xref: /freebsd/sys/kern/vfs_bio.c (revision d34048812292b714a0bf99967270d18fe3097c62)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2004 Poul-Henning Kamp
5  * Copyright (c) 1994,1997 John S. Dyson
6  * Copyright (c) 2013 The FreeBSD Foundation
7  * All rights reserved.
8  *
9  * Portions of this software were developed by Konstantin Belousov
10  * under sponsorship from the FreeBSD Foundation.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * this file contains a new buffer I/O scheme implementing a coherent
36  * VM object and buffer cache scheme.  Pains have been taken to make
37  * sure that the performance degradation associated with schemes such
38  * as this is not realized.
39  *
40  * Author:  John S. Dyson
41  * Significant help during the development and debugging phases
42  * had been provided by David Greenman, also of the FreeBSD core team.
43  *
44  * see man buf(9) for more info.
45  */
46 
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/bio.h>
53 #include <sys/bitset.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/buf.h>
57 #include <sys/devicestat.h>
58 #include <sys/eventhandler.h>
59 #include <sys/fail.h>
60 #include <sys/limits.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/mount.h>
64 #include <sys/mutex.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/proc.h>
68 #include <sys/racct.h>
69 #include <sys/resourcevar.h>
70 #include <sys/rwlock.h>
71 #include <sys/smp.h>
72 #include <sys/sysctl.h>
73 #include <sys/sysproto.h>
74 #include <sys/vmem.h>
75 #include <sys/vmmeter.h>
76 #include <sys/vnode.h>
77 #include <sys/watchdog.h>
78 #include <geom/geom.h>
79 #include <vm/vm.h>
80 #include <vm/vm_param.h>
81 #include <vm/vm_kern.h>
82 #include <vm/vm_object.h>
83 #include <vm/vm_page.h>
84 #include <vm/vm_pageout.h>
85 #include <vm/vm_pager.h>
86 #include <vm/vm_extern.h>
87 #include <vm/vm_map.h>
88 #include <vm/swap_pager.h>
89 #include "opt_swap.h"
90 
91 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
92 
93 struct	bio_ops bioops;		/* I/O operation notification */
94 
95 struct	buf_ops buf_ops_bio = {
96 	.bop_name	=	"buf_ops_bio",
97 	.bop_write	=	bufwrite,
98 	.bop_strategy	=	bufstrategy,
99 	.bop_sync	=	bufsync,
100 	.bop_bdflush	=	bufbdflush,
101 };
102 
103 struct bufqueue {
104 	struct mtx_padalign	bq_lock;
105 	TAILQ_HEAD(, buf)	bq_queue;
106 	uint8_t			bq_index;
107 	uint16_t		bq_subqueue;
108 	int			bq_len;
109 } __aligned(CACHE_LINE_SIZE);
110 
111 #define	BQ_LOCKPTR(bq)		(&(bq)->bq_lock)
112 #define	BQ_LOCK(bq)		mtx_lock(BQ_LOCKPTR((bq)))
113 #define	BQ_UNLOCK(bq)		mtx_unlock(BQ_LOCKPTR((bq)))
114 #define	BQ_ASSERT_LOCKED(bq)	mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
115 
116 struct bufdomain {
117 	struct bufqueue	bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
118 	struct bufqueue bd_dirtyq;
119 	struct bufqueue	*bd_cleanq;
120 	struct mtx_padalign bd_run_lock;
121 	/* Constants */
122 	long		bd_maxbufspace;
123 	long		bd_hibufspace;
124 	long 		bd_lobufspace;
125 	long 		bd_bufspacethresh;
126 	int		bd_hifreebuffers;
127 	int		bd_lofreebuffers;
128 	int		bd_hidirtybuffers;
129 	int		bd_lodirtybuffers;
130 	int		bd_dirtybufthresh;
131 	int		bd_lim;
132 	/* atomics */
133 	int		bd_wanted;
134 	int __aligned(CACHE_LINE_SIZE)	bd_numdirtybuffers;
135 	int __aligned(CACHE_LINE_SIZE)	bd_running;
136 	long __aligned(CACHE_LINE_SIZE) bd_bufspace;
137 	int __aligned(CACHE_LINE_SIZE)	bd_freebuffers;
138 } __aligned(CACHE_LINE_SIZE);
139 
140 #define	BD_LOCKPTR(bd)		(&(bd)->bd_cleanq->bq_lock)
141 #define	BD_LOCK(bd)		mtx_lock(BD_LOCKPTR((bd)))
142 #define	BD_UNLOCK(bd)		mtx_unlock(BD_LOCKPTR((bd)))
143 #define	BD_ASSERT_LOCKED(bd)	mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
144 #define	BD_RUN_LOCKPTR(bd)	(&(bd)->bd_run_lock)
145 #define	BD_RUN_LOCK(bd)		mtx_lock(BD_RUN_LOCKPTR((bd)))
146 #define	BD_RUN_UNLOCK(bd)	mtx_unlock(BD_RUN_LOCKPTR((bd)))
147 #define	BD_DOMAIN(bd)		(bd - bdomain)
148 
149 static struct buf *buf;		/* buffer header pool */
150 extern struct buf *swbuf;	/* Swap buffer header pool. */
151 caddr_t unmapped_buf;
152 
153 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
154 struct proc *bufdaemonproc;
155 
156 static int inmem(struct vnode *vp, daddr_t blkno);
157 static void vm_hold_free_pages(struct buf *bp, int newbsize);
158 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
159 		vm_offset_t to);
160 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
161 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
162 		vm_page_t m);
163 static void vfs_clean_pages_dirty_buf(struct buf *bp);
164 static void vfs_setdirty_locked_object(struct buf *bp);
165 static void vfs_vmio_invalidate(struct buf *bp);
166 static void vfs_vmio_truncate(struct buf *bp, int npages);
167 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
168 static int vfs_bio_clcheck(struct vnode *vp, int size,
169 		daddr_t lblkno, daddr_t blkno);
170 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
171 		void (*)(struct buf *));
172 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
173 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
174 static void buf_daemon(void);
175 static __inline void bd_wakeup(void);
176 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
177 static void bufkva_reclaim(vmem_t *, int);
178 static void bufkva_free(struct buf *);
179 static int buf_import(void *, void **, int, int, int);
180 static void buf_release(void *, void **, int);
181 static void maxbcachebuf_adjust(void);
182 static inline struct bufdomain *bufdomain(struct buf *);
183 static void bq_remove(struct bufqueue *bq, struct buf *bp);
184 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
185 static int buf_recycle(struct bufdomain *, bool kva);
186 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
187 	    const char *lockname);
188 static void bd_init(struct bufdomain *bd);
189 static int bd_flushall(struct bufdomain *bd);
190 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
191 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
192 
193 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
194 int vmiodirenable = TRUE;
195 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
196     "Use the VM system for directory writes");
197 long runningbufspace;
198 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
199     "Amount of presently outstanding async buffer io");
200 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
201     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
202 static counter_u64_t bufkvaspace;
203 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
204     "Kernel virtual memory used for buffers");
205 static long maxbufspace;
206 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
207     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
208     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
209     "Maximum allowed value of bufspace (including metadata)");
210 static long bufmallocspace;
211 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
212     "Amount of malloced memory for buffers");
213 static long maxbufmallocspace;
214 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
215     0, "Maximum amount of malloced memory for buffers");
216 static long lobufspace;
217 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
218     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
219     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
220     "Minimum amount of buffers we want to have");
221 long hibufspace;
222 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
223     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
224     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
225     "Maximum allowed value of bufspace (excluding metadata)");
226 long bufspacethresh;
227 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
228     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
229     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
230     "Bufspace consumed before waking the daemon to free some");
231 static counter_u64_t buffreekvacnt;
232 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
233     "Number of times we have freed the KVA space from some buffer");
234 static counter_u64_t bufdefragcnt;
235 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
236     "Number of times we have had to repeat buffer allocation to defragment");
237 static long lorunningspace;
238 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
239     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
240     "Minimum preferred space used for in-progress I/O");
241 static long hirunningspace;
242 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
243     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
244     "Maximum amount of space to use for in-progress I/O");
245 int dirtybufferflushes;
246 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
247     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
248 int bdwriteskip;
249 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
250     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
251 int altbufferflushes;
252 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
253     0, "Number of fsync flushes to limit dirty buffers");
254 static int recursiveflushes;
255 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
256     0, "Number of flushes skipped due to being recursive");
257 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
258 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
259     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
260     "Number of buffers that are dirty (has unwritten changes) at the moment");
261 static int lodirtybuffers;
262 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
263     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
264     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
265     "How many buffers we want to have free before bufdaemon can sleep");
266 static int hidirtybuffers;
267 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
268     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
269     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
270     "When the number of dirty buffers is considered severe");
271 int dirtybufthresh;
272 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
273     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
274     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
275     "Number of bdwrite to bawrite conversions to clear dirty buffers");
276 static int numfreebuffers;
277 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
278     "Number of free buffers");
279 static int lofreebuffers;
280 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
281     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
282     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
283    "Target number of free buffers");
284 static int hifreebuffers;
285 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
286     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
287     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
288    "Threshold for clean buffer recycling");
289 static counter_u64_t getnewbufcalls;
290 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
291    &getnewbufcalls, "Number of calls to getnewbuf");
292 static counter_u64_t getnewbufrestarts;
293 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
294     &getnewbufrestarts,
295     "Number of times getnewbuf has had to restart a buffer acquisition");
296 static counter_u64_t mappingrestarts;
297 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
298     &mappingrestarts,
299     "Number of times getblk has had to restart a buffer mapping for "
300     "unmapped buffer");
301 static counter_u64_t numbufallocfails;
302 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
303     &numbufallocfails, "Number of times buffer allocations failed");
304 static int flushbufqtarget = 100;
305 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
306     "Amount of work to do in flushbufqueues when helping bufdaemon");
307 static counter_u64_t notbufdflushes;
308 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
309     "Number of dirty buffer flushes done by the bufdaemon helpers");
310 static long barrierwrites;
311 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
312     "Number of barrier writes");
313 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
314     &unmapped_buf_allowed, 0,
315     "Permit the use of the unmapped i/o");
316 int maxbcachebuf = MAXBCACHEBUF;
317 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
318     "Maximum size of a buffer cache block");
319 
320 /*
321  * This lock synchronizes access to bd_request.
322  */
323 static struct mtx_padalign __exclusive_cache_line bdlock;
324 
325 /*
326  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
327  * waitrunningbufspace().
328  */
329 static struct mtx_padalign __exclusive_cache_line rbreqlock;
330 
331 /*
332  * Lock that protects bdirtywait.
333  */
334 static struct mtx_padalign __exclusive_cache_line bdirtylock;
335 
336 /*
337  * Wakeup point for bufdaemon, as well as indicator of whether it is already
338  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
339  * is idling.
340  */
341 static int bd_request;
342 
343 /*
344  * Request for the buf daemon to write more buffers than is indicated by
345  * lodirtybuf.  This may be necessary to push out excess dependencies or
346  * defragment the address space where a simple count of the number of dirty
347  * buffers is insufficient to characterize the demand for flushing them.
348  */
349 static int bd_speedupreq;
350 
351 /*
352  * Synchronization (sleep/wakeup) variable for active buffer space requests.
353  * Set when wait starts, cleared prior to wakeup().
354  * Used in runningbufwakeup() and waitrunningbufspace().
355  */
356 static int runningbufreq;
357 
358 /*
359  * Synchronization for bwillwrite() waiters.
360  */
361 static int bdirtywait;
362 
363 /*
364  * Definitions for the buffer free lists.
365  */
366 #define QUEUE_NONE	0	/* on no queue */
367 #define QUEUE_EMPTY	1	/* empty buffer headers */
368 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
369 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
370 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
371 
372 /* Maximum number of buffer domains. */
373 #define	BUF_DOMAINS	8
374 
375 struct bufdomainset bdlodirty;		/* Domains > lodirty */
376 struct bufdomainset bdhidirty;		/* Domains > hidirty */
377 
378 /* Configured number of clean queues. */
379 static int __read_mostly buf_domains;
380 
381 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
382 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
383 struct bufqueue __exclusive_cache_line bqempty;
384 
385 /*
386  * per-cpu empty buffer cache.
387  */
388 uma_zone_t buf_zone;
389 
390 /*
391  * Single global constant for BUF_WMESG, to avoid getting multiple references.
392  * buf_wmesg is referred from macros.
393  */
394 const char *buf_wmesg = BUF_WMESG;
395 
396 static int
397 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
398 {
399 	long value;
400 	int error;
401 
402 	value = *(long *)arg1;
403 	error = sysctl_handle_long(oidp, &value, 0, req);
404 	if (error != 0 || req->newptr == NULL)
405 		return (error);
406 	mtx_lock(&rbreqlock);
407 	if (arg1 == &hirunningspace) {
408 		if (value < lorunningspace)
409 			error = EINVAL;
410 		else
411 			hirunningspace = value;
412 	} else {
413 		KASSERT(arg1 == &lorunningspace,
414 		    ("%s: unknown arg1", __func__));
415 		if (value > hirunningspace)
416 			error = EINVAL;
417 		else
418 			lorunningspace = value;
419 	}
420 	mtx_unlock(&rbreqlock);
421 	return (error);
422 }
423 
424 static int
425 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
426 {
427 	int error;
428 	int value;
429 	int i;
430 
431 	value = *(int *)arg1;
432 	error = sysctl_handle_int(oidp, &value, 0, req);
433 	if (error != 0 || req->newptr == NULL)
434 		return (error);
435 	*(int *)arg1 = value;
436 	for (i = 0; i < buf_domains; i++)
437 		*(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
438 		    value / buf_domains;
439 
440 	return (error);
441 }
442 
443 static int
444 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
445 {
446 	long value;
447 	int error;
448 	int i;
449 
450 	value = *(long *)arg1;
451 	error = sysctl_handle_long(oidp, &value, 0, req);
452 	if (error != 0 || req->newptr == NULL)
453 		return (error);
454 	*(long *)arg1 = value;
455 	for (i = 0; i < buf_domains; i++)
456 		*(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
457 		    value / buf_domains;
458 
459 	return (error);
460 }
461 
462 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
463     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
464 static int
465 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
466 {
467 	long lvalue;
468 	int ivalue;
469 	int i;
470 
471 	lvalue = 0;
472 	for (i = 0; i < buf_domains; i++)
473 		lvalue += bdomain[i].bd_bufspace;
474 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
475 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
476 	if (lvalue > INT_MAX)
477 		/* On overflow, still write out a long to trigger ENOMEM. */
478 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
479 	ivalue = lvalue;
480 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
481 }
482 #else
483 static int
484 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
485 {
486 	long lvalue;
487 	int i;
488 
489 	lvalue = 0;
490 	for (i = 0; i < buf_domains; i++)
491 		lvalue += bdomain[i].bd_bufspace;
492 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
493 }
494 #endif
495 
496 static int
497 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
498 {
499 	int value;
500 	int i;
501 
502 	value = 0;
503 	for (i = 0; i < buf_domains; i++)
504 		value += bdomain[i].bd_numdirtybuffers;
505 	return (sysctl_handle_int(oidp, &value, 0, req));
506 }
507 
508 /*
509  *	bdirtywakeup:
510  *
511  *	Wakeup any bwillwrite() waiters.
512  */
513 static void
514 bdirtywakeup(void)
515 {
516 	mtx_lock(&bdirtylock);
517 	if (bdirtywait) {
518 		bdirtywait = 0;
519 		wakeup(&bdirtywait);
520 	}
521 	mtx_unlock(&bdirtylock);
522 }
523 
524 /*
525  *	bd_clear:
526  *
527  *	Clear a domain from the appropriate bitsets when dirtybuffers
528  *	is decremented.
529  */
530 static void
531 bd_clear(struct bufdomain *bd)
532 {
533 
534 	mtx_lock(&bdirtylock);
535 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
536 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
537 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
538 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
539 	mtx_unlock(&bdirtylock);
540 }
541 
542 /*
543  *	bd_set:
544  *
545  *	Set a domain in the appropriate bitsets when dirtybuffers
546  *	is incremented.
547  */
548 static void
549 bd_set(struct bufdomain *bd)
550 {
551 
552 	mtx_lock(&bdirtylock);
553 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
554 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
555 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
556 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
557 	mtx_unlock(&bdirtylock);
558 }
559 
560 /*
561  *	bdirtysub:
562  *
563  *	Decrement the numdirtybuffers count by one and wakeup any
564  *	threads blocked in bwillwrite().
565  */
566 static void
567 bdirtysub(struct buf *bp)
568 {
569 	struct bufdomain *bd;
570 	int num;
571 
572 	bd = bufdomain(bp);
573 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
574 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
575 		bdirtywakeup();
576 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
577 		bd_clear(bd);
578 }
579 
580 /*
581  *	bdirtyadd:
582  *
583  *	Increment the numdirtybuffers count by one and wakeup the buf
584  *	daemon if needed.
585  */
586 static void
587 bdirtyadd(struct buf *bp)
588 {
589 	struct bufdomain *bd;
590 	int num;
591 
592 	/*
593 	 * Only do the wakeup once as we cross the boundary.  The
594 	 * buf daemon will keep running until the condition clears.
595 	 */
596 	bd = bufdomain(bp);
597 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
598 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
599 		bd_wakeup();
600 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
601 		bd_set(bd);
602 }
603 
604 /*
605  *	bufspace_daemon_wakeup:
606  *
607  *	Wakeup the daemons responsible for freeing clean bufs.
608  */
609 static void
610 bufspace_daemon_wakeup(struct bufdomain *bd)
611 {
612 
613 	/*
614 	 * avoid the lock if the daemon is running.
615 	 */
616 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
617 		BD_RUN_LOCK(bd);
618 		atomic_store_int(&bd->bd_running, 1);
619 		wakeup(&bd->bd_running);
620 		BD_RUN_UNLOCK(bd);
621 	}
622 }
623 
624 /*
625  *	bufspace_daemon_wait:
626  *
627  *	Sleep until the domain falls below a limit or one second passes.
628  */
629 static void
630 bufspace_daemon_wait(struct bufdomain *bd)
631 {
632 	/*
633 	 * Re-check our limits and sleep.  bd_running must be
634 	 * cleared prior to checking the limits to avoid missed
635 	 * wakeups.  The waker will adjust one of bufspace or
636 	 * freebuffers prior to checking bd_running.
637 	 */
638 	BD_RUN_LOCK(bd);
639 	atomic_store_int(&bd->bd_running, 0);
640 	if (bd->bd_bufspace < bd->bd_bufspacethresh &&
641 	    bd->bd_freebuffers > bd->bd_lofreebuffers) {
642 		msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
643 		    "-", hz);
644 	} else {
645 		/* Avoid spurious wakeups while running. */
646 		atomic_store_int(&bd->bd_running, 1);
647 		BD_RUN_UNLOCK(bd);
648 	}
649 }
650 
651 /*
652  *	bufspace_adjust:
653  *
654  *	Adjust the reported bufspace for a KVA managed buffer, possibly
655  * 	waking any waiters.
656  */
657 static void
658 bufspace_adjust(struct buf *bp, int bufsize)
659 {
660 	struct bufdomain *bd;
661 	long space;
662 	int diff;
663 
664 	KASSERT((bp->b_flags & B_MALLOC) == 0,
665 	    ("bufspace_adjust: malloc buf %p", bp));
666 	bd = bufdomain(bp);
667 	diff = bufsize - bp->b_bufsize;
668 	if (diff < 0) {
669 		atomic_subtract_long(&bd->bd_bufspace, -diff);
670 	} else if (diff > 0) {
671 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
672 		/* Wake up the daemon on the transition. */
673 		if (space < bd->bd_bufspacethresh &&
674 		    space + diff >= bd->bd_bufspacethresh)
675 			bufspace_daemon_wakeup(bd);
676 	}
677 	bp->b_bufsize = bufsize;
678 }
679 
680 /*
681  *	bufspace_reserve:
682  *
683  *	Reserve bufspace before calling allocbuf().  metadata has a
684  *	different space limit than data.
685  */
686 static int
687 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
688 {
689 	long limit, new;
690 	long space;
691 
692 	if (metadata)
693 		limit = bd->bd_maxbufspace;
694 	else
695 		limit = bd->bd_hibufspace;
696 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
697 	new = space + size;
698 	if (new > limit) {
699 		atomic_subtract_long(&bd->bd_bufspace, size);
700 		return (ENOSPC);
701 	}
702 
703 	/* Wake up the daemon on the transition. */
704 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
705 		bufspace_daemon_wakeup(bd);
706 
707 	return (0);
708 }
709 
710 /*
711  *	bufspace_release:
712  *
713  *	Release reserved bufspace after bufspace_adjust() has consumed it.
714  */
715 static void
716 bufspace_release(struct bufdomain *bd, int size)
717 {
718 
719 	atomic_subtract_long(&bd->bd_bufspace, size);
720 }
721 
722 /*
723  *	bufspace_wait:
724  *
725  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
726  *	supplied.  bd_wanted must be set prior to polling for space.  The
727  *	operation must be re-tried on return.
