsize)
3666 			eoff = bp->b_offset + bp->b_bufsize;
3667 		m = bp->b_pages[i];
3668 		vfs_page_set_validclean(bp, foff, m);
3669 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3670 		foff = noff;
3671 	}
3672 	vfs_busy_pages_release(bp);
3673 }
3674 
3675 static void
3676 vfs_setdirty_range(struct buf *bp)
3677 {
3678 	vm_offset_t boffset;
3679 	vm_offset_t eoffset;
3680 	int i;
3681 
3682 	/*
3683 	 * test the pages to see if they have been modified directly
3684 	 * by users through the VM system.
3685 	 */
3686 	for (i = 0; i < bp->b_npages; i++)
3687 		vm_page_test_dirty(bp->b_pages[i]);
3688 
3689 	/*
3690 	 * Calculate the encompassing dirty range, boffset and eoffset,
3691 	 * (eoffset - boffset) bytes.
3692 	 */
3693 
3694 	for (i = 0; i < bp->b_npages; i++) {
3695 		if (bp->b_pages[i]->dirty)
3696 			break;
3697 	}
3698 	boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3699 
3700 	for (i = bp->b_npages - 1; i >= 0; --i) {
3701 		if (bp->b_pages[i]->dirty) {
3702 			break;
3703 		}
3704 	}
3705 	eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3706 
3707 	/*
3708 	 * Fit it to the buffer.
3709 	 */
3710 
3711 	if (eoffset > bp->b_bcount)
3712 		eoffset = bp->b_bcount;
3713 
3714 	/*
3715 	 * If we have a good dirty range, merge with the existing
3716 	 * dirty range.
3717 	 */
3718 
3719 	if (boffset < eoffset) {
3720 		if (bp->b_dirtyoff > boffset)
3721 			bp->b_dirtyoff = boffset;
3722 		if (bp->b_dirtyend < eoffset)
3723 			bp->b_dirtyend = eoffset;
3724 	}
3725 }
3726 
3727 /*
3728  * Allocate the KVA mapping for an existing buffer.
3729  * If an unmapped buffer is provided but a mapped buffer is requested, take
3730  * also care to properly setup mappings between pages and KVA.
3731  */
3732 static void
3733 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3734 {
3735 	int bsize, maxsize, need_mapping, need_kva;
3736 	off_t offset;
3737 
3738 	need_mapping = bp->b_data == unmapped_buf &&
3739 	    (gbflags & GB_UNMAPPED) == 0;
3740 	need_kva = bp->b_kvabase == unmapped_buf &&
3741 	    bp->b_data == unmapped_buf &&
3742 	    (gbflags & GB_KVAALLOC) != 0;
3743 	if (!need_mapping && !need_kva)
3744 		return;
3745 
3746 	BUF_CHECK_UNMAPPED(bp);
3747 
3748 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3749 		/*
3750 		 * Buffer is not mapped, but the KVA was already
3751 		 * reserved at the time of the instantiation.  Use the
3752 		 * allocated space.
3753 		 */
3754 		goto has_addr;
3755 	}
3756 
3757 	/*
3758 	 * Calculate the amount of the address space we would reserve
3759 	 * if the buffer was mapped.
3760 	 */
3761 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3762 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3763 	offset = blkno * bsize;
3764 	maxsize = size + (offset & PAGE_MASK);
3765 	maxsize = imax(maxsize, bsize);
3766 
3767 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3768 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3769 			/*
3770 			 * XXXKIB: defragmentation cannot
3771 			 * succeed, not sure what else to do.
3772 			 */
3773 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3774 		}
3775 		counter_u64_add(mappingrestarts, 1);
3776 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3777 	}
3778 has_addr:
3779 	if (need_mapping) {
3780 		/* b_offset is handled by bpmap_qenter. */
3781 		bp->b_data = bp->b_kvabase;
3782 		BUF_CHECK_MAPPED(bp);
3783 		bpmap_qenter(bp);
3784 	}
3785 }
3786 
3787 struct buf *
3788 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3789     int flags)
3790 {
3791 	struct buf *bp;
3792 	int error;
3793 
3794 	error = getblkx(vp, blkno, size, slpflag, slptimeo, flags, &bp);
3795 	if (error != 0)
3796 		return (NULL);
3797 	return (bp);
3798 }
3799 
3800 /*
3801  *	getblkx:
3802  *
3803  *	Get a block given a specified block and offset into a file/device.
3804  *	The buffers B_DONE bit will be cleared on return, making it almost
3805  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3806  *	return.  The caller should clear B_INVAL prior to initiating a
3807  *	READ.
3808  *
3809  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3810  *	an existing buffer.
3811  *
3812  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3813  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3814  *	and then cleared based on the backing VM.  If the previous buffer is
3815  *	non-0-sized but invalid, B_CACHE will be cleared.
3816  *
3817  *	If getblk() must create a new buffer, the new buffer is returned with
3818  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3819  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3820  *	backing VM.
3821  *
3822  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3823  *	B_CACHE bit is clear.
3824  *
3825  *	What this means, basically, is that the caller should use B_CACHE to
3826  *	determine whether the buffer is fully valid or not and should clear
3827  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3828  *	the buffer by loading its data area with something, the caller needs
3829  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3830  *	the caller should set B_CACHE ( as an optimization ), else the caller
3831  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3832  *	a write attempt or if it was a successful read.  If the caller
3833  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3834  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3835  */
3836 int
3837 getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3838     int flags, struct buf **bpp)
3839 {
3840 	struct buf *bp;
3841 	struct bufobj *bo;
3842 	daddr_t d_blkno;
3843 	int bsize, error, maxsize, vmio;
3844 	off_t offset;
3845 
3846 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3847 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3848 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3849 	ASSERT_VOP_LOCKED(vp, "getblk");
3850 	if (size > maxbcachebuf)
3851 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3852 		    maxbcachebuf);
3853 	if (!unmapped_buf_allowed)
3854 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3855 
3856 	bo = &vp->v_bufobj;
3857 	d_blkno = blkno;
3858 loop:
3859 	BO_RLOCK(bo);
3860 	bp = gbincore(bo, blkno);
3861 	if (bp != NULL) {
3862 		int lockflags;
3863 		/*
3864 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3865 		 * it must be on a queue.
3866 		 */
3867 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3868 
3869 		if ((flags & GB_LOCK_NOWAIT) != 0)
3870 			lockflags |= LK_NOWAIT;
3871 
3872 		error = BUF_TIMELOCK(bp, lockflags,
3873 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3874 
3875 		/*
3876 		 * If we slept and got the lock we have to restart in case
3877 		 * the buffer changed identities.
3878 		 */
3879 		if (error == ENOLCK)
3880 			goto loop;
3881 		/* We timed out or were interrupted. */
3882 		else if (error != 0)
3883 			return (error);
3884 		/* If recursed, assume caller knows the rules. */
3885 		else if (BUF_LOCKRECURSED(bp))
3886 			goto end;
3887 
3888 		/*
3889 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3890 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3891 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3892 		 * backing VM cache.
