xref: /freebsd/sys/kern/vfs_bio.c (revision f15e18a642cb3f7ebc747f8e9cdf11274140107d)
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 (LIST_EMPTY(&bp->b_dep)) {
2643 		bp->b_flags &= ~B_IOSTARTED;
2644 	} else {
2645 		KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2646 		    ("brelse: SU io not finished bp %p", bp));
2647 	}
2648 
2649 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2650 		BO_LOCK(bp->b_bufobj);
2651 		bp->b_vflags &= ~BV_BKGRDERR;
2652 		BO_UNLOCK(bp->b_bufobj);
2653 		bdirty(bp);
2654 	}
2655 
2656 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2657 	    (bp->b_flags & B_INVALONERR)) {
2658 		/*
2659 		 * Forced invalidation of dirty buffer contents, to be used
2660 		 * after a failed write in the rare case that the loss of the
2661 		 * contents is acceptable.  The buffer is invalidated and
2662 		 * freed.
2663 		 */
2664 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2665 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2666 	}
2667 
2668 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2669 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2670 	    !(bp->b_flags & B_INVAL)) {
2671 		/*
2672 		 * Failed write, redirty.  All errors except ENXIO (which
2673 		 * means the device is gone) are treated as being
2674 		 * transient.
2675 		 *
2676 		 * XXX Treating EIO as transient is not correct; the
2677 		 * contract with the local storage device drivers is that
2678 		 * they will only return EIO once the I/O is no longer
2679 		 * retriable.  Network I/O also respects this through the
2680 		 * guarantees of TCP and/or the internal retries of NFS.
2681 		 * ENOMEM might be transient, but we also have no way of
2682 		 * knowing when its ok to retry/reschedule.  In general,
2683 		 * this entire case should be made obsolete through better
2684 		 * error handling/recovery and resource scheduling.
2685 		 *
2686 		 * Do this also for buffers that failed with ENXIO, but have
2687 		 * non-empty dependencies - the soft updates code might need
2688 		 * to access the buffer to untangle them.
2689 		 *
2690 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2691 		 */
2692 		bp->b_ioflags &= ~BIO_ERROR;
2693 		bdirty(bp);
2694 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2695 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2696 		/*
2697 		 * Either a failed read I/O, or we were asked to free or not
2698 		 * cache the buffer, or we failed to write to a device that's
2699 		 * no longer present.
2700 		 */
2701 		bp->b_flags |= B_INVAL;
2702 		if (!LIST_EMPTY(&bp->b_dep))
2703 			buf_deallocate(bp);
2704 		if (bp->b_flags & B_DELWRI)
2705 			bdirtysub(bp);
2706 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2707 		if ((bp->b_flags & B_VMIO) == 0) {
2708 			allocbuf(bp, 0);
2709 			if (bp->b_vp)
2710 				brelvp(bp);
2711 		}
2712 	}
2713 
2714 	/*
2715 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2716 	 * is called with B_DELWRI set, the underlying pages may wind up
2717 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2718 	 * because pages associated with a B_DELWRI bp are marked clean.
2719 	 *
2720 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2721 	 * if B_DELWRI is set.
2722 	 */
2723 	if (bp->b_flags & B_DELWRI)
2724 		bp->b_flags &= ~B_RELBUF;
2725 
2726 	/*
2727 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2728 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2729 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2730 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2731 	 *
2732 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2733 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2734 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2735 	 *
2736 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2737 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2738 	 * the commit state and we cannot afford to lose the buffer. If the
2739 	 * buffer has a background write in progress, we need to keep it
2740 	 * around to prevent it from being reconstituted and starting a second
2741 	 * background write.
2742 	 */
2743 
2744 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2745 
2746 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2747 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2748 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2749 	    vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2750 		vfs_vmio_invalidate(bp);
2751 		allocbuf(bp, 0);
2752 	}
2753 
2754 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2755 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2756 		allocbuf(bp, 0);
2757 		bp->b_flags &= ~B_NOREUSE;
2758 		if (bp->b_vp != NULL)
2759 			brelvp(bp);
2760 	}
2761 
2762 	/*
2763 	 * If the buffer has junk contents signal it and eventually
2764 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2765 	 * doesn't find it.
2766 	 */
2767 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2768 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2769 		bp->b_flags |= B_INVAL;
2770 	if (bp->b_flags & B_INVAL) {
2771 		if (bp->b_flags & B_DELWRI)
2772 			bundirty(bp);
2773 		if (bp->b_vp)
2774 			brelvp(bp);
2775 	}
2776 
2777 	buf_track(bp, __func__);
2778 
2779 	/* buffers with no memory */
2780 	if (bp->b_bufsize == 0) {
2781 		buf_free(bp);
2782 		return;
2783 	}
2784 	/* buffers with junk contents */
2785 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2786 	    (bp->b_ioflags & BIO_ERROR)) {
2787 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2788 		if (bp->b_vflags & BV_BKGRDINPROG)
2789 			panic("losing buffer 2");
2790 		qindex = QUEUE_CLEAN;
2791 		bp->b_flags |= B_AGE;
2792 	/* remaining buffers */
2793 	} else if (bp->b_flags & B_DELWRI)
2794 		qindex = QUEUE_DIRTY;
2795 	else
2796 		qindex = QUEUE_CLEAN;
2797 
2798 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2799 		panic("brelse: not dirty");
2800 
2801 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2802 	bp->b_xflags &= ~(BX_CVTENXIO);
2803 	/* binsfree unlocks bp. */
2804 	binsfree(bp, qindex);
2805 }
2806 
2807 /*
2808  * Release a buffer back to the appropriate queue but do not try to free
2809  * it.  The buffer is expected to be used again soon.
2810  *
2811  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2812  * biodone() to requeue an async I/O on completion.  It is also used when
2813  * known good buffers need to be requeued but we think we may need the data
2814  * again soon.
2815  *
2816  * XXX we should be able to leave the B_RELBUF hint set on completion.
2817  */
2818 void
2819 bqrelse(struct buf *bp)
2820 {
2821 	int qindex;
2822 
2823 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2824 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2825 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2826 
2827 	qindex = QUEUE_NONE;
2828 	if (BUF_LOCKRECURSED(bp)) {
2829 		/* do not release to free list */
2830 		BUF_UNLOCK(bp);
2831 		return;
2832 	}
2833 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2834 	bp->b_xflags &= ~(BX_CVTENXIO);
2835 
2836 	if (LIST_EMPTY(&bp->b_dep)) {
2837 		bp->b_flags &= ~B_IOSTARTED;
2838 	} else {
2839 		KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2840 		    ("bqrelse: SU io not finished bp %p", bp));
2841 	}
2842 
2843 	if (bp->b_flags & B_MANAGED) {
2844 		if (bp->b_flags & B_REMFREE)
2845 			bremfreef(bp);
2846 		goto out;
2847 	}
2848 
2849 	/* buffers with stale but valid contents */
2850 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2851 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2852 		BO_LOCK(bp->b_bufobj);
2853 		bp->b_vflags &= ~BV_BKGRDERR;
2854 		BO_UNLOCK(bp->b_bufobj);
2855 		qindex = QUEUE_DIRTY;
2856 	} else {
2857 		if ((bp->b_flags & B_DELWRI) == 0 &&
2858 		    (bp->b_xflags & BX_VNDIRTY))
2859 			panic("bqrelse: not dirty");
2860 		if ((bp->b_flags & B_NOREUSE) != 0) {
2861 			brelse(bp);
2862 			return;
2863 		}
2864 		qindex = QUEUE_CLEAN;
2865 	}
2866 	buf_track(bp, __func__);
2867 	/* binsfree unlocks bp. */
2868 	binsfree(bp, qindex);
2869 	return;
2870 
2871 out:
2872 	buf_track(bp, __func__);
2873 	/* unlock */
2874 	BUF_UNLOCK(bp);
2875 }
2876 
2877 /*
2878  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2879  * restore bogus pages.