728  */
729 static void
730 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
731     int slpflag, int slptimeo)
732 {
733 	struct thread *td;
734 	int error, fl, norunbuf;
735 
736 	if ((gbflags & GB_NOWAIT_BD) != 0)
737 		return;
738 
739 	td = curthread;
740 	BD_LOCK(bd);
741 	while (bd->bd_wanted) {
742 		if (vp != NULL && vp->v_type != VCHR &&
743 		    (td->td_pflags & TDP_BUFNEED) == 0) {
744 			BD_UNLOCK(bd);
745 			/*
746 			 * getblk() is called with a vnode locked, and
747 			 * some majority of the dirty buffers may as
748 			 * well belong to the vnode.  Flushing the
749 			 * buffers there would make a progress that
750 			 * cannot be achieved by the buf_daemon, that
751 			 * cannot lock the vnode.
752 			 */
753 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
754 			    (td->td_pflags & TDP_NORUNNINGBUF);
755 
756 			/*
757 			 * Play bufdaemon.  The getnewbuf() function
758 			 * may be called while the thread owns lock
759 			 * for another dirty buffer for the same
760 			 * vnode, which makes it impossible to use
761 			 * VOP_FSYNC() there, due to the buffer lock
762 			 * recursion.
763 			 */
764 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
765 			fl = buf_flush(vp, bd, flushbufqtarget);
766 			td->td_pflags &= norunbuf;
767 			BD_LOCK(bd);
768 			if (fl != 0)
769 				continue;
770 			if (bd->bd_wanted == 0)
771 				break;
772 		}
773 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
774 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
775 		if (error != 0)
776 			break;
777 	}
778 	BD_UNLOCK(bd);
779 }
780 
781 
782 /*
783  *	bufspace_daemon:
784  *
785  *	buffer space management daemon.  Tries to maintain some marginal
786  *	amount of free buffer space so that requesting processes neither
787  *	block nor work to reclaim buffers.
788  */
789 static void
790 bufspace_daemon(void *arg)
791 {
792 	struct bufdomain *bd;
793 
794 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
795 	    SHUTDOWN_PRI_LAST + 100);
796 
797 	bd = arg;
798 	for (;;) {
799 		kthread_suspend_check();
800 
801 		/*
802 		 * Free buffers from the clean queue until we meet our
803 		 * targets.
804 		 *
805 		 * Theory of operation:  The buffer cache is most efficient
806 		 * when some free buffer headers and space are always
807 		 * available to getnewbuf().  This daemon attempts to prevent
808 		 * the excessive blocking and synchronization associated
809 		 * with shortfall.  It goes through three phases according
810 		 * demand:
811 		 *
812 		 * 1)	The daemon wakes up voluntarily once per-second
813 		 *	during idle periods when the counters are below
814 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
815 		 *
816 		 * 2)	The daemon wakes up as we cross the thresholds
817 		 *	ahead of any potential blocking.  This may bounce
818 		 *	slightly according to the rate of consumption and
819 		 *	release.
820 		 *
821 		 * 3)	The daemon and consumers are starved for working
822 		 *	clean buffers.  This is the 'bufspace' sleep below
823 		 *	which will inefficiently trade bufs with bqrelse
824 		 *	until we return to condition 2.
825 		 */
826 		while (bd->bd_bufspace > bd->bd_lobufspace ||
827 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
828 			if (buf_recycle(bd, false) != 0) {
829 				if (bd_flushall(bd))
830 					continue;
831 				/*
832 				 * Speedup dirty if we've run out of clean
833 				 * buffers.  This is possible in particular
834 				 * because softdep may held many bufs locked
835 				 * pending writes to other bufs which are
836 				 * marked for delayed write, exhausting
837 				 * clean space until they are written.
838 				 */
839 				bd_speedup();
840 				BD_LOCK(bd);
841 				if (bd->bd_wanted) {
842 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
843 					    PRIBIO|PDROP, "bufspace", hz/10);
844 				} else
845 					BD_UNLOCK(bd);
846 			}
847 			maybe_yield();
848 		}
849 		bufspace_daemon_wait(bd);
850 	}
851 }
852 
853 /*
854  *	bufmallocadjust:
855  *
856  *	Adjust the reported bufspace for a malloc managed buffer, possibly
857  *	waking any waiters.
858  */
859 static void
860 bufmallocadjust(struct buf *bp, int bufsize)
861 {
862 	int diff;
863 
864 	KASSERT((bp->b_flags & B_MALLOC) != 0,
865 	    ("bufmallocadjust: non-malloc buf %p", bp));
866 	diff = bufsize - bp->b_bufsize;
867 	if (diff < 0)
868 		atomic_subtract_long(&bufmallocspace, -diff);
869 	else
870 		atomic_add_long(&bufmallocspace, diff);
871 	bp->b_bufsize = bufsize;
872 }
873 
874 /*
875  *	runningwakeup:
876  *
877  *	Wake up processes that are waiting on asynchronous writes to fall
878  *	below lorunningspace.
879  */
880 static void
881 runningwakeup(void)
882 {
883 
884 	mtx_lock(&rbreqlock);
885 	if (runningbufreq) {
886 		runningbufreq = 0;
887 		wakeup(&runningbufreq);
888 	}
889 	mtx_unlock(&rbreqlock);
890 }
891 
892 /*
893  *	runningbufwakeup:
894  *
895  *	Decrement the outstanding write count according.
896  */
897 void
898 runningbufwakeup(struct buf *bp)
899 {
900 	long space, bspace;
901 
902 	bspace = bp->b_runningbufspace;
903 	if (bspace == 0)
904 		return;
905 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
906 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
907 	    space, bspace));
908 	bp->b_runningbufspace = 0;
909 	/*
910 	 * Only acquire the lock and wakeup on the transition from exceeding
911 	 * the threshold to falling below it.
912 	 */
913 	if (space < lorunningspace)
914 		return;
915 	if (space - bspace > lorunningspace)
916 		return;
917 	runningwakeup();
918 }
919 
920 /*
921  *	waitrunningbufspace()
922  *
923  *	runningbufspace is a measure of the amount of I/O currently
924  *	running.  This routine is used in async-write situations to
925  *	prevent creating huge backups of pending writes to a device.
926  *	Only asynchronous writes are governed by this function.
927  *
928  *	This does NOT turn an async write into a sync write.  It waits
929  *	for earlier writes to complete and generally returns before the
930  *	caller's write has reached the device.
931  */
932 void
933 waitrunningbufspace(void)
934 {
935 
936 	mtx_lock(&rbreqlock);
937 	while (runningbufspace > hirunningspace) {
938 		runningbufreq = 1;
939 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
940 	}
941 	mtx_unlock(&rbreqlock);
942 }
943 
944 
945 /*
946  *	vfs_buf_test_cache:
947  *
948  *	Called when a buffer is extended.  This function clears the B_CACHE
949  *	bit if the newly extended portion of the buffer does not contain
950  *	valid data.
951  */
952 static __inline void
953 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
954     vm_offset_t size, vm_page_t m)
955 {
956 
957 	VM_OBJECT_ASSERT_LOCKED(m->object);
958 	if (bp->b_flags & B_CACHE) {
959 		int base = (foff + off) & PAGE_MASK;
960 		if (vm_page_is_valid(m, base, size) == 0)
961 			bp->b_flags &= ~B_CACHE;
962 	}
963 }
964 
965 /* Wake up the buffer daemon if necessary */
966 static void
967 bd_wakeup(void)
968 {
969 
970 	mtx_lock(&bdlock);
971 	if (bd_request == 0) {
972 		bd_request = 1;
973 		wakeup(&bd_request);
974 	}
975 	mtx_unlock(&bdlock);
976 }
977 
978 /*
979  * Adjust the maxbcachbuf tunable.
980  */
981 static void
982 maxbcachebuf_adjust(void)
983 {
984 	int i;
985 
986 	/*
987 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
988 	 */
989 	i = 2;
990 	while (i * 2 <= maxbcachebuf)
991 		i *= 2;
992 	maxbcachebuf = i;
993 	if (maxbcachebuf < MAXBSIZE)
994 		maxbcachebuf = MAXBSIZE;
995 	if (maxbcachebuf > MAXPHYS)
996 		maxbcachebuf = MAXPHYS;
997 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
998 		printf("maxbcachebuf=%d\n", maxbcachebuf);
999 }
1000 
1001 /*
1002  * bd_speedup - speedup the buffer cache flushing code
1003  */
1004 void
1005 bd_speedup(void)
1006 {
1007 	int needwake;
1008 
1009 	mtx_lock(&bdlock);
1010 	needwake = 0;
1011 	if (bd_speedupreq == 0 || bd_request == 0)
1012 		needwake = 1;
1013 	bd_speedupreq = 1;
1014 	bd_request = 1;
1015 	if (needwake)
1016 		wakeup(&bd_request);
1017 	mtx_unlock(&bdlock);
1018 }
1019 
1020 #ifndef NSWBUF_MIN
1021 #define	NSWBUF_MIN	16
1022 #endif
1023 
1024 #ifdef __i386__
1025 #define	TRANSIENT_DENOM	5
1026 #else
1027 #define	TRANSIENT_DENOM 10
1028 #endif
1029 
1030 /*
1031  * Calculating buffer cache scaling values and reserve space for buffer
1032  * headers.  This is called during low level kernel initialization and
1033  * may be called more then once.  We CANNOT write to the memory area
1034  * being reserved at this time.
1035  */
1036 caddr_t
1037 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1038 {
1039 	int tuned_nbuf;
1040 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1041 
1042 	/*
1043 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1044 	 * PAGE_SIZE is >= 1K)
1045 	 */
1046 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1047 
1048 	maxbcachebuf_adjust();
1049 	/*
1050 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1051 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1052 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1053 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1054 	 * the buffer cache we limit the eventual kva reservation to
1055 	 * maxbcache bytes.
1056 	 *
1057 	 * factor represents the 1/4 x ram conversion.
1058 	 */
1059 	if (nbuf == 0) {
1060 		int factor = 4 * BKVASIZE / 1024;
1061 
1062 		nbuf = 50;
1063 		if (physmem_est > 4096)
1064 			nbuf += min((physmem_est - 4096) / factor,
1065 			    65536 / factor);
1066 		if (physmem_est > 65536)
1067 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1068 			    32 * 1024 * 1024 / (factor * 5));
1069 
1070 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1071 			nbuf = maxbcache / BKVASIZE;
1072 		tuned_nbuf = 1;
1073 	} else
1074 		tuned_nbuf = 0;
1075 
1076 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1077 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1078 	if (nbuf > maxbuf) {
1079 		if (!tuned_nbuf)
1080 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1081 			    maxbuf);
1082 		nbuf = maxbuf;
1083 	}
1084 
1085 	/*
1086 	 * Ideal allocation size for the transient bio submap is 10%
1087 	 * of the maximal space buffer map.  This roughly corresponds
1088 	 * to the amount of the buffer mapped for typical UFS load.
1089 	 *
1090 	 * Clip the buffer map to reserve space for the transient
1091 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1092 	 * maximum buffer map extent on the platform.
1093 	 *
1094 	 * The fall-back to the maxbuf in case of maxbcache unset,
1095 	 * allows to not trim the buffer KVA for the architectures
1096 	 * with ample KVA space.
1097 	 */
1098 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1099 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1100 		buf_sz = (long)nbuf * BKVASIZE;
1101 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1102 		    (TRANSIENT_DENOM - 1)) {
1103 			/*
1104 			 * There is more KVA than memory.  Do not
1105 			 * adjust buffer map size, and assign the rest
1106 			 * of maxbuf to transient map.
1107 			 */
1108 			biotmap_sz = maxbuf_sz - buf_sz;
1109 		} else {
1110 			/*
1111 			 * Buffer map spans all KVA we could afford on
1112 			 * this platform.  Give 10% (20% on i386) of
1113 			 * the buffer map to the transient bio map.
1114 			 */
1115 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1116 			buf_sz -= biotmap_sz;
1117 		}
1118 		if (biotmap_sz / INT_MAX > MAXPHYS)
1119 			bio_transient_maxcnt = INT_MAX;
1120 		else
1121 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1122 		/*
1123 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1124 		 * using the transient mapping.
1125 		 */
1126 		if (bio_transient_maxcnt > 1024)
1127 			bio_transient_maxcnt = 1024;
1128 		if (tuned_nbuf)
1129 			nbuf = buf_sz / BKVASIZE;
1130 	}
1131 
1132 	/*
1133 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
1134 	 * We have no less then 16 and no more then 256.
1135 	 */
1136 	nswbuf = min(nbuf / 4, 256);
1137 	TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf);
1138 	if (nswbuf < NSWBUF_MIN)
1139 		nswbuf = NSWBUF_MIN;
1140 
1141 	/*
1142 	 * Reserve space for the buffer cache buffers
1143 	 */
1144 	swbuf = (void *)v;
1145 	v = (caddr_t)(swbuf + nswbuf);
1146 	buf = (void *)v;
1147 	v = (caddr_t)(buf + 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 = &buf[i];
1172 		bzero(bp, sizeof *bp);
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", sizeof(struct buf),
1257 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1258 
1259 	/*
1260 	 * Size the clean queue according to the amount of buffer space.
1261 	 * One queue per-256mb up to the max.  More queues gives better
1262 	 * concurrency but less accurate LRU.
1263 	 */
1264 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1265 	for (i = 0 ; i < buf_domains; i++) {
1266 		struct bufdomain *bd;
1267 
1268 		bd = &bdomain[i];
1269 		bd_init(bd);
1270 		bd->bd_freebuffers = nbuf / buf_domains;
1271 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1272 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1273 		bd->bd_bufspace = 0;
1274 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1275 		bd->bd_hibufspace = hibufspace / buf_domains;
1276 		bd->bd_lobufspace = lobufspace / buf_domains;
1277 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1278 		bd->bd_numdirtybuffers = 0;
1279 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1280 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1281 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1282 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1283 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1284 	}
1285 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1286 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1287 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1288 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1289 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1290 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1291 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1292 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1293 }
1294 
1295 #ifdef INVARIANTS
1296 static inline void
1297 vfs_buf_check_mapped(struct buf *bp)
1298 {
1299 
1300 	KASSERT(bp->b_kvabase != unmapped_buf,
1301 	    ("mapped buf: b_kvabase was not updated %p", bp));
1302 	KASSERT(bp->b_data != unmapped_buf,
1303 	    ("mapped buf: b_data was not updated %p", bp));
1304 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1305 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1306 }
1307 
1308 static inline void
1309 vfs_buf_check_unmapped(struct buf *bp)
1310 {
1311 
1312 	KASSERT(bp->b_data == unmapped_buf,
1313 	    ("unmapped buf: corrupted b_data %p", bp));
1314 }
1315 
1316 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1317 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1318 #else
1319 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1320 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1321 #endif
1322 
1323 static int
1324 isbufbusy(struct buf *bp)
1325 {
1326 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1327 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1328 		return (1);
1329 	return (0);
1330 }
1331 
1332 /*
1333  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1334  */
1335 void
1336 bufshutdown(int show_busybufs)
1337 {
1338 	static int first_buf_printf = 1;
1339 	struct buf *bp;
1340 	int iter, nbusy, pbusy;
1341 #ifndef PREEMPTION
1342 	int subiter;
1343 #endif
1344 
1345 	/*
1346 	 * Sync filesystems for shutdown
1347 	 */
1348 	wdog_kern_pat(WD_LASTVAL);
1349 	sys_sync(curthread, NULL);
1350 
1351 	/*
1352 	 * With soft updates, some buffers that are
1353 	 * written will be remarked as dirty until other
1354 	 * buffers are written.
1355 	 */
1356 	for (iter = pbusy = 0; iter < 20; iter++) {
1357 		nbusy = 0;
1358 		for (bp = &buf[nbuf]; --bp >= buf; )
1359 			if (isbufbusy(bp))
1360 				nbusy++;
1361 		if (nbusy == 0) {
1362 			if (first_buf_printf)
1363 				printf("All buffers synced.");
1364 			break;
1365 		}
1366 		if (first_buf_printf) {
1367 			printf("Syncing disks, buffers remaining... ");
1368 			first_buf_printf = 0;
1369 		}
1370 		printf("%d ", nbusy);
1371 		if (nbusy < pbusy)
1372 			iter = 0;
1373 		pbusy = nbusy;
1374 
1375 		wdog_kern_pat(WD_LASTVAL);
1376 		sys_sync(curthread, NULL);
1377 
1378 #ifdef PREEMPTION
1379 		/*
1380 		 * Spin for a while to allow interrupt threads to run.
1381 		 */
1382 		DELAY(50000 * iter);
1383 #else
1384 		/*
1385 		 * Context switch several times to allow interrupt
1386 		 * threads to run.
1387 		 */
1388 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1389 			thread_lock(curthread);
1390 			mi_switch(SW_VOL, NULL);
1391 			thread_unlock(curthread);
1392 			DELAY(1000);
1393 		}
1394 #endif
1395 	}
1396 	printf("\n");
1397 	/*
1398 	 * Count only busy local buffers to prevent forcing
1399 	 * a fsck if we're just a client of a wedged NFS server
1400 	 */
1401 	nbusy = 0;
1402 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1403 		if (isbufbusy(bp)) {
1404 #if 0
1405 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1406 			if (bp->b_dev == NULL) {
1407 				TAILQ_REMOVE(&mountlist,
1408 				    bp->b_vp->v_mount, mnt_list);
1409 				continue;
1410 			}
1411 #endif
1412 			nbusy++;
1413 			if (show_busybufs > 0) {
1414 				printf(
1415 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1416 				    nbusy, bp, bp->b_vp, bp->b_flags,
1417 				    (intmax_t)bp->b_blkno,
1418 				    (intmax_t)bp->b_lblkno);
1419 				BUF_LOCKPRINTINFO(bp);
1420 				if (show_busybufs > 1)
1421 					vn_printf(bp->b_vp,
1422 					    "vnode content: ");
1423 			}
1424 		}
1425 	}
1426 	if (nbusy) {
1427 		/*
1428 		 * Failed to sync all blocks. Indicate this and don't
1429 		 * unmount filesystems (thus forcing an fsck on reboot).
1430 		 */
1431 		printf("Giving up on %d buffers\n", nbusy);
1432 		DELAY(5000000);	/* 5 seconds */
1433 	} else {
1434 		if (!first_buf_printf)
1435 			printf("Final sync complete\n");
1436 		/*
1437 		 * Unmount filesystems
1438 		 */
1439 		if (panicstr == NULL)
1440 			vfs_unmountall();
1441 	}
1442 	swapoff_all();
1443 	DELAY(100000);		/* wait for console output to finish */
1444 }
1445 
1446 static void
1447 bpmap_qenter(struct buf *bp)
1448 {
1449 
1450 	BUF_CHECK_MAPPED(bp);
1451 
1452 	/*
1453 	 * bp->b_data is relative to bp->b_offset, but
1454 	 * bp->b_offset may be offset into the first page.
1455 	 */
1456 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1457 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1458 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1459 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1460 }
1461 
1462 static inline struct bufdomain *
1463 bufdomain(struct buf *bp)
1464 {
1465 
1466 	return (&bdomain[bp->b_domain]);
1467 }
1468 
1469 static struct bufqueue *
1470 bufqueue(struct buf *bp)
1471 {
1472 
1473 	switch (bp->b_qindex) {
1474 	case QUEUE_NONE:
1475 		/* FALLTHROUGH */
1476 	case QUEUE_SENTINEL:
1477 		return (NULL);
1478 	case QUEUE_EMPTY:
1479 		return (&bqempty);
1480 	case QUEUE_DIRTY:
1481 		return (&bufdomain(bp)->bd_dirtyq);
1482 	case QUEUE_CLEAN:
1483 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1484 	default:
1485 		break;
1486 	}
1487 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1488 }
1489 
1490 /*
1491  * Return the locked bufqueue that bp is a member of.
1492  */
1493 static struct bufqueue *
1494 bufqueue_acquire(struct buf *bp)
1495 {
1496 	struct bufqueue *bq, *nbq;
1497 
1498 	/*
1499 	 * bp can be pushed from a per-cpu queue to the
1500 	 * cleanq while we're waiting on the lock.  Retry
1501 	 * if the queues don't match.
1502 	 */
1503 	bq = bufqueue(bp);
1504 	BQ_LOCK(bq);
1505 	for (;;) {
1506 		nbq = bufqueue(bp);
1507 		if (bq == nbq)
1508 			break;
1509 		BQ_UNLOCK(bq);
1510 		BQ_LOCK(nbq);
1511 		bq = nbq;
1512 	}
1513 	return (bq);
1514 }
1515 
1516 /*
1517  *	binsfree:
1518  *
1519  *	Insert the buffer into the appropriate free list.  Requires a
1520  *	locked buffer on entry and buffer is unlocked before return.
