xref: /freebsd/sys/kern/vfs_bio.c (revision d00066a5f923073113bcb0213ca7a0a9a29b9561)
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  * The blkno parameter is the logical block being requested. Normally
2131  * the mapping of logical block number to disk block address is done
2132  * by calling VOP_BMAP(). However, if the mapping is already known, the
2133  * disk block address can be passed using the dblkno parameter. If the
2134  * disk block address is not known, then the same value should be passed
2135  * for blkno and dblkno.
2136  */
2137 int
2138 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2139     daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2140     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2141 {
2142 	struct buf *bp;
2143 	struct thread *td;
2144 	int error, readwait, rv;
2145 
2146 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2147 	td = curthread;
2148 	/*
2149 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2150 	 * are specified.
2151 	 */
2152 	error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2153 	if (error != 0) {
2154 		*bpp = NULL;
2155 		return (error);
2156 	}
2157 	KASSERT(blkno == bp->b_lblkno,
2158 	    ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2159 	    (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2160 	flags &= ~GB_NOSPARSE;
2161 	*bpp = bp;
2162 
2163 	/*
2164 	 * If not found in cache, do some I/O
2165 	 */
2166 	readwait = 0;
2167 	if ((bp->b_flags & B_CACHE) == 0) {
2168 #ifdef RACCT
2169 		if (racct_enable) {
2170 			PROC_LOCK(td->td_proc);
2171 			racct_add_buf(td->td_proc, bp, 0);
2172 			PROC_UNLOCK(td->td_proc);
2173 		}
2174 #endif /* RACCT */
2175 		td->td_ru.ru_inblock++;
2176 		bp->b_iocmd = BIO_READ;
2177 		bp->b_flags &= ~B_INVAL;
2178 		if ((flags & GB_CKHASH) != 0) {
2179 			bp->b_flags |= B_CKHASH;
2180 			bp->b_ckhashcalc = ckhashfunc;
2181 		}
2182 		bp->b_ioflags &= ~BIO_ERROR;
2183 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2184 			bp->b_rcred = crhold(cred);
2185 		vfs_busy_pages(bp, 0);
2186 		bp->b_iooffset = dbtob(bp->b_blkno);
2187 		bstrategy(bp);
2188 		++readwait;
2189 	}
2190 
2191 	/*
2192 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2193 	 */
2194 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2195 
2196 	rv = 0;
2197 	if (readwait) {
2198 		rv = bufwait(bp);
2199 		if (rv != 0) {
2200 			brelse(bp);
2201 			*bpp = NULL;
2202 		}
2203 	}
2204 	return (rv);
2205 }
2206 
2207 /*
2208  * Write, release buffer on completion.  (Done by iodone
2209  * if async).  Do not bother writing anything if the buffer
2210  * is invalid.
2211  *
2212  * Note that we set B_CACHE here, indicating that buffer is
2213  * fully valid and thus cacheable.  This is true even of NFS
2214  * now so we set it generally.  This could be set either here
2215  * or in biodone() since the I/O is synchronous.  We put it
2216  * here.
2217  */
2218 int
2219 bufwrite(struct buf *bp)
2220 {
2221 	int oldflags;
2222 	struct vnode *vp;
2223 	long space;
2224 	int vp_md;
2225 
2226 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2227 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2228 		bp->b_flags |= B_INVAL | B_RELBUF;
2229 		bp->b_flags &= ~B_CACHE;
2230 		brelse(bp);
2231 		return (ENXIO);
2232 	}
2233 	if (bp->b_flags & B_INVAL) {
2234 		brelse(bp);
2235 		return (0);
2236 	}
2237 
2238 	if (bp->b_flags & B_BARRIER)
2239 		atomic_add_long(&barrierwrites, 1);
2240 
2241 	oldflags = bp->b_flags;
2242 
2243 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2244 	    ("FFS background buffer should not get here %p", bp));
2245 
2246 	vp = bp->b_vp;
2247 	if (vp)
2248 		vp_md = vp->v_vflag & VV_MD;
2249 	else
2250 		vp_md = 0;
2251 
2252 	/*
2253 	 * Mark the buffer clean.  Increment the bufobj write count
2254 	 * before bundirty() call, to prevent other thread from seeing
2255 	 * empty dirty list and zero counter for writes in progress,
2256 	 * falsely indicating that the bufobj is clean.
2257 	 */
2258 	bufobj_wref(bp->b_bufobj);
2259 	bundirty(bp);
2260 
2261 	bp->b_flags &= ~B_DONE;
2262 	bp->b_ioflags &= ~BIO_ERROR;
2263 	bp->b_flags |= B_CACHE;
2264 	bp->b_iocmd = BIO_WRITE;
2265 
2266 	vfs_busy_pages(bp, 1);
2267 
2268 	/*
2269 	 * Normal bwrites pipeline writes
2270 	 */
2271 	bp->b_runningbufspace = bp->b_bufsize;
2272 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2273 
2274 #ifdef RACCT
2275 	if (racct_enable) {
2276 		PROC_LOCK(curproc);
2277 		racct_add_buf(curproc, bp, 1);
2278 		PROC_UNLOCK(curproc);
2279 	}
2280 #endif /* RACCT */
2281 	curthread->td_ru.ru_oublock++;
2282 	if (oldflags & B_ASYNC)
2283 		BUF_KERNPROC(bp);
2284 	bp->b_iooffset = dbtob(bp->b_blkno);
2285 	buf_track(bp, __func__);
2286 	bstrategy(bp);
2287 
2288 	if ((oldflags & B_ASYNC) == 0) {
2289 		int rtval = bufwait(bp);
2290 		brelse(bp);
2291 		return (rtval);
2292 	} else if (space > hirunningspace) {
2293 		/*
2294 		 * don't allow the async write to saturate the I/O
2295 		 * system.  We will not deadlock here because
2296 		 * we are blocking waiting for I/O that is already in-progress
2297 		 * to complete. We do not block here if it is the update
2298 		 * or syncer daemon trying to clean up as that can lead
2299 		 * to deadlock.
2300 		 */
2301 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2302 			waitrunningbufspace();
2303 	}
2304 
2305 	return (0);
2306 }
2307 
2308 void
2309 bufbdflush(struct bufobj *bo, struct buf *bp)
2310 {
2311 	struct buf *nbp;
2312 
2313 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2314 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2315 		altbufferflushes++;
2316 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2317 		BO_LOCK(bo);
2318 		/*
2319 		 * Try to find a buffer to flush.
2320 		 */
2321 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2322 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2323 			    BUF_LOCK(nbp,
2324 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2325 				continue;
2326 			if (bp == nbp)
2327 				panic("bdwrite: found ourselves");
2328 			BO_UNLOCK(bo);
2329 			/* Don't countdeps with the bo lock held. */
2330 			if (buf_countdeps(nbp, 0)) {
2331 				BO_LOCK(bo);
2332 				BUF_UNLOCK(nbp);
2333 				continue;
2334 			}
2335 			if (nbp->b_flags & B_CLUSTEROK) {
2336 				vfs_bio_awrite(nbp);
2337 			} else {
2338 				bremfree(nbp);
2339 				bawrite(nbp);
2340 			}
2341 			dirtybufferflushes++;
2342 			break;
2343 		}
2344 		if (nbp == NULL)
2345 			BO_UNLOCK(bo);
2346 	}
2347 }
2348 
2349 /*
2350  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2351  * anything if the buffer is marked invalid.
2352  *
2353  * Note that since the buffer must be completely valid, we can safely
2354  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2355  * biodone() in order to prevent getblk from writing the buffer
2356  * out synchronously.
2357  */
2358 void
2359 bdwrite(struct buf *bp)
2360 {
2361 	struct thread *td = curthread;
2362 	struct vnode *vp;
2363 	struct bufobj *bo;
2364 
2365 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2366 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2367 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2368 	    ("Barrier request in delayed write %p", bp));
2369 
2370 	if (bp->b_flags & B_INVAL) {
2371 		brelse(bp);
2372 		return;
2373 	}
2374 
2375 	/*
2376 	 * If we have too many dirty buffers, don't create any more.
2377 	 * If we are wildly over our limit, then force a complete
2378 	 * cleanup. Otherwise, just keep the situation from getting
2379 	 * out of control. Note that we have to avoid a recursive
2380 	 * disaster and not try to clean up after our own cleanup!
2381 	 */
2382 	vp = bp->b_vp;
2383 	bo = bp->b_bufobj;
2384 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2385 		td->td_pflags |= TDP_INBDFLUSH;
2386 		BO_BDFLUSH(bo, bp);
2387 		td->td_pflags &= ~TDP_INBDFLUSH;
2388 	} else
2389 		recursiveflushes++;
2390 
2391 	bdirty(bp);
2392 	/*
2393 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2394 	 * true even of NFS now.
2395 	 */
2396 	bp->b_flags |= B_CACHE;
2397 
2398 	/*
2399 	 * This bmap keeps the system from needing to do the bmap later,
2400 	 * perhaps when the system is attempting to do a sync.  Since it
2401 	 * is likely that the indirect block -- or whatever other datastructure
2402 	 * that the filesystem needs is still in memory now, it is a good
2403 	 * thing to do this.  Note also, that if the pageout daemon is
2404 	 * requesting a sync -- there might not be enough memory to do
2405 	 * the bmap then...  So, this is important to do.
2406 	 */
2407 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2408 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2409 	}
2410 
2411 	buf_track(bp, __func__);
2412 
2413 	/*
2414 	 * Set the *dirty* buffer range based upon the VM system dirty
2415 	 * pages.
2416 	 *
2417 	 * Mark the buffer pages as clean.  We need to do this here to
2418 	 * satisfy the vnode_pager and the pageout daemon, so that it
2419 	 * thinks that the pages have been "cleaned".  Note that since
2420 	 * the pages are in a delayed write buffer -- the VFS layer
2421 	 * "will" see that the pages get written out on the next sync,
2422 	 * or perhaps the cluster will be completed.
2423 	 */
2424 	vfs_clean_pages_dirty_buf(bp);
2425 	bqrelse(bp);
2426 
2427 	/*
2428 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2429 	 * due to the softdep code.
2430 	 */
2431 }
2432 
2433 /*
2434  *	bdirty:
2435  *
2436  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2437  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2438  *	itself to properly update it in the dirty/clean lists.  We mark it
2439  *	B_DONE to ensure that any asynchronization of the buffer properly
2440  *	clears B_DONE ( else a panic will occur later ).
2441  *
2442  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2443  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2444  *	should only be called if the buffer is known-good.
2445  *
2446  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2447  *	count.
2448  *
2449  *	The buffer must be on QUEUE_NONE.
2450  */
2451 void
2452 bdirty(struct buf *bp)
2453 {
2454 
2455 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2456 	    bp, bp->b_vp, bp->b_flags);
2457 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2458 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2459 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2460 	bp->b_flags &= ~(B_RELBUF);
2461 	bp->b_iocmd = BIO_WRITE;
2462 
2463 	if ((bp->b_flags & B_DELWRI) == 0) {
2464 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2465 		reassignbuf(bp);
2466 		bdirtyadd(bp);
2467 	}
2468 }
2469 
2470 /*
2471  *	bundirty:
2472  *
2473  *	Clear B_DELWRI for buffer.
2474  *
2475  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2476  *	count.
2477  *
2478  *	The buffer must be on QUEUE_NONE.
2479  */
2480 
2481 void
2482 bundirty(struct buf *bp)
2483 {
2484 
2485 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2486 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2487 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2488 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2489 
2490 	if (bp->b_flags & B_DELWRI) {
2491 		bp->b_flags &= ~B_DELWRI;
2492 		reassignbuf(bp);
2493 		bdirtysub(bp);
2494 	}
2495 	/*
2496 	 * Since it is now being written, we can clear its deferred write flag.
2497 	 */
2498 	bp->b_flags &= ~B_DEFERRED;
2499 }
2500 
2501 /*
2502  *	bawrite:
2503  *
2504  *	Asynchronous write.  Start output on a buffer, but do not wait for
2505  *	it to complete.  The buffer is released when the output completes.
2506  *
2507  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2508  *	B_INVAL buffers.  Not us.
2509  */
2510 void
2511 bawrite(struct buf *bp)
2512 {
2513 
2514 	bp->b_flags |= B_ASYNC;
2515 	(void) bwrite(bp);
2516 }
2517 
2518 /*
2519  *	babarrierwrite:
2520  *
2521  *	Asynchronous barrier write.  Start output on a buffer, but do not
2522  *	wait for it to complete.  Place a write barrier after this write so
2523  *	that this buffer and all buffers written before it are committed to
2524  *	the disk before any buffers written after this write are committed
2525  *	to the disk.  The buffer is released when the output completes.
2526  */
2527 void
2528 babarrierwrite(struct buf *bp)
2529 {
2530 
2531 	bp->b_flags |= B_ASYNC | B_BARRIER;
2532 	(void) bwrite(bp);
2533 }
2534 
2535 /*
2536  *	bbarrierwrite:
2537  *
2538  *	Synchronous barrier write.  Start output on a buffer and wait for
2539  *	it to complete.  Place a write barrier after this write so that
2540  *	this buffer and all buffers written before it are committed to
2541  *	the disk before any buffers written after this write are committed
2542  *	to the disk.  The buffer is released when the output completes.