2880  */
2881 static void
2882 vfs_vmio_iodone(struct buf *bp)
2883 {
2884 	vm_ooffset_t foff;
2885 	vm_page_t m;
2886 	vm_object_t obj;
2887 	struct vnode *vp __unused;
2888 	int i, iosize, resid;
2889 	bool bogus;
2890 
2891 	obj = bp->b_bufobj->bo_object;
2892 	KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2893 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2894 	    blockcount_read(&obj->paging_in_progress), bp->b_npages));
2895 
2896 	vp = bp->b_vp;
2897 	VNPASS(vp->v_holdcnt > 0, vp);
2898 	VNPASS(vp->v_object != NULL, vp);
2899 
2900 	foff = bp->b_offset;
2901 	KASSERT(bp->b_offset != NOOFFSET,
2902 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2903 
2904 	bogus = false;
2905 	iosize = bp->b_bcount - bp->b_resid;
2906 	for (i = 0; i < bp->b_npages; i++) {
2907 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2908 		if (resid > iosize)
2909 			resid = iosize;
2910 
2911 		/*
2912 		 * cleanup bogus pages, restoring the originals
2913 		 */
2914 		m = bp->b_pages[i];
2915 		if (m == bogus_page) {
2916 			bogus = true;
2917 			m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2918 			if (m == NULL)
2919 				panic("biodone: page disappeared!");
2920 			bp->b_pages[i] = m;
2921 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2922 			/*
2923 			 * In the write case, the valid and clean bits are
2924 			 * already changed correctly ( see bdwrite() ), so we
2925 			 * only need to do this here in the read case.
2926 			 */
2927 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2928 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2929 			    "has unexpected dirty bits", m));
2930 			vfs_page_set_valid(bp, foff, m);
2931 		}
2932 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2933 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2934 		    (intmax_t)foff, (uintmax_t)m->pindex));
2935 
2936 		vm_page_sunbusy(m);
2937 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2938 		iosize -= resid;
2939 	}
2940 	vm_object_pip_wakeupn(obj, bp->b_npages);
2941 	if (bogus && buf_mapped(bp)) {
2942 		BUF_CHECK_MAPPED(bp);
2943 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2944 		    bp->b_pages, bp->b_npages);
2945 	}
2946 }
2947 
2948 /*
2949  * Perform page invalidation when a buffer is released.  The fully invalid
2950  * pages will be reclaimed later in vfs_vmio_truncate().
2951  */
2952 static void
2953 vfs_vmio_invalidate(struct buf *bp)
2954 {
2955 	vm_object_t obj;
2956 	vm_page_t m;
2957 	int flags, i, resid, poffset, presid;
2958 
2959 	if (buf_mapped(bp)) {
2960 		BUF_CHECK_MAPPED(bp);
2961 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2962 	} else
2963 		BUF_CHECK_UNMAPPED(bp);
2964 	/*
2965 	 * Get the base offset and length of the buffer.  Note that
2966 	 * in the VMIO case if the buffer block size is not
2967 	 * page-aligned then b_data pointer may not be page-aligned.
2968 	 * But our b_pages[] array *IS* page aligned.
2969 	 *
2970 	 * block sizes less then DEV_BSIZE (usually 512) are not
2971 	 * supported due to the page granularity bits (m->valid,
2972 	 * m->dirty, etc...).
2973 	 *
2974 	 * See man buf(9) for more information
2975 	 */
2976 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2977 	obj = bp->b_bufobj->bo_object;
2978 	resid = bp->b_bufsize;
2979 	poffset = bp->b_offset & PAGE_MASK;
2980 	VM_OBJECT_WLOCK(obj);
2981 	for (i = 0; i < bp->b_npages; i++) {
2982 		m = bp->b_pages[i];
2983 		if (m == bogus_page)
2984 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2985 		bp->b_pages[i] = NULL;
2986 
2987 		presid = resid > (PAGE_SIZE - poffset) ?
2988 		    (PAGE_SIZE - poffset) : resid;
2989 		KASSERT(presid >= 0, ("brelse: extra page"));
2990 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
2991 		if (pmap_page_wired_mappings(m) == 0)
2992 			vm_page_set_invalid(m, poffset, presid);
2993 		vm_page_sunbusy(m);
2994 		vm_page_release_locked(m, flags);
2995 		resid -= presid;
2996 		poffset = 0;
2997 	}
2998 	VM_OBJECT_WUNLOCK(obj);
2999 	bp->b_npages = 0;
3000 }
3001 
3002 /*
3003  * Page-granular truncation of an existing VMIO buffer.
3004  */
3005 static void
3006 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3007 {
3008 	vm_object_t obj;
3009 	vm_page_t m;
3010 	int flags, i;
3011 
3012 	if (bp->b_npages == desiredpages)
3013 		return;
3014 
3015 	if (buf_mapped(bp)) {
3016 		BUF_CHECK_MAPPED(bp);
3017 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3018 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3019 	} else
3020 		BUF_CHECK_UNMAPPED(bp);
3021 
3022 	/*
3023 	 * The object lock is needed only if we will attempt to free pages.
3024 	 */
3025 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3026 	if ((bp->b_flags & B_DIRECT) != 0) {
3027 		flags |= VPR_TRYFREE;
3028 		obj = bp->b_bufobj->bo_object;
3029 		VM_OBJECT_WLOCK(obj);
3030 	} else {
3031 		obj = NULL;
3032 	}
3033 	for (i = desiredpages; i < bp->b_npages; i++) {
3034 		m = bp->b_pages[i];
3035 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3036 		bp->b_pages[i] = NULL;
3037 		if (obj != NULL)
3038 			vm_page_release_locked(m, flags);
3039 		else
3040 			vm_page_release(m, flags);
3041 	}
3042 	if (obj != NULL)
3043 		VM_OBJECT_WUNLOCK(obj);
3044 	bp->b_npages = desiredpages;
3045 }
3046 
3047 /*
3048  * Byte granular extension of VMIO buffers.
3049  */
3050 static void
3051 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3052 {
3053 	/*
3054 	 * We are growing the buffer, possibly in a
3055 	 * byte-granular fashion.
3056 	 */
3057 	vm_object_t obj;
3058 	vm_offset_t toff;
3059 	vm_offset_t tinc;
3060 	vm_page_t m;
3061 
3062 	/*
3063 	 * Step 1, bring in the VM pages from the object, allocating
3064 	 * them if necessary.  We must clear B_CACHE if these pages
3065 	 * are not valid for the range covered by the buffer.
3066 	 */
3067 	obj = bp->b_bufobj->bo_object;
3068 	if (bp->b_npages < desiredpages) {
3069 		KASSERT(desiredpages <= atop(maxbcachebuf),
3070 		    ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3071 		    bp, desiredpages, maxbcachebuf));
3072 
3073 		/*
3074 		 * We must allocate system pages since blocking
3075 		 * here could interfere with paging I/O, no
3076 		 * matter which process we are.
3077 		 *
3078 		 * Only exclusive busy can be tested here.
3079 		 * Blocking on shared busy might lead to
3080 		 * deadlocks once allocbuf() is called after
3081 		 * pages are vfs_busy_pages().
3082 		 */
3083 		(void)vm_page_grab_pages_unlocked(obj,
3084 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3085 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3086 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3087 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3088 		bp->b_npages = desiredpages;
3089 	}
3090 
3091 	/*
3092 	 * Step 2.  We've loaded the pages into the buffer,
3093 	 * we have to figure out if we can still have B_CACHE
3094 	 * set.  Note that B_CACHE is set according to the
3095 	 * byte-granular range ( bcount and size ), not the
3096 	 * aligned range ( newbsize ).
3097 	 *
3098 	 * The VM test is against m->valid, which is DEV_BSIZE
3099 	 * aligned.  Needless to say, the validity of the data
3100 	 * needs to also be DEV_BSIZE aligned.  Note that this
3101 	 * fails with NFS if the server or some other client
3102 	 * extends the file's EOF.  If our buffer is resized,
3103 	 * B_CACHE may remain set! XXX
3104 	 */
3105 	toff = bp->b_bcount;
3106 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3107 	while ((bp->b_flags & B_CACHE) && toff < size) {
3108 		vm_pindex_t pi;
3109 
3110 		if (tinc > (size - toff))
3111 			tinc = size - toff;
3112 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3113 		m = bp->b_pages[pi];
3114 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3115 		toff += tinc;
3116 		tinc = PAGE_SIZE;
3117 	}
3118 
3119 	/*
3120 	 * Step 3, fixup the KVA pmap.
3121 	 */
3122 	if (buf_mapped(bp))
3123 		bpmap_qenter(bp);
3124 	else
3125 		BUF_CHECK_UNMAPPED(bp);
3126 }
3127 
3128 /*
3129  * Check to see if a block at a particular lbn is available for a clustered
3130  * write.
3131  */
3132 static int
3133 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3134 {
3135 	struct buf *bpa;
3136 	int match;
3137 
3138 	match = 0;
3139 
3140 	/* If the buf isn't in core skip it */
3141 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3142 		return (0);
3143 
3144 	/* If the buf is busy we don't want to wait for it */
3145 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3146 		return (0);
3147 
3148 	/* Only cluster with valid clusterable delayed write buffers */
3149 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3150 	    (B_DELWRI | B_CLUSTEROK))
3151 		goto done;
3152 
3153 	if (bpa->b_bufsize != size)
3154 		goto done;
3155 
3156 	/*
3157 	 * Check to see if it is in the expected place on disk and that the
3158 	 * block has been mapped.