1521  */
1522 static void
1523 binsfree(struct buf *bp, int qindex)
1524 {
1525 	struct bufdomain *bd;
1526 	struct bufqueue *bq;
1527 
1528 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1529 	    ("binsfree: Invalid qindex %d", qindex));
1530 	BUF_ASSERT_XLOCKED(bp);
1531 
1532 	/*
1533 	 * Handle delayed bremfree() processing.
1534 	 */
1535 	if (bp->b_flags & B_REMFREE) {
1536 		if (bp->b_qindex == qindex) {
1537 			bp->b_flags |= B_REUSE;
1538 			bp->b_flags &= ~B_REMFREE;
1539 			BUF_UNLOCK(bp);
1540 			return;
1541 		}
1542 		bq = bufqueue_acquire(bp);
1543 		bq_remove(bq, bp);
1544 		BQ_UNLOCK(bq);
1545 	}
1546 	bd = bufdomain(bp);
1547 	if (qindex == QUEUE_CLEAN) {
1548 		if (bd->bd_lim != 0)
1549 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1550 		else
1551 			bq = bd->bd_cleanq;
1552 	} else
1553 		bq = &bd->bd_dirtyq;
1554 	bq_insert(bq, bp, true);
1555 }
1556 
1557 /*
1558  * buf_free:
1559  *
1560  *	Free a buffer to the buf zone once it no longer has valid contents.
1561  */
1562 static void
1563 buf_free(struct buf *bp)
1564 {
1565 
1566 	if (bp->b_flags & B_REMFREE)
1567 		bremfreef(bp);
1568 	if (bp->b_vflags & BV_BKGRDINPROG)
1569 		panic("losing buffer 1");
1570 	if (bp->b_rcred != NOCRED) {
1571 		crfree(bp->b_rcred);
1572 		bp->b_rcred = NOCRED;
1573 	}
1574 	if (bp->b_wcred != NOCRED) {
1575 		crfree(bp->b_wcred);
1576 		bp->b_wcred = NOCRED;
1577 	}
1578 	if (!LIST_EMPTY(&bp->b_dep))
1579 		buf_deallocate(bp);
1580 	bufkva_free(bp);
1581 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1582 	BUF_UNLOCK(bp);
1583 	uma_zfree(buf_zone, bp);
1584 }
1585 
1586 /*
1587  * buf_import:
1588  *
1589  *	Import bufs into the uma cache from the buf list.  The system still
1590  *	expects a static array of bufs and much of the synchronization
1591  *	around bufs assumes type stable storage.  As a result, UMA is used
1592  *	only as a per-cpu cache of bufs still maintained on a global list.
1593  */
1594 static int
1595 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1596 {
1597 	struct buf *bp;
1598 	int i;
1599 
1600 	BQ_LOCK(&bqempty);
1601 	for (i = 0; i < cnt; i++) {
1602 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1603 		if (bp == NULL)
1604 			break;
1605 		bq_remove(&bqempty, bp);
1606 		store[i] = bp;
1607 	}
1608 	BQ_UNLOCK(&bqempty);
1609 
1610 	return (i);
1611 }
1612 
1613 /*
1614  * buf_release:
1615  *
1616  *	Release bufs from the uma cache back to the buffer queues.
1617  */
1618 static void
1619 buf_release(void *arg, void **store, int cnt)
1620 {
1621 	struct bufqueue *bq;
1622 	struct buf *bp;
1623         int i;
1624 
1625 	bq = &bqempty;
1626 	BQ_LOCK(bq);
1627         for (i = 0; i < cnt; i++) {
1628 		bp = store[i];
1629 		/* Inline bq_insert() to batch locking. */
1630 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1631 		bp->b_flags &= ~(B_AGE | B_REUSE);
1632 		bq->bq_len++;
1633 		bp->b_qindex = bq->bq_index;
1634 	}
1635 	BQ_UNLOCK(bq);
1636 }
1637 
1638 /*
1639  * buf_alloc:
1640  *
1641  *	Allocate an empty buffer header.
1642  */
1643 static struct buf *
1644 buf_alloc(struct bufdomain *bd)
1645 {
1646 	struct buf *bp;
1647 	int freebufs;
1648 
1649 	/*
1650 	 * We can only run out of bufs in the buf zone if the average buf
1651 	 * is less than BKVASIZE.  In this case the actual wait/block will
1652 	 * come from buf_reycle() failing to flush one of these small bufs.
1653 	 */
1654 	bp = NULL;
1655 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1656 	if (freebufs > 0)
1657 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1658 	if (bp == NULL) {
1659 		atomic_add_int(&bd->bd_freebuffers, 1);
1660 		bufspace_daemon_wakeup(bd);
1661 		counter_u64_add(numbufallocfails, 1);
1662 		return (NULL);
1663 	}
1664 	/*
1665 	 * Wake-up the bufspace daemon on transition below threshold.
1666 	 */
1667 	if (freebufs == bd->bd_lofreebuffers)
1668 		bufspace_daemon_wakeup(bd);
1669 
1670 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1671 		panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1672 
1673 	KASSERT(bp->b_vp == NULL,
1674 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1675 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1676 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1677 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1678 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1679 	KASSERT(bp->b_npages == 0,
1680 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1681 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1682 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1683 
1684 	bp->b_domain = BD_DOMAIN(bd);
1685 	bp->b_flags = 0;
1686 	bp->b_ioflags = 0;
1687 	bp->b_xflags = 0;
1688 	bp->b_vflags = 0;
1689 	bp->b_vp = NULL;
1690 	bp->b_blkno = bp->b_lblkno = 0;
1691 	bp->b_offset = NOOFFSET;
1692 	bp->b_iodone = 0;
1693 	bp->b_error = 0;
1694 	bp->b_resid = 0;
1695 	bp->b_bcount = 0;
1696 	bp->b_npages = 0;
1697 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1698 	bp->b_bufobj = NULL;
1699 	bp->b_data = bp->b_kvabase = unmapped_buf;
1700 	bp->b_fsprivate1 = NULL;
1701 	bp->b_fsprivate2 = NULL;
1702 	bp->b_fsprivate3 = NULL;
1703 	LIST_INIT(&bp->b_dep);
1704 
1705 	return (bp);
1706 }
1707 
1708 /*
1709  *	buf_recycle:
1710  *
1711  *	Free a buffer from the given bufqueue.  kva controls whether the
1712  *	freed buf must own some kva resources.  This is used for
1713  *	defragmenting.
1714  */
1715 static int
1716 buf_recycle(struct bufdomain *bd, bool kva)
1717 {
1718 	struct bufqueue *bq;
1719 	struct buf *bp, *nbp;
1720 
1721 	if (kva)
1722 		counter_u64_add(bufdefragcnt, 1);
1723 	nbp = NULL;
1724 	bq = bd->bd_cleanq;
1725 	BQ_LOCK(bq);
1726 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1727 	    ("buf_recycle: Locks don't match"));
1728 	nbp = TAILQ_FIRST(&bq->bq_queue);
1729 
1730 	/*
1731 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1732 	 * depending.
1733 	 */
1734 	while ((bp = nbp) != NULL) {
1735 		/*
1736 		 * Calculate next bp (we can only use it if we do not
1737 		 * release the bqlock).
1738 		 */
1739 		nbp = TAILQ_NEXT(bp, b_freelist);
1740 
1741 		/*
1742 		 * If we are defragging then we need a buffer with
1743 		 * some kva to reclaim.
1744 		 */
1745 		if (kva && bp->b_kvasize == 0)
1746 			continue;
1747 
1748 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1749 			continue;
1750 
1751 		/*
1752 		 * Implement a second chance algorithm for frequently
1753 		 * accessed buffers.
1754 		 */
1755 		if ((bp->b_flags & B_REUSE) != 0) {
1756 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1757 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1758 			bp->b_flags &= ~B_REUSE;
1759 			BUF_UNLOCK(bp);
1760 			continue;
1761 		}
1762 
1763 		/*
1764 		 * Skip buffers with background writes in progress.
1765 		 */
1766 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1767 			BUF_UNLOCK(bp);
1768 			continue;
1769 		}
1770 
1771 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1772 		    ("buf_recycle: inconsistent queue %d bp %p",
1773 		    bp->b_qindex, bp));
1774 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1775 		    ("getnewbuf: queue domain %d doesn't match request %d",
1776 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1777 		/*
1778 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1779 		 * the scan from this point on.
1780 		 */
1781 		bq_remove(bq, bp);
1782 		BQ_UNLOCK(bq);
1783 
1784 		/*
1785 		 * Requeue the background write buffer with error and
1786 		 * restart the scan.
1787 		 */
1788 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1789 			bqrelse(bp);
1790 			BQ_LOCK(bq);
1791 			nbp = TAILQ_FIRST(&bq->bq_queue);
1792 			continue;
1793 		}
1794 		bp->b_flags |= B_INVAL;
1795 		brelse(bp);
1796 		return (0);
1797 	}
1798 	bd->bd_wanted = 1;
1799 	BQ_UNLOCK(bq);
1800 
1801 	return (ENOBUFS);
1802 }
1803 
1804 /*
1805  *	bremfree:
1806  *
1807  *	Mark the buffer for removal from the appropriate free list.
1808  *
1809  */
1810 void
1811 bremfree(struct buf *bp)
1812 {
1813 
1814 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1815 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1816 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1817 	KASSERT(bp->b_qindex != QUEUE_NONE,
1818 	    ("bremfree: buffer %p not on a queue.", bp));
1819 	BUF_ASSERT_XLOCKED(bp);
1820 
1821 	bp->b_flags |= B_REMFREE;
1822 }
1823 
1824 /*
1825  *	bremfreef:
1826  *
1827  *	Force an immediate removal from a free list.  Used only in nfs when
1828  *	it abuses the b_freelist pointer.
1829  */
1830 void
1831 bremfreef(struct buf *bp)
1832 {
1833 	struct bufqueue *bq;
1834 
1835 	bq = bufqueue_acquire(bp);
1836 	bq_remove(bq, bp);
1837 	BQ_UNLOCK(bq);
1838 }
1839 
1840 static void
1841 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1842 {
1843 
1844 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1845 	TAILQ_INIT(&bq->bq_queue);
1846 	bq->bq_len = 0;
1847 	bq->bq_index = qindex;
1848 	bq->bq_subqueue = subqueue;
1849 }
1850 
1851 static void
1852 bd_init(struct bufdomain *bd)
1853 {
1854 	int i;
1855 
1856 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1857 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1858 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1859 	for (i = 0; i <= mp_maxid; i++)
1860 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1861 		    "bufq clean subqueue lock");
1862 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1863 }
1864 
1865 /*
1866  *	bq_remove:
1867  *
1868  *	Removes a buffer from the free list, must be called with the
1869  *	correct qlock held.
1870  */
1871 static void
1872 bq_remove(struct bufqueue *bq, struct buf *bp)
1873 {
1874 
1875 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1876 	    bp, bp->b_vp, bp->b_flags);
1877 	KASSERT(bp->b_qindex != QUEUE_NONE,
1878 	    ("bq_remove: buffer %p not on a queue.", bp));
1879 	KASSERT(bufqueue(bp) == bq,
1880 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1881 
1882 	BQ_ASSERT_LOCKED(bq);
1883 	if (bp->b_qindex != QUEUE_EMPTY) {
1884 		BUF_ASSERT_XLOCKED(bp);
1885 	}
1886 	KASSERT(bq->bq_len >= 1,
1887 	    ("queue %d underflow", bp->b_qindex));
1888 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1889 	bq->bq_len--;
1890 	bp->b_qindex = QUEUE_NONE;
1891 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1892 }
1893 
1894 static void
1895 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1896 {
1897 	struct buf *bp;
1898 
1899 	BQ_ASSERT_LOCKED(bq);
1900 	if (bq != bd->bd_cleanq) {
1901 		BD_LOCK(bd);
1902 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1903 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1904 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1905 			    b_freelist);
1906 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1907 		}
1908 		bd->bd_cleanq->bq_len += bq->bq_len;
1909 		bq->bq_len = 0;
1910 	}
1911 	if (bd->bd_wanted) {
1912 		bd->bd_wanted = 0;
1913 		wakeup(&bd->bd_wanted);
1914 	}
1915 	if (bq != bd->bd_cleanq)
1916 		BD_UNLOCK(bd);
1917 }
1918 
1919 static int
1920 bd_flushall(struct bufdomain *bd)
1921 {
1922 	struct bufqueue *bq;
1923 	int flushed;
1924 	int i;
1925 
1926 	if (bd->bd_lim == 0)
1927 		return (0);
1928 	flushed = 0;
1929 	for (i = 0; i <= mp_maxid; i++) {
1930 		bq = &bd->bd_subq[i];
1931 		if (bq->bq_len == 0)
1932 			continue;
1933 		BQ_LOCK(bq);
1934 		bd_flush(bd, bq);
1935 		BQ_UNLOCK(bq);
1936 		flushed++;
1937 	}
1938 
1939 	return (flushed);
1940 }
1941 
1942 static void
1943 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1944 {
1945 	struct bufdomain *bd;
1946 
1947 	if (bp->b_qindex != QUEUE_NONE)
1948 		panic("bq_insert: free buffer %p onto another queue?", bp);
1949 
1950 	bd = bufdomain(bp);
1951 	if (bp->b_flags & B_AGE) {
1952 		/* Place this buf directly on the real queue. */
1953 		if (bq->bq_index == QUEUE_CLEAN)
1954 			bq = bd->bd_cleanq;
1955 		BQ_LOCK(bq);
1956 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1957 	} else {
1958 		BQ_LOCK(bq);
1959 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1960 	}
1961 	bp->b_flags &= ~(B_AGE | B_REUSE);
1962 	bq->bq_len++;
1963 	bp->b_qindex = bq->bq_index;
1964 	bp->b_subqueue = bq->bq_subqueue;
1965 
1966 	/*
1967 	 * Unlock before we notify so that we don't wakeup a waiter that
1968 	 * fails a trylock on the buf and sleeps again.
1969 	 */
1970 	if (unlock)
1971 		BUF_UNLOCK(bp);
1972 
1973 	if (bp->b_qindex == QUEUE_CLEAN) {
1974 		/*
1975 		 * Flush the per-cpu queue and notify any waiters.
1976 		 */
1977 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1978 		    bq->bq_len >= bd->bd_lim))
1979 			bd_flush(bd, bq);
1980 	}
1981 	BQ_UNLOCK(bq);
1982 }
1983 
1984 /*
1985  *	bufkva_free:
1986  *
1987  *	Free the kva allocation for a buffer.
1988  *
1989  */
1990 static void
1991 bufkva_free(struct buf *bp)
1992 {
1993 
1994 #ifdef INVARIANTS
1995 	if (bp->b_kvasize == 0) {
1996 		KASSERT(bp->b_kvabase == unmapped_buf &&
1997 		    bp->b_data == unmapped_buf,
1998 		    ("Leaked KVA space on %p", bp));
1999 	} else if (buf_mapped(bp))
2000 		BUF_CHECK_MAPPED(bp);
2001 	else
2002 		BUF_CHECK_UNMAPPED(bp);
2003 #endif
2004 	if (bp->b_kvasize == 0)
2005 		return;
2006 
2007 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2008 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2009 	counter_u64_add(buffreekvacnt, 1);
2010 	bp->b_data = bp->b_kvabase = unmapped_buf;
2011 	bp->b_kvasize = 0;
2012 }
2013 
2014 /*
2015  *	bufkva_alloc:
2016  *
2017  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2018  */
2019 static int
2020 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2021 {
2022 	vm_offset_t addr;
2023 	int error;
2024 
2025 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2026 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2027 
2028 	bufkva_free(bp);
2029 
2030 	addr = 0;
2031 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2032 	if (error != 0) {
2033 		/*
2034 		 * Buffer map is too fragmented.  Request the caller
2035 		 * to defragment the map.
2036 		 */
2037 		return (error);
2038 	}
2039 	bp->b_kvabase = (caddr_t)addr;
2040 	bp->b_kvasize = maxsize;
2041 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2042 	if ((gbflags & GB_UNMAPPED) != 0) {
2043 		bp->b_data = unmapped_buf;
2044 		BUF_CHECK_UNMAPPED(bp);
2045 	} else {
2046 		bp->b_data = bp->b_kvabase;
2047 		BUF_CHECK_MAPPED(bp);
2048 	}
2049 	return (0);
2050 }
2051 
2052 /*
2053  *	bufkva_reclaim:
2054  *
2055  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2056  *	callback that fires to avoid returning failure.
2057  */
2058 static void
2059 bufkva_reclaim(vmem_t *vmem, int flags)
2060 {
2061 	bool done;
2062 	int q;
2063 	int i;
2064 
2065 	done = false;
2066 	for (i = 0; i < 5; i++) {
2067 		for (q = 0; q < buf_domains; q++)
2068 			if (buf_recycle(&bdomain[q], true) != 0)
2069 				done = true;
2070 		if (done)
2071 			break;
2072 	}
2073 	return;
2074 }
2075 
2076 /*
2077  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2078  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2079  * the buffer is valid and we do not have to do anything.
2080  */
2081 static void
2082 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2083     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2084 {
2085 	struct buf *rabp;
2086 	int i;
2087 
2088 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2089 		if (inmem(vp, *rablkno))
2090 			continue;
2091 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2092 		if ((rabp->b_flags & B_CACHE) != 0) {
2093 			brelse(rabp);
2094 			continue;
2095 		}
2096 		if (!TD_IS_IDLETHREAD(curthread)) {
2097 #ifdef RACCT
2098 			if (racct_enable) {
2099 				PROC_LOCK(curproc);
2100 				racct_add_buf(curproc, rabp, 0);
2101 				PROC_UNLOCK(curproc);
2102 			}
2103 #endif /* RACCT */
2104 			curthread->td_ru.ru_inblock++;
2105 		}
2106 		rabp->b_flags |= B_ASYNC;
2107 		rabp->b_flags &= ~B_INVAL;
2108 		if ((flags & GB_CKHASH) != 0) {
2109 			rabp->b_flags |= B_CKHASH;
2110 			rabp->b_ckhashcalc = ckhashfunc;
2111 		}
2112 		rabp->b_ioflags &= ~BIO_ERROR;
2113 		rabp->b_iocmd = BIO_READ;
2114 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2115 			rabp->b_rcred = crhold(cred);
2116 		vfs_busy_pages(rabp, 0);
2117 		BUF_KERNPROC(rabp);
2118 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2119 		bstrategy(rabp);
2120 	}
2121 }
2122 
2123 /*
2124  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2125  *
2126  * Get a buffer with the specified data.  Look in the cache first.  We
2127  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2128  * is set, the buffer is valid and we do not have to do anything, see
2129  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2130  *
2131  * Always return a NULL buffer pointer (in bpp) when returning an error.
2132  */
2133 int
2134 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
2135     int *rabsize, int cnt, struct ucred *cred, int flags,
2136     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2137 {
2138 	struct buf *bp;
2139 	struct thread *td;
2140 	int error, readwait, rv;
2141 
2142 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2143 	td = curthread;
2144 	/*
2145 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2146 	 * are specified.
2147 	 */
2148 	error = getblkx(vp, blkno, size, 0, 0, flags, &bp);
2149 	if (error != 0) {
2150 		*bpp = NULL;
2151 		return (error);
2152 	}
2153 	flags &= ~GB_NOSPARSE;
2154 	*bpp = bp;
2155 
2156 	/*
2157 	 * If not found in cache, do some I/O
2158 	 */
2159 	readwait = 0;
2160 	if ((bp->b_flags & B_CACHE) == 0) {
2161 		if (!TD_IS_IDLETHREAD(td)) {
2162 #ifdef RACCT
2163 			if (racct_enable) {
2164 				PROC_LOCK(td->td_proc);
2165 				racct_add_buf(td->td_proc, bp, 0);
2166 				PROC_UNLOCK(td->td_proc);
2167 			}
2168 #endif /* RACCT */
2169 			td->td_ru.ru_inblock++;
2170 		}
2171 		bp->b_iocmd = BIO_READ;
2172 		bp->b_flags &= ~B_INVAL;
2173 		if ((flags & GB_CKHASH) != 0) {
2174 			bp->b_flags |= B_CKHASH;
2175 			bp->b_ckhashcalc = ckhashfunc;
2176 		}
2177 		bp->b_ioflags &= ~BIO_ERROR;
2178 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2179 			bp->b_rcred = crhold(cred);
2180 		vfs_busy_pages(bp, 0);
2181 		bp->b_iooffset = dbtob(bp->b_blkno);
2182 		bstrategy(bp);
2183 		++readwait;
2184 	}
2185 
2186 	/*
2187 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2188 	 */
2189 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2190 
2191 	rv = 0;
2192 	if (readwait) {
2193 		rv = bufwait(bp);
2194 		if (rv != 0) {
2195 			brelse(bp);
2196 			*bpp = NULL;
2197 		}
2198 	}
2199 	return (rv);
2200 }
2201 
2202 /*
2203  * Write, release buffer on completion.  (Done by iodone
2204  * if async).  Do not bother writing anything if the buffer
2205  * is invalid.
