xref: /freebsd/sys/kern/vfs_bio.c (revision a0b9e2e854027e6ff61fb075a1309dbc71c42b54)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2004 Poul-Henning Kamp
5  * Copyright (c) 1994,1997 John S. Dyson
6  * Copyright (c) 2013 The FreeBSD Foundation
7  * All rights reserved.
8  *
9  * Portions of this software were developed by Konstantin Belousov
10  * under sponsorship from the FreeBSD Foundation.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * this file contains a new buffer I/O scheme implementing a coherent
36  * VM object and buffer cache scheme.  Pains have been taken to make
37  * sure that the performance degradation associated with schemes such
38  * as this is not realized.
39  *
40  * Author:  John S. Dyson
41  * Significant help during the development and debugging phases
42  * had been provided by David Greenman, also of the FreeBSD core team.
43  *
44  * see man buf(9) for more info.
45  */
46 
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/bio.h>
53 #include <sys/bitset.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/buf.h>
57 #include <sys/devicestat.h>
58 #include <sys/eventhandler.h>
59 #include <sys/fail.h>
60 #include <sys/ktr.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/malloc.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/proc.h>
69 #include <sys/racct.h>
70 #include <sys/refcount.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/smp.h>
74 #include <sys/sysctl.h>
75 #include <sys/syscallsubr.h>
76 #include <sys/vmem.h>
77 #include <sys/vmmeter.h>
78 #include <sys/vnode.h>
79 #include <sys/watchdog.h>
80 #include <geom/geom.h>
81 #include <vm/vm.h>
82 #include <vm/vm_param.h>
83 #include <vm/vm_kern.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_pager.h>
88 #include <vm/vm_extern.h>
89 #include <vm/vm_map.h>
90 #include <vm/swap_pager.h>
91 
92 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
93 
94 struct	bio_ops bioops;		/* I/O operation notification */
95 
96 struct	buf_ops buf_ops_bio = {
97 	.bop_name	=	"buf_ops_bio",
98 	.bop_write	=	bufwrite,
99 	.bop_strategy	=	bufstrategy,
100 	.bop_sync	=	bufsync,
101 	.bop_bdflush	=	bufbdflush,
102 };
103 
104 struct bufqueue {
105 	struct mtx_padalign	bq_lock;
106 	TAILQ_HEAD(, buf)	bq_queue;
107 	uint8_t			bq_index;
108 	uint16_t		bq_subqueue;
109 	int			bq_len;
110 } __aligned(CACHE_LINE_SIZE);
111 
112 #define	BQ_LOCKPTR(bq)		(&(bq)->bq_lock)
113 #define	BQ_LOCK(bq)		mtx_lock(BQ_LOCKPTR((bq)))
114 #define	BQ_UNLOCK(bq)		mtx_unlock(BQ_LOCKPTR((bq)))
115 #define	BQ_ASSERT_LOCKED(bq)	mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
116 
117 struct bufdomain {
118 	struct bufqueue	bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
119 	struct bufqueue bd_dirtyq;
120 	struct bufqueue	*bd_cleanq;
121 	struct mtx_padalign bd_run_lock;
122 	/* Constants */
123 	long		bd_maxbufspace;
124 	long		bd_hibufspace;
125 	long 		bd_lobufspace;
126 	long 		bd_bufspacethresh;
127 	int		bd_hifreebuffers;
128 	int		bd_lofreebuffers;
129 	int		bd_hidirtybuffers;
130 	int		bd_lodirtybuffers;
131 	int		bd_dirtybufthresh;
132 	int		bd_lim;
133 	/* atomics */
134 	int		bd_wanted;
135 	int __aligned(CACHE_LINE_SIZE)	bd_numdirtybuffers;
136 	int __aligned(CACHE_LINE_SIZE)	bd_running;
137 	long __aligned(CACHE_LINE_SIZE) bd_bufspace;
138 	int __aligned(CACHE_LINE_SIZE)	bd_freebuffers;
139 } __aligned(CACHE_LINE_SIZE);
140 
141 #define	BD_LOCKPTR(bd)		(&(bd)->bd_cleanq->bq_lock)
142 #define	BD_LOCK(bd)		mtx_lock(BD_LOCKPTR((bd)))
143 #define	BD_UNLOCK(bd)		mtx_unlock(BD_LOCKPTR((bd)))
144 #define	BD_ASSERT_LOCKED(bd)	mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
145 #define	BD_RUN_LOCKPTR(bd)	(&(bd)->bd_run_lock)
146 #define	BD_RUN_LOCK(bd)		mtx_lock(BD_RUN_LOCKPTR((bd)))
147 #define	BD_RUN_UNLOCK(bd)	mtx_unlock(BD_RUN_LOCKPTR((bd)))
148 #define	BD_DOMAIN(bd)		(bd - bdomain)
149 
150 static char *buf;		/* buffer header pool */
151 static struct buf *
152 nbufp(unsigned i)
153 {
154 	return ((struct buf *)(buf + (sizeof(struct buf) +
155 	    sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
156 }
157 
158 caddr_t __read_mostly unmapped_buf;
159 
160 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
161 struct proc *bufdaemonproc;
162 
163 static void vm_hold_free_pages(struct buf *bp, int newbsize);
164 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
165 		vm_offset_t to);
166 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
167 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
168 		vm_page_t m);
169 static void vfs_clean_pages_dirty_buf(struct buf *bp);
170 static void vfs_setdirty_range(struct buf *bp);
171 static void vfs_vmio_invalidate(struct buf *bp);
172 static void vfs_vmio_truncate(struct buf *bp, int npages);
173 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
174 static int vfs_bio_clcheck(struct vnode *vp, int size,
175 		daddr_t lblkno, daddr_t blkno);
176 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
177 		void (*)(struct buf *));
178 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
179 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
180 static void buf_daemon(void);
181 static __inline void bd_wakeup(void);
182 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
183 static void bufkva_reclaim(vmem_t *, int);
184 static void bufkva_free(struct buf *);
185 static int buf_import(void *, void **, int, int, int);
186 static void buf_release(void *, void **, int);
187 static void maxbcachebuf_adjust(void);
188 static inline struct bufdomain *bufdomain(struct buf *);
189 static void bq_remove(struct bufqueue *bq, struct buf *bp);
190 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
191 static int buf_recycle(struct bufdomain *, bool kva);
192 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
193 	    const char *lockname);
194 static void bd_init(struct bufdomain *bd);
195 static int bd_flushall(struct bufdomain *bd);
196 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
197 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
198 
199 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
200 int vmiodirenable = TRUE;
201 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
202     "Use the VM system for directory writes");
203 long runningbufspace;
204 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
205     "Amount of presently outstanding async buffer io");
206 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
207     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
208 static counter_u64_t bufkvaspace;
209 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
210     "Kernel virtual memory used for buffers");
211 static long maxbufspace;
212 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
213     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
214     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
215     "Maximum allowed value of bufspace (including metadata)");
216 static long bufmallocspace;
217 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
218     "Amount of malloced memory for buffers");
219 static long maxbufmallocspace;
220 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
221     0, "Maximum amount of malloced memory for buffers");
222 static long lobufspace;
223 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
224     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
225     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
226     "Minimum amount of buffers we want to have");
227 long hibufspace;
228 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
229     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
230     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
231     "Maximum allowed value of bufspace (excluding metadata)");
232 long bufspacethresh;
233 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
234     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
235     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
236     "Bufspace consumed before waking the daemon to free some");
237 static counter_u64_t buffreekvacnt;
238 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
239     "Number of times we have freed the KVA space from some buffer");
240 static counter_u64_t bufdefragcnt;
241 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
242     "Number of times we have had to repeat buffer allocation to defragment");
243 static long lorunningspace;
244 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
245     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
246     "Minimum preferred space used for in-progress I/O");
247 static long hirunningspace;
248 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
249     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
250     "Maximum amount of space to use for in-progress I/O");
251 int dirtybufferflushes;
252 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
253     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
254 int bdwriteskip;
255 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
256     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
257 int altbufferflushes;
258 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
259     &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
260 static int recursiveflushes;
261 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
262     &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
263 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
264 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
265     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
266     "Number of buffers that are dirty (has unwritten changes) at the moment");
267 static int lodirtybuffers;
268 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
269     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
270     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
271     "How many buffers we want to have free before bufdaemon can sleep");
272 static int hidirtybuffers;
273 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
274     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
275     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
276     "When the number of dirty buffers is considered severe");
277 int dirtybufthresh;
278 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
279     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
280     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
281     "Number of bdwrite to bawrite conversions to clear dirty buffers");
282 static int numfreebuffers;
283 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
284     "Number of free buffers");
285 static int lofreebuffers;
286 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
287     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
288     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
289    "Target number of free buffers");
290 static int hifreebuffers;
291 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
292     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
293     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
294    "Threshold for clean buffer recycling");
295 static counter_u64_t getnewbufcalls;
296 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
297    &getnewbufcalls, "Number of calls to getnewbuf");
298 static counter_u64_t getnewbufrestarts;
299 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
300     &getnewbufrestarts,
301     "Number of times getnewbuf has had to restart a buffer acquisition");
302 static counter_u64_t mappingrestarts;
303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
304     &mappingrestarts,
305     "Number of times getblk has had to restart a buffer mapping for "
306     "unmapped buffer");
307 static counter_u64_t numbufallocfails;
308 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
309     &numbufallocfails, "Number of times buffer allocations failed");
310 static int flushbufqtarget = 100;
311 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
312     "Amount of work to do in flushbufqueues when helping bufdaemon");
313 static counter_u64_t notbufdflushes;
314 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
315     "Number of dirty buffer flushes done by the bufdaemon helpers");
316 static long barrierwrites;
317 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
318     &barrierwrites, 0, "Number of barrier writes");
319 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
320     &unmapped_buf_allowed, 0,
321     "Permit the use of the unmapped i/o");
322 int maxbcachebuf = MAXBCACHEBUF;
323 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
324     "Maximum size of a buffer cache block");
325 
326 /*
327  * This lock synchronizes access to bd_request.
328  */
329 static struct mtx_padalign __exclusive_cache_line bdlock;
330 
331 /*
332  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
333  * waitrunningbufspace().
334  */
335 static struct mtx_padalign __exclusive_cache_line rbreqlock;
336 
337 /*
338  * Lock that protects bdirtywait.
339  */
340 static struct mtx_padalign __exclusive_cache_line bdirtylock;
341 
342 /*
343  * Wakeup point for bufdaemon, as well as indicator of whether it is already
344  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
345  * is idling.
346  */
347 static int bd_request;
348 
349 /*
350  * Request for the buf daemon to write more buffers than is indicated by
351  * lodirtybuf.  This may be necessary to push out excess dependencies or
352  * defragment the address space where a simple count of the number of dirty
353  * buffers is insufficient to characterize the demand for flushing them.
354  */
355 static int bd_speedupreq;
356 
357 /*
358  * Synchronization (sleep/wakeup) variable for active buffer space requests.
359  * Set when wait starts, cleared prior to wakeup().
360  * Used in runningbufwakeup() and waitrunningbufspace().
361  */
362 static int runningbufreq;
363 
364 /*
365  * Synchronization for bwillwrite() waiters.
366  */
367 static int bdirtywait;
368 
369 /*
370  * Definitions for the buffer free lists.
371  */
372 #define QUEUE_NONE	0	/* on no queue */
373 #define QUEUE_EMPTY	1	/* empty buffer headers */
374 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
375 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
376 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
377 
378 /* Maximum number of buffer domains. */
379 #define	BUF_DOMAINS	8
380 
381 struct bufdomainset bdlodirty;		/* Domains > lodirty */
382 struct bufdomainset bdhidirty;		/* Domains > hidirty */
383 
384 /* Configured number of clean queues. */
385 static int __read_mostly buf_domains;
386 
387 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
388 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
389 struct bufqueue __exclusive_cache_line bqempty;
390 
391 /*
392  * per-cpu empty buffer cache.
393  */
394 uma_zone_t buf_zone;
395 
396 /*
397  * Single global constant for BUF_WMESG, to avoid getting multiple references.
398  * buf_wmesg is referred from macros.
399  */
400 const char *buf_wmesg = BUF_WMESG;
401 
402 static int
403 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
404 {
405 	long value;
406 	int error;
407 
408 	value = *(long *)arg1;
409 	error = sysctl_handle_long(oidp, &value, 0, req);
410 	if (error != 0 || req->newptr == NULL)
411 		return (error);
412 	mtx_lock(&rbreqlock);
413 	if (arg1 == &hirunningspace) {
414 		if (value < lorunningspace)
415 			error = EINVAL;
416 		else
417 			hirunningspace = value;
418 	} else {
419 		KASSERT(arg1 == &lorunningspace,
420 		    ("%s: unknown arg1", __func__));
421 		if (value > hirunningspace)
422 			error = EINVAL;
423 		else
424 			lorunningspace = value;
425 	}
426 	mtx_unlock(&rbreqlock);
427 	return (error);
428 }
429 
430 static int
431 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
432 {
433 	int error;
434 	int value;
435 	int i;
436 
437 	value = *(int *)arg1;
438 	error = sysctl_handle_int(oidp, &value, 0, req);
439 	if (error != 0 || req->newptr == NULL)
440 		return (error);
441 	*(int *)arg1 = value;
442 	for (i = 0; i < buf_domains; i++)
443 		*(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
444 		    value / buf_domains;
445 
446 	return (error);
447 }
448 
449 static int
450 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
451 {
452 	long value;
453 	int error;
454 	int i;
455 
456 	value = *(long *)arg1;
457 	error = sysctl_handle_long(oidp, &value, 0, req);
458 	if (error != 0 || req->newptr == NULL)
459 		return (error);
460 	*(long *)arg1 = value;
461 	for (i = 0; i < buf_domains; i++)
462 		*(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
463 		    value / buf_domains;
464 
465 	return (error);
466 }
467 
468 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
469     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
470 static int
471 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
472 {
473 	long lvalue;
474 	int ivalue;
475 	int i;
476 
477 	lvalue = 0;
478 	for (i = 0; i < buf_domains; i++)
479 		lvalue += bdomain[i].bd_bufspace;
480 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
481 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
482 	if (lvalue > INT_MAX)
483 		/* On overflow, still write out a long to trigger ENOMEM. */
484 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
485 	ivalue = lvalue;
486 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
487 }
488 #else
489 static int
490 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
491 {
492 	long lvalue;
493 	int i;
494 
495 	lvalue = 0;
496 	for (i = 0; i < buf_domains; i++)
497 		lvalue += bdomain[i].bd_bufspace;
498 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
499 }
500 #endif
501 
502 static int
503 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
504 {
505 	int value;
506 	int i;
507 
508 	value = 0;
509 	for (i = 0; i < buf_domains; i++)
510 		value += bdomain[i].bd_numdirtybuffers;
511 	return (sysctl_handle_int(oidp, &value, 0, req));
512 }
513 
514 /*
515  *	bdirtywakeup:
516  *
517  *	Wakeup any bwillwrite() waiters.
518  */
519 static void
520 bdirtywakeup(void)
521 {
522 	mtx_lock(&bdirtylock);
523 	if (bdirtywait) {
524 		bdirtywait = 0;
525 		wakeup(&bdirtywait);
526 	}
527 	mtx_unlock(&bdirtylock);
528 }
529 
530 /*
531  *	bd_clear:
532  *
533  *	Clear a domain from the appropriate bitsets when dirtybuffers
534  *	is decremented.
535  */
536 static void
537 bd_clear(struct bufdomain *bd)
538 {
539 
540 	mtx_lock(&bdirtylock);
541 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
542 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
543 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
544 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
545 	mtx_unlock(&bdirtylock);
546 }
547 
548 /*
549  *	bd_set:
550  *
551  *	Set a domain in the appropriate bitsets when dirtybuffers
552  *	is incremented.
553  */
554 static void
555 bd_set(struct bufdomain *bd)
556 {
557 
558 	mtx_lock(&bdirtylock);
559 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
560 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
561 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
562 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
563 	mtx_unlock(&bdirtylock);
564 }
565 
566 /*
567  *	bdirtysub:
568  *
569  *	Decrement the numdirtybuffers count by one and wakeup any
570  *	threads blocked in bwillwrite().
571  */
572 static void
573 bdirtysub(struct buf *bp)
574 {
575 	struct bufdomain *bd;
576 	int num;
577 
578 	bd = bufdomain(bp);
579 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
580 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
581 		bdirtywakeup();
582 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
583 		bd_clear(bd);
584 }
585 
586 /*
587  *	bdirtyadd:
588  *
589  *	Increment the numdirtybuffers count by one and wakeup the buf
590  *	daemon if needed.
591  */
592 static void
593 bdirtyadd(struct buf *bp)
594 {
595 	struct bufdomain *bd;
596 	int num;
597 
598 	/*
599 	 * Only do the wakeup once as we cross the boundary.  The
600 	 * buf daemon will keep running until the condition clears.
601 	 */
602 	bd = bufdomain(bp);
603 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
604 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
605 		bd_wakeup();
606 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
607 		bd_set(bd);
608 }
609 
610 /*
611  *	bufspace_daemon_wakeup:
612  *
613  *	Wakeup the daemons responsible for freeing clean bufs.
614  */
615 static void
616 bufspace_daemon_wakeup(struct bufdomain *bd)
617 {
618 
619 	/*
620 	 * avoid the lock if the daemon is running.
621 	 */
622 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
623 		BD_RUN_LOCK(bd);
624 		atomic_store_int(&bd->bd_running, 1);
625 		wakeup(&bd->bd_running);
626 		BD_RUN_UNLOCK(bd);
627 	}
628 }
629 
630 /*
631  *	bufspace_daemon_wait:
632  *
633  *	Sleep until the domain falls below a limit or one second passes.
634  */
635 static void
636 bufspace_daemon_wait(struct bufdomain *bd)
637 {
638 	/*
639 	 * Re-check our limits and sleep.  bd_running must be
640 	 * cleared prior to checking the limits to avoid missed
641 	 * wakeups.  The waker will adjust one of bufspace or
642 	 * freebuffers prior to checking bd_running.
643 	 */
644 	BD_RUN_LOCK(bd);
645 	atomic_store_int(&bd->bd_running, 0);
646 	if (bd->bd_bufspace < bd->bd_bufspacethresh &&
647 	    bd->bd_freebuffers > bd->bd_lofreebuffers) {
648 		msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
649 		    "-", hz);
650 	} else {
651 		/* Avoid spurious wakeups while running. */
652 		atomic_store_int(&bd->bd_running, 1);
653 		BD_RUN_UNLOCK(bd);
654 	}
655 }
656 
657 /*
658  *	bufspace_adjust:
659  *
660  *	Adjust the reported bufspace for a KVA managed buffer, possibly
661  * 	waking any waiters.
662  */
663 static void
664 bufspace_adjust(struct buf *bp, int bufsize)
665 {
666 	struct bufdomain *bd;
667 	long space;
668 	int diff;
669 
670 	KASSERT((bp->b_flags & B_MALLOC) == 0,
671 	    ("bufspace_adjust: malloc buf %p", bp));
672 	bd = bufdomain(bp);
673 	diff = bufsize - bp->b_bufsize;
674 	if (diff < 0) {
675 		atomic_subtract_long(&bd->bd_bufspace, -diff);
676 	} else if (diff > 0) {
677 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
678 		/* Wake up the daemon on the transition. */
679 		if (space < bd->bd_bufspacethresh &&
680 		    space + diff >= bd->bd_bufspacethresh)
681 			bufspace_daemon_wakeup(bd);
682 	}
683 	bp->b_bufsize = bufsize;
684 }
685 
686 /*
687  *	bufspace_reserve:
688  *
689  *	Reserve bufspace before calling allocbuf().  metadata has a
690  *	different space limit than data.
691  */
692 static int
693 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
694 {
695 	long limit, new;
696 	long space;
697 
698 	if (metadata)
699 		limit = bd->bd_maxbufspace;
700 	else
701 		limit = bd->bd_hibufspace;
702 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
703 	new = space + size;
704 	if (new > limit) {
705 		atomic_subtract_long(&bd->bd_bufspace, size);
706 		return (ENOSPC);
707 	}
708 
709 	/* Wake up the daemon on the transition. */
710 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
711 		bufspace_daemon_wakeup(bd);
712 
713 	return (0);
714 }
715 
716 /*
717  *	bufspace_release:
718  *
719  *	Release reserved bufspace after bufspace_adjust() has consumed it.
720  */
721 static void
722 bufspace_release(struct bufdomain *bd, int size)
723 {
724 
725 	atomic_subtract_long(&bd->bd_bufspace, size);
726 }
727 
728 /*
729  *	bufspace_wait:
730  *
731  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
732  *	supplied.  bd_wanted must be set prior to polling for space.  The
733  *	operation must be re-tried on return.