3159 	 */
3160 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3161 		match = 1;
3162 done:
3163 	BUF_UNLOCK(bpa);
3164 	return (match);
3165 }
3166 
3167 /*
3168  *	vfs_bio_awrite:
3169  *
3170  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3171  *	This is much better then the old way of writing only one buffer at
3172  *	a time.  Note that we may not be presented with the buffers in the
3173  *	correct order, so we search for the cluster in both directions.
3174  */
3175 int
3176 vfs_bio_awrite(struct buf *bp)
3177 {
3178 	struct bufobj *bo;
3179 	int i;
3180 	int j;
3181 	daddr_t lblkno = bp->b_lblkno;
3182 	struct vnode *vp = bp->b_vp;
3183 	int ncl;
3184 	int nwritten;
3185 	int size;
3186 	int maxcl;
3187 	int gbflags;
3188 
3189 	bo = &vp->v_bufobj;
3190 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3191 	/*
3192 	 * right now we support clustered writing only to regular files.  If
3193 	 * we find a clusterable block we could be in the middle of a cluster
3194 	 * rather then at the beginning.
3195 	 */
3196 	if ((vp->v_type == VREG) &&
3197 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3198 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3199 		size = vp->v_mount->mnt_stat.f_iosize;
3200 		maxcl = maxphys / size;
3201 
3202 		BO_RLOCK(bo);
3203 		for (i = 1; i < maxcl; i++)
3204 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3205 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3206 				break;
3207 
3208 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3209 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3210 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3211 				break;
3212 		BO_RUNLOCK(bo);
3213 		--j;
3214 		ncl = i + j;
3215 		/*
3216 		 * this is a possible cluster write
3217 		 */
3218 		if (ncl != 1) {
3219 			BUF_UNLOCK(bp);
3220 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3221 			    gbflags);
3222 			return (nwritten);
3223 		}
3224 	}
3225 	bremfree(bp);
3226 	bp->b_flags |= B_ASYNC;
3227 	/*
3228 	 * default (old) behavior, writing out only one block
3229 	 *
3230 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3231 	 */
3232 	nwritten = bp->b_bufsize;
3233 	(void) bwrite(bp);
3234 
3235 	return (nwritten);
3236 }
3237 
3238 /*
3239  *	getnewbuf_kva:
3240  *
3241  *	Allocate KVA for an empty buf header according to gbflags.
3242  */
3243 static int
3244 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3245 {
3246 
3247 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3248 		/*
3249 		 * In order to keep fragmentation sane we only allocate kva
3250 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3251 		 */
3252 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3253 
3254 		if (maxsize != bp->b_kvasize &&
3255 		    bufkva_alloc(bp, maxsize, gbflags))
3256 			return (ENOSPC);
3257 	}
3258 	return (0);
3259 }
3260 
3261 /*
3262  *	getnewbuf:
3263  *
3264  *	Find and initialize a new buffer header, freeing up existing buffers
3265  *	in the bufqueues as necessary.  The new buffer is returned locked.
3266  *
3267  *	We block if:
3268  *		We have insufficient buffer headers
3269  *		We have insufficient buffer space
3270  *		buffer_arena is too fragmented ( space reservation fails )
3271  *		If we have to flush dirty buffers ( but we try to avoid this )
3272  *
3273  *	The caller is responsible for releasing the reserved bufspace after
3274  *	allocbuf() is called.
3275  */
3276 static struct buf *
3277 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3278 {
3279 	struct bufdomain *bd;
3280 	struct buf *bp;
3281 	bool metadata, reserved;
3282 
3283 	bp = NULL;
3284 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3285 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3286 	if (!unmapped_buf_allowed)
3287 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3288 
3289 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3290 	    vp->v_type == VCHR)
3291 		metadata = true;
3292 	else
3293 		metadata = false;
3294 	if (vp == NULL)
3295 		bd = &bdomain[0];
3296 	else
3297 		bd = &bdomain[vp->v_bufobj.bo_domain];
3298 
3299 	counter_u64_add(getnewbufcalls, 1);
3300 	reserved = false;
3301 	do {
3302 		if (reserved == false &&
3303 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3304 			counter_u64_add(getnewbufrestarts, 1);
3305 			continue;
3306 		}
3307 		reserved = true;
3308 		if ((bp = buf_alloc(bd)) == NULL) {
3309 			counter_u64_add(getnewbufrestarts, 1);
3310 			continue;
3311 		}
3312 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3313 			return (bp);
3314 		break;
3315 	} while (buf_recycle(bd, false) == 0);
3316 
3317 	if (reserved)
3318 		bufspace_release(bd, maxsize);
3319 	if (bp != NULL) {
3320 		bp->b_flags |= B_INVAL;
3321 		brelse(bp);
3322 	}
3323 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3324 
3325 	return (NULL);
3326 }
3327 
3328 /*
3329  *	buf_daemon:
3330  *
3331  *	buffer flushing daemon.  Buffers are normally flushed by the
3332  *	update daemon but if it cannot keep up this process starts to
3333  *	take the load in an attempt to prevent getnewbuf() from blocking.
3334  */
3335 static struct kproc_desc buf_kp = {
3336 	"bufdaemon",
3337 	buf_daemon,
3338 	&bufdaemonproc
3339 };
3340 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3341 
3342 static int
3343 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3344 {
3345 	int flushed;
3346 
3347 	flushed = flushbufqueues(vp, bd, target, 0);
3348 	if (flushed == 0) {
3349 		/*
3350 		 * Could not find any buffers without rollback
3351 		 * dependencies, so just write the first one
3352 		 * in the hopes of eventually making progress.
3353 		 */
3354 		if (vp != NULL && target > 2)
3355 			target /= 2;
3356 		flushbufqueues(vp, bd, target, 1);
3357 	}
3358 	return (flushed);
3359 }
3360 
3361 static void
3362 buf_daemon()
3363 {
3364 	struct bufdomain *bd;
3365 	int speedupreq;
3366 	int lodirty;
3367 	int i;
3368 
3369 	/*
3370 	 * This process needs to be suspended prior to shutdown sync.
3371 	 */
3372 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3373 	    SHUTDOWN_PRI_LAST + 100);
3374 
3375 	/*
3376 	 * Start the buf clean daemons as children threads.
3377 	 */
3378 	for (i = 0 ; i < buf_domains; i++) {
3379 		int error;
3380 
3381 		error = kthread_add((void (*)(void *))bufspace_daemon,
3382 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3383 		if (error)
3384 			panic("error %d spawning bufspace daemon", error);
3385 	}
3386 
3387 	/*
3388 	 * This process is allowed to take the buffer cache to the limit
3389 	 */
3390 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3391 	mtx_lock(&bdlock);
3392 	for (;;) {
3393 		bd_request = 0;
3394 		mtx_unlock(&bdlock);
3395 
3396 		kthread_suspend_check();
3397 
3398 		/*
3399 		 * Save speedupreq for this pass and reset to capture new
3400 		 * requests.
3401 		 */
3402 		speedupreq = bd_speedupreq;
3403 		bd_speedupreq = 0;
3404 
3405 		/*
3406 		 * Flush each domain sequentially according to its level and
3407 		 * the speedup request.
3408 		 */
3409 		for (i = 0; i < buf_domains; i++) {
3410 			bd = &bdomain[i];
3411 			if (speedupreq)
3412 				lodirty = bd->bd_numdirtybuffers / 2;
3413 			else
3414 				lodirty = bd->bd_lodirtybuffers;
3415 			while (bd->bd_numdirtybuffers > lodirty) {
3416 				if (buf_flush(NULL, bd,
3417 				    bd->bd_numdirtybuffers - lodirty) == 0)
3418 					break;
3419 				kern_yield(PRI_USER);
3420 			}
3421 		}
3422 
3423 		/*
3424 		 * Only clear bd_request if we have reached our low water
3425 		 * mark.  The buf_daemon normally waits 1 second and
3426 		 * then incrementally flushes any dirty buffers that have
3427 		 * built up, within reason.
3428 		 *
3429 		 * If we were unable to hit our low water mark and couldn't
3430 		 * find any flushable buffers, we sleep for a short period
3431 		 * to avoid endless loops on unlockable buffers.
3432 		 */
3433 		mtx_lock(&bdlock);
3434 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3435 			/*
3436 			 * We reached our low water mark, reset the
3437 			 * request and sleep until we are needed again.
3438 			 * The sleep is just so the suspend code works.
3439 			 */
3440 			bd_request = 0;
3441 			/*
3442 			 * Do an extra wakeup in case dirty threshold
3443 			 * changed via sysctl and the explicit transition
3444 			 * out of shortfall was missed.