2206  *
2207  * Note that we set B_CACHE here, indicating that buffer is
2208  * fully valid and thus cacheable.  This is true even of NFS
2209  * now so we set it generally.  This could be set either here
2210  * or in biodone() since the I/O is synchronous.  We put it
2211  * here.
2212  */
2213 int
2214 bufwrite(struct buf *bp)
2215 {
2216 	int oldflags;
2217 	struct vnode *vp;
2218 	long space;
2219 	int vp_md;
2220 
2221 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2222 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2223 		bp->b_flags |= B_INVAL | B_RELBUF;
2224 		bp->b_flags &= ~B_CACHE;
2225 		brelse(bp);
2226 		return (ENXIO);
2227 	}
2228 	if (bp->b_flags & B_INVAL) {
2229 		brelse(bp);
2230 		return (0);
2231 	}
2232 
2233 	if (bp->b_flags & B_BARRIER)
2234 		barrierwrites++;
2235 
2236 	oldflags = bp->b_flags;
2237 
2238 	BUF_ASSERT_HELD(bp);
2239 
2240 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2241 	    ("FFS background buffer should not get here %p", bp));
2242 
2243 	vp = bp->b_vp;
2244 	if (vp)
2245 		vp_md = vp->v_vflag & VV_MD;
2246 	else
2247 		vp_md = 0;
2248 
2249 	/*
2250 	 * Mark the buffer clean.  Increment the bufobj write count
2251 	 * before bundirty() call, to prevent other thread from seeing
2252 	 * empty dirty list and zero counter for writes in progress,
2253 	 * falsely indicating that the bufobj is clean.
2254 	 */
2255 	bufobj_wref(bp->b_bufobj);
2256 	bundirty(bp);
2257 
2258 	bp->b_flags &= ~B_DONE;
2259 	bp->b_ioflags &= ~BIO_ERROR;
2260 	bp->b_flags |= B_CACHE;
2261 	bp->b_iocmd = BIO_WRITE;
2262 
2263 	vfs_busy_pages(bp, 1);
2264 
2265 	/*
2266 	 * Normal bwrites pipeline writes
2267 	 */
2268 	bp->b_runningbufspace = bp->b_bufsize;
2269 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2270 
2271 	if (!TD_IS_IDLETHREAD(curthread)) {
2272 #ifdef RACCT
2273 		if (racct_enable) {
2274 			PROC_LOCK(curproc);
2275 			racct_add_buf(curproc, bp, 1);
2276 			PROC_UNLOCK(curproc);
2277 		}
2278 #endif /* RACCT */
2279 		curthread->td_ru.ru_oublock++;
2280 	}
2281 	if (oldflags & B_ASYNC)
2282 		BUF_KERNPROC(bp);
2283 	bp->b_iooffset = dbtob(bp->b_blkno);
2284 	buf_track(bp, __func__);
2285 	bstrategy(bp);
2286 
2287 	if ((oldflags & B_ASYNC) == 0) {
2288 		int rtval = bufwait(bp);
2289 		brelse(bp);
2290 		return (rtval);
2291 	} else if (space > hirunningspace) {
2292 		/*
2293 		 * don't allow the async write to saturate the I/O
2294 		 * system.  We will not deadlock here because
2295 		 * we are blocking waiting for I/O that is already in-progress
2296 		 * to complete. We do not block here if it is the update
2297 		 * or syncer daemon trying to clean up as that can lead
2298 		 * to deadlock.
2299 		 */
2300 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2301 			waitrunningbufspace();
2302 	}
2303 
2304 	return (0);
2305 }
2306 
2307 void
2308 bufbdflush(struct bufobj *bo, struct buf *bp)
2309 {
2310 	struct buf *nbp;
2311 
2312 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2313 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2314 		altbufferflushes++;
2315 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2316 		BO_LOCK(bo);
2317 		/*
2318 		 * Try to find a buffer to flush.
2319 		 */
2320 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2321 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2322 			    BUF_LOCK(nbp,
2323 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2324 				continue;
2325 			if (bp == nbp)
2326 				panic("bdwrite: found ourselves");
2327 			BO_UNLOCK(bo);
2328 			/* Don't countdeps with the bo lock held. */
2329 			if (buf_countdeps(nbp, 0)) {
2330 				BO_LOCK(bo);
2331 				BUF_UNLOCK(nbp);
2332 				continue;
2333 			}
2334 			if (nbp->b_flags & B_CLUSTEROK) {
2335 				vfs_bio_awrite(nbp);
2336 			} else {
2337 				bremfree(nbp);
2338 				bawrite(nbp);
2339 			}
2340 			dirtybufferflushes++;
2341 			break;
2342 		}
2343 		if (nbp == NULL)
2344 			BO_UNLOCK(bo);
2345 	}
2346 }
2347 
2348 /*
2349  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2350  * anything if the buffer is marked invalid.
2351  *
2352  * Note that since the buffer must be completely valid, we can safely
2353  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2354  * biodone() in order to prevent getblk from writing the buffer
2355  * out synchronously.
2356  */
2357 void
2358 bdwrite(struct buf *bp)
2359 {
2360 	struct thread *td = curthread;
2361 	struct vnode *vp;
2362 	struct bufobj *bo;
2363 
2364 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2365 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2366 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2367 	    ("Barrier request in delayed write %p", bp));
2368 	BUF_ASSERT_HELD(bp);
2369 
2370 	if (bp->b_flags & B_INVAL) {
2371 		brelse(bp);
2372 		return;
2373 	}
2374 
2375 	/*
2376 	 * If we have too many dirty buffers, don't create any more.
2377 	 * If we are wildly over our limit, then force a complete
2378 	 * cleanup. Otherwise, just keep the situation from getting
2379 	 * out of control. Note that we have to avoid a recursive
2380 	 * disaster and not try to clean up after our own cleanup!
2381 	 */
2382 	vp = bp->b_vp;
2383 	bo = bp->b_bufobj;
2384 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2385 		td->td_pflags |= TDP_INBDFLUSH;
2386 		BO_BDFLUSH(bo, bp);
2387 		td->td_pflags &= ~TDP_INBDFLUSH;
2388 	} else
2389 		recursiveflushes++;
2390 
2391 	bdirty(bp);
2392 	/*
2393 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2394 	 * true even of NFS now.
2395 	 */
2396 	bp->b_flags |= B_CACHE;
2397 
2398 	/*
2399 	 * This bmap keeps the system from needing to do the bmap later,
2400 	 * perhaps when the system is attempting to do a sync.  Since it
2401 	 * is likely that the indirect block -- or whatever other datastructure
2402 	 * that the filesystem needs is still in memory now, it is a good
2403 	 * thing to do this.  Note also, that if the pageout daemon is
2404 	 * requesting a sync -- there might not be enough memory to do
2405 	 * the bmap then...  So, this is important to do.
2406 	 */
2407 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2408 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2409 	}
2410 
2411 	buf_track(bp, __func__);
2412 
2413 	/*
2414 	 * Set the *dirty* buffer range based upon the VM system dirty
2415 	 * pages.
2416 	 *
2417 	 * Mark the buffer pages as clean.  We need to do this here to
2418 	 * satisfy the vnode_pager and the pageout daemon, so that it
2419 	 * thinks that the pages have been "cleaned".  Note that since
2420 	 * the pages are in a delayed write buffer -- the VFS layer
2421 	 * "will" see that the pages get written out on the next sync,
2422 	 * or perhaps the cluster will be completed.
2423 	 */
2424 	vfs_clean_pages_dirty_buf(bp);
2425 	bqrelse(bp);
2426 
2427 	/*
2428 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2429 	 * due to the softdep code.
2430 	 */
2431 }
2432 
2433 /*
2434  *	bdirty:
2435  *
2436  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2437  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2438  *	itself to properly update it in the dirty/clean lists.  We mark it
2439  *	B_DONE to ensure that any asynchronization of the buffer properly
2440  *	clears B_DONE ( else a panic will occur later ).
2441  *
2442  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2443  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2444  *	should only be called if the buffer is known-good.
2445  *
2446  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2447  *	count.
2448  *
2449  *	The buffer must be on QUEUE_NONE.
2450  */
2451 void
2452 bdirty(struct buf *bp)
2453 {
2454 
2455 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2456 	    bp, bp->b_vp, bp->b_flags);
2457 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2458 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2459 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2460 	BUF_ASSERT_HELD(bp);
2461 	bp->b_flags &= ~(B_RELBUF);
2462 	bp->b_iocmd = BIO_WRITE;
2463 
2464 	if ((bp->b_flags & B_DELWRI) == 0) {
2465 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2466 		reassignbuf(bp);
2467 		bdirtyadd(bp);
2468 	}
2469 }
2470 
2471 /*
2472  *	bundirty:
2473  *
2474  *	Clear B_DELWRI for buffer.
2475  *
2476  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2477  *	count.
2478  *
2479  *	The buffer must be on QUEUE_NONE.
2480  */
2481 
2482 void
2483 bundirty(struct buf *bp)
2484 {
2485 
2486 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2487 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2488 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2489 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2490 	BUF_ASSERT_HELD(bp);
2491 
2492 	if (bp->b_flags & B_DELWRI) {
2493 		bp->b_flags &= ~B_DELWRI;
2494 		reassignbuf(bp);
2495 		bdirtysub(bp);
2496 	}
2497 	/*
2498 	 * Since it is now being written, we can clear its deferred write flag.
2499 	 */
2500 	bp->b_flags &= ~B_DEFERRED;
2501 }
2502 
2503 /*
2504  *	bawrite:
2505  *
2506  *	Asynchronous write.  Start output on a buffer, but do not wait for
2507  *	it to complete.  The buffer is released when the output completes.
2508  *
2509  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2510  *	B_INVAL buffers.  Not us.
2511  */
2512 void
2513 bawrite(struct buf *bp)
2514 {
2515 
2516 	bp->b_flags |= B_ASYNC;
2517 	(void) bwrite(bp);
2518 }
2519 
2520 /*
2521  *	babarrierwrite:
2522  *
2523  *	Asynchronous barrier write.  Start output on a buffer, but do not
2524  *	wait for it to complete.  Place a write barrier after this write so
2525  *	that this buffer and all buffers written before it are committed to
2526  *	the disk before any buffers written after this write are committed
2527  *	to the disk.  The buffer is released when the output completes.
2528  */
2529 void
2530 babarrierwrite(struct buf *bp)
2531 {
2532 
2533 	bp->b_flags |= B_ASYNC | B_BARRIER;
2534 	(void) bwrite(bp);
2535 }
2536 
2537 /*
2538  *	bbarrierwrite:
2539  *
2540  *	Synchronous barrier write.  Start output on a buffer and wait for
2541  *	it to complete.  Place a write barrier after this write so that
2542  *	this buffer and all buffers written before it are committed to
2543  *	the disk before any buffers written after this write are committed
2544  *	to the disk.  The buffer is released when the output completes.
2545  */
2546 int
2547 bbarrierwrite(struct buf *bp)
2548 {
2549 
2550 	bp->b_flags |= B_BARRIER;
2551 	return (bwrite(bp));
2552 }
2553 
2554 /*
2555  *	bwillwrite:
2556  *
2557  *	Called prior to the locking of any vnodes when we are expecting to
2558  *	write.  We do not want to starve the buffer cache with too many
2559  *	dirty buffers so we block here.  By blocking prior to the locking
2560  *	of any vnodes we attempt to avoid the situation where a locked vnode
2561  *	prevents the various system daemons from flushing related buffers.
2562  */
2563 void
2564 bwillwrite(void)
2565 {
2566 
2567 	if (buf_dirty_count_severe()) {
2568 		mtx_lock(&bdirtylock);
2569 		while (buf_dirty_count_severe()) {
2570 			bdirtywait = 1;
2571 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2572 			    "flswai", 0);
2573 		}
2574 		mtx_unlock(&bdirtylock);
2575 	}
2576 }
2577 
2578 /*
2579  * Return true if we have too many dirty buffers.
2580  */
2581 int
2582 buf_dirty_count_severe(void)
2583 {
2584 
2585 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2586 }
2587 
2588 /*
2589  *	brelse:
2590  *
2591  *	Release a busy buffer and, if requested, free its resources.  The
2592  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2593  *	to be accessed later as a cache entity or reused for other purposes.
2594  */
2595 void
2596 brelse(struct buf *bp)
2597 {
2598 	struct mount *v_mnt;
2599 	int qindex;
2600 
2601 	/*
2602 	 * Many functions erroneously call brelse with a NULL bp under rare
2603 	 * error conditions. Simply return when called with a NULL bp.
2604 	 */
2605 	if (bp == NULL)
2606 		return;
2607 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2608 	    bp, bp->b_vp, bp->b_flags);
2609 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2610 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2611 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2612 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2613 
2614 	if (BUF_LOCKRECURSED(bp)) {
2615 		/*
2616 		 * Do not process, in particular, do not handle the
2617 		 * B_INVAL/B_RELBUF and do not release to free list.
2618 		 */
2619 		BUF_UNLOCK(bp);
2620 		return;
2621 	}
2622 
2623 	if (bp->b_flags & B_MANAGED) {
2624 		bqrelse(bp);
2625 		return;
2626 	}
2627 
2628 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2629 		BO_LOCK(bp->b_bufobj);
2630 		bp->b_vflags &= ~BV_BKGRDERR;
2631 		BO_UNLOCK(bp->b_bufobj);
2632 		bdirty(bp);
2633 	}
2634 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2635 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2636 	    !(bp->b_flags & B_INVAL)) {
2637 		/*
2638 		 * Failed write, redirty.  All errors except ENXIO (which
2639 		 * means the device is gone) are treated as being
2640 		 * transient.
2641 		 *
2642 		 * XXX Treating EIO as transient is not correct; the
2643 		 * contract with the local storage device drivers is that
2644 		 * they will only return EIO once the I/O is no longer
2645 		 * retriable.  Network I/O also respects this through the
2646 		 * guarantees of TCP and/or the internal retries of NFS.
2647 		 * ENOMEM might be transient, but we also have no way of
2648 		 * knowing when its ok to retry/reschedule.  In general,
2649 		 * this entire case should be made obsolete through better
2650 		 * error handling/recovery and resource scheduling.
2651 		 *
2652 		 * Do this also for buffers that failed with ENXIO, but have
2653 		 * non-empty dependencies - the soft updates code might need
2654 		 * to access the buffer to untangle them.
2655 		 *
2656 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2657 		 */
2658 		bp->b_ioflags &= ~BIO_ERROR;
2659 		bdirty(bp);
2660 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2661 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2662 		/*
2663 		 * Either a failed read I/O, or we were asked to free or not
2664 		 * cache the buffer, or we failed to write to a device that's
2665 		 * no longer present.
2666 		 */
2667 		bp->b_flags |= B_INVAL;
2668 		if (!LIST_EMPTY(&bp->b_dep))
2669 			buf_deallocate(bp);
2670 		if (bp->b_flags & B_DELWRI)
2671 			bdirtysub(bp);
2672 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2673 		if ((bp->b_flags & B_VMIO) == 0) {
2674 			allocbuf(bp, 0);
2675 			if (bp->b_vp)
2676 				brelvp(bp);
2677 		}
2678 	}
2679 
2680 	/*
2681 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2682 	 * is called with B_DELWRI set, the underlying pages may wind up
2683 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2684 	 * because pages associated with a B_DELWRI bp are marked clean.
2685 	 *
2686 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2687 	 * if B_DELWRI is set.
2688 	 */
2689 	if (bp->b_flags & B_DELWRI)
2690 		bp->b_flags &= ~B_RELBUF;
2691 
2692 	/*
2693 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2694 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2695 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2696 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2697 	 *
2698 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2699 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2700 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2701 	 *
2702 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2703 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2704 	 * the commit state and we cannot afford to lose the buffer. If the
2705 	 * buffer has a background write in progress, we need to keep it
2706 	 * around to prevent it from being reconstituted and starting a second
2707 	 * background write.
2708 	 */
2709 
2710 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2711 
2712 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2713 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2714 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2715 	    vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2716 		vfs_vmio_invalidate(bp);
2717 		allocbuf(bp, 0);
2718 	}
2719 
2720 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2721 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2722 		allocbuf(bp, 0);
2723 		bp->b_flags &= ~B_NOREUSE;
2724 		if (bp->b_vp != NULL)
2725 			brelvp(bp);
2726 	}
2727 
2728 	/*
2729 	 * If the buffer has junk contents signal it and eventually
2730 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2731 	 * doesn't find it.
2732 	 */
2733 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2734 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2735 		bp->b_flags |= B_INVAL;
2736 	if (bp->b_flags & B_INVAL) {
2737 		if (bp->b_flags & B_DELWRI)
2738 			bundirty(bp);
2739 		if (bp->b_vp)
2740 			brelvp(bp);
2741 	}
2742 
2743 	buf_track(bp, __func__);
2744 
2745 	/* buffers with no memory */
2746 	if (bp->b_bufsize == 0) {
2747 		buf_free(bp);
2748 		return;
2749 	}
2750 	/* buffers with junk contents */
2751 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2752 	    (bp->b_ioflags & BIO_ERROR)) {
2753 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2754 		if (bp->b_vflags & BV_BKGRDINPROG)
2755 			panic("losing buffer 2");
2756 		qindex = QUEUE_CLEAN;
2757 		bp->b_flags |= B_AGE;
2758 	/* remaining buffers */
2759 	} else if (bp->b_flags & B_DELWRI)
2760 		qindex = QUEUE_DIRTY;
2761 	else
2762 		qindex = QUEUE_CLEAN;
2763 
2764 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2765 		panic("brelse: not dirty");
2766 
2767 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2768 	/* binsfree unlocks bp. */
2769 	binsfree(bp, qindex);
2770 }
2771 
2772 /*
2773  * Release a buffer back to the appropriate queue but do not try to free
2774  * it.  The buffer is expected to be used again soon.
2775  *
2776  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2777  * biodone() to requeue an async I/O on completion.  It is also used when
2778  * known good buffers need to be requeued but we think we may need the data
2779  * again soon.
2780  *
2781  * XXX we should be able to leave the B_RELBUF hint set on completion.
2782  */
2783 void
2784 bqrelse(struct buf *bp)
2785 {
2786 	int qindex;
2787 
2788 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2789 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2790 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2791 
2792 	qindex = QUEUE_NONE;
2793 	if (BUF_LOCKRECURSED(bp)) {
2794 		/* do not release to free list */
2795 		BUF_UNLOCK(bp);
2796 		return;
2797 	}
2798 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2799 
2800 	if (bp->b_flags & B_MANAGED) {
2801 		if (bp->b_flags & B_REMFREE)
2802 			bremfreef(bp);
2803 		goto out;
2804 	}
2805 
2806 	/* buffers with stale but valid contents */
2807 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2808 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2809 		BO_LOCK(bp->b_bufobj);
2810 		bp->b_vflags &= ~BV_BKGRDERR;
2811 		BO_UNLOCK(bp->b_bufobj);
2812 		qindex = QUEUE_DIRTY;
2813 	} else {
2814 		if ((bp->b_flags & B_DELWRI) == 0 &&
2815 		    (bp->b_xflags & BX_VNDIRTY))
2816 			panic("bqrelse: not dirty");
2817 		if ((bp->b_flags & B_NOREUSE) != 0) {
2818 			brelse(bp);
2819 			return;
2820 		}
2821 		qindex = QUEUE_CLEAN;
2822 	}
2823 	buf_track(bp, __func__);
2824 	/* binsfree unlocks bp. */
2825 	binsfree(bp, qindex);
2826 	return;
2827 
2828 out:
2829 	buf_track(bp, __func__);
2830 	/* unlock */
2831 	BUF_UNLOCK(bp);
2832 }
2833 
2834 /*
2835  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2836  * restore bogus pages.
2837  */
2838 static void
2839 vfs_vmio_iodone(struct buf *bp)
2840 {
2841 	vm_ooffset_t foff;
2842 	vm_page_t m;
2843 	vm_object_t obj;
2844 	struct vnode *vp __unused;
2845 	int i, iosize, resid;
2846 	bool bogus;
2847 
2848 	obj = bp->b_bufobj->bo_object;
2849 	KASSERT(obj->paging_in_progress >= bp->b_npages,
2850 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2851 	    obj->paging_in_progress, bp->b_npages));
2852 
2853 	vp = bp->b_vp;
2854 	KASSERT(vp->v_holdcnt > 0,
2855 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2856 	KASSERT(vp->v_object != NULL,
2857 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2858 
2859 	foff = bp->b_offset;
2860 	KASSERT(bp->b_offset != NOOFFSET,
2861 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2862 
2863 	bogus = false;
2864 	iosize = bp->b_bcount - bp->b_resid;
2865 	VM_OBJECT_WLOCK(obj);
2866 	for (i = 0; i < bp->b_npages; i++) {
2867 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2868 		if (resid > iosize)
2869 			resid = iosize;
2870 
2871 		/*
2872 		 * cleanup bogus pages, restoring the originals
2873 		 */
2874 		m = bp->b_pages[i];
2875 		if (m == bogus_page) {
2876 			bogus = true;
2877 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2878 			if (m == NULL)
2879 				panic("biodone: page disappeared!");
2880 			bp->b_pages[i] = m;
2881 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2882 			/*
2883 			 * In the write case, the valid and clean bits are
2884 			 * already changed correctly ( see bdwrite() ), so we
2885 			 * only need to do this here in the read case.