734  */
735 static void
736 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
737     int slpflag, int slptimeo)
738 {
739 	struct thread *td;
740 	int error, fl, norunbuf;
741 
742 	if ((gbflags & GB_NOWAIT_BD) != 0)
743 		return;
744 
745 	td = curthread;
746 	BD_LOCK(bd);
747 	while (bd->bd_wanted) {
748 		if (vp != NULL && vp->v_type != VCHR &&
749 		    (td->td_pflags & TDP_BUFNEED) == 0) {
750 			BD_UNLOCK(bd);
751 			/*
752 			 * getblk() is called with a vnode locked, and
753 			 * some majority of the dirty buffers may as
754 			 * well belong to the vnode.  Flushing the
755 			 * buffers there would make a progress that
756 			 * cannot be achieved by the buf_daemon, that
757 			 * cannot lock the vnode.
758 			 */
759 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
760 			    (td->td_pflags & TDP_NORUNNINGBUF);
761 
762 			/*
763 			 * Play bufdaemon.  The getnewbuf() function
764 			 * may be called while the thread owns lock
765 			 * for another dirty buffer for the same
766 			 * vnode, which makes it impossible to use
767 			 * VOP_FSYNC() there, due to the buffer lock
768 			 * recursion.
769 			 */
770 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
771 			fl = buf_flush(vp, bd, flushbufqtarget);
772 			td->td_pflags &= norunbuf;
773 			BD_LOCK(bd);
774 			if (fl != 0)
775 				continue;
776 			if (bd->bd_wanted == 0)
777 				break;
778 		}
779 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
780 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
781 		if (error != 0)
782 			break;
783 	}
784 	BD_UNLOCK(bd);
785 }
786 
787 /*
788  *	bufspace_daemon:
789  *
790  *	buffer space management daemon.  Tries to maintain some marginal
791  *	amount of free buffer space so that requesting processes neither
792  *	block nor work to reclaim buffers.
793  */
794 static void
795 bufspace_daemon(void *arg)
796 {
797 	struct bufdomain *bd;
798 
799 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
800 	    SHUTDOWN_PRI_LAST + 100);
801 
802 	bd = arg;
803 	for (;;) {
804 		kthread_suspend_check();
805 
806 		/*
807 		 * Free buffers from the clean queue until we meet our
808 		 * targets.
809 		 *
810 		 * Theory of operation:  The buffer cache is most efficient
811 		 * when some free buffer headers and space are always
812 		 * available to getnewbuf().  This daemon attempts to prevent
813 		 * the excessive blocking and synchronization associated
814 		 * with shortfall.  It goes through three phases according
815 		 * demand:
816 		 *
817 		 * 1)	The daemon wakes up voluntarily once per-second
818 		 *	during idle periods when the counters are below
819 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
820 		 *
821 		 * 2)	The daemon wakes up as we cross the thresholds
822 		 *	ahead of any potential blocking.  This may bounce
823 		 *	slightly according to the rate of consumption and
824 		 *	release.
825 		 *
826 		 * 3)	The daemon and consumers are starved for working
827 		 *	clean buffers.  This is the 'bufspace' sleep below
828 		 *	which will inefficiently trade bufs with bqrelse
829 		 *	until we return to condition 2.
830 		 */
831 		while (bd->bd_bufspace > bd->bd_lobufspace ||
832 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
833 			if (buf_recycle(bd, false) != 0) {
834 				if (bd_flushall(bd))
835 					continue;
836 				/*
837 				 * Speedup dirty if we've run out of clean
838 				 * buffers.  This is possible in particular
839 				 * because softdep may held many bufs locked
840 				 * pending writes to other bufs which are
841 				 * marked for delayed write, exhausting
842 				 * clean space until they are written.
843 				 */
844 				bd_speedup();
845 				BD_LOCK(bd);
846 				if (bd->bd_wanted) {
847 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
848 					    PRIBIO|PDROP, "bufspace", hz/10);
849 				} else
850 					BD_UNLOCK(bd);
851 			}
852 			maybe_yield();
853 		}
854 		bufspace_daemon_wait(bd);
855 	}
856 }
857 
858 /*
859  *	bufmallocadjust:
860  *
861  *	Adjust the reported bufspace for a malloc managed buffer, possibly
862  *	waking any waiters.
863  */
864 static void
865 bufmallocadjust(struct buf *bp, int bufsize)
866 {
867 	int diff;
868 
869 	KASSERT((bp->b_flags & B_MALLOC) != 0,
870 	    ("bufmallocadjust: non-malloc buf %p", bp));
871 	diff = bufsize - bp->b_bufsize;
872 	if (diff < 0)
873 		atomic_subtract_long(&bufmallocspace, -diff);
874 	else
875 		atomic_add_long(&bufmallocspace, diff);
876 	bp->b_bufsize = bufsize;
877 }
878 
879 /*
880  *	runningwakeup:
881  *
882  *	Wake up processes that are waiting on asynchronous writes to fall
883  *	below lorunningspace.
884  */
885 static void
886 runningwakeup(void)
887 {
888 
889 	mtx_lock(&rbreqlock);
890 	if (runningbufreq) {
891 		runningbufreq = 0;
892 		wakeup(&runningbufreq);
893 	}
894 	mtx_unlock(&rbreqlock);
895 }
896 
897 /*
898  *	runningbufwakeup:
899  *
900  *	Decrement the outstanding write count according.
901  */
902 void
903 runningbufwakeup(struct buf *bp)
904 {
905 	long space, bspace;
906 
907 	bspace = bp->b_runningbufspace;
908 	if (bspace == 0)
909 		return;
910 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
911 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
912 	    space, bspace));
913 	bp->b_runningbufspace = 0;
914 	/*
915 	 * Only acquire the lock and wakeup on the transition from exceeding
916 	 * the threshold to falling below it.
917 	 */
918 	if (space < lorunningspace)
919 		return;
920 	if (space - bspace > lorunningspace)
921 		return;
922 	runningwakeup();
923 }
924 
925 /*
926  *	waitrunningbufspace()
927  *
928  *	runningbufspace is a measure of the amount of I/O currently
929  *	running.  This routine is used in async-write situations to
930  *	prevent creating huge backups of pending writes to a device.
931  *	Only asynchronous writes are governed by this function.
932  *
933  *	This does NOT turn an async write into a sync write.  It waits
934  *	for earlier writes to complete and generally returns before the
935  *	caller's write has reached the device.
936  */
937 void
938 waitrunningbufspace(void)
939 {
940 
941 	mtx_lock(&rbreqlock);
942 	while (runningbufspace > hirunningspace) {
943 		runningbufreq = 1;
944 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
945 	}
946 	mtx_unlock(&rbreqlock);
947 }
948 
949 /*
950  *	vfs_buf_test_cache:
951  *
952  *	Called when a buffer is extended.  This function clears the B_CACHE
953  *	bit if the newly extended portion of the buffer does not contain
954  *	valid data.
955  */
956 static __inline void
957 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
958     vm_offset_t size, vm_page_t m)
959 {
960 
961 	/*
962 	 * This function and its results are protected by higher level
963 	 * synchronization requiring vnode and buf locks to page in and
964 	 * validate pages.
965 	 */
966 	if (bp->b_flags & B_CACHE) {
967 		int base = (foff + off) & PAGE_MASK;
968 		if (vm_page_is_valid(m, base, size) == 0)
969 			bp->b_flags &= ~B_CACHE;
970 	}
971 }
972 
973 /* Wake up the buffer daemon if necessary */
974 static void
975 bd_wakeup(void)
976 {
977 
978 	mtx_lock(&bdlock);
979 	if (bd_request == 0) {
980 		bd_request = 1;
981 		wakeup(&bd_request);
982 	}
983 	mtx_unlock(&bdlock);
984 }
985 
986 /*
987  * Adjust the maxbcachbuf tunable.
988  */
989 static void
990 maxbcachebuf_adjust(void)
991 {
992 	int i;
993 
994 	/*
995 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
996 	 */
997 	i = 2;
998 	while (i * 2 <= maxbcachebuf)
999 		i *= 2;
1000 	maxbcachebuf = i;
1001 	if (maxbcachebuf < MAXBSIZE)
1002 		maxbcachebuf = MAXBSIZE;
1003 	if (maxbcachebuf > maxphys)
1004 		maxbcachebuf = maxphys;
1005 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1006 		printf("maxbcachebuf=%d\n", maxbcachebuf);
1007 }
1008 
1009 /*
1010  * bd_speedup - speedup the buffer cache flushing code
1011  */
1012 void
1013 bd_speedup(void)
1014 {
1015 	int needwake;
1016 
1017 	mtx_lock(&bdlock);
1018 	needwake = 0;
1019 	if (bd_speedupreq == 0 || bd_request == 0)
1020 		needwake = 1;
1021 	bd_speedupreq = 1;
1022 	bd_request = 1;
1023 	if (needwake)
1024 		wakeup(&bd_request);
1025 	mtx_unlock(&bdlock);
1026 }
1027 
1028 #ifdef __i386__
1029 #define	TRANSIENT_DENOM	5
1030 #else
1031 #define	TRANSIENT_DENOM 10
1032 #endif
1033 
1034 /*
1035  * Calculating buffer cache scaling values and reserve space for buffer
1036  * headers.  This is called during low level kernel initialization and
1037  * may be called more then once.  We CANNOT write to the memory area
1038  * being reserved at this time.
1039  */
1040 caddr_t
1041 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1042 {
1043 	int tuned_nbuf;
1044 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1045 
1046 	/*
1047 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1048 	 * PAGE_SIZE is >= 1K)
1049 	 */
1050 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1051 
1052 	maxbcachebuf_adjust();
1053 	/*
1054 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1055 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1056 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1057 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1058 	 * the buffer cache we limit the eventual kva reservation to
1059 	 * maxbcache bytes.
1060 	 *
1061 	 * factor represents the 1/4 x ram conversion.
1062 	 */
1063 	if (nbuf == 0) {
1064 		int factor = 4 * BKVASIZE / 1024;
1065 
1066 		nbuf = 50;
1067 		if (physmem_est > 4096)
1068 			nbuf += min((physmem_est - 4096) / factor,
1069 			    65536 / factor);
1070 		if (physmem_est > 65536)
1071 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1072 			    32 * 1024 * 1024 / (factor * 5));
1073 
1074 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1075 			nbuf = maxbcache / BKVASIZE;
1076 		tuned_nbuf = 1;
1077 	} else
1078 		tuned_nbuf = 0;
1079 
1080 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1081 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1082 	if (nbuf > maxbuf) {
1083 		if (!tuned_nbuf)
1084 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1085 			    maxbuf);
1086 		nbuf = maxbuf;
1087 	}
1088 
1089 	/*
1090 	 * Ideal allocation size for the transient bio submap is 10%
1091 	 * of the maximal space buffer map.  This roughly corresponds
1092 	 * to the amount of the buffer mapped for typical UFS load.
1093 	 *
1094 	 * Clip the buffer map to reserve space for the transient
1095 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1096 	 * maximum buffer map extent on the platform.
1097 	 *
1098 	 * The fall-back to the maxbuf in case of maxbcache unset,
1099 	 * allows to not trim the buffer KVA for the architectures
1100 	 * with ample KVA space.
1101 	 */
1102 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1103 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1104 		buf_sz = (long)nbuf * BKVASIZE;
1105 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1106 		    (TRANSIENT_DENOM - 1)) {
1107 			/*
1108 			 * There is more KVA than memory.  Do not
1109 			 * adjust buffer map size, and assign the rest
1110 			 * of maxbuf to transient map.
1111 			 */
1112 			biotmap_sz = maxbuf_sz - buf_sz;
1113 		} else {
1114 			/*
1115 			 * Buffer map spans all KVA we could afford on
1116 			 * this platform.  Give 10% (20% on i386) of
1117 			 * the buffer map to the transient bio map.
1118 			 */
1119 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1120 			buf_sz -= biotmap_sz;
1121 		}
1122 		if (biotmap_sz / INT_MAX > maxphys)
1123 			bio_transient_maxcnt = INT_MAX;
1124 		else
1125 			bio_transient_maxcnt = biotmap_sz / maxphys;
1126 		/*
1127 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1128 		 * using the transient mapping.
1129 		 */
1130 		if (bio_transient_maxcnt > 1024)
1131 			bio_transient_maxcnt = 1024;
1132 		if (tuned_nbuf)
1133 			nbuf = buf_sz / BKVASIZE;
1134 	}
1135 
1136 	if (nswbuf == 0) {
1137 		nswbuf = min(nbuf / 4, 256);
1138 		if (nswbuf < NSWBUF_MIN)
1139 			nswbuf = NSWBUF_MIN;
1140 	}
1141 
1142 	/*
1143 	 * Reserve space for the buffer cache buffers
1144 	 */
1145 	buf = (char *)v;
1146 	v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1147 	    atop(maxbcachebuf)) * nbuf;
1148 
1149 	return (v);
1150 }
1151 
1152 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1153 void
1154 bufinit(void)
1155 {
1156 	struct buf *bp;
1157 	int i;
1158 
1159 	KASSERT(maxbcachebuf >= MAXBSIZE,
1160 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1161 	    MAXBSIZE));
1162 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1163 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1164 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1165 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1166 
1167 	unmapped_buf = (caddr_t)kva_alloc(maxphys);
1168 
1169 	/* finally, initialize each buffer header and stick on empty q */
1170 	for (i = 0; i < nbuf; i++) {
1171 		bp = nbufp(i);
1172 		bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1173 		bp->b_flags = B_INVAL;
1174 		bp->b_rcred = NOCRED;
1175 		bp->b_wcred = NOCRED;
1176 		bp->b_qindex = QUEUE_NONE;
1177 		bp->b_domain = -1;
1178 		bp->b_subqueue = mp_maxid + 1;
1179 		bp->b_xflags = 0;
1180 		bp->b_data = bp->b_kvabase = unmapped_buf;
1181 		LIST_INIT(&bp->b_dep);
1182 		BUF_LOCKINIT(bp);
1183 		bq_insert(&bqempty, bp, false);
1184 	}
1185 
1186 	/*
1187 	 * maxbufspace is the absolute maximum amount of buffer space we are
1188 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1189 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1190 	 * used by most other requests.  The differential is required to
1191 	 * ensure that metadata deadlocks don't occur.
1192 	 *
1193 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1194 	 * this may result in KVM fragmentation which is not handled optimally
1195 	 * by the system. XXX This is less true with vmem.  We could use
1196 	 * PAGE_SIZE.
1197 	 */
1198 	maxbufspace = (long)nbuf * BKVASIZE;
1199 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1200 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1201 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1202 
1203 	/*
1204 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1205 	 * arbitrarily and may need further tuning. It corresponds to
1206 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1207 	 * which fits with many RAID controllers' tagged queuing limits.
1208 	 * The lower 1 MiB limit is the historical upper limit for
1209 	 * hirunningspace.
1210 	 */
1211 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1212 	    16 * 1024 * 1024), 1024 * 1024);
1213 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1214 
1215 	/*
1216 	 * Limit the amount of malloc memory since it is wired permanently into
1217 	 * the kernel space.  Even though this is accounted for in the buffer
1218 	 * allocation, we don't want the malloced region to grow uncontrolled.
1219 	 * The malloc scheme improves memory utilization significantly on
1220 	 * average (small) directories.
1221 	 */
1222 	maxbufmallocspace = hibufspace / 20;
1223 
1224 	/*
1225 	 * Reduce the chance of a deadlock occurring by limiting the number
1226 	 * of delayed-write dirty buffers we allow to stack up.
1227 	 */
1228 	hidirtybuffers = nbuf / 4 + 20;
1229 	dirtybufthresh = hidirtybuffers * 9 / 10;
1230 	/*
1231 	 * To support extreme low-memory systems, make sure hidirtybuffers
1232 	 * cannot eat up all available buffer space.  This occurs when our
1233 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1234 	 * buffer space assuming BKVASIZE'd buffers.
1235 	 */
1236 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1237 		hidirtybuffers >>= 1;
1238 	}
1239 	lodirtybuffers = hidirtybuffers / 2;
1240 
1241 	/*
1242 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1243 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1244 	 * are counted as free but will be unavailable to threads executing
1245 	 * on other cpus.
1246 	 *
1247 	 * hifreebuffers is the free target for the bufspace daemon.  This
1248 	 * should be set appropriately to limit work per-iteration.
1249 	 */
1250 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1251 	hifreebuffers = (3 * lofreebuffers) / 2;
1252 	numfreebuffers = nbuf;
1253 
1254 	/* Setup the kva and free list allocators. */
1255 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1256 	buf_zone = uma_zcache_create("buf free cache",
1257 	    sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1258 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1259 
1260 	/*
1261 	 * Size the clean queue according to the amount of buffer space.
1262 	 * One queue per-256mb up to the max.  More queues gives better
1263 	 * concurrency but less accurate LRU.
1264 	 */
1265 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1266 	for (i = 0 ; i < buf_domains; i++) {
1267 		struct bufdomain *bd;
1268 
1269 		bd = &bdomain[i];
1270 		bd_init(bd);
1271 		bd->bd_freebuffers = nbuf / buf_domains;
1272 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1273 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1274 		bd->bd_bufspace = 0;
1275 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1276 		bd->bd_hibufspace = hibufspace / buf_domains;
1277 		bd->bd_lobufspace = lobufspace / buf_domains;
1278 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1279 		bd->bd_numdirtybuffers = 0;
1280 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1281 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1282 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1283 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1284 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1285 	}
1286 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1287 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1288 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1289 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1290 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1291 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1292 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1293 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1294 }
1295 
1296 #ifdef INVARIANTS
1297 static inline void
1298 vfs_buf_check_mapped(struct buf *bp)
1299 {
1300 
1301 	KASSERT(bp->b_kvabase != unmapped_buf,
1302 	    ("mapped buf: b_kvabase was not updated %p", bp));
1303 	KASSERT(bp->b_data != unmapped_buf,
1304 	    ("mapped buf: b_data was not updated %p", bp));
1305 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1306 	    maxphys, ("b_data + b_offset unmapped %p", bp));
1307 }
1308 
1309 static inline void
1310 vfs_buf_check_unmapped(struct buf *bp)
1311 {
1312 
1313 	KASSERT(bp->b_data == unmapped_buf,
1314 	    ("unmapped buf: corrupted b_data %p", bp));
1315 }
1316 
1317 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1318 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1319 #else
1320 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1321 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1322 #endif
1323 
1324 static int
1325 isbufbusy(struct buf *bp)
1326 {
1327 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1328 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1329 		return (1);
1330 	return (0);
1331 }
1332 
1333 /*
1334  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1335  */
1336 void
1337 bufshutdown(int show_busybufs)
1338 {
1339 	static int first_buf_printf = 1;
1340 	struct buf *bp;
1341 	int i, iter, nbusy, pbusy;
1342 #ifndef PREEMPTION
1343 	int subiter;
1344 #endif
1345 
1346 	/*
1347 	 * Sync filesystems for shutdown
1348 	 */
1349 	wdog_kern_pat(WD_LASTVAL);
1350 	kern_sync(curthread);
1351 
1352 	/*
1353 	 * With soft updates, some buffers that are
1354 	 * written will be remarked as dirty until other
1355 	 * buffers are written.
1356 	 */
1357 	for (iter = pbusy = 0; iter < 20; iter++) {
1358 		nbusy = 0;
1359 		for (i = nbuf - 1; i >= 0; i--) {
1360 			bp = nbufp(i);
1361 			if (isbufbusy(bp))
1362 				nbusy++;
1363 		}
1364 		if (nbusy == 0) {
1365 			if (first_buf_printf)
1366 				printf("All buffers synced.");
1367 			break;
1368 		}
1369 		if (first_buf_printf) {
1370 			printf("Syncing disks, buffers remaining... ");
1371 			first_buf_printf = 0;
1372 		}
1373 		printf("%d ", nbusy);
1374 		if (nbusy < pbusy)
1375 			iter = 0;
1376 		pbusy = nbusy;
1377 
1378 		wdog_kern_pat(WD_LASTVAL);
1379 		kern_sync(curthread);
1380 
1381 #ifdef PREEMPTION
1382 		/*
1383 		 * Spin for a while to allow interrupt threads to run.
1384 		 */
1385 		DELAY(50000 * iter);
1386 #else
1387 		/*
1388 		 * Context switch several times to allow interrupt
1389 		 * threads to run.
