xref: /freebsd/sys/kern/vfs_bio.c (revision a0409676120c1e558d0ade943019934e0f15118d)
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/syscallsubr.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 char *buf;		/* buffer header pool */
151 static struct buf *
152 nbufp(unsigned i)
153 {
154 	return ((struct buf *)(buf + (sizeof(struct buf) +
155 	    sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
156 }
157 
158 caddr_t __read_mostly unmapped_buf;
159 
160 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
161 struct proc *bufdaemonproc;
162 
163 static void vm_hold_free_pages(struct buf *bp, int newbsize);
164 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
165 		vm_offset_t to);
166 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
167 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
168 		vm_page_t m);
169 static void vfs_clean_pages_dirty_buf(struct buf *bp);
170 static void vfs_setdirty_range(struct buf *bp);
171 static void vfs_vmio_invalidate(struct buf *bp);
172 static void vfs_vmio_truncate(struct buf *bp, int npages);
173 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
174 static int vfs_bio_clcheck(struct vnode *vp, int size,
175 		daddr_t lblkno, daddr_t blkno);
176 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
177 		void (*)(struct buf *));
178 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
179 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
180 static void buf_daemon(void);
181 static __inline void bd_wakeup(void);
182 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
183 static void bufkva_reclaim(vmem_t *, int);
184 static void bufkva_free(struct buf *);
185 static int buf_import(void *, void **, int, int, int);
186 static void buf_release(void *, void **, int);
187 static void maxbcachebuf_adjust(void);
188 static inline struct bufdomain *bufdomain(struct buf *);
189 static void bq_remove(struct bufqueue *bq, struct buf *bp);
190 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
191 static int buf_recycle(struct bufdomain *, bool kva);
192 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
193 	    const char *lockname);
194 static void bd_init(struct bufdomain *bd);
195 static int bd_flushall(struct bufdomain *bd);
196 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
197 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
198 
199 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
200 int vmiodirenable = TRUE;
201 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
202     "Use the VM system for directory writes");
203 long runningbufspace;
204 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
205     "Amount of presently outstanding async buffer io");
206 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
207     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
208 static counter_u64_t bufkvaspace;
209 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
210     "Kernel virtual memory used for buffers");
211 static long maxbufspace;
212 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
213     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
214     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
215     "Maximum allowed value of bufspace (including metadata)");
216 static long bufmallocspace;
217 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
218     "Amount of malloced memory for buffers");
219 static long maxbufmallocspace;
220 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
221     0, "Maximum amount of malloced memory for buffers");
222 static long lobufspace;
223 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
224     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
225     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
226     "Minimum amount of buffers we want to have");
227 long hibufspace;
228 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
229     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
230     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
231     "Maximum allowed value of bufspace (excluding metadata)");
232 long bufspacethresh;
233 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
234     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
235     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
236     "Bufspace consumed before waking the daemon to free some");
237 static counter_u64_t buffreekvacnt;
238 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
239     "Number of times we have freed the KVA space from some buffer");
240 static counter_u64_t bufdefragcnt;
241 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
242     "Number of times we have had to repeat buffer allocation to defragment");
243 static long lorunningspace;
244 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
245     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
246     "Minimum preferred space used for in-progress I/O");
247 static long hirunningspace;
248 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
249     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
250     "Maximum amount of space to use for in-progress I/O");
251 int dirtybufferflushes;
252 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
253     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
254 int bdwriteskip;
255 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
256     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
257 int altbufferflushes;
258 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
259     &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
260 static int recursiveflushes;
261 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
262     &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
263 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
264 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
265     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
266     "Number of buffers that are dirty (has unwritten changes) at the moment");
267 static int lodirtybuffers;
268 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
269     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
270     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
271     "How many buffers we want to have free before bufdaemon can sleep");
272 static int hidirtybuffers;
273 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
274     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
275     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
276     "When the number of dirty buffers is considered severe");
277 int dirtybufthresh;
278 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
279     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
280     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
281     "Number of bdwrite to bawrite conversions to clear dirty buffers");
282 static int numfreebuffers;
283 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
284     "Number of free buffers");
285 static int lofreebuffers;
286 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
287     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
288     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
289    "Target number of free buffers");
290 static int hifreebuffers;
291 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
292     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
293     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
294    "Threshold for clean buffer recycling");
295 static counter_u64_t getnewbufcalls;
296 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
297    &getnewbufcalls, "Number of calls to getnewbuf");
298 static counter_u64_t getnewbufrestarts;
299 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
300     &getnewbufrestarts,
301     "Number of times getnewbuf has had to restart a buffer acquisition");
302 static counter_u64_t mappingrestarts;
303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
304     &mappingrestarts,
305     "Number of times getblk has had to restart a buffer mapping for "
306     "unmapped buffer");
307 static counter_u64_t numbufallocfails;
308 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
309     &numbufallocfails, "Number of times buffer allocations failed");
310 static int flushbufqtarget = 100;
311 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
312     "Amount of work to do in flushbufqueues when helping bufdaemon");
313 static counter_u64_t notbufdflushes;
314 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
315     "Number of dirty buffer flushes done by the bufdaemon helpers");
316 static long barrierwrites;
317 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
318     &barrierwrites, 0, "Number of barrier writes");
319 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
320     &unmapped_buf_allowed, 0,
321     "Permit the use of the unmapped i/o");
322 int maxbcachebuf = MAXBCACHEBUF;
323 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
324     "Maximum size of a buffer cache block");
325 
326 /*
327  * This lock synchronizes access to bd_request.
328  */
329 static struct mtx_padalign __exclusive_cache_line bdlock;
330 
331 /*
332  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
333  * waitrunningbufspace().
334  */
335 static struct mtx_padalign __exclusive_cache_line rbreqlock;
336 
337 /*
338  * Lock that protects bdirtywait.
339  */
340 static struct mtx_padalign __exclusive_cache_line bdirtylock;
341 
342 /*
343  * Wakeup point for bufdaemon, as well as indicator of whether it is already
344  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
345  * is idling.
346  */
347 static int bd_request;
348 
349 /*
350  * Request for the buf daemon to write more buffers than is indicated by
351  * lodirtybuf.  This may be necessary to push out excess dependencies or
352  * defragment the address space where a simple count of the number of dirty
353  * buffers is insufficient to characterize the demand for flushing them.
354  */
355 static int bd_speedupreq;
356 
357 /*
358  * Synchronization (sleep/wakeup) variable for active buffer space requests.
359  * Set when wait starts, cleared prior to wakeup().
360  * Used in runningbufwakeup() and waitrunningbufspace().
361  */
362 static int runningbufreq;
363 
364 /*
365  * Synchronization for bwillwrite() waiters.
366  */
367 static int bdirtywait;
368 
369 /*
370  * Definitions for the buffer free lists.
371  */
372 #define QUEUE_NONE	0	/* on no queue */
373 #define QUEUE_EMPTY	1	/* empty buffer headers */
374 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
375 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
376 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
377 
378 /* Maximum number of buffer domains. */
379 #define	BUF_DOMAINS	8
380 
381 struct bufdomainset bdlodirty;		/* Domains > lodirty */
382 struct bufdomainset bdhidirty;		/* Domains > hidirty */
383 
384 /* Configured number of clean queues. */
385 static int __read_mostly buf_domains;
386 
387 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
388 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
389 struct bufqueue __exclusive_cache_line bqempty;
390 
391 /*
392  * per-cpu empty buffer cache.
393  */
394 uma_zone_t buf_zone;
395 
396 /*
397  * Single global constant for BUF_WMESG, to avoid getting multiple references.
398  * buf_wmesg is referred from macros.
399  */
400 const char *buf_wmesg = BUF_WMESG;
401 
402 static int
403 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
404 {
405 	long value;
406 	int error;
407 
408 	value = *(long *)arg1;
409 	error = sysctl_handle_long(oidp, &value, 0, req);
410 	if (error != 0 || req->newptr == NULL)
411 		return (error);
412 	mtx_lock(&rbreqlock);
413 	if (arg1 == &hirunningspace) {
414 		if (value < lorunningspace)
415 			error = EINVAL;
416 		else
417 			hirunningspace = value;
418 	} else {
419 		KASSERT(arg1 == &lorunningspace,
420 		    ("%s: unknown arg1", __func__));
421 		if (value > hirunningspace)
422 			error = EINVAL;
423 		else
424 			lorunningspace = value;
425 	}
426 	mtx_unlock(&rbreqlock);
427 	return (error);
428 }
429 
430 static int
431 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
432 {
433 	int error;
434 	int value;
435 	int i;
436 
437 	value = *(int *)arg1;
438 	error = sysctl_handle_int(oidp, &value, 0, req);
439 	if (error != 0 || req->newptr == NULL)
440 		return (error);
441 	*(int *)arg1 = value;
442 	for (i = 0; i < buf_domains; i++)
443 		*(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
444 		    value / buf_domains;
445 
446 	return (error);
447 }
448 
449 static int
450 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
451 {
452 	long value;
453 	int error;
454 	int i;
455 
456 	value = *(long *)arg1;
457 	error = sysctl_handle_long(oidp, &value, 0, req);
458 	if (error != 0 || req->newptr == NULL)
459 		return (error);
460 	*(long *)arg1 = value;
461 	for (i = 0; i < buf_domains; i++)
462 		*(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
463 		    value / buf_domains;
464 
465 	return (error);
466 }
467 
468 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
469     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
470 static int
471 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
472 {
473 	long lvalue;
474 	int ivalue;
475 	int i;
476 
477 	lvalue = 0;
478 	for (i = 0; i < buf_domains; i++)
479 		lvalue += bdomain[i].bd_bufspace;
480 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
481 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
482 	if (lvalue > INT_MAX)
483 		/* On overflow, still write out a long to trigger ENOMEM. */
484 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
485 	ivalue = lvalue;
486 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
487 }
488 #else
489 static int
490 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
491 {
492 	long lvalue;
493 	int i;
494 
495 	lvalue = 0;
496 	for (i = 0; i < buf_domains; i++)
497 		lvalue += bdomain[i].bd_bufspace;
498 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
499 }
500 #endif
501 
502 static int
503 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
504 {
505 	int value;
506 	int i;
507 
508 	value = 0;
509 	for (i = 0; i < buf_domains; i++)
510 		value += bdomain[i].bd_numdirtybuffers;
511 	return (sysctl_handle_int(oidp, &value, 0, req));
512 }
513 
514 /*
515  *	bdirtywakeup:
516  *
517  *	Wakeup any bwillwrite() waiters.
518  */
519 static void
520 bdirtywakeup(void)
521 {
522 	mtx_lock(&bdirtylock);
523 	if (bdirtywait) {
524 		bdirtywait = 0;
525 		wakeup(&bdirtywait);
526 	}
527 	mtx_unlock(&bdirtylock);
528 }
529 
530 /*
531  *	bd_clear:
532  *
533  *	Clear a domain from the appropriate bitsets when dirtybuffers
534  *	is decremented.
535  */
536 static void
537 bd_clear(struct bufdomain *bd)
538 {
539 
540 	mtx_lock(&bdirtylock);
541 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
542 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
543 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
544 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
545 	mtx_unlock(&bdirtylock);
546 }
547 
548 /*
549  *	bd_set:
550  *
551  *	Set a domain in the appropriate bitsets when dirtybuffers
552  *	is incremented.
553  */
554 static void
555 bd_set(struct bufdomain *bd)
556 {
557 
558 	mtx_lock(&bdirtylock);
559 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
560 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
561 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
562 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
563 	mtx_unlock(&bdirtylock);
564 }
565 
566 /*
567  *	bdirtysub:
568  *
569  *	Decrement the numdirtybuffers count by one and wakeup any
570  *	threads blocked in bwillwrite().
571  */
572 static void
573 bdirtysub(struct buf *bp)
574 {
575 	struct bufdomain *bd;
576 	int num;
577 
578 	bd = bufdomain(bp);
579 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
580 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
581 		bdirtywakeup();
582 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
583 		bd_clear(bd);
584 }
585 
586 /*
587  *	bdirtyadd:
588  *
589  *	Increment the numdirtybuffers count by one and wakeup the buf
590  *	daemon if needed.
591  */
592 static void
593 bdirtyadd(struct buf *bp)
594 {
595 	struct bufdomain *bd;
596 	int num;
597 
598 	/*
599 	 * Only do the wakeup once as we cross the boundary.  The
600 	 * buf daemon will keep running until the condition clears.
601 	 */
602 	bd = bufdomain(bp);
603 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
604 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
605 		bd_wakeup();
606 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
607 		bd_set(bd);
608 }
609 
610 /*
611  *	bufspace_daemon_wakeup:
612  *
613  *	Wakeup the daemons responsible for freeing clean bufs.
614  */
615 static void
616 bufspace_daemon_wakeup(struct bufdomain *bd)
617 {
618 
619 	/*
620 	 * avoid the lock if the daemon is running.
621 	 */
622 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
623 		BD_RUN_LOCK(bd);
624 		atomic_store_int(&bd->bd_running, 1);
625 		wakeup(&bd->bd_running);
626 		BD_RUN_UNLOCK(bd);
627 	}
628 }
629 
630 /*
631  *	bufspace_daemon_wait:
632  *
633  *	Sleep until the domain falls below a limit or one second passes.
634  */
635 static void
636 bufspace_daemon_wait(struct bufdomain *bd)
637 {
638 	/*
639 	 * Re-check our limits and sleep.  bd_running must be
640 	 * cleared prior to checking the limits to avoid missed
641 	 * wakeups.  The waker will adjust one of bufspace or
642 	 * freebuffers prior to checking bd_running.
643 	 */
644 	BD_RUN_LOCK(bd);
645 	atomic_store_int(&bd->bd_running, 0);
646 	if (bd->bd_bufspace < bd->bd_bufspacethresh &&
647 	    bd->bd_freebuffers > bd->bd_lofreebuffers) {
648 		msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
649 		    "-", hz);
650 	} else {
651 		/* Avoid spurious wakeups while running. */
652 		atomic_store_int(&bd->bd_running, 1);
653 		BD_RUN_UNLOCK(bd);
654 	}
655 }
656 
657 /*
658  *	bufspace_adjust:
659  *
660  *	Adjust the reported bufspace for a KVA managed buffer, possibly
661  * 	waking any waiters.
662  */
663 static void
664 bufspace_adjust(struct buf *bp, int bufsize)
665 {
666 	struct bufdomain *bd;
667 	long space;
668 	int diff;
669 
670 	KASSERT((bp->b_flags & B_MALLOC) == 0,
671 	    ("bufspace_adjust: malloc buf %p", bp));
672 	bd = bufdomain(bp);
673 	diff = bufsize - bp->b_bufsize;
674 	if (diff < 0) {
675 		atomic_subtract_long(&bd->bd_bufspace, -diff);
676 	} else if (diff > 0) {
677 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
678 		/* Wake up the daemon on the transition. */
679 		if (space < bd->bd_bufspacethresh &&
680 		    space + diff >= bd->bd_bufspacethresh)
681 			bufspace_daemon_wakeup(bd);
682 	}
683 	bp->b_bufsize = bufsize;
684 }
685 
686 /*
687  *	bufspace_reserve:
688  *
689  *	Reserve bufspace before calling allocbuf().  metadata has a
690  *	different space limit than data.
691  */
692 static int
693 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
694 {
695 	long limit, new;
696 	long space;
697 
698 	if (metadata)
699 		limit = bd->bd_maxbufspace;
700 	else
701 		limit = bd->bd_hibufspace;
702 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
703 	new = space + size;
704 	if (new > limit) {
705 		atomic_subtract_long(&bd->bd_bufspace, size);
706 		return (ENOSPC);
707 	}
708 
709 	/* Wake up the daemon on the transition. */
710 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
711 		bufspace_daemon_wakeup(bd);
712 
713 	return (0);
714 }
715 
716 /*
717  *	bufspace_release:
718  *
719  *	Release reserved bufspace after bufspace_adjust() has consumed it.
720  */
721 static void
722 bufspace_release(struct bufdomain *bd, int size)
723 {
724 
725 	atomic_subtract_long(&bd->bd_bufspace, size);
726 }
727 
728 /*
729  *	bufspace_wait:
730  *
731  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
732  *	supplied.  bd_wanted must be set prior to polling for space.  The
733  *	operation must be re-tried on return.
734  */
735 static void
736 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
737     int slpflag, int slptimeo)
738 {
739 	struct thread *td;
740 	int error, fl, norunbuf;
741 
742 	if ((gbflags & GB_NOWAIT_BD) != 0)
743 		return;
744 
745 	td = curthread;
746 	BD_LOCK(bd);
747 	while (bd->bd_wanted) {
748 		if (vp != NULL && vp->v_type != VCHR &&
749 		    (td->td_pflags & TDP_BUFNEED) == 0) {
750 			BD_UNLOCK(bd);
751 			/*
752 			 * getblk() is called with a vnode locked, and
753 			 * some majority of the dirty buffers may as
754 			 * well belong to the vnode.  Flushing the
755 			 * buffers there would make a progress that
756 			 * cannot be achieved by the buf_daemon, that
757 			 * cannot lock the vnode.
758 			 */
759 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
760 			    (td->td_pflags & TDP_NORUNNINGBUF);
761 
762 			/*
763 			 * Play bufdaemon.  The getnewbuf() function
764 			 * may be called while the thread owns lock
765 			 * for another dirty buffer for the same
766 			 * vnode, which makes it impossible to use
767 			 * VOP_FSYNC() there, due to the buffer lock
768 			 * recursion.
769 			 */
770 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
771 			fl = buf_flush(vp, bd, flushbufqtarget);
772 			td->td_pflags &= norunbuf;
773 			BD_LOCK(bd);
774 			if (fl != 0)
775 				continue;
776 			if (bd->bd_wanted == 0)
777 				break;
778 		}
779 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
780 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
781 		if (error != 0)
782 			break;
783 	}
784 	BD_UNLOCK(bd);
785 }
786 
787 /*
788  *	bufspace_daemon:
789  *
790  *	buffer space management daemon.  Tries to maintain some marginal
791  *	amount of free buffer space so that requesting processes neither
792  *	block nor work to reclaim buffers.
793  */
794 static void
795 bufspace_daemon(void *arg)
796 {
797 	struct bufdomain *bd;
798 
799 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
800 	    SHUTDOWN_PRI_LAST + 100);
801 
802 	bd = arg;
803 	for (;;) {
804 		kthread_suspend_check();
805 
806 		/*
807 		 * Free buffers from the clean queue until we meet our
808 		 * targets.
809 		 *
810 		 * Theory of operation:  The buffer cache is most efficient
811 		 * when some free buffer headers and space are always
812 		 * available to getnewbuf().  This daemon attempts to prevent
813 		 * the excessive blocking and synchronization associated
814 		 * with shortfall.  It goes through three phases according
815 		 * demand:
816 		 *
817 		 * 1)	The daemon wakes up voluntarily once per-second
818 		 *	during idle periods when the counters are below
819 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
820 		 *
821 		 * 2)	The daemon wakes up as we cross the thresholds
822 		 *	ahead of any potential blocking.  This may bounce
823 		 *	slightly according to the rate of consumption and
824 		 *	release.
825 		 *
826 		 * 3)	The daemon and consumers are starved for working
827 		 *	clean buffers.  This is the 'bufspace' sleep below
828 		 *	which will inefficiently trade bufs with bqrelse
829 		 *	until we return to condition 2.
830 		 */
831 		while (bd->bd_bufspace > bd->bd_lobufspace ||
832 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
833 			if (buf_recycle(bd, false) != 0) {
834 				if (bd_flushall(bd))
835 					continue;
836 				/*
837 				 * Speedup dirty if we've run out of clean
838 				 * buffers.  This is possible in particular
839 				 * because softdep may held many bufs locked
840 				 * pending writes to other bufs which are
841 				 * marked for delayed write, exhausting
842 				 * clean space until they are written.
843 				 */
844 				bd_speedup();
845 				BD_LOCK(bd);
846 				if (bd->bd_wanted) {
847 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
848 					    PRIBIO|PDROP, "bufspace", hz/10);
849 				} else
850 					BD_UNLOCK(bd);
851 			}
852 			maybe_yield();
853 		}
854 		bufspace_daemon_wait(bd);
855 	}
856 }
857 
858 /*
859  *	bufmallocadjust:
860  *
861  *	Adjust the reported bufspace for a malloc managed buffer, possibly
862  *	waking any waiters.
863  */
864 static void
865 bufmallocadjust(struct buf *bp, int bufsize)
866 {
867 	int diff;
868 
869 	KASSERT((bp->b_flags & B_MALLOC) != 0,
870 	    ("bufmallocadjust: non-malloc buf %p", bp));
871 	diff = bufsize - bp->b_bufsize;
872 	if (diff < 0)
873 		atomic_subtract_long(&bufmallocspace, -diff);
874 	else
875 		atomic_add_long(&bufmallocspace, diff);
876 	bp->b_bufsize = bufsize;
877 }
878 
879 /*
880  *	runningwakeup:
881  *
882  *	Wake up processes that are waiting on asynchronous writes to fall
883  *	below lorunningspace.