3893 		 */
3894 		if (bp->b_flags & B_INVAL)
3895 			bp->b_flags &= ~B_CACHE;
3896 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3897 			bp->b_flags |= B_CACHE;
3898 		if (bp->b_flags & B_MANAGED)
3899 			MPASS(bp->b_qindex == QUEUE_NONE);
3900 		else
3901 			bremfree(bp);
3902 
3903 		/*
3904 		 * check for size inconsistencies for non-VMIO case.
3905 		 */
3906 		if (bp->b_bcount != size) {
3907 			if ((bp->b_flags & B_VMIO) == 0 ||
3908 			    (size > bp->b_kvasize)) {
3909 				if (bp->b_flags & B_DELWRI) {
3910 					bp->b_flags |= B_NOCACHE;
3911 					bwrite(bp);
3912 				} else {
3913 					if (LIST_EMPTY(&bp->b_dep)) {
3914 						bp->b_flags |= B_RELBUF;
3915 						brelse(bp);
3916 					} else {
3917 						bp->b_flags |= B_NOCACHE;
3918 						bwrite(bp);
3919 					}
3920 				}
3921 				goto loop;
3922 			}
3923 		}
3924 
3925 		/*
3926 		 * Handle the case of unmapped buffer which should
3927 		 * become mapped, or the buffer for which KVA
3928 		 * reservation is requested.
3929 		 */
3930 		bp_unmapped_get_kva(bp, blkno, size, flags);
3931 
3932 		/*
3933 		 * If the size is inconsistent in the VMIO case, we can resize
3934 		 * the buffer.  This might lead to B_CACHE getting set or
3935 		 * cleared.  If the size has not changed, B_CACHE remains
3936 		 * unchanged from its previous state.
3937 		 */
3938 		allocbuf(bp, size);
3939 
3940 		KASSERT(bp->b_offset != NOOFFSET,
3941 		    ("getblk: no buffer offset"));
3942 
3943 		/*
3944 		 * A buffer with B_DELWRI set and B_CACHE clear must
3945 		 * be committed before we can return the buffer in
3946 		 * order to prevent the caller from issuing a read
3947 		 * ( due to B_CACHE not being set ) and overwriting
3948 		 * it.
3949 		 *
3950 		 * Most callers, including NFS and FFS, need this to
3951 		 * operate properly either because they assume they
3952 		 * can issue a read if B_CACHE is not set, or because
3953 		 * ( for example ) an uncached B_DELWRI might loop due
3954 		 * to softupdates re-dirtying the buffer.  In the latter
3955 		 * case, B_CACHE is set after the first write completes,
3956 		 * preventing further loops.
3957 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3958 		 * above while extending the buffer, we cannot allow the
3959 		 * buffer to remain with B_CACHE set after the write
3960 		 * completes or it will represent a corrupt state.  To
3961 		 * deal with this we set B_NOCACHE to scrap the buffer
3962 		 * after the write.
3963 		 *
3964 		 * We might be able to do something fancy, like setting
3965 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3966 		 * so the below call doesn't set B_CACHE, but that gets real
3967 		 * confusing.  This is much easier.
3968 		 */
3969 
3970 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3971 			bp->b_flags |= B_NOCACHE;
3972 			bwrite(bp);
3973 			goto loop;
3974 		}
3975 		bp->b_flags &= ~B_DONE;
3976 	} else {
3977 		/*
3978 		 * Buffer is not in-core, create new buffer.  The buffer
3979 		 * returned by getnewbuf() is locked.  Note that the returned
3980 		 * buffer is also considered valid (not marked B_INVAL).
3981 		 */
3982 		BO_RUNLOCK(bo);
3983 		/*
3984 		 * If the user does not want us to create the buffer, bail out
3985 		 * here.
3986 		 */
3987 		if (flags & GB_NOCREAT)
3988 			return (EEXIST);
3989 
3990 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3991 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3992 		offset = blkno * bsize;
3993 		vmio = vp->v_object != NULL;
3994 		if (vmio) {
3995 			maxsize = size + (offset & PAGE_MASK);
3996 		} else {
3997 			maxsize = size;
3998 			/* Do not allow non-VMIO notmapped buffers. */
3999 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4000 		}
4001 		maxsize = imax(maxsize, bsize);
4002 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4003 		    !vn_isdisk(vp, NULL)) {
4004 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4005 			KASSERT(error != EOPNOTSUPP,
4006 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4007 			    vp));
4008 			if (error != 0)
4009 				return (error);
4010 			if (d_blkno == -1)
4011 				return (EJUSTRETURN);
4012 		}
4013 
4014 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4015 		if (bp == NULL) {
4016 			if (slpflag || slptimeo)
4017 				return (ETIMEDOUT);
4018 			/*
4019 			 * XXX This is here until the sleep path is diagnosed
4020 			 * enough to work under very low memory conditions.
4021 			 *
4022 			 * There's an issue on low memory, 4BSD+non-preempt
4023 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4024 			 * exhaustion occurs without sleeping for buffer
4025 			 * reclaimation.  This just sticks in a loop and
4026 			 * constantly attempts to allocate a buffer, which
4027 			 * hits exhaustion and tries to wakeup bufdaemon.
4028 			 * This never happens because we never yield.
4029 			 *
4030 			 * The real solution is to identify and fix these cases
4031 			 * so we aren't effectively busy-waiting in a loop
4032 			 * until the reclaimation path has cycles to run.
4033 			 */
4034 			kern_yield(PRI_USER);
4035 			goto loop;
4036 		}
4037 
4038 		/*
4039 		 * This code is used to make sure that a buffer is not
4040 		 * created while the getnewbuf routine is blocked.
4041 		 * This can be a problem whether the vnode is locked or not.
4042 		 * If the buffer is created out from under us, we have to
4043 		 * throw away the one we just created.
4044 		 *
4045 		 * Note: this must occur before we associate the buffer
4046 		 * with the vp especially considering limitations in
4047 		 * the splay tree implementation when dealing with duplicate
4048 		 * lblkno's.
4049 		 */
4050 		BO_LOCK(bo);
4051 		if (gbincore(bo, blkno)) {
4052 			BO_UNLOCK(bo);
4053 			bp->b_flags |= B_INVAL;
4054 			bufspace_release(bufdomain(bp), maxsize);
4055 			brelse(bp);
4056 			goto loop;
4057 		}
4058 
4059 		/*
4060 		 * Insert the buffer into the hash, so that it can
4061 		 * be found by incore.
4062 		 */
4063 		bp->b_lblkno = blkno;
4064 		bp->b_blkno = d_blkno;
4065 		bp->b_offset = offset;
4066 		bgetvp(vp, bp);
4067 		BO_UNLOCK(bo);
4068 
4069 		/*
4070 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4071 		 * buffer size starts out as 0, B_CACHE will be set by
4072 		 * allocbuf() for the VMIO case prior to it testing the
4073 		 * backing store for validity.