2543  */
2544 int
2545 bbarrierwrite(struct buf *bp)
2546 {
2547 
2548 	bp->b_flags |= B_BARRIER;
2549 	return (bwrite(bp));
2550 }
2551 
2552 /*
2553  *	bwillwrite:
2554  *
2555  *	Called prior to the locking of any vnodes when we are expecting to
2556  *	write.  We do not want to starve the buffer cache with too many
2557  *	dirty buffers so we block here.  By blocking prior to the locking
2558  *	of any vnodes we attempt to avoid the situation where a locked vnode
2559  *	prevents the various system daemons from flushing related buffers.
2560  */
2561 void
2562 bwillwrite(void)
2563 {
2564 
2565 	if (buf_dirty_count_severe()) {
2566 		mtx_lock(&bdirtylock);
2567 		while (buf_dirty_count_severe()) {
2568 			bdirtywait = 1;
2569 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2570 			    "flswai", 0);
2571 		}
2572 		mtx_unlock(&bdirtylock);
2573 	}
2574 }
2575 
2576 /*
2577  * Return true if we have too many dirty buffers.
2578  */
2579 int
2580 buf_dirty_count_severe(void)
2581 {
2582 
2583 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2584 }
2585 
2586 /*
2587  *	brelse:
2588  *
2589  *	Release a busy buffer and, if requested, free its resources.  The
2590  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2591  *	to be accessed later as a cache entity or reused for other purposes.
2592  */
2593 void
2594 brelse(struct buf *bp)
2595 {
2596 	struct mount *v_mnt;
2597 	int qindex;
2598 
2599 	/*
2600 	 * Many functions erroneously call brelse with a NULL bp under rare
2601 	 * error conditions. Simply return when called with a NULL bp.
2602 	 */
2603 	if (bp == NULL)
2604 		return;
2605 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2606 	    bp, bp->b_vp, bp->b_flags);
2607 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2608 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2609 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2610 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2611 
2612 	if (BUF_LOCKRECURSED(bp)) {
2613 		/*
2614 		 * Do not process, in particular, do not handle the
2615 		 * B_INVAL/B_RELBUF and do not release to free list.
2616 		 */
2617 		BUF_UNLOCK(bp);
2618 		return;
2619 	}
2620 
2621 	if (bp->b_flags & B_MANAGED) {
2622 		bqrelse(bp);
2623 		return;
2624 	}
2625 
2626 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2627 		BO_LOCK(bp->b_bufobj);
2628 		bp->b_vflags &= ~BV_BKGRDERR;
2629 		BO_UNLOCK(bp->b_bufobj);
2630 		bdirty(bp);
2631 	}
2632 
2633 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2634 	    (bp->b_flags & B_INVALONERR)) {
2635 		/*
2636 		 * Forced invalidation of dirty buffer contents, to be used
2637 		 * after a failed write in the rare case that the loss of the
2638 		 * contents is acceptable.  The buffer is invalidated and
2639 		 * freed.
2640 		 */
2641 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2642 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2643 	}
2644 
2645 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2646 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2647 	    !(bp->b_flags & B_INVAL)) {
2648 		/*
2649 		 * Failed write, redirty.  All errors except ENXIO (which
2650 		 * means the device is gone) are treated as being
2651 		 * transient.
2652 		 *
2653 		 * XXX Treating EIO as transient is not correct; the
2654 		 * contract with the local storage device drivers is that
2655 		 * they will only return EIO once the I/O is no longer
2656 		 * retriable.  Network I/O also respects this through the
2657 		 * guarantees of TCP and/or the internal retries of NFS.
2658 		 * ENOMEM might be transient, but we also have no way of
2659 		 * knowing when its ok to retry/reschedule.  In general,
2660 		 * this entire case should be made obsolete through better
2661 		 * error handling/recovery and resource scheduling.
2662 		 *
2663 		 * Do this also for buffers that failed with ENXIO, but have
2664 		 * non-empty dependencies - the soft updates code might need
2665 		 * to access the buffer to untangle them.
2666 		 *
2667 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2668 		 */
2669 		bp->b_ioflags &= ~BIO_ERROR;
2670 		bdirty(bp);
2671 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2672 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2673 		/*
2674 		 * Either a failed read I/O, or we were asked to free or not
2675 		 * cache the buffer, or we failed to write to a device that's
2676 		 * no longer present.
2677 		 */
2678 		bp->b_flags |= B_INVAL;
2679 		if (!LIST_EMPTY(&bp->b_dep))
2680 			buf_deallocate(bp);
2681 		if (bp->b_flags & B_DELWRI)
2682 			bdirtysub(bp);
2683 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2684 		if ((bp->b_flags & B_VMIO) == 0) {
2685 			allocbuf(bp, 0);
2686 			if (bp->b_vp)
2687 				brelvp(bp);
2688 		}
2689 	}
2690 
2691 	/*
2692 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2693 	 * is called with B_DELWRI set, the underlying pages may wind up
2694 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2695 	 * because pages associated with a B_DELWRI bp are marked clean.
2696 	 *
2697 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2698 	 * if B_DELWRI is set.
2699 	 */
2700 	if (bp->b_flags & B_DELWRI)
2701 		bp->b_flags &= ~B_RELBUF;
2702 
2703 	/*
2704 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2705 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2706 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2707 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2708 	 *
2709 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2710 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2711 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2712 	 *
2713 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2714 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2715 	 * the commit state and we cannot afford to lose the buffer. If the
2716 	 * buffer has a background write in progress, we need to keep it
2717 	 * around to prevent it from being reconstituted and starting a second
2718 	 * background write.
2719 	 */
2720 
2721 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2722 
2723 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2724 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2725 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2726 	    vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2727 		vfs_vmio_invalidate(bp);
2728 		allocbuf(bp, 0);
2729 	}
2730 
2731 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2732 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2733 		allocbuf(bp, 0);
2734 		bp->b_flags &= ~B_NOREUSE;
2735 		if (bp->b_vp != NULL)
2736 			brelvp(bp);
2737 	}
2738 
2739 	/*
2740 	 * If the buffer has junk contents signal it and eventually
2741 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2742 	 * doesn't find it.
2743 	 */
2744 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2745 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2746 		bp->b_flags |= B_INVAL;
2747 	if (bp->b_flags & B_INVAL) {
2748 		if (bp->b_flags & B_DELWRI)
2749 			bundirty(bp);
2750 		if (bp->b_vp)
2751 			brelvp(bp);
2752 	}
2753 
2754 	buf_track(bp, __func__);
2755 
2756 	/* buffers with no memory */
2757 	if (bp->b_bufsize == 0) {
2758 		buf_free(bp);
2759 		return;
2760 	}
2761 	/* buffers with junk contents */
2762 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2763 	    (bp->b_ioflags & BIO_ERROR)) {
2764 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2765 		if (bp->b_vflags & BV_BKGRDINPROG)
2766 			panic("losing buffer 2");
2767 		qindex = QUEUE_CLEAN;
2768 		bp->b_flags |= B_AGE;
2769 	/* remaining buffers */
2770 	} else if (bp->b_flags & B_DELWRI)
2771 		qindex = QUEUE_DIRTY;
2772 	else
2773 		qindex = QUEUE_CLEAN;
2774 
2775 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2776 		panic("brelse: not dirty");
2777 
2778 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2779 	/* binsfree unlocks bp. */
2780 	binsfree(bp, qindex);
2781 }
2782 
2783 /*
2784  * Release a buffer back to the appropriate queue but do not try to free
2785  * it.  The buffer is expected to be used again soon.
2786  *
2787  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2788  * biodone() to requeue an async I/O on completion.  It is also used when
2789  * known good buffers need to be requeued but we think we may need the data
2790  * again soon.
2791  *
2792  * XXX we should be able to leave the B_RELBUF hint set on completion.
2793  */
2794 void
2795 bqrelse(struct buf *bp)
2796 {
2797 	int qindex;
2798 
2799 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2800 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2801 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2802 
2803 	qindex = QUEUE_NONE;
2804 	if (BUF_LOCKRECURSED(bp)) {
2805 		/* do not release to free list */
2806 		BUF_UNLOCK(bp);
2807 		return;
2808 	}
2809 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2810 
2811 	if (bp->b_flags & B_MANAGED) {
2812 		if (bp->b_flags & B_REMFREE)
2813 			bremfreef(bp);
2814 		goto out;
2815 	}
2816 
2817 	/* buffers with stale but valid contents */
2818 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2819 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2820 		BO_LOCK(bp->b_bufobj);
2821 		bp->b_vflags &= ~BV_BKGRDERR;
2822 		BO_UNLOCK(bp->b_bufobj);
2823 		qindex = QUEUE_DIRTY;
2824 	} else {
2825 		if ((bp->b_flags & B_DELWRI) == 0 &&
2826 		    (bp->b_xflags & BX_VNDIRTY))
2827 			panic("bqrelse: not dirty");
2828 		if ((bp->b_flags & B_NOREUSE) != 0) {
2829 			brelse(bp);
2830 			return;
2831 		}
2832 		qindex = QUEUE_CLEAN;
2833 	}
2834 	buf_track(bp, __func__);
2835 	/* binsfree unlocks bp. */
2836 	binsfree(bp, qindex);
2837 	return;
2838 
2839 out:
2840 	buf_track(bp, __func__);
2841 	/* unlock */
2842 	BUF_UNLOCK(bp);
2843 }
2844 
2845 /*
2846  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2847  * restore bogus pages.
2848  */
2849 static void
2850 vfs_vmio_iodone(struct buf *bp)
2851 {
2852 	vm_ooffset_t foff;
2853 	vm_page_t m;
2854 	vm_object_t obj;
2855 	struct vnode *vp __unused;
2856 	int i, iosize, resid;
2857 	bool bogus;
2858 
2859 	obj = bp->b_bufobj->bo_object;
2860 	KASSERT(REFCOUNT_COUNT(obj->paging_in_progress) >= bp->b_npages,
2861 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2862 	    REFCOUNT_COUNT(obj->paging_in_progress), bp->b_npages));
2863 
2864 	vp = bp->b_vp;
2865 	KASSERT(vp->v_holdcnt > 0,
2866 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2867 	KASSERT(vp->v_object != NULL,
2868 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2869 
2870 	foff = bp->b_offset;
2871 	KASSERT(bp->b_offset != NOOFFSET,
2872 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2873 
2874 	bogus = false;
2875 	iosize = bp->b_bcount - bp->b_resid;
2876 	for (i = 0; i < bp->b_npages; i++) {
2877 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2878 		if (resid > iosize)
2879 			resid = iosize;
2880 
2881 		/*
2882 		 * cleanup bogus pages, restoring the originals
2883 		 */
2884 		m = bp->b_pages[i];
2885 		if (m == bogus_page) {
2886 			if (bogus == false) {
2887 				bogus = true;
2888 				VM_OBJECT_RLOCK(obj);
2889 			}
2890 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2891 			if (m == NULL)
2892 				panic("biodone: page disappeared!");
2893 			bp->b_pages[i] = m;
2894 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2895 			/*
2896 			 * In the write case, the valid and clean bits are
2897 			 * already changed correctly ( see bdwrite() ), so we
2898 			 * only need to do this here in the read case.
2899 			 */
2900 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2901 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2902 			    "has unexpected dirty bits", m));
2903 			vfs_page_set_valid(bp, foff, m);
2904 		}
2905 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2906 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2907 		    (intmax_t)foff, (uintmax_t)m->pindex));
2908 
2909 		vm_page_sunbusy(m);
2910 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2911 		iosize -= resid;
2912 	}
2913 	if (bogus)
2914 		VM_OBJECT_RUNLOCK(obj);
2915 	vm_object_pip_wakeupn(obj, bp->b_npages);
2916 	if (bogus && buf_mapped(bp)) {
2917 		BUF_CHECK_MAPPED(bp);
2918 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2919 		    bp->b_pages, bp->b_npages);
2920 	}
2921 }
2922 
2923 /*
2924  * Perform page invalidation when a buffer is released.  The fully invalid
2925  * pages will be reclaimed later in vfs_vmio_truncate().
2926  */
2927 static void
2928 vfs_vmio_invalidate(struct buf *bp)
2929 {
2930 	vm_object_t obj;
2931 	vm_page_t m;
2932 	int flags, i, resid, poffset, presid;
2933 
2934 	if (buf_mapped(bp)) {
2935 		BUF_CHECK_MAPPED(bp);
2936 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2937 	} else
2938 		BUF_CHECK_UNMAPPED(bp);
2939 	/*
2940 	 * Get the base offset and length of the buffer.  Note that
2941 	 * in the VMIO case if the buffer block size is not
2942 	 * page-aligned then b_data pointer may not be page-aligned.
2943 	 * But our b_pages[] array *IS* page aligned.
2944 	 *
2945 	 * block sizes less then DEV_BSIZE (usually 512) are not
2946 	 * supported due to the page granularity bits (m->valid,
2947 	 * m->dirty, etc...).
2948 	 *
2949 	 * See man buf(9) for more information
2950 	 */
2951 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2952 	obj = bp->b_bufobj->bo_object;
2953 	resid = bp->b_bufsize;
2954 	poffset = bp->b_offset & PAGE_MASK;
2955 	VM_OBJECT_WLOCK(obj);
2956 	for (i = 0; i < bp->b_npages; i++) {
2957 		m = bp->b_pages[i];
2958 		if (m == bogus_page)
2959 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2960 		bp->b_pages[i] = NULL;
2961 
2962 		presid = resid > (PAGE_SIZE - poffset) ?