884  */
885 static void
886 runningwakeup(void)
887 {
888 
889 	mtx_lock(&rbreqlock);
890 	if (runningbufreq) {
891 		runningbufreq = 0;
892 		wakeup(&runningbufreq);
893 	}
894 	mtx_unlock(&rbreqlock);
895 }
896 
897 /*
898  *	runningbufwakeup:
899  *
900  *	Decrement the outstanding write count according.
901  */
902 void
903 runningbufwakeup(struct buf *bp)
904 {
905 	long space, bspace;
906 
907 	bspace = bp->b_runningbufspace;
908 	if (bspace == 0)
909 		return;
910 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
911 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
912 	    space, bspace));
913 	bp->b_runningbufspace = 0;
914 	/*
915 	 * Only acquire the lock and wakeup on the transition from exceeding
916 	 * the threshold to falling below it.
917 	 */
918 	if (space < lorunningspace)
919 		return;
920 	if (space - bspace > lorunningspace)
921 		return;
922 	runningwakeup();
923 }
924 
925 /*
926  *	waitrunningbufspace()
927  *
928  *	runningbufspace is a measure of the amount of I/O currently
929  *	running.  This routine is used in async-write situations to
930  *	prevent creating huge backups of pending writes to a device.
931  *	Only asynchronous writes are governed by this function.
932  *
933  *	This does NOT turn an async write into a sync write.  It waits
934  *	for earlier writes to complete and generally returns before the
935  *	caller's write has reached the device.
936  */
937 void
938 waitrunningbufspace(void)
939 {
940 
941 	mtx_lock(&rbreqlock);
942 	while (runningbufspace > hirunningspace) {
943 		runningbufreq = 1;
944 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
945 	}
946 	mtx_unlock(&rbreqlock);
947 }
948 
949 /*
950  *	vfs_buf_test_cache:
951  *
952  *	Called when a buffer is extended.  This function clears the B_CACHE
953  *	bit if the newly extended portion of the buffer does not contain
954  *	valid data.
955  */
956 static __inline void
957 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
958     vm_offset_t size, vm_page_t m)
959 {
960 
961 	/*
962 	 * This function and its results are protected by higher level
963 	 * synchronization requiring vnode and buf locks to page in and
964 	 * validate pages.
965 	 */
966 	if (bp->b_flags & B_CACHE) {
967 		int base = (foff + off) & PAGE_MASK;
968 		if (vm_page_is_valid(m, base, size) == 0)
969 			bp->b_flags &= ~B_CACHE;
970 	}
971 }
972 
973 /* Wake up the buffer daemon if necessary */
974 static void
975 bd_wakeup(void)
976 {
977 
978 	mtx_lock(&bdlock);
979 	if (bd_request == 0) {
980 		bd_request = 1;
981 		wakeup(&bd_request);
982 	}
983 	mtx_unlock(&bdlock);
984 }
985 
986 /*
987  * Adjust the maxbcachbuf tunable.
988  */
989 static void
990 maxbcachebuf_adjust(void)
991 {
992 	int i;
993 
994 	/*
995 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
996 	 */
997 	i = 2;
998 	while (i * 2 <= maxbcachebuf)
999 		i *= 2;
1000 	maxbcachebuf = i;
1001 	if (maxbcachebuf < MAXBSIZE)
1002 		maxbcachebuf = MAXBSIZE;
1003 	if (maxbcachebuf > maxphys)
1004 		maxbcachebuf = maxphys;
1005 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1006 		printf("maxbcachebuf=%d\n", maxbcachebuf);
1007 }
1008 
1009 /*
1010  * bd_speedup - speedup the buffer cache flushing code
1011  */
1012 void
1013 bd_speedup(void)
1014 {
1015 	int needwake;
1016 
1017 	mtx_lock(&bdlock);
1018 	needwake = 0;
1019 	if (bd_speedupreq == 0 || bd_request == 0)
1020 		needwake = 1;
1021 	bd_speedupreq = 1;
1022 	bd_request = 1;
1023 	if (needwake)
1024 		wakeup(&bd_request);
1025 	mtx_unlock(&bdlock);
1026 }
1027 
1028 #ifdef __i386__
1029 #define	TRANSIENT_DENOM	5
1030 #else
1031 #define	TRANSIENT_DENOM 10
1032 #endif
1033 
1034 /*
1035  * Calculating buffer cache scaling values and reserve space for buffer
1036  * headers.  This is called during low level kernel initialization and
1037  * may be called more then once.  We CANNOT write to the memory area
1038  * being reserved at this time.
1039  */
1040 caddr_t
1041 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1042 {
1043 	int tuned_nbuf;
1044 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1045 
1046 	/*
1047 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1048 	 * PAGE_SIZE is >= 1K)
1049 	 */
1050 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1051 
1052 	maxbcachebuf_adjust();
1053 	/*
1054 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1055 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1056 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1057 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1058 	 * the buffer cache we limit the eventual kva reservation to
1059 	 * maxbcache bytes.
1060 	 *
1061 	 * factor represents the 1/4 x ram conversion.
1062 	 */
1063 	if (nbuf == 0) {
1064 		int factor = 4 * BKVASIZE / 1024;
1065 
1066 		nbuf = 50;
1067 		if (physmem_est > 4096)
1068 			nbuf += min((physmem_est - 4096) / factor,
1069 			    65536 / factor);
1070 		if (physmem_est > 65536)
1071 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1072 			    32 * 1024 * 1024 / (factor * 5));
1073 
1074 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1075 			nbuf = maxbcache / BKVASIZE;
1076 		tuned_nbuf = 1;
1077 	} else
1078 		tuned_nbuf = 0;
1079 
1080 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1081 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1082 	if (nbuf > maxbuf) {
1083 		if (!tuned_nbuf)
1084 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1085 			    maxbuf);
1086 		nbuf = maxbuf;
1087 	}
1088 
1089 	/*
1090 	 * Ideal allocation size for the transient bio submap is 10%
1091 	 * of the maximal space buffer map.  This roughly corresponds
1092 	 * to the amount of the buffer mapped for typical UFS load.
1093 	 *
1094 	 * Clip the buffer map to reserve space for the transient
1095 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1096 	 * maximum buffer map extent on the platform.
1097 	 *
1098 	 * The fall-back to the maxbuf in case of maxbcache unset,
1099 	 * allows to not trim the buffer KVA for the architectures
1100 	 * with ample KVA space.
1101 	 */
1102 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1103 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1104 		buf_sz = (long)nbuf * BKVASIZE;
1105 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1106 		    (TRANSIENT_DENOM - 1)) {
1107 			/*
1108 			 * There is more KVA than memory.  Do not
1109 			 * adjust buffer map size, and assign the rest
1110 			 * of maxbuf to transient map.
1111 			 */
1112 			biotmap_sz = maxbuf_sz - buf_sz;
1113 		} else {
1114 			/*
1115 			 * Buffer map spans all KVA we could afford on
1116 			 * this platform.  Give 10% (20% on i386) of
1117 			 * the buffer map to the transient bio map.
1118 			 */
1119 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1120 			buf_sz -= biotmap_sz;
1121 		}
1122 		if (biotmap_sz / INT_MAX > maxphys)
1123 			bio_transient_maxcnt = INT_MAX;
1124 		else
1125 			bio_transient_maxcnt = biotmap_sz / maxphys;
1126 		/*
1127 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1128 		 * using the transient mapping.
1129 		 */
1130 		if (bio_transient_maxcnt > 1024)
1131 			bio_transient_maxcnt = 1024;
1132 		if (tuned_nbuf)
1133 			nbuf = buf_sz / BKVASIZE;
1134 	}
1135 
1136 	if (nswbuf == 0) {
1137 		nswbuf = min(nbuf / 4, 256);
1138 		if (nswbuf < NSWBUF_MIN)
1139 			nswbuf = NSWBUF_MIN;
1140 	}
1141 
1142 	/*
1143 	 * Reserve space for the buffer cache buffers
1144 	 */
1145 	buf = (char *)v;
1146 	v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1147 	    atop(maxbcachebuf)) * nbuf;
1148 
1149 	return (v);
1150 }
1151 
1152 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1153 void
1154 bufinit(void)
1155 {
1156 	struct buf *bp;
1157 	int i;
1158 
1159 	KASSERT(maxbcachebuf >= MAXBSIZE,
1160 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1161 	    MAXBSIZE));
1162 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1163 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1164 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1165 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1166 
1167 	unmapped_buf = (caddr_t)kva_alloc(maxphys);
1168 
1169 	/* finally, initialize each buffer header and stick on empty q */
1170 	for (i = 0; i < nbuf; i++) {
1171 		bp = nbufp(i);
1172 		bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1173 		bp->b_flags = B_INVAL;
1174 		bp->b_rcred = NOCRED;
1175 		bp->b_wcred = NOCRED;
1176 		bp->b_qindex = QUEUE_NONE;
1177 		bp->b_domain = -1;
1178 		bp->b_subqueue = mp_maxid + 1;
1179 		bp->b_xflags = 0;
1180 		bp->b_data = bp->b_kvabase = unmapped_buf;
1181 		LIST_INIT(&bp->b_dep);
1182 		BUF_LOCKINIT(bp);
1183 		bq_insert(&bqempty, bp, false);
1184 	}
1185 
1186 	/*
1187 	 * maxbufspace is the absolute maximum amount of buffer space we are
1188 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1189 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1190 	 * used by most other requests.  The differential is required to
1191 	 * ensure that metadata deadlocks don't occur.
1192 	 *
1193 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1194 	 * this may result in KVM fragmentation which is not handled optimally
1195 	 * by the system. XXX This is less true with vmem.  We could use
1196 	 * PAGE_SIZE.
1197 	 */
1198 	maxbufspace = (long)nbuf * BKVASIZE;
1199 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1200 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1201 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1202 
1203 	/*
1204 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1205 	 * arbitrarily and may need further tuning. It corresponds to
1206 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1207 	 * which fits with many RAID controllers' tagged queuing limits.
1208 	 * The lower 1 MiB limit is the historical upper limit for
1209 	 * hirunningspace.
1210 	 */
1211 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1212 	    16 * 1024 * 1024), 1024 * 1024);
1213 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1214 
1215 	/*
1216 	 * Limit the amount of malloc memory since it is wired permanently into
1217 	 * the kernel space.  Even though this is accounted for in the buffer
1218 	 * allocation, we don't want the malloced region to grow uncontrolled.
1219 	 * The malloc scheme improves memory utilization significantly on
1220 	 * average (small) directories.
1221 	 */
1222 	maxbufmallocspace = hibufspace / 20;
1223 
1224 	/*
1225 	 * Reduce the chance of a deadlock occurring by limiting the number
1226 	 * of delayed-write dirty buffers we allow to stack up.
1227 	 */
1228 	hidirtybuffers = nbuf / 4 + 20;
1229 	dirtybufthresh = hidirtybuffers * 9 / 10;
1230 	/*
1231 	 * To support extreme low-memory systems, make sure hidirtybuffers
1232 	 * cannot eat up all available buffer space.  This occurs when our
1233 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1234 	 * buffer space assuming BKVASIZE'd buffers.
1235 	 */
1236 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1237 		hidirtybuffers >>= 1;
1238 	}
1239 	lodirtybuffers = hidirtybuffers / 2;
1240 
1241 	/*
1242 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1243 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1244 	 * are counted as free but will be unavailable to threads executing
1245 	 * on other cpus.
1246 	 *
1247 	 * hifreebuffers is the free target for the bufspace daemon.  This
1248 	 * should be set appropriately to limit work per-iteration.
1249 	 */
1250 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1251 	hifreebuffers = (3 * lofreebuffers) / 2;
1252 	numfreebuffers = nbuf;
1253 
1254 	/* Setup the kva and free list allocators. */
1255 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1256 	buf_zone = uma_zcache_create("buf free cache",
1257 	    sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1258 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1259 
1260 	/*
1261 	 * Size the clean queue according to the amount of buffer space.
1262 	 * One queue per-256mb up to the max.  More queues gives better
1263 	 * concurrency but less accurate LRU.
1264 	 */
1265 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1266 	for (i = 0 ; i < buf_domains; i++) {
1267 		struct bufdomain *bd;
1268 
1269 		bd = &bdomain[i];
1270 		bd_init(bd);
1271 		bd->bd_freebuffers = nbuf / buf_domains;
1272 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1273 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1274 		bd->bd_bufspace = 0;
1275 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1276 		bd->bd_hibufspace = hibufspace / buf_domains;
1277 		bd->bd_lobufspace = lobufspace / buf_domains;
1278 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1279 		bd->bd_numdirtybuffers = 0;
1280 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1281 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1282 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1283 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1284 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1285 	}
1286 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1287 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1288 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1289 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1290 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1291 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1292 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1293 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1294 }
1295 
1296 #ifdef INVARIANTS
1297 static inline void
1298 vfs_buf_check_mapped(struct buf *bp)
1299 {
1300 
1301 	KASSERT(bp->b_kvabase != unmapped_buf,
1302 	    ("mapped buf: b_kvabase was not updated %p", bp));
1303 	KASSERT(bp->b_data != unmapped_buf,
1304 	    ("mapped buf: b_data was not updated %p", bp));
1305 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1306 	    maxphys, ("b_data + b_offset unmapped %p", bp));
1307 }
1308 
1309 static inline void
1310 vfs_buf_check_unmapped(struct buf *bp)
1311 {
1312 
1313 	KASSERT(bp->b_data == unmapped_buf,
1314 	    ("unmapped buf: corrupted b_data %p", bp));
1315 }
1316 
1317 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1318 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1319 #else
1320 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1321 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1322 #endif
1323 
1324 static int
1325 isbufbusy(struct buf *bp)
1326 {
1327 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1328 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1329 		return (1);
1330 	return (0);
1331 }
1332 
1333 /*
1334  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1335  */
1336 void
1337 bufshutdown(int show_busybufs)
1338 {
1339 	static int first_buf_printf = 1;
1340 	struct buf *bp;
1341 	int i, iter, nbusy, pbusy;
1342 #ifndef PREEMPTION
1343 	int subiter;
1344 #endif
1345 
1346 	/*
1347 	 * Sync filesystems for shutdown
1348 	 */
1349 	wdog_kern_pat(WD_LASTVAL);
1350 	kern_sync(curthread);
1351 
1352 	/*
1353 	 * With soft updates, some buffers that are
1354 	 * written will be remarked as dirty until other
1355 	 * buffers are written.
1356 	 */
1357 	for (iter = pbusy = 0; iter < 20; iter++) {
1358 		nbusy = 0;
1359 		for (i = nbuf - 1; i >= 0; i--) {
1360 			bp = nbufp(i);
1361 			if (isbufbusy(bp))
1362 				nbusy++;
1363 		}
1364 		if (nbusy == 0) {
1365 			if (first_buf_printf)
1366 				printf("All buffers synced.");
1367 			break;
1368 		}
1369 		if (first_buf_printf) {
1370 			printf("Syncing disks, buffers remaining... ");
1371 			first_buf_printf = 0;
1372 		}
1373 		printf("%d ", nbusy);
1374 		if (nbusy < pbusy)
1375 			iter = 0;
1376 		pbusy = nbusy;
1377 
1378 		wdog_kern_pat(WD_LASTVAL);
1379 		kern_sync(curthread);
1380 
1381 #ifdef PREEMPTION
1382 		/*
1383 		 * Spin for a while to allow interrupt threads to run.
1384 		 */
1385 		DELAY(50000 * iter);
1386 #else
1387 		/*
1388 		 * Context switch several times to allow interrupt
1389 		 * threads to run.
1390 		 */
1391 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1392 			thread_lock(curthread);
1393 			mi_switch(SW_VOL);
1394 			DELAY(1000);
1395 		}
1396 #endif
1397 	}
1398 	printf("\n");
1399 	/*
1400 	 * Count only busy local buffers to prevent forcing
1401 	 * a fsck if we're just a client of a wedged NFS server
1402 	 */
1403 	nbusy = 0;
1404 	for (i = nbuf - 1; i >= 0; i--) {
1405 		bp = nbufp(i);
1406 		if (isbufbusy(bp)) {
1407 #if 0
1408 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1409 			if (bp->b_dev == NULL) {
1410 				TAILQ_REMOVE(&mountlist,
1411 				    bp->b_vp->v_mount, mnt_list);
1412 				continue;
1413 			}
1414 #endif
1415 			nbusy++;
1416 			if (show_busybufs > 0) {
1417 				printf(
1418 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1419 				    nbusy, bp, bp->b_vp, bp->b_flags,
1420 				    (intmax_t)bp->b_blkno,
1421 				    (intmax_t)bp->b_lblkno);
1422 				BUF_LOCKPRINTINFO(bp);
1423 				if (show_busybufs > 1)
1424 					vn_printf(bp->b_vp,
1425 					    "vnode content: ");
1426 			}
1427 		}
1428 	}
1429 	if (nbusy) {
1430 		/*
1431 		 * Failed to sync all blocks. Indicate this and don't
1432 		 * unmount filesystems (thus forcing an fsck on reboot).
1433 		 */
1434 		printf("Giving up on %d buffers\n", nbusy);
1435 		DELAY(5000000);	/* 5 seconds */
1436 	} else {
1437 		if (!first_buf_printf)
1438 			printf("Final sync complete\n");
1439 		/*
1440 		 * Unmount filesystems
1441 		 */
1442 		if (!KERNEL_PANICKED())
1443 			vfs_unmountall();
1444 	}
1445 	swapoff_all();
1446 	DELAY(100000);		/* wait for console output to finish */
1447 }
1448 
1449 static void
1450 bpmap_qenter(struct buf *bp)
1451 {
1452 
1453 	BUF_CHECK_MAPPED(bp);
1454 
1455 	/*
1456 	 * bp->b_data is relative to bp->b_offset, but
1457 	 * bp->b_offset may be offset into the first page.
1458 	 */
1459 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1460 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1461 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1462 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1463 }
1464 
1465 static inline struct bufdomain *
1466 bufdomain(struct buf *bp)
1467 {
1468 
1469 	return (&bdomain[bp->b_domain]);
1470 }
1471 
1472 static struct bufqueue *
1473 bufqueue(struct buf *bp)
1474 {
1475 
1476 	switch (bp->b_qindex) {
1477 	case QUEUE_NONE:
1478 		/* FALLTHROUGH */
1479 	case QUEUE_SENTINEL:
1480 		return (NULL);
1481 	case QUEUE_EMPTY:
1482 		return (&bqempty);
1483 	case QUEUE_DIRTY:
1484 		return (&bufdomain(bp)->bd_dirtyq);
1485 	case QUEUE_CLEAN:
1486 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1487 	default:
1488 		break;
1489 	}
1490 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1491 }
1492 
1493 /*
1494  * Return the locked bufqueue that bp is a member of.
1495  */
1496 static struct bufqueue *
1497 bufqueue_acquire(struct buf *bp)
1498 {
1499 	struct bufqueue *bq, *nbq;
1500 
1501 	/*
1502 	 * bp can be pushed from a per-cpu queue to the
1503 	 * cleanq while we're waiting on the lock.  Retry
1504 	 * if the queues don't match.
1505 	 */
1506 	bq = bufqueue(bp);
1507 	BQ_LOCK(bq);
1508 	for (;;) {
1509 		nbq = bufqueue(bp);
1510 		if (bq == nbq)
1511 			break;
1512 		BQ_UNLOCK(bq);
1513 		BQ_LOCK(nbq);
1514 		bq = nbq;
1515 	}
1516 	return (bq);
1517 }
1518 
1519 /*
1520  *	binsfree:
1521  *
1522  *	Insert the buffer into the appropriate free list.  Requires a
1523  *	locked buffer on entry and buffer is unlocked before return.
1524  */
1525 static void
1526 binsfree(struct buf *bp, int qindex)
1527 {
1528 	struct bufdomain *bd;
1529 	struct bufqueue *bq;
1530 
1531 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1532 	    ("binsfree: Invalid qindex %d", qindex));
1533 	BUF_ASSERT_XLOCKED(bp);
1534 
1535 	/*
1536 	 * Handle delayed bremfree() processing.
1537 	 */
1538 	if (bp->b_flags & B_REMFREE) {
1539 		if (bp->b_qindex == qindex) {
1540 			bp->b_flags |= B_REUSE;
1541 			bp->b_flags &= ~B_REMFREE;
1542 			BUF_UNLOCK(bp);
1543 			return;
1544 		}
1545 		bq = bufqueue_acquire(bp);
1546 		bq_remove(bq, bp);
1547 		BQ_UNLOCK(bq);
1548 	}
1549 	bd = bufdomain(bp);
1550 	if (qindex == QUEUE_CLEAN) {
1551 		if (bd->bd_lim != 0)
1552 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1553 		else
1554 			bq = bd->bd_cleanq;
1555 	} else
1556 		bq = &bd->bd_dirtyq;
1557 	bq_insert(bq, bp, true);
1558 }
1559 
1560 /*
1561  * buf_free:
1562  *
1563  *	Free a buffer to the buf zone once it no longer has valid contents.