4074 		 */
4075 
4076 		if (vmio) {
4077 			bp->b_flags |= B_VMIO;
4078 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4079 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4080 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4081 		} else {
4082 			bp->b_flags &= ~B_VMIO;
4083 			KASSERT(bp->b_bufobj->bo_object == NULL,
4084 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4085 			    bp, bp->b_bufobj->bo_object));
4086 			BUF_CHECK_MAPPED(bp);
4087 		}
4088 
4089 		allocbuf(bp, size);
4090 		bufspace_release(bufdomain(bp), maxsize);
4091 		bp->b_flags &= ~B_DONE;
4092 	}
4093 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4094 end:
4095 	buf_track(bp, __func__);
4096 	KASSERT(bp->b_bufobj == bo,
4097 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4098 	*bpp = bp;
4099 	return (0);
4100 }
4101 
4102 /*
4103  * Get an empty, disassociated buffer of given size.  The buffer is initially
4104  * set to B_INVAL.
4105  */
4106 struct buf *
4107 geteblk(int size, int flags)
4108 {
4109 	struct buf *bp;
4110 	int maxsize;
4111 
4112 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4113 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4114 		if ((flags & GB_NOWAIT_BD) &&
4115 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4116 			return (NULL);
4117 	}
4118 	allocbuf(bp, size);
4119 	bufspace_release(bufdomain(bp), maxsize);
4120 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4121 	return (bp);
4122 }
4123 
4124 /*
4125  * Truncate the backing store for a non-vmio buffer.
4126  */
4127 static void
4128 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4129 {
4130 
4131 	if (bp->b_flags & B_MALLOC) {
4132 		/*
4133 		 * malloced buffers are not shrunk
4134 		 */
4135 		if (newbsize == 0) {
4136 			bufmallocadjust(bp, 0);
4137 			free(bp->b_data, M_BIOBUF);
4138 			bp->b_data = bp->b_kvabase;
4139 			bp->b_flags &= ~B_MALLOC;
4140 		}
4141 		return;
4142 	}
4143 	vm_hold_free_pages(bp, newbsize);
4144 	bufspace_adjust(bp, newbsize);
4145 }
4146 
4147 /*
4148  * Extend the backing for a non-VMIO buffer.
4149  */
4150 static void
4151 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4152 {
4153 	caddr_t origbuf;
4154 	int origbufsize;
4155 
4156 	/*
4157 	 * We only use malloced memory on the first allocation.
4158 	 * and revert to page-allocated memory when the buffer
4159 	 * grows.
4160 	 *
4161 	 * There is a potential smp race here that could lead
4162 	 * to bufmallocspace slightly passing the max.  It
4163 	 * is probably extremely rare and not worth worrying
4164 	 * over.
4165 	 */
4166 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4167 	    bufmallocspace < maxbufmallocspace) {
4168 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4169 		bp->b_flags |= B_MALLOC;
4170 		bufmallocadjust(bp, newbsize);
4171 		return;
4172 	}
4173 
4174 	/*
4175 	 * If the buffer is growing on its other-than-first
4176 	 * allocation then we revert to the page-allocation
4177 	 * scheme.
4178 	 */
4179 	origbuf = NULL;
4180 	origbufsize = 0;
4181 	if (bp->b_flags & B_MALLOC) {
4182 		origbuf = bp->b_data;
4183 		origbufsize = bp->b_bufsize;
4184 		bp->b_data = bp->b_kvabase;
4185 		bufmallocadjust(bp, 0);
4186 		bp->b_flags &= ~B_MALLOC;
4187 		newbsize = round_page(newbsize);
4188 	}
4189 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4190 	    (vm_offset_t) bp->b_data + newbsize);
4191 	if (origbuf != NULL) {
4192 		bcopy(origbuf, bp->b_data, origbufsize);
4193 		free(origbuf, M_BIOBUF);
4194 	}
4195 	bufspace_adjust(bp, newbsize);
4196 }
4197 
4198 /*
4199  * This code constitutes the buffer memory from either anonymous system
4200  * memory (in the case of non-VMIO operations) or from an associated
4201  * VM object (in the case of VMIO operations).  This code is able to
4202  * resize a buffer up or down.
4203  *
4204  * Note that this code is tricky, and has many complications to resolve
4205  * deadlock or inconsistent data situations.  Tread lightly!!!
4206  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4207  * the caller.  Calling this code willy nilly can result in the loss of data.
4208  *
4209  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4210  * B_CACHE for the non-VMIO case.
4211  */
4212 int
4213 allocbuf(struct buf *bp, int size)
4214 {
4215 	int newbsize;
4216 
4217 	if (bp->b_bcount == size)
4218 		return (1);
4219 
4220 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4221 		panic("allocbuf: buffer too small");
4222 
4223 	newbsize = roundup2(size, DEV_BSIZE);
4224 	if ((bp->b_flags & B_VMIO) == 0) {
4225 		if ((bp->b_flags & B_MALLOC) == 0)
4226 			newbsize = round_page(newbsize);
4227 		/*
4228 		 * Just get anonymous memory from the kernel.  Don't
4229 		 * mess with B_CACHE.
4230 		 */
4231 		if (newbsize < bp->b_bufsize)
4232 			vfs_nonvmio_truncate(bp, newbsize);
4233 		else if (newbsize > bp->b_bufsize)
4234 			vfs_nonvmio_extend(bp, newbsize);
4235 	} else {
4236 		int desiredpages;
4237 
4238 		desiredpages = (size == 0) ? 0 :
4239 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4240 
4241 		if (bp->b_flags & B_MALLOC)
4242 			panic("allocbuf: VMIO buffer can't be malloced");
4243 		/*
4244 		 * Set B_CACHE initially if buffer is 0 length or will become
4245 		 * 0-length.
4246 		 */
4247 		if (size == 0 || bp->b_bufsize == 0)
4248 			bp->b_flags |= B_CACHE;
4249 
4250 		if (newbsize < bp->b_bufsize)
4251 			vfs_vmio_truncate(bp, desiredpages);
4252 		/* XXX This looks as if it should be newbsize > b_bufsize */
4253 		else if (size > bp->b_bcount)
4254 			vfs_vmio_extend(bp, desiredpages, size);
4255 		bufspace_adjust(bp, newbsize);
4256 	}
4257 	bp->b_bcount = size;		/* requested buffer size. */
4258 	return (1);
4259 }
4260 
4261 extern int inflight_transient_maps;
4262 
4263 static struct bio_queue nondump_bios;
4264 
4265 void
4266 biodone(struct bio *bp)
4267 {
4268 	struct mtx *mtxp;
4269 	void (*done)(struct bio *);
4270 	vm_offset_t start, end;
4271 
4272 	biotrack(bp, __func__);
4273 
4274 	/*
4275 	 * Avoid completing I/O when dumping after a panic since that may
4276 	 * result in a deadlock in the filesystem or pager code.  Note that
4277 	 * this doesn't affect dumps that were started manually since we aim
4278 	 * to keep the system usable after it has been resumed.
4279 	 */
4280 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4281 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4282 		return;
4283 	}
4284 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4285 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4286 		bp->bio_flags |= BIO_UNMAPPED;
4287 		start = trunc_page((vm_offset_t)bp->bio_data);
4288 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4289 		bp->bio_data = unmapped_buf;
4290 		pmap_qremove(start, atop(end - start));
4291 		vmem_free(transient_arena, start, end - start);
4292 		atomic_add_int(&inflight_transient_maps, -1);
4293 	}
4294 	done = bp->bio_done;
4295 	if (done == NULL) {
4296 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4297 		mtx_lock(mtxp);
4298 		bp->bio_flags |= BIO_DONE;
4299 		wakeup(bp);
4300 		mtx_unlock(mtxp);
4301 	} else
4302 		done(bp);
4303 }
4304 
4305 /*
4306  * Wait for a BIO to finish.