2886 			 */
2887 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2888 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2889 			    "has unexpected dirty bits", m));
2890 			vfs_page_set_valid(bp, foff, m);
2891 		}
2892 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2893 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2894 		    (intmax_t)foff, (uintmax_t)m->pindex));
2895 
2896 		vm_page_sunbusy(m);
2897 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2898 		iosize -= resid;
2899 	}
2900 	vm_object_pip_wakeupn(obj, bp->b_npages);
2901 	VM_OBJECT_WUNLOCK(obj);
2902 	if (bogus && buf_mapped(bp)) {
2903 		BUF_CHECK_MAPPED(bp);
2904 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2905 		    bp->b_pages, bp->b_npages);
2906 	}
2907 }
2908 
2909 /*
2910  * Unwire a page held by a buf and either free it or update the page queues to
2911  * reflect its recent use.
2912  */
2913 static void
2914 vfs_vmio_unwire(struct buf *bp, vm_page_t m)
2915 {
2916 	bool freed;
2917 
2918 	vm_page_lock(m);
2919 	if (vm_page_unwire_noq(m)) {
2920 		if ((bp->b_flags & B_DIRECT) != 0)
2921 			freed = vm_page_try_to_free(m);
2922 		else
2923 			freed = false;
2924 		if (!freed) {
2925 			/*
2926 			 * Use a racy check of the valid bits to determine
2927 			 * whether we can accelerate reclamation of the page.
2928 			 * The valid bits will be stable unless the page is
2929 			 * being mapped or is referenced by multiple buffers,
2930 			 * and in those cases we expect races to be rare.  At
2931 			 * worst we will either accelerate reclamation of a
2932 			 * valid page and violate LRU, or unnecessarily defer
2933 			 * reclamation of an invalid page.
2934 			 *
2935 			 * The B_NOREUSE flag marks data that is not expected to
2936 			 * be reused, so accelerate reclamation in that case
2937 			 * too.  Otherwise, maintain LRU.
2938 			 */
2939 			if (m->valid == 0 || (bp->b_flags & B_NOREUSE) != 0)
2940 				vm_page_deactivate_noreuse(m);
2941 			else if (vm_page_active(m))
2942 				vm_page_reference(m);
2943 			else
2944 				vm_page_deactivate(m);
2945 		}
2946 	}
2947 	vm_page_unlock(m);
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 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 	obj = bp->b_bufobj->bo_object;
2979 	resid = bp->b_bufsize;
2980 	poffset = bp->b_offset & PAGE_MASK;
2981 	VM_OBJECT_WLOCK(obj);
2982 	for (i = 0; i < bp->b_npages; i++) {
2983 		m = bp->b_pages[i];
2984 		if (m == bogus_page)
2985 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2986 		bp->b_pages[i] = NULL;
2987 
2988 		presid = resid > (PAGE_SIZE - poffset) ?
2989 		    (PAGE_SIZE - poffset) : resid;
2990 		KASSERT(presid >= 0, ("brelse: extra page"));
2991 		while (vm_page_xbusied(m)) {
2992 			vm_page_lock(m);
2993 			VM_OBJECT_WUNLOCK(obj);
2994 			vm_page_busy_sleep(m, "mbncsh", true);
2995 			VM_OBJECT_WLOCK(obj);
2996 		}
2997 		if (pmap_page_wired_mappings(m) == 0)
2998 			vm_page_set_invalid(m, poffset, presid);
2999 		vfs_vmio_unwire(bp, m);
3000 		resid -= presid;
3001 		poffset = 0;
3002 	}
3003 	VM_OBJECT_WUNLOCK(obj);
3004 	bp->b_npages = 0;
3005 }
3006 
3007 /*
3008  * Page-granular truncation of an existing VMIO buffer.
3009  */
3010 static void
3011 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3012 {
3013 	vm_object_t obj;
3014 	vm_page_t m;
3015 	int i;
3016 
3017 	if (bp->b_npages == desiredpages)
3018 		return;
3019 
3020 	if (buf_mapped(bp)) {
3021 		BUF_CHECK_MAPPED(bp);
3022 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3023 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3024 	} else
3025 		BUF_CHECK_UNMAPPED(bp);
3026 
3027 	/*
3028 	 * The object lock is needed only if we will attempt to free pages.
3029 	 */
3030 	obj = (bp->b_flags & B_DIRECT) != 0 ? bp->b_bufobj->bo_object : NULL;
3031 	if (obj != NULL)
3032 		VM_OBJECT_WLOCK(obj);
3033 	for (i = desiredpages; i < bp->b_npages; i++) {
3034 		m = bp->b_pages[i];
3035 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3036 		bp->b_pages[i] = NULL;
3037 		vfs_vmio_unwire(bp, m);
3038 	}
3039 	if (obj != NULL)
3040 		VM_OBJECT_WUNLOCK(obj);
3041 	bp->b_npages = desiredpages;
3042 }
3043 
3044 /*
3045  * Byte granular extension of VMIO buffers.
3046  */
3047 static void
3048 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3049 {
3050 	/*
3051 	 * We are growing the buffer, possibly in a
3052 	 * byte-granular fashion.
3053 	 */
3054 	vm_object_t obj;
3055 	vm_offset_t toff;
3056 	vm_offset_t tinc;
3057 	vm_page_t m;
3058 
3059 	/*
3060 	 * Step 1, bring in the VM pages from the object, allocating
3061 	 * them if necessary.  We must clear B_CACHE if these pages
3062 	 * are not valid for the range covered by the buffer.
3063 	 */
3064 	obj = bp->b_bufobj->bo_object;
3065 	VM_OBJECT_WLOCK(obj);
3066 	if (bp->b_npages < desiredpages) {
3067 		/*
3068 		 * We must allocate system pages since blocking
3069 		 * here could interfere with paging I/O, no
3070 		 * matter which process we are.
3071 		 *
3072 		 * Only exclusive busy can be tested here.
3073 		 * Blocking on shared busy might lead to
3074 		 * deadlocks once allocbuf() is called after
3075 		 * pages are vfs_busy_pages().
3076 		 */
3077 		(void)vm_page_grab_pages(obj,
3078 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3079 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3080 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3081 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3082 		bp->b_npages = desiredpages;
3083 	}
3084 
3085 	/*
3086 	 * Step 2.  We've loaded the pages into the buffer,
3087 	 * we have to figure out if we can still have B_CACHE
3088 	 * set.  Note that B_CACHE is set according to the
3089 	 * byte-granular range ( bcount and size ), not the
3090 	 * aligned range ( newbsize ).
3091 	 *
3092 	 * The VM test is against m->valid, which is DEV_BSIZE
3093 	 * aligned.  Needless to say, the validity of the data
3094 	 * needs to also be DEV_BSIZE aligned.  Note that this
3095 	 * fails with NFS if the server or some other client
3096 	 * extends the file's EOF.  If our buffer is resized,
3097 	 * B_CACHE may remain set! XXX
3098 	 */
3099 	toff = bp->b_bcount;
3100 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3101 	while ((bp->b_flags & B_CACHE) && toff < size) {
3102 		vm_pindex_t pi;
3103 
3104 		if (tinc > (size - toff))
3105 			tinc = size - toff;
3106 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3107 		m = bp->b_pages[pi];
3108 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3109 		toff += tinc;
3110 		tinc = PAGE_SIZE;
3111 	}
3112 	VM_OBJECT_WUNLOCK(obj);
3113 
3114 	/*
3115 	 * Step 3, fixup the KVA pmap.
3116 	 */
3117 	if (buf_mapped(bp))
3118 		bpmap_qenter(bp);
3119 	else
3120 		BUF_CHECK_UNMAPPED(bp);
3121 }
3122 
3123 /*
3124  * Check to see if a block at a particular lbn is available for a clustered
3125  * write.
3126  */
3127 static int
3128 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3129 {
3130 	struct buf *bpa;
3131 	int match;
3132 
3133 	match = 0;
3134 
3135 	/* If the buf isn't in core skip it */
3136 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3137 		return (0);
3138 
3139 	/* If the buf is busy we don't want to wait for it */
3140 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3141 		return (0);
3142 
3143 	/* Only cluster with valid clusterable delayed write buffers */
3144 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3145 	    (B_DELWRI | B_CLUSTEROK))
3146 		goto done;
3147 
3148 	if (bpa->b_bufsize != size)
3149 		goto done;
3150 
3151 	/*
3152 	 * Check to see if it is in the expected place on disk and that the
3153 	 * block has been mapped.
3154 	 */
3155 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3156 		match = 1;
3157 done:
3158 	BUF_UNLOCK(bpa);
3159 	return (match);
3160 }
3161 
3162 /*
3163  *	vfs_bio_awrite:
3164  *
3165  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3166  *	This is much better then the old way of writing only one buffer at
3167  *	a time.  Note that we may not be presented with the buffers in the
3168  *	correct order, so we search for the cluster in both directions.
3169  */
3170 int
3171 vfs_bio_awrite(struct buf *bp)
3172 {
3173 	struct bufobj *bo;
3174 	int i;
3175 	int j;
3176 	daddr_t lblkno = bp->b_lblkno;
3177 	struct vnode *vp = bp->b_vp;
3178 	int ncl;
3179 	int nwritten;
3180 	int size;
3181 	int maxcl;
3182 	int gbflags;
3183 
3184 	bo = &vp->v_bufobj;
3185 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3186 	/*
3187 	 * right now we support clustered writing only to regular files.  If
3188 	 * we find a clusterable block we could be in the middle of a cluster
3189 	 * rather then at the beginning.
3190 	 */
3191 	if ((vp->v_type == VREG) &&
3192 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3193 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3194 
3195 		size = vp->v_mount->mnt_stat.f_iosize;
3196 		maxcl = MAXPHYS / size;
3197 
3198 		BO_RLOCK(bo);
3199 		for (i = 1; i < maxcl; i++)
3200 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3201 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3202 				break;
3203 
3204 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3205 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3206 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3207 				break;
3208 		BO_RUNLOCK(bo);
3209 		--j;
3210 		ncl = i + j;
3211 		/*
3212 		 * this is a possible cluster write
3213 		 */
3214 		if (ncl != 1) {
3215 			BUF_UNLOCK(bp);
3216 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3217 			    gbflags);
3218 			return (nwritten);
3219 		}
3220 	}
3221 	bremfree(bp);
3222 	bp->b_flags |= B_ASYNC;
3223 	/*
3224 	 * default (old) behavior, writing out only one block
3225 	 *
3226 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3227 	 */
3228 	nwritten = bp->b_bufsize;
3229 	(void) bwrite(bp);
3230 
3231 	return (nwritten);
3232 }
3233 
3234 /*
3235  *	getnewbuf_kva:
3236  *
3237  *	Allocate KVA for an empty buf header according to gbflags.
3238  */
3239 static int
3240 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3241 {
3242 
3243 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3244 		/*
3245 		 * In order to keep fragmentation sane we only allocate kva
3246 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3247 		 */
3248 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3249 
3250 		if (maxsize != bp->b_kvasize &&
3251 		    bufkva_alloc(bp, maxsize, gbflags))
3252 			return (ENOSPC);
3253 	}
3254 	return (0);
3255 }
3256 
3257 /*
3258  *	getnewbuf:
3259  *
3260  *	Find and initialize a new buffer header, freeing up existing buffers
3261  *	in the bufqueues as necessary.  The new buffer is returned locked.
3262  *
3263  *	We block if:
3264  *		We have insufficient buffer headers
3265  *		We have insufficient buffer space
3266  *		buffer_arena is too fragmented ( space reservation fails )
3267  *		If we have to flush dirty buffers ( but we try to avoid this )
3268  *
3269  *	The caller is responsible for releasing the reserved bufspace after
3270  *	allocbuf() is called.
3271  */
3272 static struct buf *
3273 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3274 {
3275 	struct bufdomain *bd;
3276 	struct buf *bp;
3277 	bool metadata, reserved;
3278 
3279 	bp = NULL;
3280 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3281 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3282 	if (!unmapped_buf_allowed)
3283 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3284 
3285 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3286 	    vp->v_type == VCHR)
3287 		metadata = true;
3288 	else
3289 		metadata = false;
3290 	if (vp == NULL)
3291 		bd = &bdomain[0];
3292 	else
3293 		bd = &bdomain[vp->v_bufobj.bo_domain];
3294 
3295 	counter_u64_add(getnewbufcalls, 1);
3296 	reserved = false;
3297 	do {
3298 		if (reserved == false &&
3299 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3300 			counter_u64_add(getnewbufrestarts, 1);
3301 			continue;
3302 		}
3303 		reserved = true;
3304 		if ((bp = buf_alloc(bd)) == NULL) {
3305 			counter_u64_add(getnewbufrestarts, 1);
3306 			continue;
3307 		}
3308 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3309 			return (bp);
3310 		break;
3311 	} while (buf_recycle(bd, false) == 0);
3312 
3313 	if (reserved)
3314 		bufspace_release(bd, maxsize);
3315 	if (bp != NULL) {
3316 		bp->b_flags |= B_INVAL;
3317 		brelse(bp);
3318 	}
3319 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3320 
3321 	return (NULL);
3322 }
3323 
3324 /*
3325  *	buf_daemon:
3326  *
3327  *	buffer flushing daemon.  Buffers are normally flushed by the
3328  *	update daemon but if it cannot keep up this process starts to
3329  *	take the load in an attempt to prevent getnewbuf() from blocking.
3330  */
3331 static struct kproc_desc buf_kp = {
3332 	"bufdaemon",
3333 	buf_daemon,
3334 	&bufdaemonproc
3335 };
3336 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3337 
3338 static int
3339 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3340 {
3341 	int flushed;
3342 
3343 	flushed = flushbufqueues(vp, bd, target, 0);
3344 	if (flushed == 0) {
3345 		/*
3346 		 * Could not find any buffers without rollback
3347 		 * dependencies, so just write the first one
3348 		 * in the hopes of eventually making progress.
3349 		 */
3350 		if (vp != NULL && target > 2)
3351 			target /= 2;
3352 		flushbufqueues(vp, bd, target, 1);
3353 	}
3354 	return (flushed);
3355 }
3356 
3357 static void
3358 buf_daemon()
3359 {
3360 	struct bufdomain *bd;
3361 	int speedupreq;
3362 	int lodirty;
3363 	int i;
3364 
3365 	/*
3366 	 * This process needs to be suspended prior to shutdown sync.
3367 	 */
3368 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3369 	    SHUTDOWN_PRI_LAST + 100);
3370 
3371 	/*
3372 	 * Start the buf clean daemons as children threads.
3373 	 */
3374 	for (i = 0 ; i < buf_domains; i++) {
3375 		int error;
3376 
3377 		error = kthread_add((void (*)(void *))bufspace_daemon,
3378 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3379 		if (error)
3380 			panic("error %d spawning bufspace daemon", error);
3381 	}
3382 
3383 	/*
3384 	 * This process is allowed to take the buffer cache to the limit
3385 	 */
3386 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3387 	mtx_lock(&bdlock);
3388 	for (;;) {
3389 		bd_request = 0;
3390 		mtx_unlock(&bdlock);
3391 
3392 		kthread_suspend_check();
3393 
3394 		/*
3395 		 * Save speedupreq for this pass and reset to capture new
3396 		 * requests.
3397 		 */
3398 		speedupreq = bd_speedupreq;
3399 		bd_speedupreq = 0;
3400 
3401 		/*
3402 		 * Flush each domain sequentially according to its level and
3403 		 * the speedup request.
3404 		 */
3405 		for (i = 0; i < buf_domains; i++) {
3406 			bd = &bdomain[i];
3407 			if (speedupreq)
3408 				lodirty = bd->bd_numdirtybuffers / 2;
3409 			else
3410 				lodirty = bd->bd_lodirtybuffers;
3411 			while (bd->bd_numdirtybuffers > lodirty) {
3412 				if (buf_flush(NULL, bd,
3413 				    bd->bd_numdirtybuffers - lodirty) == 0)
3414 					break;
3415 				kern_yield(PRI_USER);
3416 			}
3417 		}
3418 
3419 		/*
3420 		 * Only clear bd_request if we have reached our low water
3421 		 * mark.  The buf_daemon normally waits 1 second and
3422 		 * then incrementally flushes any dirty buffers that have
3423 		 * built up, within reason.
3424 		 *
3425 		 * If we were unable to hit our low water mark and couldn't
3426 		 * find any flushable buffers, we sleep for a short period
3427 		 * to avoid endless loops on unlockable buffers.
3428 		 */
3429 		mtx_lock(&bdlock);
3430 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3431 			/*
3432 			 * We reached our low water mark, reset the
3433 			 * request and sleep until we are needed again.
3434 			 * The sleep is just so the suspend code works.
3435 			 */
3436 			bd_request = 0;
3437 			/*
3438 			 * Do an extra wakeup in case dirty threshold
3439 			 * changed via sysctl and the explicit transition
3440 			 * out of shortfall was missed.
3441 			 */
3442 			bdirtywakeup();
3443 			if (runningbufspace <= lorunningspace)
3444 				runningwakeup();
3445 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3446 		} else {
3447 			/*
3448 			 * We couldn't find any flushable dirty buffers but
3449 			 * still have too many dirty buffers, we
3450 			 * have to sleep and try again.  (rare)
3451 			 */
3452 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3453 		}
3454 	}
3455 }
3456 
3457 /*
3458  *	flushbufqueues:
3459  *
3460  *	Try to flush a buffer in the dirty queue.  We must be careful to
3461  *	free up B_INVAL buffers instead of write them, which NFS is
3462  *	particularly sensitive to.
3463  */
3464 static int flushwithdeps = 0;
3465 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
3466     0, "Number of buffers flushed with dependecies that require rollbacks");
3467 
3468 static int
3469 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3470     int flushdeps)
3471 {
3472 	struct bufqueue *bq;
3473 	struct buf *sentinel;
3474 	struct vnode *vp;
3475 	struct mount *mp;
3476 	struct buf *bp;
3477 	int hasdeps;
3478 	int flushed;
3479 	int error;
3480 	bool unlock;
3481 
3482 	flushed = 0;
3483 	bq = &bd->bd_dirtyq;
3484 	bp = NULL;
3485 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3486 	sentinel->b_qindex = QUEUE_SENTINEL;
3487 	BQ_LOCK(bq);
3488 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3489 	BQ_UNLOCK(bq);
3490 	while (flushed != target) {
3491 		maybe_yield();
3492 		BQ_LOCK(bq);
3493 		bp = TAILQ_NEXT(sentinel, b_freelist);
3494 		if (bp != NULL) {
3495 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3496 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3497 			    b_freelist);
3498 		} else {
3499 			BQ_UNLOCK(bq);
3500 			break;
3501 		}
3502 		/*
3503 		 * Skip sentinels inserted by other invocations of the
3504 		 * flushbufqueues(), taking care to not reorder them.
3505 		 *
3506 		 * Only flush the buffers that belong to the
3507 		 * vnode locked by the curthread.
3508 		 */
3509 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3510 		    bp->b_vp != lvp)) {
3511 			BQ_UNLOCK(bq);
3512 			continue;
3513 		}
3514 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3515 		BQ_UNLOCK(bq);
3516 		if (error != 0)
3517 			continue;
3518 
3519 		/*
3520 		 * BKGRDINPROG can only be set with the buf and bufobj
3521 		 * locks both held.  We tolerate a race to clear it here.
3522 		 */
3523 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3524 		    (bp->b_flags & B_DELWRI) == 0) {
3525 			BUF_UNLOCK(bp);
3526 			continue;
3527 		}
3528 		if (bp->b_flags & B_INVAL) {
3529 			bremfreef(bp);
3530 			brelse(bp);
3531 			flushed++;
3532 			continue;
3533 		}
3534 
3535 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3536 			if (flushdeps == 0) {
3537 				BUF_UNLOCK(bp);
3538 				continue;
3539 			}
3540 			hasdeps = 1;
3541 		} else
3542 			hasdeps = 0;
3543 		/*
3544 		 * We must hold the lock on a vnode before writing
3545 		 * one of its buffers. Otherwise we may confuse, or
3546 		 * in the case of a snapshot vnode, deadlock the
3547 		 * system.
3548 		 *
3549 		 * The lock order here is the reverse of the normal
3550 		 * of vnode followed by buf lock.  This is ok because
3551 		 * the NOWAIT will prevent deadlock.