1564  */
1565 static void
1566 buf_free(struct buf *bp)
1567 {
1568 
1569 	if (bp->b_flags & B_REMFREE)
1570 		bremfreef(bp);
1571 	if (bp->b_vflags & BV_BKGRDINPROG)
1572 		panic("losing buffer 1");
1573 	if (bp->b_rcred != NOCRED) {
1574 		crfree(bp->b_rcred);
1575 		bp->b_rcred = NOCRED;
1576 	}
1577 	if (bp->b_wcred != NOCRED) {
1578 		crfree(bp->b_wcred);
1579 		bp->b_wcred = NOCRED;
1580 	}
1581 	if (!LIST_EMPTY(&bp->b_dep))
1582 		buf_deallocate(bp);
1583 	bufkva_free(bp);
1584 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1585 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
1586 	BUF_UNLOCK(bp);
1587 	uma_zfree(buf_zone, bp);
1588 }
1589 
1590 /*
1591  * buf_import:
1592  *
1593  *	Import bufs into the uma cache from the buf list.  The system still
1594  *	expects a static array of bufs and much of the synchronization
1595  *	around bufs assumes type stable storage.  As a result, UMA is used
1596  *	only as a per-cpu cache of bufs still maintained on a global list.
1597  */
1598 static int
1599 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1600 {
1601 	struct buf *bp;
1602 	int i;
1603 
1604 	BQ_LOCK(&bqempty);
1605 	for (i = 0; i < cnt; i++) {
1606 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1607 		if (bp == NULL)
1608 			break;
1609 		bq_remove(&bqempty, bp);
1610 		store[i] = bp;
1611 	}
1612 	BQ_UNLOCK(&bqempty);
1613 
1614 	return (i);
1615 }
1616 
1617 /*
1618  * buf_release:
1619  *
1620  *	Release bufs from the uma cache back to the buffer queues.
1621  */
1622 static void
1623 buf_release(void *arg, void **store, int cnt)
1624 {
1625 	struct bufqueue *bq;
1626 	struct buf *bp;
1627         int i;
1628 
1629 	bq = &bqempty;
1630 	BQ_LOCK(bq);
1631         for (i = 0; i < cnt; i++) {
1632 		bp = store[i];
1633 		/* Inline bq_insert() to batch locking. */
1634 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1635 		bp->b_flags &= ~(B_AGE | B_REUSE);
1636 		bq->bq_len++;
1637 		bp->b_qindex = bq->bq_index;
1638 	}
1639 	BQ_UNLOCK(bq);
1640 }
1641 
1642 /*
1643  * buf_alloc:
1644  *
1645  *	Allocate an empty buffer header.
1646  */
1647 static struct buf *
1648 buf_alloc(struct bufdomain *bd)
1649 {
1650 	struct buf *bp;
1651 	int freebufs, error;
1652 
1653 	/*
1654 	 * We can only run out of bufs in the buf zone if the average buf
1655 	 * is less than BKVASIZE.  In this case the actual wait/block will
1656 	 * come from buf_reycle() failing to flush one of these small bufs.
1657 	 */
1658 	bp = NULL;
1659 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1660 	if (freebufs > 0)
1661 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1662 	if (bp == NULL) {
1663 		atomic_add_int(&bd->bd_freebuffers, 1);
1664 		bufspace_daemon_wakeup(bd);
1665 		counter_u64_add(numbufallocfails, 1);
1666 		return (NULL);
1667 	}
1668 	/*
1669 	 * Wake-up the bufspace daemon on transition below threshold.
1670 	 */
1671 	if (freebufs == bd->bd_lofreebuffers)
1672 		bufspace_daemon_wakeup(bd);
1673 
1674 	error = BUF_LOCK(bp, LK_EXCLUSIVE, NULL);
1675 	KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1676 	    error));
1677 	(void)error;
1678 
1679 	KASSERT(bp->b_vp == NULL,
1680 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1681 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1682 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1683 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1684 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1685 	KASSERT(bp->b_npages == 0,
1686 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1687 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1688 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1689 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
1690 
1691 	bp->b_domain = BD_DOMAIN(bd);
1692 	bp->b_flags = 0;
1693 	bp->b_ioflags = 0;
1694 	bp->b_xflags = 0;
1695 	bp->b_vflags = 0;
1696 	bp->b_vp = NULL;
1697 	bp->b_blkno = bp->b_lblkno = 0;
1698 	bp->b_offset = NOOFFSET;
1699 	bp->b_iodone = 0;
1700 	bp->b_error = 0;
1701 	bp->b_resid = 0;
1702 	bp->b_bcount = 0;
1703 	bp->b_npages = 0;
1704 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1705 	bp->b_bufobj = NULL;
1706 	bp->b_data = bp->b_kvabase = unmapped_buf;
1707 	bp->b_fsprivate1 = NULL;
1708 	bp->b_fsprivate2 = NULL;
1709 	bp->b_fsprivate3 = NULL;
1710 	LIST_INIT(&bp->b_dep);
1711 
1712 	return (bp);
1713 }
1714 
1715 /*
1716  *	buf_recycle:
1717  *
1718  *	Free a buffer from the given bufqueue.  kva controls whether the
1719  *	freed buf must own some kva resources.  This is used for
1720  *	defragmenting.
1721  */
1722 static int
1723 buf_recycle(struct bufdomain *bd, bool kva)
1724 {
1725 	struct bufqueue *bq;
1726 	struct buf *bp, *nbp;
1727 
1728 	if (kva)
1729 		counter_u64_add(bufdefragcnt, 1);
1730 	nbp = NULL;
1731 	bq = bd->bd_cleanq;
1732 	BQ_LOCK(bq);
1733 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1734 	    ("buf_recycle: Locks don't match"));
1735 	nbp = TAILQ_FIRST(&bq->bq_queue);
1736 
1737 	/*
1738 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1739 	 * depending.
1740 	 */
1741 	while ((bp = nbp) != NULL) {
1742 		/*
1743 		 * Calculate next bp (we can only use it if we do not
1744 		 * release the bqlock).
1745 		 */
1746 		nbp = TAILQ_NEXT(bp, b_freelist);
1747 
1748 		/*
1749 		 * If we are defragging then we need a buffer with
1750 		 * some kva to reclaim.
1751 		 */
1752 		if (kva && bp->b_kvasize == 0)
1753 			continue;
1754 
1755 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1756 			continue;
1757 
1758 		/*
1759 		 * Implement a second chance algorithm for frequently
1760 		 * accessed buffers.
1761 		 */
1762 		if ((bp->b_flags & B_REUSE) != 0) {
1763 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1764 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1765 			bp->b_flags &= ~B_REUSE;
1766 			BUF_UNLOCK(bp);
1767 			continue;
1768 		}
1769 
1770 		/*
1771 		 * Skip buffers with background writes in progress.
1772 		 */
1773 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1774 			BUF_UNLOCK(bp);
1775 			continue;
1776 		}
1777 
1778 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1779 		    ("buf_recycle: inconsistent queue %d bp %p",
1780 		    bp->b_qindex, bp));
1781 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1782 		    ("getnewbuf: queue domain %d doesn't match request %d",
1783 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1784 		/*
1785 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1786 		 * the scan from this point on.
1787 		 */
1788 		bq_remove(bq, bp);
1789 		BQ_UNLOCK(bq);
1790 
1791 		/*
1792 		 * Requeue the background write buffer with error and
1793 		 * restart the scan.
1794 		 */
1795 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1796 			bqrelse(bp);
1797 			BQ_LOCK(bq);
1798 			nbp = TAILQ_FIRST(&bq->bq_queue);
1799 			continue;
1800 		}
1801 		bp->b_flags |= B_INVAL;
1802 		brelse(bp);
1803 		return (0);
1804 	}
1805 	bd->bd_wanted = 1;
1806 	BQ_UNLOCK(bq);
1807 
1808 	return (ENOBUFS);
1809 }
1810 
1811 /*
1812  *	bremfree:
1813  *
1814  *	Mark the buffer for removal from the appropriate free list.
1815  *
1816  */
1817 void
1818 bremfree(struct buf *bp)
1819 {
1820 
1821 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1822 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1823 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1824 	KASSERT(bp->b_qindex != QUEUE_NONE,
1825 	    ("bremfree: buffer %p not on a queue.", bp));
1826 	BUF_ASSERT_XLOCKED(bp);
1827 
1828 	bp->b_flags |= B_REMFREE;
1829 }
1830 
1831 /*
1832  *	bremfreef:
1833  *
1834  *	Force an immediate removal from a free list.  Used only in nfs when
1835  *	it abuses the b_freelist pointer.
1836  */
1837 void
1838 bremfreef(struct buf *bp)
1839 {
1840 	struct bufqueue *bq;
1841 
1842 	bq = bufqueue_acquire(bp);
1843 	bq_remove(bq, bp);
1844 	BQ_UNLOCK(bq);
1845 }
1846 
1847 static void
1848 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1849 {
1850 
1851 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1852 	TAILQ_INIT(&bq->bq_queue);
1853 	bq->bq_len = 0;
1854 	bq->bq_index = qindex;
1855 	bq->bq_subqueue = subqueue;
1856 }
1857 
1858 static void
1859 bd_init(struct bufdomain *bd)
1860 {
1861 	int i;
1862 
1863 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1864 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1865 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1866 	for (i = 0; i <= mp_maxid; i++)
1867 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1868 		    "bufq clean subqueue lock");
1869 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1870 }
1871 
1872 /*
1873  *	bq_remove:
1874  *
1875  *	Removes a buffer from the free list, must be called with the
1876  *	correct qlock held.
1877  */
1878 static void
1879 bq_remove(struct bufqueue *bq, struct buf *bp)
1880 {
1881 
1882 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1883 	    bp, bp->b_vp, bp->b_flags);
1884 	KASSERT(bp->b_qindex != QUEUE_NONE,
1885 	    ("bq_remove: buffer %p not on a queue.", bp));
1886 	KASSERT(bufqueue(bp) == bq,
1887 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1888 
1889 	BQ_ASSERT_LOCKED(bq);
1890 	if (bp->b_qindex != QUEUE_EMPTY) {
1891 		BUF_ASSERT_XLOCKED(bp);
1892 	}
1893 	KASSERT(bq->bq_len >= 1,
1894 	    ("queue %d underflow", bp->b_qindex));
1895 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1896 	bq->bq_len--;
1897 	bp->b_qindex = QUEUE_NONE;
1898 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1899 }
1900 
1901 static void
1902 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1903 {
1904 	struct buf *bp;
1905 
1906 	BQ_ASSERT_LOCKED(bq);
1907 	if (bq != bd->bd_cleanq) {
1908 		BD_LOCK(bd);
1909 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1910 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1911 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1912 			    b_freelist);
1913 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1914 		}
1915 		bd->bd_cleanq->bq_len += bq->bq_len;
1916 		bq->bq_len = 0;
1917 	}
1918 	if (bd->bd_wanted) {
1919 		bd->bd_wanted = 0;
1920 		wakeup(&bd->bd_wanted);
1921 	}
1922 	if (bq != bd->bd_cleanq)
1923 		BD_UNLOCK(bd);
1924 }
1925 
1926 static int
1927 bd_flushall(struct bufdomain *bd)
1928 {
1929 	struct bufqueue *bq;
1930 	int flushed;
1931 	int i;
1932 
1933 	if (bd->bd_lim == 0)
1934 		return (0);
1935 	flushed = 0;
1936 	for (i = 0; i <= mp_maxid; i++) {
1937 		bq = &bd->bd_subq[i];
1938 		if (bq->bq_len == 0)
1939 			continue;
1940 		BQ_LOCK(bq);
1941 		bd_flush(bd, bq);
1942 		BQ_UNLOCK(bq);
1943 		flushed++;
1944 	}
1945 
1946 	return (flushed);
1947 }
1948 
1949 static void
1950 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1951 {
1952 	struct bufdomain *bd;
1953 
1954 	if (bp->b_qindex != QUEUE_NONE)
1955 		panic("bq_insert: free buffer %p onto another queue?", bp);
1956 
1957 	bd = bufdomain(bp);
1958 	if (bp->b_flags & B_AGE) {
1959 		/* Place this buf directly on the real queue. */
1960 		if (bq->bq_index == QUEUE_CLEAN)
1961 			bq = bd->bd_cleanq;
1962 		BQ_LOCK(bq);
1963 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1964 	} else {
1965 		BQ_LOCK(bq);
1966 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1967 	}
1968 	bp->b_flags &= ~(B_AGE | B_REUSE);
1969 	bq->bq_len++;
1970 	bp->b_qindex = bq->bq_index;
1971 	bp->b_subqueue = bq->bq_subqueue;
1972 
1973 	/*
1974 	 * Unlock before we notify so that we don't wakeup a waiter that
1975 	 * fails a trylock on the buf and sleeps again.
1976 	 */
1977 	if (unlock)
1978 		BUF_UNLOCK(bp);
1979 
1980 	if (bp->b_qindex == QUEUE_CLEAN) {
1981 		/*
1982 		 * Flush the per-cpu queue and notify any waiters.
1983 		 */
1984 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1985 		    bq->bq_len >= bd->bd_lim))
1986 			bd_flush(bd, bq);
1987 	}
1988 	BQ_UNLOCK(bq);
1989 }
1990 
1991 /*
1992  *	bufkva_free:
1993  *
1994  *	Free the kva allocation for a buffer.
1995  *
1996  */
1997 static void
1998 bufkva_free(struct buf *bp)
1999 {
2000 
2001 #ifdef INVARIANTS
2002 	if (bp->b_kvasize == 0) {
2003 		KASSERT(bp->b_kvabase == unmapped_buf &&
2004 		    bp->b_data == unmapped_buf,
2005 		    ("Leaked KVA space on %p", bp));
2006 	} else if (buf_mapped(bp))
2007 		BUF_CHECK_MAPPED(bp);
2008 	else
2009 		BUF_CHECK_UNMAPPED(bp);
2010 #endif
2011 	if (bp->b_kvasize == 0)
2012 		return;
2013 
2014 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2015 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2016 	counter_u64_add(buffreekvacnt, 1);
2017 	bp->b_data = bp->b_kvabase = unmapped_buf;
2018 	bp->b_kvasize = 0;
2019 }
2020 
2021 /*
2022  *	bufkva_alloc:
2023  *
2024  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2025  */
2026 static int
2027 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2028 {
2029 	vm_offset_t addr;
2030 	int error;
2031 
2032 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2033 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2034 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
2035 	KASSERT(maxsize <= maxbcachebuf,
2036 	    ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2037 
2038 	bufkva_free(bp);
2039 
2040 	addr = 0;
2041 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2042 	if (error != 0) {
2043 		/*
2044 		 * Buffer map is too fragmented.  Request the caller
2045 		 * to defragment the map.
2046 		 */
2047 		return (error);
2048 	}
2049 	bp->b_kvabase = (caddr_t)addr;
2050 	bp->b_kvasize = maxsize;
2051 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2052 	if ((gbflags & GB_UNMAPPED) != 0) {
2053 		bp->b_data = unmapped_buf;
2054 		BUF_CHECK_UNMAPPED(bp);
2055 	} else {
2056 		bp->b_data = bp->b_kvabase;
2057 		BUF_CHECK_MAPPED(bp);
2058 	}
2059 	return (0);
2060 }
2061 
2062 /*
2063  *	bufkva_reclaim:
2064  *
2065  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2066  *	callback that fires to avoid returning failure.
2067  */
2068 static void
2069 bufkva_reclaim(vmem_t *vmem, int flags)
2070 {
2071 	bool done;
2072 	int q;
2073 	int i;
2074 
2075 	done = false;
2076 	for (i = 0; i < 5; i++) {
2077 		for (q = 0; q < buf_domains; q++)
2078 			if (buf_recycle(&bdomain[q], true) != 0)
2079 				done = true;
2080 		if (done)
2081 			break;
2082 	}
2083 	return;
2084 }
2085 
2086 /*
2087  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2088  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2089  * the buffer is valid and we do not have to do anything.
2090  */
2091 static void
2092 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2093     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2094 {
2095 	struct buf *rabp;
2096 	struct thread *td;
2097 	int i;
2098 
2099 	td = curthread;
2100 
2101 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2102 		if (inmem(vp, *rablkno))
2103 			continue;
2104 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2105 		if ((rabp->b_flags & B_CACHE) != 0) {
2106 			brelse(rabp);
2107 			continue;
2108 		}
2109 #ifdef RACCT
2110 		if (racct_enable) {
2111 			PROC_LOCK(curproc);
2112 			racct_add_buf(curproc, rabp, 0);
2113 			PROC_UNLOCK(curproc);
2114 		}
2115 #endif /* RACCT */
2116 		td->td_ru.ru_inblock++;
2117 		rabp->b_flags |= B_ASYNC;
2118 		rabp->b_flags &= ~B_INVAL;
2119 		if ((flags & GB_CKHASH) != 0) {
2120 			rabp->b_flags |= B_CKHASH;
2121 			rabp->b_ckhashcalc = ckhashfunc;
2122 		}
2123 		rabp->b_ioflags &= ~BIO_ERROR;
2124 		rabp->b_iocmd = BIO_READ;
2125 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2126 			rabp->b_rcred = crhold(cred);
2127 		vfs_busy_pages(rabp, 0);
2128 		BUF_KERNPROC(rabp);
2129 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2130 		bstrategy(rabp);
2131 	}
2132 }
2133 
2134 /*
2135  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2136  *
2137  * Get a buffer with the specified data.  Look in the cache first.  We
2138  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2139  * is set, the buffer is valid and we do not have to do anything, see
2140  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2141  *
2142  * Always return a NULL buffer pointer (in bpp) when returning an error.
2143  *
2144  * The blkno parameter is the logical block being requested. Normally
2145  * the mapping of logical block number to disk block address is done
2146  * by calling VOP_BMAP(). However, if the mapping is already known, the
2147  * disk block address can be passed using the dblkno parameter. If the
2148  * disk block address is not known, then the same value should be passed
2149  * for blkno and dblkno.
2150  */
2151 int
2152 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2153     daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2154     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2155 {
2156 	struct buf *bp;
2157 	struct thread *td;
2158 	int error, readwait, rv;
2159 
2160 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2161 	td = curthread;
2162 	/*
2163 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2164 	 * are specified.
2165 	 */
2166 	error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2167 	if (error != 0) {
2168 		*bpp = NULL;
2169 		return (error);
2170 	}
2171 	KASSERT(blkno == bp->b_lblkno,
2172 	    ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2173 	    (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2174 	flags &= ~GB_NOSPARSE;
2175 	*bpp = bp;
2176 
2177 	/*
2178 	 * If not found in cache, do some I/O
2179 	 */
2180 	readwait = 0;
2181 	if ((bp->b_flags & B_CACHE) == 0) {
2182 #ifdef RACCT
2183 		if (racct_enable) {
2184 			PROC_LOCK(td->td_proc);
2185 			racct_add_buf(td->td_proc, bp, 0);
2186 			PROC_UNLOCK(td->td_proc);
2187 		}
2188 #endif /* RACCT */
2189 		td->td_ru.ru_inblock++;
2190 		bp->b_iocmd = BIO_READ;
2191 		bp->b_flags &= ~B_INVAL;
2192 		if ((flags & GB_CKHASH) != 0) {
2193 			bp->b_flags |= B_CKHASH;
2194 			bp->b_ckhashcalc = ckhashfunc;
2195 		}
2196 		if ((flags & GB_CVTENXIO) != 0)
2197 			bp->b_xflags |= BX_CVTENXIO;
2198 		bp->b_ioflags &= ~BIO_ERROR;
2199 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2200 			bp->b_rcred = crhold(cred);
2201 		vfs_busy_pages(bp, 0);
2202 		bp->b_iooffset = dbtob(bp->b_blkno);
2203 		bstrategy(bp);
2204 		++readwait;
2205 	}
2206 
2207 	/*
2208 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2209 	 */
2210 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2211 
2212 	rv = 0;
2213 	if (readwait) {
2214 		rv = bufwait(bp);
2215 		if (rv != 0) {
2216 			brelse(bp);
2217 			*bpp = NULL;
2218 		}
2219 	}
2220 	return (rv);
2221 }
2222 
2223 /*
2224  * Write, release buffer on completion.  (Done by iodone
2225  * if async).  Do not bother writing anything if the buffer
2226  * is invalid.
2227  *
2228  * Note that we set B_CACHE here, indicating that buffer is
2229  * fully valid and thus cacheable.  This is true even of NFS
2230  * now so we set it generally.  This could be set either here
2231  * or in biodone() since the I/O is synchronous.  We put it
2232  * here.