4307  */
4308 int
4309 biowait(struct bio *bp, const char *wchan)
4310 {
4311 	struct mtx *mtxp;
4312 
4313 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4314 	mtx_lock(mtxp);
4315 	while ((bp->bio_flags & BIO_DONE) == 0)
4316 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4317 	mtx_unlock(mtxp);
4318 	if (bp->bio_error != 0)
4319 		return (bp->bio_error);
4320 	if (!(bp->bio_flags & BIO_ERROR))
4321 		return (0);
4322 	return (EIO);
4323 }
4324 
4325 void
4326 biofinish(struct bio *bp, struct devstat *stat, int error)
4327 {
4328 
4329 	if (error) {
4330 		bp->bio_error = error;
4331 		bp->bio_flags |= BIO_ERROR;
4332 	}
4333 	if (stat != NULL)
4334 		devstat_end_transaction_bio(stat, bp);
4335 	biodone(bp);
4336 }
4337 
4338 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4339 void
4340 biotrack_buf(struct bio *bp, const char *location)
4341 {
4342 
4343 	buf_track(bp->bio_track_bp, location);
4344 }
4345 #endif
4346 
4347 /*
4348  *	bufwait:
4349  *
4350  *	Wait for buffer I/O completion, returning error status.  The buffer
4351  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4352  *	error and cleared.
4353  */
4354 int
4355 bufwait(struct buf *bp)
4356 {
4357 	if (bp->b_iocmd == BIO_READ)
4358 		bwait(bp, PRIBIO, "biord");
4359 	else
4360 		bwait(bp, PRIBIO, "biowr");
4361 	if (bp->b_flags & B_EINTR) {
4362 		bp->b_flags &= ~B_EINTR;
4363 		return (EINTR);
4364 	}
4365 	if (bp->b_ioflags & BIO_ERROR) {
4366 		return (bp->b_error ? bp->b_error : EIO);
4367 	} else {
4368 		return (0);
4369 	}
4370 }
4371 
4372 /*
4373  *	bufdone:
4374  *
4375  *	Finish I/O on a buffer, optionally calling a completion function.
4376  *	This is usually called from an interrupt so process blocking is
4377  *	not allowed.
4378  *
4379  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4380  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4381  *	assuming B_INVAL is clear.
4382  *
4383  *	For the VMIO case, we set B_CACHE if the op was a read and no
4384  *	read error occurred, or if the op was a write.  B_CACHE is never
4385  *	set if the buffer is invalid or otherwise uncacheable.
4386  *
4387  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4388  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4389  *	in the biodone routine.
4390  */
4391 void
4392 bufdone(struct buf *bp)
4393 {
4394 	struct bufobj *dropobj;
4395 	void    (*biodone)(struct buf *);
4396 
4397 	buf_track(bp, __func__);
4398 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4399 	dropobj = NULL;
4400 
4401 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4402 
4403 	runningbufwakeup(bp);
4404 	if (bp->b_iocmd == BIO_WRITE)
4405 		dropobj = bp->b_bufobj;
4406 	/* call optional completion function if requested */
4407 	if (bp->b_iodone != NULL) {
4408 		biodone = bp->b_iodone;
4409 		bp->b_iodone = NULL;
4410 		(*biodone) (bp);
4411 		if (dropobj)
4412 			bufobj_wdrop(dropobj);
4413 		return;
4414 	}
4415 	if (bp->b_flags & B_VMIO) {
4416 		/*
4417 		 * Set B_CACHE if the op was a normal read and no error
4418 		 * occurred.  B_CACHE is set for writes in the b*write()
4419 		 * routines.
4420 		 */
4421 		if (bp->b_iocmd == BIO_READ &&
4422 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4423 		    !(bp->b_ioflags & BIO_ERROR))
4424 			bp->b_flags |= B_CACHE;
4425 		vfs_vmio_iodone(bp);
4426 	}
4427 	if (!LIST_EMPTY(&bp->b_dep))
4428 		buf_complete(bp);
4429 	if ((bp->b_flags & B_CKHASH) != 0) {
4430 		KASSERT(bp->b_iocmd == BIO_READ,
4431 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4432 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4433 		(*bp->b_ckhashcalc)(bp);
4434 	}
4435 	/*
4436 	 * For asynchronous completions, release the buffer now. The brelse
4437 	 * will do a wakeup there if necessary - so no need to do a wakeup
4438 	 * here in the async case. The sync case always needs to do a wakeup.
4439 	 */
4440 	if (bp->b_flags & B_ASYNC) {
4441 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4442 		    (bp->b_ioflags & BIO_ERROR))
4443 			brelse(bp);
4444 		else
4445 			bqrelse(bp);
4446 	} else
4447 		bdone(bp);
4448 	if (dropobj)
4449 		bufobj_wdrop(dropobj);
4450 }
4451 
4452 /*
4453  * This routine is called in lieu of iodone in the case of
4454  * incomplete I/O.  This keeps the busy status for pages
4455  * consistent.
4456  */
4457 void
4458 vfs_unbusy_pages(struct buf *bp)
4459 {
4460 	int i;
4461 	vm_object_t obj;
4462 	vm_page_t m;
4463 	bool bogus;
4464 
4465 	runningbufwakeup(bp);
4466 	if (!(bp->b_flags & B_VMIO))
4467 		return;
4468 
4469 	obj = bp->b_bufobj->bo_object;
4470 	bogus = false;
4471 	for (i = 0; i < bp->b_npages; i++) {
4472 		m = bp->b_pages[i];
4473 		if (m == bogus_page) {
4474 			if (bogus == false) {
4475 				bogus = true;
4476 				VM_OBJECT_RLOCK(obj);
4477 			}
4478 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4479 			if (!m)
4480 				panic("vfs_unbusy_pages: page missing\n");
4481 			bp->b_pages[i] = m;
4482 			if (buf_mapped(bp)) {
4483 				BUF_CHECK_MAPPED(bp);
4484 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4485 				    bp->b_pages, bp->b_npages);
4486 			} else
4487 				BUF_CHECK_UNMAPPED(bp);
4488 		}
4489 		vm_page_sunbusy(m);
4490 	}
4491 	if (bogus)
4492 		VM_OBJECT_RUNLOCK(obj);
4493 	vm_object_pip_wakeupn(obj, bp->b_npages);
4494 }
4495 
4496 /*
4497  * vfs_page_set_valid:
4498  *
4499  *	Set the valid bits in a page based on the supplied offset.   The
4500  *	range is restricted to the buffer's size.
4501  *
4502  *	This routine is typically called after a read completes.
4503  */
4504 static void
4505 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4506 {
4507 	vm_ooffset_t eoff;
4508 
4509 	/*
4510 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4511 	 * page boundary and eoff is not greater than the end of the buffer.