1390 		 */
1391 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1392 			thread_lock(curthread);
1393 			mi_switch(SW_VOL);
1394 			DELAY(1000);
1395 		}
1396 #endif
1397 	}
1398 	printf("\n");
1399 	/*
1400 	 * Count only busy local buffers to prevent forcing
1401 	 * a fsck if we're just a client of a wedged NFS server
1402 	 */
1403 	nbusy = 0;
1404 	for (i = nbuf - 1; i >= 0; i--) {
1405 		bp = nbufp(i);
1406 		if (isbufbusy(bp)) {
1407 #if 0
1408 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1409 			if (bp->b_dev == NULL) {
1410 				TAILQ_REMOVE(&mountlist,
1411 				    bp->b_vp->v_mount, mnt_list);
1412 				continue;
1413 			}
1414 #endif
1415 			nbusy++;
1416 			if (show_busybufs > 0) {
1417 				printf(
1418 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1419 				    nbusy, bp, bp->b_vp, bp->b_flags,
1420 				    (intmax_t)bp->b_blkno,
1421 				    (intmax_t)bp->b_lblkno);
1422 				BUF_LOCKPRINTINFO(bp);
1423 				if (show_busybufs > 1)
1424 					vn_printf(bp->b_vp,
1425 					    "vnode content: ");
1426 			}
1427 		}
1428 	}
1429 	if (nbusy) {
1430 		/*
1431 		 * Failed to sync all blocks. Indicate this and don't
1432 		 * unmount filesystems (thus forcing an fsck on reboot).
1433 		 */
1434 		printf("Giving up on %d buffers\n", nbusy);
1435 		DELAY(5000000);	/* 5 seconds */
1436 	} else {
1437 		if (!first_buf_printf)
1438 			printf("Final sync complete\n");
1439 		/*
1440 		 * Unmount filesystems
1441 		 */
1442 		if (!KERNEL_PANICKED())
1443 			vfs_unmountall();
1444 	}
1445 	swapoff_all();
1446 	DELAY(100000);		/* wait for console output to finish */
1447 }
1448 
1449 static void
1450 bpmap_qenter(struct buf *bp)
1451 {
1452 
1453 	BUF_CHECK_MAPPED(bp);
1454 
1455 	/*
1456 	 * bp->b_data is relative to bp->b_offset, but
1457 	 * bp->b_offset may be offset into the first page.
1458 	 */
1459 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1460 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1461 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1462 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1463 }
1464 
1465 static inline struct bufdomain *
1466 bufdomain(struct buf *bp)
1467 {
1468 
1469 	return (&bdomain[bp->b_domain]);
1470 }
1471 
1472 static struct bufqueue *
1473 bufqueue(struct buf *bp)
1474 {
1475 
1476 	switch (bp->b_qindex) {
1477 	case QUEUE_NONE:
1478 		/* FALLTHROUGH */
1479 	case QUEUE_SENTINEL:
1480 		return (NULL);
1481 	case QUEUE_EMPTY:
1482 		return (&bqempty);
1483 	case QUEUE_DIRTY:
1484 		return (&bufdomain(bp)->bd_dirtyq);
1485 	case QUEUE_CLEAN:
1486 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1487 	default:
1488 		break;
1489 	}
1490 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1491 }
1492 
1493 /*
1494  * Return the locked bufqueue that bp is a member of.
1495  */
1496 static struct bufqueue *
1497 bufqueue_acquire(struct buf *bp)
1498 {
1499 	struct bufqueue *bq, *nbq;
1500 
1501 	/*
1502 	 * bp can be pushed from a per-cpu queue to the
1503 	 * cleanq while we're waiting on the lock.  Retry
1504 	 * if the queues don't match.
1505 	 */
1506 	bq = bufqueue(bp);
1507 	BQ_LOCK(bq);
1508 	for (;;) {
1509 		nbq = bufqueue(bp);
1510 		if (bq == nbq)
1511 			break;
1512 		BQ_UNLOCK(bq);
1513 		BQ_LOCK(nbq);
1514 		bq = nbq;
1515 	}
1516 	return (bq);
1517 }
1518 
1519 /*
1520  *	binsfree:
1521  *
1522  *	Insert the buffer into the appropriate free list.  Requires a
1523  *	locked buffer on entry and buffer is unlocked before return.
1524  */
1525 static void
1526 binsfree(struct buf *bp, int qindex)
1527 {
1528 	struct bufdomain *bd;
1529 	struct bufqueue *bq;
1530 
1531 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1532 	    ("binsfree: Invalid qindex %d", qindex));
1533 	BUF_ASSERT_XLOCKED(bp);
1534 
1535 	/*
1536 	 * Handle delayed bremfree() processing.
1537 	 */
1538 	if (bp->b_flags & B_REMFREE) {
1539 		if (bp->b_qindex == qindex) {
1540 			bp->b_flags |= B_REUSE;
1541 			bp->b_flags &= ~B_REMFREE;
1542 			BUF_UNLOCK(bp);
1543 			return;
1544 		}
1545 		bq = bufqueue_acquire(bp);
1546 		bq_remove(bq, bp);
1547 		BQ_UNLOCK(bq);
1548 	}
1549 	bd = bufdomain(bp);
1550 	if (qindex == QUEUE_CLEAN) {
1551 		if (bd->bd_lim != 0)
1552 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1553 		else
1554 			bq = bd->bd_cleanq;
1555 	} else
1556 		bq = &bd->bd_dirtyq;
1557 	bq_insert(bq, bp, true);
1558 }
1559 
1560 /*
1561  * buf_free:
1562  *
1563  *	Free a buffer to the buf zone once it no longer has valid contents.
1564  */
1565 static void
1566 buf_free(struct buf *bp)
1567 {
1568 
1569 	if (bp->b_flags & B_REMFREE)
1570 		bremfreef(bp);
1571 	if (bp->b_vflags & BV_BKGRDINPROG)
1572 		panic("losing buffer 1");
1573 	if (bp->b_rcred != NOCRED) {
1574 		crfree(bp->b_rcred);
1575 		bp->b_rcred = NOCRED;
1576 	}
1577 	if (bp->b_wcred != NOCRED) {
1578 		crfree(bp->b_wcred);
1579 		bp->b_wcred = NOCRED;
1580 	}
1581 	if (!LIST_EMPTY(&bp->b_dep))
1582 		buf_deallocate(bp);
1583 	bufkva_free(bp);
1584 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1585 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
1586 	BUF_UNLOCK(bp);
1587 	uma_zfree(buf_zone, bp);
1588 }
1589 
1590 /*
1591  * buf_import:
1592  *
1593  *	Import bufs into the uma cache from the buf list.  The system still
1594  *	expects a static array of bufs and much of the synchronization
1595  *	around bufs assumes type stable storage.  As a result, UMA is used
1596  *	only as a per-cpu cache of bufs still maintained on a global list.
1597  */
1598 static int
1599 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1600 {
1601 	struct buf *bp;
1602 	int i;
1603 
1604 	BQ_LOCK(&bqempty);
1605 	for (i = 0; i < cnt; i++) {
1606 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1607 		if (bp == NULL)
1608 			break;
1609 		bq_remove(&bqempty, bp);
1610 		store[i] = bp;
1611 	}
1612 	BQ_UNLOCK(&bqempty);
1613 
1614 	return (i);
1615 }
1616 
1617 /*
1618  * buf_release:
1619  *
1620  *	Release bufs from the uma cache back to the buffer queues.
1621  */
1622 static void
1623 buf_release(void *arg, void **store, int cnt)
1624 {
1625 	struct bufqueue *bq;
1626 	struct buf *bp;
1627         int i;
1628 
1629 	bq = &bqempty;
1630 	BQ_LOCK(bq);
1631         for (i = 0; i < cnt; i++) {
1632 		bp = store[i];
1633 		/* Inline bq_insert() to batch locking. */
1634 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1635 		bp->b_flags &= ~(B_AGE | B_REUSE);
1636 		bq->bq_len++;
1637 		bp->b_qindex = bq->bq_index;
1638 	}
1639 	BQ_UNLOCK(bq);
1640 }
1641 
1642 /*
1643  * buf_alloc:
1644  *
1645  *	Allocate an empty buffer header.
1646  */
1647 static struct buf *
1648 buf_alloc(struct bufdomain *bd)
1649 {
1650 	struct buf *bp;
1651 	int freebufs, error;
1652 
1653 	/*
1654 	 * We can only run out of bufs in the buf zone if the average buf
1655 	 * is less than BKVASIZE.  In this case the actual wait/block will
1656 	 * come from buf_reycle() failing to flush one of these small bufs.
1657 	 */
1658 	bp = NULL;
1659 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1660 	if (freebufs > 0)
1661 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1662 	if (bp == NULL) {
1663 		atomic_add_int(&bd->bd_freebuffers, 1);
1664 		bufspace_daemon_wakeup(bd);
1665 		counter_u64_add(numbufallocfails, 1);
1666 		return (NULL);
1667 	}
1668 	/*
1669 	 * Wake-up the bufspace daemon on transition below threshold.
1670 	 */
1671 	if (freebufs == bd->bd_lofreebuffers)
1672 		bufspace_daemon_wakeup(bd);
1673 
1674 	error = BUF_LOCK(bp, LK_EXCLUSIVE, NULL);
1675 	KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1676 	    error));
1677 	(void)error;
1678 
1679 	KASSERT(bp->b_vp == NULL,
1680 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1681 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1682 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1683 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1684 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1685 	KASSERT(bp->b_npages == 0,
1686 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1687 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1688 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1689 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
1690 
1691 	bp->b_domain = BD_DOMAIN(bd);
1692 	bp->b_flags = 0;
1693 	bp->b_ioflags = 0;
1694 	bp->b_xflags = 0;
1695 	bp->b_vflags = 0;
1696 	bp->b_vp = NULL;
1697 	bp->b_blkno = bp->b_lblkno = 0;
1698 	bp->b_offset = NOOFFSET;
1699 	bp->b_iodone = 0;
1700 	bp->b_error = 0;
1701 	bp->b_resid = 0;
1702 	bp->b_bcount = 0;
1703 	bp->b_npages = 0;
1704 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1705 	bp->b_bufobj = NULL;
1706 	bp->b_data = bp->b_kvabase = unmapped_buf;
1707 	bp->b_fsprivate1 = NULL;
1708 	bp->b_fsprivate2 = NULL;
1709 	bp->b_fsprivate3 = NULL;
1710 	LIST_INIT(&bp->b_dep);
1711 
1712 	return (bp);
1713 }
1714 
1715 /*
1716  *	buf_recycle:
1717  *
1718  *	Free a buffer from the given bufqueue.  kva controls whether the
1719  *	freed buf must own some kva resources.  This is used for
1720  *	defragmenting.
1721  */
1722 static int
1723 buf_recycle(struct bufdomain *bd, bool kva)
1724 {
1725 	struct bufqueue *bq;
1726 	struct buf *bp, *nbp;
1727 
1728 	if (kva)
1729 		counter_u64_add(bufdefragcnt, 1);
1730 	nbp = NULL;
1731 	bq = bd->bd_cleanq;
1732 	BQ_LOCK(bq);
1733 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1734 	    ("buf_recycle: Locks don't match"));
1735 	nbp = TAILQ_FIRST(&bq->bq_queue);
1736 
1737 	/*
1738 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1739 	 * depending.
1740 	 */
1741 	while ((bp = nbp) != NULL) {
1742 		/*
1743 		 * Calculate next bp (we can only use it if we do not
1744 		 * release the bqlock).
1745 		 */
1746 		nbp = TAILQ_NEXT(bp, b_freelist);
1747 
1748 		/*
1749 		 * If we are defragging then we need a buffer with
1750 		 * some kva to reclaim.
1751 		 */
1752 		if (kva && bp->b_kvasize == 0)
1753 			continue;
1754 
1755 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1756 			continue;
1757 
1758 		/*
1759 		 * Implement a second chance algorithm for frequently
1760 		 * accessed buffers.
1761 		 */
1762 		if ((bp->b_flags & B_REUSE) != 0) {
1763 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1764 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1765 			bp->b_flags &= ~B_REUSE;
1766 			BUF_UNLOCK(bp);
1767 			continue;
1768 		}
1769 
1770 		/*
1771 		 * Skip buffers with background writes in progress.
1772 		 */
1773 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1774 			BUF_UNLOCK(bp);
1775 			continue;
1776 		}
1777 
1778 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1779 		    ("buf_recycle: inconsistent queue %d bp %p",
1780 		    bp->b_qindex, bp));
1781 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1782 		    ("getnewbuf: queue domain %d doesn't match request %d",
1783 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1784 		/*
1785 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1786 		 * the scan from this point on.
1787 		 */
1788 		bq_remove(bq, bp);
1789 		BQ_UNLOCK(bq);
1790 
1791 		/*
1792 		 * Requeue the background write buffer with error and
1793 		 * restart the scan.
1794 		 */
1795 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1796 			bqrelse(bp);
1797 			BQ_LOCK(bq);
1798 			nbp = TAILQ_FIRST(&bq->bq_queue);
1799 			continue;
1800 		}
1801 		bp->b_flags |= B_INVAL;
1802 		brelse(bp);
1803 		return (0);
1804 	}
1805 	bd->bd_wanted = 1;
1806 	BQ_UNLOCK(bq);
1807 
1808 	return (ENOBUFS);
1809 }
1810 
1811 /*
1812  *	bremfree:
1813  *
1814  *	Mark the buffer for removal from the appropriate free list.
1815  *
1816  */
1817 void
1818 bremfree(struct buf *bp)
1819 {
1820 
1821 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1822 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1823 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1824 	KASSERT(bp->b_qindex != QUEUE_NONE,
1825 	    ("bremfree: buffer %p not on a queue.", bp));
1826 	BUF_ASSERT_XLOCKED(bp);
1827 
1828 	bp->b_flags |= B_REMFREE;
1829 }
1830 
1831 /*
1832  *	bremfreef:
1833  *
1834  *	Force an immediate removal from a free list.  Used only in nfs when
1835  *	it abuses the b_freelist pointer.
1836  */
1837 void
1838 bremfreef(struct buf *bp)
1839 {
1840 	struct bufqueue *bq;
1841 
1842 	bq = bufqueue_acquire(bp);
1843 	bq_remove(bq, bp);
1844 	BQ_UNLOCK(bq);
1845 }
1846 
1847 static void
1848 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1849 {
1850 
1851 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1852 	TAILQ_INIT(&bq->bq_queue);
1853 	bq->bq_len = 0;
1854 	bq->bq_index = qindex;
1855 	bq->bq_subqueue = subqueue;
1856 }
1857 
1858 static void
1859 bd_init(struct bufdomain *bd)
1860 {
1861 	int i;
1862 
1863 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1864 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1865 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1866 	for (i = 0; i <= mp_maxid; i++)
1867 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1868 		    "bufq clean subqueue lock");
1869 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1870 }
1871 
1872 /*
1873  *	bq_remove:
1874  *
1875  *	Removes a buffer from the free list, must be called with the
1876  *	correct qlock held.
1877  */
1878 static void
1879 bq_remove(struct bufqueue *bq, struct buf *bp)
1880 {
1881 
1882 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1883 	    bp, bp->b_vp, bp->b_flags);
1884 	KASSERT(bp->b_qindex != QUEUE_NONE,
1885 	    ("bq_remove: buffer %p not on a queue.", bp));
1886 	KASSERT(bufqueue(bp) == bq,
1887 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1888 
1889 	BQ_ASSERT_LOCKED(bq);
1890 	if (bp->b_qindex != QUEUE_EMPTY) {
1891 		BUF_ASSERT_XLOCKED(bp);
1892 	}
1893 	KASSERT(bq->bq_len >= 1,
1894 	    ("queue %d underflow", bp->b_qindex));
1895 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1896 	bq->bq_len--;
1897 	bp->b_qindex = QUEUE_NONE;
1898 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1899 }
1900 
1901 static void
1902 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1903 {
1904 	struct buf *bp;
1905 
1906 	BQ_ASSERT_LOCKED(bq);
1907 	if (bq != bd->bd_cleanq) {
1908 		BD_LOCK(bd);
1909 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1910 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1911 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1912 			    b_freelist);
1913 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1914 		}
1915 		bd->bd_cleanq->bq_len += bq->bq_len;
1916 		bq->bq_len = 0;
1917 	}
1918 	if (bd->bd_wanted) {
1919 		bd->bd_wanted = 0;
1920 		wakeup(&bd->bd_wanted);
1921 	}
1922 	if (bq != bd->bd_cleanq)
1923 		BD_UNLOCK(bd);
1924 }
1925 
1926 static int
1927 bd_flushall(struct bufdomain *bd)
1928 {
1929 	struct bufqueue *bq;
1930 	int flushed;
1931 	int i;
1932 
1933 	if (bd->bd_lim == 0)
1934 		return (0);
1935 	flushed = 0;
1936 	for (i = 0; i <= mp_maxid; i++) {
1937 		bq = &bd->bd_subq[i];
1938 		if (bq->bq_len == 0)
1939 			continue;
1940 		BQ_LOCK(bq);
1941 		bd_flush(bd, bq);
1942 		BQ_UNLOCK(bq);
1943 		flushed++;
1944 	}
1945 
1946 	return (flushed);
1947 }
1948 
1949 static void
1950 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1951 {
1952 	struct bufdomain *bd;
1953 
1954 	if (bp->b_qindex != QUEUE_NONE)
1955 		panic("bq_insert: free buffer %p onto another queue?", bp);
1956 
1957 	bd = bufdomain(bp);
1958 	if (bp->b_flags & B_AGE) {
1959 		/* Place this buf directly on the real queue. */
1960 		if (bq->bq_index == QUEUE_CLEAN)
1961 			bq = bd->bd_cleanq;
1962 		BQ_LOCK(bq);
1963 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1964 	} else {
1965 		BQ_LOCK(bq);
1966 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1967 	}
1968 	bp->b_flags &= ~(B_AGE | B_REUSE);
1969 	bq->bq_len++;
1970 	bp->b_qindex = bq->bq_index;
1971 	bp->b_subqueue = bq->bq_subqueue;
1972 
1973 	/*
1974 	 * Unlock before we notify so that we don't wakeup a waiter that
1975 	 * fails a trylock on the buf and sleeps again.
1976 	 */
1977 	if (unlock)
1978 		BUF_UNLOCK(bp);
1979 
1980 	if (bp->b_qindex == QUEUE_CLEAN) {
1981 		/*
1982 		 * Flush the per-cpu queue and notify any waiters.
1983 		 */
1984 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1985 		    bq->bq_len >= bd->bd_lim))
1986 			bd_flush(bd, bq);
1987 	}
1988 	BQ_UNLOCK(bq);
1989 }
1990 
1991 /*
1992  *	bufkva_free:
1993  *
1994  *	Free the kva allocation for a buffer.
1995  *
1996  */
1997 static void
1998 bufkva_free(struct buf *bp)
1999 {
2000 
2001 #ifdef INVARIANTS
2002 	if (bp->b_kvasize == 0) {
2003 		KASSERT(bp->b_kvabase == unmapped_buf &&
2004 		    bp->b_data == unmapped_buf,
2005 		    ("Leaked KVA space on %p", bp));
2006 	} else if (buf_mapped(bp))
2007 		BUF_CHECK_MAPPED(bp);
2008 	else
2009 		BUF_CHECK_UNMAPPED(bp);
2010 #endif
2011 	if (bp->b_kvasize == 0)
2012 		return;
2013 
2014 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2015 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2016 	counter_u64_add(buffreekvacnt, 1);
2017 	bp->b_data = bp->b_kvabase = unmapped_buf;
2018 	bp->b_kvasize = 0;
2019 }
2020 
2021 /*
2022  *	bufkva_alloc:
2023  *
2024  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2025  */
2026 static int
2027 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2028 {
2029 	vm_offset_t addr;
2030 	int error;
2031 
2032 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2033 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2034 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
2035 	KASSERT(maxsize <= maxbcachebuf,
2036 	    ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2037 
2038 	bufkva_free(bp);
2039 
2040 	addr = 0;
2041 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2042 	if (error != 0) {
2043 		/*
2044 		 * Buffer map is too fragmented.  Request the caller
2045 		 * to defragment the map.
2046 		 */
2047 		return (error);
2048 	}
2049 	bp->b_kvabase = (caddr_t)addr;
2050 	bp->b_kvasize = maxsize;
2051 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2052 	if ((gbflags & GB_UNMAPPED) != 0) {
2053 		bp->b_data = unmapped_buf;
2054 		BUF_CHECK_UNMAPPED(bp);
2055 	} else {
2056 		bp->b_data = bp->b_kvabase;
2057 		BUF_CHECK_MAPPED(bp);
2058 	}
2059 	return (0);
2060 }
2061 
2062 /*
2063  *	bufkva_reclaim:
2064  *
2065  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2066  *	callback that fires to avoid returning failure.