2233  */
2234 int
2235 bufwrite(struct buf *bp)
2236 {
2237 	int oldflags;
2238 	struct vnode *vp;
2239 	long space;
2240 	int vp_md;
2241 
2242 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2243 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2244 		bp->b_flags |= B_INVAL | B_RELBUF;
2245 		bp->b_flags &= ~B_CACHE;
2246 		brelse(bp);
2247 		return (ENXIO);
2248 	}
2249 	if (bp->b_flags & B_INVAL) {
2250 		brelse(bp);
2251 		return (0);
2252 	}
2253 
2254 	if (bp->b_flags & B_BARRIER)
2255 		atomic_add_long(&barrierwrites, 1);
2256 
2257 	oldflags = bp->b_flags;
2258 
2259 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2260 	    ("FFS background buffer should not get here %p", bp));
2261 
2262 	vp = bp->b_vp;
2263 	if (vp)
2264 		vp_md = vp->v_vflag & VV_MD;
2265 	else
2266 		vp_md = 0;
2267 
2268 	/*
2269 	 * Mark the buffer clean.  Increment the bufobj write count
2270 	 * before bundirty() call, to prevent other thread from seeing
2271 	 * empty dirty list and zero counter for writes in progress,
2272 	 * falsely indicating that the bufobj is clean.
2273 	 */
2274 	bufobj_wref(bp->b_bufobj);
2275 	bundirty(bp);
2276 
2277 	bp->b_flags &= ~B_DONE;
2278 	bp->b_ioflags &= ~BIO_ERROR;
2279 	bp->b_flags |= B_CACHE;
2280 	bp->b_iocmd = BIO_WRITE;
2281 
2282 	vfs_busy_pages(bp, 1);
2283 
2284 	/*
2285 	 * Normal bwrites pipeline writes
2286 	 */
2287 	bp->b_runningbufspace = bp->b_bufsize;
2288 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2289 
2290 #ifdef RACCT
2291 	if (racct_enable) {
2292 		PROC_LOCK(curproc);
2293 		racct_add_buf(curproc, bp, 1);
2294 		PROC_UNLOCK(curproc);
2295 	}
2296 #endif /* RACCT */
2297 	curthread->td_ru.ru_oublock++;
2298 	if (oldflags & B_ASYNC)
2299 		BUF_KERNPROC(bp);
2300 	bp->b_iooffset = dbtob(bp->b_blkno);
2301 	buf_track(bp, __func__);
2302 	bstrategy(bp);
2303 
2304 	if ((oldflags & B_ASYNC) == 0) {
2305 		int rtval = bufwait(bp);
2306 		brelse(bp);
2307 		return (rtval);
2308 	} else if (space > hirunningspace) {
2309 		/*
2310 		 * don't allow the async write to saturate the I/O
2311 		 * system.  We will not deadlock here because
2312 		 * we are blocking waiting for I/O that is already in-progress
2313 		 * to complete. We do not block here if it is the update
2314 		 * or syncer daemon trying to clean up as that can lead
2315 		 * to deadlock.
2316 		 */
2317 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2318 			waitrunningbufspace();
2319 	}
2320 
2321 	return (0);
2322 }
2323 
2324 void
2325 bufbdflush(struct bufobj *bo, struct buf *bp)
2326 {
2327 	struct buf *nbp;
2328 	struct bufdomain *bd;
2329 
2330 	bd = &bdomain[bo->bo_domain];
2331 	if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2332 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2333 		altbufferflushes++;
2334 	} else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2335 		BO_LOCK(bo);
2336 		/*
2337 		 * Try to find a buffer to flush.
2338 		 */
2339 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2340 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2341 			    BUF_LOCK(nbp,
2342 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2343 				continue;
2344 			if (bp == nbp)
2345 				panic("bdwrite: found ourselves");
2346 			BO_UNLOCK(bo);
2347 			/* Don't countdeps with the bo lock held. */
2348 			if (buf_countdeps(nbp, 0)) {
2349 				BO_LOCK(bo);
2350 				BUF_UNLOCK(nbp);
2351 				continue;
2352 			}
2353 			if (nbp->b_flags & B_CLUSTEROK) {
2354 				vfs_bio_awrite(nbp);
2355 			} else {
2356 				bremfree(nbp);
2357 				bawrite(nbp);
2358 			}
2359 			dirtybufferflushes++;
2360 			break;
2361 		}
2362 		if (nbp == NULL)
2363 			BO_UNLOCK(bo);
2364 	}
2365 }
2366 
2367 /*
2368  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2369  * anything if the buffer is marked invalid.
2370  *
2371  * Note that since the buffer must be completely valid, we can safely
2372  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2373  * biodone() in order to prevent getblk from writing the buffer
2374  * out synchronously.
2375  */
2376 void
2377 bdwrite(struct buf *bp)
2378 {
2379 	struct thread *td = curthread;
2380 	struct vnode *vp;
2381 	struct bufobj *bo;
2382 
2383 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2384 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2385 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2386 	    ("Barrier request in delayed write %p", bp));
2387 
2388 	if (bp->b_flags & B_INVAL) {
2389 		brelse(bp);
2390 		return;
2391 	}
2392 
2393 	/*
2394 	 * If we have too many dirty buffers, don't create any more.
2395 	 * If we are wildly over our limit, then force a complete
2396 	 * cleanup. Otherwise, just keep the situation from getting
2397 	 * out of control. Note that we have to avoid a recursive
2398 	 * disaster and not try to clean up after our own cleanup!
2399 	 */
2400 	vp = bp->b_vp;
2401 	bo = bp->b_bufobj;
2402 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2403 		td->td_pflags |= TDP_INBDFLUSH;
2404 		BO_BDFLUSH(bo, bp);
2405 		td->td_pflags &= ~TDP_INBDFLUSH;
2406 	} else
2407 		recursiveflushes++;
2408 
2409 	bdirty(bp);
2410 	/*
2411 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2412 	 * true even of NFS now.
2413 	 */
2414 	bp->b_flags |= B_CACHE;
2415 
2416 	/*
2417 	 * This bmap keeps the system from needing to do the bmap later,
2418 	 * perhaps when the system is attempting to do a sync.  Since it
2419 	 * is likely that the indirect block -- or whatever other datastructure
2420 	 * that the filesystem needs is still in memory now, it is a good
2421 	 * thing to do this.  Note also, that if the pageout daemon is
2422 	 * requesting a sync -- there might not be enough memory to do
2423 	 * the bmap then...  So, this is important to do.
2424 	 */
2425 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2426 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2427 	}
2428 
2429 	buf_track(bp, __func__);
2430 
2431 	/*
2432 	 * Set the *dirty* buffer range based upon the VM system dirty
2433 	 * pages.
2434 	 *
2435 	 * Mark the buffer pages as clean.  We need to do this here to
2436 	 * satisfy the vnode_pager and the pageout daemon, so that it
2437 	 * thinks that the pages have been "cleaned".  Note that since
2438 	 * the pages are in a delayed write buffer -- the VFS layer
2439 	 * "will" see that the pages get written out on the next sync,
2440 	 * or perhaps the cluster will be completed.
2441 	 */
2442 	vfs_clean_pages_dirty_buf(bp);
2443 	bqrelse(bp);
2444 
2445 	/*
2446 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2447 	 * due to the softdep code.
2448 	 */
2449 }
2450 
2451 /*
2452  *	bdirty:
2453  *
2454  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2455  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2456  *	itself to properly update it in the dirty/clean lists.  We mark it
2457  *	B_DONE to ensure that any asynchronization of the buffer properly
2458  *	clears B_DONE ( else a panic will occur later ).
2459  *
2460  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2461  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2462  *	should only be called if the buffer is known-good.
2463  *
2464  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2465  *	count.
2466  *
2467  *	The buffer must be on QUEUE_NONE.
2468  */
2469 void
2470 bdirty(struct buf *bp)
2471 {
2472 
2473 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2474 	    bp, bp->b_vp, bp->b_flags);
2475 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2476 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2477 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2478 	bp->b_flags &= ~(B_RELBUF);
2479 	bp->b_iocmd = BIO_WRITE;
2480 
2481 	if ((bp->b_flags & B_DELWRI) == 0) {
2482 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2483 		reassignbuf(bp);
2484 		bdirtyadd(bp);
2485 	}
2486 }
2487 
2488 /*
2489  *	bundirty:
2490  *
2491  *	Clear B_DELWRI for buffer.
2492  *
2493  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2494  *	count.
2495  *
2496  *	The buffer must be on QUEUE_NONE.
2497  */
2498 
2499 void
2500 bundirty(struct buf *bp)
2501 {
2502 
2503 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2504 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2505 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2506 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2507 
2508 	if (bp->b_flags & B_DELWRI) {
2509 		bp->b_flags &= ~B_DELWRI;
2510 		reassignbuf(bp);
2511 		bdirtysub(bp);
2512 	}
2513 	/*
2514 	 * Since it is now being written, we can clear its deferred write flag.
2515 	 */
2516 	bp->b_flags &= ~B_DEFERRED;
2517 }
2518 
2519 /*
2520  *	bawrite:
2521  *
2522  *	Asynchronous write.  Start output on a buffer, but do not wait for
2523  *	it to complete.  The buffer is released when the output completes.
2524  *
2525  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2526  *	B_INVAL buffers.  Not us.
2527  */
2528 void
2529 bawrite(struct buf *bp)
2530 {
2531 
2532 	bp->b_flags |= B_ASYNC;
2533 	(void) bwrite(bp);
2534 }
2535 
2536 /*
2537  *	babarrierwrite:
2538  *
2539  *	Asynchronous barrier write.  Start output on a buffer, but do not
2540  *	wait for it to complete.  Place a write barrier after this write so
2541  *	that this buffer and all buffers written before it are committed to
2542  *	the disk before any buffers written after this write are committed
2543  *	to the disk.  The buffer is released when the output completes.
2544  */
2545 void
2546 babarrierwrite(struct buf *bp)
2547 {
2548 
2549 	bp->b_flags |= B_ASYNC | B_BARRIER;
2550 	(void) bwrite(bp);
2551 }
2552 
2553 /*
2554  *	bbarrierwrite:
2555  *
2556  *	Synchronous barrier write.  Start output on a buffer and wait for
2557  *	it to complete.  Place a write barrier after this write so that
2558  *	this buffer and all buffers written before it are committed to
2559  *	the disk before any buffers written after this write are committed
2560  *	to the disk.  The buffer is released when the output completes.
2561  */
2562 int
2563 bbarrierwrite(struct buf *bp)
2564 {
2565 
2566 	bp->b_flags |= B_BARRIER;
2567 	return (bwrite(bp));
2568 }
2569 
2570 /*
2571  *	bwillwrite:
2572  *
2573  *	Called prior to the locking of any vnodes when we are expecting to
2574  *	write.  We do not want to starve the buffer cache with too many
2575  *	dirty buffers so we block here.  By blocking prior to the locking
2576  *	of any vnodes we attempt to avoid the situation where a locked vnode
2577  *	prevents the various system daemons from flushing related buffers.
2578  */
2579 void
2580 bwillwrite(void)
2581 {
2582 
2583 	if (buf_dirty_count_severe()) {
2584 		mtx_lock(&bdirtylock);
2585 		while (buf_dirty_count_severe()) {
2586 			bdirtywait = 1;
2587 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2588 			    "flswai", 0);
2589 		}
2590 		mtx_unlock(&bdirtylock);
2591 	}
2592 }
2593 
2594 /*
2595  * Return true if we have too many dirty buffers.
2596  */
2597 int
2598 buf_dirty_count_severe(void)
2599 {
2600 
2601 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2602 }
2603 
2604 /*
2605  *	brelse:
2606  *
2607  *	Release a busy buffer and, if requested, free its resources.  The
2608  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2609  *	to be accessed later as a cache entity or reused for other purposes.
2610  */
2611 void
2612 brelse(struct buf *bp)
2613 {
2614 	struct mount *v_mnt;
2615 	int qindex;
2616 
2617 	/*
2618 	 * Many functions erroneously call brelse with a NULL bp under rare
2619 	 * error conditions. Simply return when called with a NULL bp.
2620 	 */
2621 	if (bp == NULL)
2622 		return;
2623 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2624 	    bp, bp->b_vp, bp->b_flags);
2625 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2626 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2627 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2628 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2629 
2630 	if (BUF_LOCKRECURSED(bp)) {
2631 		/*
2632 		 * Do not process, in particular, do not handle the
2633 		 * B_INVAL/B_RELBUF and do not release to free list.
2634 		 */
2635 		BUF_UNLOCK(bp);
2636 		return;
2637 	}
2638 
2639 	if (bp->b_flags & B_MANAGED) {
2640 		bqrelse(bp);
2641 		return;
2642 	}
2643 
2644 	if (LIST_EMPTY(&bp->b_dep)) {
2645 		bp->b_flags &= ~B_IOSTARTED;
2646 	} else {
2647 		KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2648 		    ("brelse: SU io not finished bp %p", bp));
2649 	}
2650 
2651 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2652 		BO_LOCK(bp->b_bufobj);
2653 		bp->b_vflags &= ~BV_BKGRDERR;
2654 		BO_UNLOCK(bp->b_bufobj);
2655 		bdirty(bp);
2656 	}
2657 
2658 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2659 	    (bp->b_flags & B_INVALONERR)) {
2660 		/*
2661 		 * Forced invalidation of dirty buffer contents, to be used
2662 		 * after a failed write in the rare case that the loss of the
2663 		 * contents is acceptable.  The buffer is invalidated and
2664 		 * freed.
2665 		 */
2666 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2667 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2668 	}
2669 
2670 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2671 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2672 	    !(bp->b_flags & B_INVAL)) {
2673 		/*
2674 		 * Failed write, redirty.  All errors except ENXIO (which
2675 		 * means the device is gone) are treated as being
2676 		 * transient.
2677 		 *
2678 		 * XXX Treating EIO as transient is not correct; the
2679 		 * contract with the local storage device drivers is that
2680 		 * they will only return EIO once the I/O is no longer
2681 		 * retriable.  Network I/O also respects this through the
2682 		 * guarantees of TCP and/or the internal retries of NFS.
2683 		 * ENOMEM might be transient, but we also have no way of
2684 		 * knowing when its ok to retry/reschedule.  In general,
2685 		 * this entire case should be made obsolete through better
2686 		 * error handling/recovery and resource scheduling.
2687 		 *
2688 		 * Do this also for buffers that failed with ENXIO, but have
2689 		 * non-empty dependencies - the soft updates code might need
2690 		 * to access the buffer to untangle them.
2691 		 *
2692 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2693 		 */
2694 		bp->b_ioflags &= ~BIO_ERROR;
2695 		bdirty(bp);
2696 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2697 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2698 		/*
2699 		 * Either a failed read I/O, or we were asked to free or not
2700 		 * cache the buffer, or we failed to write to a device that's
2701 		 * no longer present.
2702 		 */
2703 		bp->b_flags |= B_INVAL;
2704 		if (!LIST_EMPTY(&bp->b_dep))
2705 			buf_deallocate(bp);
2706 		if (bp->b_flags & B_DELWRI)
2707 			bdirtysub(bp);
2708 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2709 		if ((bp->b_flags & B_VMIO) == 0) {
2710 			allocbuf(bp, 0);
2711 			if (bp->b_vp)
2712 				brelvp(bp);
2713 		}
2714 	}
2715 
2716 	/*
2717 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2718 	 * is called with B_DELWRI set, the underlying pages may wind up
2719 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2720 	 * because pages associated with a B_DELWRI bp are marked clean.
2721 	 *
2722 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2723 	 * if B_DELWRI is set.
2724 	 */
2725 	if (bp->b_flags & B_DELWRI)
2726 		bp->b_flags &= ~B_RELBUF;
2727 
2728 	/*
2729 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2730 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2731 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2732 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2733 	 *
2734 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2735 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2736 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2737 	 *
2738 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2739 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2740 	 * the commit state and we cannot afford to lose the buffer. If the
2741 	 * buffer has a background write in progress, we need to keep it
2742 	 * around to prevent it from being reconstituted and starting a second
2743 	 * background write.
2744 	 */
2745 
2746 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2747 
2748 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2749 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2750 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2751 	    vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2752 		vfs_vmio_invalidate(bp);
2753 		allocbuf(bp, 0);
2754 	}
2755 
2756 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2757 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2758 		allocbuf(bp, 0);
2759 		bp->b_flags &= ~B_NOREUSE;
2760 		if (bp->b_vp != NULL)
2761 			brelvp(bp);
2762 	}
2763 
2764 	/*
2765 	 * If the buffer has junk contents signal it and eventually
2766 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2767 	 * doesn't find it.
2768 	 */
2769 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2770 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2771 		bp->b_flags |= B_INVAL;
2772 	if (bp->b_flags & B_INVAL) {
2773 		if (bp->b_flags & B_DELWRI)
2774 			bundirty(bp);
2775 		if (bp->b_vp)
2776 			brelvp(bp);
2777 	}
2778 
2779 	buf_track(bp, __func__);
2780 
2781 	/* buffers with no memory */
2782 	if (bp->b_bufsize == 0) {
2783 		buf_free(bp);
2784 		return;
2785 	}
2786 	/* buffers with junk contents */
2787 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2788 	    (bp->b_ioflags & BIO_ERROR)) {
2789 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2790 		if (bp->b_vflags & BV_BKGRDINPROG)
2791 			panic("losing buffer 2");
2792 		qindex = QUEUE_CLEAN;
2793 		bp->b_flags |= B_AGE;
2794 	/* remaining buffers */
2795 	} else if (bp->b_flags & B_DELWRI)
2796 		qindex = QUEUE_DIRTY;
2797 	else
2798 		qindex = QUEUE_CLEAN;
2799 
2800 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2801 		panic("brelse: not dirty");
2802 
2803 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2804 	bp->b_xflags &= ~(BX_CVTENXIO);
2805 	/* binsfree unlocks bp. */
2806 	binsfree(bp, qindex);
2807 }
2808 
2809 /*
2810  * Release a buffer back to the appropriate queue but do not try to free
2811  * it.  The buffer is expected to be used again soon.
2812  *
2813  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2814  * biodone() to requeue an async I/O on completion.  It is also used when
2815  * known good buffers need to be requeued but we think we may need the data
2816  * again soon.
2817  *
2818  * XXX we should be able to leave the B_RELBUF hint set on completion.
2819  */
2820 void
2821 bqrelse(struct buf *bp)
2822 {
2823 	int qindex;
2824 
2825 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2826 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2827 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2828 
2829 	qindex = QUEUE_NONE;
2830 	if (BUF_LOCKRECURSED(bp)) {
2831 		/* do not release to free list */
2832 		BUF_UNLOCK(bp);
2833 		return;
2834 	}
2835 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2836 	bp->b_xflags &= ~(BX_CVTENXIO);
2837 
2838 	if (LIST_EMPTY(&bp->b_dep)) {
2839 		bp->b_flags &= ~B_IOSTARTED;
2840 	} else {
2841 		KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2842 		    ("bqrelse: SU io not finished bp %p", bp));
2843 	}
2844 
2845 	if (bp->b_flags & B_MANAGED) {
2846 		if (bp->b_flags & B_REMFREE)
2847 			bremfreef(bp);
2848 		goto out;
2849 	}
2850 
2851 	/* buffers with stale but valid contents */
2852 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2853 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2854 		BO_LOCK(bp->b_bufobj);
2855 		bp->b_vflags &= ~BV_BKGRDERR;
2856 		BO_UNLOCK(bp->b_bufobj);
2857 		qindex = QUEUE_DIRTY;
2858 	} else {
2859 		if ((bp->b_flags & B_DELWRI) == 0 &&
2860 		    (bp->b_xflags & BX_VNDIRTY))
2861 			panic("bqrelse: not dirty");
2862 		if ((bp->b_flags & B_NOREUSE) != 0) {
2863 			brelse(bp);
2864 			return;
2865 		}
2866 		qindex = QUEUE_CLEAN;
2867 	}
2868 	buf_track(bp, __func__);
2869 	/* binsfree unlocks bp. */
2870 	binsfree(bp, qindex);
2871 	return;
2872 
2873 out:
2874 	buf_track(bp, __func__);
2875 	/* unlock */
2876 	BUF_UNLOCK(bp);
2877 }
2878 
2879 /*
2880  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2881  * restore bogus pages.