3552 		 */
3553 		vp = bp->b_vp;
3554 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3555 			BUF_UNLOCK(bp);
3556 			continue;
3557 		}
3558 		if (lvp == NULL) {
3559 			unlock = true;
3560 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3561 		} else {
3562 			ASSERT_VOP_LOCKED(vp, "getbuf");
3563 			unlock = false;
3564 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3565 			    vn_lock(vp, LK_TRYUPGRADE);
3566 		}
3567 		if (error == 0) {
3568 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3569 			    bp, bp->b_vp, bp->b_flags);
3570 			if (curproc == bufdaemonproc) {
3571 				vfs_bio_awrite(bp);
3572 			} else {
3573 				bremfree(bp);
3574 				bwrite(bp);
3575 				counter_u64_add(notbufdflushes, 1);
3576 			}
3577 			vn_finished_write(mp);
3578 			if (unlock)
3579 				VOP_UNLOCK(vp, 0);
3580 			flushwithdeps += hasdeps;
3581 			flushed++;
3582 
3583 			/*
3584 			 * Sleeping on runningbufspace while holding
3585 			 * vnode lock leads to deadlock.
3586 			 */
3587 			if (curproc == bufdaemonproc &&
3588 			    runningbufspace > hirunningspace)
3589 				waitrunningbufspace();
3590 			continue;
3591 		}
3592 		vn_finished_write(mp);
3593 		BUF_UNLOCK(bp);
3594 	}
3595 	BQ_LOCK(bq);
3596 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3597 	BQ_UNLOCK(bq);
3598 	free(sentinel, M_TEMP);
3599 	return (flushed);
3600 }
3601 
3602 /*
3603  * Check to see if a block is currently memory resident.
3604  */
3605 struct buf *
3606 incore(struct bufobj *bo, daddr_t blkno)
3607 {
3608 	struct buf *bp;
3609 
3610 	BO_RLOCK(bo);
3611 	bp = gbincore(bo, blkno);
3612 	BO_RUNLOCK(bo);
3613 	return (bp);
3614 }
3615 
3616 /*
3617  * Returns true if no I/O is needed to access the
3618  * associated VM object.  This is like incore except
3619  * it also hunts around in the VM system for the data.
3620  */
3621 
3622 static int
3623 inmem(struct vnode * vp, daddr_t blkno)
3624 {
3625 	vm_object_t obj;
3626 	vm_offset_t toff, tinc, size;
3627 	vm_page_t m;
3628 	vm_ooffset_t off;
3629 
3630 	ASSERT_VOP_LOCKED(vp, "inmem");
3631 
3632 	if (incore(&vp->v_bufobj, blkno))
3633 		return 1;
3634 	if (vp->v_mount == NULL)
3635 		return 0;
3636 	obj = vp->v_object;
3637 	if (obj == NULL)
3638 		return (0);
3639 
3640 	size = PAGE_SIZE;
3641 	if (size > vp->v_mount->mnt_stat.f_iosize)
3642 		size = vp->v_mount->mnt_stat.f_iosize;
3643 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3644 
3645 	VM_OBJECT_RLOCK(obj);
3646 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3647 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3648 		if (!m)
3649 			goto notinmem;
3650 		tinc = size;
3651 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3652 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3653 		if (vm_page_is_valid(m,
3654 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3655 			goto notinmem;
3656 	}
3657 	VM_OBJECT_RUNLOCK(obj);
3658 	return 1;
3659 
3660 notinmem:
3661 	VM_OBJECT_RUNLOCK(obj);
3662 	return (0);
3663 }
3664 
3665 /*
3666  * Set the dirty range for a buffer based on the status of the dirty
3667  * bits in the pages comprising the buffer.  The range is limited
3668  * to the size of the buffer.
3669  *
3670  * Tell the VM system that the pages associated with this buffer
3671  * are clean.  This is used for delayed writes where the data is
3672  * going to go to disk eventually without additional VM intevention.
3673  *
3674  * Note that while we only really need to clean through to b_bcount, we
3675  * just go ahead and clean through to b_bufsize.
3676  */
3677 static void
3678 vfs_clean_pages_dirty_buf(struct buf *bp)
3679 {
3680 	vm_ooffset_t foff, noff, eoff;
3681 	vm_page_t m;
3682 	int i;
3683 
3684 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3685 		return;
3686 
3687 	foff = bp->b_offset;
3688 	KASSERT(bp->b_offset != NOOFFSET,
3689 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3690 
3691 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3692 	vfs_drain_busy_pages(bp);
3693 	vfs_setdirty_locked_object(bp);
3694 	for (i = 0; i < bp->b_npages; i++) {
3695 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3696 		eoff = noff;
3697 		if (eoff > bp->b_offset + bp->b_bufsize)
3698 			eoff = bp->b_offset + bp->b_bufsize;
3699 		m = bp->b_pages[i];
3700 		vfs_page_set_validclean(bp, foff, m);
3701 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3702 		foff = noff;
3703 	}
3704 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3705 }
3706 
3707 static void
3708 vfs_setdirty_locked_object(struct buf *bp)
3709 {
3710 	vm_object_t object;
3711 	int i;
3712 
3713 	object = bp->b_bufobj->bo_object;
3714 	VM_OBJECT_ASSERT_WLOCKED(object);
3715 
3716 	/*
3717 	 * We qualify the scan for modified pages on whether the
3718 	 * object has been flushed yet.
3719 	 */
3720 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3721 		vm_offset_t boffset;
3722 		vm_offset_t eoffset;
3723 
3724 		/*
3725 		 * test the pages to see if they have been modified directly
3726 		 * by users through the VM system.
3727 		 */
3728 		for (i = 0; i < bp->b_npages; i++)
3729 			vm_page_test_dirty(bp->b_pages[i]);
3730 
3731 		/*
3732 		 * Calculate the encompassing dirty range, boffset and eoffset,
3733 		 * (eoffset - boffset) bytes.
3734 		 */
3735 
3736 		for (i = 0; i < bp->b_npages; i++) {
3737 			if (bp->b_pages[i]->dirty)
3738 				break;
3739 		}
3740 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3741 
3742 		for (i = bp->b_npages - 1; i >= 0; --i) {
3743 			if (bp->b_pages[i]->dirty) {
3744 				break;
3745 			}
3746 		}
3747 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3748 
3749 		/*
3750 		 * Fit it to the buffer.
3751 		 */
3752 
3753 		if (eoffset > bp->b_bcount)
3754 			eoffset = bp->b_bcount;
3755 
3756 		/*
3757 		 * If we have a good dirty range, merge with the existing
3758 		 * dirty range.
3759 		 */
3760 
3761 		if (boffset < eoffset) {
3762 			if (bp->b_dirtyoff > boffset)
3763 				bp->b_dirtyoff = boffset;
3764 			if (bp->b_dirtyend < eoffset)
3765 				bp->b_dirtyend = eoffset;
3766 		}
3767 	}
3768 }
3769 
3770 /*
3771  * Allocate the KVA mapping for an existing buffer.
3772  * If an unmapped buffer is provided but a mapped buffer is requested, take
3773  * also care to properly setup mappings between pages and KVA.
3774  */
3775 static void
3776 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3777 {
3778 	int bsize, maxsize, need_mapping, need_kva;
3779 	off_t offset;
3780 
3781 	need_mapping = bp->b_data == unmapped_buf &&
3782 	    (gbflags & GB_UNMAPPED) == 0;
3783 	need_kva = bp->b_kvabase == unmapped_buf &&
3784 	    bp->b_data == unmapped_buf &&
3785 	    (gbflags & GB_KVAALLOC) != 0;
3786 	if (!need_mapping && !need_kva)
3787 		return;
3788 
3789 	BUF_CHECK_UNMAPPED(bp);
3790 
3791 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3792 		/*
3793 		 * Buffer is not mapped, but the KVA was already
3794 		 * reserved at the time of the instantiation.  Use the
3795 		 * allocated space.
3796 		 */
3797 		goto has_addr;
3798 	}
3799 
3800 	/*
3801 	 * Calculate the amount of the address space we would reserve
3802 	 * if the buffer was mapped.
3803 	 */
3804 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3805 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3806 	offset = blkno * bsize;
3807 	maxsize = size + (offset & PAGE_MASK);
3808 	maxsize = imax(maxsize, bsize);
3809 
3810 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3811 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3812 			/*
3813 			 * XXXKIB: defragmentation cannot
3814 			 * succeed, not sure what else to do.
3815 			 */
3816 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3817 		}
3818 		counter_u64_add(mappingrestarts, 1);
3819 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3820 	}
3821 has_addr:
3822 	if (need_mapping) {
3823 		/* b_offset is handled by bpmap_qenter. */
3824 		bp->b_data = bp->b_kvabase;
3825 		BUF_CHECK_MAPPED(bp);
3826 		bpmap_qenter(bp);
3827 	}
3828 }
3829 
3830 struct buf *
3831 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3832     int flags)
3833 {
3834 	struct buf *bp;
3835 	int error;
3836 
3837 	error = getblkx(vp, blkno, size, slpflag, slptimeo, flags, &bp);
3838 	if (error != 0)
3839 		return (NULL);
3840 	return (bp);
3841 }
3842 
3843 /*
3844  *	getblkx:
3845  *
3846  *	Get a block given a specified block and offset into a file/device.
3847  *	The buffers B_DONE bit will be cleared on return, making it almost
3848  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3849  *	return.  The caller should clear B_INVAL prior to initiating a
3850  *	READ.
3851  *
3852  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3853  *	an existing buffer.
3854  *
3855  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3856  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3857  *	and then cleared based on the backing VM.  If the previous buffer is
3858  *	non-0-sized but invalid, B_CACHE will be cleared.
3859  *
3860  *	If getblk() must create a new buffer, the new buffer is returned with
3861  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3862  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3863  *	backing VM.
3864  *
3865  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3866  *	B_CACHE bit is clear.
3867  *
3868  *	What this means, basically, is that the caller should use B_CACHE to
3869  *	determine whether the buffer is fully valid or not and should clear
3870  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3871  *	the buffer by loading its data area with something, the caller needs
3872  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3873  *	the caller should set B_CACHE ( as an optimization ), else the caller
3874  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3875  *	a write attempt or if it was a successful read.  If the caller
3876  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3877  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3878  */
3879 int
3880 getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3881     int flags, struct buf **bpp)
3882 {
3883 	struct buf *bp;
3884 	struct bufobj *bo;
3885 	daddr_t d_blkno;
3886 	int bsize, error, maxsize, vmio;
3887 	off_t offset;
3888 
3889 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3890 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3891 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3892 	ASSERT_VOP_LOCKED(vp, "getblk");
3893 	if (size > maxbcachebuf)
3894 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3895 		    maxbcachebuf);
3896 	if (!unmapped_buf_allowed)
3897 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3898 
3899 	bo = &vp->v_bufobj;
3900 	d_blkno = blkno;
3901 loop:
3902 	BO_RLOCK(bo);
3903 	bp = gbincore(bo, blkno);
3904 	if (bp != NULL) {
3905 		int lockflags;
3906 		/*
3907 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3908 		 * it must be on a queue.
3909 		 */
3910 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3911 
3912 		if ((flags & GB_LOCK_NOWAIT) != 0)
3913 			lockflags |= LK_NOWAIT;
3914 
3915 		error = BUF_TIMELOCK(bp, lockflags,
3916 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3917 
3918 		/*
3919 		 * If we slept and got the lock we have to restart in case
3920 		 * the buffer changed identities.
3921 		 */
3922 		if (error == ENOLCK)
3923 			goto loop;
3924 		/* We timed out or were interrupted. */
3925 		else if (error != 0)
3926 			return (error);
3927 		/* If recursed, assume caller knows the rules. */
3928 		else if (BUF_LOCKRECURSED(bp))
3929 			goto end;
3930 
3931 		/*
3932 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3933 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3934 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3935 		 * backing VM cache.
3936 		 */
3937 		if (bp->b_flags & B_INVAL)
3938 			bp->b_flags &= ~B_CACHE;
3939 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3940 			bp->b_flags |= B_CACHE;
3941 		if (bp->b_flags & B_MANAGED)
3942 			MPASS(bp->b_qindex == QUEUE_NONE);
3943 		else
3944 			bremfree(bp);
3945 
3946 		/*
3947 		 * check for size inconsistencies for non-VMIO case.
3948 		 */
3949 		if (bp->b_bcount != size) {
3950 			if ((bp->b_flags & B_VMIO) == 0 ||
3951 			    (size > bp->b_kvasize)) {
3952 				if (bp->b_flags & B_DELWRI) {
3953 					bp->b_flags |= B_NOCACHE;
3954 					bwrite(bp);
3955 				} else {
3956 					if (LIST_EMPTY(&bp->b_dep)) {
3957 						bp->b_flags |= B_RELBUF;
3958 						brelse(bp);
3959 					} else {
3960 						bp->b_flags |= B_NOCACHE;
3961 						bwrite(bp);
3962 					}
3963 				}
3964 				goto loop;
3965 			}
3966 		}
3967 
3968 		/*
3969 		 * Handle the case of unmapped buffer which should
3970 		 * become mapped, or the buffer for which KVA
3971 		 * reservation is requested.
3972 		 */
3973 		bp_unmapped_get_kva(bp, blkno, size, flags);
3974 
3975 		/*
3976 		 * If the size is inconsistent in the VMIO case, we can resize
3977 		 * the buffer.  This might lead to B_CACHE getting set or
3978 		 * cleared.  If the size has not changed, B_CACHE remains
3979 		 * unchanged from its previous state.
3980 		 */
3981 		allocbuf(bp, size);
3982 
3983 		KASSERT(bp->b_offset != NOOFFSET,
3984 		    ("getblk: no buffer offset"));
3985 
3986 		/*
3987 		 * A buffer with B_DELWRI set and B_CACHE clear must
3988 		 * be committed before we can return the buffer in
3989 		 * order to prevent the caller from issuing a read
3990 		 * ( due to B_CACHE not being set ) and overwriting
3991 		 * it.
3992 		 *
3993 		 * Most callers, including NFS and FFS, need this to
3994 		 * operate properly either because they assume they
3995 		 * can issue a read if B_CACHE is not set, or because
3996 		 * ( for example ) an uncached B_DELWRI might loop due
3997 		 * to softupdates re-dirtying the buffer.  In the latter
3998 		 * case, B_CACHE is set after the first write completes,
3999 		 * preventing further loops.
4000 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
4001 		 * above while extending the buffer, we cannot allow the
4002 		 * buffer to remain with B_CACHE set after the write
4003 		 * completes or it will represent a corrupt state.  To
4004 		 * deal with this we set B_NOCACHE to scrap the buffer
4005 		 * after the write.
4006 		 *
4007 		 * We might be able to do something fancy, like setting
4008 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
4009 		 * so the below call doesn't set B_CACHE, but that gets real
4010 		 * confusing.  This is much easier.
4011 		 */
4012 
4013 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4014 			bp->b_flags |= B_NOCACHE;
4015 			bwrite(bp);
4016 			goto loop;
4017 		}
4018 		bp->b_flags &= ~B_DONE;
4019 	} else {
4020 		/*
4021 		 * Buffer is not in-core, create new buffer.  The buffer
4022 		 * returned by getnewbuf() is locked.  Note that the returned
4023 		 * buffer is also considered valid (not marked B_INVAL).
4024 		 */
4025 		BO_RUNLOCK(bo);
4026 		/*
4027 		 * If the user does not want us to create the buffer, bail out
4028 		 * here.
4029 		 */
4030 		if (flags & GB_NOCREAT)
4031 			return (EEXIST);
4032 		if (bdomain[bo->bo_domain].bd_freebuffers == 0 &&
4033 		    TD_IS_IDLETHREAD(curthread))
4034 			return (EBUSY);
4035 
4036 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
4037 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4038 		offset = blkno * bsize;
4039 		vmio = vp->v_object != NULL;
4040 		if (vmio) {
4041 			maxsize = size + (offset & PAGE_MASK);
4042 		} else {
4043 			maxsize = size;
4044 			/* Do not allow non-VMIO notmapped buffers. */
4045 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4046 		}
4047 		maxsize = imax(maxsize, bsize);
4048 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4049 		    !vn_isdisk(vp, NULL)) {
4050 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4051 			KASSERT(error != EOPNOTSUPP,
4052 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4053 			    vp));
4054 			if (error != 0)
4055 				return (error);
4056 			if (d_blkno == -1)
4057 				return (EJUSTRETURN);
4058 		}
4059 
4060 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4061 		if (bp == NULL) {
4062 			if (slpflag || slptimeo)
4063 				return (ETIMEDOUT);
4064 			/*
4065 			 * XXX This is here until the sleep path is diagnosed
4066 			 * enough to work under very low memory conditions.
4067 			 *
4068 			 * There's an issue on low memory, 4BSD+non-preempt
4069 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4070 			 * exhaustion occurs without sleeping for buffer
4071 			 * reclaimation.  This just sticks in a loop and
4072 			 * constantly attempts to allocate a buffer, which
4073 			 * hits exhaustion and tries to wakeup bufdaemon.
4074 			 * This never happens because we never yield.
4075 			 *
4076 			 * The real solution is to identify and fix these cases
4077 			 * so we aren't effectively busy-waiting in a loop
4078 			 * until the reclaimation path has cycles to run.
4079 			 */
4080 			kern_yield(PRI_USER);
4081 			goto loop;
4082 		}
4083 
4084 		/*
4085 		 * This code is used to make sure that a buffer is not
4086 		 * created while the getnewbuf routine is blocked.
4087 		 * This can be a problem whether the vnode is locked or not.
4088 		 * If the buffer is created out from under us, we have to
4089 		 * throw away the one we just created.
4090 		 *
4091 		 * Note: this must occur before we associate the buffer
4092 		 * with the vp especially considering limitations in
4093 		 * the splay tree implementation when dealing with duplicate
4094 		 * lblkno's.
4095 		 */
4096 		BO_LOCK(bo);
4097 		if (gbincore(bo, blkno)) {
4098 			BO_UNLOCK(bo);
4099 			bp->b_flags |= B_INVAL;
4100 			bufspace_release(bufdomain(bp), maxsize);
4101 			brelse(bp);
4102 			goto loop;
4103 		}
4104 
4105 		/*
4106 		 * Insert the buffer into the hash, so that it can
4107 		 * be found by incore.
4108 		 */
4109 		bp->b_lblkno = blkno;
4110 		bp->b_blkno = d_blkno;
4111 		bp->b_offset = offset;
4112 		bgetvp(vp, bp);
4113 		BO_UNLOCK(bo);
4114 
4115 		/*
4116 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4117 		 * buffer size starts out as 0, B_CACHE will be set by
4118 		 * allocbuf() for the VMIO case prior to it testing the
4119 		 * backing store for validity.
4120 		 */
4121 
4122 		if (vmio) {
4123 			bp->b_flags |= B_VMIO;
4124 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4125 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4126 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4127 		} else {
4128 			bp->b_flags &= ~B_VMIO;
4129 			KASSERT(bp->b_bufobj->bo_object == NULL,
4130 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4131 			    bp, bp->b_bufobj->bo_object));
4132 			BUF_CHECK_MAPPED(bp);
4133 		}
4134 
4135 		allocbuf(bp, size);
4136 		bufspace_release(bufdomain(bp), maxsize);
4137 		bp->b_flags &= ~B_DONE;
4138 	}
4139 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4140 	BUF_ASSERT_HELD(bp);
4141 end:
4142 	buf_track(bp, __func__);
4143 	KASSERT(bp->b_bufobj == bo,
4144 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4145 	*bpp = bp;
4146 	return (0);
4147 }
4148 
4149 /*
4150  * Get an empty, disassociated buffer of given size.  The buffer is initially
4151  * set to B_INVAL.
4152  */
4153 struct buf *
4154 geteblk(int size, int flags)
4155 {
4156 	struct buf *bp;
4157 	int maxsize;
4158 
4159 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4160 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4161 		if ((flags & GB_NOWAIT_BD) &&
4162 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4163 			return (NULL);
4164 	}
4165 	allocbuf(bp, size);
4166 	bufspace_release(bufdomain(bp), maxsize);
4167 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4168 	BUF_ASSERT_HELD(bp);
4169 	return (bp);
4170 }
4171 
4172 /*
4173  * Truncate the backing store for a non-vmio buffer.
4174  */
4175 static void
4176 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4177 {
4178 
4179 	if (bp->b_flags & B_MALLOC) {
4180 		/*
4181 		 * malloced buffers are not shrunk
4182 		 */
4183 		if (newbsize == 0) {
4184 			bufmallocadjust(bp, 0);
4185 			free(bp->b_data, M_BIOBUF);
4186 			bp->b_data = bp->b_kvabase;
4187 			bp->b_flags &= ~B_MALLOC;
4188 		}
4189 		return;
4190 	}
4191 	vm_hold_free_pages(bp, newbsize);
4192 	bufspace_adjust(bp, newbsize);
4193 }
4194 
4195 /*
4196  * Extend the backing for a non-VMIO buffer.