2963 		    (PAGE_SIZE - poffset) : resid;
2964 		KASSERT(presid >= 0, ("brelse: extra page"));
2965 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
2966 		if (pmap_page_wired_mappings(m) == 0)
2967 			vm_page_set_invalid(m, poffset, presid);
2968 		vm_page_sunbusy(m);
2969 		vm_page_release_locked(m, flags);
2970 		resid -= presid;
2971 		poffset = 0;
2972 	}
2973 	VM_OBJECT_WUNLOCK(obj);
2974 	bp->b_npages = 0;
2975 }
2976 
2977 /*
2978  * Page-granular truncation of an existing VMIO buffer.
2979  */
2980 static void
2981 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2982 {
2983 	vm_object_t obj;
2984 	vm_page_t m;
2985 	int flags, i;
2986 
2987 	if (bp->b_npages == desiredpages)
2988 		return;
2989 
2990 	if (buf_mapped(bp)) {
2991 		BUF_CHECK_MAPPED(bp);
2992 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2993 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2994 	} else
2995 		BUF_CHECK_UNMAPPED(bp);
2996 
2997 	/*
2998 	 * The object lock is needed only if we will attempt to free pages.
2999 	 */
3000 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3001 	if ((bp->b_flags & B_DIRECT) != 0) {
3002 		flags |= VPR_TRYFREE;
3003 		obj = bp->b_bufobj->bo_object;
3004 		VM_OBJECT_WLOCK(obj);
3005 	} else {
3006 		obj = NULL;
3007 	}
3008 	for (i = desiredpages; i < bp->b_npages; i++) {
3009 		m = bp->b_pages[i];
3010 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3011 		bp->b_pages[i] = NULL;
3012 		if (obj != NULL)
3013 			vm_page_release_locked(m, flags);
3014 		else
3015 			vm_page_release(m, flags);
3016 	}
3017 	if (obj != NULL)
3018 		VM_OBJECT_WUNLOCK(obj);
3019 	bp->b_npages = desiredpages;
3020 }
3021 
3022 /*
3023  * Byte granular extension of VMIO buffers.
3024  */
3025 static void
3026 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3027 {
3028 	/*
3029 	 * We are growing the buffer, possibly in a
3030 	 * byte-granular fashion.
3031 	 */
3032 	vm_object_t obj;
3033 	vm_offset_t toff;
3034 	vm_offset_t tinc;
3035 	vm_page_t m;
3036 
3037 	/*
3038 	 * Step 1, bring in the VM pages from the object, allocating
3039 	 * them if necessary.  We must clear B_CACHE if these pages
3040 	 * are not valid for the range covered by the buffer.
3041 	 */
3042 	obj = bp->b_bufobj->bo_object;
3043 	if (bp->b_npages < desiredpages) {
3044 		/*
3045 		 * We must allocate system pages since blocking
3046 		 * here could interfere with paging I/O, no
3047 		 * matter which process we are.
3048 		 *
3049 		 * Only exclusive busy can be tested here.
3050 		 * Blocking on shared busy might lead to
3051 		 * deadlocks once allocbuf() is called after
3052 		 * pages are vfs_busy_pages().
3053 		 */
3054 		VM_OBJECT_WLOCK(obj);
3055 		(void)vm_page_grab_pages(obj,
3056 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3057 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3058 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3059 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3060 		VM_OBJECT_WUNLOCK(obj);
3061 		bp->b_npages = desiredpages;
3062 	}
3063 
3064 	/*
3065 	 * Step 2.  We've loaded the pages into the buffer,
3066 	 * we have to figure out if we can still have B_CACHE
3067 	 * set.  Note that B_CACHE is set according to the
3068 	 * byte-granular range ( bcount and size ), not the
3069 	 * aligned range ( newbsize ).
3070 	 *
3071 	 * The VM test is against m->valid, which is DEV_BSIZE
3072 	 * aligned.  Needless to say, the validity of the data
3073 	 * needs to also be DEV_BSIZE aligned.  Note that this
3074 	 * fails with NFS if the server or some other client
3075 	 * extends the file's EOF.  If our buffer is resized,
3076 	 * B_CACHE may remain set! XXX
3077 	 */
3078 	toff = bp->b_bcount;
3079 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3080 	while ((bp->b_flags & B_CACHE) && toff < size) {
3081 		vm_pindex_t pi;
3082 
3083 		if (tinc > (size - toff))
3084 			tinc = size - toff;
3085 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3086 		m = bp->b_pages[pi];
3087 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3088 		toff += tinc;
3089 		tinc = PAGE_SIZE;
3090 	}
3091 
3092 	/*
3093 	 * Step 3, fixup the KVA pmap.
3094 	 */
3095 	if (buf_mapped(bp))
3096 		bpmap_qenter(bp);
3097 	else
3098 		BUF_CHECK_UNMAPPED(bp);
3099 }
3100 
3101 /*
3102  * Check to see if a block at a particular lbn is available for a clustered
3103  * write.
3104  */
3105 static int
3106 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3107 {
3108 	struct buf *bpa;
3109 	int match;
3110 
3111 	match = 0;
3112 
3113 	/* If the buf isn't in core skip it */
3114 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3115 		return (0);
3116 
3117 	/* If the buf is busy we don't want to wait for it */
3118 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3119 		return (0);
3120 
3121 	/* Only cluster with valid clusterable delayed write buffers */
3122 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3123 	    (B_DELWRI | B_CLUSTEROK))
3124 		goto done;
3125 
3126 	if (bpa->b_bufsize != size)
3127 		goto done;
3128 
3129 	/*
3130 	 * Check to see if it is in the expected place on disk and that the
3131 	 * block has been mapped.
3132 	 */
3133 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3134 		match = 1;
3135 done:
3136 	BUF_UNLOCK(bpa);
3137 	return (match);
3138 }
3139 
3140 /*
3141  *	vfs_bio_awrite:
3142  *
3143  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3144  *	This is much better then the old way of writing only one buffer at
3145  *	a time.  Note that we may not be presented with the buffers in the
3146  *	correct order, so we search for the cluster in both directions.
3147  */
3148 int
3149 vfs_bio_awrite(struct buf *bp)
3150 {
3151 	struct bufobj *bo;
3152 	int i;
3153 	int j;
3154 	daddr_t lblkno = bp->b_lblkno;
3155 	struct vnode *vp = bp->b_vp;
3156 	int ncl;
3157 	int nwritten;
3158 	int size;
3159 	int maxcl;
3160 	int gbflags;
3161 
3162 	bo = &vp->v_bufobj;
3163 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3164 	/*
3165 	 * right now we support clustered writing only to regular files.  If
3166 	 * we find a clusterable block we could be in the middle of a cluster
3167 	 * rather then at the beginning.
3168 	 */
3169 	if ((vp->v_type == VREG) &&
3170 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3171 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3172 
3173 		size = vp->v_mount->mnt_stat.f_iosize;
3174 		maxcl = MAXPHYS / size;
3175 
3176 		BO_RLOCK(bo);
3177 		for (i = 1; i < maxcl; i++)
3178 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3179 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3180 				break;
3181 
3182 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3183 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3184 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3185 				break;
3186 		BO_RUNLOCK(bo);
3187 		--j;
3188 		ncl = i + j;
3189 		/*
3190 		 * this is a possible cluster write
3191 		 */
3192 		if (ncl != 1) {
3193 			BUF_UNLOCK(bp);
3194 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3195 			    gbflags);
3196 			return (nwritten);
3197 		}
3198 	}
3199 	bremfree(bp);
3200 	bp->b_flags |= B_ASYNC;
3201 	/*
3202 	 * default (old) behavior, writing out only one block
3203 	 *
3204 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3205 	 */
3206 	nwritten = bp->b_bufsize;
3207 	(void) bwrite(bp);
3208 
3209 	return (nwritten);
3210 }
3211 
3212 /*
3213  *	getnewbuf_kva:
3214  *
3215  *	Allocate KVA for an empty buf header according to gbflags.
3216  */
3217 static int
3218 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3219 {
3220 
3221 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3222 		/*
3223 		 * In order to keep fragmentation sane we only allocate kva
3224 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3225 		 */
3226 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3227 
3228 		if (maxsize != bp->b_kvasize &&
3229 		    bufkva_alloc(bp, maxsize, gbflags))
3230 			return (ENOSPC);
3231 	}
3232 	return (0);
3233 }
3234 
3235 /*
3236  *	getnewbuf:
3237  *
3238  *	Find and initialize a new buffer header, freeing up existing buffers
3239  *	in the bufqueues as necessary.  The new buffer is returned locked.
3240  *
3241  *	We block if:
3242  *		We have insufficient buffer headers
3243  *		We have insufficient buffer space
3244  *		buffer_arena is too fragmented ( space reservation fails )
3245  *		If we have to flush dirty buffers ( but we try to avoid this )
3246  *
3247  *	The caller is responsible for releasing the reserved bufspace after
3248  *	allocbuf() is called.
3249  */
3250 static struct buf *
3251 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3252 {
3253 	struct bufdomain *bd;
3254 	struct buf *bp;
3255 	bool metadata, reserved;
3256 
3257 	bp = NULL;
3258 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3259 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3260 	if (!unmapped_buf_allowed)
3261 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3262 
3263 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3264 	    vp->v_type == VCHR)
3265 		metadata = true;
3266 	else
3267 		metadata = false;
3268 	if (vp == NULL)
3269 		bd = &bdomain[0];
3270 	else
3271 		bd = &bdomain[vp->v_bufobj.bo_domain];
3272 
3273 	counter_u64_add(getnewbufcalls, 1);
3274 	reserved = false;
3275 	do {
3276 		if (reserved == false &&
3277 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3278 			counter_u64_add(getnewbufrestarts, 1);
3279 			continue;
3280 		}
3281 		reserved = true;
3282 		if ((bp = buf_alloc(bd)) == NULL) {
3283 			counter_u64_add(getnewbufrestarts, 1);
3284 			continue;
3285 		}
3286 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3287 			return (bp);
3288 		break;
3289 	} while (buf_recycle(bd, false) == 0);
3290 
3291 	if (reserved)
3292 		bufspace_release(bd, maxsize);
3293 	if (bp != NULL) {
3294 		bp->b_flags |= B_INVAL;
3295 		brelse(bp);
3296 	}
3297 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3298 
3299 	return (NULL);
3300 }
3301 
3302 /*
3303  *	buf_daemon:
3304  *
3305  *	buffer flushing daemon.  Buffers are normally flushed by the
3306  *	update daemon but if it cannot keep up this process starts to
3307  *	take the load in an attempt to prevent getnewbuf() from blocking.
3308  */
3309 static struct kproc_desc buf_kp = {
3310 	"bufdaemon",
3311 	buf_daemon,
3312 	&bufdaemonproc
3313 };
3314 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3315 
3316 static int
3317 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3318 {
3319 	int flushed;
3320 
3321 	flushed = flushbufqueues(vp, bd, target, 0);
3322 	if (flushed == 0) {
3323 		/*
3324 		 * Could not find any buffers without rollback
3325 		 * dependencies, so just write the first one
3326 		 * in the hopes of eventually making progress.
3327 		 */
3328 		if (vp != NULL && target > 2)
3329 			target /= 2;
3330 		flushbufqueues(vp, bd, target, 1);
3331 	}
3332 	return (flushed);
3333 }
3334 
3335 static void
3336 buf_daemon()
3337 {
3338 	struct bufdomain *bd;
3339 	int speedupreq;
3340 	int lodirty;
3341 	int i;
3342 
3343 	/*
3344 	 * This process needs to be suspended prior to shutdown sync.
3345 	 */
3346 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3347 	    SHUTDOWN_PRI_LAST + 100);
3348 
3349 	/*
3350 	 * Start the buf clean daemons as children threads.
3351 	 */
3352 	for (i = 0 ; i < buf_domains; i++) {
3353 		int error;
3354 
3355 		error = kthread_add((void (*)(void *))bufspace_daemon,
3356 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3357 		if (error)
3358 			panic("error %d spawning bufspace daemon", error);
3359 	}
3360 
3361 	/*
3362 	 * This process is allowed to take the buffer cache to the limit
3363 	 */
3364 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3365 	mtx_lock(&bdlock);
3366 	for (;;) {
3367 		bd_request = 0;
3368 		mtx_unlock(&bdlock);
3369 
3370 		kthread_suspend_check();
3371 
3372 		/*
3373 		 * Save speedupreq for this pass and reset to capture new
3374 		 * requests.
3375 		 */
3376 		speedupreq = bd_speedupreq;
3377 		bd_speedupreq = 0;
3378 
3379 		/*
3380 		 * Flush each domain sequentially according to its level and
3381 		 * the speedup request.
3382 		 */
3383 		for (i = 0; i < buf_domains; i++) {
3384 			bd = &bdomain[i];
3385 			if (speedupreq)
3386 				lodirty = bd->bd_numdirtybuffers / 2;
3387 			else
3388 				lodirty = bd->bd_lodirtybuffers;
3389 			while (bd->bd_numdirtybuffers > lodirty) {
3390 				if (buf_flush(NULL, bd,
3391 				    bd->bd_numdirtybuffers - lodirty) == 0)
3392 					break;
3393 				kern_yield(PRI_USER);
3394 			}
3395 		}
3396 
3397 		/*
3398 		 * Only clear bd_request if we have reached our low water
3399 		 * mark.  The buf_daemon normally waits 1 second and
3400 		 * then incrementally flushes any dirty buffers that have
3401 		 * built up, within reason.