3445 			 */
3446 			bdirtywakeup();
3447 			if (runningbufspace <= lorunningspace)
3448 				runningwakeup();
3449 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3450 		} else {
3451 			/*
3452 			 * We couldn't find any flushable dirty buffers but
3453 			 * still have too many dirty buffers, we
3454 			 * have to sleep and try again.  (rare)
3455 			 */
3456 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3457 		}
3458 	}
3459 }
3460 
3461 /*
3462  *	flushbufqueues:
3463  *
3464  *	Try to flush a buffer in the dirty queue.  We must be careful to
3465  *	free up B_INVAL buffers instead of write them, which NFS is
3466  *	particularly sensitive to.
3467  */
3468 static int flushwithdeps = 0;
3469 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3470     &flushwithdeps, 0,
3471     "Number of buffers flushed with dependecies that require rollbacks");
3472 
3473 static int
3474 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3475     int flushdeps)
3476 {
3477 	struct bufqueue *bq;
3478 	struct buf *sentinel;
3479 	struct vnode *vp;
3480 	struct mount *mp;
3481 	struct buf *bp;
3482 	int hasdeps;
3483 	int flushed;
3484 	int error;
3485 	bool unlock;
3486 
3487 	flushed = 0;
3488 	bq = &bd->bd_dirtyq;
3489 	bp = NULL;
3490 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3491 	sentinel->b_qindex = QUEUE_SENTINEL;
3492 	BQ_LOCK(bq);
3493 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3494 	BQ_UNLOCK(bq);
3495 	while (flushed != target) {
3496 		maybe_yield();
3497 		BQ_LOCK(bq);
3498 		bp = TAILQ_NEXT(sentinel, b_freelist);
3499 		if (bp != NULL) {
3500 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3501 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3502 			    b_freelist);
3503 		} else {
3504 			BQ_UNLOCK(bq);
3505 			break;
3506 		}
3507 		/*
3508 		 * Skip sentinels inserted by other invocations of the
3509 		 * flushbufqueues(), taking care to not reorder them.
3510 		 *
3511 		 * Only flush the buffers that belong to the
3512 		 * vnode locked by the curthread.
3513 		 */
3514 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3515 		    bp->b_vp != lvp)) {
3516 			BQ_UNLOCK(bq);
3517 			continue;
3518 		}
3519 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3520 		BQ_UNLOCK(bq);
3521 		if (error != 0)
3522 			continue;
3523 
3524 		/*
3525 		 * BKGRDINPROG can only be set with the buf and bufobj
3526 		 * locks both held.  We tolerate a race to clear it here.
3527 		 */
3528 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3529 		    (bp->b_flags & B_DELWRI) == 0) {
3530 			BUF_UNLOCK(bp);
3531 			continue;
3532 		}
3533 		if (bp->b_flags & B_INVAL) {
3534 			bremfreef(bp);
3535 			brelse(bp);
3536 			flushed++;
3537 			continue;
3538 		}
3539 
3540 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3541 			if (flushdeps == 0) {
3542 				BUF_UNLOCK(bp);
3543 				continue;
3544 			}
3545 			hasdeps = 1;
3546 		} else
3547 			hasdeps = 0;
3548 		/*
3549 		 * We must hold the lock on a vnode before writing
3550 		 * one of its buffers. Otherwise we may confuse, or
3551 		 * in the case of a snapshot vnode, deadlock the
3552 		 * system.
3553 		 *
3554 		 * The lock order here is the reverse of the normal
3555 		 * of vnode followed by buf lock.  This is ok because
3556 		 * the NOWAIT will prevent deadlock.
3557 		 */
3558 		vp = bp->b_vp;
3559 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3560 			BUF_UNLOCK(bp);
3561 			continue;
3562 		}
3563 		if (lvp == NULL) {
3564 			unlock = true;
3565 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3566 		} else {
3567 			ASSERT_VOP_LOCKED(vp, "getbuf");
3568 			unlock = false;
3569 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3570 			    vn_lock(vp, LK_TRYUPGRADE);
3571 		}
3572 		if (error == 0) {
3573 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3574 			    bp, bp->b_vp, bp->b_flags);
3575 			if (curproc == bufdaemonproc) {
3576 				vfs_bio_awrite(bp);
3577 			} else {
3578 				bremfree(bp);
3579 				bwrite(bp);
3580 				counter_u64_add(notbufdflushes, 1);
3581 			}
3582 			vn_finished_write(mp);
3583 			if (unlock)
3584 				VOP_UNLOCK(vp);
3585 			flushwithdeps += hasdeps;
3586 			flushed++;
3587 
3588 			/*
3589 			 * Sleeping on runningbufspace while holding
3590 			 * vnode lock leads to deadlock.
3591 			 */
3592 			if (curproc == bufdaemonproc &&
3593 			    runningbufspace > hirunningspace)
3594 				waitrunningbufspace();
3595 			continue;
3596 		}
3597 		vn_finished_write(mp);
3598 		BUF_UNLOCK(bp);
3599 	}
3600 	BQ_LOCK(bq);
3601 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3602 	BQ_UNLOCK(bq);
3603 	free(sentinel, M_TEMP);
3604 	return (flushed);
3605 }
3606 
3607 /*
3608  * Check to see if a block is currently memory resident.
3609  */
3610 struct buf *
3611 incore(struct bufobj *bo, daddr_t blkno)
3612 {
3613 	return (gbincore_unlocked(bo, blkno));
3614 }
3615 
3616 /*
3617  * Returns true if no I/O is needed to access the
3618  * associated VM object.  This is like incore except
3619  * it also hunts around in the VM system for the data.
3620  */
3621 bool
3622 inmem(struct vnode * vp, daddr_t blkno)
3623 {
3624 	vm_object_t obj;
3625 	vm_offset_t toff, tinc, size;
3626 	vm_page_t m, n;
3627 	vm_ooffset_t off;
3628 	int valid;
3629 
3630 	ASSERT_VOP_LOCKED(vp, "inmem");
3631 
3632 	if (incore(&vp->v_bufobj, blkno))
3633 		return (true);
3634 	if (vp->v_mount == NULL)
3635 		return (false);
3636 	obj = vp->v_object;
3637 	if (obj == NULL)
3638 		return (false);
3639 
3640 	size = PAGE_SIZE;
3641 	if (size > vp->v_mount->mnt_stat.f_iosize)
3642 		size = vp->v_mount->mnt_stat.f_iosize;
3643 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3644 
3645 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3646 		m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3647 recheck:
3648 		if (m == NULL)
3649 			return (false);
3650 
3651 		tinc = size;
3652 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3653 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3654 		/*
3655 		 * Consider page validity only if page mapping didn't change
3656 		 * during the check.
3657 		 */
3658 		valid = vm_page_is_valid(m,
3659 		    (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3660 		n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3661 		if (m != n) {
3662 			m = n;
3663 			goto recheck;
3664 		}
3665 		if (!valid)
3666 			return (false);
3667 	}
3668 	return (true);
3669 }
3670 
3671 /*
3672  * Set the dirty range for a buffer based on the status of the dirty
3673  * bits in the pages comprising the buffer.  The range is limited
3674  * to the size of the buffer.
3675  *
3676  * Tell the VM system that the pages associated with this buffer
3677  * are clean.  This is used for delayed writes where the data is
3678  * going to go to disk eventually without additional VM intevention.
3679  *
3680  * Note that while we only really need to clean through to b_bcount, we
3681  * just go ahead and clean through to b_bufsize.
3682  */
3683 static void
3684 vfs_clean_pages_dirty_buf(struct buf *bp)
3685 {
3686 	vm_ooffset_t foff, noff, eoff;
3687 	vm_page_t m;
3688 	int i;
3689 
3690 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3691 		return;
3692 
3693 	foff = bp->b_offset;
3694 	KASSERT(bp->b_offset != NOOFFSET,
3695 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3696 
3697 	vfs_busy_pages_acquire(bp);
3698 	vfs_setdirty_range(bp);
3699 	for (i = 0; i < bp->b_npages; i++) {
3700 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3701 		eoff = noff;
3702 		if (eoff > bp->b_offset + bp->b_bufsize)
3703 			eoff = bp->b_offset + bp->b_bufsize;
3704 		m = bp->b_pages[i];
3705 		vfs_page_set_validclean(bp, foff, m);
3706 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3707 		foff = noff;
3708 	}
3709 	vfs_busy_pages_release(bp);
3710 }
3711 
3712 static void
3713 vfs_setdirty_range(struct buf *bp)
3714 {
3715 	vm_offset_t boffset;
3716 	vm_offset_t eoffset;
3717 	int i;
3718 
3719 	/*
3720 	 * test the pages to see if they have been modified directly
3721 	 * by users through the VM system.