2882  */
2883 static void
2884 vfs_vmio_iodone(struct buf *bp)
2885 {
2886 	vm_ooffset_t foff;
2887 	vm_page_t m;
2888 	vm_object_t obj;
2889 	struct vnode *vp __unused;
2890 	int i, iosize, resid;
2891 	bool bogus;
2892 
2893 	obj = bp->b_bufobj->bo_object;
2894 	KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2895 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2896 	    blockcount_read(&obj->paging_in_progress), bp->b_npages));
2897 
2898 	vp = bp->b_vp;
2899 	VNPASS(vp->v_holdcnt > 0, vp);
2900 	VNPASS(vp->v_object != NULL, vp);
2901 
2902 	foff = bp->b_offset;
2903 	KASSERT(bp->b_offset != NOOFFSET,
2904 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2905 
2906 	bogus = false;
2907 	iosize = bp->b_bcount - bp->b_resid;
2908 	for (i = 0; i < bp->b_npages; i++) {
2909 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2910 		if (resid > iosize)
2911 			resid = iosize;
2912 
2913 		/*
2914 		 * cleanup bogus pages, restoring the originals
2915 		 */
2916 		m = bp->b_pages[i];
2917 		if (m == bogus_page) {
2918 			bogus = true;
2919 			m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2920 			if (m == NULL)
2921 				panic("biodone: page disappeared!");
2922 			bp->b_pages[i] = m;
2923 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2924 			/*
2925 			 * In the write case, the valid and clean bits are
2926 			 * already changed correctly ( see bdwrite() ), so we
2927 			 * only need to do this here in the read case.
2928 			 */
2929 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2930 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2931 			    "has unexpected dirty bits", m));
2932 			vfs_page_set_valid(bp, foff, m);
2933 		}
2934 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2935 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2936 		    (intmax_t)foff, (uintmax_t)m->pindex));
2937 
2938 		vm_page_sunbusy(m);
2939 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2940 		iosize -= resid;
2941 	}
2942 	vm_object_pip_wakeupn(obj, bp->b_npages);
2943 	if (bogus && buf_mapped(bp)) {
2944 		BUF_CHECK_MAPPED(bp);
2945 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2946 		    bp->b_pages, bp->b_npages);
2947 	}
2948 }
2949 
2950 /*
2951  * Perform page invalidation when a buffer is released.  The fully invalid
2952  * pages will be reclaimed later in vfs_vmio_truncate().
2953  */
2954 static void
2955 vfs_vmio_invalidate(struct buf *bp)
2956 {
2957 	vm_object_t obj;
2958 	vm_page_t m;
2959 	int flags, i, resid, poffset, presid;
2960 
2961 	if (buf_mapped(bp)) {
2962 		BUF_CHECK_MAPPED(bp);
2963 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2964 	} else
2965 		BUF_CHECK_UNMAPPED(bp);
2966 	/*
2967 	 * Get the base offset and length of the buffer.  Note that
2968 	 * in the VMIO case if the buffer block size is not
2969 	 * page-aligned then b_data pointer may not be page-aligned.
2970 	 * But our b_pages[] array *IS* page aligned.
2971 	 *
2972 	 * block sizes less then DEV_BSIZE (usually 512) are not
2973 	 * supported due to the page granularity bits (m->valid,
2974 	 * m->dirty, etc...).
2975 	 *
2976 	 * See man buf(9) for more information
2977 	 */
2978 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2979 	obj = bp->b_bufobj->bo_object;
2980 	resid = bp->b_bufsize;
2981 	poffset = bp->b_offset & PAGE_MASK;
2982 	VM_OBJECT_WLOCK(obj);
2983 	for (i = 0; i < bp->b_npages; i++) {
2984 		m = bp->b_pages[i];
2985 		if (m == bogus_page)
2986 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2987 		bp->b_pages[i] = NULL;
2988 
2989 		presid = resid > (PAGE_SIZE - poffset) ?
2990 		    (PAGE_SIZE - poffset) : resid;
2991 		KASSERT(presid >= 0, ("brelse: extra page"));
2992 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
2993 		if (pmap_page_wired_mappings(m) == 0)
2994 			vm_page_set_invalid(m, poffset, presid);
2995 		vm_page_sunbusy(m);
2996 		vm_page_release_locked(m, flags);
2997 		resid -= presid;
2998 		poffset = 0;
2999 	}
3000 	VM_OBJECT_WUNLOCK(obj);
3001 	bp->b_npages = 0;
3002 }
3003 
3004 /*
3005  * Page-granular truncation of an existing VMIO buffer.
3006  */
3007 static void
3008 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3009 {
3010 	vm_object_t obj;
3011 	vm_page_t m;
3012 	int flags, i;
3013 
3014 	if (bp->b_npages == desiredpages)
3015 		return;
3016 
3017 	if (buf_mapped(bp)) {
3018 		BUF_CHECK_MAPPED(bp);
3019 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3020 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3021 	} else
3022 		BUF_CHECK_UNMAPPED(bp);
3023 
3024 	/*
3025 	 * The object lock is needed only if we will attempt to free pages.
3026 	 */
3027 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3028 	if ((bp->b_flags & B_DIRECT) != 0) {
3029 		flags |= VPR_TRYFREE;
3030 		obj = bp->b_bufobj->bo_object;
3031 		VM_OBJECT_WLOCK(obj);
3032 	} else {
3033 		obj = NULL;
3034 	}
3035 	for (i = desiredpages; i < bp->b_npages; i++) {
3036 		m = bp->b_pages[i];
3037 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3038 		bp->b_pages[i] = NULL;
3039 		if (obj != NULL)
3040 			vm_page_release_locked(m, flags);
3041 		else
3042 			vm_page_release(m, flags);
3043 	}
3044 	if (obj != NULL)
3045 		VM_OBJECT_WUNLOCK(obj);
3046 	bp->b_npages = desiredpages;
3047 }
3048 
3049 /*
3050  * Byte granular extension of VMIO buffers.
3051  */
3052 static void
3053 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3054 {
3055 	/*
3056 	 * We are growing the buffer, possibly in a
3057 	 * byte-granular fashion.
3058 	 */
3059 	vm_object_t obj;
3060 	vm_offset_t toff;
3061 	vm_offset_t tinc;
3062 	vm_page_t m;
3063 
3064 	/*
3065 	 * Step 1, bring in the VM pages from the object, allocating
3066 	 * them if necessary.  We must clear B_CACHE if these pages
3067 	 * are not valid for the range covered by the buffer.
3068 	 */
3069 	obj = bp->b_bufobj->bo_object;
3070 	if (bp->b_npages < desiredpages) {
3071 		KASSERT(desiredpages <= atop(maxbcachebuf),
3072 		    ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3073 		    bp, desiredpages, maxbcachebuf));
3074 
3075 		/*
3076 		 * We must allocate system pages since blocking
3077 		 * here could interfere with paging I/O, no
3078 		 * matter which process we are.
3079 		 *
3080 		 * Only exclusive busy can be tested here.
3081 		 * Blocking on shared busy might lead to
3082 		 * deadlocks once allocbuf() is called after
3083 		 * pages are vfs_busy_pages().
3084 		 */
3085 		(void)vm_page_grab_pages_unlocked(obj,
3086 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3087 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3088 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3089 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3090 		bp->b_npages = desiredpages;
3091 	}
3092 
3093 	/*
3094 	 * Step 2.  We've loaded the pages into the buffer,
3095 	 * we have to figure out if we can still have B_CACHE
3096 	 * set.  Note that B_CACHE is set according to the
3097 	 * byte-granular range ( bcount and size ), not the
3098 	 * aligned range ( newbsize ).
3099 	 *
3100 	 * The VM test is against m->valid, which is DEV_BSIZE
3101 	 * aligned.  Needless to say, the validity of the data
3102 	 * needs to also be DEV_BSIZE aligned.  Note that this
3103 	 * fails with NFS if the server or some other client
3104 	 * extends the file's EOF.  If our buffer is resized,
3105 	 * B_CACHE may remain set! XXX
3106 	 */
3107 	toff = bp->b_bcount;
3108 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3109 	while ((bp->b_flags & B_CACHE) && toff < size) {
3110 		vm_pindex_t pi;
3111 
3112 		if (tinc > (size - toff))
3113 			tinc = size - toff;
3114 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3115 		m = bp->b_pages[pi];
3116 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3117 		toff += tinc;
3118 		tinc = PAGE_SIZE;
3119 	}
3120 
3121 	/*
3122 	 * Step 3, fixup the KVA pmap.
3123 	 */
3124 	if (buf_mapped(bp))
3125 		bpmap_qenter(bp);
3126 	else
3127 		BUF_CHECK_UNMAPPED(bp);
3128 }
3129 
3130 /*
3131  * Check to see if a block at a particular lbn is available for a clustered
3132  * write.
3133  */
3134 static int
3135 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3136 {
3137 	struct buf *bpa;
3138 	int match;
3139 
3140 	match = 0;
3141 
3142 	/* If the buf isn't in core skip it */
3143 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3144 		return (0);
3145 
3146 	/* If the buf is busy we don't want to wait for it */
3147 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3148 		return (0);
3149 
3150 	/* Only cluster with valid clusterable delayed write buffers */
3151 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3152 	    (B_DELWRI | B_CLUSTEROK))
3153 		goto done;
3154 
3155 	if (bpa->b_bufsize != size)
3156 		goto done;
3157 
3158 	/*
3159 	 * Check to see if it is in the expected place on disk and that the
3160 	 * block has been mapped.
3161 	 */
3162 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3163 		match = 1;
3164 done:
3165 	BUF_UNLOCK(bpa);
3166 	return (match);
3167 }
3168 
3169 /*
3170  *	vfs_bio_awrite:
3171  *
3172  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3173  *	This is much better then the old way of writing only one buffer at
3174  *	a time.  Note that we may not be presented with the buffers in the
3175  *	correct order, so we search for the cluster in both directions.
3176  */
3177 int
3178 vfs_bio_awrite(struct buf *bp)
3179 {
3180 	struct bufobj *bo;
3181 	int i;
3182 	int j;
3183 	daddr_t lblkno = bp->b_lblkno;
3184 	struct vnode *vp = bp->b_vp;
3185 	int ncl;
3186 	int nwritten;
3187 	int size;
3188 	int maxcl;
3189 	int gbflags;
3190 
3191 	bo = &vp->v_bufobj;
3192 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3193 	/*
3194 	 * right now we support clustered writing only to regular files.  If
3195 	 * we find a clusterable block we could be in the middle of a cluster
3196 	 * rather then at the beginning.
3197 	 */
3198 	if ((vp->v_type == VREG) &&
3199 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3200 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3201 		size = vp->v_mount->mnt_stat.f_iosize;
3202 		maxcl = maxphys / size;
3203 
3204 		BO_RLOCK(bo);
3205 		for (i = 1; i < maxcl; i++)
3206 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3207 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3208 				break;
3209 
3210 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3211 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3212 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3213 				break;
3214 		BO_RUNLOCK(bo);
3215 		--j;
3216 		ncl = i + j;
3217 		/*
3218 		 * this is a possible cluster write
3219 		 */
3220 		if (ncl != 1) {
3221 			BUF_UNLOCK(bp);
3222 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3223 			    gbflags);
3224 			return (nwritten);
3225 		}
3226 	}
3227 	bremfree(bp);
3228 	bp->b_flags |= B_ASYNC;
3229 	/*
3230 	 * default (old) behavior, writing out only one block
3231 	 *
3232 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3233 	 */
3234 	nwritten = bp->b_bufsize;
3235 	(void) bwrite(bp);
3236 
3237 	return (nwritten);
3238 }
3239 
3240 /*
3241  *	getnewbuf_kva:
3242  *
3243  *	Allocate KVA for an empty buf header according to gbflags.
3244  */
3245 static int
3246 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3247 {
3248 
3249 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3250 		/*
3251 		 * In order to keep fragmentation sane we only allocate kva
3252 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3253 		 */
3254 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3255 
3256 		if (maxsize != bp->b_kvasize &&
3257 		    bufkva_alloc(bp, maxsize, gbflags))
3258 			return (ENOSPC);
3259 	}
3260 	return (0);
3261 }
3262 
3263 /*
3264  *	getnewbuf:
3265  *
3266  *	Find and initialize a new buffer header, freeing up existing buffers
3267  *	in the bufqueues as necessary.  The new buffer is returned locked.
3268  *
3269  *	We block if:
3270  *		We have insufficient buffer headers
3271  *		We have insufficient buffer space
3272  *		buffer_arena is too fragmented ( space reservation fails )
3273  *		If we have to flush dirty buffers ( but we try to avoid this )
3274  *
3275  *	The caller is responsible for releasing the reserved bufspace after
3276  *	allocbuf() is called.
3277  */
3278 static struct buf *
3279 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3280 {
3281 	struct bufdomain *bd;
3282 	struct buf *bp;
3283 	bool metadata, reserved;
3284 
3285 	bp = NULL;
3286 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3287 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3288 	if (!unmapped_buf_allowed)
3289 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3290 
3291 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3292 	    vp->v_type == VCHR)
3293 		metadata = true;
3294 	else
3295 		metadata = false;
3296 	if (vp == NULL)
3297 		bd = &bdomain[0];
3298 	else
3299 		bd = &bdomain[vp->v_bufobj.bo_domain];
3300 
3301 	counter_u64_add(getnewbufcalls, 1);
3302 	reserved = false;
3303 	do {
3304 		if (reserved == false &&
3305 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3306 			counter_u64_add(getnewbufrestarts, 1);
3307 			continue;
3308 		}
3309 		reserved = true;
3310 		if ((bp = buf_alloc(bd)) == NULL) {
3311 			counter_u64_add(getnewbufrestarts, 1);
3312 			continue;
3313 		}
3314 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3315 			return (bp);
3316 		break;
3317 	} while (buf_recycle(bd, false) == 0);
3318 
3319 	if (reserved)
3320 		bufspace_release(bd, maxsize);
3321 	if (bp != NULL) {
3322 		bp->b_flags |= B_INVAL;
3323 		brelse(bp);
3324 	}
3325 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3326 
3327 	return (NULL);
3328 }
3329 
3330 /*
3331  *	buf_daemon:
3332  *
3333  *	buffer flushing daemon.  Buffers are normally flushed by the
3334  *	update daemon but if it cannot keep up this process starts to
3335  *	take the load in an attempt to prevent getnewbuf() from blocking.
3336  */
3337 static struct kproc_desc buf_kp = {
3338 	"bufdaemon",
3339 	buf_daemon,
3340 	&bufdaemonproc
3341 };
3342 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3343 
3344 static int
3345 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3346 {
3347 	int flushed;
3348 
3349 	flushed = flushbufqueues(vp, bd, target, 0);
3350 	if (flushed == 0) {
3351 		/*
3352 		 * Could not find any buffers without rollback
3353 		 * dependencies, so just write the first one
3354 		 * in the hopes of eventually making progress.
3355 		 */
3356 		if (vp != NULL && target > 2)
3357 			target /= 2;
3358 		flushbufqueues(vp, bd, target, 1);
3359 	}
3360 	return (flushed);
3361 }
3362 
3363 static void
3364 buf_daemon()
3365 {
3366 	struct bufdomain *bd;
3367 	int speedupreq;
3368 	int lodirty;
3369 	int i;
3370 
3371 	/*
3372 	 * This process needs to be suspended prior to shutdown sync.
3373 	 */
3374 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3375 	    SHUTDOWN_PRI_LAST + 100);
3376 
3377 	/*
3378 	 * Start the buf clean daemons as children threads.
3379 	 */
3380 	for (i = 0 ; i < buf_domains; i++) {
3381 		int error;
3382 
3383 		error = kthread_add((void (*)(void *))bufspace_daemon,
3384 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3385 		if (error)
3386 			panic("error %d spawning bufspace daemon", error);
3387 	}
3388 
3389 	/*
3390 	 * This process is allowed to take the buffer cache to the limit
3391 	 */
3392 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3393 	mtx_lock(&bdlock);
3394 	for (;;) {
3395 		bd_request = 0;
3396 		mtx_unlock(&bdlock);
3397 
3398 		kthread_suspend_check();
3399 
3400 		/*
3401 		 * Save speedupreq for this pass and reset to capture new
3402 		 * requests.
3403 		 */
3404 		speedupreq = bd_speedupreq;
3405 		bd_speedupreq = 0;
3406 
3407 		/*
3408 		 * Flush each domain sequentially according to its level and
3409 		 * the speedup request.
3410 		 */
3411 		for (i = 0; i < buf_domains; i++) {
3412 			bd = &bdomain[i];
3413 			if (speedupreq)
3414 				lodirty = bd->bd_numdirtybuffers / 2;
3415 			else
3416 				lodirty = bd->bd_lodirtybuffers;
3417 			while (bd->bd_numdirtybuffers > lodirty) {
3418 				if (buf_flush(NULL, bd,
3419 				    bd->bd_numdirtybuffers - lodirty) == 0)
3420 					break;
3421 				kern_yield(PRI_USER);
3422 			}
3423 		}
3424 
3425 		/*
3426 		 * Only clear bd_request if we have reached our low water
3427 		 * mark.  The buf_daemon normally waits 1 second and
3428 		 * then incrementally flushes any dirty buffers that have
3429 		 * built up, within reason.
3430 		 *
3431 		 * If we were unable to hit our low water mark and couldn't
3432 		 * find any flushable buffers, we sleep for a short period
3433 		 * to avoid endless loops on unlockable buffers.
3434 		 */
3435 		mtx_lock(&bdlock);
3436 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3437 			/*
3438 			 * We reached our low water mark, reset the
3439 			 * request and sleep until we are needed again.
3440 			 * The sleep is just so the suspend code works.
3441 			 */
3442 			bd_request = 0;
3443 			/*
3444 			 * Do an extra wakeup in case dirty threshold
3445 			 * changed via sysctl and the explicit transition
3446 			 * out of shortfall was missed.
3447 			 */
3448 			bdirtywakeup();
3449 			if (runningbufspace <= lorunningspace)
3450 				runningwakeup();
3451 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3452 		} else {
3453 			/*
3454 			 * We couldn't find any flushable dirty buffers but
3455 			 * still have too many dirty buffers, we
3456 			 * have to sleep and try again.  (rare)
3457 			 */
3458 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3459 		}
3460 	}
3461 }
3462 
3463 /*
3464  *	flushbufqueues:
3465  *
3466  *	Try to flush a buffer in the dirty queue.  We must be careful to
3467  *	free up B_INVAL buffers instead of write them, which NFS is
3468  *	particularly sensitive to.
3469  */
3470 static int flushwithdeps = 0;
3471 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3472     &flushwithdeps, 0,
3473     "Number of buffers flushed with dependecies that require rollbacks");
3474 
3475 static int
3476 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3477     int flushdeps)
3478 {
3479 	struct bufqueue *bq;
3480 	struct buf *sentinel;
3481 	struct vnode *vp;
3482 	struct mount *mp;
3483 	struct buf *bp;
3484 	int hasdeps;
3485 	int flushed;
3486 	int error;
3487 	bool unlock;
3488 
3489 	flushed = 0;
3490 	bq = &bd->bd_dirtyq;
3491 	bp = NULL;
3492 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3493 	sentinel->b_qindex = QUEUE_SENTINEL;
3494 	BQ_LOCK(bq);
3495 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3496 	BQ_UNLOCK(bq);
3497 	while (flushed != target) {
3498 		maybe_yield();
3499 		BQ_LOCK(bq);
3500 		bp = TAILQ_NEXT(sentinel, b_freelist);
3501 		if (bp != NULL) {
3502 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3503 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3504 			    b_freelist);
3505 		} else {
3506 			BQ_UNLOCK(bq);
3507 			break;
3508 		}
3509 		/*
3510 		 * Skip sentinels inserted by other invocations of the
3511 		 * flushbufqueues(), taking care to not reorder them.
3512 		 *
3513 		 * Only flush the buffers that belong to the
3514 		 * vnode locked by the curthread.
3515 		 */
3516 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3517 		    bp->b_vp != lvp)) {
3518 			BQ_UNLOCK(bq);
3519 			continue;
3520 		}
3521 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3522 		BQ_UNLOCK(bq);
3523 		if (error != 0)
3524 			continue;
3525 
3526 		/*
3527 		 * BKGRDINPROG can only be set with the buf and bufobj
3528 		 * locks both held.  We tolerate a race to clear it here.
3529 		 */
3530 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3531 		    (bp->b_flags & B_DELWRI) == 0) {
3532 			BUF_UNLOCK(bp);
3533 			continue;
3534 		}
3535 		if (bp->b_flags & B_INVAL) {
3536 			bremfreef(bp);
3537 			brelse(bp);
3538 			flushed++;
3539 			continue;
3540 		}
3541 
3542 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3543 			if (flushdeps == 0) {
3544 				BUF_UNLOCK(bp);
3545 				continue;
3546 			}
3547 			hasdeps = 1;
3548 		} else
3549 			hasdeps = 0;
3550 		/*
3551 		 * We must hold the lock on a vnode before writing
3552 		 * one of its buffers. Otherwise we may confuse, or
3553 		 * in the case of a snapshot vnode, deadlock the
3554 		 * system.
3555 		 *
3556 		 * The lock order here is the reverse of the normal
3557 		 * of vnode followed by buf lock.  This is ok because
3558 		 * the NOWAIT will prevent deadlock.