4512 	 * The end of the buffer, in this case, is our file EOF, not the
4513 	 * allocation size of the buffer.
4514 	 */
4515 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4516 	if (eoff > bp->b_offset + bp->b_bcount)
4517 		eoff = bp->b_offset + bp->b_bcount;
4518 
4519 	/*
4520 	 * Set valid range.  This is typically the entire buffer and thus the
4521 	 * entire page.
4522 	 */
4523 	if (eoff > off)
4524 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4525 }
4526 
4527 /*
4528  * vfs_page_set_validclean:
4529  *
4530  *	Set the valid bits and clear the dirty bits in a page based on the
4531  *	supplied offset.   The range is restricted to the buffer's size.
4532  */
4533 static void
4534 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4535 {
4536 	vm_ooffset_t soff, eoff;
4537 
4538 	/*
4539 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4540 	 * page boundary or cross the end of the buffer.  The end of the
4541 	 * buffer, in this case, is our file EOF, not the allocation size
4542 	 * of the buffer.
4543 	 */
4544 	soff = off;
4545 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4546 	if (eoff > bp->b_offset + bp->b_bcount)
4547 		eoff = bp->b_offset + bp->b_bcount;
4548 
4549 	/*
4550 	 * Set valid range.  This is typically the entire buffer and thus the
4551 	 * entire page.
4552 	 */
4553 	if (eoff > soff) {
4554 		vm_page_set_validclean(
4555 		    m,
4556 		   (vm_offset_t) (soff & PAGE_MASK),
4557 		   (vm_offset_t) (eoff - soff)
4558 		);
4559 	}
4560 }
4561 
4562 /*
4563  * Acquire a shared busy on all pages in the buf.
4564  */
4565 void
4566 vfs_busy_pages_acquire(struct buf *bp)
4567 {
4568 	int i;
4569 
4570 	for (i = 0; i < bp->b_npages; i++)
4571 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4572 }
4573 
4574 void
4575 vfs_busy_pages_release(struct buf *bp)
4576 {
4577 	int i;
4578 
4579 	for (i = 0; i < bp->b_npages; i++)
4580 		vm_page_sunbusy(bp->b_pages[i]);
4581 }
4582 
4583 /*
4584  * This routine is called before a device strategy routine.
4585  * It is used to tell the VM system that paging I/O is in
4586  * progress, and treat the pages associated with the buffer
4587  * almost as being exclusive busy.  Also the object paging_in_progress
4588  * flag is handled to make sure that the object doesn't become
4589  * inconsistent.
4590  *
4591  * Since I/O has not been initiated yet, certain buffer flags
4592  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4593  * and should be ignored.
4594  */
4595 void
4596 vfs_busy_pages(struct buf *bp, int clear_modify)
4597 {
4598 	vm_object_t obj;
4599 	vm_ooffset_t foff;
4600 	vm_page_t m;
4601 	int i;
4602 	bool bogus;
4603 
4604 	if (!(bp->b_flags & B_VMIO))
4605 		return;
4606 
4607 	obj = bp->b_bufobj->bo_object;
4608 	foff = bp->b_offset;
4609 	KASSERT(bp->b_offset != NOOFFSET,
4610 	    ("vfs_busy_pages: no buffer offset"));
4611 	if ((bp->b_flags & B_CLUSTER) == 0) {
4612 		vm_object_pip_add(obj, bp->b_npages);
4613 		vfs_busy_pages_acquire(bp);
4614 	}
4615 	if (bp->b_bufsize != 0)
4616 		vfs_setdirty_range(bp);
4617 	bogus = false;
4618 	for (i = 0; i < bp->b_npages; i++) {
4619 		m = bp->b_pages[i];
4620 		vm_page_assert_sbusied(m);
4621 
4622 		/*
4623 		 * When readying a buffer for a read ( i.e
4624 		 * clear_modify == 0 ), it is important to do
4625 		 * bogus_page replacement for valid pages in
4626 		 * partially instantiated buffers.  Partially
4627 		 * instantiated buffers can, in turn, occur when
4628 		 * reconstituting a buffer from its VM backing store
4629 		 * base.  We only have to do this if B_CACHE is
4630 		 * clear ( which causes the I/O to occur in the
4631 		 * first place ).  The replacement prevents the read
4632 		 * I/O from overwriting potentially dirty VM-backed
4633 		 * pages.  XXX bogus page replacement is, uh, bogus.
4634 		 * It may not work properly with small-block devices.
4635 		 * We need to find a better way.
4636 		 */
4637 		if (clear_modify) {
4638 			pmap_remove_write(m);
4639 			vfs_page_set_validclean(bp, foff, m);
4640 		} else if (vm_page_all_valid(m) &&
4641 		    (bp->b_flags & B_CACHE) == 0) {
4642 			bp->b_pages[i] = bogus_page;
4643 			bogus = true;
4644 		}
4645 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4646 	}
4647 	if (bogus && buf_mapped(bp)) {
4648 		BUF_CHECK_MAPPED(bp);
4649 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4650 		    bp->b_pages, bp->b_npages);
4651 	}
4652 }
4653 
4654 /*
4655  *	vfs_bio_set_valid:
4656  *
4657  *	Set the range within the buffer to valid.  The range is
4658  *	relative to the beginning of the buffer, b_offset.  Note that
4659  *	b_offset itself may be offset from the beginning of the first
4660  *	page.
4661  */
4662 void
4663 vfs_bio_set_valid(struct buf *bp, int base, int size)
4664 {
4665 	int i, n;
4666 	vm_page_t m;
4667 
4668 	if (!(bp->b_flags & B_VMIO))
4669 		return;
4670 
4671 	/*
4672 	 * Fixup base to be relative to beginning of first page.
4673 	 * Set initial n to be the maximum number of bytes in the
4674 	 * first page that can be validated.
4675 	 */
4676 	base += (bp->b_offset & PAGE_MASK);
4677 	n = PAGE_SIZE - (base & PAGE_MASK);
4678 
4679 	/*
4680 	 * Busy may not be strictly necessary here because the pages are
4681 	 * unlikely to be fully valid and the vnode lock will synchronize
4682 	 * their access via getpages.  It is grabbed for consistency with
4683 	 * other page validation.
4684 	 */
4685 	vfs_busy_pages_acquire(bp);
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 	vfs_busy_pages_release(bp);
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 	vfs_busy_pages_acquire(bp);
4722 	sa = bp->b_offset & PAGE_MASK;
4723 	slide = 0;
4724 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4725 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4726 		ea = slide & PAGE_MASK;
4727 		if (ea == 0)
4728 			ea = PAGE_SIZE;
4729 		if (bp->b_pages[i] == bogus_page)
4730 			continue;
4731 		j = sa / DEV_BSIZE;
4732 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4733 		if ((bp->b_pages[i]->valid & mask) == mask)
4734 			continue;
4735 		if ((bp->b_pages[i]->valid & mask) == 0)
4736 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4737 		else {
4738 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4739 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4740 					pmap_zero_page_area(bp->b_pages[i],
4741 					    sa, DEV_BSIZE);
4742 				}
4743 			}
4744 		}
4745 		vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4746 		    roundup2(ea - sa, DEV_BSIZE));
4747 	}
4748 	vfs_busy_pages_release(bp);
4749 	bp->b_resid = 0;
4750 }
4751 
4752 void
4753 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4754 {
4755 	vm_page_t m;
4756 	int i, n;
4757 
4758 	if (buf_mapped(bp)) {
4759 		BUF_CHECK_MAPPED(bp);
4760 		bzero(bp->b_data + base, size);
4761 	} else {
4762 		BUF_CHECK_UNMAPPED(bp);
4763 		n = PAGE_SIZE - (base & PAGE_MASK);
4764 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4765 			m = bp->b_pages[i];
4766 			if (n > size)
4767 				n = size;
4768 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4769 			base += n;
4770 			size -= n;
4771 			n = PAGE_SIZE;
4772 		}
4773 	}
4774 }
4775 
4776 /*
4777  * Update buffer flags based on I/O request parameters, optionally releasing the
4778  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4779  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4780  * I/O).  Otherwise the buffer is released to the cache.