4197  */
4198 static void
4199 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4200 {
4201 	caddr_t origbuf;
4202 	int origbufsize;
4203 
4204 	/*
4205 	 * We only use malloced memory on the first allocation.
4206 	 * and revert to page-allocated memory when the buffer
4207 	 * grows.
4208 	 *
4209 	 * There is a potential smp race here that could lead
4210 	 * to bufmallocspace slightly passing the max.  It
4211 	 * is probably extremely rare and not worth worrying
4212 	 * over.
4213 	 */
4214 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4215 	    bufmallocspace < maxbufmallocspace) {
4216 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4217 		bp->b_flags |= B_MALLOC;
4218 		bufmallocadjust(bp, newbsize);
4219 		return;
4220 	}
4221 
4222 	/*
4223 	 * If the buffer is growing on its other-than-first
4224 	 * allocation then we revert to the page-allocation
4225 	 * scheme.
4226 	 */
4227 	origbuf = NULL;
4228 	origbufsize = 0;
4229 	if (bp->b_flags & B_MALLOC) {
4230 		origbuf = bp->b_data;
4231 		origbufsize = bp->b_bufsize;
4232 		bp->b_data = bp->b_kvabase;
4233 		bufmallocadjust(bp, 0);
4234 		bp->b_flags &= ~B_MALLOC;
4235 		newbsize = round_page(newbsize);
4236 	}
4237 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4238 	    (vm_offset_t) bp->b_data + newbsize);
4239 	if (origbuf != NULL) {
4240 		bcopy(origbuf, bp->b_data, origbufsize);
4241 		free(origbuf, M_BIOBUF);
4242 	}
4243 	bufspace_adjust(bp, newbsize);
4244 }
4245 
4246 /*
4247  * This code constitutes the buffer memory from either anonymous system
4248  * memory (in the case of non-VMIO operations) or from an associated
4249  * VM object (in the case of VMIO operations).  This code is able to
4250  * resize a buffer up or down.
4251  *
4252  * Note that this code is tricky, and has many complications to resolve
4253  * deadlock or inconsistent data situations.  Tread lightly!!!
4254  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4255  * the caller.  Calling this code willy nilly can result in the loss of data.
4256  *
4257  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4258  * B_CACHE for the non-VMIO case.
4259  */
4260 int
4261 allocbuf(struct buf *bp, int size)
4262 {
4263 	int newbsize;
4264 
4265 	BUF_ASSERT_HELD(bp);
4266 
4267 	if (bp->b_bcount == size)
4268 		return (1);
4269 
4270 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4271 		panic("allocbuf: buffer too small");
4272 
4273 	newbsize = roundup2(size, DEV_BSIZE);
4274 	if ((bp->b_flags & B_VMIO) == 0) {
4275 		if ((bp->b_flags & B_MALLOC) == 0)
4276 			newbsize = round_page(newbsize);
4277 		/*
4278 		 * Just get anonymous memory from the kernel.  Don't
4279 		 * mess with B_CACHE.
4280 		 */
4281 		if (newbsize < bp->b_bufsize)
4282 			vfs_nonvmio_truncate(bp, newbsize);
4283 		else if (newbsize > bp->b_bufsize)
4284 			vfs_nonvmio_extend(bp, newbsize);
4285 	} else {
4286 		int desiredpages;
4287 
4288 		desiredpages = (size == 0) ? 0 :
4289 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4290 
4291 		if (bp->b_flags & B_MALLOC)
4292 			panic("allocbuf: VMIO buffer can't be malloced");
4293 		/*
4294 		 * Set B_CACHE initially if buffer is 0 length or will become
4295 		 * 0-length.
4296 		 */
4297 		if (size == 0 || bp->b_bufsize == 0)
4298 			bp->b_flags |= B_CACHE;
4299 
4300 		if (newbsize < bp->b_bufsize)
4301 			vfs_vmio_truncate(bp, desiredpages);
4302 		/* XXX This looks as if it should be newbsize > b_bufsize */
4303 		else if (size > bp->b_bcount)
4304 			vfs_vmio_extend(bp, desiredpages, size);
4305 		bufspace_adjust(bp, newbsize);
4306 	}
4307 	bp->b_bcount = size;		/* requested buffer size. */
4308 	return (1);
4309 }
4310 
4311 extern int inflight_transient_maps;
4312 
4313 static struct bio_queue nondump_bios;
4314 
4315 void
4316 biodone(struct bio *bp)
4317 {
4318 	struct mtx *mtxp;
4319 	void (*done)(struct bio *);
4320 	vm_offset_t start, end;
4321 
4322 	biotrack(bp, __func__);
4323 
4324 	/*
4325 	 * Avoid completing I/O when dumping after a panic since that may
4326 	 * result in a deadlock in the filesystem or pager code.  Note that
4327 	 * this doesn't affect dumps that were started manually since we aim
4328 	 * to keep the system usable after it has been resumed.
4329 	 */
4330 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4331 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4332 		return;
4333 	}
4334 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4335 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4336 		bp->bio_flags |= BIO_UNMAPPED;
4337 		start = trunc_page((vm_offset_t)bp->bio_data);
4338 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4339 		bp->bio_data = unmapped_buf;
4340 		pmap_qremove(start, atop(end - start));
4341 		vmem_free(transient_arena, start, end - start);
4342 		atomic_add_int(&inflight_transient_maps, -1);
4343 	}
4344 	done = bp->bio_done;
4345 	if (done == NULL) {
4346 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4347 		mtx_lock(mtxp);
4348 		bp->bio_flags |= BIO_DONE;
4349 		wakeup(bp);
4350 		mtx_unlock(mtxp);
4351 	} else
4352 		done(bp);
4353 }
4354 
4355 /*
4356  * Wait for a BIO to finish.
4357  */
4358 int
4359 biowait(struct bio *bp, const char *wchan)
4360 {
4361 	struct mtx *mtxp;
4362 
4363 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4364 	mtx_lock(mtxp);
4365 	while ((bp->bio_flags & BIO_DONE) == 0)
4366 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4367 	mtx_unlock(mtxp);
4368 	if (bp->bio_error != 0)
4369 		return (bp->bio_error);
4370 	if (!(bp->bio_flags & BIO_ERROR))
4371 		return (0);
4372 	return (EIO);
4373 }
4374 
4375 void
4376 biofinish(struct bio *bp, struct devstat *stat, int error)
4377 {
4378 
4379 	if (error) {
4380 		bp->bio_error = error;
4381 		bp->bio_flags |= BIO_ERROR;
4382 	}
4383 	if (stat != NULL)
4384 		devstat_end_transaction_bio(stat, bp);
4385 	biodone(bp);
4386 }
4387 
4388 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4389 void
4390 biotrack_buf(struct bio *bp, const char *location)
4391 {
4392 
4393 	buf_track(bp->bio_track_bp, location);
4394 }
4395 #endif
4396 
4397 /*
4398  *	bufwait:
4399  *
4400  *	Wait for buffer I/O completion, returning error status.  The buffer
4401  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4402  *	error and cleared.
4403  */
4404 int
4405 bufwait(struct buf *bp)
4406 {
4407 	if (bp->b_iocmd == BIO_READ)
4408 		bwait(bp, PRIBIO, "biord");
4409 	else
4410 		bwait(bp, PRIBIO, "biowr");
4411 	if (bp->b_flags & B_EINTR) {
4412 		bp->b_flags &= ~B_EINTR;
4413 		return (EINTR);
4414 	}
4415 	if (bp->b_ioflags & BIO_ERROR) {
4416 		return (bp->b_error ? bp->b_error : EIO);
4417 	} else {
4418 		return (0);
4419 	}
4420 }
4421 
4422 /*
4423  *	bufdone:
4424  *
4425  *	Finish I/O on a buffer, optionally calling a completion function.
4426  *	This is usually called from an interrupt so process blocking is
4427  *	not allowed.
4428  *
4429  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4430  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4431  *	assuming B_INVAL is clear.
4432  *
4433  *	For the VMIO case, we set B_CACHE if the op was a read and no
4434  *	read error occurred, or if the op was a write.  B_CACHE is never
4435  *	set if the buffer is invalid or otherwise uncacheable.
4436  *
4437  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
4438  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4439  *	in the biodone routine.
4440  */
4441 void
4442 bufdone(struct buf *bp)
4443 {
4444 	struct bufobj *dropobj;
4445 	void    (*biodone)(struct buf *);
4446 
4447 	buf_track(bp, __func__);
4448 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4449 	dropobj = NULL;
4450 
4451 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4452 	BUF_ASSERT_HELD(bp);
4453 
4454 	runningbufwakeup(bp);
4455 	if (bp->b_iocmd == BIO_WRITE)
4456 		dropobj = bp->b_bufobj;
4457 	/* call optional completion function if requested */
4458 	if (bp->b_iodone != NULL) {
4459 		biodone = bp->b_iodone;
4460 		bp->b_iodone = NULL;
4461 		(*biodone) (bp);
4462 		if (dropobj)
4463 			bufobj_wdrop(dropobj);
4464 		return;
4465 	}
4466 	if (bp->b_flags & B_VMIO) {
4467 		/*
4468 		 * Set B_CACHE if the op was a normal read and no error
4469 		 * occurred.  B_CACHE is set for writes in the b*write()
4470 		 * routines.
4471 		 */
4472 		if (bp->b_iocmd == BIO_READ &&
4473 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4474 		    !(bp->b_ioflags & BIO_ERROR))
4475 			bp->b_flags |= B_CACHE;
4476 		vfs_vmio_iodone(bp);
4477 	}
4478 	if (!LIST_EMPTY(&bp->b_dep))
4479 		buf_complete(bp);
4480 	if ((bp->b_flags & B_CKHASH) != 0) {
4481 		KASSERT(bp->b_iocmd == BIO_READ,
4482 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4483 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4484 		(*bp->b_ckhashcalc)(bp);
4485 	}
4486 	/*
4487 	 * For asynchronous completions, release the buffer now. The brelse
4488 	 * will do a wakeup there if necessary - so no need to do a wakeup
4489 	 * here in the async case. The sync case always needs to do a wakeup.
4490 	 */
4491 	if (bp->b_flags & B_ASYNC) {
4492 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4493 		    (bp->b_ioflags & BIO_ERROR))
4494 			brelse(bp);
4495 		else
4496 			bqrelse(bp);
4497 	} else
4498 		bdone(bp);
4499 	if (dropobj)
4500 		bufobj_wdrop(dropobj);
4501 }
4502 
4503 /*
4504  * This routine is called in lieu of iodone in the case of
4505  * incomplete I/O.  This keeps the busy status for pages
4506  * consistent.
4507  */
4508 void
4509 vfs_unbusy_pages(struct buf *bp)
4510 {
4511 	int i;
4512 	vm_object_t obj;
4513 	vm_page_t m;
4514 
4515 	runningbufwakeup(bp);
4516 	if (!(bp->b_flags & B_VMIO))
4517 		return;
4518 
4519 	obj = bp->b_bufobj->bo_object;
4520 	VM_OBJECT_WLOCK(obj);
4521 	for (i = 0; i < bp->b_npages; i++) {
4522 		m = bp->b_pages[i];
4523 		if (m == bogus_page) {
4524 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4525 			if (!m)
4526 				panic("vfs_unbusy_pages: page missing\n");
4527 			bp->b_pages[i] = m;
4528 			if (buf_mapped(bp)) {
4529 				BUF_CHECK_MAPPED(bp);
4530 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4531 				    bp->b_pages, bp->b_npages);
4532 			} else
4533 				BUF_CHECK_UNMAPPED(bp);
4534 		}
4535 		vm_page_sunbusy(m);
4536 	}
4537 	vm_object_pip_wakeupn(obj, bp->b_npages);
4538 	VM_OBJECT_WUNLOCK(obj);
4539 }
4540 
4541 /*
4542  * vfs_page_set_valid:
4543  *
4544  *	Set the valid bits in a page based on the supplied offset.   The
4545  *	range is restricted to the buffer's size.
4546  *
4547  *	This routine is typically called after a read completes.
4548  */
4549 static void
4550 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4551 {
4552 	vm_ooffset_t eoff;
4553 
4554 	/*
4555 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4556 	 * page boundary and eoff is not greater than the end of the buffer.
4557 	 * The end of the buffer, in this case, is our file EOF, not the
4558 	 * allocation size of the buffer.
4559 	 */
4560 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4561 	if (eoff > bp->b_offset + bp->b_bcount)
4562 		eoff = bp->b_offset + bp->b_bcount;
4563 
4564 	/*
4565 	 * Set valid range.  This is typically the entire buffer and thus the
4566 	 * entire page.
4567 	 */
4568 	if (eoff > off)
4569 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4570 }
4571 
4572 /*
4573  * vfs_page_set_validclean:
4574  *
4575  *	Set the valid bits and clear the dirty bits in a page based on the
4576  *	supplied offset.   The range is restricted to the buffer's size.
4577  */
4578 static void
4579 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4580 {
4581 	vm_ooffset_t soff, eoff;
4582 
4583 	/*
4584 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4585 	 * page boundary or cross the end of the buffer.  The end of the
4586 	 * buffer, in this case, is our file EOF, not the allocation size
4587 	 * of the buffer.
4588 	 */
4589 	soff = off;
4590 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4591 	if (eoff > bp->b_offset + bp->b_bcount)
4592 		eoff = bp->b_offset + bp->b_bcount;
4593 
4594 	/*
4595 	 * Set valid range.  This is typically the entire buffer and thus the
4596 	 * entire page.
4597 	 */
4598 	if (eoff > soff) {
4599 		vm_page_set_validclean(
4600 		    m,
4601 		   (vm_offset_t) (soff & PAGE_MASK),
4602 		   (vm_offset_t) (eoff - soff)
4603 		);
4604 	}
4605 }
4606 
4607 /*
4608  * Ensure that all buffer pages are not exclusive busied.  If any page is
4609  * exclusive busy, drain it.
4610  */
4611 void
4612 vfs_drain_busy_pages(struct buf *bp)
4613 {
4614 	vm_page_t m;
4615 	int i, last_busied;
4616 
4617 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4618 	last_busied = 0;
4619 	for (i = 0; i < bp->b_npages; i++) {
4620 		m = bp->b_pages[i];
4621 		if (vm_page_xbusied(m)) {
4622 			for (; last_busied < i; last_busied++)
4623 				vm_page_sbusy(bp->b_pages[last_busied]);
4624 			while (vm_page_xbusied(m)) {
4625 				vm_page_lock(m);
4626 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4627 				vm_page_busy_sleep(m, "vbpage", true);
4628 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4629 			}
4630 		}
4631 	}
4632 	for (i = 0; i < last_busied; i++)
4633 		vm_page_sunbusy(bp->b_pages[i]);
4634 }
4635 
4636 /*
4637  * This routine is called before a device strategy routine.
4638  * It is used to tell the VM system that paging I/O is in
4639  * progress, and treat the pages associated with the buffer
4640  * almost as being exclusive busy.  Also the object paging_in_progress
4641  * flag is handled to make sure that the object doesn't become
4642  * inconsistent.
4643  *
4644  * Since I/O has not been initiated yet, certain buffer flags
4645  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4646  * and should be ignored.
4647  */
4648 void
4649 vfs_busy_pages(struct buf *bp, int clear_modify)
4650 {
4651 	vm_object_t obj;
4652 	vm_ooffset_t foff;
4653 	vm_page_t m;
4654 	int i;
4655 	bool bogus;
4656 
4657 	if (!(bp->b_flags & B_VMIO))
4658 		return;
4659 
4660 	obj = bp->b_bufobj->bo_object;
4661 	foff = bp->b_offset;
4662 	KASSERT(bp->b_offset != NOOFFSET,
4663 	    ("vfs_busy_pages: no buffer offset"));
4664 	VM_OBJECT_WLOCK(obj);
4665 	vfs_drain_busy_pages(bp);
4666 	if (bp->b_bufsize != 0)
4667 		vfs_setdirty_locked_object(bp);
4668 	bogus = false;
4669 	for (i = 0; i < bp->b_npages; i++) {
4670 		m = bp->b_pages[i];
4671 
4672 		if ((bp->b_flags & B_CLUSTER) == 0) {
4673 			vm_object_pip_add(obj, 1);
4674 			vm_page_sbusy(m);
4675 		}
4676 		/*
4677 		 * When readying a buffer for a read ( i.e
4678 		 * clear_modify == 0 ), it is important to do
4679 		 * bogus_page replacement for valid pages in
4680 		 * partially instantiated buffers.  Partially
4681 		 * instantiated buffers can, in turn, occur when
4682 		 * reconstituting a buffer from its VM backing store
4683 		 * base.  We only have to do this if B_CACHE is
4684 		 * clear ( which causes the I/O to occur in the
4685 		 * first place ).  The replacement prevents the read
4686 		 * I/O from overwriting potentially dirty VM-backed
4687 		 * pages.  XXX bogus page replacement is, uh, bogus.
4688 		 * It may not work properly with small-block devices.
4689 		 * We need to find a better way.
4690 		 */
4691 		if (clear_modify) {
4692 			pmap_remove_write(m);
4693 			vfs_page_set_validclean(bp, foff, m);
4694 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4695 		    (bp->b_flags & B_CACHE) == 0) {
4696 			bp->b_pages[i] = bogus_page;
4697 			bogus = true;
4698 		}
4699 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4700 	}
4701 	VM_OBJECT_WUNLOCK(obj);
4702 	if (bogus && buf_mapped(bp)) {
4703 		BUF_CHECK_MAPPED(bp);
4704 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4705 		    bp->b_pages, bp->b_npages);
4706 	}
4707 }
4708 
4709 /*
4710  *	vfs_bio_set_valid:
4711  *
4712  *	Set the range within the buffer to valid.  The range is
4713  *	relative to the beginning of the buffer, b_offset.  Note that
4714  *	b_offset itself may be offset from the beginning of the first
4715  *	page.
4716  */
4717 void
4718 vfs_bio_set_valid(struct buf *bp, int base, int size)
4719 {
4720 	int i, n;
4721 	vm_page_t m;
4722 
4723 	if (!(bp->b_flags & B_VMIO))
4724 		return;
4725 
4726 	/*
4727 	 * Fixup base to be relative to beginning of first page.
4728 	 * Set initial n to be the maximum number of bytes in the
4729 	 * first page that can be validated.
4730 	 */
4731 	base += (bp->b_offset & PAGE_MASK);
4732 	n = PAGE_SIZE - (base & PAGE_MASK);
4733 
4734 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4735 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4736 		m = bp->b_pages[i];
4737 		if (n > size)
4738 			n = size;
4739 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4740 		base += n;
4741 		size -= n;
4742 		n = PAGE_SIZE;
4743 	}
4744 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4745 }
4746 
4747 /*
4748  *	vfs_bio_clrbuf:
4749  *
4750  *	If the specified buffer is a non-VMIO buffer, clear the entire
4751  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4752  *	validate only the previously invalid portions of the buffer.
4753  *	This routine essentially fakes an I/O, so we need to clear
4754  *	BIO_ERROR and B_INVAL.
4755  *
4756  *	Note that while we only theoretically need to clear through b_bcount,
4757  *	we go ahead and clear through b_bufsize.
4758  */
4759 void
4760 vfs_bio_clrbuf(struct buf *bp)
4761 {
4762 	int i, j, mask, sa, ea, slide;
4763 
4764 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4765 		clrbuf(bp);
4766 		return;
4767 	}
4768 	bp->b_flags &= ~B_INVAL;
4769 	bp->b_ioflags &= ~BIO_ERROR;
4770 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4771 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4772 	    (bp->b_offset & PAGE_MASK) == 0) {
4773 		if (bp->b_pages[0] == bogus_page)
4774 			goto unlock;
4775 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4776 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4777 		if ((bp->b_pages[0]->valid & mask) == mask)
4778 			goto unlock;
4779 		if ((bp->b_pages[0]->valid & mask) == 0) {
4780 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4781 			bp->b_pages[0]->valid |= mask;
4782 			goto unlock;
4783 		}
4784 	}
4785 	sa = bp->b_offset & PAGE_MASK;
4786 	slide = 0;
4787 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4788 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4789 		ea = slide & PAGE_MASK;
4790 		if (ea == 0)
4791 			ea = PAGE_SIZE;
4792 		if (bp->b_pages[i] == bogus_page)
4793 			continue;
4794 		j = sa / DEV_BSIZE;
4795 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4796 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4797 		if ((bp->b_pages[i]->valid & mask) == mask)
4798 			continue;
4799 		if ((bp->b_pages[i]->valid & mask) == 0)
4800 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4801 		else {
4802 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4803 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4804 					pmap_zero_page_area(bp->b_pages[i],
4805 					    sa, DEV_BSIZE);
4806 				}
4807 			}
4808 		}
4809 		bp->b_pages[i]->valid |= mask;
4810 	}
4811 unlock:
4812 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4813 	bp->b_resid = 0;
4814 }
4815 
4816 void
4817 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4818 {
4819 	vm_page_t m;
4820 	int i, n;
4821 
4822 	if (buf_mapped(bp)) {
4823 		BUF_CHECK_MAPPED(bp);
4824 		bzero(bp->b_data + base, size);
4825 	} else {
4826 		BUF_CHECK_UNMAPPED(bp);
4827 		n = PAGE_SIZE - (base & PAGE_MASK);
4828 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4829 			m = bp->b_pages[i];
4830 			if (n > size)
4831 				n = size;
4832 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4833 			base += n;
4834 			size -= n;
4835 			n = PAGE_SIZE;
4836 		}
4837 	}
4838 }
4839 
4840 /*
4841  * Update buffer flags based on I/O request parameters, optionally releasing the
4842  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4843  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4844  * I/O).  Otherwise the buffer is released to the cache.