2067  */
2068 static void
2069 bufkva_reclaim(vmem_t *vmem, int flags)
2070 {
2071 	bool done;
2072 	int q;
2073 	int i;
2074 
2075 	done = false;
2076 	for (i = 0; i < 5; i++) {
2077 		for (q = 0; q < buf_domains; q++)
2078 			if (buf_recycle(&bdomain[q], true) != 0)
2079 				done = true;
2080 		if (done)
2081 			break;
2082 	}
2083 	return;
2084 }
2085 
2086 /*
2087  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2088  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2089  * the buffer is valid and we do not have to do anything.
2090  */
2091 static void
2092 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2093     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2094 {
2095 	struct buf *rabp;
2096 	struct thread *td;
2097 	int i;
2098 
2099 	td = curthread;
2100 
2101 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2102 		if (inmem(vp, *rablkno))
2103 			continue;
2104 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2105 		if ((rabp->b_flags & B_CACHE) != 0) {
2106 			brelse(rabp);
2107 			continue;
2108 		}
2109 #ifdef RACCT
2110 		if (racct_enable) {
2111 			PROC_LOCK(curproc);
2112 			racct_add_buf(curproc, rabp, 0);
2113 			PROC_UNLOCK(curproc);
2114 		}
2115 #endif /* RACCT */
2116 		td->td_ru.ru_inblock++;
2117 		rabp->b_flags |= B_ASYNC;
2118 		rabp->b_flags &= ~B_INVAL;
2119 		if ((flags & GB_CKHASH) != 0) {
2120 			rabp->b_flags |= B_CKHASH;
2121 			rabp->b_ckhashcalc = ckhashfunc;
2122 		}
2123 		rabp->b_ioflags &= ~BIO_ERROR;
2124 		rabp->b_iocmd = BIO_READ;
2125 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2126 			rabp->b_rcred = crhold(cred);
2127 		vfs_busy_pages(rabp, 0);
2128 		BUF_KERNPROC(rabp);
2129 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2130 		bstrategy(rabp);
2131 	}
2132 }
2133 
2134 /*
2135  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2136  *
2137  * Get a buffer with the specified data.  Look in the cache first.  We
2138  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2139  * is set, the buffer is valid and we do not have to do anything, see
2140  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2141  *
2142  * Always return a NULL buffer pointer (in bpp) when returning an error.
2143  *
2144  * The blkno parameter is the logical block being requested. Normally
2145  * the mapping of logical block number to disk block address is done
2146  * by calling VOP_BMAP(). However, if the mapping is already known, the
2147  * disk block address can be passed using the dblkno parameter. If the
2148  * disk block address is not known, then the same value should be passed
2149  * for blkno and dblkno.
2150  */
2151 int
2152 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2153     daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2154     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2155 {
2156 	struct buf *bp;
2157 	struct thread *td;
2158 	int error, readwait, rv;
2159 
2160 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2161 	td = curthread;
2162 	/*
2163 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2164 	 * are specified.
2165 	 */
2166 	error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2167 	if (error != 0) {
2168 		*bpp = NULL;
2169 		return (error);
2170 	}
2171 	KASSERT(blkno == bp->b_lblkno,
2172 	    ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2173 	    (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2174 	flags &= ~GB_NOSPARSE;
2175 	*bpp = bp;
2176 
2177 	/*
2178 	 * If not found in cache, do some I/O
2179 	 */
2180 	readwait = 0;
2181 	if ((bp->b_flags & B_CACHE) == 0) {
2182 #ifdef RACCT
2183 		if (racct_enable) {
2184 			PROC_LOCK(td->td_proc);
2185 			racct_add_buf(td->td_proc, bp, 0);
2186 			PROC_UNLOCK(td->td_proc);
2187 		}
2188 #endif /* RACCT */
2189 		td->td_ru.ru_inblock++;
2190 		bp->b_iocmd = BIO_READ;
2191 		bp->b_flags &= ~B_INVAL;
2192 		if ((flags & GB_CKHASH) != 0) {
2193 			bp->b_flags |= B_CKHASH;
2194 			bp->b_ckhashcalc = ckhashfunc;
2195 		}
2196 		if ((flags & GB_CVTENXIO) != 0)
2197 			bp->b_xflags |= BX_CVTENXIO;
2198 		bp->b_ioflags &= ~BIO_ERROR;
2199 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2200 			bp->b_rcred = crhold(cred);
2201 		vfs_busy_pages(bp, 0);
2202 		bp->b_iooffset = dbtob(bp->b_blkno);
2203 		bstrategy(bp);
2204 		++readwait;
2205 	}
2206 
2207 	/*
2208 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2209 	 */
2210 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2211 
2212 	rv = 0;
2213 	if (readwait) {
2214 		rv = bufwait(bp);
2215 		if (rv != 0) {
2216 			brelse(bp);
2217 			*bpp = NULL;
2218 		}
2219 	}
2220 	return (rv);
2221 }
2222 
2223 /*
2224  * Write, release buffer on completion.  (Done by iodone
2225  * if async).  Do not bother writing anything if the buffer
2226  * is invalid.
2227  *
2228  * Note that we set B_CACHE here, indicating that buffer is
2229  * fully valid and thus cacheable.  This is true even of NFS
2230  * now so we set it generally.  This could be set either here
2231  * or in biodone() since the I/O is synchronous.  We put it
2232  * here.
2233  */
2234 int
2235 bufwrite(struct buf *bp)
2236 {
2237 	int oldflags;
2238 	struct vnode *vp;
2239 	long space;
2240 	int vp_md;
2241 
2242 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2243 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2244 		bp->b_flags |= B_INVAL | B_RELBUF;
2245 		bp->b_flags &= ~B_CACHE;
2246 		brelse(bp);
2247 		return (ENXIO);
2248 	}
2249 	if (bp->b_flags & B_INVAL) {
2250 		brelse(bp);
2251 		return (0);
2252 	}
2253 
2254 	if (bp->b_flags & B_BARRIER)
2255 		atomic_add_long(&barrierwrites, 1);
2256 
2257 	oldflags = bp->b_flags;
2258 
2259 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2260 	    ("FFS background buffer should not get here %p", bp));
2261 
2262 	vp = bp->b_vp;
2263 	if (vp)
2264 		vp_md = vp->v_vflag & VV_MD;
2265 	else
2266 		vp_md = 0;
2267 
2268 	/*
2269 	 * Mark the buffer clean.  Increment the bufobj write count
2270 	 * before bundirty() call, to prevent other thread from seeing
2271 	 * empty dirty list and zero counter for writes in progress,
2272 	 * falsely indicating that the bufobj is clean.
2273 	 */
2274 	bufobj_wref(bp->b_bufobj);
2275 	bundirty(bp);
2276 
2277 	bp->b_flags &= ~B_DONE;
2278 	bp->b_ioflags &= ~BIO_ERROR;
2279 	bp->b_flags |= B_CACHE;
2280 	bp->b_iocmd = BIO_WRITE;
2281 
2282 	vfs_busy_pages(bp, 1);
2283 
2284 	/*
2285 	 * Normal bwrites pipeline writes
2286 	 */
2287 	bp->b_runningbufspace = bp->b_bufsize;
2288 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2289 
2290 #ifdef RACCT
2291 	if (racct_enable) {
2292 		PROC_LOCK(curproc);
2293 		racct_add_buf(curproc, bp, 1);
2294 		PROC_UNLOCK(curproc);
2295 	}
2296 #endif /* RACCT */
2297 	curthread->td_ru.ru_oublock++;
2298 	if (oldflags & B_ASYNC)
2299 		BUF_KERNPROC(bp);
2300 	bp->b_iooffset = dbtob(bp->b_blkno);
2301 	buf_track(bp, __func__);
2302 	bstrategy(bp);
2303 
2304 	if ((oldflags & B_ASYNC) == 0) {
2305 		int rtval = bufwait(bp);
2306 		brelse(bp);
2307 		return (rtval);
2308 	} else if (space > hirunningspace) {
2309 		/*
2310 		 * don't allow the async write to saturate the I/O
2311 		 * system.  We will not deadlock here because
2312 		 * we are blocking waiting for I/O that is already in-progress
2313 		 * to complete. We do not block here if it is the update
2314 		 * or syncer daemon trying to clean up as that can lead
2315 		 * to deadlock.
2316 		 */
2317 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2318 			waitrunningbufspace();
2319 	}
2320 
2321 	return (0);
2322 }
2323 
2324 void
2325 bufbdflush(struct bufobj *bo, struct buf *bp)
2326 {
2327 	struct buf *nbp;
2328 
2329 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2330 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2331 		altbufferflushes++;
2332 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2333 		BO_LOCK(bo);
2334 		/*
2335 		 * Try to find a buffer to flush.
2336 		 */
2337 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2338 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2339 			    BUF_LOCK(nbp,
2340 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2341 				continue;
2342 			if (bp == nbp)
2343 				panic("bdwrite: found ourselves");
2344 			BO_UNLOCK(bo);
2345 			/* Don't countdeps with the bo lock held. */
2346 			if (buf_countdeps(nbp, 0)) {
2347 				BO_LOCK(bo);
2348 				BUF_UNLOCK(nbp);
2349 				continue;
2350 			}
2351 			if (nbp->b_flags & B_CLUSTEROK) {
2352 				vfs_bio_awrite(nbp);
2353 			} else {
2354 				bremfree(nbp);
2355 				bawrite(nbp);
2356 			}
2357 			dirtybufferflushes++;
2358 			break;
2359 		}
2360 		if (nbp == NULL)
2361 			BO_UNLOCK(bo);
2362 	}
2363 }
2364 
2365 /*
2366  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2367  * anything if the buffer is marked invalid.
2368  *
2369  * Note that since the buffer must be completely valid, we can safely
2370  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2371  * biodone() in order to prevent getblk from writing the buffer
2372  * out synchronously.
2373  */
2374 void
2375 bdwrite(struct buf *bp)
2376 {
2377 	struct thread *td = curthread;
2378 	struct vnode *vp;
2379 	struct bufobj *bo;
2380 
2381 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2382 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2383 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2384 	    ("Barrier request in delayed write %p", bp));
2385 
2386 	if (bp->b_flags & B_INVAL) {
2387 		brelse(bp);
2388 		return;
2389 	}
2390 
2391 	/*
2392 	 * If we have too many dirty buffers, don't create any more.
2393 	 * If we are wildly over our limit, then force a complete
2394 	 * cleanup. Otherwise, just keep the situation from getting
2395 	 * out of control. Note that we have to avoid a recursive
2396 	 * disaster and not try to clean up after our own cleanup!
2397 	 */
2398 	vp = bp->b_vp;
2399 	bo = bp->b_bufobj;
2400 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2401 		td->td_pflags |= TDP_INBDFLUSH;
2402 		BO_BDFLUSH(bo, bp);
2403 		td->td_pflags &= ~TDP_INBDFLUSH;
2404 	} else
2405 		recursiveflushes++;
2406 
2407 	bdirty(bp);
2408 	/*
2409 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2410 	 * true even of NFS now.
2411 	 */
2412 	bp->b_flags |= B_CACHE;
2413 
2414 	/*
2415 	 * This bmap keeps the system from needing to do the bmap later,
2416 	 * perhaps when the system is attempting to do a sync.  Since it
2417 	 * is likely that the indirect block -- or whatever other datastructure
2418 	 * that the filesystem needs is still in memory now, it is a good
2419 	 * thing to do this.  Note also, that if the pageout daemon is
2420 	 * requesting a sync -- there might not be enough memory to do
2421 	 * the bmap then...  So, this is important to do.
2422 	 */
2423 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2424 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2425 	}
2426 
2427 	buf_track(bp, __func__);
2428 
2429 	/*
2430 	 * Set the *dirty* buffer range based upon the VM system dirty
2431 	 * pages.
2432 	 *
2433 	 * Mark the buffer pages as clean.  We need to do this here to
2434 	 * satisfy the vnode_pager and the pageout daemon, so that it
2435 	 * thinks that the pages have been "cleaned".  Note that since
2436 	 * the pages are in a delayed write buffer -- the VFS layer
2437 	 * "will" see that the pages get written out on the next sync,
2438 	 * or perhaps the cluster will be completed.
2439 	 */
2440 	vfs_clean_pages_dirty_buf(bp);
2441 	bqrelse(bp);
2442 
2443 	/*
2444 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2445 	 * due to the softdep code.
2446 	 */
2447 }
2448 
2449 /*
2450  *	bdirty:
2451  *
2452  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2453  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2454  *	itself to properly update it in the dirty/clean lists.  We mark it
2455  *	B_DONE to ensure that any asynchronization of the buffer properly
2456  *	clears B_DONE ( else a panic will occur later ).
2457  *
2458  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2459  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2460  *	should only be called if the buffer is known-good.
2461  *
2462  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2463  *	count.
2464  *
2465  *	The buffer must be on QUEUE_NONE.
2466  */
2467 void
2468 bdirty(struct buf *bp)
2469 {
2470 
2471 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2472 	    bp, bp->b_vp, bp->b_flags);
2473 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2474 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2475 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2476 	bp->b_flags &= ~(B_RELBUF);
2477 	bp->b_iocmd = BIO_WRITE;
2478 
2479 	if ((bp->b_flags & B_DELWRI) == 0) {
2480 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2481 		reassignbuf(bp);
2482 		bdirtyadd(bp);
2483 	}
2484 }
2485 
2486 /*
2487  *	bundirty:
2488  *
2489  *	Clear B_DELWRI for buffer.
2490  *
2491  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2492  *	count.
2493  *
2494  *	The buffer must be on QUEUE_NONE.
2495  */
2496 
2497 void
2498 bundirty(struct buf *bp)
2499 {
2500 
2501 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2502 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2503 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2504 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2505 
2506 	if (bp->b_flags & B_DELWRI) {
2507 		bp->b_flags &= ~B_DELWRI;
2508 		reassignbuf(bp);
2509 		bdirtysub(bp);
2510 	}
2511 	/*
2512 	 * Since it is now being written, we can clear its deferred write flag.
2513 	 */
2514 	bp->b_flags &= ~B_DEFERRED;
2515 }
2516 
2517 /*
2518  *	bawrite:
2519  *
2520  *	Asynchronous write.  Start output on a buffer, but do not wait for
2521  *	it to complete.  The buffer is released when the output completes.
2522  *
2523  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2524  *	B_INVAL buffers.  Not us.
2525  */
2526 void
2527 bawrite(struct buf *bp)
2528 {
2529 
2530 	bp->b_flags |= B_ASYNC;
2531 	(void) bwrite(bp);
2532 }
2533 
2534 /*
2535  *	babarrierwrite:
2536  *
2537  *	Asynchronous barrier write.  Start output on a buffer, but do not
2538  *	wait for it to complete.  Place a write barrier after this write so
2539  *	that this buffer and all buffers written before it are committed to
2540  *	the disk before any buffers written after this write are committed
2541  *	to the disk.  The buffer is released when the output completes.
2542  */
2543 void
2544 babarrierwrite(struct buf *bp)
2545 {
2546 
2547 	bp->b_flags |= B_ASYNC | B_BARRIER;
2548 	(void) bwrite(bp);
2549 }
2550 
2551 /*
2552  *	bbarrierwrite:
2553  *
2554  *	Synchronous barrier write.  Start output on a buffer and wait for
2555  *	it to complete.  Place a write barrier after this write so that
2556  *	this buffer and all buffers written before it are committed to
2557  *	the disk before any buffers written after this write are committed
2558  *	to the disk.  The buffer is released when the output completes.
2559  */
2560 int
2561 bbarrierwrite(struct buf *bp)
2562 {
2563 
2564 	bp->b_flags |= B_BARRIER;
2565 	return (bwrite(bp));
2566 }
2567 
2568 /*
2569  *	bwillwrite:
2570  *
2571  *	Called prior to the locking of any vnodes when we are expecting to
2572  *	write.  We do not want to starve the buffer cache with too many
2573  *	dirty buffers so we block here.  By blocking prior to the locking
2574  *	of any vnodes we attempt to avoid the situation where a locked vnode
2575  *	prevents the various system daemons from flushing related buffers.
2576  */
2577 void
2578 bwillwrite(void)
2579 {
2580 
2581 	if (buf_dirty_count_severe()) {
2582 		mtx_lock(&bdirtylock);
2583 		while (buf_dirty_count_severe()) {
2584 			bdirtywait = 1;
2585 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2586 			    "flswai", 0);
2587 		}
2588 		mtx_unlock(&bdirtylock);
2589 	}
2590 }
2591 
2592 /*
2593  * Return true if we have too many dirty buffers.
2594  */
2595 int
2596 buf_dirty_count_severe(void)
2597 {
2598 
2599 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2600 }
2601 
2602 /*
2603  *	brelse:
2604  *
2605  *	Release a busy buffer and, if requested, free its resources.  The
2606  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2607  *	to be accessed later as a cache entity or reused for other purposes.
2608  */
2609 void
2610 brelse(struct buf *bp)
2611 {
2612 	struct mount *v_mnt;
2613 	int qindex;
2614 
2615 	/*
2616 	 * Many functions erroneously call brelse with a NULL bp under rare
2617 	 * error conditions. Simply return when called with a NULL bp.
2618 	 */
2619 	if (bp == NULL)
2620 		return;
2621 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2622 	    bp, bp->b_vp, bp->b_flags);
2623 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2624 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2625 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2626 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2627 
2628 	if (BUF_LOCKRECURSED(bp)) {
2629 		/*
2630 		 * Do not process, in particular, do not handle the
2631 		 * B_INVAL/B_RELBUF and do not release to free list.
2632 		 */
2633 		BUF_UNLOCK(bp);
2634 		return;
2635 	}
2636 
2637 	if (bp->b_flags & B_MANAGED) {
2638 		bqrelse(bp);
2639 		return;
2640 	}
2641 
2642 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2643 		BO_LOCK(bp->b_bufobj);
2644 		bp->b_vflags &= ~BV_BKGRDERR;
2645 		BO_UNLOCK(bp->b_bufobj);
2646 		bdirty(bp);
2647 	}
2648 
2649 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2650 	    (bp->b_flags & B_INVALONERR)) {
2651 		/*
2652 		 * Forced invalidation of dirty buffer contents, to be used
2653 		 * after a failed write in the rare case that the loss of the
2654 		 * contents is acceptable.  The buffer is invalidated and
2655 		 * freed.
2656 		 */
2657 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2658 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2659 	}
2660 
2661 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2662 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2663 	    !(bp->b_flags & B_INVAL)) {
2664 		/*
2665 		 * Failed write, redirty.  All errors except ENXIO (which
2666 		 * means the device is gone) are treated as being
2667 		 * transient.
2668 		 *
2669 		 * XXX Treating EIO as transient is not correct; the
2670 		 * contract with the local storage device drivers is that
2671 		 * they will only return EIO once the I/O is no longer
2672 		 * retriable.  Network I/O also respects this through the
2673 		 * guarantees of TCP and/or the internal retries of NFS.
2674 		 * ENOMEM might be transient, but we also have no way of
2675 		 * knowing when its ok to retry/reschedule.  In general,
2676 		 * this entire case should be made obsolete through better
2677 		 * error handling/recovery and resource scheduling.
2678 		 *
2679 		 * Do this also for buffers that failed with ENXIO, but have
2680 		 * non-empty dependencies - the soft updates code might need
2681 		 * to access the buffer to untangle them.
2682 		 *
2683 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2684 		 */
2685 		bp->b_ioflags &= ~BIO_ERROR;
2686 		bdirty(bp);
2687 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2688 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2689 		/*
2690 		 * Either a failed read I/O, or we were asked to free or not
2691 		 * cache the buffer, or we failed to write to a device that's
2692 		 * no longer present.
2693 		 */
2694 		bp->b_flags |= B_INVAL;
2695 		if (!LIST_EMPTY(&bp->b_dep))
2696 			buf_deallocate(bp);
2697 		if (bp->b_flags & B_DELWRI)
2698 			bdirtysub(bp);
2699 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2700 		if ((bp->b_flags & B_VMIO) == 0) {
2701 			allocbuf(bp, 0);
2702 			if (bp->b_vp)
2703 				brelvp(bp);
2704 		}
2705 	}
2706 
2707 	/*
2708 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2709 	 * is called with B_DELWRI set, the underlying pages may wind up
2710 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2711 	 * because pages associated with a B_DELWRI bp are marked clean.
2712 	 *
2713 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2714 	 * if B_DELWRI is set.