3722 	 */
3723 	for (i = 0; i < bp->b_npages; i++)
3724 		vm_page_test_dirty(bp->b_pages[i]);
3725 
3726 	/*
3727 	 * Calculate the encompassing dirty range, boffset and eoffset,
3728 	 * (eoffset - boffset) bytes.
3729 	 */
3730 
3731 	for (i = 0; i < bp->b_npages; i++) {
3732 		if (bp->b_pages[i]->dirty)
3733 			break;
3734 	}
3735 	boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3736 
3737 	for (i = bp->b_npages - 1; i >= 0; --i) {
3738 		if (bp->b_pages[i]->dirty) {
3739 			break;
3740 		}
3741 	}
3742 	eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3743 
3744 	/*
3745 	 * Fit it to the buffer.
3746 	 */
3747 
3748 	if (eoffset > bp->b_bcount)
3749 		eoffset = bp->b_bcount;
3750 
3751 	/*
3752 	 * If we have a good dirty range, merge with the existing
3753 	 * dirty range.
3754 	 */
3755 
3756 	if (boffset < eoffset) {
3757 		if (bp->b_dirtyoff > boffset)
3758 			bp->b_dirtyoff = boffset;
3759 		if (bp->b_dirtyend < eoffset)
3760 			bp->b_dirtyend = eoffset;
3761 	}
3762 }
3763 
3764 /*
3765  * Allocate the KVA mapping for an existing buffer.
3766  * If an unmapped buffer is provided but a mapped buffer is requested, take
3767  * also care to properly setup mappings between pages and KVA.
3768  */
3769 static void
3770 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3771 {
3772 	int bsize, maxsize, need_mapping, need_kva;
3773 	off_t offset;
3774 
3775 	need_mapping = bp->b_data == unmapped_buf &&
3776 	    (gbflags & GB_UNMAPPED) == 0;
3777 	need_kva = bp->b_kvabase == unmapped_buf &&
3778 	    bp->b_data == unmapped_buf &&
3779 	    (gbflags & GB_KVAALLOC) != 0;
3780 	if (!need_mapping && !need_kva)
3781 		return;
3782 
3783 	BUF_CHECK_UNMAPPED(bp);
3784 
3785 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3786 		/*
3787 		 * Buffer is not mapped, but the KVA was already
3788 		 * reserved at the time of the instantiation.  Use the
3789 		 * allocated space.
3790 		 */
3791 		goto has_addr;
3792 	}
3793 
3794 	/*
3795 	 * Calculate the amount of the address space we would reserve
3796 	 * if the buffer was mapped.
3797 	 */
3798 	bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3799 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3800 	offset = blkno * bsize;
3801 	maxsize = size + (offset & PAGE_MASK);
3802 	maxsize = imax(maxsize, bsize);
3803 
3804 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3805 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3806 			/*
3807 			 * XXXKIB: defragmentation cannot
3808 			 * succeed, not sure what else to do.
3809 			 */
3810 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3811 		}
3812 		counter_u64_add(mappingrestarts, 1);
3813 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3814 	}
3815 has_addr:
3816 	if (need_mapping) {
3817 		/* b_offset is handled by bpmap_qenter. */
3818 		bp->b_data = bp->b_kvabase;
3819 		BUF_CHECK_MAPPED(bp);
3820 		bpmap_qenter(bp);
3821 	}
3822 }
3823 
3824 struct buf *
3825 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3826     int flags)
3827 {
3828 	struct buf *bp;
3829 	int error;
3830 
3831 	error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3832 	if (error != 0)
3833 		return (NULL);
3834 	return (bp);
3835 }
3836 
3837 /*
3838  *	getblkx:
3839  *
3840  *	Get a block given a specified block and offset into a file/device.
3841  *	The buffers B_DONE bit will be cleared on return, making it almost
3842  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3843  *	return.  The caller should clear B_INVAL prior to initiating a
3844  *	READ.
3845  *
3846  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3847  *	an existing buffer.
3848  *
3849  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3850  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3851  *	and then cleared based on the backing VM.  If the previous buffer is
3852  *	non-0-sized but invalid, B_CACHE will be cleared.
3853  *
3854  *	If getblk() must create a new buffer, the new buffer is returned with
3855  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3856  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3857  *	backing VM.
3858  *
3859  *	getblk() also forces a bwrite() for any B_DELWRI buffer whose
3860  *	B_CACHE bit is clear.
3861  *
3862  *	What this means, basically, is that the caller should use B_CACHE to
3863  *	determine whether the buffer is fully valid or not and should clear
3864  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3865  *	the buffer by loading its data area with something, the caller needs
3866  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3867  *	the caller should set B_CACHE ( as an optimization ), else the caller
3868  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3869  *	a write attempt or if it was a successful read.  If the caller
3870  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3871  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3872  *
3873  *	The blkno parameter is the logical block being requested. Normally
3874  *	the mapping of logical block number to disk block address is done
3875  *	by calling VOP_BMAP(). However, if the mapping is already known, the
3876  *	disk block address can be passed using the dblkno parameter. If the
3877  *	disk block address is not known, then the same value should be passed
3878  *	for blkno and dblkno.
3879  */
3880 int
3881 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3882     int slptimeo, int flags, struct buf **bpp)
3883 {
3884 	struct buf *bp;
3885 	struct bufobj *bo;
3886 	daddr_t d_blkno;
3887 	int bsize, error, maxsize, vmio;
3888 	off_t offset;
3889 
3890 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3891 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3892 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3893 	ASSERT_VOP_LOCKED(vp, "getblk");
3894 	if (size > maxbcachebuf)
3895 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3896 		    maxbcachebuf);
3897 	if (!unmapped_buf_allowed)
3898 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3899 
3900 	bo = &vp->v_bufobj;
3901 	d_blkno = dblkno;
3902 
3903 	/* Attempt lockless lookup first. */
3904 	bp = gbincore_unlocked(bo, blkno);
3905 	if (bp == NULL) {
3906 		/*
3907 		 * With GB_NOCREAT we must be sure about not finding the buffer
3908 		 * as it may have been reassigned during unlocked lookup.
3909 		 */
3910 		if ((flags & GB_NOCREAT) != 0)
3911 			goto loop;
3912 		goto newbuf_unlocked;
3913 	}
3914 
3915 	error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3916 	    0);
3917 	if (error != 0)
3918 		goto loop;
3919 
3920 	/* Verify buf identify has not changed since lookup. */
3921 	if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3922 		goto foundbuf_fastpath;
3923 
3924 	/* It changed, fallback to locked lookup. */
3925 	BUF_UNLOCK_RAW(bp);
3926 
3927 loop:
3928 	BO_RLOCK(bo);
3929 	bp = gbincore(bo, blkno);
3930 	if (bp != NULL) {
3931 		int lockflags;
3932 
3933 		/*
3934 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3935 		 * it must be on a queue.
3936 		 */
3937 		lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3938 		    ((flags & GB_LOCK_NOWAIT) ? LK_NOWAIT : LK_SLEEPFAIL);
3939 
3940 		error = BUF_TIMELOCK(bp, lockflags,
3941 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3942 
3943 		/*
3944 		 * If we slept and got the lock we have to restart in case
3945 		 * the buffer changed identities.
3946 		 */
3947 		if (error == ENOLCK)
3948 			goto loop;
3949 		/* We timed out or were interrupted. */
3950 		else if (error != 0)
3951 			return (error);
3952 
3953 foundbuf_fastpath:
3954 		/* If recursed, assume caller knows the rules. */
3955 		if (BUF_LOCKRECURSED(bp))
3956 			goto end;
3957 
3958 		/*
3959 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3960 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3961 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3962 		 * backing VM cache.
3963 		 */
3964 		if (bp->b_flags & B_INVAL)
3965 			bp->b_flags &= ~B_CACHE;
3966 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3967 			bp->b_flags |= B_CACHE;
3968 		if (bp->b_flags & B_MANAGED)
3969 			MPASS(bp->b_qindex == QUEUE_NONE);
3970 		else
3971 			bremfree(bp);
3972 
3973 		/*
3974 		 * check for size inconsistencies for non-VMIO case.
3975 		 */
3976 		if (bp->b_bcount != size) {
3977 			if ((bp->b_flags & B_VMIO) == 0 ||
3978 			    (size > bp->b_kvasize)) {
3979 				if (bp->b_flags & B_DELWRI) {
3980 					bp->b_flags |= B_NOCACHE;
3981 					bwrite(bp);
3982 				} else {
3983 					if (LIST_EMPTY(&bp->b_dep)) {
3984 						bp->b_flags |= B_RELBUF;
3985 						brelse(bp);
3986 					} else {
3987 						bp->b_flags |= B_NOCACHE;
3988 						bwrite(bp);
3989 					}
3990 				}
3991 				goto loop;
3992 			}
3993 		}
3994 
3995 		/*
3996 		 * Handle the case of unmapped buffer which should
3997 		 * become mapped, or the buffer for which KVA
3998 		 * reservation is requested.