4845  */
4846 static void
4847 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4848 {
4849 
4850 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4851 	    ("buf %p non-VMIO noreuse", bp));
4852 
4853 	if ((ioflag & IO_DIRECT) != 0)
4854 		bp->b_flags |= B_DIRECT;
4855 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4856 		bp->b_flags |= B_RELBUF;
4857 		if ((ioflag & IO_NOREUSE) != 0)
4858 			bp->b_flags |= B_NOREUSE;
4859 		if (release)
4860 			brelse(bp);
4861 	} else if (release)
4862 		bqrelse(bp);
4863 }
4864 
4865 void
4866 vfs_bio_brelse(struct buf *bp, int ioflag)
4867 {
4868 
4869 	b_io_dismiss(bp, ioflag, true);
4870 }
4871 
4872 void
4873 vfs_bio_set_flags(struct buf *bp, int ioflag)
4874 {
4875 
4876 	b_io_dismiss(bp, ioflag, false);
4877 }
4878 
4879 /*
4880  * vm_hold_load_pages and vm_hold_free_pages get pages into
4881  * a buffers address space.  The pages are anonymous and are
4882  * not associated with a file object.
4883  */
4884 static void
4885 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4886 {
4887 	vm_offset_t pg;
4888 	vm_page_t p;
4889 	int index;
4890 
4891 	BUF_CHECK_MAPPED(bp);
4892 
4893 	to = round_page(to);
4894 	from = round_page(from);
4895 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4896 
4897 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4898 		/*
4899 		 * note: must allocate system pages since blocking here
4900 		 * could interfere with paging I/O, no matter which
4901 		 * process we are.
4902 		 */
4903 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4904 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4905 		    VM_ALLOC_WAITOK);
4906 		pmap_qenter(pg, &p, 1);
4907 		bp->b_pages[index] = p;
4908 	}
4909 	bp->b_npages = index;
4910 }
4911 
4912 /* Return pages associated with this buf to the vm system */
4913 static void
4914 vm_hold_free_pages(struct buf *bp, int newbsize)
4915 {
4916 	vm_offset_t from;
4917 	vm_page_t p;
4918 	int index, newnpages;
4919 
4920 	BUF_CHECK_MAPPED(bp);
4921 
4922 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4923 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4924 	if (bp->b_npages > newnpages)
4925 		pmap_qremove(from, bp->b_npages - newnpages);
4926 	for (index = newnpages; index < bp->b_npages; index++) {
4927 		p = bp->b_pages[index];
4928 		bp->b_pages[index] = NULL;
4929 		p->wire_count--;
4930 		vm_page_free(p);
4931 	}
4932 	vm_wire_sub(bp->b_npages - newnpages);
4933 	bp->b_npages = newnpages;
4934 }
4935 
4936 /*
4937  * Map an IO request into kernel virtual address space.
4938  *
4939  * All requests are (re)mapped into kernel VA space.
4940  * Notice that we use b_bufsize for the size of the buffer
4941  * to be mapped.  b_bcount might be modified by the driver.
4942  *
4943  * Note that even if the caller determines that the address space should
4944  * be valid, a race or a smaller-file mapped into a larger space may
4945  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4946  * check the return value.
4947  *
4948  * This function only works with pager buffers.
4949  */
4950 int
4951 vmapbuf(struct buf *bp, int mapbuf)
4952 {
4953 	vm_prot_t prot;
4954 	int pidx;
4955 
4956 	if (bp->b_bufsize < 0)
4957 		return (-1);
4958 	prot = VM_PROT_READ;
4959 	if (bp->b_iocmd == BIO_READ)
4960 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4961 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4962 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4963 	    btoc(MAXPHYS))) < 0)
4964 		return (-1);
4965 	bp->b_npages = pidx;
4966 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4967 	if (mapbuf || !unmapped_buf_allowed) {
4968 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4969 		bp->b_data = bp->b_kvabase + bp->b_offset;
4970 	} else
4971 		bp->b_data = unmapped_buf;
4972 	return(0);
4973 }
4974 
4975 /*
4976  * Free the io map PTEs associated with this IO operation.
4977  * We also invalidate the TLB entries and restore the original b_addr.
4978  *
4979  * This function only works with pager buffers.
4980  */
4981 void
4982 vunmapbuf(struct buf *bp)
4983 {
4984 	int npages;
4985 
4986 	npages = bp->b_npages;
4987 	if (buf_mapped(bp))
4988 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4989 	vm_page_unhold_pages(bp->b_pages, npages);
4990 
4991 	bp->b_data = unmapped_buf;
4992 }
4993 
4994 void
4995 bdone(struct buf *bp)
4996 {
4997 	struct mtx *mtxp;
4998 
4999 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5000 	mtx_lock(mtxp);
5001 	bp->b_flags |= B_DONE;
5002 	wakeup(bp);
5003 	mtx_unlock(mtxp);
5004 }
5005 
5006 void
5007 bwait(struct buf *bp, u_char pri, const char *wchan)
5008 {
5009 	struct mtx *mtxp;
5010 
5011 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5012 	mtx_lock(mtxp);
5013 	while ((bp->b_flags & B_DONE) == 0)
5014 		msleep(bp, mtxp, pri, wchan, 0);
5015 	mtx_unlock(mtxp);
5016 }
5017 
5018 int
5019 bufsync(struct bufobj *bo, int waitfor)
5020 {
5021 
5022 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5023 }
5024 
5025 void
5026 bufstrategy(struct bufobj *bo, struct buf *bp)
5027 {
5028 	int i __unused;
5029 	struct vnode *vp;
5030 
5031 	vp = bp->b_vp;
5032 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5033 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5034 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5035 	i = VOP_STRATEGY(vp, bp);
5036 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5037 }
5038 
5039 /*
5040  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
5041  */
5042 void
5043 bufobj_init(struct bufobj *bo, void *private)
5044 {
5045 	static volatile int bufobj_cleanq;
5046 
5047         bo->bo_domain =
5048             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5049         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5050         bo->bo_private = private;
5051         TAILQ_INIT(&bo->bo_clean.bv_hd);
5052         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5053 }
5054 
5055 void
5056 bufobj_wrefl(struct bufobj *bo)
5057 {
5058 
5059 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5060 	ASSERT_BO_WLOCKED(bo);
5061 	bo->bo_numoutput++;
5062 }
5063 
5064 void
5065 bufobj_wref(struct bufobj *bo)
5066 {
5067 
5068 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5069 	BO_LOCK(bo);
5070 	bo->bo_numoutput++;
5071 	BO_UNLOCK(bo);
5072 }
5073 
5074 void
5075 bufobj_wdrop(struct bufobj *bo)
5076 {
5077 
5078 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5079 	BO_LOCK(bo);
5080 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5081 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5082 		bo->bo_flag &= ~BO_WWAIT;
5083 		wakeup(&bo->bo_numoutput);
5084 	}
5085 	BO_UNLOCK(bo);
5086 }
5087 
5088 int
5089 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5090 {
5091 	int error;
5092 
5093 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5094 	ASSERT_BO_WLOCKED(bo);
5095 	error = 0;
5096 	while (bo->bo_numoutput) {
5097 		bo->bo_flag |= BO_WWAIT;
5098 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5099 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5100 		if (error)
5101 			break;
5102 	}
5103 	return (error);
5104 }
5105 
5106 /*
5107  * Set bio_data or bio_ma for struct bio from the struct buf.
5108  */
5109 void
5110 bdata2bio(struct buf *bp, struct bio *bip)
5111 {
5112 
5113 	if (!buf_mapped(bp)) {
5114 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5115 		bip->bio_ma = bp->b_pages;
5116 		bip->bio_ma_n = bp->b_npages;
5117 		bip->bio_data = unmapped_buf;
5118 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5119 		bip->bio_flags |= BIO_UNMAPPED;
5120 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5121 		    PAGE_SIZE == bp->b_npages,
5122 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5123 		    (long long)bip->bio_length, bip->bio_ma_n));
5124 	} else {
5125 		bip->bio_data = bp->b_data;
5126 		bip->bio_ma = NULL;
5127 	}
5128 }
5129 
5130 /*
5131  * The MIPS pmap code currently doesn't handle aliased pages.
5132  * The VIPT caches may not handle page aliasing themselves, leading
5133  * to data corruption.
5134  *
5135  * As such, this code makes a system extremely unhappy if said
5136  * system doesn't support unaliasing the above situation in hardware.
5137  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5138  * this feature at build time, so it has to be handled in software.
5139  *
5140  * Once the MIPS pmap/cache code grows to support this function on
5141  * earlier chips, it should be flipped back off.
5142  */
5143 #ifdef	__mips__
5144 static int buf_pager_relbuf = 1;
5145 #else
5146 static int buf_pager_relbuf = 0;
5147 #endif
5148 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5149     &buf_pager_relbuf, 0,
5150     "Make buffer pager release buffers after reading");
5151 
5152 /*
5153  * The buffer pager.  It uses buffer reads to validate pages.
5154  *
5155  * In contrast to the generic local pager from vm/vnode_pager.c, this
5156  * pager correctly and easily handles volumes where the underlying
5157  * device block size is greater than the machine page size.  The
5158  * buffer cache transparently extends the requested page run to be
5159  * aligned at the block boundary, and does the necessary bogus page
5160  * replacements in the addends to avoid obliterating already valid
5161  * pages.
5162  *
5163  * The only non-trivial issue is that the exclusive busy state for
5164  * pages, which is assumed by the vm_pager_getpages() interface, is
5165  * incompatible with the VMIO buffer cache's desire to share-busy the
5166  * pages.  This function performs a trivial downgrade of the pages'
5167  * state before reading buffers, and a less trivial upgrade from the
5168  * shared-busy to excl-busy state after the read.
5169  */
5170 int
5171 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5172     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5173     vbg_get_blksize_t get_blksize)
5174 {
5175 	vm_page_t m;
5176 	vm_object_t object;
5177 	struct buf *bp;
5178 	struct mount *mp;
5179 	daddr_t lbn, lbnp;
5180 	vm_ooffset_t la, lb, poff, poffe;
5181 	long bsize;
5182 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5183 	bool redo, lpart;
5184 
5185 	object = vp->v_object;
5186 	mp = vp->v_mount;
5187 	error = 0;
5188 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5189 	if (la >= object->un_pager.vnp.vnp_size)
5190 		return (VM_PAGER_BAD);
5191 
5192 	/*
5193 	 * Change the meaning of la from where the last requested page starts
5194 	 * to where it ends, because that's the end of the requested region
5195 	 * and the start of the potential read-ahead region.
5196 	 */
5197 	la += PAGE_SIZE;
5198 	lpart = la > object->un_pager.vnp.vnp_size;
5199 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5200 
5201 	/*
5202 	 * Calculate read-ahead, behind and total pages.
5203 	 */
5204 	pgsin = count;
5205 	lb = IDX_TO_OFF(ma[0]->pindex);
5206 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5207 	pgsin += pgsin_b;
5208 	if (rbehind != NULL)
5209 		*rbehind = pgsin_b;
5210 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5211 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5212 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5213 		    PAGE_SIZE) - la);
5214 	pgsin += pgsin_a;
5215 	if (rahead != NULL)
5216 		*rahead = pgsin_a;
5217 	VM_CNT_INC(v_vnodein);
5218 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5219 
5220 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5221 	    != 0) ? GB_UNMAPPED : 0;
5222 	VM_OBJECT_WLOCK(object);
5223 again:
5224 	for (i = 0; i < count; i++)
5225 		vm_page_busy_downgrade(ma[i]);
5226 	VM_OBJECT_WUNLOCK(object);
5227 
5228 	lbnp = -1;
5229 	for (i = 0; i < count; i++) {
5230 		m = ma[i];
5231 
5232 		/*
5233 		 * Pages are shared busy and the object lock is not
5234 		 * owned, which together allow for the pages'
5235 		 * invalidation.  The racy test for validity avoids
5236 		 * useless creation of the buffer for the most typical
5237 		 * case when invalidation is not used in redo or for
5238 		 * parallel read.  The shared->excl upgrade loop at
5239 		 * the end of the function catches the race in a
5240 		 * reliable way (protected by the object lock).
5241 		 */
5242 		if (m->valid == VM_PAGE_BITS_ALL)
5243 			continue;
5244 
5245 		poff = IDX_TO_OFF(m->pindex);
5246 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5247 		for (; poff < poffe; poff += bsize) {
5248 			lbn = get_lblkno(vp, poff);
5249 			if (lbn == lbnp)
5250 				goto next_page;
5251 			lbnp = lbn;
5252 
5253 			bsize = get_blksize(vp, lbn);
5254 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5255 			    br_flags, &bp);
5256 			if (error != 0)
5257 				goto end_pages;
5258 			if (LIST_EMPTY(&bp->b_dep)) {
5259 				/*
5260 				 * Invalidation clears m->valid, but
5261 				 * may leave B_CACHE flag if the
5262 				 * buffer existed at the invalidation
5263 				 * time.  In this case, recycle the
5264 				 * buffer to do real read on next
5265 				 * bread() after redo.
5266 				 *
5267 				 * Otherwise B_RELBUF is not strictly
5268 				 * necessary, enable to reduce buf
5269 				 * cache pressure.
5270 				 */
5271 				if (buf_pager_relbuf ||
5272 				    m->valid != VM_PAGE_BITS_ALL)
5273 					bp->b_flags |= B_RELBUF;
5274 
5275 				bp->b_flags &= ~B_NOCACHE;
5276 				brelse(bp);
5277 			} else {
5278 				bqrelse(bp);
5279 			}
5280 		}
5281 		KASSERT(1 /* racy, enable for debugging */ ||
5282 		    m->valid == VM_PAGE_BITS_ALL || i == count - 1,
5283 		    ("buf %d %p invalid", i, m));
5284 		if (i == count - 1 && lpart) {
5285 			VM_OBJECT_WLOCK(object);
5286 			if (m->valid != 0 &&
5287 			    m->valid != VM_PAGE_BITS_ALL)
5288 				vm_page_zero_invalid(m, TRUE);
5289 			VM_OBJECT_WUNLOCK(object);
5290 		}
5291 next_page:;
5292 	}
5293 end_pages:
5294 
5295 	VM_OBJECT_WLOCK(object);
5296 	redo = false;
5297 	for (i = 0; i < count; i++) {
5298 		vm_page_sunbusy(ma[i]);
5299 		ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
5300 
5301 		/*
5302 		 * Since the pages were only sbusy while neither the
5303 		 * buffer nor the object lock was held by us, or
5304 		 * reallocated while vm_page_grab() slept for busy
5305 		 * relinguish, they could have been invalidated.
5306 		 * Recheck the valid bits and re-read as needed.
5307 		 *
5308 		 * Note that the last page is made fully valid in the
5309 		 * read loop, and partial validity for the page at
5310 		 * index count - 1 could mean that the page was
5311 		 * invalidated or removed, so we must restart for
5312 		 * safety as well.
5313 		 */
5314 		if (ma[i]->valid != VM_PAGE_BITS_ALL)
5315 			redo = true;
5316 	}
5317 	if (redo && error == 0)
5318 		goto again;
5319 	VM_OBJECT_WUNLOCK(object);
5320 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5321 }
5322 
5323 #include "opt_ddb.h"
5324 #ifdef DDB
5325 #include <ddb/ddb.h>
5326 
5327 /* DDB command to show buffer data */
5328 DB_SHOW_COMMAND(buffer, db_show_buffer)
5329 {
5330 	/* get args */
5331 	struct buf *bp = (struct buf *)addr;
5332 #ifdef FULL_BUF_TRACKING
5333 	uint32_t i, j;
5334 #endif
5335 
5336 	if (!have_addr) {
5337 		db_printf("usage: show buffer <addr>\n");
5338 		return;
5339 	}
5340 
5341 	db_printf("buf at %p\n", bp);
5342 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
5343 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
5344 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
5345 	db_printf(
5346 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5347 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
5348 	    "b_dep = %p\n",
5349 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5350 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5351 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
5352 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5353 	    bp->b_kvabase, bp->b_kvasize);
5354 	if (bp->b_npages) {
5355 		int i;
5356 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5357 		for (i = 0; i < bp->b_npages; i++) {
5358 			vm_page_t m;
5359 			m = bp->b_pages[i];
5360 			if (m != NULL)
5361 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5362 				    (u_long)m->pindex,
5363 				    (u_long)VM_PAGE_TO_PHYS(m));
5364 			else
5365 				db_printf("( ??? )");
5366 			if ((i + 1) < bp->b_npages)
5367 				db_printf(",");
5368 		}
5369 		db_printf("\n");
5370 	}
5371 	BUF_LOCKPRINTINFO(bp);
5372 #if defined(FULL_BUF_TRACKING)
5373 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5374 
5375 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5376 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5377 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5378 			continue;
5379 		db_printf(" %2u: %s\n", j,
5380 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5381 	}
5382 #elif defined(BUF_TRACKING)
5383 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5384 #endif
5385 	db_printf(" ");
5386 }
5387 
5388 DB_SHOW_COMMAND(bufqueues, bufqueues)
5389 {
5390 	struct bufdomain *bd;
5391 	struct buf *bp;
5392 	long total;
5393 	int i, j, cnt;
5394 
5395 	db_printf("bqempty: %d\n", bqempty.bq_len);
5396 
5397 	for (i = 0; i < buf_domains; i++) {
5398 		bd = &bdomain[i];
5399 		db_printf("Buf domain %d\n", i);
5400 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5401 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5402 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5403 		db_printf("\n");
5404 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5405 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5406 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5407 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5408 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5409 		db_printf("\n");
5410 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5411 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5412 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5413 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5414 		db_printf("\n");
5415 		total = 0;
5416 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5417 			total += bp->b_bufsize;
5418 		db_printf("\tcleanq count\t%d (%ld)\n",
5419 		    bd->bd_cleanq->bq_len, total);
5420 		total = 0;
5421 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5422 			total += bp->b_bufsize;
5423 		db_printf("\tdirtyq count\t%d (%ld)\n",
5424 		    bd->bd_dirtyq.bq_len, total);
5425 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5426 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5427 		db_printf("\tCPU ");
5428 		for (j = 0; j <= mp_maxid; j++)
5429 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5430 		db_printf("\n");
5431 		cnt = 0;
5432 		total = 0;
5433 		for (j = 0; j < nbuf; j++)
5434 			if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5435 				cnt++;
5436 				total += buf[j].b_bufsize;
5437 			}
5438 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5439 		cnt = 0;
5440 		total = 0;
5441 		for (j = 0; j < nbuf; j++)
5442 			if (buf[j].b_domain == i) {
5443 				cnt++;
5444 				total += buf[j].b_bufsize;
5445 			}
5446 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5447 	}
5448 }
5449 
5450 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5451 {
5452 	struct buf *bp;
5453 	int i;
5454 
5455 	for (i = 0; i < nbuf; i++) {
5456 		bp = &buf[i];
5457 		if (BUF_ISLOCKED(bp)) {
5458 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5459 			db_printf("\n");
5460 			if (db_pager_quit)
5461 				break;
5462 		}
5463 	}
5464 }
5465 
5466 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5467 {
5468 	struct vnode *vp;
5469 	struct buf *bp;
5470 
5471 	if (!have_addr) {
5472 		db_printf("usage: show vnodebufs <addr>\n");
5473 		return;
5474 	}
5475 	vp = (struct vnode *)addr;
5476 	db_printf("Clean buffers:\n");
5477 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5478 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5479 		db_printf("\n");
5480 	}
5481 	db_printf("Dirty buffers:\n");
5482 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5483 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5484 		db_printf("\n");
5485 	}
5486 }
5487 
5488 DB_COMMAND(countfreebufs, db_coundfreebufs)
5489 {
5490 	struct buf *bp;
5491 	int i, used = 0, nfree = 0;
5492 
5493 	if (have_addr) {
5494 		db_printf("usage: countfreebufs\n");
5495 		return;
5496 	}
5497 
5498 	for (i = 0; i < nbuf; i++) {
5499 		bp = &buf[i];
5500 		if (bp->b_qindex == QUEUE_EMPTY)
5501 			nfree++;
5502 		else
5503 			used++;
5504 	}
5505 
5506 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5507 	    nfree + used);
5508 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5509 }
5510 #endif /* DDB */
5511