4781  */
4782 static void
4783 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4784 {
4785 
4786 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4787 	    ("buf %p non-VMIO noreuse", bp));
4788 
4789 	if ((ioflag & IO_DIRECT) != 0)
4790 		bp->b_flags |= B_DIRECT;
4791 	if ((ioflag & IO_EXT) != 0)
4792 		bp->b_xflags |= BX_ALTDATA;
4793 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4794 		bp->b_flags |= B_RELBUF;
4795 		if ((ioflag & IO_NOREUSE) != 0)
4796 			bp->b_flags |= B_NOREUSE;
4797 		if (release)
4798 			brelse(bp);
4799 	} else if (release)
4800 		bqrelse(bp);
4801 }
4802 
4803 void
4804 vfs_bio_brelse(struct buf *bp, int ioflag)
4805 {
4806 
4807 	b_io_dismiss(bp, ioflag, true);
4808 }
4809 
4810 void
4811 vfs_bio_set_flags(struct buf *bp, int ioflag)
4812 {
4813 
4814 	b_io_dismiss(bp, ioflag, false);
4815 }
4816 
4817 /*
4818  * vm_hold_load_pages and vm_hold_free_pages get pages into
4819  * a buffers address space.  The pages are anonymous and are
4820  * not associated with a file object.
4821  */
4822 static void
4823 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4824 {
4825 	vm_offset_t pg;
4826 	vm_page_t p;
4827 	int index;
4828 
4829 	BUF_CHECK_MAPPED(bp);
4830 
4831 	to = round_page(to);
4832 	from = round_page(from);
4833 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4834 
4835 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4836 		/*
4837 		 * note: must allocate system pages since blocking here
4838 		 * could interfere with paging I/O, no matter which
4839 		 * process we are.
4840 		 */
4841 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4842 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4843 		    VM_ALLOC_WAITOK);
4844 		pmap_qenter(pg, &p, 1);
4845 		bp->b_pages[index] = p;
4846 	}
4847 	bp->b_npages = index;
4848 }
4849 
4850 /* Return pages associated with this buf to the vm system */
4851 static void
4852 vm_hold_free_pages(struct buf *bp, int newbsize)
4853 {
4854 	vm_offset_t from;
4855 	vm_page_t p;
4856 	int index, newnpages;
4857 
4858 	BUF_CHECK_MAPPED(bp);
4859 
4860 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4861 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4862 	if (bp->b_npages > newnpages)
4863 		pmap_qremove(from, bp->b_npages - newnpages);
4864 	for (index = newnpages; index < bp->b_npages; index++) {
4865 		p = bp->b_pages[index];
4866 		bp->b_pages[index] = NULL;
4867 		vm_page_unwire_noq(p);
4868 		vm_page_free(p);
4869 	}
4870 	bp->b_npages = newnpages;
4871 }
4872 
4873 /*
4874  * Map an IO request into kernel virtual address space.
4875  *
4876  * All requests are (re)mapped into kernel VA space.
4877  * Notice that we use b_bufsize for the size of the buffer
4878  * to be mapped.  b_bcount might be modified by the driver.
4879  *
4880  * Note that even if the caller determines that the address space should
4881  * be valid, a race or a smaller-file mapped into a larger space may
4882  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4883  * check the return value.
4884  *
4885  * This function only works with pager buffers.
4886  */
4887 int
4888 vmapbuf(struct buf *bp, int mapbuf)
4889 {
4890 	vm_prot_t prot;
4891 	int pidx;
4892 
4893 	if (bp->b_bufsize < 0)
4894 		return (-1);
4895 	prot = VM_PROT_READ;
4896 	if (bp->b_iocmd == BIO_READ)
4897 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4898 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4899 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4900 	    btoc(MAXPHYS))) < 0)
4901 		return (-1);
4902 	bp->b_npages = pidx;
4903 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4904 	if (mapbuf || !unmapped_buf_allowed) {
4905 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4906 		bp->b_data = bp->b_kvabase + bp->b_offset;
4907 	} else
4908 		bp->b_data = unmapped_buf;
4909 	return(0);
4910 }
4911 
4912 /*
4913  * Free the io map PTEs associated with this IO operation.
4914  * We also invalidate the TLB entries and restore the original b_addr.
4915  *
4916  * This function only works with pager buffers.
4917  */
4918 void
4919 vunmapbuf(struct buf *bp)
4920 {
4921 	int npages;
4922 
4923 	npages = bp->b_npages;
4924 	if (buf_mapped(bp))
4925 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4926 	vm_page_unhold_pages(bp->b_pages, npages);
4927 
4928 	bp->b_data = unmapped_buf;
4929 }
4930 
4931 void
4932 bdone(struct buf *bp)
4933 {
4934 	struct mtx *mtxp;
4935 
4936 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4937 	mtx_lock(mtxp);
4938 	bp->b_flags |= B_DONE;
4939 	wakeup(bp);
4940 	mtx_unlock(mtxp);
4941 }
4942 
4943 void
4944 bwait(struct buf *bp, u_char pri, const char *wchan)
4945 {
4946 	struct mtx *mtxp;
4947 
4948 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4949 	mtx_lock(mtxp);
4950 	while ((bp->b_flags & B_DONE) == 0)
4951 		msleep(bp, mtxp, pri, wchan, 0);
4952 	mtx_unlock(mtxp);
4953 }
4954 
4955 int
4956 bufsync(struct bufobj *bo, int waitfor)
4957 {
4958 
4959 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
4960 }
4961 
4962 void
4963 bufstrategy(struct bufobj *bo, struct buf *bp)
4964 {
4965 	int i __unused;
4966 	struct vnode *vp;
4967 
4968 	vp = bp->b_vp;
4969 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4970 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4971 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4972 	i = VOP_STRATEGY(vp, bp);
4973 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4974 }
4975 
4976 /*
4977  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
4978  */
4979 void
4980 bufobj_init(struct bufobj *bo, void *private)
4981 {
4982 	static volatile int bufobj_cleanq;
4983 
4984         bo->bo_domain =
4985             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
4986         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
4987         bo->bo_private = private;
4988         TAILQ_INIT(&bo->bo_clean.bv_hd);
4989         TAILQ_INIT(&bo->bo_dirty.bv_hd);
4990 }
4991 
4992 void
4993 bufobj_wrefl(struct bufobj *bo)
4994 {
4995 
4996 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4997 	ASSERT_BO_WLOCKED(bo);
4998 	bo->bo_numoutput++;
4999 }
5000 
5001 void
5002 bufobj_wref(struct bufobj *bo)
5003 {
5004 
5005 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5006 	BO_LOCK(bo);
5007 	bo->bo_numoutput++;
5008 	BO_UNLOCK(bo);
5009 }
5010 
5011 void
5012 bufobj_wdrop(struct bufobj *bo)
5013 {
5014 
5015 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5016 	BO_LOCK(bo);
5017 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5018 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5019 		bo->bo_flag &= ~BO_WWAIT;
5020 		wakeup(&bo->bo_numoutput);
5021 	}
5022 	BO_UNLOCK(bo);
5023 }
5024 
5025 int
5026 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5027 {
5028 	int error;
5029 
5030 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5031 	ASSERT_BO_WLOCKED(bo);
5032 	error = 0;
5033 	while (bo->bo_numoutput) {
5034 		bo->bo_flag |= BO_WWAIT;
5035 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5036 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5037 		if (error)
5038 			break;
5039 	}
5040 	return (error);
5041 }
5042 
5043 /*
5044  * Set bio_data or bio_ma for struct bio from the struct buf.