3999 		 */
4000 		bp_unmapped_get_kva(bp, blkno, size, flags);
4001 
4002 		/*
4003 		 * If the size is inconsistent in the VMIO case, we can resize
4004 		 * the buffer.  This might lead to B_CACHE getting set or
4005 		 * cleared.  If the size has not changed, B_CACHE remains
4006 		 * unchanged from its previous state.
4007 		 */
4008 		allocbuf(bp, size);
4009 
4010 		KASSERT(bp->b_offset != NOOFFSET,
4011 		    ("getblk: no buffer offset"));
4012 
4013 		/*
4014 		 * A buffer with B_DELWRI set and B_CACHE clear must
4015 		 * be committed before we can return the buffer in
4016 		 * order to prevent the caller from issuing a read
4017 		 * ( due to B_CACHE not being set ) and overwriting
4018 		 * it.
4019 		 *
4020 		 * Most callers, including NFS and FFS, need this to
4021 		 * operate properly either because they assume they
4022 		 * can issue a read if B_CACHE is not set, or because
4023 		 * ( for example ) an uncached B_DELWRI might loop due
4024 		 * to softupdates re-dirtying the buffer.  In the latter
4025 		 * case, B_CACHE is set after the first write completes,
4026 		 * preventing further loops.
4027 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
4028 		 * above while extending the buffer, we cannot allow the
4029 		 * buffer to remain with B_CACHE set after the write
4030 		 * completes or it will represent a corrupt state.  To
4031 		 * deal with this we set B_NOCACHE to scrap the buffer
4032 		 * after the write.
4033 		 *
4034 		 * We might be able to do something fancy, like setting
4035 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
4036 		 * so the below call doesn't set B_CACHE, but that gets real
4037 		 * confusing.  This is much easier.
4038 		 */
4039 
4040 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4041 			bp->b_flags |= B_NOCACHE;
4042 			bwrite(bp);
4043 			goto loop;
4044 		}
4045 		bp->b_flags &= ~B_DONE;
4046 	} else {
4047 		/*
4048 		 * Buffer is not in-core, create new buffer.  The buffer
4049 		 * returned by getnewbuf() is locked.  Note that the returned
4050 		 * buffer is also considered valid (not marked B_INVAL).
4051 		 */
4052 		BO_RUNLOCK(bo);
4053 newbuf_unlocked:
4054 		/*
4055 		 * If the user does not want us to create the buffer, bail out
4056 		 * here.
4057 		 */
4058 		if (flags & GB_NOCREAT)
4059 			return (EEXIST);
4060 
4061 		bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4062 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4063 		offset = blkno * bsize;
4064 		vmio = vp->v_object != NULL;
4065 		if (vmio) {
4066 			maxsize = size + (offset & PAGE_MASK);
4067 		} else {
4068 			maxsize = size;
4069 			/* Do not allow non-VMIO notmapped buffers. */
4070 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4071 		}
4072 		maxsize = imax(maxsize, bsize);
4073 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4074 		    !vn_isdisk(vp)) {
4075 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4076 			KASSERT(error != EOPNOTSUPP,
4077 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4078 			    vp));
4079 			if (error != 0)
4080 				return (error);
4081 			if (d_blkno == -1)
4082 				return (EJUSTRETURN);
4083 		}
4084 
4085 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4086 		if (bp == NULL) {
4087 			if (slpflag || slptimeo)
4088 				return (ETIMEDOUT);
4089 			/*
4090 			 * XXX This is here until the sleep path is diagnosed
4091 			 * enough to work under very low memory conditions.
4092 			 *
4093 			 * There's an issue on low memory, 4BSD+non-preempt
4094 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4095 			 * exhaustion occurs without sleeping for buffer
4096 			 * reclaimation.  This just sticks in a loop and
4097 			 * constantly attempts to allocate a buffer, which
4098 			 * hits exhaustion and tries to wakeup bufdaemon.
4099 			 * This never happens because we never yield.
4100 			 *
4101 			 * The real solution is to identify and fix these cases
4102 			 * so we aren't effectively busy-waiting in a loop
4103 			 * until the reclaimation path has cycles to run.
4104 			 */
4105 			kern_yield(PRI_USER);
4106 			goto loop;
4107 		}
4108 
4109 		/*
4110 		 * This code is used to make sure that a buffer is not
4111 		 * created while the getnewbuf routine is blocked.
4112 		 * This can be a problem whether the vnode is locked or not.
4113 		 * If the buffer is created out from under us, we have to
4114 		 * throw away the one we just created.
4115 		 *
4116 		 * Note: this must occur before we associate the buffer
4117 		 * with the vp especially considering limitations in
4118 		 * the splay tree implementation when dealing with duplicate
4119 		 * lblkno's.
4120 		 */
4121 		BO_LOCK(bo);
4122 		if (gbincore(bo, blkno)) {
4123 			BO_UNLOCK(bo);
4124 			bp->b_flags |= B_INVAL;
4125 			bufspace_release(bufdomain(bp), maxsize);
4126 			brelse(bp);
4127 			goto loop;
4128 		}
4129 
4130 		/*
4131 		 * Insert the buffer into the hash, so that it can
4132 		 * be found by incore.
4133 		 */
4134 		bp->b_lblkno = blkno;
4135 		bp->b_blkno = d_blkno;
4136 		bp->b_offset = offset;
4137 		bgetvp(vp, bp);
4138 		BO_UNLOCK(bo);
4139 
4140 		/*
4141 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4142 		 * buffer size starts out as 0, B_CACHE will be set by
4143 		 * allocbuf() for the VMIO case prior to it testing the
4144 		 * backing store for validity.
4145 		 */
4146 
4147 		if (vmio) {
4148 			bp->b_flags |= B_VMIO;
4149 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4150 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4151 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4152 		} else {
4153 			bp->b_flags &= ~B_VMIO;
4154 			KASSERT(bp->b_bufobj->bo_object == NULL,
4155 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4156 			    bp, bp->b_bufobj->bo_object));
4157 			BUF_CHECK_MAPPED(bp);
4158 		}
4159 
4160 		allocbuf(bp, size);
4161 		bufspace_release(bufdomain(bp), maxsize);
4162 		bp->b_flags &= ~B_DONE;
4163 	}
4164 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4165 end:
4166 	buf_track(bp, __func__);
4167 	KASSERT(bp->b_bufobj == bo,
4168 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4169 	*bpp = bp;
4170 	return (0);
4171 }
4172 
4173 /*
4174  * Get an empty, disassociated buffer of given size.  The buffer is initially
4175  * set to B_INVAL.
4176  */
4177 struct buf *
4178 geteblk(int size, int flags)
4179 {
4180 	struct buf *bp;
4181 	int maxsize;
4182 
4183 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4184 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4185 		if ((flags & GB_NOWAIT_BD) &&
4186 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4187 			return (NULL);
4188 	}
4189 	allocbuf(bp, size);
4190 	bufspace_release(bufdomain(bp), maxsize);
4191 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4192 	return (bp);
4193 }
4194 
4195 /*
4196  * Truncate the backing store for a non-vmio buffer.
4197  */
4198 static void
4199 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4200 {
4201 
4202 	if (bp->b_flags & B_MALLOC) {
4203 		/*
4204 		 * malloced buffers are not shrunk
4205 		 */
4206 		if (newbsize == 0) {
4207 			bufmallocadjust(bp, 0);
4208 			free(bp->b_data, M_BIOBUF);
4209 			bp->b_data = bp->b_kvabase;
4210 			bp->b_flags &= ~B_MALLOC;
4211 		}
4212 		return;
4213 	}
4214 	vm_hold_free_pages(bp, newbsize);
4215 	bufspace_adjust(bp, newbsize);
4216 }
4217 
4218 /*
4219  * Extend the backing for a non-VMIO buffer.
4220  */
4221 static void
4222 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4223 {
4224 	caddr_t origbuf;
4225 	int origbufsize;
4226 
4227 	/*
4228 	 * We only use malloced memory on the first allocation.
4229 	 * and revert to page-allocated memory when the buffer
4230 	 * grows.
4231 	 *
4232 	 * There is a potential smp race here that could lead
4233 	 * to bufmallocspace slightly passing the max.  It
4234 	 * is probably extremely rare and not worth worrying
4235 	 * over.