3559 		 */
3560 		vp = bp->b_vp;
3561 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3562 			BUF_UNLOCK(bp);
3563 			continue;
3564 		}
3565 		if (lvp == NULL) {
3566 			unlock = true;
3567 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3568 		} else {
3569 			ASSERT_VOP_LOCKED(vp, "getbuf");
3570 			unlock = false;
3571 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3572 			    vn_lock(vp, LK_TRYUPGRADE);
3573 		}
3574 		if (error == 0) {
3575 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3576 			    bp, bp->b_vp, bp->b_flags);
3577 			if (curproc == bufdaemonproc) {
3578 				vfs_bio_awrite(bp);
3579 			} else {
3580 				bremfree(bp);
3581 				bwrite(bp);
3582 				counter_u64_add(notbufdflushes, 1);
3583 			}
3584 			vn_finished_write(mp);
3585 			if (unlock)
3586 				VOP_UNLOCK(vp);
3587 			flushwithdeps += hasdeps;
3588 			flushed++;
3589 
3590 			/*
3591 			 * Sleeping on runningbufspace while holding
3592 			 * vnode lock leads to deadlock.
3593 			 */
3594 			if (curproc == bufdaemonproc &&
3595 			    runningbufspace > hirunningspace)
3596 				waitrunningbufspace();
3597 			continue;
3598 		}
3599 		vn_finished_write(mp);
3600 		BUF_UNLOCK(bp);
3601 	}
3602 	BQ_LOCK(bq);
3603 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3604 	BQ_UNLOCK(bq);
3605 	free(sentinel, M_TEMP);
3606 	return (flushed);
3607 }
3608 
3609 /*
3610  * Check to see if a block is currently memory resident.
3611  */
3612 struct buf *
3613 incore(struct bufobj *bo, daddr_t blkno)
3614 {
3615 	return (gbincore_unlocked(bo, blkno));
3616 }
3617 
3618 /*
3619  * Returns true if no I/O is needed to access the
3620  * associated VM object.  This is like incore except
3621  * it also hunts around in the VM system for the data.
3622  */
3623 bool
3624 inmem(struct vnode * vp, daddr_t blkno)
3625 {
3626 	vm_object_t obj;
3627 	vm_offset_t toff, tinc, size;
3628 	vm_page_t m, n;
3629 	vm_ooffset_t off;
3630 	int valid;
3631 
3632 	ASSERT_VOP_LOCKED(vp, "inmem");
3633 
3634 	if (incore(&vp->v_bufobj, blkno))
3635 		return (true);
3636 	if (vp->v_mount == NULL)
3637 		return (false);
3638 	obj = vp->v_object;
3639 	if (obj == NULL)
3640 		return (false);
3641 
3642 	size = PAGE_SIZE;
3643 	if (size > vp->v_mount->mnt_stat.f_iosize)
3644 		size = vp->v_mount->mnt_stat.f_iosize;
3645 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3646 
3647 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3648 		m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3649 recheck:
3650 		if (m == NULL)
3651 			return (false);
3652 
3653 		tinc = size;
3654 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3655 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3656 		/*
3657 		 * Consider page validity only if page mapping didn't change
3658 		 * during the check.
3659 		 */
3660 		valid = vm_page_is_valid(m,
3661 		    (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3662 		n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3663 		if (m != n) {
3664 			m = n;
3665 			goto recheck;
3666 		}
3667 		if (!valid)
3668 			return (false);
3669 	}
3670 	return (true);
3671 }
3672 
3673 /*
3674  * Set the dirty range for a buffer based on the status of the dirty
3675  * bits in the pages comprising the buffer.  The range is limited
3676  * to the size of the buffer.
3677  *
3678  * Tell the VM system that the pages associated with this buffer
3679  * are clean.  This is used for delayed writes where the data is
3680  * going to go to disk eventually without additional VM intevention.
3681  *
3682  * Note that while we only really need to clean through to b_bcount, we
3683  * just go ahead and clean through to b_bufsize.
3684  */
3685 static void
3686 vfs_clean_pages_dirty_buf(struct buf *bp)
3687 {
3688 	vm_ooffset_t foff, noff, eoff;
3689 	vm_page_t m;
3690 	int i;
3691 
3692 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3693 		return;
3694 
3695 	foff = bp->b_offset;
3696 	KASSERT(bp->b_offset != NOOFFSET,
3697 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3698 
3699 	vfs_busy_pages_acquire(bp);
3700 	vfs_setdirty_range(bp);
3701 	for (i = 0; i < bp->b_npages; i++) {
3702 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3703 		eoff = noff;
3704 		if (eoff > bp->b_offset + bp->b_bufsize)
3705 			eoff = bp->b_offset + bp->b_bufsize;
3706 		m = bp->b_pages[i];
3707 		vfs_page_set_validclean(bp, foff, m);
3708 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3709 		foff = noff;
3710 	}
3711 	vfs_busy_pages_release(bp);
3712 }
3713 
3714 static void
3715 vfs_setdirty_range(struct buf *bp)
3716 {
3717 	vm_offset_t boffset;
3718 	vm_offset_t eoffset;
3719 	int i;
3720 
3721 	/*
3722 	 * test the pages to see if they have been modified directly
3723 	 * by users through the VM system.
3724 	 */
3725 	for (i = 0; i < bp->b_npages; i++)
3726 		vm_page_test_dirty(bp->b_pages[i]);
3727 
3728 	/*
3729 	 * Calculate the encompassing dirty range, boffset and eoffset,
3730 	 * (eoffset - boffset) bytes.
3731 	 */
3732 
3733 	for (i = 0; i < bp->b_npages; i++) {
3734 		if (bp->b_pages[i]->dirty)
3735 			break;
3736 	}
3737 	boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3738 
3739 	for (i = bp->b_npages - 1; i >= 0; --i) {
3740 		if (bp->b_pages[i]->dirty) {
3741 			break;
3742 		}
3743 	}
3744 	eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3745 
3746 	/*
3747 	 * Fit it to the buffer.
3748 	 */
3749 
3750 	if (eoffset > bp->b_bcount)
3751 		eoffset = bp->b_bcount;
3752 
3753 	/*
3754 	 * If we have a good dirty range, merge with the existing
3755 	 * dirty range.
3756 	 */
3757 
3758 	if (boffset < eoffset) {
3759 		if (bp->b_dirtyoff > boffset)
3760 			bp->b_dirtyoff = boffset;
3761 		if (bp->b_dirtyend < eoffset)
3762 			bp->b_dirtyend = eoffset;
3763 	}
3764 }
3765 
3766 /*
3767  * Allocate the KVA mapping for an existing buffer.
3768  * If an unmapped buffer is provided but a mapped buffer is requested, take
3769  * also care to properly setup mappings between pages and KVA.
3770  */
3771 static void
3772 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3773 {
3774 	int bsize, maxsize, need_mapping, need_kva;
3775 	off_t offset;
3776 
3777 	need_mapping = bp->b_data == unmapped_buf &&
3778 	    (gbflags & GB_UNMAPPED) == 0;
3779 	need_kva = bp->b_kvabase == unmapped_buf &&
3780 	    bp->b_data == unmapped_buf &&
3781 	    (gbflags & GB_KVAALLOC) != 0;
3782 	if (!need_mapping && !need_kva)
3783 		return;
3784 
3785 	BUF_CHECK_UNMAPPED(bp);
3786 
3787 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3788 		/*
3789 		 * Buffer is not mapped, but the KVA was already
3790 		 * reserved at the time of the instantiation.  Use the
3791 		 * allocated space.
3792 		 */
3793 		goto has_addr;
3794 	}
3795 
3796 	/*
3797 	 * Calculate the amount of the address space we would reserve
3798 	 * if the buffer was mapped.
3799 	 */
3800 	bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3801 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3802 	offset = blkno * bsize;
3803 	maxsize = size + (offset & PAGE_MASK);
3804 	maxsize = imax(maxsize, bsize);
3805 
3806 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3807 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3808 			/*
3809 			 * XXXKIB: defragmentation cannot
3810 			 * succeed, not sure what else to do.
3811 			 */
3812 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3813 		}
3814 		counter_u64_add(mappingrestarts, 1);
3815 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3816 	}
3817 has_addr:
3818 	if (need_mapping) {
3819 		/* b_offset is handled by bpmap_qenter. */
3820 		bp->b_data = bp->b_kvabase;
3821 		BUF_CHECK_MAPPED(bp);
3822 		bpmap_qenter(bp);
3823 	}
3824 }
3825 
3826 struct buf *
3827 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3828     int flags)
3829 {
3830 	struct buf *bp;
3831 	int error;
3832 
3833 	error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3834 	if (error != 0)
3835 		return (NULL);
3836 	return (bp);
3837 }
3838 
3839 /*
3840  *	getblkx:
3841  *
3842  *	Get a block given a specified block and offset into a file/device.
3843  *	The buffers B_DONE bit will be cleared on return, making it almost
3844  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3845  *	return.  The caller should clear B_INVAL prior to initiating a
3846  *	READ.
3847  *
3848  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3849  *	an existing buffer.
3850  *
3851  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3852  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3853  *	and then cleared based on the backing VM.  If the previous buffer is
3854  *	non-0-sized but invalid, B_CACHE will be cleared.
3855  *
3856  *	If getblk() must create a new buffer, the new buffer is returned with
3857  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3858  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3859  *	backing VM.
3860  *
3861  *	getblk() also forces a bwrite() for any B_DELWRI buffer whose
3862  *	B_CACHE bit is clear.
3863  *
3864  *	What this means, basically, is that the caller should use B_CACHE to
3865  *	determine whether the buffer is fully valid or not and should clear
3866  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3867  *	the buffer by loading its data area with something, the caller needs
3868  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3869  *	the caller should set B_CACHE ( as an optimization ), else the caller
3870  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3871  *	a write attempt or if it was a successful read.  If the caller
3872  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3873  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3874  *
3875  *	The blkno parameter is the logical block being requested. Normally
3876  *	the mapping of logical block number to disk block address is done
3877  *	by calling VOP_BMAP(). However, if the mapping is already known, the
3878  *	disk block address can be passed using the dblkno parameter. If the
3879  *	disk block address is not known, then the same value should be passed
3880  *	for blkno and dblkno.
3881  */
3882 int
3883 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3884     int slptimeo, int flags, struct buf **bpp)
3885 {
3886 	struct buf *bp;
3887 	struct bufobj *bo;
3888 	daddr_t d_blkno;
3889 	int bsize, error, maxsize, vmio;
3890 	off_t offset;
3891 
3892 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3893 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3894 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3895 	ASSERT_VOP_LOCKED(vp, "getblk");
3896 	if (size > maxbcachebuf)
3897 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3898 		    maxbcachebuf);
3899 	if (!unmapped_buf_allowed)
3900 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3901 
3902 	bo = &vp->v_bufobj;
3903 	d_blkno = dblkno;
3904 
3905 	/* Attempt lockless lookup first. */
3906 	bp = gbincore_unlocked(bo, blkno);
3907 	if (bp == NULL) {
3908 		/*
3909 		 * With GB_NOCREAT we must be sure about not finding the buffer
3910 		 * as it may have been reassigned during unlocked lookup.
3911 		 */
3912 		if ((flags & GB_NOCREAT) != 0)
3913 			goto loop;
3914 		goto newbuf_unlocked;
3915 	}
3916 
3917 	error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3918 	    0);
3919 	if (error != 0)
3920 		goto loop;
3921 
3922 	/* Verify buf identify has not changed since lookup. */
3923 	if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3924 		goto foundbuf_fastpath;
3925 
3926 	/* It changed, fallback to locked lookup. */
3927 	BUF_UNLOCK_RAW(bp);
3928 
3929 loop:
3930 	BO_RLOCK(bo);
3931 	bp = gbincore(bo, blkno);
3932 	if (bp != NULL) {
3933 		int lockflags;
3934 
3935 		/*
3936 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3937 		 * it must be on a queue.
3938 		 */
3939 		lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3940 		    ((flags & GB_LOCK_NOWAIT) ? LK_NOWAIT : LK_SLEEPFAIL);
3941 
3942 		error = BUF_TIMELOCK(bp, lockflags,
3943 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3944 
3945 		/*
3946 		 * If we slept and got the lock we have to restart in case
3947 		 * the buffer changed identities.
3948 		 */
3949 		if (error == ENOLCK)
3950 			goto loop;
3951 		/* We timed out or were interrupted. */
3952 		else if (error != 0)
3953 			return (error);
3954 
3955 foundbuf_fastpath:
3956 		/* If recursed, assume caller knows the rules. */
3957 		if (BUF_LOCKRECURSED(bp))
3958 			goto end;
3959 
3960 		/*
3961 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3962 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3963 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3964 		 * backing VM cache.
3965 		 */
3966 		if (bp->b_flags & B_INVAL)
3967 			bp->b_flags &= ~B_CACHE;
3968 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3969 			bp->b_flags |= B_CACHE;
3970 		if (bp->b_flags & B_MANAGED)
3971 			MPASS(bp->b_qindex == QUEUE_NONE);
3972 		else
3973 			bremfree(bp);
3974 
3975 		/*
3976 		 * check for size inconsistencies for non-VMIO case.
3977 		 */
3978 		if (bp->b_bcount != size) {
3979 			if ((bp->b_flags & B_VMIO) == 0 ||
3980 			    (size > bp->b_kvasize)) {
3981 				if (bp->b_flags & B_DELWRI) {
3982 					bp->b_flags |= B_NOCACHE;
3983 					bwrite(bp);
3984 				} else {
3985 					if (LIST_EMPTY(&bp->b_dep)) {
3986 						bp->b_flags |= B_RELBUF;
3987 						brelse(bp);
3988 					} else {
3989 						bp->b_flags |= B_NOCACHE;
3990 						bwrite(bp);
3991 					}
3992 				}
3993 				goto loop;
3994 			}
3995 		}
3996 
3997 		/*
3998 		 * Handle the case of unmapped buffer which should
3999 		 * become mapped, or the buffer for which KVA
4000 		 * reservation is requested.
4001 		 */
4002 		bp_unmapped_get_kva(bp, blkno, size, flags);
4003 
4004 		/*
4005 		 * If the size is inconsistent in the VMIO case, we can resize
4006 		 * the buffer.  This might lead to B_CACHE getting set or
4007 		 * cleared.  If the size has not changed, B_CACHE remains
4008 		 * unchanged from its previous state.
4009 		 */
4010 		allocbuf(bp, size);
4011 
4012 		KASSERT(bp->b_offset != NOOFFSET,
4013 		    ("getblk: no buffer offset"));
4014 
4015 		/*
4016 		 * A buffer with B_DELWRI set and B_CACHE clear must
4017 		 * be committed before we can return the buffer in
4018 		 * order to prevent the caller from issuing a read
4019 		 * ( due to B_CACHE not being set ) and overwriting
4020 		 * it.
4021 		 *
4022 		 * Most callers, including NFS and FFS, need this to
4023 		 * operate properly either because they assume they
4024 		 * can issue a read if B_CACHE is not set, or because
4025 		 * ( for example ) an uncached B_DELWRI might loop due
4026 		 * to softupdates re-dirtying the buffer.  In the latter
4027 		 * case, B_CACHE is set after the first write completes,
4028 		 * preventing further loops.
4029 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
4030 		 * above while extending the buffer, we cannot allow the
4031 		 * buffer to remain with B_CACHE set after the write
4032 		 * completes or it will represent a corrupt state.  To
4033 		 * deal with this we set B_NOCACHE to scrap the buffer
4034 		 * after the write.
4035 		 *
4036 		 * We might be able to do something fancy, like setting
4037 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
4038 		 * so the below call doesn't set B_CACHE, but that gets real
4039 		 * confusing.  This is much easier.
4040 		 */
4041 
4042 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4043 			bp->b_flags |= B_NOCACHE;
4044 			bwrite(bp);
4045 			goto loop;
4046 		}
4047 		bp->b_flags &= ~B_DONE;
4048 	} else {
4049 		/*
4050 		 * Buffer is not in-core, create new buffer.  The buffer
4051 		 * returned by getnewbuf() is locked.  Note that the returned
4052 		 * buffer is also considered valid (not marked B_INVAL).
4053 		 */
4054 		BO_RUNLOCK(bo);
4055 newbuf_unlocked:
4056 		/*
4057 		 * If the user does not want us to create the buffer, bail out
4058 		 * here.
4059 		 */
4060 		if (flags & GB_NOCREAT)
4061 			return (EEXIST);
4062 
4063 		bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4064 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4065 		offset = blkno * bsize;
4066 		vmio = vp->v_object != NULL;
4067 		if (vmio) {
4068 			maxsize = size + (offset & PAGE_MASK);
4069 		} else {
4070 			maxsize = size;
4071 			/* Do not allow non-VMIO notmapped buffers. */
4072 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4073 		}
4074 		maxsize = imax(maxsize, bsize);
4075 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4076 		    !vn_isdisk(vp)) {
4077 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4078 			KASSERT(error != EOPNOTSUPP,
4079 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4080 			    vp));
4081 			if (error != 0)
4082 				return (error);
4083 			if (d_blkno == -1)
4084 				return (EJUSTRETURN);
4085 		}
4086 
4087 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4088 		if (bp == NULL) {
4089 			if (slpflag || slptimeo)
4090 				return (ETIMEDOUT);
4091 			/*
4092 			 * XXX This is here until the sleep path is diagnosed
4093 			 * enough to work under very low memory conditions.
4094 			 *
4095 			 * There's an issue on low memory, 4BSD+non-preempt
4096 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4097 			 * exhaustion occurs without sleeping for buffer
4098 			 * reclaimation.  This just sticks in a loop and
4099 			 * constantly attempts to allocate a buffer, which
4100 			 * hits exhaustion and tries to wakeup bufdaemon.
4101 			 * This never happens because we never yield.
4102 			 *
4103 			 * The real solution is to identify and fix these cases
4104 			 * so we aren't effectively busy-waiting in a loop
4105 			 * until the reclaimation path has cycles to run.
4106 			 */
4107 			kern_yield(PRI_USER);
4108 			goto loop;
4109 		}
4110 
4111 		/*
4112 		 * This code is used to make sure that a buffer is not
4113 		 * created while the getnewbuf routine is blocked.
4114 		 * This can be a problem whether the vnode is locked or not.
4115 		 * If the buffer is created out from under us, we have to
4116 		 * throw away the one we just created.
4117 		 *
4118 		 * Note: this must occur before we associate the buffer
4119 		 * with the vp especially considering limitations in
4120 		 * the splay tree implementation when dealing with duplicate
4121 		 * lblkno's.
4122 		 */
4123 		BO_LOCK(bo);
4124 		if (gbincore(bo, blkno)) {
4125 			BO_UNLOCK(bo);
4126 			bp->b_flags |= B_INVAL;
4127 			bufspace_release(bufdomain(bp), maxsize);
4128 			brelse(bp);
4129 			goto loop;
4130 		}
4131 
4132 		/*
4133 		 * Insert the buffer into the hash, so that it can
4134 		 * be found by incore.
4135 		 */
4136 		bp->b_lblkno = blkno;
4137 		bp->b_blkno = d_blkno;
4138 		bp->b_offset = offset;
4139 		bgetvp(vp, bp);
4140 		BO_UNLOCK(bo);
4141 
4142 		/*
4143 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4144 		 * buffer size starts out as 0, B_CACHE will be set by
4145 		 * allocbuf() for the VMIO case prior to it testing the
4146 		 * backing store for validity.
4147 		 */
4148 
4149 		if (vmio) {
4150 			bp->b_flags |= B_VMIO;
4151 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4152 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4153 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4154 		} else {
4155 			bp->b_flags &= ~B_VMIO;
4156 			KASSERT(bp->b_bufobj->bo_object == NULL,
4157 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4158 			    bp, bp->b_bufobj->bo_object));
4159 			BUF_CHECK_MAPPED(bp);
4160 		}
4161 
4162 		allocbuf(bp, size);
4163 		bufspace_release(bufdomain(bp), maxsize);
4164 		bp->b_flags &= ~B_DONE;
4165 	}
4166 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4167 end:
4168 	buf_track(bp, __func__);
4169 	KASSERT(bp->b_bufobj == bo,
4170 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4171 	*bpp = bp;
4172 	return (0);
4173 }
4174 
4175 /*
4176  * Get an empty, disassociated buffer of given size.  The buffer is initially
4177  * set to B_INVAL.
4178  */
4179 struct buf *
4180 geteblk(int size, int flags)
4181 {
4182 	struct buf *bp;
4183 	int maxsize;
4184 
4185 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4186 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4187 		if ((flags & GB_NOWAIT_BD) &&
4188 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4189 			return (NULL);
4190 	}
4191 	allocbuf(bp, size);
4192 	bufspace_release(bufdomain(bp), maxsize);
4193 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4194 	return (bp);
4195 }
4196 
4197 /*
4198  * Truncate the backing store for a non-vmio buffer.