5045  */
5046 void
5047 bdata2bio(struct buf *bp, struct bio *bip)
5048 {
5049 
5050 	if (!buf_mapped(bp)) {
5051 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5052 		bip->bio_ma = bp->b_pages;
5053 		bip->bio_ma_n = bp->b_npages;
5054 		bip->bio_data = unmapped_buf;
5055 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5056 		bip->bio_flags |= BIO_UNMAPPED;
5057 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5058 		    PAGE_SIZE == bp->b_npages,
5059 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5060 		    (long long)bip->bio_length, bip->bio_ma_n));
5061 	} else {
5062 		bip->bio_data = bp->b_data;
5063 		bip->bio_ma = NULL;
5064 	}
5065 }
5066 
5067 /*
5068  * The MIPS pmap code currently doesn't handle aliased pages.
5069  * The VIPT caches may not handle page aliasing themselves, leading
5070  * to data corruption.
5071  *
5072  * As such, this code makes a system extremely unhappy if said
5073  * system doesn't support unaliasing the above situation in hardware.
5074  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5075  * this feature at build time, so it has to be handled in software.
5076  *
5077  * Once the MIPS pmap/cache code grows to support this function on
5078  * earlier chips, it should be flipped back off.
5079  */
5080 #ifdef	__mips__
5081 static int buf_pager_relbuf = 1;
5082 #else
5083 static int buf_pager_relbuf = 0;
5084 #endif
5085 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5086     &buf_pager_relbuf, 0,
5087     "Make buffer pager release buffers after reading");
5088 
5089 /*
5090  * The buffer pager.  It uses buffer reads to validate pages.
5091  *
5092  * In contrast to the generic local pager from vm/vnode_pager.c, this
5093  * pager correctly and easily handles volumes where the underlying
5094  * device block size is greater than the machine page size.  The
5095  * buffer cache transparently extends the requested page run to be
5096  * aligned at the block boundary, and does the necessary bogus page
5097  * replacements in the addends to avoid obliterating already valid
5098  * pages.
5099  *
5100  * The only non-trivial issue is that the exclusive busy state for
5101  * pages, which is assumed by the vm_pager_getpages() interface, is
5102  * incompatible with the VMIO buffer cache's desire to share-busy the
5103  * pages.  This function performs a trivial downgrade of the pages'
5104  * state before reading buffers, and a less trivial upgrade from the
5105  * shared-busy to excl-busy state after the read.
5106  */
5107 int
5108 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5109     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5110     vbg_get_blksize_t get_blksize)
5111 {
5112 	vm_page_t m;
5113 	vm_object_t object;
5114 	struct buf *bp;
5115 	struct mount *mp;
5116 	daddr_t lbn, lbnp;
5117 	vm_ooffset_t la, lb, poff, poffe;
5118 	long bsize;
5119 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5120 	bool redo, lpart;
5121 
5122 	object = vp->v_object;
5123 	mp = vp->v_mount;
5124 	error = 0;
5125 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5126 	if (la >= object->un_pager.vnp.vnp_size)
5127 		return (VM_PAGER_BAD);
5128 
5129 	/*
5130 	 * Change the meaning of la from where the last requested page starts
5131 	 * to where it ends, because that's the end of the requested region
5132 	 * and the start of the potential read-ahead region.
5133 	 */
5134 	la += PAGE_SIZE;
5135 	lpart = la > object->un_pager.vnp.vnp_size;
5136 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5137 
5138 	/*
5139 	 * Calculate read-ahead, behind and total pages.
5140 	 */
5141 	pgsin = count;
5142 	lb = IDX_TO_OFF(ma[0]->pindex);
5143 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5144 	pgsin += pgsin_b;
5145 	if (rbehind != NULL)
5146 		*rbehind = pgsin_b;
5147 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5148 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5149 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5150 		    PAGE_SIZE) - la);
5151 	pgsin += pgsin_a;
5152 	if (rahead != NULL)
5153 		*rahead = pgsin_a;
5154 	VM_CNT_INC(v_vnodein);
5155 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5156 
5157 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5158 	    != 0) ? GB_UNMAPPED : 0;
5159 again:
5160 	for (i = 0; i < count; i++)
5161 		vm_page_busy_downgrade(ma[i]);
5162 
5163 	lbnp = -1;
5164 	for (i = 0; i < count; i++) {
5165 		m = ma[i];
5166 
5167 		/*
5168 		 * Pages are shared busy and the object lock is not
5169 		 * owned, which together allow for the pages'
5170 		 * invalidation.  The racy test for validity avoids
5171 		 * useless creation of the buffer for the most typical
5172 		 * case when invalidation is not used in redo or for
5173 		 * parallel read.  The shared->excl upgrade loop at
5174 		 * the end of the function catches the race in a
5175 		 * reliable way (protected by the object lock).
5176 		 */
5177 		if (vm_page_all_valid(m))
5178 			continue;
5179 
5180 		poff = IDX_TO_OFF(m->pindex);
5181 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5182 		for (; poff < poffe; poff += bsize) {
5183 			lbn = get_lblkno(vp, poff);
5184 			if (lbn == lbnp)
5185 				goto next_page;
5186 			lbnp = lbn;
5187 
5188 			bsize = get_blksize(vp, lbn);
5189 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5190 			    br_flags, &bp);
5191 			if (error != 0)
5192 				goto end_pages;
5193 			if (LIST_EMPTY(&bp->b_dep)) {
5194 				/*
5195 				 * Invalidation clears m->valid, but
5196 				 * may leave B_CACHE flag if the
5197 				 * buffer existed at the invalidation
5198 				 * time.  In this case, recycle the
5199 				 * buffer to do real read on next
5200 				 * bread() after redo.
5201 				 *
5202 				 * Otherwise B_RELBUF is not strictly
5203 				 * necessary, enable to reduce buf
5204 				 * cache pressure.