4236 	 */
4237 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4238 	    bufmallocspace < maxbufmallocspace) {
4239 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4240 		bp->b_flags |= B_MALLOC;
4241 		bufmallocadjust(bp, newbsize);
4242 		return;
4243 	}
4244 
4245 	/*
4246 	 * If the buffer is growing on its other-than-first
4247 	 * allocation then we revert to the page-allocation
4248 	 * scheme.
4249 	 */
4250 	origbuf = NULL;
4251 	origbufsize = 0;
4252 	if (bp->b_flags & B_MALLOC) {
4253 		origbuf = bp->b_data;
4254 		origbufsize = bp->b_bufsize;
4255 		bp->b_data = bp->b_kvabase;
4256 		bufmallocadjust(bp, 0);
4257 		bp->b_flags &= ~B_MALLOC;
4258 		newbsize = round_page(newbsize);
4259 	}
4260 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4261 	    (vm_offset_t) bp->b_data + newbsize);
4262 	if (origbuf != NULL) {
4263 		bcopy(origbuf, bp->b_data, origbufsize);
4264 		free(origbuf, M_BIOBUF);
4265 	}
4266 	bufspace_adjust(bp, newbsize);
4267 }
4268 
4269 /*
4270  * This code constitutes the buffer memory from either anonymous system
4271  * memory (in the case of non-VMIO operations) or from an associated
4272  * VM object (in the case of VMIO operations).  This code is able to
4273  * resize a buffer up or down.
4274  *
4275  * Note that this code is tricky, and has many complications to resolve
4276  * deadlock or inconsistent data situations.  Tread lightly!!!
4277  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4278  * the caller.  Calling this code willy nilly can result in the loss of data.
4279  *
4280  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4281  * B_CACHE for the non-VMIO case.
4282  */
4283 int
4284 allocbuf(struct buf *bp, int size)
4285 {
4286 	int newbsize;
4287 
4288 	if (bp->b_bcount == size)
4289 		return (1);
4290 
4291 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4292 		panic("allocbuf: buffer too small");
4293 
4294 	newbsize = roundup2(size, DEV_BSIZE);
4295 	if ((bp->b_flags & B_VMIO) == 0) {
4296 		if ((bp->b_flags & B_MALLOC) == 0)
4297 			newbsize = round_page(newbsize);
4298 		/*
4299 		 * Just get anonymous memory from the kernel.  Don't
4300 		 * mess with B_CACHE.
4301 		 */
4302 		if (newbsize < bp->b_bufsize)
4303 			vfs_nonvmio_truncate(bp, newbsize);
4304 		else if (newbsize > bp->b_bufsize)
4305 			vfs_nonvmio_extend(bp, newbsize);
4306 	} else {
4307 		int desiredpages;
4308 
4309 		desiredpages = (size == 0) ? 0 :
4310 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4311 
4312 		if (bp->b_flags & B_MALLOC)
4313 			panic("allocbuf: VMIO buffer can't be malloced");
4314 		/*
4315 		 * Set B_CACHE initially if buffer is 0 length or will become
4316 		 * 0-length.
4317 		 */
4318 		if (size == 0 || bp->b_bufsize == 0)
4319 			bp->b_flags |= B_CACHE;
4320 
4321 		if (newbsize < bp->b_bufsize)
4322 			vfs_vmio_truncate(bp, desiredpages);
4323 		/* XXX This looks as if it should be newbsize > b_bufsize */
4324 		else if (size > bp->b_bcount)
4325 			vfs_vmio_extend(bp, desiredpages, size);
4326 		bufspace_adjust(bp, newbsize);
4327 	}
4328 	bp->b_bcount = size;		/* requested buffer size. */
4329 	return (1);
4330 }
4331 
4332 extern int inflight_transient_maps;
4333 
4334 static struct bio_queue nondump_bios;
4335 
4336 void
4337 biodone(struct bio *bp)
4338 {
4339 	struct mtx *mtxp;
4340 	void (*done)(struct bio *);
4341 	vm_offset_t start, end;
4342 
4343 	biotrack(bp, __func__);
4344 
4345 	/*
4346 	 * Avoid completing I/O when dumping after a panic since that may
4347 	 * result in a deadlock in the filesystem or pager code.  Note that
4348 	 * this doesn't affect dumps that were started manually since we aim
4349 	 * to keep the system usable after it has been resumed.
4350 	 */
4351 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4352 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4353 		return;
4354 	}
4355 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4356 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4357 		bp->bio_flags |= BIO_UNMAPPED;
4358 		start = trunc_page((vm_offset_t)bp->bio_data);
4359 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4360 		bp->bio_data = unmapped_buf;
4361 		pmap_qremove(start, atop(end - start));
4362 		vmem_free(transient_arena, start, end - start);
4363 		atomic_add_int(&inflight_transient_maps, -1);
4364 	}
4365 	done = bp->bio_done;
4366 	if (done == NULL) {
4367 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4368 		mtx_lock(mtxp);
4369 		bp->bio_flags |= BIO_DONE;
4370 		wakeup(bp);
4371 		mtx_unlock(mtxp);
4372 	} else
4373 		done(bp);
4374 }
4375 
4376 /*
4377  * Wait for a BIO to finish.
4378  */
4379 int
4380 biowait(struct bio *bp, const char *wchan)
4381 {
4382 	struct mtx *mtxp;
4383 
4384 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4385 	mtx_lock(mtxp);
4386 	while ((bp->bio_flags & BIO_DONE) == 0)
4387 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4388 	mtx_unlock(mtxp);
4389 	if (bp->bio_error != 0)
4390 		return (bp->bio_error);
4391 	if (!(bp->bio_flags & BIO_ERROR))
4392 		return (0);
4393 	return (EIO);
4394 }
4395 
4396 void
4397 biofinish(struct bio *bp, struct devstat *stat, int error)
4398 {
4399 
4400 	if (error) {
4401 		bp->bio_error = error;
4402 		bp->bio_flags |= BIO_ERROR;
4403 	}
4404 	if (stat != NULL)
4405 		devstat_end_transaction_bio(stat, bp);
4406 	biodone(bp);
4407 }
4408 
4409 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4410 void
4411 biotrack_buf(struct bio *bp, const char *location)
4412 {
4413 
4414 	buf_track(bp->bio_track_bp, location);
4415 }
4416 #endif
4417 
4418 /*
4419  *	bufwait:
4420  *
4421  *	Wait for buffer I/O completion, returning error status.  The buffer
4422  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4423  *	error and cleared.
4424  */
4425 int
4426 bufwait(struct buf *bp)
4427 {
4428 	if (bp->b_iocmd == BIO_READ)
4429 		bwait(bp, PRIBIO, "biord");
4430 	else
4431 		bwait(bp, PRIBIO, "biowr");
4432 	if (bp->b_flags & B_EINTR) {
4433 		bp->b_flags &= ~B_EINTR;
4434 		return (EINTR);
4435 	}
4436 	if (bp->b_ioflags & BIO_ERROR) {
4437 		return (bp->b_error ? bp->b_error : EIO);
4438 	} else {
4439 		return (0);
4440 	}
4441 }
4442 
4443 /*
4444  *	bufdone:
4445  *
4446  *	Finish I/O on a buffer, optionally calling a completion function.
4447  *	This is usually called from an interrupt so process blocking is
4448  *	not allowed.
4449  *
4450  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4451  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4452  *	assuming B_INVAL is clear.
4453  *
4454  *	For the VMIO case, we set B_CACHE if the op was a read and no
4455  *	read error occurred, or if the op was a write.  B_CACHE is never
4456  *	set if the buffer is invalid or otherwise uncacheable.
4457  *
4458  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4459  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4460  *	in the biodone routine.
4461  */
4462 void
4463 bufdone(struct buf *bp)
4464 {
4465 	struct bufobj *dropobj;
4466 	void    (*biodone)(struct buf *);
4467 
4468 	buf_track(bp, __func__);
4469 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4470 	dropobj = NULL;
4471 
4472 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4473 
4474 	runningbufwakeup(bp);
4475 	if (bp->b_iocmd == BIO_WRITE)
4476 		dropobj = bp->b_bufobj;
4477 	/* call optional completion function if requested */
4478 	if (bp->b_iodone != NULL) {
4479 		biodone = bp->b_iodone;
4480 		bp->b_iodone = NULL;
4481 		(*biodone) (bp);
4482 		if (dropobj)
4483 			bufobj_wdrop(dropobj);
4484 		return;
4485 	}
4486 	if (bp->b_flags & B_VMIO) {
4487 		/*
4488 		 * Set B_CACHE if the op was a normal read and no error
4489 		 * occurred.  B_CACHE is set for writes in the b*write()
4490 		 * routines.