3402 		 *
3403 		 * If we were unable to hit our low water mark and couldn't
3404 		 * find any flushable buffers, we sleep for a short period
3405 		 * to avoid endless loops on unlockable buffers.
3406 		 */
3407 		mtx_lock(&bdlock);
3408 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3409 			/*
3410 			 * We reached our low water mark, reset the
3411 			 * request and sleep until we are needed again.
3412 			 * The sleep is just so the suspend code works.
3413 			 */
3414 			bd_request = 0;
3415 			/*
3416 			 * Do an extra wakeup in case dirty threshold
3417 			 * changed via sysctl and the explicit transition
3418 			 * out of shortfall was missed.
3419 			 */
3420 			bdirtywakeup();
3421 			if (runningbufspace <= lorunningspace)
3422 				runningwakeup();
3423 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3424 		} else {
3425 			/*
3426 			 * We couldn't find any flushable dirty buffers but
3427 			 * still have too many dirty buffers, we
3428 			 * have to sleep and try again.  (rare)
3429 			 */
3430 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3431 		}
3432 	}
3433 }
3434 
3435 /*
3436  *	flushbufqueues:
3437  *
3438  *	Try to flush a buffer in the dirty queue.  We must be careful to
3439  *	free up B_INVAL buffers instead of write them, which NFS is
3440  *	particularly sensitive to.
3441  */
3442 static int flushwithdeps = 0;
3443 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3444     &flushwithdeps, 0,
3445     "Number of buffers flushed with dependecies that require rollbacks");
3446 
3447 static int
3448 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3449     int flushdeps)
3450 {
3451 	struct bufqueue *bq;
3452 	struct buf *sentinel;
3453 	struct vnode *vp;
3454 	struct mount *mp;
3455 	struct buf *bp;
3456 	int hasdeps;
3457 	int flushed;
3458 	int error;
3459 	bool unlock;
3460 
3461 	flushed = 0;
3462 	bq = &bd->bd_dirtyq;
3463 	bp = NULL;
3464 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3465 	sentinel->b_qindex = QUEUE_SENTINEL;
3466 	BQ_LOCK(bq);
3467 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3468 	BQ_UNLOCK(bq);
3469 	while (flushed != target) {
3470 		maybe_yield();
3471 		BQ_LOCK(bq);
3472 		bp = TAILQ_NEXT(sentinel, b_freelist);
3473 		if (bp != NULL) {
3474 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3475 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3476 			    b_freelist);
3477 		} else {
3478 			BQ_UNLOCK(bq);
3479 			break;
3480 		}
3481 		/*
3482 		 * Skip sentinels inserted by other invocations of the
3483 		 * flushbufqueues(), taking care to not reorder them.
3484 		 *
3485 		 * Only flush the buffers that belong to the
3486 		 * vnode locked by the curthread.
3487 		 */
3488 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3489 		    bp->b_vp != lvp)) {
3490 			BQ_UNLOCK(bq);
3491 			continue;
3492 		}
3493 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3494 		BQ_UNLOCK(bq);
3495 		if (error != 0)
3496 			continue;
3497 
3498 		/*
3499 		 * BKGRDINPROG can only be set with the buf and bufobj
3500 		 * locks both held.  We tolerate a race to clear it here.
3501 		 */
3502 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3503 		    (bp->b_flags & B_DELWRI) == 0) {
3504 			BUF_UNLOCK(bp);
3505 			continue;
3506 		}
3507 		if (bp->b_flags & B_INVAL) {
3508 			bremfreef(bp);
3509 			brelse(bp);
3510 			flushed++;
3511 			continue;
3512 		}
3513 
3514 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3515 			if (flushdeps == 0) {
3516 				BUF_UNLOCK(bp);
3517 				continue;
3518 			}
3519 			hasdeps = 1;
3520 		} else
3521 			hasdeps = 0;
3522 		/*
3523 		 * We must hold the lock on a vnode before writing
3524 		 * one of its buffers. Otherwise we may confuse, or
3525 		 * in the case of a snapshot vnode, deadlock the
3526 		 * system.
3527 		 *
3528 		 * The lock order here is the reverse of the normal
3529 		 * of vnode followed by buf lock.  This is ok because
3530 		 * the NOWAIT will prevent deadlock.
3531 		 */
3532 		vp = bp->b_vp;
3533 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3534 			BUF_UNLOCK(bp);
3535 			continue;
3536 		}
3537 		if (lvp == NULL) {
3538 			unlock = true;
3539 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3540 		} else {
3541 			ASSERT_VOP_LOCKED(vp, "getbuf");
3542 			unlock = false;
3543 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3544 			    vn_lock(vp, LK_TRYUPGRADE);
3545 		}
3546 		if (error == 0) {
3547 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3548 			    bp, bp->b_vp, bp->b_flags);
3549 			if (curproc == bufdaemonproc) {
3550 				vfs_bio_awrite(bp);
3551 			} else {
3552 				bremfree(bp);
3553 				bwrite(bp);
3554 				counter_u64_add(notbufdflushes, 1);
3555 			}
3556 			vn_finished_write(mp);
3557 			if (unlock)
3558 				VOP_UNLOCK(vp, 0);
3559 			flushwithdeps += hasdeps;
3560 			flushed++;
3561 
3562 			/*
3563 			 * Sleeping on runningbufspace while holding
3564 			 * vnode lock leads to deadlock.
3565 			 */
3566 			if (curproc == bufdaemonproc &&
3567 			    runningbufspace > hirunningspace)
3568 				waitrunningbufspace();
3569 			continue;
3570 		}
3571 		vn_finished_write(mp);
3572 		BUF_UNLOCK(bp);
3573 	}
3574 	BQ_LOCK(bq);
3575 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3576 	BQ_UNLOCK(bq);
3577 	free(sentinel, M_TEMP);
3578 	return (flushed);
3579 }
3580 
3581 /*
3582  * Check to see if a block is currently memory resident.
3583  */
3584 struct buf *
3585 incore(struct bufobj *bo, daddr_t blkno)
3586 {
3587 	struct buf *bp;
3588 
3589 	BO_RLOCK(bo);
3590 	bp = gbincore(bo, blkno);
3591 	BO_RUNLOCK(bo);
3592 	return (bp);
3593 }
3594 
3595 /*
3596  * Returns true if no I/O is needed to access the
3597  * associated VM object.  This is like incore except
3598  * it also hunts around in the VM system for the data.
3599  */
3600 
3601 static int
3602 inmem(struct vnode * vp, daddr_t blkno)
3603 {
3604 	vm_object_t obj;
3605 	vm_offset_t toff, tinc, size;
3606 	vm_page_t m;
3607 	vm_ooffset_t off;
3608 
3609 	ASSERT_VOP_LOCKED(vp, "inmem");
3610 
3611 	if (incore(&vp->v_bufobj, blkno))
3612 		return 1;
3613 	if (vp->v_mount == NULL)
3614 		return 0;
3615 	obj = vp->v_object;
3616 	if (obj == NULL)
3617 		return (0);
3618 
3619 	size = PAGE_SIZE;
3620 	if (size > vp->v_mount->mnt_stat.f_iosize)
3621 		size = vp->v_mount->mnt_stat.f_iosize;
3622 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3623 
3624 	VM_OBJECT_RLOCK(obj);
3625 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3626 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3627 		if (!m)
3628 			goto notinmem;
3629 		tinc = size;
3630 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3631 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3632 		if (vm_page_is_valid(m,
3633 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3634 			goto notinmem;
3635 	}
3636 	VM_OBJECT_RUNLOCK(obj);
3637 	return 1;
3638 
3639 notinmem:
3640 	VM_OBJECT_RUNLOCK(obj);
3641 	return (0);
3642 }
3643 
3644 /*
3645  * Set the dirty range for a buffer based on the status of the dirty
3646  * bits in the pages comprising the buffer.  The range is limited
3647  * to the size of the buffer.
3648  *
3649  * Tell the VM system that the pages associated with this buffer
3650  * are clean.  This is used for delayed writes where the data is
3651  * going to go to disk eventually without additional VM intevention.
3652  *
3653  * Note that while we only really need to clean through to b_bcount, we
3654  * just go ahead and clean through to b_bufsize.
3655  */
3656 static void
3657 vfs_clean_pages_dirty_buf(struct buf *bp)
3658 {
3659 	vm_ooffset_t foff, noff, eoff;
3660 	vm_page_t m;
3661 	int i;
3662 
3663 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3664 		return;
3665 
3666 	foff = bp->b_offset;
3667 	KASSERT(bp->b_offset != NOOFFSET,
3668 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3669 
3670 	vfs_busy_pages_acquire(bp);
3671 	vfs_setdirty_range(bp);
3672 	for (i = 0; i < bp->b_npages; i++) {
3673 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3674 		eoff = noff;
3675 		if (eoff > bp->b_offset + bp->b_bufsize)
3676 			eoff = bp->b_offset + bp->b_bufsize;
3677 		m = bp->b_pages[i];
3678 		vfs_page_set_validclean(bp, foff, m);
3679 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3680 		foff = noff;
3681 	}
3682 	vfs_busy_pages_release(bp);
3683 }
3684 
3685 static void
3686 vfs_setdirty_range(struct buf *bp)
3687 {
3688 	vm_offset_t boffset;
3689 	vm_offset_t eoffset;
3690 	int i;
3691 
3692 	/*
3693 	 * test the pages to see if they have been modified directly
3694 	 * by users through the VM system.
3695 	 */
3696 	for (i = 0; i < bp->b_npages; i++)
3697 		vm_page_test_dirty(bp->b_pages[i]);
3698 
3699 	/*
3700 	 * Calculate the encompassing dirty range, boffset and eoffset,
3701 	 * (eoffset - boffset) bytes.
3702 	 */
3703 
3704 	for (i = 0; i < bp->b_npages; i++) {
3705 		if (bp->b_pages[i]->dirty)
3706 			break;
3707 	}
3708 	boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3709 
3710 	for (i = bp->b_npages - 1; i >= 0; --i) {
3711 		if (bp->b_pages[i]->dirty) {
3712 			break;
3713 		}
3714 	}
3715 	eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3716 
3717 	/*
3718 	 * Fit it to the buffer.
3719 	 */
3720 
3721 	if (eoffset > bp->b_bcount)
3722 		eoffset = bp->b_bcount;
3723 
3724 	/*
3725 	 * If we have a good dirty range, merge with the existing
3726 	 * dirty range.
3727 	 */
3728 
3729 	if (boffset < eoffset) {
3730 		if (bp->b_dirtyoff > boffset)
3731 			bp->b_dirtyoff = boffset;
3732 		if (bp->b_dirtyend < eoffset)
3733 			bp->b_dirtyend = eoffset;
3734 	}
3735 }
3736 
3737 /*
3738  * Allocate the KVA mapping for an existing buffer.
3739  * If an unmapped buffer is provided but a mapped buffer is requested, take
3740  * also care to properly setup mappings between pages and KVA.
3741  */
3742 static void
3743 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3744 {
3745 	int bsize, maxsize, need_mapping, need_kva;
3746 	off_t offset;
3747 
3748 	need_mapping = bp->b_data == unmapped_buf &&
3749 	    (gbflags & GB_UNMAPPED) == 0;
3750 	need_kva = bp->b_kvabase == unmapped_buf &&
3751 	    bp->b_data == unmapped_buf &&
3752 	    (gbflags & GB_KVAALLOC) != 0;
3753 	if (!need_mapping && !need_kva)
3754 		return;
3755 
3756 	BUF_CHECK_UNMAPPED(bp);
3757 
3758 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3759 		/*
3760 		 * Buffer is not mapped, but the KVA was already
3761 		 * reserved at the time of the instantiation.  Use the
3762 		 * allocated space.
3763 		 */
3764 		goto has_addr;
3765 	}
3766 
3767 	/*
3768 	 * Calculate the amount of the address space we would reserve
3769 	 * if the buffer was mapped.
3770 	 */
3771 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3772 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3773 	offset = blkno * bsize;
3774 	maxsize = size + (offset & PAGE_MASK);
3775 	maxsize = imax(maxsize, bsize);
3776 
3777 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3778 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3779 			/*
3780 			 * XXXKIB: defragmentation cannot
3781 			 * succeed, not sure what else to do.
3782 			 */
3783 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3784 		}
3785 		counter_u64_add(mappingrestarts, 1);
3786 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3787 	}
3788 has_addr:
3789 	if (need_mapping) {
3790 		/* b_offset is handled by bpmap_qenter. */
3791 		bp->b_data = bp->b_kvabase;
3792 		BUF_CHECK_MAPPED(bp);
3793 		bpmap_qenter(bp);
3794 	}
3795 }
3796 
3797 struct buf *
3798 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3799     int flags)
3800 {
3801 	struct buf *bp;
3802 	int error;
3803 
3804 	error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3805 	if (error != 0)
3806 		return (NULL);
3807 	return (bp);
3808 }
3809 
3810 /*
3811  *	getblkx:
3812  *
3813  *	Get a block given a specified block and offset into a file/device.
3814  *	The buffers B_DONE bit will be cleared on return, making it almost
3815  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3816  *	return.  The caller should clear B_INVAL prior to initiating a
3817  *	READ.