2715 	 */
2716 	if (bp->b_flags & B_DELWRI)
2717 		bp->b_flags &= ~B_RELBUF;
2718 
2719 	/*
2720 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2721 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2722 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2723 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2724 	 *
2725 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2726 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2727 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2728 	 *
2729 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2730 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2731 	 * the commit state and we cannot afford to lose the buffer. If the
2732 	 * buffer has a background write in progress, we need to keep it
2733 	 * around to prevent it from being reconstituted and starting a second
2734 	 * background write.
2735 	 */
2736 
2737 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2738 
2739 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2740 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2741 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2742 	    vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2743 		vfs_vmio_invalidate(bp);
2744 		allocbuf(bp, 0);
2745 	}
2746 
2747 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2748 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2749 		allocbuf(bp, 0);
2750 		bp->b_flags &= ~B_NOREUSE;
2751 		if (bp->b_vp != NULL)
2752 			brelvp(bp);
2753 	}
2754 
2755 	/*
2756 	 * If the buffer has junk contents signal it and eventually
2757 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2758 	 * doesn't find it.
2759 	 */
2760 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2761 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2762 		bp->b_flags |= B_INVAL;
2763 	if (bp->b_flags & B_INVAL) {
2764 		if (bp->b_flags & B_DELWRI)
2765 			bundirty(bp);
2766 		if (bp->b_vp)
2767 			brelvp(bp);
2768 	}
2769 
2770 	buf_track(bp, __func__);
2771 
2772 	/* buffers with no memory */
2773 	if (bp->b_bufsize == 0) {
2774 		buf_free(bp);
2775 		return;
2776 	}
2777 	/* buffers with junk contents */
2778 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2779 	    (bp->b_ioflags & BIO_ERROR)) {
2780 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2781 		if (bp->b_vflags & BV_BKGRDINPROG)
2782 			panic("losing buffer 2");
2783 		qindex = QUEUE_CLEAN;
2784 		bp->b_flags |= B_AGE;
2785 	/* remaining buffers */
2786 	} else if (bp->b_flags & B_DELWRI)
2787 		qindex = QUEUE_DIRTY;
2788 	else
2789 		qindex = QUEUE_CLEAN;
2790 
2791 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2792 		panic("brelse: not dirty");
2793 
2794 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2795 	bp->b_xflags &= ~(BX_CVTENXIO);
2796 	/* binsfree unlocks bp. */
2797 	binsfree(bp, qindex);
2798 }
2799 
2800 /*
2801  * Release a buffer back to the appropriate queue but do not try to free
2802  * it.  The buffer is expected to be used again soon.
2803  *
2804  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2805  * biodone() to requeue an async I/O on completion.  It is also used when
2806  * known good buffers need to be requeued but we think we may need the data
2807  * again soon.
2808  *
2809  * XXX we should be able to leave the B_RELBUF hint set on completion.
2810  */
2811 void
2812 bqrelse(struct buf *bp)
2813 {
2814 	int qindex;
2815 
2816 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2817 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2818 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2819 
2820 	qindex = QUEUE_NONE;
2821 	if (BUF_LOCKRECURSED(bp)) {
2822 		/* do not release to free list */
2823 		BUF_UNLOCK(bp);
2824 		return;
2825 	}
2826 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2827 	bp->b_xflags &= ~(BX_CVTENXIO);
2828 
2829 	if (bp->b_flags & B_MANAGED) {
2830 		if (bp->b_flags & B_REMFREE)
2831 			bremfreef(bp);
2832 		goto out;
2833 	}
2834 
2835 	/* buffers with stale but valid contents */
2836 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2837 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2838 		BO_LOCK(bp->b_bufobj);
2839 		bp->b_vflags &= ~BV_BKGRDERR;
2840 		BO_UNLOCK(bp->b_bufobj);
2841 		qindex = QUEUE_DIRTY;
2842 	} else {
2843 		if ((bp->b_flags & B_DELWRI) == 0 &&
2844 		    (bp->b_xflags & BX_VNDIRTY))
2845 			panic("bqrelse: not dirty");
2846 		if ((bp->b_flags & B_NOREUSE) != 0) {
2847 			brelse(bp);
2848 			return;
2849 		}
2850 		qindex = QUEUE_CLEAN;
2851 	}
2852 	buf_track(bp, __func__);
2853 	/* binsfree unlocks bp. */
2854 	binsfree(bp, qindex);
2855 	return;
2856 
2857 out:
2858 	buf_track(bp, __func__);
2859 	/* unlock */
2860 	BUF_UNLOCK(bp);
2861 }
2862 
2863 /*
2864  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2865  * restore bogus pages.
2866  */
2867 static void
2868 vfs_vmio_iodone(struct buf *bp)
2869 {
2870 	vm_ooffset_t foff;
2871 	vm_page_t m;
2872 	vm_object_t obj;
2873 	struct vnode *vp __unused;
2874 	int i, iosize, resid;
2875 	bool bogus;
2876 
2877 	obj = bp->b_bufobj->bo_object;
2878 	KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2879 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2880 	    blockcount_read(&obj->paging_in_progress), bp->b_npages));
2881 
2882 	vp = bp->b_vp;
2883 	VNPASS(vp->v_holdcnt > 0, vp);
2884 	VNPASS(vp->v_object != NULL, vp);
2885 
2886 	foff = bp->b_offset;
2887 	KASSERT(bp->b_offset != NOOFFSET,
2888 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2889 
2890 	bogus = false;
2891 	iosize = bp->b_bcount - bp->b_resid;
2892 	for (i = 0; i < bp->b_npages; i++) {
2893 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2894 		if (resid > iosize)
2895 			resid = iosize;
2896 
2897 		/*
2898 		 * cleanup bogus pages, restoring the originals
2899 		 */
2900 		m = bp->b_pages[i];
2901 		if (m == bogus_page) {
2902 			bogus = true;
2903 			m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2904 			if (m == NULL)
2905 				panic("biodone: page disappeared!");
2906 			bp->b_pages[i] = m;
2907 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2908 			/*
2909 			 * In the write case, the valid and clean bits are
2910 			 * already changed correctly ( see bdwrite() ), so we
2911 			 * only need to do this here in the read case.
2912 			 */
2913 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2914 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2915 			    "has unexpected dirty bits", m));
2916 			vfs_page_set_valid(bp, foff, m);
2917 		}
2918 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2919 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2920 		    (intmax_t)foff, (uintmax_t)m->pindex));
2921 
2922 		vm_page_sunbusy(m);
2923 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2924 		iosize -= resid;
2925 	}
2926 	vm_object_pip_wakeupn(obj, bp->b_npages);
2927 	if (bogus && buf_mapped(bp)) {
2928 		BUF_CHECK_MAPPED(bp);
2929 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2930 		    bp->b_pages, bp->b_npages);
2931 	}
2932 }
2933 
2934 /*
2935  * Perform page invalidation when a buffer is released.  The fully invalid
2936  * pages will be reclaimed later in vfs_vmio_truncate().
2937  */
2938 static void
2939 vfs_vmio_invalidate(struct buf *bp)
2940 {
2941 	vm_object_t obj;
2942 	vm_page_t m;
2943 	int flags, i, resid, poffset, presid;
2944 
2945 	if (buf_mapped(bp)) {
2946 		BUF_CHECK_MAPPED(bp);
2947 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2948 	} else
2949 		BUF_CHECK_UNMAPPED(bp);
2950 	/*
2951 	 * Get the base offset and length of the buffer.  Note that
2952 	 * in the VMIO case if the buffer block size is not
2953 	 * page-aligned then b_data pointer may not be page-aligned.
2954 	 * But our b_pages[] array *IS* page aligned.
2955 	 *
2956 	 * block sizes less then DEV_BSIZE (usually 512) are not
2957 	 * supported due to the page granularity bits (m->valid,
2958 	 * m->dirty, etc...).
2959 	 *
2960 	 * See man buf(9) for more information
2961 	 */
2962 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2963 	obj = bp->b_bufobj->bo_object;
2964 	resid = bp->b_bufsize;
2965 	poffset = bp->b_offset & PAGE_MASK;
2966 	VM_OBJECT_WLOCK(obj);
2967 	for (i = 0; i < bp->b_npages; i++) {
2968 		m = bp->b_pages[i];
2969 		if (m == bogus_page)
2970 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2971 		bp->b_pages[i] = NULL;
2972 
2973 		presid = resid > (PAGE_SIZE - poffset) ?
2974 		    (PAGE_SIZE - poffset) : resid;
2975 		KASSERT(presid >= 0, ("brelse: extra page"));
2976 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
2977 		if (pmap_page_wired_mappings(m) == 0)
2978 			vm_page_set_invalid(m, poffset, presid);
2979 		vm_page_sunbusy(m);
2980 		vm_page_release_locked(m, flags);
2981 		resid -= presid;
2982 		poffset = 0;
2983 	}
2984 	VM_OBJECT_WUNLOCK(obj);
2985 	bp->b_npages = 0;
2986 }
2987 
2988 /*
2989  * Page-granular truncation of an existing VMIO buffer.
2990  */
2991 static void
2992 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2993 {
2994 	vm_object_t obj;
2995 	vm_page_t m;
2996 	int flags, i;
2997 
2998 	if (bp->b_npages == desiredpages)
2999 		return;
3000 
3001 	if (buf_mapped(bp)) {
3002 		BUF_CHECK_MAPPED(bp);
3003 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3004 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3005 	} else
3006 		BUF_CHECK_UNMAPPED(bp);
3007 
3008 	/*
3009 	 * The object lock is needed only if we will attempt to free pages.
3010 	 */
3011 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3012 	if ((bp->b_flags & B_DIRECT) != 0) {
3013 		flags |= VPR_TRYFREE;
3014 		obj = bp->b_bufobj->bo_object;
3015 		VM_OBJECT_WLOCK(obj);
3016 	} else {
3017 		obj = NULL;
3018 	}
3019 	for (i = desiredpages; i < bp->b_npages; i++) {
3020 		m = bp->b_pages[i];
3021 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3022 		bp->b_pages[i] = NULL;
3023 		if (obj != NULL)
3024 			vm_page_release_locked(m, flags);
3025 		else
3026 			vm_page_release(m, flags);
3027 	}
3028 	if (obj != NULL)
3029 		VM_OBJECT_WUNLOCK(obj);
3030 	bp->b_npages = desiredpages;
3031 }
3032 
3033 /*
3034  * Byte granular extension of VMIO buffers.
3035  */
3036 static void
3037 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3038 {
3039 	/*
3040 	 * We are growing the buffer, possibly in a
3041 	 * byte-granular fashion.
3042 	 */
3043 	vm_object_t obj;
3044 	vm_offset_t toff;
3045 	vm_offset_t tinc;
3046 	vm_page_t m;
3047 
3048 	/*
3049 	 * Step 1, bring in the VM pages from the object, allocating
3050 	 * them if necessary.  We must clear B_CACHE if these pages
3051 	 * are not valid for the range covered by the buffer.
3052 	 */
3053 	obj = bp->b_bufobj->bo_object;
3054 	if (bp->b_npages < desiredpages) {
3055 		KASSERT(desiredpages <= atop(maxbcachebuf),
3056 		    ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3057 		    bp, desiredpages, maxbcachebuf));
3058 
3059 		/*
3060 		 * We must allocate system pages since blocking
3061 		 * here could interfere with paging I/O, no
3062 		 * matter which process we are.
3063 		 *
3064 		 * Only exclusive busy can be tested here.
3065 		 * Blocking on shared busy might lead to
3066 		 * deadlocks once allocbuf() is called after
3067 		 * pages are vfs_busy_pages().
3068 		 */
3069 		(void)vm_page_grab_pages_unlocked(obj,
3070 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3071 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3072 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3073 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3074 		bp->b_npages = desiredpages;
3075 	}
3076 
3077 	/*
3078 	 * Step 2.  We've loaded the pages into the buffer,
3079 	 * we have to figure out if we can still have B_CACHE
3080 	 * set.  Note that B_CACHE is set according to the
3081 	 * byte-granular range ( bcount and size ), not the
3082 	 * aligned range ( newbsize ).
3083 	 *
3084 	 * The VM test is against m->valid, which is DEV_BSIZE
3085 	 * aligned.  Needless to say, the validity of the data
3086 	 * needs to also be DEV_BSIZE aligned.  Note that this
3087 	 * fails with NFS if the server or some other client
3088 	 * extends the file's EOF.  If our buffer is resized,
3089 	 * B_CACHE may remain set! XXX
3090 	 */
3091 	toff = bp->b_bcount;
3092 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3093 	while ((bp->b_flags & B_CACHE) && toff < size) {
3094 		vm_pindex_t pi;
3095 
3096 		if (tinc > (size - toff))
3097 			tinc = size - toff;
3098 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3099 		m = bp->b_pages[pi];
3100 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3101 		toff += tinc;
3102 		tinc = PAGE_SIZE;
3103 	}
3104 
3105 	/*
3106 	 * Step 3, fixup the KVA pmap.
3107 	 */
3108 	if (buf_mapped(bp))
3109 		bpmap_qenter(bp);
3110 	else
3111 		BUF_CHECK_UNMAPPED(bp);
3112 }
3113 
3114 /*
3115  * Check to see if a block at a particular lbn is available for a clustered
3116  * write.
3117  */
3118 static int
3119 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3120 {
3121 	struct buf *bpa;
3122 	int match;
3123 
3124 	match = 0;
3125 
3126 	/* If the buf isn't in core skip it */
3127 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3128 		return (0);
3129 
3130 	/* If the buf is busy we don't want to wait for it */
3131 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3132 		return (0);
3133 
3134 	/* Only cluster with valid clusterable delayed write buffers */
3135 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3136 	    (B_DELWRI | B_CLUSTEROK))
3137 		goto done;
3138 
3139 	if (bpa->b_bufsize != size)
3140 		goto done;
3141 
3142 	/*
3143 	 * Check to see if it is in the expected place on disk and that the
3144 	 * block has been mapped.
3145 	 */
3146 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3147 		match = 1;
3148 done:
3149 	BUF_UNLOCK(bpa);
3150 	return (match);
3151 }
3152 
3153 /*
3154  *	vfs_bio_awrite:
3155  *
3156  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3157  *	This is much better then the old way of writing only one buffer at
3158  *	a time.  Note that we may not be presented with the buffers in the
3159  *	correct order, so we search for the cluster in both directions.
3160  */
3161 int
3162 vfs_bio_awrite(struct buf *bp)
3163 {
3164 	struct bufobj *bo;
3165 	int i;
3166 	int j;
3167 	daddr_t lblkno = bp->b_lblkno;
3168 	struct vnode *vp = bp->b_vp;
3169 	int ncl;
3170 	int nwritten;
3171 	int size;
3172 	int maxcl;
3173 	int gbflags;
3174 
3175 	bo = &vp->v_bufobj;
3176 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3177 	/*
3178 	 * right now we support clustered writing only to regular files.  If
3179 	 * we find a clusterable block we could be in the middle of a cluster
3180 	 * rather then at the beginning.
3181 	 */
3182 	if ((vp->v_type == VREG) &&
3183 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3184 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3185 		size = vp->v_mount->mnt_stat.f_iosize;
3186 		maxcl = maxphys / size;
3187 
3188 		BO_RLOCK(bo);
3189 		for (i = 1; i < maxcl; i++)
3190 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3191 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3192 				break;
3193 
3194 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3195 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3196 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3197 				break;
3198 		BO_RUNLOCK(bo);
3199 		--j;
3200 		ncl = i + j;
3201 		/*
3202 		 * this is a possible cluster write
3203 		 */
3204 		if (ncl != 1) {
3205 			BUF_UNLOCK(bp);
3206 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3207 			    gbflags);
3208 			return (nwritten);
3209 		}
3210 	}
3211 	bremfree(bp);
3212 	bp->b_flags |= B_ASYNC;
3213 	/*
3214 	 * default (old) behavior, writing out only one block
3215 	 *
3216 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3217 	 */
3218 	nwritten = bp->b_bufsize;
3219 	(void) bwrite(bp);
3220 
3221 	return (nwritten);
3222 }
3223 
3224 /*
3225  *	getnewbuf_kva:
3226  *
3227  *	Allocate KVA for an empty buf header according to gbflags.
3228  */
3229 static int
3230 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3231 {
3232 
3233 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3234 		/*
3235 		 * In order to keep fragmentation sane we only allocate kva
3236 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3237 		 */
3238 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3239 
3240 		if (maxsize != bp->b_kvasize &&
3241 		    bufkva_alloc(bp, maxsize, gbflags))
3242 			return (ENOSPC);
3243 	}
3244 	return (0);
3245 }
3246 
3247 /*
3248  *	getnewbuf:
3249  *
3250  *	Find and initialize a new buffer header, freeing up existing buffers
3251  *	in the bufqueues as necessary.  The new buffer is returned locked.
3252  *
3253  *	We block if:
3254  *		We have insufficient buffer headers
3255  *		We have insufficient buffer space
3256  *		buffer_arena is too fragmented ( space reservation fails )
3257  *		If we have to flush dirty buffers ( but we try to avoid this )
3258  *
3259  *	The caller is responsible for releasing the reserved bufspace after
3260  *	allocbuf() is called.
3261  */
3262 static struct buf *
3263 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3264 {
3265 	struct bufdomain *bd;
3266 	struct buf *bp;
3267 	bool metadata, reserved;
3268 
3269 	bp = NULL;
3270 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3271 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3272 	if (!unmapped_buf_allowed)
3273 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3274 
3275 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3276 	    vp->v_type == VCHR)
3277 		metadata = true;
3278 	else
3279 		metadata = false;
3280 	if (vp == NULL)
3281 		bd = &bdomain[0];
3282 	else
3283 		bd = &bdomain[vp->v_bufobj.bo_domain];
3284 
3285 	counter_u64_add(getnewbufcalls, 1);
3286 	reserved = false;
3287 	do {
3288 		if (reserved == false &&
3289 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3290 			counter_u64_add(getnewbufrestarts, 1);
3291 			continue;
3292 		}
3293 		reserved = true;
3294 		if ((bp = buf_alloc(bd)) == NULL) {
3295 			counter_u64_add(getnewbufrestarts, 1);
3296 			continue;
3297 		}
3298 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3299 			return (bp);
3300 		break;
3301 	} while (buf_recycle(bd, false) == 0);
3302 
3303 	if (reserved)
3304 		bufspace_release(bd, maxsize);
3305 	if (bp != NULL) {
3306 		bp->b_flags |= B_INVAL;
3307 		brelse(bp);
3308 	}
3309 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3310 
3311 	return (NULL);
3312 }
3313 
3314 /*
3315  *	buf_daemon:
3316  *
3317  *	buffer flushing daemon.  Buffers are normally flushed by the
3318  *	update daemon but if it cannot keep up this process starts to
3319  *	take the load in an attempt to prevent getnewbuf() from blocking.
3320  */
3321 static struct kproc_desc buf_kp = {
3322 	"bufdaemon",
3323 	buf_daemon,
3324 	&bufdaemonproc
3325 };
3326 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3327 
3328 static int
3329 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3330 {
3331 	int flushed;
3332 
3333 	flushed = flushbufqueues(vp, bd, target, 0);
3334 	if (flushed == 0) {
3335 		/*
3336 		 * Could not find any buffers without rollback
3337 		 * dependencies, so just write the first one
3338 		 * in the hopes of eventually making progress.
3339 		 */
3340 		if (vp != NULL && target > 2)
3341 			target /= 2;
3342 		flushbufqueues(vp, bd, target, 1);
3343 	}
3344 	return (flushed);
3345 }
3346 
3347 static void
3348 buf_daemon()
3349 {
3350 	struct bufdomain *bd;
3351 	int speedupreq;
3352 	int lodirty;
3353 	int i;
3354 
3355 	/*
3356 	 * This process needs to be suspended prior to shutdown sync.
3357 	 */
3358 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3359 	    SHUTDOWN_PRI_LAST + 100);
3360 
3361 	/*
3362 	 * Start the buf clean daemons as children threads.
3363 	 */
3364 	for (i = 0 ; i < buf_domains; i++) {
3365 		int error;
3366 
3367 		error = kthread_add((void (*)(void *))bufspace_daemon,
3368 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3369 		if (error)
3370 			panic("error %d spawning bufspace daemon", error);
3371 	}
3372 
3373 	/*
3374 	 * This process is allowed to take the buffer cache to the limit
3375 	 */
3376 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3377 	mtx_lock(&bdlock);
3378 	for (;;) {
3379 		bd_request = 0;
3380 		mtx_unlock(&bdlock);
3381 
3382 		kthread_suspend_check();
3383 
3384 		/*
3385 		 * Save speedupreq for this pass and reset to capture new
3386 		 * requests.