4199  */
4200 static void
4201 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4202 {
4203 
4204 	if (bp->b_flags & B_MALLOC) {
4205 		/*
4206 		 * malloced buffers are not shrunk
4207 		 */
4208 		if (newbsize == 0) {
4209 			bufmallocadjust(bp, 0);
4210 			free(bp->b_data, M_BIOBUF);
4211 			bp->b_data = bp->b_kvabase;
4212 			bp->b_flags &= ~B_MALLOC;
4213 		}
4214 		return;
4215 	}
4216 	vm_hold_free_pages(bp, newbsize);
4217 	bufspace_adjust(bp, newbsize);
4218 }
4219 
4220 /*
4221  * Extend the backing for a non-VMIO buffer.
4222  */
4223 static void
4224 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4225 {
4226 	caddr_t origbuf;
4227 	int origbufsize;
4228 
4229 	/*
4230 	 * We only use malloced memory on the first allocation.
4231 	 * and revert to page-allocated memory when the buffer
4232 	 * grows.
4233 	 *
4234 	 * There is a potential smp race here that could lead
4235 	 * to bufmallocspace slightly passing the max.  It
4236 	 * is probably extremely rare and not worth worrying
4237 	 * over.
4238 	 */
4239 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4240 	    bufmallocspace < maxbufmallocspace) {
4241 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4242 		bp->b_flags |= B_MALLOC;
4243 		bufmallocadjust(bp, newbsize);
4244 		return;
4245 	}
4246 
4247 	/*
4248 	 * If the buffer is growing on its other-than-first
4249 	 * allocation then we revert to the page-allocation
4250 	 * scheme.
4251 	 */
4252 	origbuf = NULL;
4253 	origbufsize = 0;
4254 	if (bp->b_flags & B_MALLOC) {
4255 		origbuf = bp->b_data;
4256 		origbufsize = bp->b_bufsize;
4257 		bp->b_data = bp->b_kvabase;
4258 		bufmallocadjust(bp, 0);
4259 		bp->b_flags &= ~B_MALLOC;
4260 		newbsize = round_page(newbsize);
4261 	}
4262 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4263 	    (vm_offset_t) bp->b_data + newbsize);
4264 	if (origbuf != NULL) {
4265 		bcopy(origbuf, bp->b_data, origbufsize);
4266 		free(origbuf, M_BIOBUF);
4267 	}
4268 	bufspace_adjust(bp, newbsize);
4269 }
4270 
4271 /*
4272  * This code constitutes the buffer memory from either anonymous system
4273  * memory (in the case of non-VMIO operations) or from an associated
4274  * VM object (in the case of VMIO operations).  This code is able to
4275  * resize a buffer up or down.
4276  *
4277  * Note that this code is tricky, and has many complications to resolve
4278  * deadlock or inconsistent data situations.  Tread lightly!!!
4279  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4280  * the caller.  Calling this code willy nilly can result in the loss of data.
4281  *
4282  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4283  * B_CACHE for the non-VMIO case.
4284  */
4285 int
4286 allocbuf(struct buf *bp, int size)
4287 {
4288 	int newbsize;
4289 
4290 	if (bp->b_bcount == size)
4291 		return (1);
4292 
4293 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4294 		panic("allocbuf: buffer too small");
4295 
4296 	newbsize = roundup2(size, DEV_BSIZE);
4297 	if ((bp->b_flags & B_VMIO) == 0) {
4298 		if ((bp->b_flags & B_MALLOC) == 0)
4299 			newbsize = round_page(newbsize);
4300 		/*
4301 		 * Just get anonymous memory from the kernel.  Don't
4302 		 * mess with B_CACHE.
4303 		 */
4304 		if (newbsize < bp->b_bufsize)
4305 			vfs_nonvmio_truncate(bp, newbsize);
4306 		else if (newbsize > bp->b_bufsize)
4307 			vfs_nonvmio_extend(bp, newbsize);
4308 	} else {
4309 		int desiredpages;
4310 
4311 		desiredpages = (size == 0) ? 0 :
4312 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4313 
4314 		if (bp->b_flags & B_MALLOC)
4315 			panic("allocbuf: VMIO buffer can't be malloced");
4316 		/*
4317 		 * Set B_CACHE initially if buffer is 0 length or will become
4318 		 * 0-length.
4319 		 */
4320 		if (size == 0 || bp->b_bufsize == 0)
4321 			bp->b_flags |= B_CACHE;
4322 
4323 		if (newbsize < bp->b_bufsize)
4324 			vfs_vmio_truncate(bp, desiredpages);
4325 		/* XXX This looks as if it should be newbsize > b_bufsize */
4326 		else if (size > bp->b_bcount)
4327 			vfs_vmio_extend(bp, desiredpages, size);
4328 		bufspace_adjust(bp, newbsize);
4329 	}
4330 	bp->b_bcount = size;		/* requested buffer size. */
4331 	return (1);
4332 }
4333 
4334 extern int inflight_transient_maps;
4335 
4336 static struct bio_queue nondump_bios;
4337 
4338 void
4339 biodone(struct bio *bp)
4340 {
4341 	struct mtx *mtxp;
4342 	void (*done)(struct bio *);
4343 	vm_offset_t start, end;
4344 
4345 	biotrack(bp, __func__);
4346 
4347 	/*
4348 	 * Avoid completing I/O when dumping after a panic since that may
4349 	 * result in a deadlock in the filesystem or pager code.  Note that
4350 	 * this doesn't affect dumps that were started manually since we aim
4351 	 * to keep the system usable after it has been resumed.
4352 	 */
4353 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4354 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4355 		return;
4356 	}
4357 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4358 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4359 		bp->bio_flags |= BIO_UNMAPPED;
4360 		start = trunc_page((vm_offset_t)bp->bio_data);
4361 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4362 		bp->bio_data = unmapped_buf;
4363 		pmap_qremove(start, atop(end - start));
4364 		vmem_free(transient_arena, start, end - start);
4365 		atomic_add_int(&inflight_transient_maps, -1);
4366 	}
4367 	done = bp->bio_done;
4368 	if (done == NULL) {
4369 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4370 		mtx_lock(mtxp);
4371 		bp->bio_flags |= BIO_DONE;
4372 		wakeup(bp);
4373 		mtx_unlock(mtxp);
4374 	} else
4375 		done(bp);
4376 }
4377 
4378 /*
4379  * Wait for a BIO to finish.
4380  */
4381 int
4382 biowait(struct bio *bp, const char *wchan)
4383 {
4384 	struct mtx *mtxp;
4385 
4386 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4387 	mtx_lock(mtxp);
4388 	while ((bp->bio_flags & BIO_DONE) == 0)
4389 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4390 	mtx_unlock(mtxp);
4391 	if (bp->bio_error != 0)
4392 		return (bp->bio_error);
4393 	if (!(bp->bio_flags & BIO_ERROR))
4394 		return (0);
4395 	return (EIO);
4396 }
4397 
4398 void
4399 biofinish(struct bio *bp, struct devstat *stat, int error)
4400 {
4401 
4402 	if (error) {
4403 		bp->bio_error = error;
4404 		bp->bio_flags |= BIO_ERROR;
4405 	}
4406 	if (stat != NULL)
4407 		devstat_end_transaction_bio(stat, bp);
4408 	biodone(bp);
4409 }
4410 
4411 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4412 void
4413 biotrack_buf(struct bio *bp, const char *location)
4414 {
4415 
4416 	buf_track(bp->bio_track_bp, location);
4417 }
4418 #endif
4419 
4420 /*
4421  *	bufwait:
4422  *
4423  *	Wait for buffer I/O completion, returning error status.  The buffer
4424  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4425  *	error and cleared.
4426  */
4427 int
4428 bufwait(struct buf *bp)
4429 {
4430 	if (bp->b_iocmd == BIO_READ)
4431 		bwait(bp, PRIBIO, "biord");
4432 	else
4433 		bwait(bp, PRIBIO, "biowr");
4434 	if (bp->b_flags & B_EINTR) {
4435 		bp->b_flags &= ~B_EINTR;
4436 		return (EINTR);
4437 	}
4438 	if (bp->b_ioflags & BIO_ERROR) {
4439 		return (bp->b_error ? bp->b_error : EIO);
4440 	} else {
4441 		return (0);
4442 	}
4443 }
4444 
4445 /*
4446  *	bufdone:
4447  *
4448  *	Finish I/O on a buffer, optionally calling a completion function.
4449  *	This is usually called from an interrupt so process blocking is
4450  *	not allowed.
4451  *
4452  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4453  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4454  *	assuming B_INVAL is clear.
4455  *
4456  *	For the VMIO case, we set B_CACHE if the op was a read and no
4457  *	read error occurred, or if the op was a write.  B_CACHE is never
4458  *	set if the buffer is invalid or otherwise uncacheable.
4459  *
4460  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4461  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4462  *	in the biodone routine.
4463  */
4464 void
4465 bufdone(struct buf *bp)
4466 {
4467 	struct bufobj *dropobj;
4468 	void    (*biodone)(struct buf *);
4469 
4470 	buf_track(bp, __func__);
4471 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4472 	dropobj = NULL;
4473 
4474 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4475 
4476 	runningbufwakeup(bp);
4477 	if (bp->b_iocmd == BIO_WRITE)
4478 		dropobj = bp->b_bufobj;
4479 	/* call optional completion function if requested */
4480 	if (bp->b_iodone != NULL) {
4481 		biodone = bp->b_iodone;
4482 		bp->b_iodone = NULL;
4483 		(*biodone) (bp);
4484 		if (dropobj)
4485 			bufobj_wdrop(dropobj);
4486 		return;
4487 	}
4488 	if (bp->b_flags & B_VMIO) {
4489 		/*
4490 		 * Set B_CACHE if the op was a normal read and no error
4491 		 * occurred.  B_CACHE is set for writes in the b*write()
4492 		 * routines.
4493 		 */
4494 		if (bp->b_iocmd == BIO_READ &&
4495 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4496 		    !(bp->b_ioflags & BIO_ERROR))
4497 			bp->b_flags |= B_CACHE;
4498 		vfs_vmio_iodone(bp);
4499 	}
4500 	if (!LIST_EMPTY(&bp->b_dep))
4501 		buf_complete(bp);
4502 	if ((bp->b_flags & B_CKHASH) != 0) {
4503 		KASSERT(bp->b_iocmd == BIO_READ,
4504 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4505 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4506 		(*bp->b_ckhashcalc)(bp);
4507 	}
4508 	/*
4509 	 * For asynchronous completions, release the buffer now. The brelse
4510 	 * will do a wakeup there if necessary - so no need to do a wakeup
4511 	 * here in the async case. The sync case always needs to do a wakeup.
4512 	 */
4513 	if (bp->b_flags & B_ASYNC) {
4514 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4515 		    (bp->b_ioflags & BIO_ERROR))
4516 			brelse(bp);
4517 		else
4518 			bqrelse(bp);
4519 	} else
4520 		bdone(bp);
4521 	if (dropobj)
4522 		bufobj_wdrop(dropobj);
4523 }
4524 
4525 /*
4526  * This routine is called in lieu of iodone in the case of
4527  * incomplete I/O.  This keeps the busy status for pages
4528  * consistent.
4529  */
4530 void
4531 vfs_unbusy_pages(struct buf *bp)
4532 {
4533 	int i;
4534 	vm_object_t obj;
4535 	vm_page_t m;
4536 
4537 	runningbufwakeup(bp);
4538 	if (!(bp->b_flags & B_VMIO))
4539 		return;
4540 
4541 	obj = bp->b_bufobj->bo_object;
4542 	for (i = 0; i < bp->b_npages; i++) {
4543 		m = bp->b_pages[i];
4544 		if (m == bogus_page) {
4545 			m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4546 			if (!m)
4547 				panic("vfs_unbusy_pages: page missing\n");
4548 			bp->b_pages[i] = m;
4549 			if (buf_mapped(bp)) {
4550 				BUF_CHECK_MAPPED(bp);
4551 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4552 				    bp->b_pages, bp->b_npages);
4553 			} else
4554 				BUF_CHECK_UNMAPPED(bp);
4555 		}
4556 		vm_page_sunbusy(m);
4557 	}
4558 	vm_object_pip_wakeupn(obj, bp->b_npages);
4559 }
4560 
4561 /*
4562  * vfs_page_set_valid:
4563  *
4564  *	Set the valid bits in a page based on the supplied offset.   The
4565  *	range is restricted to the buffer's size.
4566  *
4567  *	This routine is typically called after a read completes.
4568  */
4569 static void
4570 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4571 {
4572 	vm_ooffset_t eoff;
4573 
4574 	/*
4575 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4576 	 * page boundary and eoff is not greater than the end of the buffer.
4577 	 * The end of the buffer, in this case, is our file EOF, not the
4578 	 * allocation size of the buffer.
4579 	 */
4580 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4581 	if (eoff > bp->b_offset + bp->b_bcount)
4582 		eoff = bp->b_offset + bp->b_bcount;
4583 
4584 	/*
4585 	 * Set valid range.  This is typically the entire buffer and thus the
4586 	 * entire page.
4587 	 */
4588 	if (eoff > off)
4589 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4590 }
4591 
4592 /*
4593  * vfs_page_set_validclean:
4594  *
4595  *	Set the valid bits and clear the dirty bits in a page based on the
4596  *	supplied offset.   The range is restricted to the buffer's size.
4597  */
4598 static void
4599 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4600 {
4601 	vm_ooffset_t soff, eoff;
4602 
4603 	/*
4604 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4605 	 * page boundary or cross the end of the buffer.  The end of the
4606 	 * buffer, in this case, is our file EOF, not the allocation size
4607 	 * of the buffer.
4608 	 */
4609 	soff = off;
4610 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4611 	if (eoff > bp->b_offset + bp->b_bcount)
4612 		eoff = bp->b_offset + bp->b_bcount;
4613 
4614 	/*
4615 	 * Set valid range.  This is typically the entire buffer and thus the
4616 	 * entire page.
4617 	 */
4618 	if (eoff > soff) {
4619 		vm_page_set_validclean(
4620 		    m,
4621 		   (vm_offset_t) (soff & PAGE_MASK),
4622 		   (vm_offset_t) (eoff - soff)
4623 		);
4624 	}
4625 }
4626 
4627 /*
4628  * Acquire a shared busy on all pages in the buf.
4629  */
4630 void
4631 vfs_busy_pages_acquire(struct buf *bp)
4632 {
4633 	int i;
4634 
4635 	for (i = 0; i < bp->b_npages; i++)
4636 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4637 }
4638 
4639 void
4640 vfs_busy_pages_release(struct buf *bp)
4641 {
4642 	int i;
4643 
4644 	for (i = 0; i < bp->b_npages; i++)
4645 		vm_page_sunbusy(bp->b_pages[i]);
4646 }
4647 
4648 /*
4649  * This routine is called before a device strategy routine.
4650  * It is used to tell the VM system that paging I/O is in
4651  * progress, and treat the pages associated with the buffer
4652  * almost as being exclusive busy.  Also the object paging_in_progress
4653  * flag is handled to make sure that the object doesn't become
4654  * inconsistent.
4655  *
4656  * Since I/O has not been initiated yet, certain buffer flags
4657  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4658  * and should be ignored.
4659  */
4660 void
4661 vfs_busy_pages(struct buf *bp, int clear_modify)
4662 {
4663 	vm_object_t obj;
4664 	vm_ooffset_t foff;
4665 	vm_page_t m;
4666 	int i;
4667 	bool bogus;
4668 
4669 	if (!(bp->b_flags & B_VMIO))
4670 		return;
4671 
4672 	obj = bp->b_bufobj->bo_object;
4673 	foff = bp->b_offset;
4674 	KASSERT(bp->b_offset != NOOFFSET,
4675 	    ("vfs_busy_pages: no buffer offset"));
4676 	if ((bp->b_flags & B_CLUSTER) == 0) {
4677 		vm_object_pip_add(obj, bp->b_npages);
4678 		vfs_busy_pages_acquire(bp);
4679 	}
4680 	if (bp->b_bufsize != 0)
4681 		vfs_setdirty_range(bp);
4682 	bogus = false;
4683 	for (i = 0; i < bp->b_npages; i++) {
4684 		m = bp->b_pages[i];
4685 		vm_page_assert_sbusied(m);
4686 
4687 		/*
4688 		 * When readying a buffer for a read ( i.e
4689 		 * clear_modify == 0 ), it is important to do
4690 		 * bogus_page replacement for valid pages in
4691 		 * partially instantiated buffers.  Partially
4692 		 * instantiated buffers can, in turn, occur when
4693 		 * reconstituting a buffer from its VM backing store
4694 		 * base.  We only have to do this if B_CACHE is
4695 		 * clear ( which causes the I/O to occur in the
4696 		 * first place ).  The replacement prevents the read
4697 		 * I/O from overwriting potentially dirty VM-backed
4698 		 * pages.  XXX bogus page replacement is, uh, bogus.
4699 		 * It may not work properly with small-block devices.
4700 		 * We need to find a better way.
4701 		 */
4702 		if (clear_modify) {
4703 			pmap_remove_write(m);
4704 			vfs_page_set_validclean(bp, foff, m);
4705 		} else if (vm_page_all_valid(m) &&
4706 		    (bp->b_flags & B_CACHE) == 0) {
4707 			bp->b_pages[i] = bogus_page;
4708 			bogus = true;
4709 		}
4710 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4711 	}
4712 	if (bogus && buf_mapped(bp)) {
4713 		BUF_CHECK_MAPPED(bp);
4714 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4715 		    bp->b_pages, bp->b_npages);
4716 	}
4717 }
4718 
4719 /*
4720  *	vfs_bio_set_valid:
4721  *
4722  *	Set the range within the buffer to valid.  The range is
4723  *	relative to the beginning of the buffer, b_offset.  Note that
4724  *	b_offset itself may be offset from the beginning of the first
4725  *	page.
4726  */
4727 void
4728 vfs_bio_set_valid(struct buf *bp, int base, int size)
4729 {
4730 	int i, n;
4731 	vm_page_t m;
4732 
4733 	if (!(bp->b_flags & B_VMIO))
4734 		return;
4735 
4736 	/*
4737 	 * Fixup base to be relative to beginning of first page.
4738 	 * Set initial n to be the maximum number of bytes in the
4739 	 * first page that can be validated.
4740 	 */
4741 	base += (bp->b_offset & PAGE_MASK);
4742 	n = PAGE_SIZE - (base & PAGE_MASK);
4743 
4744 	/*
4745 	 * Busy may not be strictly necessary here because the pages are
4746 	 * unlikely to be fully valid and the vnode lock will synchronize
4747 	 * their access via getpages.  It is grabbed for consistency with
4748 	 * other page validation.
4749 	 */
4750 	vfs_busy_pages_acquire(bp);
4751 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4752 		m = bp->b_pages[i];
4753 		if (n > size)
4754 			n = size;
4755 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4756 		base += n;
4757 		size -= n;
4758 		n = PAGE_SIZE;
4759 	}
4760 	vfs_busy_pages_release(bp);
4761 }
4762 
4763 /*
4764  *	vfs_bio_clrbuf:
4765  *
4766  *	If the specified buffer is a non-VMIO buffer, clear the entire
4767  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4768  *	validate only the previously invalid portions of the buffer.
4769  *	This routine essentially fakes an I/O, so we need to clear
4770  *	BIO_ERROR and B_INVAL.
4771  *
4772  *	Note that while we only theoretically need to clear through b_bcount,
4773  *	we go ahead and clear through b_bufsize.
4774  */
4775 void
4776 vfs_bio_clrbuf(struct buf *bp)
4777 {
4778 	int i, j, mask, sa, ea, slide;
4779 
4780 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4781 		clrbuf(bp);
4782 		return;
4783 	}
4784 	bp->b_flags &= ~B_INVAL;
4785 	bp->b_ioflags &= ~BIO_ERROR;
4786 	vfs_busy_pages_acquire(bp);
4787 	sa = bp->b_offset & PAGE_MASK;
4788 	slide = 0;
4789 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4790 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4791 		ea = slide & PAGE_MASK;
4792 		if (ea == 0)
4793 			ea = PAGE_SIZE;
4794 		if (bp->b_pages[i] == bogus_page)
4795 			continue;
4796 		j = sa / DEV_BSIZE;
4797 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4798 		if ((bp->b_pages[i]->valid & mask) == mask)
4799 			continue;
4800 		if ((bp->b_pages[i]->valid & mask) == 0)
4801 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4802 		else {
4803 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4804 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4805 					pmap_zero_page_area(bp->b_pages[i],
4806 					    sa, DEV_BSIZE);
4807 				}
4808 			}
4809 		}
4810 		vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4811 		    roundup2(ea - sa, DEV_BSIZE));
4812 	}
4813 	vfs_busy_pages_release(bp);
4814 	bp->b_resid = 0;
4815 }
4816 
4817 void
4818 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4819 {
4820 	vm_page_t m;
4821 	int i, n;
4822 
4823 	if (buf_mapped(bp)) {
4824 		BUF_CHECK_MAPPED(bp);
4825 		bzero(bp->b_data + base, size);
4826 	} else {
4827 		BUF_CHECK_UNMAPPED(bp);
4828 		n = PAGE_SIZE - (base & PAGE_MASK);
4829 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4830 			m = bp->b_pages[i];
4831 			if (n > size)
4832 				n = size;
4833 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4834 			base += n;
4835 			size -= n;
4836 			n = PAGE_SIZE;
4837 		}
4838 	}
4839 }
4840 
4841 /*
4842  * Update buffer flags based on I/O request parameters, optionally releasing the
4843  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4844  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4845  * I/O).  Otherwise the buffer is released to the cache.