5205 				 */
5206 				if (buf_pager_relbuf ||
5207 				    !vm_page_all_valid(m))
5208 					bp->b_flags |= B_RELBUF;
5209 
5210 				bp->b_flags &= ~B_NOCACHE;
5211 				brelse(bp);
5212 			} else {
5213 				bqrelse(bp);
5214 			}
5215 		}
5216 		KASSERT(1 /* racy, enable for debugging */ ||
5217 		    vm_page_all_valid(m) || i == count - 1,
5218 		    ("buf %d %p invalid", i, m));
5219 		if (i == count - 1 && lpart) {
5220 			if (!vm_page_none_valid(m) &&
5221 			    !vm_page_all_valid(m))
5222 				vm_page_zero_invalid(m, TRUE);
5223 		}
5224 next_page:;
5225 	}
5226 end_pages:
5227 
5228 	VM_OBJECT_WLOCK(object);
5229 	redo = false;
5230 	for (i = 0; i < count; i++) {
5231 		vm_page_sunbusy(ma[i]);
5232 		ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
5233 
5234 		/*
5235 		 * Since the pages were only sbusy while neither the
5236 		 * buffer nor the object lock was held by us, or
5237 		 * reallocated while vm_page_grab() slept for busy
5238 		 * relinguish, they could have been invalidated.
5239 		 * Recheck the valid bits and re-read as needed.
5240 		 *
5241 		 * Note that the last page is made fully valid in the
5242 		 * read loop, and partial validity for the page at
5243 		 * index count - 1 could mean that the page was
5244 		 * invalidated or removed, so we must restart for
5245 		 * safety as well.
5246 		 */
5247 		if (!vm_page_all_valid(ma[i]))
5248 			redo = true;
5249 	}
5250 	VM_OBJECT_WUNLOCK(object);
5251 	if (redo && error == 0)
5252 		goto again;
5253 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5254 }
5255 
5256 #include "opt_ddb.h"
5257 #ifdef DDB
5258 #include <ddb/ddb.h>
5259 
5260 /* DDB command to show buffer data */
5261 DB_SHOW_COMMAND(buffer, db_show_buffer)
5262 {
5263 	/* get args */
5264 	struct buf *bp = (struct buf *)addr;
5265 #ifdef FULL_BUF_TRACKING
5266 	uint32_t i, j;
5267 #endif
5268 
5269 	if (!have_addr) {
5270 		db_printf("usage: show buffer <addr>\n");
5271 		return;
5272 	}
5273 
5274 	db_printf("buf at %p\n", bp);
5275 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5276 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5277 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5278 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5279 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5280 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5281 	db_printf(
5282 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5283 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5284 	    "b_vp = %p, b_dep = %p\n",
5285 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5286 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5287 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5288 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5289 	    bp->b_kvabase, bp->b_kvasize);
5290 	if (bp->b_npages) {
5291 		int i;
5292 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5293 		for (i = 0; i < bp->b_npages; i++) {
5294 			vm_page_t m;
5295 			m = bp->b_pages[i];
5296 			if (m != NULL)
5297 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5298 				    (u_long)m->pindex,
5299 				    (u_long)VM_PAGE_TO_PHYS(m));
5300 			else
5301 				db_printf("( ??? )");
5302 			if ((i + 1) < bp->b_npages)
5303 				db_printf(",");
5304 		}
5305 		db_printf("\n");
5306 	}
5307 	BUF_LOCKPRINTINFO(bp);
5308 #if defined(FULL_BUF_TRACKING)
5309 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5310 
5311 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5312 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5313 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5314 			continue;
5315 		db_printf(" %2u: %s\n", j,
5316 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5317 	}
5318 #elif defined(BUF_TRACKING)
5319 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5320 #endif
5321 	db_printf(" ");
5322 }
5323 
5324 DB_SHOW_COMMAND(bufqueues, bufqueues)
5325 {
5326 	struct bufdomain *bd;
5327 	struct buf *bp;
5328 	long total;
5329 	int i, j, cnt;
5330 
5331 	db_printf("bqempty: %d\n", bqempty.bq_len);
5332 
5333 	for (i = 0; i < buf_domains; i++) {
5334 		bd = &bdomain[i];
5335 		db_printf("Buf domain %d\n", i);
5336 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5337 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5338 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5339 		db_printf("\n");
5340 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5341 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5342 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5343 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5344 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5345 		db_printf("\n");
5346 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5347 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5348 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5349 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5350 		db_printf("\n");
5351 		total = 0;
5352 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5353 			total += bp->b_bufsize;
5354 		db_printf("\tcleanq count\t%d (%ld)\n",
5355 		    bd->bd_cleanq->bq_len, total);
5356 		total = 0;
5357 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5358 			total += bp->b_bufsize;
5359 		db_printf("\tdirtyq count\t%d (%ld)\n",
5360 		    bd->bd_dirtyq.bq_len, total);
5361 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5362 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5363 		db_printf("\tCPU ");
5364 		for (j = 0; j <= mp_maxid; j++)
5365 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5366 		db_printf("\n");
5367 		cnt = 0;
5368 		total = 0;
5369 		for (j = 0; j < nbuf; j++)
5370 			if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5371 				cnt++;
5372 				total += buf[j].b_bufsize;
5373 			}
5374 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5375 		cnt = 0;
5376 		total = 0;
5377 		for (j = 0; j < nbuf; j++)
5378 			if (buf[j].b_domain == i) {
5379 				cnt++;
5380 				total += buf[j].b_bufsize;
5381 			}
5382 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5383 	}
5384 }
5385 
5386 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5387 {
5388 	struct buf *bp;
5389 	int i;
5390 
5391 	for (i = 0; i < nbuf; i++) {
5392 		bp = &buf[i];
5393 		if (BUF_ISLOCKED(bp)) {
5394 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5395 			db_printf("\n");
5396 			if (db_pager_quit)
5397 				break;
5398 		}
5399 	}
5400 }
5401 
5402 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5403 {
5404 	struct vnode *vp;
5405 	struct buf *bp;
5406 
5407 	if (!have_addr) {
5408 		db_printf("usage: show vnodebufs <addr>\n");
5409 		return;
5410 	}
5411 	vp = (struct vnode *)addr;
5412 	db_printf("Clean buffers:\n");
5413 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5414 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5415 		db_printf("\n");
5416 	}
5417 	db_printf("Dirty buffers:\n");
5418 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5419 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5420 		db_printf("\n");
5421 	}
5422 }
5423 
5424 DB_COMMAND(countfreebufs, db_coundfreebufs)
5425 {
5426 	struct buf *bp;
5427 	int i, used = 0, nfree = 0;
5428 
5429 	if (have_addr) {
5430 		db_printf("usage: countfreebufs\n");
5431 		return;
5432 	}
5433 
5434 	for (i = 0; i < nbuf; i++) {
5435 		bp = &buf[i];
5436 		if (bp->b_qindex == QUEUE_EMPTY)
5437 			nfree++;
5438 		else
5439 			used++;
5440 	}
5441 
5442 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5443 	    nfree + used);
5444 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5445 }
5446 #endif /* DDB */
5447