3387 		 */
3388 		speedupreq = bd_speedupreq;
3389 		bd_speedupreq = 0;
3390 
3391 		/*
3392 		 * Flush each domain sequentially according to its level and
3393 		 * the speedup request.
3394 		 */
3395 		for (i = 0; i < buf_domains; i++) {
3396 			bd = &bdomain[i];
3397 			if (speedupreq)
3398 				lodirty = bd->bd_numdirtybuffers / 2;
3399 			else
3400 				lodirty = bd->bd_lodirtybuffers;
3401 			while (bd->bd_numdirtybuffers > lodirty) {
3402 				if (buf_flush(NULL, bd,
3403 				    bd->bd_numdirtybuffers - lodirty) == 0)
3404 					break;
3405 				kern_yield(PRI_USER);
3406 			}
3407 		}
3408 
3409 		/*
3410 		 * Only clear bd_request if we have reached our low water
3411 		 * mark.  The buf_daemon normally waits 1 second and
3412 		 * then incrementally flushes any dirty buffers that have
3413 		 * built up, within reason.
3414 		 *
3415 		 * If we were unable to hit our low water mark and couldn't
3416 		 * find any flushable buffers, we sleep for a short period
3417 		 * to avoid endless loops on unlockable buffers.
3418 		 */
3419 		mtx_lock(&bdlock);
3420 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3421 			/*
3422 			 * We reached our low water mark, reset the
3423 			 * request and sleep until we are needed again.
3424 			 * The sleep is just so the suspend code works.
3425 			 */
3426 			bd_request = 0;
3427 			/*
3428 			 * Do an extra wakeup in case dirty threshold
3429 			 * changed via sysctl and the explicit transition
3430 			 * out of shortfall was missed.
3431 			 */
3432 			bdirtywakeup();
3433 			if (runningbufspace <= lorunningspace)
3434 				runningwakeup();
3435 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3436 		} else {
3437 			/*
3438 			 * We couldn't find any flushable dirty buffers but
3439 			 * still have too many dirty buffers, we
3440 			 * have to sleep and try again.  (rare)
3441 			 */
3442 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3443 		}
3444 	}
3445 }
3446 
3447 /*
3448  *	flushbufqueues:
3449  *
3450  *	Try to flush a buffer in the dirty queue.  We must be careful to
3451  *	free up B_INVAL buffers instead of write them, which NFS is
3452  *	particularly sensitive to.
3453  */
3454 static int flushwithdeps = 0;
3455 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3456     &flushwithdeps, 0,
3457     "Number of buffers flushed with dependecies that require rollbacks");
3458 
3459 static int
3460 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3461     int flushdeps)
3462 {
3463 	struct bufqueue *bq;
3464 	struct buf *sentinel;
3465 	struct vnode *vp;
3466 	struct mount *mp;
3467 	struct buf *bp;
3468 	int hasdeps;
3469 	int flushed;
3470 	int error;
3471 	bool unlock;
3472 
3473 	flushed = 0;
3474 	bq = &bd->bd_dirtyq;
3475 	bp = NULL;
3476 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3477 	sentinel->b_qindex = QUEUE_SENTINEL;
3478 	BQ_LOCK(bq);
3479 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3480 	BQ_UNLOCK(bq);
3481 	while (flushed != target) {
3482 		maybe_yield();
3483 		BQ_LOCK(bq);
3484 		bp = TAILQ_NEXT(sentinel, b_freelist);
3485 		if (bp != NULL) {
3486 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3487 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3488 			    b_freelist);
3489 		} else {
3490 			BQ_UNLOCK(bq);
3491 			break;
3492 		}
3493 		/*
3494 		 * Skip sentinels inserted by other invocations of the
3495 		 * flushbufqueues(), taking care to not reorder them.
3496 		 *
3497 		 * Only flush the buffers that belong to the
3498 		 * vnode locked by the curthread.
3499 		 */
3500 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3501 		    bp->b_vp != lvp)) {
3502 			BQ_UNLOCK(bq);
3503 			continue;
3504 		}
3505 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3506 		BQ_UNLOCK(bq);
3507 		if (error != 0)
3508 			continue;
3509 
3510 		/*
3511 		 * BKGRDINPROG can only be set with the buf and bufobj
3512 		 * locks both held.  We tolerate a race to clear it here.
3513 		 */
3514 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3515 		    (bp->b_flags & B_DELWRI) == 0) {
3516 			BUF_UNLOCK(bp);
3517 			continue;
3518 		}
3519 		if (bp->b_flags & B_INVAL) {
3520 			bremfreef(bp);
3521 			brelse(bp);
3522 			flushed++;
3523 			continue;
3524 		}
3525 
3526 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3527 			if (flushdeps == 0) {
3528 				BUF_UNLOCK(bp);
3529 				continue;
3530 			}
3531 			hasdeps = 1;
3532 		} else
3533 			hasdeps = 0;
3534 		/*
3535 		 * We must hold the lock on a vnode before writing
3536 		 * one of its buffers. Otherwise we may confuse, or
3537 		 * in the case of a snapshot vnode, deadlock the
3538 		 * system.
3539 		 *
3540 		 * The lock order here is the reverse of the normal
3541 		 * of vnode followed by buf lock.  This is ok because
3542 		 * the NOWAIT will prevent deadlock.
3543 		 */
3544 		vp = bp->b_vp;
3545 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3546 			BUF_UNLOCK(bp);
3547 			continue;
3548 		}
3549 		if (lvp == NULL) {
3550 			unlock = true;
3551 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3552 		} else {
3553 			ASSERT_VOP_LOCKED(vp, "getbuf");
3554 			unlock = false;
3555 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3556 			    vn_lock(vp, LK_TRYUPGRADE);
3557 		}
3558 		if (error == 0) {
3559 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3560 			    bp, bp->b_vp, bp->b_flags);
3561 			if (curproc == bufdaemonproc) {
3562 				vfs_bio_awrite(bp);
3563 			} else {
3564 				bremfree(bp);
3565 				bwrite(bp);
3566 				counter_u64_add(notbufdflushes, 1);
3567 			}
3568 			vn_finished_write(mp);
3569 			if (unlock)
3570 				VOP_UNLOCK(vp);
3571 			flushwithdeps += hasdeps;
3572 			flushed++;
3573 
3574 			/*
3575 			 * Sleeping on runningbufspace while holding
3576 			 * vnode lock leads to deadlock.
3577 			 */
3578 			if (curproc == bufdaemonproc &&
3579 			    runningbufspace > hirunningspace)
3580 				waitrunningbufspace();
3581 			continue;
3582 		}
3583 		vn_finished_write(mp);
3584 		BUF_UNLOCK(bp);
3585 	}
3586 	BQ_LOCK(bq);
3587 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3588 	BQ_UNLOCK(bq);
3589 	free(sentinel, M_TEMP);
3590 	return (flushed);
3591 }
3592 
3593 /*
3594  * Check to see if a block is currently memory resident.
3595  */
3596 struct buf *
3597 incore(struct bufobj *bo, daddr_t blkno)
3598 {
3599 	return (gbincore_unlocked(bo, blkno));
3600 }
3601 
3602 /*
3603  * Returns true if no I/O is needed to access the
3604  * associated VM object.  This is like incore except
3605  * it also hunts around in the VM system for the data.
3606  */
3607 bool
3608 inmem(struct vnode * vp, daddr_t blkno)
3609 {
3610 	vm_object_t obj;
3611 	vm_offset_t toff, tinc, size;
3612 	vm_page_t m, n;
3613 	vm_ooffset_t off;
3614 	int valid;
3615 
3616 	ASSERT_VOP_LOCKED(vp, "inmem");
3617 
3618 	if (incore(&vp->v_bufobj, blkno))
3619 		return (true);
3620 	if (vp->v_mount == NULL)
3621 		return (false);
3622 	obj = vp->v_object;
3623 	if (obj == NULL)
3624 		return (false);
3625 
3626 	size = PAGE_SIZE;
3627 	if (size > vp->v_mount->mnt_stat.f_iosize)
3628 		size = vp->v_mount->mnt_stat.f_iosize;
3629 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3630 
3631 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3632 		m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3633 recheck:
3634 		if (m == NULL)
3635 			return (false);
3636 
3637 		tinc = size;
3638 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3639 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3640 		/*
3641 		 * Consider page validity only if page mapping didn't change
3642 		 * during the check.
3643 		 */
3644 		valid = vm_page_is_valid(m,
3645 		    (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3646 		n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3647 		if (m != n) {
3648 			m = n;
3649 			goto recheck;
3650 		}
3651 		if (!valid)
3652 			return (false);
3653 	}
3654 	return (true);
3655 }
3656 
3657 /*
3658  * Set the dirty range for a buffer based on the status of the dirty
3659  * bits in the pages comprising the buffer.  The range is limited
3660  * to the size of the buffer.
3661  *
3662  * Tell the VM system that the pages associated with this buffer
3663  * are clean.  This is used for delayed writes where the data is
3664  * going to go to disk eventually without additional VM intevention.
3665  *
3666  * Note that while we only really need to clean through to b_bcount, we
3667  * just go ahead and clean through to b_bufsize.
3668  */
3669 static void
3670 vfs_clean_pages_dirty_buf(struct buf *bp)
3671 {
3672 	vm_ooffset_t foff, noff, eoff;
3673 	vm_page_t m;
3674 	int i;
3675 
3676 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3677 		return;
3678 
3679 	foff = bp->b_offset;
3680 	KASSERT(bp->b_offset != NOOFFSET,
3681 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3682 
3683 	vfs_busy_pages_acquire(bp);
3684 	vfs_setdirty_range(bp);
3685 	for (i = 0; i < bp->b_npages; i++) {
3686 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3687 		eoff = noff;
3688 		if (eoff > bp->b_offset + bp->b_bufsize)
3689 			eoff = bp->b_offset + bp->b_bufsize;
3690 		m = bp->b_pages[i];
3691 		vfs_page_set_validclean(bp, foff, m);
3692 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3693 		foff = noff;
3694 	}
3695 	vfs_busy_pages_release(bp);
3696 }
3697 
3698 static void
3699 vfs_setdirty_range(struct buf *bp)
3700 {
3701 	vm_offset_t boffset;
3702 	vm_offset_t eoffset;
3703 	int i;
3704 
3705 	/*
3706 	 * test the pages to see if they have been modified directly
3707 	 * by users through the VM system.
3708 	 */
3709 	for (i = 0; i < bp->b_npages; i++)
3710 		vm_page_test_dirty(bp->b_pages[i]);
3711 
3712 	/*
3713 	 * Calculate the encompassing dirty range, boffset and eoffset,
3714 	 * (eoffset - boffset) bytes.
3715 	 */
3716 
3717 	for (i = 0; i < bp->b_npages; i++) {
3718 		if (bp->b_pages[i]->dirty)
3719 			break;
3720 	}
3721 	boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3722 
3723 	for (i = bp->b_npages - 1; i >= 0; --i) {
3724 		if (bp->b_pages[i]->dirty) {
3725 			break;
3726 		}
3727 	}
3728 	eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3729 
3730 	/*
3731 	 * Fit it to the buffer.
3732 	 */
3733 
3734 	if (eoffset > bp->b_bcount)
3735 		eoffset = bp->b_bcount;
3736 
3737 	/*
3738 	 * If we have a good dirty range, merge with the existing
3739 	 * dirty range.
3740 	 */
3741 
3742 	if (boffset < eoffset) {
3743 		if (bp->b_dirtyoff > boffset)
3744 			bp->b_dirtyoff = boffset;
3745 		if (bp->b_dirtyend < eoffset)
3746 			bp->b_dirtyend = eoffset;
3747 	}
3748 }
3749 
3750 /*
3751  * Allocate the KVA mapping for an existing buffer.
3752  * If an unmapped buffer is provided but a mapped buffer is requested, take
3753  * also care to properly setup mappings between pages and KVA.
3754  */
3755 static void
3756 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3757 {
3758 	int bsize, maxsize, need_mapping, need_kva;
3759 	off_t offset;
3760 
3761 	need_mapping = bp->b_data == unmapped_buf &&
3762 	    (gbflags & GB_UNMAPPED) == 0;
3763 	need_kva = bp->b_kvabase == unmapped_buf &&
3764 	    bp->b_data == unmapped_buf &&
3765 	    (gbflags & GB_KVAALLOC) != 0;
3766 	if (!need_mapping && !need_kva)
3767 		return;
3768 
3769 	BUF_CHECK_UNMAPPED(bp);
3770 
3771 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3772 		/*
3773 		 * Buffer is not mapped, but the KVA was already
3774 		 * reserved at the time of the instantiation.  Use the
3775 		 * allocated space.
3776 		 */
3777 		goto has_addr;
3778 	}
3779 
3780 	/*
3781 	 * Calculate the amount of the address space we would reserve
3782 	 * if the buffer was mapped.
3783 	 */
3784 	bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3785 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3786 	offset = blkno * bsize;
3787 	maxsize = size + (offset & PAGE_MASK);
3788 	maxsize = imax(maxsize, bsize);
3789 
3790 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3791 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3792 			/*
3793 			 * XXXKIB: defragmentation cannot
3794 			 * succeed, not sure what else to do.
3795 			 */
3796 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3797 		}
3798 		counter_u64_add(mappingrestarts, 1);
3799 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3800 	}
3801 has_addr:
3802 	if (need_mapping) {
3803 		/* b_offset is handled by bpmap_qenter. */
3804 		bp->b_data = bp->b_kvabase;
3805 		BUF_CHECK_MAPPED(bp);
3806 		bpmap_qenter(bp);
3807 	}
3808 }
3809 
3810 struct buf *
3811 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3812     int flags)
3813 {
3814 	struct buf *bp;
3815 	int error;
3816 
3817 	error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3818 	if (error != 0)
3819 		return (NULL);
3820 	return (bp);
3821 }
3822 
3823 /*
3824  *	getblkx:
3825  *
3826  *	Get a block given a specified block and offset into a file/device.
3827  *	The buffers B_DONE bit will be cleared on return, making it almost
3828  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3829  *	return.  The caller should clear B_INVAL prior to initiating a
3830  *	READ.
3831  *
3832  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3833  *	an existing buffer.
3834  *
3835  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3836  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3837  *	and then cleared based on the backing VM.  If the previous buffer is
3838  *	non-0-sized but invalid, B_CACHE will be cleared.
3839  *
3840  *	If getblk() must create a new buffer, the new buffer is returned with
3841  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3842  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3843  *	backing VM.
3844  *
3845  *	getblk() also forces a bwrite() for any B_DELWRI buffer whose
3846  *	B_CACHE bit is clear.
3847  *
3848  *	What this means, basically, is that the caller should use B_CACHE to
3849  *	determine whether the buffer is fully valid or not and should clear
3850  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3851  *	the buffer by loading its data area with something, the caller needs
3852  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3853  *	the caller should set B_CACHE ( as an optimization ), else the caller
3854  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3855  *	a write attempt or if it was a successful read.  If the caller
3856  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3857  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3858  *
3859  *	The blkno parameter is the logical block being requested. Normally
3860  *	the mapping of logical block number to disk block address is done
3861  *	by calling VOP_BMAP(). However, if the mapping is already known, the
3862  *	disk block address can be passed using the dblkno parameter. If the
3863  *	disk block address is not known, then the same value should be passed
3864  *	for blkno and dblkno.
3865  */
3866 int
3867 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3868     int slptimeo, int flags, struct buf **bpp)
3869 {
3870 	struct buf *bp;
3871 	struct bufobj *bo;
3872 	daddr_t d_blkno;
3873 	int bsize, error, maxsize, vmio;
3874 	off_t offset;
3875 
3876 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3877 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3878 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3879 	ASSERT_VOP_LOCKED(vp, "getblk");
3880 	if (size > maxbcachebuf)
3881 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3882 		    maxbcachebuf);
3883 	if (!unmapped_buf_allowed)
3884 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3885 
3886 	bo = &vp->v_bufobj;
3887 	d_blkno = dblkno;
3888 
3889 	/* Attempt lockless lookup first. */
3890 	bp = gbincore_unlocked(bo, blkno);
3891 	if (bp == NULL)
3892 		goto newbuf_unlocked;
3893 
3894 	error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3895 	    0);
3896 	if (error != 0)
3897 		goto loop;
3898 
3899 	/* Verify buf identify has not changed since lookup. */
3900 	if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3901 		goto foundbuf_fastpath;
3902 
3903 	/* It changed, fallback to locked lookup. */
3904 	BUF_UNLOCK_RAW(bp);
3905 
3906 loop:
3907 	BO_RLOCK(bo);
3908 	bp = gbincore(bo, blkno);
3909 	if (bp != NULL) {
3910 		int lockflags;
3911 
3912 		/*
3913 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3914 		 * it must be on a queue.
3915 		 */
3916 		lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3917 		    ((flags & GB_LOCK_NOWAIT) ? LK_NOWAIT : LK_SLEEPFAIL);
3918 
3919 		error = BUF_TIMELOCK(bp, lockflags,
3920 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3921 
3922 		/*
3923 		 * If we slept and got the lock we have to restart in case
3924 		 * the buffer changed identities.
3925 		 */
3926 		if (error == ENOLCK)
3927 			goto loop;
3928 		/* We timed out or were interrupted. */
3929 		else if (error != 0)
3930 			return (error);
3931 
3932 foundbuf_fastpath:
3933 		/* If recursed, assume caller knows the rules. */
3934 		if (BUF_LOCKRECURSED(bp))
3935 			goto end;
3936 
3937 		/*
3938 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3939 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3940 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3941 		 * backing VM cache.
3942 		 */
3943 		if (bp->b_flags & B_INVAL)
3944 			bp->b_flags &= ~B_CACHE;
3945 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3946 			bp->b_flags |= B_CACHE;
3947 		if (bp->b_flags & B_MANAGED)
3948 			MPASS(bp->b_qindex == QUEUE_NONE);
3949 		else
3950 			bremfree(bp);
3951 
3952 		/*
3953 		 * check for size inconsistencies for non-VMIO case.
3954 		 */
3955 		if (bp->b_bcount != size) {
3956 			if ((bp->b_flags & B_VMIO) == 0 ||
3957 			    (size > bp->b_kvasize)) {
3958 				if (bp->b_flags & B_DELWRI) {
3959 					bp->b_flags |= B_NOCACHE;
3960 					bwrite(bp);
3961 				} else {
3962 					if (LIST_EMPTY(&bp->b_dep)) {
3963 						bp->b_flags |= B_RELBUF;
3964 						brelse(bp);
3965 					} else {
3966 						bp->b_flags |= B_NOCACHE;
3967 						bwrite(bp);
3968 					}
3969 				}
3970 				goto loop;
3971 			}
3972 		}
3973 
3974 		/*
3975 		 * Handle the case of unmapped buffer which should
3976 		 * become mapped, or the buffer for which KVA
3977 		 * reservation is requested.
3978 		 */
3979 		bp_unmapped_get_kva(bp, blkno, size, flags);
3980 
3981 		/*
3982 		 * If the size is inconsistent in the VMIO case, we can resize
3983 		 * the buffer.  This might lead to B_CACHE getting set or
3984 		 * cleared.  If the size has not changed, B_CACHE remains
3985 		 * unchanged from its previous state.
3986 		 */
3987 		allocbuf(bp, size);
3988 
3989 		KASSERT(bp->b_offset != NOOFFSET,
3990 		    ("getblk: no buffer offset"));
3991 
3992 		/*
3993 		 * A buffer with B_DELWRI set and B_CACHE clear must
3994 		 * be committed before we can return the buffer in
3995 		 * order to prevent the caller from issuing a read
3996 		 * ( due to B_CACHE not being set ) and overwriting
3997 		 * it.
3998 		 *
3999 		 * Most callers, including NFS and FFS, need this to
4000 		 * operate properly either because they assume they
4001 		 * can issue a read if B_CACHE is not set, or because
4002 		 * ( for example ) an uncached B_DELWRI might loop due
4003 		 * to softupdates re-dirtying the buffer.  In the latter
4004 		 * case, B_CACHE is set after the first write completes,
4005 		 * preventing further loops.
4006 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
4007 		 * above while extending the buffer, we cannot allow the
4008 		 * buffer to remain with B_CACHE set after the write
4009 		 * completes or it will represent a corrupt state.  To
4010 		 * deal with this we set B_NOCACHE to scrap the buffer
4011 		 * after the write.
4012 		 *
4013 		 * We might be able to do something fancy, like setting
4014 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
4015 		 * so the below call doesn't set B_CACHE, but that gets real
4016 		 * confusing.  This is much easier.