3818  *
3819  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3820  *	an existing buffer.
3821  *
3822  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3823  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3824  *	and then cleared based on the backing VM.  If the previous buffer is
3825  *	non-0-sized but invalid, B_CACHE will be cleared.
3826  *
3827  *	If getblk() must create a new buffer, the new buffer is returned with
3828  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3829  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3830  *	backing VM.
3831  *
3832  *	getblk() also forces a bwrite() for any B_DELWRI buffer whose
3833  *	B_CACHE bit is clear.
3834  *
3835  *	What this means, basically, is that the caller should use B_CACHE to
3836  *	determine whether the buffer is fully valid or not and should clear
3837  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3838  *	the buffer by loading its data area with something, the caller needs
3839  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3840  *	the caller should set B_CACHE ( as an optimization ), else the caller
3841  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3842  *	a write attempt or if it was a successful read.  If the caller
3843  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3844  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3845  *
3846  *	The blkno parameter is the logical block being requested. Normally
3847  *	the mapping of logical block number to disk block address is done
3848  *	by calling VOP_BMAP(). However, if the mapping is already known, the
3849  *	disk block address can be passed using the dblkno parameter. If the
3850  *	disk block address is not known, then the same value should be passed
3851  *	for blkno and dblkno.
3852  */
3853 int
3854 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3855     int slptimeo, int flags, struct buf **bpp)
3856 {
3857 	struct buf *bp;
3858 	struct bufobj *bo;
3859 	daddr_t d_blkno;
3860 	int bsize, error, maxsize, vmio;
3861 	off_t offset;
3862 
3863 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3864 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3865 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3866 	ASSERT_VOP_LOCKED(vp, "getblk");
3867 	if (size > maxbcachebuf)
3868 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3869 		    maxbcachebuf);
3870 	if (!unmapped_buf_allowed)
3871 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3872 
3873 	bo = &vp->v_bufobj;
3874 	d_blkno = dblkno;
3875 loop:
3876 	BO_RLOCK(bo);
3877 	bp = gbincore(bo, blkno);
3878 	if (bp != NULL) {
3879 		int lockflags;
3880 		/*
3881 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3882 		 * it must be on a queue.
3883 		 */
3884 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3885 
3886 		if ((flags & GB_LOCK_NOWAIT) != 0)
3887 			lockflags |= LK_NOWAIT;
3888 
3889 		error = BUF_TIMELOCK(bp, lockflags,
3890 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3891 
3892 		/*
3893 		 * If we slept and got the lock we have to restart in case
3894 		 * the buffer changed identities.
3895 		 */
3896 		if (error == ENOLCK)
3897 			goto loop;
3898 		/* We timed out or were interrupted. */
3899 		else if (error != 0)
3900 			return (error);
3901 		/* If recursed, assume caller knows the rules. */
3902 		else if (BUF_LOCKRECURSED(bp))
3903 			goto end;
3904 
3905 		/*
3906 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3907 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3908 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3909 		 * backing VM cache.
3910 		 */
3911 		if (bp->b_flags & B_INVAL)
3912 			bp->b_flags &= ~B_CACHE;
3913 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3914 			bp->b_flags |= B_CACHE;
3915 		if (bp->b_flags & B_MANAGED)
3916 			MPASS(bp->b_qindex == QUEUE_NONE);
3917 		else
3918 			bremfree(bp);
3919 
3920 		/*
3921 		 * check for size inconsistencies for non-VMIO case.
3922 		 */
3923 		if (bp->b_bcount != size) {
3924 			if ((bp->b_flags & B_VMIO) == 0 ||
3925 			    (size > bp->b_kvasize)) {
3926 				if (bp->b_flags & B_DELWRI) {
3927 					bp->b_flags |= B_NOCACHE;
3928 					bwrite(bp);
3929 				} else {
3930 					if (LIST_EMPTY(&bp->b_dep)) {
3931 						bp->b_flags |= B_RELBUF;
3932 						brelse(bp);
3933 					} else {
3934 						bp->b_flags |= B_NOCACHE;
3935 						bwrite(bp);
3936 					}
3937 				}
3938 				goto loop;
3939 			}
3940 		}
3941 
3942 		/*
3943 		 * Handle the case of unmapped buffer which should
3944 		 * become mapped, or the buffer for which KVA
3945 		 * reservation is requested.
3946 		 */
3947 		bp_unmapped_get_kva(bp, blkno, size, flags);
3948 
3949 		/*
3950 		 * If the size is inconsistent in the VMIO case, we can resize
3951 		 * the buffer.  This might lead to B_CACHE getting set or
3952 		 * cleared.  If the size has not changed, B_CACHE remains
3953 		 * unchanged from its previous state.
3954 		 */
3955 		allocbuf(bp, size);
3956 
3957 		KASSERT(bp->b_offset != NOOFFSET,
3958 		    ("getblk: no buffer offset"));
3959 
3960 		/*
3961 		 * A buffer with B_DELWRI set and B_CACHE clear must
3962 		 * be committed before we can return the buffer in
3963 		 * order to prevent the caller from issuing a read
3964 		 * ( due to B_CACHE not being set ) and overwriting
3965 		 * it.
3966 		 *
3967 		 * Most callers, including NFS and FFS, need this to
3968 		 * operate properly either because they assume they
3969 		 * can issue a read if B_CACHE is not set, or because
3970 		 * ( for example ) an uncached B_DELWRI might loop due
3971 		 * to softupdates re-dirtying the buffer.  In the latter
3972 		 * case, B_CACHE is set after the first write completes,
3973 		 * preventing further loops.
3974 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3975 		 * above while extending the buffer, we cannot allow the
3976 		 * buffer to remain with B_CACHE set after the write
3977 		 * completes or it will represent a corrupt state.  To
3978 		 * deal with this we set B_NOCACHE to scrap the buffer
3979 		 * after the write.
3980 		 *
3981 		 * We might be able to do something fancy, like setting
3982 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3983 		 * so the below call doesn't set B_CACHE, but that gets real
3984 		 * confusing.  This is much easier.
3985 		 */
3986 
3987 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3988 			bp->b_flags |= B_NOCACHE;
3989 			bwrite(bp);
3990 			goto loop;
3991 		}
3992 		bp->b_flags &= ~B_DONE;
3993 	} else {
3994 		/*
3995 		 * Buffer is not in-core, create new buffer.  The buffer
3996 		 * returned by getnewbuf() is locked.  Note that the returned
3997 		 * buffer is also considered valid (not marked B_INVAL).
3998 		 */
3999 		BO_RUNLOCK(bo);
4000 		/*
4001 		 * If the user does not want us to create the buffer, bail out
4002 		 * here.
4003 		 */
4004 		if (flags & GB_NOCREAT)
4005 			return (EEXIST);
4006 
4007 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
4008 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4009 		offset = blkno * bsize;
4010 		vmio = vp->v_object != NULL;
4011 		if (vmio) {
4012 			maxsize = size + (offset & PAGE_MASK);
4013 		} else {
4014 			maxsize = size;
4015 			/* Do not allow non-VMIO notmapped buffers. */
4016 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4017 		}
4018 		maxsize = imax(maxsize, bsize);
4019 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4020 		    !vn_isdisk(vp, NULL)) {
4021 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4022 			KASSERT(error != EOPNOTSUPP,
4023 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4024 			    vp));
4025 			if (error != 0)
4026 				return (error);
4027 			if (d_blkno == -1)
4028 				return (EJUSTRETURN);
4029 		}
4030 
4031 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4032 		if (bp == NULL) {
4033 			if (slpflag || slptimeo)
4034 				return (ETIMEDOUT);
4035 			/*
4036 			 * XXX This is here until the sleep path is diagnosed
4037 			 * enough to work under very low memory conditions.
4038 			 *
4039 			 * There's an issue on low memory, 4BSD+non-preempt
4040 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4041 			 * exhaustion occurs without sleeping for buffer
4042 			 * reclaimation.  This just sticks in a loop and
4043 			 * constantly attempts to allocate a buffer, which
4044 			 * hits exhaustion and tries to wakeup bufdaemon.
4045 			 * This never happens because we never yield.
4046 			 *
4047 			 * The real solution is to identify and fix these cases
4048 			 * so we aren't effectively busy-waiting in a loop
4049 			 * until the reclaimation path has cycles to run.
4050 			 */
4051 			kern_yield(PRI_USER);
4052 			goto loop;
4053 		}
4054 
4055 		/*
4056 		 * This code is used to make sure that a buffer is not
4057 		 * created while the getnewbuf routine is blocked.
4058 		 * This can be a problem whether the vnode is locked or not.
4059 		 * If the buffer is created out from under us, we have to
4060 		 * throw away the one we just created.
4061 		 *
4062 		 * Note: this must occur before we associate the buffer
4063 		 * with the vp especially considering limitations in
4064 		 * the splay tree implementation when dealing with duplicate
4065 		 * lblkno's.
4066 		 */
4067 		BO_LOCK(bo);
4068 		if (gbincore(bo, blkno)) {
4069 			BO_UNLOCK(bo);
4070 			bp->b_flags |= B_INVAL;
4071 			bufspace_release(bufdomain(bp), maxsize);
4072 			brelse(bp);
4073 			goto loop;
4074 		}
4075 
4076 		/*
4077 		 * Insert the buffer into the hash, so that it can
4078 		 * be found by incore.
4079 		 */
4080 		bp->b_lblkno = blkno;
4081 		bp->b_blkno = d_blkno;
4082 		bp->b_offset = offset;
4083 		bgetvp(vp, bp);
4084 		BO_UNLOCK(bo);
4085 
4086 		/*
4087 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4088 		 * buffer size starts out as 0, B_CACHE will be set by
4089 		 * allocbuf() for the VMIO case prior to it testing the
4090 		 * backing store for validity.
4091 		 */
4092 
4093 		if (vmio) {
4094 			bp->b_flags |= B_VMIO;
4095 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4096 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4097 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4098 		} else {
4099 			bp->b_flags &= ~B_VMIO;
4100 			KASSERT(bp->b_bufobj->bo_object == NULL,
4101 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4102 			    bp, bp->b_bufobj->bo_object));
4103 			BUF_CHECK_MAPPED(bp);
4104 		}
4105 
4106 		allocbuf(bp, size);
4107 		bufspace_release(bufdomain(bp), maxsize);
4108 		bp->b_flags &= ~B_DONE;
4109 	}
4110 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4111 end:
4112 	buf_track(bp, __func__);
4113 	KASSERT(bp->b_bufobj == bo,
4114 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4115 	*bpp = bp;
4116 	return (0);
4117 }
4118 
4119 /*
4120  * Get an empty, disassociated buffer of given size.  The buffer is initially
4121  * set to B_INVAL.
4122  */
4123 struct buf *
4124 geteblk(int size, int flags)
4125 {
4126 	struct buf *bp;
4127 	int maxsize;
4128 
4129 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4130 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4131 		if ((flags & GB_NOWAIT_BD) &&
4132 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4133 			return (NULL);
4134 	}
4135 	allocbuf(bp, size);
4136 	bufspace_release(bufdomain(bp), maxsize);
4137 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4138 	return (bp);
4139 }
4140 
4141 /*
4142  * Truncate the backing store for a non-vmio buffer.
4143  */
4144 static void
4145 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4146 {
4147 
4148 	if (bp->b_flags & B_MALLOC) {
4149 		/*
4150 		 * malloced buffers are not shrunk
4151 		 */
4152 		if (newbsize == 0) {
4153 			bufmallocadjust(bp, 0);
4154 			free(bp->b_data, M_BIOBUF);
4155 			bp->b_data = bp->b_kvabase;
4156 			bp->b_flags &= ~B_MALLOC;
4157 		}
4158 		return;
4159 	}
4160 	vm_hold_free_pages(bp, newbsize);
4161 	bufspace_adjust(bp, newbsize);
4162 }
4163 
4164 /*
4165  * Extend the backing for a non-VMIO buffer.
4166  */
4167 static void
4168 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4169 {
4170 	caddr_t origbuf;
4171 	int origbufsize;
4172 
4173 	/*
4174 	 * We only use malloced memory on the first allocation.
4175 	 * and revert to page-allocated memory when the buffer
4176 	 * grows.
4177 	 *
4178 	 * There is a potential smp race here that could lead
4179 	 * to bufmallocspace slightly passing the max.  It
4180 	 * is probably extremely rare and not worth worrying
4181 	 * over.
4182 	 */
4183 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4184 	    bufmallocspace < maxbufmallocspace) {
4185 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4186 		bp->b_flags |= B_MALLOC;
4187 		bufmallocadjust(bp, newbsize);
4188 		return;
4189 	}
4190 
4191 	/*
4192 	 * If the buffer is growing on its other-than-first
4193 	 * allocation then we revert to the page-allocation
4194 	 * scheme.
4195 	 */
4196 	origbuf = NULL;
4197 	origbufsize = 0;
4198 	if (bp->b_flags & B_MALLOC) {
4199 		origbuf = bp->b_data;
4200 		origbufsize = bp->b_bufsize;
4201 		bp->b_data = bp->b_kvabase;
4202 		bufmallocadjust(bp, 0);
4203 		bp->b_flags &= ~B_MALLOC;
4204 		newbsize = round_page(newbsize);
4205 	}
4206 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4207 	    (vm_offset_t) bp->b_data + newbsize);
4208 	if (origbuf != NULL) {
4209 		bcopy(origbuf, bp->b_data, origbufsize);
4210 		free(origbuf, M_BIOBUF);
4211 	}
4212 	bufspace_adjust(bp, newbsize);
4213 }
4214 
4215 /*
4216  * This code constitutes the buffer memory from either anonymous system
4217  * memory (in the case of non-VMIO operations) or from an associated
4218  * VM object (in the case of VMIO operations).  This code is able to
4219  * resize a buffer up or down.