4846  */
4847 static void
4848 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4849 {
4850 
4851 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4852 	    ("buf %p non-VMIO noreuse", bp));
4853 
4854 	if ((ioflag & IO_DIRECT) != 0)
4855 		bp->b_flags |= B_DIRECT;
4856 	if ((ioflag & IO_EXT) != 0)
4857 		bp->b_xflags |= BX_ALTDATA;
4858 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4859 		bp->b_flags |= B_RELBUF;
4860 		if ((ioflag & IO_NOREUSE) != 0)
4861 			bp->b_flags |= B_NOREUSE;
4862 		if (release)
4863 			brelse(bp);
4864 	} else if (release)
4865 		bqrelse(bp);
4866 }
4867 
4868 void
4869 vfs_bio_brelse(struct buf *bp, int ioflag)
4870 {
4871 
4872 	b_io_dismiss(bp, ioflag, true);
4873 }
4874 
4875 void
4876 vfs_bio_set_flags(struct buf *bp, int ioflag)
4877 {
4878 
4879 	b_io_dismiss(bp, ioflag, false);
4880 }
4881 
4882 /*
4883  * vm_hold_load_pages and vm_hold_free_pages get pages into
4884  * a buffers address space.  The pages are anonymous and are
4885  * not associated with a file object.
4886  */
4887 static void
4888 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4889 {
4890 	vm_offset_t pg;
4891 	vm_page_t p;
4892 	int index;
4893 
4894 	BUF_CHECK_MAPPED(bp);
4895 
4896 	to = round_page(to);
4897 	from = round_page(from);
4898 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4899 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
4900 	KASSERT(to - from <= maxbcachebuf,
4901 	    ("vm_hold_load_pages too large %p %#jx %#jx %u",
4902 	    bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4903 
4904 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4905 		/*
4906 		 * note: must allocate system pages since blocking here
4907 		 * could interfere with paging I/O, no matter which
4908 		 * process we are.
4909 		 */
4910 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4911 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4912 		    VM_ALLOC_WAITOK);
4913 		pmap_qenter(pg, &p, 1);
4914 		bp->b_pages[index] = p;
4915 	}
4916 	bp->b_npages = index;
4917 }
4918 
4919 /* Return pages associated with this buf to the vm system */
4920 static void
4921 vm_hold_free_pages(struct buf *bp, int newbsize)
4922 {
4923 	vm_offset_t from;
4924 	vm_page_t p;
4925 	int index, newnpages;
4926 
4927 	BUF_CHECK_MAPPED(bp);
4928 
4929 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4930 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4931 	if (bp->b_npages > newnpages)
4932 		pmap_qremove(from, bp->b_npages - newnpages);
4933 	for (index = newnpages; index < bp->b_npages; index++) {
4934 		p = bp->b_pages[index];
4935 		bp->b_pages[index] = NULL;
4936 		vm_page_unwire_noq(p);
4937 		vm_page_free(p);
4938 	}
4939 	bp->b_npages = newnpages;
4940 }
4941 
4942 /*
4943  * Map an IO request into kernel virtual address space.
4944  *
4945  * All requests are (re)mapped into kernel VA space.
4946  * Notice that we use b_bufsize for the size of the buffer
4947  * to be mapped.  b_bcount might be modified by the driver.
4948  *
4949  * Note that even if the caller determines that the address space should
4950  * be valid, a race or a smaller-file mapped into a larger space may
4951  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4952  * check the return value.
4953  *
4954  * This function only works with pager buffers.
4955  */
4956 int
4957 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
4958 {
4959 	vm_prot_t prot;
4960 	int pidx;
4961 
4962 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
4963 	prot = VM_PROT_READ;
4964 	if (bp->b_iocmd == BIO_READ)
4965 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4966 	pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4967 	    (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
4968 	if (pidx < 0)
4969 		return (-1);
4970 	bp->b_bufsize = len;
4971 	bp->b_npages = pidx;
4972 	bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
4973 	if (mapbuf || !unmapped_buf_allowed) {
4974 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4975 		bp->b_data = bp->b_kvabase + bp->b_offset;
4976 	} else
4977 		bp->b_data = unmapped_buf;
4978 	return (0);
4979 }
4980 
4981 /*
4982  * Free the io map PTEs associated with this IO operation.
4983  * We also invalidate the TLB entries and restore the original b_addr.
4984  *
4985  * This function only works with pager buffers.
4986  */
4987 void
4988 vunmapbuf(struct buf *bp)
4989 {
4990 	int npages;
4991 
4992 	npages = bp->b_npages;
4993 	if (buf_mapped(bp))
4994 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4995 	vm_page_unhold_pages(bp->b_pages, npages);
4996 
4997 	bp->b_data = unmapped_buf;
4998 }
4999 
5000 void
5001 bdone(struct buf *bp)
5002 {
5003 	struct mtx *mtxp;
5004 
5005 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5006 	mtx_lock(mtxp);
5007 	bp->b_flags |= B_DONE;
5008 	wakeup(bp);
5009 	mtx_unlock(mtxp);
5010 }
5011 
5012 void
5013 bwait(struct buf *bp, u_char pri, const char *wchan)
5014 {
5015 	struct mtx *mtxp;
5016 
5017 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5018 	mtx_lock(mtxp);
5019 	while ((bp->b_flags & B_DONE) == 0)
5020 		msleep(bp, mtxp, pri, wchan, 0);
5021 	mtx_unlock(mtxp);
5022 }
5023 
5024 int
5025 bufsync(struct bufobj *bo, int waitfor)
5026 {
5027 
5028 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5029 }
5030 
5031 void
5032 bufstrategy(struct bufobj *bo, struct buf *bp)
5033 {
5034 	int i __unused;
5035 	struct vnode *vp;
5036 
5037 	vp = bp->b_vp;
5038 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5039 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5040 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5041 	i = VOP_STRATEGY(vp, bp);
5042 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5043 }
5044 
5045 /*
5046  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
5047  */
5048 void
5049 bufobj_init(struct bufobj *bo, void *private)
5050 {
5051 	static volatile int bufobj_cleanq;
5052 
5053         bo->bo_domain =
5054             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5055         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5056         bo->bo_private = private;
5057         TAILQ_INIT(&bo->bo_clean.bv_hd);
5058         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5059 }
5060 
5061 void
5062 bufobj_wrefl(struct bufobj *bo)
5063 {
5064 
5065 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5066 	ASSERT_BO_WLOCKED(bo);
5067 	bo->bo_numoutput++;
5068 }
5069 
5070 void
5071 bufobj_wref(struct bufobj *bo)
5072 {
5073 
5074 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5075 	BO_LOCK(bo);
5076 	bo->bo_numoutput++;
5077 	BO_UNLOCK(bo);
5078 }
5079 
5080 void
5081 bufobj_wdrop(struct bufobj *bo)
5082 {
5083 
5084 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5085 	BO_LOCK(bo);
5086 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5087 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5088 		bo->bo_flag &= ~BO_WWAIT;
5089 		wakeup(&bo->bo_numoutput);
5090 	}
5091 	BO_UNLOCK(bo);
5092 }
5093 
5094 int
5095 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5096 {
5097 	int error;
5098 
5099 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5100 	ASSERT_BO_WLOCKED(bo);
5101 	error = 0;
5102 	while (bo->bo_numoutput) {
5103 		bo->bo_flag |= BO_WWAIT;
5104 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5105 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5106 		if (error)
5107 			break;
5108 	}
5109 	return (error);
5110 }
5111 
5112 /*
5113  * Set bio_data or bio_ma for struct bio from the struct buf.
5114  */
5115 void
5116 bdata2bio(struct buf *bp, struct bio *bip)
5117 {
5118 
5119 	if (!buf_mapped(bp)) {
5120 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5121 		bip->bio_ma = bp->b_pages;
5122 		bip->bio_ma_n = bp->b_npages;
5123 		bip->bio_data = unmapped_buf;
5124 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5125 		bip->bio_flags |= BIO_UNMAPPED;
5126 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5127 		    PAGE_SIZE == bp->b_npages,
5128 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5129 		    (long long)bip->bio_length, bip->bio_ma_n));
5130 	} else {
5131 		bip->bio_data = bp->b_data;
5132 		bip->bio_ma = NULL;
5133 	}
5134 }
5135 
5136 /*
5137  * The MIPS pmap code currently doesn't handle aliased pages.
5138  * The VIPT caches may not handle page aliasing themselves, leading
5139  * to data corruption.
5140  *
5141  * As such, this code makes a system extremely unhappy if said
5142  * system doesn't support unaliasing the above situation in hardware.
5143  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5144  * this feature at build time, so it has to be handled in software.
5145  *
5146  * Once the MIPS pmap/cache code grows to support this function on
5147  * earlier chips, it should be flipped back off.
5148  */
5149 #ifdef	__mips__
5150 static int buf_pager_relbuf = 1;
5151 #else
5152 static int buf_pager_relbuf = 0;
5153 #endif
5154 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5155     &buf_pager_relbuf, 0,
5156     "Make buffer pager release buffers after reading");
5157 
5158 /*
5159  * The buffer pager.  It uses buffer reads to validate pages.
5160  *
5161  * In contrast to the generic local pager from vm/vnode_pager.c, this
5162  * pager correctly and easily handles volumes where the underlying
5163  * device block size is greater than the machine page size.  The
5164  * buffer cache transparently extends the requested page run to be
5165  * aligned at the block boundary, and does the necessary bogus page
5166  * replacements in the addends to avoid obliterating already valid
5167  * pages.
5168  *
5169  * The only non-trivial issue is that the exclusive busy state for
5170  * pages, which is assumed by the vm_pager_getpages() interface, is
5171  * incompatible with the VMIO buffer cache's desire to share-busy the
5172  * pages.  This function performs a trivial downgrade of the pages'
5173  * state before reading buffers, and a less trivial upgrade from the
5174  * shared-busy to excl-busy state after the read.
5175  */
5176 int
5177 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5178     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5179     vbg_get_blksize_t get_blksize)
5180 {
5181 	vm_page_t m;
5182 	vm_object_t object;
5183 	struct buf *bp;
5184 	struct mount *mp;
5185 	daddr_t lbn, lbnp;
5186 	vm_ooffset_t la, lb, poff, poffe;
5187 	long bsize;
5188 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5189 	bool redo, lpart;
5190 
5191 	object = vp->v_object;
5192 	mp = vp->v_mount;
5193 	error = 0;
5194 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5195 	if (la >= object->un_pager.vnp.vnp_size)
5196 		return (VM_PAGER_BAD);
5197 
5198 	/*
5199 	 * Change the meaning of la from where the last requested page starts
5200 	 * to where it ends, because that's the end of the requested region
5201 	 * and the start of the potential read-ahead region.
5202 	 */
5203 	la += PAGE_SIZE;
5204 	lpart = la > object->un_pager.vnp.vnp_size;
5205 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5206 
5207 	/*
5208 	 * Calculate read-ahead, behind and total pages.
5209 	 */
5210 	pgsin = count;
5211 	lb = IDX_TO_OFF(ma[0]->pindex);
5212 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5213 	pgsin += pgsin_b;
5214 	if (rbehind != NULL)
5215 		*rbehind = pgsin_b;
5216 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5217 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5218 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5219 		    PAGE_SIZE) - la);
5220 	pgsin += pgsin_a;
5221 	if (rahead != NULL)
5222 		*rahead = pgsin_a;
5223 	VM_CNT_INC(v_vnodein);
5224 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5225 
5226 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5227 	    != 0) ? GB_UNMAPPED : 0;
5228 again:
5229 	for (i = 0; i < count; i++) {
5230 		if (ma[i] != bogus_page)
5231 			vm_page_busy_downgrade(ma[i]);
5232 	}
5233 
5234 	lbnp = -1;
5235 	for (i = 0; i < count; i++) {
5236 		m = ma[i];
5237 		if (m == bogus_page)
5238 			continue;
5239 
5240 		/*
5241 		 * Pages are shared busy and the object lock is not
5242 		 * owned, which together allow for the pages'
5243 		 * invalidation.  The racy test for validity avoids
5244 		 * useless creation of the buffer for the most typical
5245 		 * case when invalidation is not used in redo or for
5246 		 * parallel read.  The shared->excl upgrade loop at
5247 		 * the end of the function catches the race in a
5248 		 * reliable way (protected by the object lock).
5249 		 */
5250 		if (vm_page_all_valid(m))
5251 			continue;
5252 
5253 		poff = IDX_TO_OFF(m->pindex);
5254 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5255 		for (; poff < poffe; poff += bsize) {
5256 			lbn = get_lblkno(vp, poff);
5257 			if (lbn == lbnp)
5258 				goto next_page;
5259 			lbnp = lbn;
5260 
5261 			bsize = get_blksize(vp, lbn);
5262 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5263 			    br_flags, &bp);
5264 			if (error != 0)
5265 				goto end_pages;
5266 			if (bp->b_rcred == curthread->td_ucred) {
5267 				crfree(bp->b_rcred);
5268 				bp->b_rcred = NOCRED;
5269 			}
5270 			if (LIST_EMPTY(&bp->b_dep)) {
5271 				/*
5272 				 * Invalidation clears m->valid, but
5273 				 * may leave B_CACHE flag if the
5274 				 * buffer existed at the invalidation
5275 				 * time.  In this case, recycle the
5276 				 * buffer to do real read on next
5277 				 * bread() after redo.
5278 				 *
5279 				 * Otherwise B_RELBUF is not strictly
5280 				 * necessary, enable to reduce buf
5281 				 * cache pressure.
5282 				 */
5283 				if (buf_pager_relbuf ||
5284 				    !vm_page_all_valid(m))
5285 					bp->b_flags |= B_RELBUF;
5286 
5287 				bp->b_flags &= ~B_NOCACHE;
5288 				brelse(bp);
5289 			} else {
5290 				bqrelse(bp);
5291 			}
5292 		}
5293 		KASSERT(1 /* racy, enable for debugging */ ||
5294 		    vm_page_all_valid(m) || i == count - 1,
5295 		    ("buf %d %p invalid", i, m));
5296 		if (i == count - 1 && lpart) {
5297 			if (!vm_page_none_valid(m) &&
5298 			    !vm_page_all_valid(m))
5299 				vm_page_zero_invalid(m, TRUE);
5300 		}
5301 next_page:;
5302 	}
5303 end_pages:
5304 
5305 	redo = false;
5306 	for (i = 0; i < count; i++) {
5307 		if (ma[i] == bogus_page)
5308 			continue;
5309 		if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5310 			vm_page_sunbusy(ma[i]);
5311 			ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5312 			    VM_ALLOC_NORMAL);
5313 		}
5314 
5315 		/*
5316 		 * Since the pages were only sbusy while neither the
5317 		 * buffer nor the object lock was held by us, or
5318 		 * reallocated while vm_page_grab() slept for busy
5319 		 * relinguish, they could have been invalidated.
5320 		 * Recheck the valid bits and re-read as needed.
5321 		 *
5322 		 * Note that the last page is made fully valid in the
5323 		 * read loop, and partial validity for the page at
5324 		 * index count - 1 could mean that the page was
5325 		 * invalidated or removed, so we must restart for
5326 		 * safety as well.
5327 		 */
5328 		if (!vm_page_all_valid(ma[i]))
5329 			redo = true;
5330 	}
5331 	if (redo && error == 0)
5332 		goto again;
5333 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5334 }
5335 
5336 #include "opt_ddb.h"
5337 #ifdef DDB
5338 #include <ddb/ddb.h>
5339 
5340 /* DDB command to show buffer data */
5341 DB_SHOW_COMMAND(buffer, db_show_buffer)
5342 {
5343 	/* get args */
5344 	struct buf *bp = (struct buf *)addr;
5345 #ifdef FULL_BUF_TRACKING
5346 	uint32_t i, j;
5347 #endif
5348 
5349 	if (!have_addr) {
5350 		db_printf("usage: show buffer <addr>\n");
5351 		return;
5352 	}
5353 
5354 	db_printf("buf at %p\n", bp);
5355 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5356 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5357 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5358 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5359 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5360 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5361 	db_printf(
5362 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5363 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5364 	    "b_vp = %p, b_dep = %p\n",
5365 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5366 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5367 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5368 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5369 	    bp->b_kvabase, bp->b_kvasize);
5370 	if (bp->b_npages) {
5371 		int i;
5372 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5373 		for (i = 0; i < bp->b_npages; i++) {
5374 			vm_page_t m;
5375 			m = bp->b_pages[i];
5376 			if (m != NULL)
5377 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5378 				    (u_long)m->pindex,
5379 				    (u_long)VM_PAGE_TO_PHYS(m));
5380 			else
5381 				db_printf("( ??? )");
5382 			if ((i + 1) < bp->b_npages)
5383 				db_printf(",");
5384 		}
5385 		db_printf("\n");
5386 	}
5387 	BUF_LOCKPRINTINFO(bp);
5388 #if defined(FULL_BUF_TRACKING)
5389 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5390 
5391 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5392 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5393 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5394 			continue;
5395 		db_printf(" %2u: %s\n", j,
5396 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5397 	}
5398 #elif defined(BUF_TRACKING)
5399 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5400 #endif
5401 	db_printf(" ");
5402 }
5403 
5404 DB_SHOW_COMMAND(bufqueues, bufqueues)
5405 {
5406 	struct bufdomain *bd;
5407 	struct buf *bp;
5408 	long total;
5409 	int i, j, cnt;
5410 
5411 	db_printf("bqempty: %d\n", bqempty.bq_len);
5412 
5413 	for (i = 0; i < buf_domains; i++) {
5414 		bd = &bdomain[i];
5415 		db_printf("Buf domain %d\n", i);
5416 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5417 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5418 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5419 		db_printf("\n");
5420 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5421 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5422 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5423 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5424 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5425 		db_printf("\n");
5426 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5427 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5428 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5429 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5430 		db_printf("\n");
5431 		total = 0;
5432 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5433 			total += bp->b_bufsize;
5434 		db_printf("\tcleanq count\t%d (%ld)\n",
5435 		    bd->bd_cleanq->bq_len, total);
5436 		total = 0;
5437 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5438 			total += bp->b_bufsize;
5439 		db_printf("\tdirtyq count\t%d (%ld)\n",
5440 		    bd->bd_dirtyq.bq_len, total);
5441 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5442 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5443 		db_printf("\tCPU ");
5444 		for (j = 0; j <= mp_maxid; j++)
5445 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5446 		db_printf("\n");
5447 		cnt = 0;
5448 		total = 0;
5449 		for (j = 0; j < nbuf; j++) {
5450 			bp = nbufp(j);
5451 			if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5452 				cnt++;
5453 				total += bp->b_bufsize;
5454 			}
5455 		}
5456 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5457 		cnt = 0;
5458 		total = 0;
5459 		for (j = 0; j < nbuf; j++) {
5460 			bp = nbufp(j);
5461 			if (bp->b_domain == i) {
5462 				cnt++;
5463 				total += bp->b_bufsize;
5464 			}
5465 		}
5466 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5467 	}
5468 }
5469 
5470 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5471 {
5472 	struct buf *bp;
5473 	int i;
5474 
5475 	for (i = 0; i < nbuf; i++) {
5476 		bp = nbufp(i);
5477 		if (BUF_ISLOCKED(bp)) {
5478 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5479 			db_printf("\n");
5480 			if (db_pager_quit)
5481 				break;
5482 		}
5483 	}
5484 }
5485 
5486 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5487 {
5488 	struct vnode *vp;
5489 	struct buf *bp;
5490 
5491 	if (!have_addr) {
5492 		db_printf("usage: show vnodebufs <addr>\n");
5493 		return;
5494 	}
5495 	vp = (struct vnode *)addr;
5496 	db_printf("Clean buffers:\n");
5497 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5498 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5499 		db_printf("\n");
5500 	}
5501 	db_printf("Dirty buffers:\n");
5502 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5503 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5504 		db_printf("\n");
5505 	}
5506 }
5507 
5508 DB_COMMAND(countfreebufs, db_coundfreebufs)
5509 {
5510 	struct buf *bp;
5511 	int i, used = 0, nfree = 0;
5512 
5513 	if (have_addr) {
5514 		db_printf("usage: countfreebufs\n");
5515 		return;
5516 	}
5517 
5518 	for (i = 0; i < nbuf; i++) {
5519 		bp = nbufp(i);
5520 		if (bp->b_qindex == QUEUE_EMPTY)
5521 			nfree++;
5522 		else
5523 			used++;
5524 	}
5525 
5526 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5527 	    nfree + used);
5528 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5529 }
5530 #endif /* DDB */
5531