4017 		 */
4018 
4019 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4020 			bp->b_flags |= B_NOCACHE;
4021 			bwrite(bp);
4022 			goto loop;
4023 		}
4024 		bp->b_flags &= ~B_DONE;
4025 	} else {
4026 		/*
4027 		 * Buffer is not in-core, create new buffer.  The buffer
4028 		 * returned by getnewbuf() is locked.  Note that the returned
4029 		 * buffer is also considered valid (not marked B_INVAL).
4030 		 */
4031 		BO_RUNLOCK(bo);
4032 newbuf_unlocked:
4033 		/*
4034 		 * If the user does not want us to create the buffer, bail out
4035 		 * here.
4036 		 */
4037 		if (flags & GB_NOCREAT)
4038 			return (EEXIST);
4039 
4040 		bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4041 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4042 		offset = blkno * bsize;
4043 		vmio = vp->v_object != NULL;
4044 		if (vmio) {
4045 			maxsize = size + (offset & PAGE_MASK);
4046 		} else {
4047 			maxsize = size;
4048 			/* Do not allow non-VMIO notmapped buffers. */
4049 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4050 		}
4051 		maxsize = imax(maxsize, bsize);
4052 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4053 		    !vn_isdisk(vp)) {
4054 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4055 			KASSERT(error != EOPNOTSUPP,
4056 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4057 			    vp));
4058 			if (error != 0)
4059 				return (error);
4060 			if (d_blkno == -1)
4061 				return (EJUSTRETURN);
4062 		}
4063 
4064 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4065 		if (bp == NULL) {
4066 			if (slpflag || slptimeo)
4067 				return (ETIMEDOUT);
4068 			/*
4069 			 * XXX This is here until the sleep path is diagnosed
4070 			 * enough to work under very low memory conditions.
4071 			 *
4072 			 * There's an issue on low memory, 4BSD+non-preempt
4073 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4074 			 * exhaustion occurs without sleeping for buffer
4075 			 * reclaimation.  This just sticks in a loop and
4076 			 * constantly attempts to allocate a buffer, which
4077 			 * hits exhaustion and tries to wakeup bufdaemon.
4078 			 * This never happens because we never yield.
4079 			 *
4080 			 * The real solution is to identify and fix these cases
4081 			 * so we aren't effectively busy-waiting in a loop
4082 			 * until the reclaimation path has cycles to run.
4083 			 */
4084 			kern_yield(PRI_USER);
4085 			goto loop;
4086 		}
4087 
4088 		/*
4089 		 * This code is used to make sure that a buffer is not
4090 		 * created while the getnewbuf routine is blocked.
4091 		 * This can be a problem whether the vnode is locked or not.
4092 		 * If the buffer is created out from under us, we have to
4093 		 * throw away the one we just created.
4094 		 *
4095 		 * Note: this must occur before we associate the buffer
4096 		 * with the vp especially considering limitations in
4097 		 * the splay tree implementation when dealing with duplicate
4098 		 * lblkno's.
4099 		 */
4100 		BO_LOCK(bo);
4101 		if (gbincore(bo, blkno)) {
4102 			BO_UNLOCK(bo);
4103 			bp->b_flags |= B_INVAL;
4104 			bufspace_release(bufdomain(bp), maxsize);
4105 			brelse(bp);
4106 			goto loop;
4107 		}
4108 
4109 		/*
4110 		 * Insert the buffer into the hash, so that it can
4111 		 * be found by incore.
4112 		 */
4113 		bp->b_lblkno = blkno;
4114 		bp->b_blkno = d_blkno;
4115 		bp->b_offset = offset;
4116 		bgetvp(vp, bp);
4117 		BO_UNLOCK(bo);
4118 
4119 		/*
4120 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4121 		 * buffer size starts out as 0, B_CACHE will be set by
4122 		 * allocbuf() for the VMIO case prior to it testing the
4123 		 * backing store for validity.
4124 		 */
4125 
4126 		if (vmio) {
4127 			bp->b_flags |= B_VMIO;
4128 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4129 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4130 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4131 		} else {
4132 			bp->b_flags &= ~B_VMIO;
4133 			KASSERT(bp->b_bufobj->bo_object == NULL,
4134 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4135 			    bp, bp->b_bufobj->bo_object));
4136 			BUF_CHECK_MAPPED(bp);
4137 		}
4138 
4139 		allocbuf(bp, size);
4140 		bufspace_release(bufdomain(bp), maxsize);
4141 		bp->b_flags &= ~B_DONE;
4142 	}
4143 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4144 end:
4145 	buf_track(bp, __func__);
4146 	KASSERT(bp->b_bufobj == bo,
4147 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4148 	*bpp = bp;
4149 	return (0);
4150 }
4151 
4152 /*
4153  * Get an empty, disassociated buffer of given size.  The buffer is initially
4154  * set to B_INVAL.
4155  */
4156 struct buf *
4157 geteblk(int size, int flags)
4158 {
4159 	struct buf *bp;
4160 	int maxsize;
4161 
4162 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4163 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4164 		if ((flags & GB_NOWAIT_BD) &&
4165 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4166 			return (NULL);
4167 	}
4168 	allocbuf(bp, size);
4169 	bufspace_release(bufdomain(bp), maxsize);
4170 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4171 	return (bp);
4172 }
4173 
4174 /*
4175  * Truncate the backing store for a non-vmio buffer.
4176  */
4177 static void
4178 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4179 {
4180 
4181 	if (bp->b_flags & B_MALLOC) {
4182 		/*
4183 		 * malloced buffers are not shrunk
4184 		 */
4185 		if (newbsize == 0) {
4186 			bufmallocadjust(bp, 0);
4187 			free(bp->b_data, M_BIOBUF);
4188 			bp->b_data = bp->b_kvabase;
4189 			bp->b_flags &= ~B_MALLOC;
4190 		}
4191 		return;
4192 	}
4193 	vm_hold_free_pages(bp, newbsize);
4194 	bufspace_adjust(bp, newbsize);
4195 }
4196 
4197 /*
4198  * Extend the backing for a non-VMIO buffer.
4199  */
4200 static void
4201 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4202 {
4203 	caddr_t origbuf;
4204 	int origbufsize;
4205 
4206 	/*
4207 	 * We only use malloced memory on the first allocation.
4208 	 * and revert to page-allocated memory when the buffer
4209 	 * grows.
4210 	 *
4211 	 * There is a potential smp race here that could lead
4212 	 * to bufmallocspace slightly passing the max.  It
4213 	 * is probably extremely rare and not worth worrying
4214 	 * over.
4215 	 */
4216 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4217 	    bufmallocspace < maxbufmallocspace) {
4218 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4219 		bp->b_flags |= B_MALLOC;
4220 		bufmallocadjust(bp, newbsize);
4221 		return;
4222 	}
4223 
4224 	/*
4225 	 * If the buffer is growing on its other-than-first
4226 	 * allocation then we revert to the page-allocation
4227 	 * scheme.
4228 	 */
4229 	origbuf = NULL;
4230 	origbufsize = 0;
4231 	if (bp->b_flags & B_MALLOC) {
4232 		origbuf = bp->b_data;
4233 		origbufsize = bp->b_bufsize;
4234 		bp->b_data = bp->b_kvabase;
4235 		bufmallocadjust(bp, 0);
4236 		bp->b_flags &= ~B_MALLOC;
4237 		newbsize = round_page(newbsize);
4238 	}
4239 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4240 	    (vm_offset_t) bp->b_data + newbsize);
4241 	if (origbuf != NULL) {
4242 		bcopy(origbuf, bp->b_data, origbufsize);
4243 		free(origbuf, M_BIOBUF);
4244 	}
4245 	bufspace_adjust(bp, newbsize);
4246 }
4247 
4248 /*
4249  * This code constitutes the buffer memory from either anonymous system
4250  * memory (in the case of non-VMIO operations) or from an associated
4251  * VM object (in the case of VMIO operations).  This code is able to
4252  * resize a buffer up or down.
4253  *
4254  * Note that this code is tricky, and has many complications to resolve
4255  * deadlock or inconsistent data situations.  Tread lightly!!!
4256  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4257  * the caller.  Calling this code willy nilly can result in the loss of data.
4258  *
4259  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4260  * B_CACHE for the non-VMIO case.
4261  */
4262 int
4263 allocbuf(struct buf *bp, int size)
4264 {
4265 	int newbsize;
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  *	bufdone 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 
4453 	runningbufwakeup(bp);
4454 	if (bp->b_iocmd == BIO_WRITE)
4455 		dropobj = bp->b_bufobj;
4456 	/* call optional completion function if requested */
4457 	if (bp->b_iodone != NULL) {
4458 		biodone = bp->b_iodone;
4459 		bp->b_iodone = NULL;
4460 		(*biodone) (bp);
4461 		if (dropobj)
4462 			bufobj_wdrop(dropobj);
4463 		return;
4464 	}
4465 	if (bp->b_flags & B_VMIO) {
4466 		/*
4467 		 * Set B_CACHE if the op was a normal read and no error
4468 		 * occurred.  B_CACHE is set for writes in the b*write()
4469 		 * routines.
4470 		 */
4471 		if (bp->b_iocmd == BIO_READ &&
4472 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4473 		    !(bp->b_ioflags & BIO_ERROR))
4474 			bp->b_flags |= B_CACHE;
4475 		vfs_vmio_iodone(bp);
4476 	}
4477 	if (!LIST_EMPTY(&bp->b_dep))
4478 		buf_complete(bp);
4479 	if ((bp->b_flags & B_CKHASH) != 0) {
4480 		KASSERT(bp->b_iocmd == BIO_READ,
4481 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4482 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4483 		(*bp->b_ckhashcalc)(bp);
4484 	}
4485 	/*
4486 	 * For asynchronous completions, release the buffer now. The brelse
4487 	 * will do a wakeup there if necessary - so no need to do a wakeup
4488 	 * here in the async case. The sync case always needs to do a wakeup.
4489 	 */
4490 	if (bp->b_flags & B_ASYNC) {
4491 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4492 		    (bp->b_ioflags & BIO_ERROR))
4493 			brelse(bp);
4494 		else
4495 			bqrelse(bp);
4496 	} else
4497 		bdone(bp);
4498 	if (dropobj)
4499 		bufobj_wdrop(dropobj);
4500 }
4501 
4502 /*
4503  * This routine is called in lieu of iodone in the case of
4504  * incomplete I/O.  This keeps the busy status for pages
4505  * consistent.
4506  */
4507 void
4508 vfs_unbusy_pages(struct buf *bp)
4509 {
4510 	int i;
4511 	vm_object_t obj;
4512 	vm_page_t m;
4513 
4514 	runningbufwakeup(bp);
4515 	if (!(bp->b_flags & B_VMIO))
4516 		return;
4517 
4518 	obj = bp->b_bufobj->bo_object;
4519 	for (i = 0; i < bp->b_npages; i++) {
4520 		m = bp->b_pages[i];
4521 		if (m == bogus_page) {
4522 			m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4523 			if (!m)
4524 				panic("vfs_unbusy_pages: page missing\n");
4525 			bp->b_pages[i] = m;
4526 			if (buf_mapped(bp)) {
4527 				BUF_CHECK_MAPPED(bp);
4528 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4529 				    bp->b_pages, bp->b_npages);
4530 			} else
4531 				BUF_CHECK_UNMAPPED(bp);
4532 		}
4533 		vm_page_sunbusy(m);
4534 	}
4535 	vm_object_pip_wakeupn(obj, bp->b_npages);
4536 }
4537 
4538 /*
4539  * vfs_page_set_valid:
4540  *
4541  *	Set the valid bits in a page based on the supplied offset.   The
4542  *	range is restricted to the buffer's size.
4543  *
4544  *	This routine is typically called after a read completes.
4545  */
4546 static void
4547 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4548 {
4549 	vm_ooffset_t eoff;
4550 
4551 	/*
4552 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4553 	 * page boundary and eoff is not greater than the end of the buffer.
4554 	 * The end of the buffer, in this case, is our file EOF, not the
4555 	 * allocation size of the buffer.
4556 	 */
4557 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4558 	if (eoff > bp->b_offset + bp->b_bcount)
4559 		eoff = bp->b_offset + bp->b_bcount;
4560 
4561 	/*
4562 	 * Set valid range.  This is typically the entire buffer and thus the
4563 	 * entire page.
4564 	 */
4565 	if (eoff > off)
4566 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4567 }
4568 
4569 /*
4570  * vfs_page_set_validclean:
4571  *
4572  *	Set the valid bits and clear the dirty bits in a page based on the
4573  *	supplied offset.   The range is restricted to the buffer's size.
4574  */
4575 static void
4576 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4577 {
4578 	vm_ooffset_t soff, eoff;
4579 
4580 	/*
4581 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4582 	 * page boundary or cross the end of the buffer.  The end of the
4583 	 * buffer, in this case, is our file EOF, not the allocation size
4584 	 * of the buffer.
4585 	 */
4586 	soff = off;
4587 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4588 	if (eoff > bp->b_offset + bp->b_bcount)
4589 		eoff = bp->b_offset + bp->b_bcount;
4590 
4591 	/*
4592 	 * Set valid range.  This is typically the entire buffer and thus the
4593 	 * entire page.
4594 	 */
4595 	if (eoff > soff) {
4596 		vm_page_set_validclean(
4597 		    m,
4598 		   (vm_offset_t) (soff & PAGE_MASK),
4599 		   (vm_offset_t) (eoff - soff)
4600 		);
4601 	}
4602 }
4603 
4604 /*
4605  * Acquire a shared busy on all pages in the buf.
4606  */
4607 void
4608 vfs_busy_pages_acquire(struct buf *bp)
4609 {
4610 	int i;
4611 
4612 	for (i = 0; i < bp->b_npages; i++)
4613 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4614 }
4615 
4616 void
4617 vfs_busy_pages_release(struct buf *bp)
4618 {
4619 	int i;
4620 
4621 	for (i = 0; i < bp->b_npages; i++)
4622 		vm_page_sunbusy(bp->b_pages[i]);
4623 }
4624 
4625 /*
4626  * This routine is called before a device strategy routine.
4627  * It is used to tell the VM system that paging I/O is in
4628  * progress, and treat the pages associated with the buffer
4629  * almost as being exclusive busy.  Also the object paging_in_progress
4630  * flag is handled to make sure that the object doesn't become
4631  * inconsistent.
4632  *
4633  * Since I/O has not been initiated yet, certain buffer flags
4634  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4635  * and should be ignored.
4636  */
4637 void
4638 vfs_busy_pages(struct buf *bp, int clear_modify)
4639 {
4640 	vm_object_t obj;
4641 	vm_ooffset_t foff;
4642 	vm_page_t m;
4643 	int i;
4644 	bool bogus;
4645 
4646 	if (!(bp->b_flags & B_VMIO))
4647 		return;
4648 
4649 	obj = bp->b_bufobj->bo_object;
4650 	foff = bp->b_offset;
4651 	KASSERT(bp->b_offset != NOOFFSET,
4652 	    ("vfs_busy_pages: no buffer offset"));
4653 	if ((bp->b_flags & B_CLUSTER) == 0) {
4654 		vm_object_pip_add(obj, bp->b_npages);
4655 		vfs_busy_pages_acquire(bp);
4656 	}
4657 	if (bp->b_bufsize != 0)
4658 		vfs_setdirty_range(bp);
4659 	bogus = false;
4660 	for (i = 0; i < bp->b_npages; i++) {
4661 		m = bp->b_pages[i];
4662 		vm_page_assert_sbusied(m);
4663 
4664 		/*
4665 		 * When readying a buffer for a read ( i.e
4666 		 * clear_modify == 0 ), it is important to do
4667 		 * bogus_page replacement for valid pages in
4668 		 * partially instantiated buffers.  Partially
4669 		 * instantiated buffers can, in turn, occur when
4670 		 * reconstituting a buffer from its VM backing store
4671 		 * base.  We only have to do this if B_CACHE is
4672 		 * clear ( which causes the I/O to occur in the
4673 		 * first place ).  The replacement prevents the read
4674 		 * I/O from overwriting potentially dirty VM-backed
4675 		 * pages.  XXX bogus page replacement is, uh, bogus.
4676 		 * It may not work properly with small-block devices.
4677 		 * We need to find a better way.
4678 		 */
4679 		if (clear_modify) {
4680 			pmap_remove_write(m);
4681 			vfs_page_set_validclean(bp, foff, m);
4682 		} else if (vm_page_all_valid(m) &&
4683 		    (bp->b_flags & B_CACHE) == 0) {
4684 			bp->b_pages[i] = bogus_page;
4685 			bogus = true;
4686 		}
4687 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4688 	}
4689 	if (bogus && buf_mapped(bp)) {
4690 		BUF_CHECK_MAPPED(bp);
4691 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4692 		    bp->b_pages, bp->b_npages);
4693 	}
4694 }
4695 
4696 /*
4697  *	vfs_bio_set_valid:
4698  *
4699  *	Set the range within the buffer to valid.  The range is
4700  *	relative to the beginning of the buffer, b_offset.  Note that
4701  *	b_offset itself may be offset from the beginning of the first
4702  *	page.
4703  */
4704 void
4705 vfs_bio_set_valid(struct buf *bp, int base, int size)
4706 {
4707 	int i, n;
4708 	vm_page_t m;
4709 
4710 	if (!(bp->b_flags & B_VMIO))
4711 		return;
4712 
4713 	/*
4714 	 * Fixup base to be relative to beginning of first page.
4715 	 * Set initial n to be the maximum number of bytes in the
4716 	 * first page that can be validated.
4717 	 */
4718 	base += (bp->b_offset & PAGE_MASK);
4719 	n = PAGE_SIZE - (base & PAGE_MASK);
4720 
4721 	/*
4722 	 * Busy may not be strictly necessary here because the pages are
4723 	 * unlikely to be fully valid and the vnode lock will synchronize
4724 	 * their access via getpages.  It is grabbed for consistency with
4725 	 * other page validation.
4726 	 */
4727 	vfs_busy_pages_acquire(bp);
4728 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4729 		m = bp->b_pages[i];
4730 		if (n > size)
4731 			n = size;
4732 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4733 		base += n;
4734 		size -= n;
4735 		n = PAGE_SIZE;
4736 	}
4737 	vfs_busy_pages_release(bp);
4738 }
4739 
4740 /*
4741  *	vfs_bio_clrbuf:
4742  *
4743  *	If the specified buffer is a non-VMIO buffer, clear the entire
4744  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4745  *	validate only the previously invalid portions of the buffer.
4746  *	This routine essentially fakes an I/O, so we need to clear
4747  *	BIO_ERROR and B_INVAL.
4748  *
4749  *	Note that while we only theoretically need to clear through b_bcount,
4750  *	we go ahead and clear through b_bufsize.
4751  */
4752 void
4753 vfs_bio_clrbuf(struct buf *bp)
4754 {
4755 	int i, j, mask, sa, ea, slide;
4756 
4757 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4758 		clrbuf(bp);
4759 		return;
4760 	}
4761 	bp->b_flags &= ~B_INVAL;
4762 	bp->b_ioflags &= ~BIO_ERROR;
4763 	vfs_busy_pages_acquire(bp);
4764 	sa = bp->b_offset & PAGE_MASK;
4765 	slide = 0;
4766 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4767 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4768 		ea = slide & PAGE_MASK;
4769 		if (ea == 0)
4770 			ea = PAGE_SIZE;
4771 		if (bp->b_pages[i] == bogus_page)
4772 			continue;
4773 		j = sa / DEV_BSIZE;
4774 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4775 		if ((bp->b_pages[i]->valid & mask) == mask)
4776 			continue;
4777 		if ((bp->b_pages[i]->valid & mask) == 0)
4778 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4779 		else {
4780 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4781 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4782 					pmap_zero_page_area(bp->b_pages[i],
4783 					    sa, DEV_BSIZE);
4784 				}
4785 			}
4786 		}
4787 		vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4788 		    roundup2(ea - sa, DEV_BSIZE));
4789 	}
4790 	vfs_busy_pages_release(bp);
4791 	bp->b_resid = 0;
4792 }
4793 
4794 void
4795 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4796 {
4797 	vm_page_t m;
4798 	int i, n;
4799 
4800 	if (buf_mapped(bp)) {
4801 		BUF_CHECK_MAPPED(bp);
4802 		bzero(bp->b_data + base, size);
4803 	} else {
4804 		BUF_CHECK_UNMAPPED(bp);
4805 		n = PAGE_SIZE - (base & PAGE_MASK);
4806 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4807 			m = bp->b_pages[i];
4808 			if (n > size)
4809 				n = size;
4810 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4811 			base += n;
4812 			size -= n;
4813 			n = PAGE_SIZE;
4814 		}
4815 	}
4816 }
4817 
4818 /*
4819  * Update buffer flags based on I/O request parameters, optionally releasing the
4820  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4821  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4822  * I/O).  Otherwise the buffer is released to the cache.