4220  *
4221  * Note that this code is tricky, and has many complications to resolve
4222  * deadlock or inconsistent data situations.  Tread lightly!!!
4223  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4224  * the caller.  Calling this code willy nilly can result in the loss of data.
4225  *
4226  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4227  * B_CACHE for the non-VMIO case.
4228  */
4229 int
4230 allocbuf(struct buf *bp, int size)
4231 {
4232 	int newbsize;
4233 
4234 	if (bp->b_bcount == size)
4235 		return (1);
4236 
4237 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4238 		panic("allocbuf: buffer too small");
4239 
4240 	newbsize = roundup2(size, DEV_BSIZE);
4241 	if ((bp->b_flags & B_VMIO) == 0) {
4242 		if ((bp->b_flags & B_MALLOC) == 0)
4243 			newbsize = round_page(newbsize);
4244 		/*
4245 		 * Just get anonymous memory from the kernel.  Don't
4246 		 * mess with B_CACHE.
4247 		 */
4248 		if (newbsize < bp->b_bufsize)
4249 			vfs_nonvmio_truncate(bp, newbsize);
4250 		else if (newbsize > bp->b_bufsize)
4251 			vfs_nonvmio_extend(bp, newbsize);
4252 	} else {
4253 		int desiredpages;
4254 
4255 		desiredpages = (size == 0) ? 0 :
4256 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4257 
4258 		if (bp->b_flags & B_MALLOC)
4259 			panic("allocbuf: VMIO buffer can't be malloced");
4260 		/*
4261 		 * Set B_CACHE initially if buffer is 0 length or will become
4262 		 * 0-length.
4263 		 */
4264 		if (size == 0 || bp->b_bufsize == 0)
4265 			bp->b_flags |= B_CACHE;
4266 
4267 		if (newbsize < bp->b_bufsize)
4268 			vfs_vmio_truncate(bp, desiredpages);
4269 		/* XXX This looks as if it should be newbsize > b_bufsize */
4270 		else if (size > bp->b_bcount)
4271 			vfs_vmio_extend(bp, desiredpages, size);
4272 		bufspace_adjust(bp, newbsize);
4273 	}
4274 	bp->b_bcount = size;		/* requested buffer size. */
4275 	return (1);
4276 }
4277 
4278 extern int inflight_transient_maps;
4279 
4280 static struct bio_queue nondump_bios;
4281 
4282 void
4283 biodone(struct bio *bp)
4284 {
4285 	struct mtx *mtxp;
4286 	void (*done)(struct bio *);
4287 	vm_offset_t start, end;
4288 
4289 	biotrack(bp, __func__);
4290 
4291 	/*
4292 	 * Avoid completing I/O when dumping after a panic since that may
4293 	 * result in a deadlock in the filesystem or pager code.  Note that
4294 	 * this doesn't affect dumps that were started manually since we aim
4295 	 * to keep the system usable after it has been resumed.
4296 	 */
4297 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4298 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4299 		return;
4300 	}
4301 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4302 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4303 		bp->bio_flags |= BIO_UNMAPPED;
4304 		start = trunc_page((vm_offset_t)bp->bio_data);
4305 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4306 		bp->bio_data = unmapped_buf;
4307 		pmap_qremove(start, atop(end - start));
4308 		vmem_free(transient_arena, start, end - start);
4309 		atomic_add_int(&inflight_transient_maps, -1);
4310 	}
4311 	done = bp->bio_done;
4312 	if (done == NULL) {
4313 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4314 		mtx_lock(mtxp);
4315 		bp->bio_flags |= BIO_DONE;
4316 		wakeup(bp);
4317 		mtx_unlock(mtxp);
4318 	} else
4319 		done(bp);
4320 }
4321 
4322 /*
4323  * Wait for a BIO to finish.
4324  */
4325 int
4326 biowait(struct bio *bp, const char *wchan)
4327 {
4328 	struct mtx *mtxp;
4329 
4330 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4331 	mtx_lock(mtxp);
4332 	while ((bp->bio_flags & BIO_DONE) == 0)
4333 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4334 	mtx_unlock(mtxp);
4335 	if (bp->bio_error != 0)
4336 		return (bp->bio_error);
4337 	if (!(bp->bio_flags & BIO_ERROR))
4338 		return (0);
4339 	return (EIO);
4340 }
4341 
4342 void
4343 biofinish(struct bio *bp, struct devstat *stat, int error)
4344 {
4345 
4346 	if (error) {
4347 		bp->bio_error = error;
4348 		bp->bio_flags |= BIO_ERROR;
4349 	}
4350 	if (stat != NULL)
4351 		devstat_end_transaction_bio(stat, bp);
4352 	biodone(bp);
4353 }
4354 
4355 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4356 void
4357 biotrack_buf(struct bio *bp, const char *location)
4358 {
4359 
4360 	buf_track(bp->bio_track_bp, location);
4361 }
4362 #endif
4363 
4364 /*
4365  *	bufwait:
4366  *
4367  *	Wait for buffer I/O completion, returning error status.  The buffer
4368  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4369  *	error and cleared.
4370  */
4371 int
4372 bufwait(struct buf *bp)
4373 {
4374 	if (bp->b_iocmd == BIO_READ)
4375 		bwait(bp, PRIBIO, "biord");
4376 	else
4377 		bwait(bp, PRIBIO, "biowr");
4378 	if (bp->b_flags & B_EINTR) {
4379 		bp->b_flags &= ~B_EINTR;
4380 		return (EINTR);
4381 	}
4382 	if (bp->b_ioflags & BIO_ERROR) {
4383 		return (bp->b_error ? bp->b_error : EIO);
4384 	} else {
4385 		return (0);
4386 	}
4387 }
4388 
4389 /*
4390  *	bufdone:
4391  *
4392  *	Finish I/O on a buffer, optionally calling a completion function.
4393  *	This is usually called from an interrupt so process blocking is
4394  *	not allowed.
4395  *
4396  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4397  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4398  *	assuming B_INVAL is clear.
4399  *
4400  *	For the VMIO case, we set B_CACHE if the op was a read and no
4401  *	read error occurred, or if the op was a write.  B_CACHE is never
4402  *	set if the buffer is invalid or otherwise uncacheable.
4403  *
4404  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4405  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4406  *	in the biodone routine.
4407  */
4408 void
4409 bufdone(struct buf *bp)
4410 {
4411 	struct bufobj *dropobj;
4412 	void    (*biodone)(struct buf *);
4413 
4414 	buf_track(bp, __func__);
4415 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4416 	dropobj = NULL;
4417 
4418 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4419 
4420 	runningbufwakeup(bp);
4421 	if (bp->b_iocmd == BIO_WRITE)
4422 		dropobj = bp->b_bufobj;
4423 	/* call optional completion function if requested */
4424 	if (bp->b_iodone != NULL) {
4425 		biodone = bp->b_iodone;
4426 		bp->b_iodone = NULL;
4427 		(*biodone) (bp);
4428 		if (dropobj)
4429 			bufobj_wdrop(dropobj);
4430 		return;
4431 	}
4432 	if (bp->b_flags & B_VMIO) {
4433 		/*
4434 		 * Set B_CACHE if the op was a normal read and no error
4435 		 * occurred.  B_CACHE is set for writes in the b*write()
4436 		 * routines.
4437 		 */
4438 		if (bp->b_iocmd == BIO_READ &&
4439 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4440 		    !(bp->b_ioflags & BIO_ERROR))
4441 			bp->b_flags |= B_CACHE;
4442 		vfs_vmio_iodone(bp);
4443 	}
4444 	if (!LIST_EMPTY(&bp->b_dep))
4445 		buf_complete(bp);
4446 	if ((bp->b_flags & B_CKHASH) != 0) {
4447 		KASSERT(bp->b_iocmd == BIO_READ,
4448 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4449 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4450 		(*bp->b_ckhashcalc)(bp);
4451 	}
4452 	/*
4453 	 * For asynchronous completions, release the buffer now. The brelse
4454 	 * will do a wakeup there if necessary - so no need to do a wakeup
4455 	 * here in the async case. The sync case always needs to do a wakeup.
4456 	 */
4457 	if (bp->b_flags & B_ASYNC) {
4458 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4459 		    (bp->b_ioflags & BIO_ERROR))
4460 			brelse(bp);
4461 		else
4462 			bqrelse(bp);
4463 	} else
4464 		bdone(bp);
4465 	if (dropobj)
4466 		bufobj_wdrop(dropobj);
4467 }
4468 
4469 /*
4470  * This routine is called in lieu of iodone in the case of
4471  * incomplete I/O.  This keeps the busy status for pages
4472  * consistent.
4473  */
4474 void
4475 vfs_unbusy_pages(struct buf *bp)
4476 {
4477 	int i;
4478 	vm_object_t obj;
4479 	vm_page_t m;
4480 	bool bogus;
4481 
4482 	runningbufwakeup(bp);
4483 	if (!(bp->b_flags & B_VMIO))
4484 		return;
4485 
4486 	obj = bp->b_bufobj->bo_object;
4487 	bogus = false;
4488 	for (i = 0; i < bp->b_npages; i++) {
4489 		m = bp->b_pages[i];
4490 		if (m == bogus_page) {
4491 			if (bogus == false) {
4492 				bogus = true;
4493 				VM_OBJECT_RLOCK(obj);
4494 			}
4495 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4496 			if (!m)
4497 				panic("vfs_unbusy_pages: page missing\n");
4498 			bp->b_pages[i] = m;
4499 			if (buf_mapped(bp)) {
4500 				BUF_CHECK_MAPPED(bp);
4501 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4502 				    bp->b_pages, bp->b_npages);
4503 			} else
4504 				BUF_CHECK_UNMAPPED(bp);
4505 		}
4506 		vm_page_sunbusy(m);
4507 	}
4508 	if (bogus)
4509 		VM_OBJECT_RUNLOCK(obj);
4510 	vm_object_pip_wakeupn(obj, bp->b_npages);
4511 }
4512 
4513 /*
4514  * vfs_page_set_valid:
4515  *
4516  *	Set the valid bits in a page based on the supplied offset.   The
4517  *	range is restricted to the buffer's size.
4518  *
4519  *	This routine is typically called after a read completes.
4520  */
4521 static void
4522 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4523 {
4524 	vm_ooffset_t eoff;
4525 
4526 	/*
4527 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4528 	 * page boundary and eoff is not greater than the end of the buffer.
4529 	 * The end of the buffer, in this case, is our file EOF, not the
4530 	 * allocation size of the buffer.
4531 	 */
4532 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4533 	if (eoff > bp->b_offset + bp->b_bcount)
4534 		eoff = bp->b_offset + bp->b_bcount;
4535 
4536 	/*
4537 	 * Set valid range.  This is typically the entire buffer and thus the
4538 	 * entire page.
4539 	 */
4540 	if (eoff > off)
4541 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4542 }
4543 
4544 /*
4545  * vfs_page_set_validclean:
4546  *
4547  *	Set the valid bits and clear the dirty bits in a page based on the
4548  *	supplied offset.   The range is restricted to the buffer's size.
4549  */
4550 static void
4551 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4552 {
4553 	vm_ooffset_t soff, eoff;
4554 
4555 	/*
4556 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4557 	 * page boundary or cross the end of the buffer.  The end of the
4558 	 * buffer, in this case, is our file EOF, not the allocation size
4559 	 * of the buffer.
4560 	 */
4561 	soff = off;
4562 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4563 	if (eoff > bp->b_offset + bp->b_bcount)
4564 		eoff = bp->b_offset + bp->b_bcount;
4565 
4566 	/*
4567 	 * Set valid range.  This is typically the entire buffer and thus the
4568 	 * entire page.
4569 	 */
4570 	if (eoff > soff) {
4571 		vm_page_set_validclean(
4572 		    m,
4573 		   (vm_offset_t) (soff & PAGE_MASK),
4574 		   (vm_offset_t) (eoff - soff)
4575 		);
4576 	}
4577 }
4578 
4579 /*
4580  * Acquire a shared busy on all pages in the buf.
4581  */
4582 void
4583 vfs_busy_pages_acquire(struct buf *bp)
4584 {
4585 	int i;
4586 
4587 	for (i = 0; i < bp->b_npages; i++)
4588 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4589 }
4590 
4591 void
4592 vfs_busy_pages_release(struct buf *bp)
4593 {
4594 	int i;
4595 
4596 	for (i = 0; i < bp->b_npages; i++)
4597 		vm_page_sunbusy(bp->b_pages[i]);
4598 }
4599 
4600 /*
4601  * This routine is called before a device strategy routine.
4602  * It is used to tell the VM system that paging I/O is in
4603  * progress, and treat the pages associated with the buffer
4604  * almost as being exclusive busy.  Also the object paging_in_progress
4605  * flag is handled to make sure that the object doesn't become
4606  * inconsistent.
4607  *
4608  * Since I/O has not been initiated yet, certain buffer flags
4609  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4610  * and should be ignored.
4611  */
4612 void
4613 vfs_busy_pages(struct buf *bp, int clear_modify)
4614 {
4615 	vm_object_t obj;
4616 	vm_ooffset_t foff;
4617 	vm_page_t m;
4618 	int i;
4619 	bool bogus;
4620 
4621 	if (!(bp->b_flags & B_VMIO))
4622 		return;
4623 
4624 	obj = bp->b_bufobj->bo_object;
4625 	foff = bp->b_offset;
4626 	KASSERT(bp->b_offset != NOOFFSET,
4627 	    ("vfs_busy_pages: no buffer offset"));
4628 	if ((bp->b_flags & B_CLUSTER) == 0) {
4629 		vm_object_pip_add(obj, bp->b_npages);
4630 		vfs_busy_pages_acquire(bp);
4631 	}
4632 	if (bp->b_bufsize != 0)
4633 		vfs_setdirty_range(bp);
4634 	bogus = false;
4635 	for (i = 0; i < bp->b_npages; i++) {
4636 		m = bp->b_pages[i];
4637 		vm_page_assert_sbusied(m);
4638 
4639 		/*
4640 		 * When readying a buffer for a read ( i.e
4641 		 * clear_modify == 0 ), it is important to do
4642 		 * bogus_page replacement for valid pages in
4643 		 * partially instantiated buffers.  Partially
4644 		 * instantiated buffers can, in turn, occur when
4645 		 * reconstituting a buffer from its VM backing store
4646 		 * base.  We only have to do this if B_CACHE is
4647 		 * clear ( which causes the I/O to occur in the
4648 		 * first place ).  The replacement prevents the read
4649 		 * I/O from overwriting potentially dirty VM-backed
4650 		 * pages.  XXX bogus page replacement is, uh, bogus.
4651 		 * It may not work properly with small-block devices.
4652 		 * We need to find a better way.
4653 		 */
4654 		if (clear_modify) {
4655 			pmap_remove_write(m);
4656 			vfs_page_set_validclean(bp, foff, m);
4657 		} else if (vm_page_all_valid(m) &&
4658 		    (bp->b_flags & B_CACHE) == 0) {
4659 			bp->b_pages[i] = bogus_page;
4660 			bogus = true;
4661 		}
4662 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4663 	}
4664 	if (bogus && buf_mapped(bp)) {
4665 		BUF_CHECK_MAPPED(bp);
4666 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4667 		    bp->b_pages, bp->b_npages);
4668 	}
4669 }
4670 
4671 /*
4672  *	vfs_bio_set_valid:
4673  *
4674  *	Set the range within the buffer to valid.  The range is
4675  *	relative to the beginning of the buffer, b_offset.  Note that
4676  *	b_offset itself may be offset from the beginning of the first
4677  *	page.
4678  */
4679 void
4680 vfs_bio_set_valid(struct buf *bp, int base, int size)
4681 {
4682 	int i, n;
4683 	vm_page_t m;
4684 
4685 	if (!(bp->b_flags & B_VMIO))
4686 		return;
4687 
4688 	/*
4689 	 * Fixup base to be relative to beginning of first page.
4690 	 * Set initial n to be the maximum number of bytes in the
4691 	 * first page that can be validated.
4692 	 */
4693 	base += (bp->b_offset & PAGE_MASK);
4694 	n = PAGE_SIZE - (base & PAGE_MASK);
4695 
4696 	/*
4697 	 * Busy may not be strictly necessary here because the pages are
4698 	 * unlikely to be fully valid and the vnode lock will synchronize
4699 	 * their access via getpages.  It is grabbed for consistency with
4700 	 * other page validation.
4701 	 */
4702 	vfs_busy_pages_acquire(bp);
4703 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4704 		m = bp->b_pages[i];
4705 		if (n > size)
4706 			n = size;
4707 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4708 		base += n;
4709 		size -= n;
4710 		n = PAGE_SIZE;
4711 	}
4712 	vfs_busy_pages_release(bp);
4713 }
4714 
4715 /*
4716  *	vfs_bio_clrbuf:
4717  *
4718  *	If the specified buffer is a non-VMIO buffer, clear the entire
4719  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4720  *	validate only the previously invalid portions of the buffer.
4721  *	This routine essentially fakes an I/O, so we need to clear
4722  *	BIO_ERROR and B_INVAL.
4723  *
4724  *	Note that while we only theoretically need to clear through b_bcount,
4725  *	we go ahead and clear through b_bufsize.
4726  */
4727 void
4728 vfs_bio_clrbuf(struct buf *bp)
4729 {
4730 	int i, j, mask, sa, ea, slide;
4731 
4732 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4733 		clrbuf(bp);
4734 		return;
4735 	}
4736 	bp->b_flags &= ~B_INVAL;
4737 	bp->b_ioflags &= ~BIO_ERROR;
4738 	vfs_busy_pages_acquire(bp);
4739 	sa = bp->b_offset & PAGE_MASK;
4740 	slide = 0;
4741 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4742 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4743 		ea = slide & PAGE_MASK;
4744 		if (ea == 0)
4745 			ea = PAGE_SIZE;
4746 		if (bp->b_pages[i] == bogus_page)
4747 			continue;
4748 		j = sa / DEV_BSIZE;
4749 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4750 		if ((bp->b_pages[i]->valid & mask) == mask)
4751 			continue;
4752 		if ((bp->b_pages[i]->valid & mask) == 0)
4753 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4754 		else {
4755 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4756 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4757 					pmap_zero_page_area(bp->b_pages[i],
4758 					    sa, DEV_BSIZE);
4759 				}
4760 			}
4761 		}
4762 		vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4763 		    roundup2(ea - sa, DEV_BSIZE));
4764 	}
4765 	vfs_busy_pages_release(bp);
4766 	bp->b_resid = 0;
4767 }
4768 
4769 void
4770 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4771 {
4772 	vm_page_t m;
4773 	int i, n;
4774 
4775 	if (buf_mapped(bp)) {
4776 		BUF_CHECK_MAPPED(bp);
4777 		bzero(bp->b_data + base, size);
4778 	} else {
4779 		BUF_CHECK_UNMAPPED(bp);
4780 		n = PAGE_SIZE - (base & PAGE_MASK);
4781 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4782 			m = bp->b_pages[i];
4783 			if (n > size)
4784 				n = size;
4785 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4786 			base += n;
4787 			size -= n;
4788 			n = PAGE_SIZE;
4789 		}
4790 	}
4791 }
4792 
4793 /*
4794  * Update buffer flags based on I/O request parameters, optionally releasing the
4795  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4796  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4797  * I/O).  Otherwise the buffer is released to the cache.
4798  */
4799 static void
4800 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4801 {
4802 
4803 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4804 	    ("buf %p non-VMIO noreuse", bp));
4805 
4806 	if ((ioflag & IO_DIRECT) != 0)
4807 		bp->b_flags |= B_DIRECT;
4808 	if ((ioflag & IO_EXT) != 0)
4809 		bp->b_xflags |= BX_ALTDATA;
4810 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4811 		bp->b_flags |= B_RELBUF;
4812 		if ((ioflag & IO_NOREUSE) != 0)
4813 			bp->b_flags |= B_NOREUSE;
4814 		if (release)
4815 			brelse(bp);
4816 	} else if (release)
4817 		bqrelse(bp);
4818 }
4819 
4820 void
4821 vfs_bio_brelse(struct buf *bp, int ioflag)
4822 {
4823 
4824 	b_io_dismiss(bp, ioflag, true);
4825 }
4826 
4827 void
4828 vfs_bio_set_flags(struct buf *bp, int ioflag)
4829 {
4830 
4831 	b_io_dismiss(bp, ioflag, false);
4832 }
4833 
4834 /*
4835  * vm_hold_load_pages and vm_hold_free_pages get pages into
4836  * a buffers address space.  The pages are anonymous and are
4837  * not associated with a file object.
4838  */
4839 static void
4840 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4841 {
4842 	vm_offset_t pg;
4843 	vm_page_t p;
4844 	int index;
4845 
4846 	BUF_CHECK_MAPPED(bp);
4847 
4848 	to = round_page(to);
4849 	from = round_page(from);
4850 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4851 
4852 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4853 		/*
4854 		 * note: must allocate system pages since blocking here
4855 		 * could interfere with paging I/O, no matter which
4856 		 * process we are.
4857 		 */
4858 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4859 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4860 		    VM_ALLOC_WAITOK);
4861 		pmap_qenter(pg, &p, 1);
4862 		bp->b_pages[index] = p;
4863 	}
4864 	bp->b_npages = index;
4865 }
4866 
4867 /* Return pages associated with this buf to the vm system */
4868 static void
4869 vm_hold_free_pages(struct buf *bp, int newbsize)
4870 {
4871 	vm_offset_t from;
4872 	vm_page_t p;
4873 	int index, newnpages;
4874 
4875 	BUF_CHECK_MAPPED(bp);
4876 
4877 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4878 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4879 	if (bp->b_npages > newnpages)
4880 		pmap_qremove(from, bp->b_npages - newnpages);
4881 	for (index = newnpages; index < bp->b_npages; index++) {
4882 		p = bp->b_pages[index];
4883 		bp->b_pages[index] = NULL;
4884 		vm_page_unwire_noq(p);
4885 		vm_page_free(p);
4886 	}
4887 	bp->b_npages = newnpages;
4888 }
4889 
4890 /*
4891  * Map an IO request into kernel virtual address space.
4892  *
4893  * All requests are (re)mapped into kernel VA space.
4894  * Notice that we use b_bufsize for the size of the buffer
4895  * to be mapped.  b_bcount might be modified by the driver.
4896  *
4897  * Note that even if the caller determines that the address space should
4898  * be valid, a race or a smaller-file mapped into a larger space may
4899  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4900  * check the return value.
4901  *
4902  * This function only works with pager buffers.
4903  */
4904 int
4905 vmapbuf(struct buf *bp, int mapbuf)
4906 {
4907 	vm_prot_t prot;
4908 	int pidx;
4909 
4910 	if (bp->b_bufsize < 0)
4911 		return (-1);
4912 	prot = VM_PROT_READ;
4913 	if (bp->b_iocmd == BIO_READ)
4914 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4915 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4916 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4917 	    btoc(MAXPHYS))) < 0)
4918 		return (-1);
4919 	bp->b_npages = pidx;
4920 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4921 	if (mapbuf || !unmapped_buf_allowed) {
4922 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4923 		bp->b_data = bp->b_kvabase + bp->b_offset;
4924 	} else
4925 		bp->b_data = unmapped_buf;
4926 	return(0);
4927 }
4928 
4929 /*
4930  * Free the io map PTEs associated with this IO operation.
4931  * We also invalidate the TLB entries and restore the original b_addr.
4932  *
4933  * This function only works with pager buffers.
4934  */
4935 void
4936 vunmapbuf(struct buf *bp)
4937 {
4938 	int npages;
4939 
4940 	npages = bp->b_npages;
4941 	if (buf_mapped(bp))
4942 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4943 	vm_page_unhold_pages(bp->b_pages, npages);
4944 
4945 	bp->b_data = unmapped_buf;
4946 }
4947 
4948 void
4949 bdone(struct buf *bp)
4950 {
4951 	struct mtx *mtxp;
4952 
4953 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4954 	mtx_lock(mtxp);
4955 	bp->b_flags |= B_DONE;
4956 	wakeup(bp);
4957 	mtx_unlock(mtxp);
4958 }
4959 
4960 void
4961 bwait(struct buf *bp, u_char pri, const char *wchan)
4962 {
4963 	struct mtx *mtxp;
4964 
4965 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4966 	mtx_lock(mtxp);
4967 	while ((bp->b_flags & B_DONE) == 0)
4968 		msleep(bp, mtxp, pri, wchan, 0);
4969 	mtx_unlock(mtxp);
4970 }
4971 
4972 int
4973 bufsync(struct bufobj *bo, int waitfor)
4974 {
4975 
4976 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
4977 }
4978 
4979 void
4980 bufstrategy(struct bufobj *bo, struct buf *bp)
4981 {
4982 	int i __unused;
4983 	struct vnode *vp;
4984 
4985 	vp = bp->b_vp;
4986 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4987 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4988 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4989 	i = VOP_STRATEGY(vp, bp);
4990 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4991 }
4992 
4993 /*
4994  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
4995  */
4996 void
4997 bufobj_init(struct bufobj *bo, void *private)
4998 {
4999 	static volatile int bufobj_cleanq;
5000 
5001         bo->bo_domain =
5002             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5003         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5004         bo->bo_private = private;
5005         TAILQ_INIT(&bo->bo_clean.bv_hd);
5006         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5007 }
5008 
5009 void
5010 bufobj_wrefl(struct bufobj *bo)
5011 {
5012 
5013 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5014 	ASSERT_BO_WLOCKED(bo);
5015 	bo->bo_numoutput++;
5016 }
5017 
5018 void
5019 bufobj_wref(struct bufobj *bo)
5020 {
5021 
5022 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5023 	BO_LOCK(bo);
5024 	bo->bo_numoutput++;
5025 	BO_UNLOCK(bo);
5026 }
5027 
5028 void
5029 bufobj_wdrop(struct bufobj *bo)
5030 {
5031 
5032 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5033 	BO_LOCK(bo);
5034 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5035 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5036 		bo->bo_flag &= ~BO_WWAIT;
5037 		wakeup(&bo->bo_numoutput);
5038 	}
5039 	BO_UNLOCK(bo);
5040 }
5041 
5042 int
5043 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5044 {
5045 	int error;
5046 
5047 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5048 	ASSERT_BO_WLOCKED(bo);
5049 	error = 0;
5050 	while (bo->bo_numoutput) {
5051 		bo->bo_flag |= BO_WWAIT;
5052 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5053 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5054 		if (error)
5055 			break;
5056 	}
5057 	return (error);
5058 }
5059 
5060 /*
5061  * Set bio_data or bio_ma for struct bio from the struct buf.