4823  */
4824 static void
4825 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4826 {
4827 
4828 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4829 	    ("buf %p non-VMIO noreuse", bp));
4830 
4831 	if ((ioflag & IO_DIRECT) != 0)
4832 		bp->b_flags |= B_DIRECT;
4833 	if ((ioflag & IO_EXT) != 0)
4834 		bp->b_xflags |= BX_ALTDATA;
4835 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4836 		bp->b_flags |= B_RELBUF;
4837 		if ((ioflag & IO_NOREUSE) != 0)
4838 			bp->b_flags |= B_NOREUSE;
4839 		if (release)
4840 			brelse(bp);
4841 	} else if (release)
4842 		bqrelse(bp);
4843 }
4844 
4845 void
4846 vfs_bio_brelse(struct buf *bp, int ioflag)
4847 {
4848 
4849 	b_io_dismiss(bp, ioflag, true);
4850 }
4851 
4852 void
4853 vfs_bio_set_flags(struct buf *bp, int ioflag)
4854 {
4855 
4856 	b_io_dismiss(bp, ioflag, false);
4857 }
4858 
4859 /*
4860  * vm_hold_load_pages and vm_hold_free_pages get pages into
4861  * a buffers address space.  The pages are anonymous and are
4862  * not associated with a file object.
4863  */
4864 static void
4865 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4866 {
4867 	vm_offset_t pg;
4868 	vm_page_t p;
4869 	int index;
4870 
4871 	BUF_CHECK_MAPPED(bp);
4872 
4873 	to = round_page(to);
4874 	from = round_page(from);
4875 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4876 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
4877 	KASSERT(to - from <= maxbcachebuf,
4878 	    ("vm_hold_load_pages too large %p %#jx %#jx %u",
4879 	    bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4880 
4881 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4882 		/*
4883 		 * note: must allocate system pages since blocking here
4884 		 * could interfere with paging I/O, no matter which
4885 		 * process we are.
4886 		 */
4887 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4888 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4889 		    VM_ALLOC_WAITOK);
4890 		pmap_qenter(pg, &p, 1);
4891 		bp->b_pages[index] = p;
4892 	}
4893 	bp->b_npages = index;
4894 }
4895 
4896 /* Return pages associated with this buf to the vm system */
4897 static void
4898 vm_hold_free_pages(struct buf *bp, int newbsize)
4899 {
4900 	vm_offset_t from;
4901 	vm_page_t p;
4902 	int index, newnpages;
4903 
4904 	BUF_CHECK_MAPPED(bp);
4905 
4906 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4907 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4908 	if (bp->b_npages > newnpages)
4909 		pmap_qremove(from, bp->b_npages - newnpages);
4910 	for (index = newnpages; index < bp->b_npages; index++) {
4911 		p = bp->b_pages[index];
4912 		bp->b_pages[index] = NULL;
4913 		vm_page_unwire_noq(p);
4914 		vm_page_free(p);
4915 	}
4916 	bp->b_npages = newnpages;
4917 }
4918 
4919 /*
4920  * Map an IO request into kernel virtual address space.
4921  *
4922  * All requests are (re)mapped into kernel VA space.
4923  * Notice that we use b_bufsize for the size of the buffer
4924  * to be mapped.  b_bcount might be modified by the driver.
4925  *
4926  * Note that even if the caller determines that the address space should
4927  * be valid, a race or a smaller-file mapped into a larger space may
4928  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4929  * check the return value.
4930  *
4931  * This function only works with pager buffers.
4932  */
4933 int
4934 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
4935 {
4936 	vm_prot_t prot;
4937 	int pidx;
4938 
4939 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
4940 	prot = VM_PROT_READ;
4941 	if (bp->b_iocmd == BIO_READ)
4942 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4943 	pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4944 	    (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
4945 	if (pidx < 0)
4946 		return (-1);
4947 	bp->b_bufsize = len;
4948 	bp->b_npages = pidx;
4949 	bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
4950 	if (mapbuf || !unmapped_buf_allowed) {
4951 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4952 		bp->b_data = bp->b_kvabase + bp->b_offset;
4953 	} else
4954 		bp->b_data = unmapped_buf;
4955 	return (0);
4956 }
4957 
4958 /*
4959  * Free the io map PTEs associated with this IO operation.
4960  * We also invalidate the TLB entries and restore the original b_addr.
4961  *
4962  * This function only works with pager buffers.
4963  */
4964 void
4965 vunmapbuf(struct buf *bp)
4966 {
4967 	int npages;
4968 
4969 	npages = bp->b_npages;
4970 	if (buf_mapped(bp))
4971 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4972 	vm_page_unhold_pages(bp->b_pages, npages);
4973 
4974 	bp->b_data = unmapped_buf;
4975 }
4976 
4977 void
4978 bdone(struct buf *bp)
4979 {
4980 	struct mtx *mtxp;
4981 
4982 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4983 	mtx_lock(mtxp);
4984 	bp->b_flags |= B_DONE;
4985 	wakeup(bp);
4986 	mtx_unlock(mtxp);
4987 }
4988 
4989 void
4990 bwait(struct buf *bp, u_char pri, const char *wchan)
4991 {
4992 	struct mtx *mtxp;
4993 
4994 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4995 	mtx_lock(mtxp);
4996 	while ((bp->b_flags & B_DONE) == 0)
4997 		msleep(bp, mtxp, pri, wchan, 0);
4998 	mtx_unlock(mtxp);
4999 }
5000 
5001 int
5002 bufsync(struct bufobj *bo, int waitfor)
5003 {
5004 
5005 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5006 }
5007 
5008 void
5009 bufstrategy(struct bufobj *bo, struct buf *bp)
5010 {
5011 	int i __unused;
5012 	struct vnode *vp;
5013 
5014 	vp = bp->b_vp;
5015 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5016 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5017 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5018 	i = VOP_STRATEGY(vp, bp);
5019 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5020 }
5021 
5022 /*
5023  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
5024  */
5025 void
5026 bufobj_init(struct bufobj *bo, void *private)
5027 {
5028 	static volatile int bufobj_cleanq;
5029 
5030         bo->bo_domain =
5031             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5032         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5033         bo->bo_private = private;
5034         TAILQ_INIT(&bo->bo_clean.bv_hd);
5035         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5036 }
5037 
5038 void
5039 bufobj_wrefl(struct bufobj *bo)
5040 {
5041 
5042 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5043 	ASSERT_BO_WLOCKED(bo);
5044 	bo->bo_numoutput++;
5045 }
5046 
5047 void
5048 bufobj_wref(struct bufobj *bo)
5049 {
5050 
5051 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5052 	BO_LOCK(bo);
5053 	bo->bo_numoutput++;
5054 	BO_UNLOCK(bo);
5055 }
5056 
5057 void
5058 bufobj_wdrop(struct bufobj *bo)
5059 {
5060 
5061 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5062 	BO_LOCK(bo);
5063 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5064 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5065 		bo->bo_flag &= ~BO_WWAIT;
5066 		wakeup(&bo->bo_numoutput);
5067 	}
5068 	BO_UNLOCK(bo);
5069 }
5070 
5071 int
5072 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5073 {
5074 	int error;
5075 
5076 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5077 	ASSERT_BO_WLOCKED(bo);
5078 	error = 0;
5079 	while (bo->bo_numoutput) {
5080 		bo->bo_flag |= BO_WWAIT;
5081 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5082 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5083 		if (error)
5084 			break;
5085 	}
5086 	return (error);
5087 }
5088 
5089 /*
5090  * Set bio_data or bio_ma for struct bio from the struct buf.
5091  */
5092 void
5093 bdata2bio(struct buf *bp, struct bio *bip)
5094 {
5095 
5096 	if (!buf_mapped(bp)) {
5097 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5098 		bip->bio_ma = bp->b_pages;
5099 		bip->bio_ma_n = bp->b_npages;
5100 		bip->bio_data = unmapped_buf;
5101 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5102 		bip->bio_flags |= BIO_UNMAPPED;
5103 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5104 		    PAGE_SIZE == bp->b_npages,
5105 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5106 		    (long long)bip->bio_length, bip->bio_ma_n));
5107 	} else {
5108 		bip->bio_data = bp->b_data;
5109 		bip->bio_ma = NULL;
5110 	}
5111 }
5112 
5113 /*
5114  * The MIPS pmap code currently doesn't handle aliased pages.
5115  * The VIPT caches may not handle page aliasing themselves, leading
5116  * to data corruption.
5117  *
5118  * As such, this code makes a system extremely unhappy if said
5119  * system doesn't support unaliasing the above situation in hardware.
5120  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5121  * this feature at build time, so it has to be handled in software.
5122  *
5123  * Once the MIPS pmap/cache code grows to support this function on
5124  * earlier chips, it should be flipped back off.
5125  */
5126 #ifdef	__mips__
5127 static int buf_pager_relbuf = 1;
5128 #else
5129 static int buf_pager_relbuf = 0;
5130 #endif
5131 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5132     &buf_pager_relbuf, 0,
5133     "Make buffer pager release buffers after reading");
5134 
5135 /*
5136  * The buffer pager.  It uses buffer reads to validate pages.
5137  *
5138  * In contrast to the generic local pager from vm/vnode_pager.c, this
5139  * pager correctly and easily handles volumes where the underlying
5140  * device block size is greater than the machine page size.  The
5141  * buffer cache transparently extends the requested page run to be
5142  * aligned at the block boundary, and does the necessary bogus page
5143  * replacements in the addends to avoid obliterating already valid
5144  * pages.
5145  *
5146  * The only non-trivial issue is that the exclusive busy state for
5147  * pages, which is assumed by the vm_pager_getpages() interface, is
5148  * incompatible with the VMIO buffer cache's desire to share-busy the
5149  * pages.  This function performs a trivial downgrade of the pages'
5150  * state before reading buffers, and a less trivial upgrade from the
5151  * shared-busy to excl-busy state after the read.
5152  */
5153 int
5154 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5155     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5156     vbg_get_blksize_t get_blksize)
5157 {
5158 	vm_page_t m;
5159 	vm_object_t object;
5160 	struct buf *bp;
5161 	struct mount *mp;
5162 	daddr_t lbn, lbnp;
5163 	vm_ooffset_t la, lb, poff, poffe;
5164 	long bsize;
5165 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5166 	bool redo, lpart;
5167 
5168 	object = vp->v_object;
5169 	mp = vp->v_mount;
5170 	error = 0;
5171 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5172 	if (la >= object->un_pager.vnp.vnp_size)
5173 		return (VM_PAGER_BAD);
5174 
5175 	/*
5176 	 * Change the meaning of la from where the last requested page starts
5177 	 * to where it ends, because that's the end of the requested region
5178 	 * and the start of the potential read-ahead region.
5179 	 */
5180 	la += PAGE_SIZE;
5181 	lpart = la > object->un_pager.vnp.vnp_size;
5182 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5183 
5184 	/*
5185 	 * Calculate read-ahead, behind and total pages.
5186 	 */
5187 	pgsin = count;
5188 	lb = IDX_TO_OFF(ma[0]->pindex);
5189 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5190 	pgsin += pgsin_b;
5191 	if (rbehind != NULL)
5192 		*rbehind = pgsin_b;
5193 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5194 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5195 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5196 		    PAGE_SIZE) - la);
5197 	pgsin += pgsin_a;
5198 	if (rahead != NULL)
5199 		*rahead = pgsin_a;
5200 	VM_CNT_INC(v_vnodein);
5201 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5202 
5203 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5204 	    != 0) ? GB_UNMAPPED : 0;
5205 again:
5206 	for (i = 0; i < count; i++) {
5207 		if (ma[i] != bogus_page)
5208 			vm_page_busy_downgrade(ma[i]);
5209 	}
5210 
5211 	lbnp = -1;
5212 	for (i = 0; i < count; i++) {
5213 		m = ma[i];
5214 		if (m == bogus_page)
5215 			continue;
5216 
5217 		/*
5218 		 * Pages are shared busy and the object lock is not
5219 		 * owned, which together allow for the pages'
5220 		 * invalidation.  The racy test for validity avoids
5221 		 * useless creation of the buffer for the most typical
5222 		 * case when invalidation is not used in redo or for
5223 		 * parallel read.  The shared->excl upgrade loop at
5224 		 * the end of the function catches the race in a
5225 		 * reliable way (protected by the object lock).
5226 		 */
5227 		if (vm_page_all_valid(m))
5228 			continue;
5229 
5230 		poff = IDX_TO_OFF(m->pindex);
5231 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5232 		for (; poff < poffe; poff += bsize) {
5233 			lbn = get_lblkno(vp, poff);
5234 			if (lbn == lbnp)
5235 				goto next_page;
5236 			lbnp = lbn;
5237 
5238 			bsize = get_blksize(vp, lbn);
5239 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5240 			    br_flags, &bp);
5241 			if (error != 0)
5242 				goto end_pages;
5243 			if (bp->b_rcred == curthread->td_ucred) {
5244 				crfree(bp->b_rcred);
5245 				bp->b_rcred = NOCRED;
5246 			}
5247 			if (LIST_EMPTY(&bp->b_dep)) {
5248 				/*
5249 				 * Invalidation clears m->valid, but
5250 				 * may leave B_CACHE flag if the
5251 				 * buffer existed at the invalidation
5252 				 * time.  In this case, recycle the
5253 				 * buffer to do real read on next
5254 				 * bread() after redo.
5255 				 *
5256 				 * Otherwise B_RELBUF is not strictly
5257 				 * necessary, enable to reduce buf
5258 				 * cache pressure.
5259 				 */
5260 				if (buf_pager_relbuf ||
5261 				    !vm_page_all_valid(m))
5262 					bp->b_flags |= B_RELBUF;
5263 
5264 				bp->b_flags &= ~B_NOCACHE;
5265 				brelse(bp);
5266 			} else {
5267 				bqrelse(bp);
5268 			}
5269 		}
5270 		KASSERT(1 /* racy, enable for debugging */ ||
5271 		    vm_page_all_valid(m) || i == count - 1,
5272 		    ("buf %d %p invalid", i, m));
5273 		if (i == count - 1 && lpart) {
5274 			if (!vm_page_none_valid(m) &&
5275 			    !vm_page_all_valid(m))
5276 				vm_page_zero_invalid(m, TRUE);
5277 		}
5278 next_page:;
5279 	}
5280 end_pages:
5281 
5282 	redo = false;
5283 	for (i = 0; i < count; i++) {
5284 		if (ma[i] == bogus_page)
5285 			continue;
5286 		if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5287 			vm_page_sunbusy(ma[i]);
5288 			ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5289 			    VM_ALLOC_NORMAL);
5290 		}
5291 
5292 		/*
5293 		 * Since the pages were only sbusy while neither the
5294 		 * buffer nor the object lock was held by us, or
5295 		 * reallocated while vm_page_grab() slept for busy
5296 		 * relinguish, they could have been invalidated.
5297 		 * Recheck the valid bits and re-read as needed.
5298 		 *
5299 		 * Note that the last page is made fully valid in the
5300 		 * read loop, and partial validity for the page at
5301 		 * index count - 1 could mean that the page was
5302 		 * invalidated or removed, so we must restart for
5303 		 * safety as well.
5304 		 */
5305 		if (!vm_page_all_valid(ma[i]))
5306 			redo = true;
5307 	}
5308 	if (redo && error == 0)
5309 		goto again;
5310 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5311 }
5312 
5313 #include "opt_ddb.h"
5314 #ifdef DDB
5315 #include <ddb/ddb.h>
5316 
5317 /* DDB command to show buffer data */
5318 DB_SHOW_COMMAND(buffer, db_show_buffer)
5319 {
5320 	/* get args */
5321 	struct buf *bp = (struct buf *)addr;
5322 #ifdef FULL_BUF_TRACKING
5323 	uint32_t i, j;
5324 #endif
5325 
5326 	if (!have_addr) {
5327 		db_printf("usage: show buffer <addr>\n");
5328 		return;
5329 	}
5330 
5331 	db_printf("buf at %p\n", bp);
5332 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5333 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5334 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5335 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5336 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5337 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5338 	db_printf(
5339 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5340 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5341 	    "b_vp = %p, b_dep = %p\n",
5342 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5343 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5344 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5345 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5346 	    bp->b_kvabase, bp->b_kvasize);
5347 	if (bp->b_npages) {
5348 		int i;
5349 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5350 		for (i = 0; i < bp->b_npages; i++) {
5351 			vm_page_t m;
5352 			m = bp->b_pages[i];
5353 			if (m != NULL)
5354 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5355 				    (u_long)m->pindex,
5356 				    (u_long)VM_PAGE_TO_PHYS(m));
5357 			else
5358 				db_printf("( ??? )");
5359 			if ((i + 1) < bp->b_npages)
5360 				db_printf(",");
5361 		}
5362 		db_printf("\n");
5363 	}
5364 	BUF_LOCKPRINTINFO(bp);
5365 #if defined(FULL_BUF_TRACKING)
5366 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5367 
5368 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5369 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5370 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5371 			continue;
5372 		db_printf(" %2u: %s\n", j,
5373 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5374 	}
5375 #elif defined(BUF_TRACKING)
5376 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5377 #endif
5378 	db_printf(" ");
5379 }
5380 
5381 DB_SHOW_COMMAND(bufqueues, bufqueues)
5382 {
5383 	struct bufdomain *bd;
5384 	struct buf *bp;
5385 	long total;
5386 	int i, j, cnt;
5387 
5388 	db_printf("bqempty: %d\n", bqempty.bq_len);
5389 
5390 	for (i = 0; i < buf_domains; i++) {
5391 		bd = &bdomain[i];
5392 		db_printf("Buf domain %d\n", i);
5393 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5394 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5395 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5396 		db_printf("\n");
5397 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5398 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5399 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5400 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5401 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5402 		db_printf("\n");
5403 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5404 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5405 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5406 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5407 		db_printf("\n");
5408 		total = 0;
5409 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5410 			total += bp->b_bufsize;
5411 		db_printf("\tcleanq count\t%d (%ld)\n",
5412 		    bd->bd_cleanq->bq_len, total);
5413 		total = 0;
5414 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5415 			total += bp->b_bufsize;
5416 		db_printf("\tdirtyq count\t%d (%ld)\n",
5417 		    bd->bd_dirtyq.bq_len, total);
5418 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5419 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5420 		db_printf("\tCPU ");
5421 		for (j = 0; j <= mp_maxid; j++)
5422 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5423 		db_printf("\n");
5424 		cnt = 0;
5425 		total = 0;
5426 		for (j = 0; j < nbuf; j++) {
5427 			bp = nbufp(j);
5428 			if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5429 				cnt++;
5430 				total += bp->b_bufsize;
5431 			}
5432 		}
5433 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5434 		cnt = 0;
5435 		total = 0;
5436 		for (j = 0; j < nbuf; j++) {
5437 			bp = nbufp(j);
5438 			if (bp->b_domain == i) {
5439 				cnt++;
5440 				total += bp->b_bufsize;
5441 			}
5442 		}
5443 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5444 	}
5445 }
5446 
5447 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5448 {
5449 	struct buf *bp;
5450 	int i;
5451 
5452 	for (i = 0; i < nbuf; i++) {
5453 		bp = nbufp(i);
5454 		if (BUF_ISLOCKED(bp)) {
5455 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5456 			db_printf("\n");
5457 			if (db_pager_quit)
5458 				break;
5459 		}
5460 	}
5461 }
5462 
5463 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5464 {
5465 	struct vnode *vp;
5466 	struct buf *bp;
5467 
5468 	if (!have_addr) {
5469 		db_printf("usage: show vnodebufs <addr>\n");
5470 		return;
5471 	}
5472 	vp = (struct vnode *)addr;
5473 	db_printf("Clean buffers:\n");
5474 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5475 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5476 		db_printf("\n");
5477 	}
5478 	db_printf("Dirty buffers:\n");
5479 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5480 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5481 		db_printf("\n");
5482 	}
5483 }
5484 
5485 DB_COMMAND(countfreebufs, db_coundfreebufs)
5486 {
5487 	struct buf *bp;
5488 	int i, used = 0, nfree = 0;
5489 
5490 	if (have_addr) {
5491 		db_printf("usage: countfreebufs\n");
5492 		return;
5493 	}
5494 
5495 	for (i = 0; i < nbuf; i++) {
5496 		bp = nbufp(i);
5497 		if (bp->b_qindex == QUEUE_EMPTY)
5498 			nfree++;
5499 		else
5500 			used++;
5501 	}
5502 
5503 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5504 	    nfree + used);
5505 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5506 }
5507 #endif /* DDB */
5508