5062  */
5063 void
5064 bdata2bio(struct buf *bp, struct bio *bip)
5065 {
5066 
5067 	if (!buf_mapped(bp)) {
5068 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5069 		bip->bio_ma = bp->b_pages;
5070 		bip->bio_ma_n = bp->b_npages;
5071 		bip->bio_data = unmapped_buf;
5072 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5073 		bip->bio_flags |= BIO_UNMAPPED;
5074 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5075 		    PAGE_SIZE == bp->b_npages,
5076 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5077 		    (long long)bip->bio_length, bip->bio_ma_n));
5078 	} else {
5079 		bip->bio_data = bp->b_data;
5080 		bip->bio_ma = NULL;
5081 	}
5082 }
5083 
5084 /*
5085  * The MIPS pmap code currently doesn't handle aliased pages.
5086  * The VIPT caches may not handle page aliasing themselves, leading
5087  * to data corruption.
5088  *
5089  * As such, this code makes a system extremely unhappy if said
5090  * system doesn't support unaliasing the above situation in hardware.
5091  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5092  * this feature at build time, so it has to be handled in software.
5093  *
5094  * Once the MIPS pmap/cache code grows to support this function on
5095  * earlier chips, it should be flipped back off.
5096  */
5097 #ifdef	__mips__
5098 static int buf_pager_relbuf = 1;
5099 #else
5100 static int buf_pager_relbuf = 0;
5101 #endif
5102 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5103     &buf_pager_relbuf, 0,
5104     "Make buffer pager release buffers after reading");
5105 
5106 /*
5107  * The buffer pager.  It uses buffer reads to validate pages.
5108  *
5109  * In contrast to the generic local pager from vm/vnode_pager.c, this
5110  * pager correctly and easily handles volumes where the underlying
5111  * device block size is greater than the machine page size.  The
5112  * buffer cache transparently extends the requested page run to be
5113  * aligned at the block boundary, and does the necessary bogus page
5114  * replacements in the addends to avoid obliterating already valid
5115  * pages.
5116  *
5117  * The only non-trivial issue is that the exclusive busy state for
5118  * pages, which is assumed by the vm_pager_getpages() interface, is
5119  * incompatible with the VMIO buffer cache's desire to share-busy the
5120  * pages.  This function performs a trivial downgrade of the pages'
5121  * state before reading buffers, and a less trivial upgrade from the
5122  * shared-busy to excl-busy state after the read.
5123  */
5124 int
5125 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5126     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5127     vbg_get_blksize_t get_blksize)
5128 {
5129 	vm_page_t m;
5130 	vm_object_t object;
5131 	struct buf *bp;
5132 	struct mount *mp;
5133 	daddr_t lbn, lbnp;
5134 	vm_ooffset_t la, lb, poff, poffe;
5135 	long bsize;
5136 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5137 	bool redo, lpart;
5138 
5139 	object = vp->v_object;
5140 	mp = vp->v_mount;
5141 	error = 0;
5142 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5143 	if (la >= object->un_pager.vnp.vnp_size)
5144 		return (VM_PAGER_BAD);
5145 
5146 	/*
5147 	 * Change the meaning of la from where the last requested page starts
5148 	 * to where it ends, because that's the end of the requested region
5149 	 * and the start of the potential read-ahead region.
5150 	 */
5151 	la += PAGE_SIZE;
5152 	lpart = la > object->un_pager.vnp.vnp_size;
5153 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5154 
5155 	/*
5156 	 * Calculate read-ahead, behind and total pages.
5157 	 */
5158 	pgsin = count;
5159 	lb = IDX_TO_OFF(ma[0]->pindex);
5160 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5161 	pgsin += pgsin_b;
5162 	if (rbehind != NULL)
5163 		*rbehind = pgsin_b;
5164 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5165 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5166 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5167 		    PAGE_SIZE) - la);
5168 	pgsin += pgsin_a;
5169 	if (rahead != NULL)
5170 		*rahead = pgsin_a;
5171 	VM_CNT_INC(v_vnodein);
5172 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5173 
5174 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5175 	    != 0) ? GB_UNMAPPED : 0;
5176 again:
5177 	for (i = 0; i < count; i++)
5178 		vm_page_busy_downgrade(ma[i]);
5179 
5180 	lbnp = -1;
5181 	for (i = 0; i < count; i++) {
5182 		m = ma[i];
5183 
5184 		/*
5185 		 * Pages are shared busy and the object lock is not
5186 		 * owned, which together allow for the pages'
5187 		 * invalidation.  The racy test for validity avoids
5188 		 * useless creation of the buffer for the most typical
5189 		 * case when invalidation is not used in redo or for
5190 		 * parallel read.  The shared->excl upgrade loop at
5191 		 * the end of the function catches the race in a
5192 		 * reliable way (protected by the object lock).
5193 		 */
5194 		if (vm_page_all_valid(m))
5195 			continue;
5196 
5197 		poff = IDX_TO_OFF(m->pindex);
5198 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5199 		for (; poff < poffe; poff += bsize) {
5200 			lbn = get_lblkno(vp, poff);
5201 			if (lbn == lbnp)
5202 				goto next_page;
5203 			lbnp = lbn;
5204 
5205 			bsize = get_blksize(vp, lbn);
5206 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5207 			    br_flags, &bp);
5208 			if (error != 0)
5209 				goto end_pages;
5210 			if (LIST_EMPTY(&bp->b_dep)) {
5211 				/*
5212 				 * Invalidation clears m->valid, but
5213 				 * may leave B_CACHE flag if the
5214 				 * buffer existed at the invalidation
5215 				 * time.  In this case, recycle the
5216 				 * buffer to do real read on next
5217 				 * bread() after redo.
5218 				 *
5219 				 * Otherwise B_RELBUF is not strictly
5220 				 * necessary, enable to reduce buf
5221 				 * cache pressure.
5222 				 */
5223 				if (buf_pager_relbuf ||
5224 				    !vm_page_all_valid(m))
5225 					bp->b_flags |= B_RELBUF;
5226 
5227 				bp->b_flags &= ~B_NOCACHE;
5228 				brelse(bp);
5229 			} else {
5230 				bqrelse(bp);
5231 			}
5232 		}
5233 		KASSERT(1 /* racy, enable for debugging */ ||
5234 		    vm_page_all_valid(m) || i == count - 1,
5235 		    ("buf %d %p invalid", i, m));
5236 		if (i == count - 1 && lpart) {
5237 			if (!vm_page_none_valid(m) &&
5238 			    !vm_page_all_valid(m))
5239 				vm_page_zero_invalid(m, TRUE);
5240 		}
5241 next_page:;
5242 	}
5243 end_pages:
5244 
5245 	VM_OBJECT_WLOCK(object);
5246 	redo = false;
5247 	for (i = 0; i < count; i++) {
5248 		vm_page_sunbusy(ma[i]);
5249 		ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
5250 
5251 		/*
5252 		 * Since the pages were only sbusy while neither the
5253 		 * buffer nor the object lock was held by us, or
5254 		 * reallocated while vm_page_grab() slept for busy
5255 		 * relinguish, they could have been invalidated.
5256 		 * Recheck the valid bits and re-read as needed.
5257 		 *
5258 		 * Note that the last page is made fully valid in the
5259 		 * read loop, and partial validity for the page at
5260 		 * index count - 1 could mean that the page was
5261 		 * invalidated or removed, so we must restart for
5262 		 * safety as well.
5263 		 */
5264 		if (!vm_page_all_valid(ma[i]))
5265 			redo = true;
5266 	}
5267 	VM_OBJECT_WUNLOCK(object);
5268 	if (redo && error == 0)
5269 		goto again;
5270 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5271 }
5272 
5273 #include "opt_ddb.h"
5274 #ifdef DDB
5275 #include <ddb/ddb.h>
5276 
5277 /* DDB command to show buffer data */
5278 DB_SHOW_COMMAND(buffer, db_show_buffer)
5279 {
5280 	/* get args */
5281 	struct buf *bp = (struct buf *)addr;
5282 #ifdef FULL_BUF_TRACKING
5283 	uint32_t i, j;
5284 #endif
5285 
5286 	if (!have_addr) {
5287 		db_printf("usage: show buffer <addr>\n");
5288 		return;
5289 	}
5290 
5291 	db_printf("buf at %p\n", bp);
5292 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5293 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5294 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5295 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5296 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5297 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5298 	db_printf(
5299 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5300 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5301 	    "b_vp = %p, b_dep = %p\n",
5302 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5303 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5304 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5305 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5306 	    bp->b_kvabase, bp->b_kvasize);
5307 	if (bp->b_npages) {
5308 		int i;
5309 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5310 		for (i = 0; i < bp->b_npages; i++) {
5311 			vm_page_t m;
5312 			m = bp->b_pages[i];
5313 			if (m != NULL)
5314 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5315 				    (u_long)m->pindex,
5316 				    (u_long)VM_PAGE_TO_PHYS(m));
5317 			else
5318 				db_printf("( ??? )");
5319 			if ((i + 1) < bp->b_npages)
5320 				db_printf(",");
5321 		}
5322 		db_printf("\n");
5323 	}
5324 	BUF_LOCKPRINTINFO(bp);
5325 #if defined(FULL_BUF_TRACKING)
5326 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5327 
5328 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5329 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5330 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5331 			continue;
5332 		db_printf(" %2u: %s\n", j,
5333 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5334 	}
5335 #elif defined(BUF_TRACKING)
5336 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5337 #endif
5338 	db_printf(" ");
5339 }
5340 
5341 DB_SHOW_COMMAND(bufqueues, bufqueues)
5342 {
5343 	struct bufdomain *bd;
5344 	struct buf *bp;
5345 	long total;
5346 	int i, j, cnt;
5347 
5348 	db_printf("bqempty: %d\n", bqempty.bq_len);
5349 
5350 	for (i = 0; i < buf_domains; i++) {
5351 		bd = &bdomain[i];
5352 		db_printf("Buf domain %d\n", i);
5353 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5354 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5355 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5356 		db_printf("\n");
5357 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5358 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5359 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5360 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5361 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5362 		db_printf("\n");
5363 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5364 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5365 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5366 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5367 		db_printf("\n");
5368 		total = 0;
5369 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5370 			total += bp->b_bufsize;
5371 		db_printf("\tcleanq count\t%d (%ld)\n",
5372 		    bd->bd_cleanq->bq_len, total);
5373 		total = 0;
5374 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5375 			total += bp->b_bufsize;
5376 		db_printf("\tdirtyq count\t%d (%ld)\n",
5377 		    bd->bd_dirtyq.bq_len, total);
5378 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5379 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5380 		db_printf("\tCPU ");
5381 		for (j = 0; j <= mp_maxid; j++)
5382 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5383 		db_printf("\n");
5384 		cnt = 0;
5385 		total = 0;
5386 		for (j = 0; j < nbuf; j++)
5387 			if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5388 				cnt++;
5389 				total += buf[j].b_bufsize;
5390 			}
5391 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5392 		cnt = 0;
5393 		total = 0;
5394 		for (j = 0; j < nbuf; j++)
5395 			if (buf[j].b_domain == i) {
5396 				cnt++;
5397 				total += buf[j].b_bufsize;
5398 			}
5399 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5400 	}
5401 }
5402 
5403 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5404 {
5405 	struct buf *bp;
5406 	int i;
5407 
5408 	for (i = 0; i < nbuf; i++) {
5409 		bp = &buf[i];
5410 		if (BUF_ISLOCKED(bp)) {
5411 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5412 			db_printf("\n");
5413 			if (db_pager_quit)
5414 				break;
5415 		}
5416 	}
5417 }
5418 
5419 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5420 {
5421 	struct vnode *vp;
5422 	struct buf *bp;
5423 
5424 	if (!have_addr) {
5425 		db_printf("usage: show vnodebufs <addr>\n");
5426 		return;
5427 	}
5428 	vp = (struct vnode *)addr;
5429 	db_printf("Clean buffers:\n");
5430 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5431 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5432 		db_printf("\n");
5433 	}
5434 	db_printf("Dirty buffers:\n");
5435 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5436 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5437 		db_printf("\n");
5438 	}
5439 }
5440 
5441 DB_COMMAND(countfreebufs, db_coundfreebufs)
5442 {
5443 	struct buf *bp;
5444 	int i, used = 0, nfree = 0;
5445 
5446 	if (have_addr) {
5447 		db_printf("usage: countfreebufs\n");
5448 		return;
5449 	}
5450 
5451 	for (i = 0; i < nbuf; i++) {
5452 		bp = &buf[i];
5453 		if (bp->b_qindex == QUEUE_EMPTY)
5454 			nfree++;
5455 		else
5456 			used++;
5457 	}
5458 
5459 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5460 	    nfree + used);
5461 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5462 }
5463 #endif /* DDB */
5464