4491 		 */
4492 		if (bp->b_iocmd == BIO_READ &&
4493 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4494 		    !(bp->b_ioflags & BIO_ERROR))
4495 			bp->b_flags |= B_CACHE;
4496 		vfs_vmio_iodone(bp);
4497 	}
4498 	if (!LIST_EMPTY(&bp->b_dep))
4499 		buf_complete(bp);
4500 	if ((bp->b_flags & B_CKHASH) != 0) {
4501 		KASSERT(bp->b_iocmd == BIO_READ,
4502 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4503 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4504 		(*bp->b_ckhashcalc)(bp);
4505 	}
4506 	/*
4507 	 * For asynchronous completions, release the buffer now. The brelse
4508 	 * will do a wakeup there if necessary - so no need to do a wakeup
4509 	 * here in the async case. The sync case always needs to do a wakeup.
4510 	 */
4511 	if (bp->b_flags & B_ASYNC) {
4512 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4513 		    (bp->b_ioflags & BIO_ERROR))
4514 			brelse(bp);
4515 		else
4516 			bqrelse(bp);
4517 	} else
4518 		bdone(bp);
4519 	if (dropobj)
4520 		bufobj_wdrop(dropobj);
4521 }
4522 
4523 /*
4524  * This routine is called in lieu of iodone in the case of
4525  * incomplete I/O.  This keeps the busy status for pages
4526  * consistent.
4527  */
4528 void
4529 vfs_unbusy_pages(struct buf *bp)
4530 {
4531 	int i;
4532 	vm_object_t obj;
4533 	vm_page_t m;
4534 
4535 	runningbufwakeup(bp);
4536 	if (!(bp->b_flags & B_VMIO))
4537 		return;
4538 
4539 	obj = bp->b_bufobj->bo_object;
4540 	for (i = 0; i < bp->b_npages; i++) {
4541 		m = bp->b_pages[i];
4542 		if (m == bogus_page) {
4543 			m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4544 			if (!m)
4545 				panic("vfs_unbusy_pages: page missing\n");
4546 			bp->b_pages[i] = m;
4547 			if (buf_mapped(bp)) {
4548 				BUF_CHECK_MAPPED(bp);
4549 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4550 				    bp->b_pages, bp->b_npages);
4551 			} else
4552 				BUF_CHECK_UNMAPPED(bp);
4553 		}
4554 		vm_page_sunbusy(m);
4555 	}
4556 	vm_object_pip_wakeupn(obj, bp->b_npages);
4557 }
4558 
4559 /*
4560  * vfs_page_set_valid:
4561  *
4562  *	Set the valid bits in a page based on the supplied offset.   The
4563  *	range is restricted to the buffer's size.
4564  *
4565  *	This routine is typically called after a read completes.
4566  */
4567 static void
4568 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4569 {
4570 	vm_ooffset_t eoff;
4571 
4572 	/*
4573 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4574 	 * page boundary and eoff is not greater than the end of the buffer.
4575 	 * The end of the buffer, in this case, is our file EOF, not the
4576 	 * allocation size of the buffer.
4577 	 */
4578 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4579 	if (eoff > bp->b_offset + bp->b_bcount)
4580 		eoff = bp->b_offset + bp->b_bcount;
4581 
4582 	/*
4583 	 * Set valid range.  This is typically the entire buffer and thus the
4584 	 * entire page.
4585 	 */
4586 	if (eoff > off)
4587 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4588 }
4589 
4590 /*
4591  * vfs_page_set_validclean:
4592  *
4593  *	Set the valid bits and clear the dirty bits in a page based on the
4594  *	supplied offset.   The range is restricted to the buffer's size.
4595  */
4596 static void
4597 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4598 {
4599 	vm_ooffset_t soff, eoff;
4600 
4601 	/*
4602 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4603 	 * page boundary or cross the end of the buffer.  The end of the
4604 	 * buffer, in this case, is our file EOF, not the allocation size
4605 	 * of the buffer.
4606 	 */
4607 	soff = off;
4608 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4609 	if (eoff > bp->b_offset + bp->b_bcount)
4610 		eoff = bp->b_offset + bp->b_bcount;
4611 
4612 	/*
4613 	 * Set valid range.  This is typically the entire buffer and thus the
4614 	 * entire page.
4615 	 */
4616 	if (eoff > soff) {
4617 		vm_page_set_validclean(
4618 		    m,
4619 		   (vm_offset_t) (soff & PAGE_MASK),
4620 		   (vm_offset_t) (eoff - soff)
4621 		);
4622 	}
4623 }
4624 
4625 /*
4626  * Acquire a shared busy on all pages in the buf.
4627  */
4628 void
4629 vfs_busy_pages_acquire(struct buf *bp)
4630 {
4631 	int i;
4632 
4633 	for (i = 0; i < bp->b_npages; i++)
4634 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4635 }
4636 
4637 void
4638 vfs_busy_pages_release(struct buf *bp)
4639 {
4640 	int i;
4641 
4642 	for (i = 0; i < bp->b_npages; i++)
4643 		vm_page_sunbusy(bp->b_pages[i]);
4644 }
4645 
4646 /*
4647  * This routine is called before a device strategy routine.
4648  * It is used to tell the VM system that paging I/O is in
4649  * progress, and treat the pages associated with the buffer
4650  * almost as being exclusive busy.  Also the object paging_in_progress
4651  * flag is handled to make sure that the object doesn't become
4652  * inconsistent.
4653  *
4654  * Since I/O has not been initiated yet, certain buffer flags
4655  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4656  * and should be ignored.
4657  */
4658 void
4659 vfs_busy_pages(struct buf *bp, int clear_modify)
4660 {
4661 	vm_object_t obj;
4662 	vm_ooffset_t foff;
4663 	vm_page_t m;
4664 	int i;
4665 	bool bogus;
4666 
4667 	if (!(bp->b_flags & B_VMIO))
4668 		return;
4669 
4670 	obj = bp->b_bufobj->bo_object;
4671 	foff = bp->b_offset;
4672 	KASSERT(bp->b_offset != NOOFFSET,
4673 	    ("vfs_busy_pages: no buffer offset"));
4674 	if ((bp->b_flags & B_CLUSTER) == 0) {
4675 		vm_object_pip_add(obj, bp->b_npages);
4676 		vfs_busy_pages_acquire(bp);
4677 	}
4678 	if (bp->b_bufsize != 0)
4679 		vfs_setdirty_range(bp);
4680 	bogus = false;
4681 	for (i = 0; i < bp->b_npages; i++) {
4682 		m = bp->b_pages[i];
4683 		vm_page_assert_sbusied(m);
4684 
4685 		/*
4686 		 * When readying a buffer for a read ( i.e
4687 		 * clear_modify == 0 ), it is important to do
4688 		 * bogus_page replacement for valid pages in
4689 		 * partially instantiated buffers.  Partially
4690 		 * instantiated buffers can, in turn, occur when
4691 		 * reconstituting a buffer from its VM backing store
4692 		 * base.  We only have to do this if B_CACHE is
4693 		 * clear ( which causes the I/O to occur in the
4694 		 * first place ).  The replacement prevents the read
4695 		 * I/O from overwriting potentially dirty VM-backed
4696 		 * pages.  XXX bogus page replacement is, uh, bogus.
4697 		 * It may not work properly with small-block devices.
4698 		 * We need to find a better way.
4699 		 */
4700 		if (clear_modify) {
4701 			pmap_remove_write(m);
4702 			vfs_page_set_validclean(bp, foff, m);
4703 		} else if (vm_page_all_valid(m) &&
4704 		    (bp->b_flags & B_CACHE) == 0) {
4705 			bp->b_pages[i] = bogus_page;
4706 			bogus = true;
4707 		}
4708 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4709 	}
4710 	if (bogus && buf_mapped(bp)) {
4711 		BUF_CHECK_MAPPED(bp);
4712 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4713 		    bp->b_pages, bp->b_npages);
4714 	}
4715 }
4716 
4717 /*
4718  *	vfs_bio_set_valid:
4719  *
4720  *	Set the range within the buffer to valid.  The range is
4721  *	relative to the beginning of the buffer, b_offset.  Note that
4722  *	b_offset itself may be offset from the beginning of the first
4723  *	page.
4724  */
4725 void
4726 vfs_bio_set_valid(struct buf *bp, int base, int size)
4727 {
4728 	int i, n;
4729 	vm_page_t m;
4730 
4731 	if (!(bp->b_flags & B_VMIO))
4732 		return;
4733 
4734 	/*
4735 	 * Fixup base to be relative to beginning of first page.
4736 	 * Set initial n to be the maximum number of bytes in the
4737 	 * first page that can be validated.
4738 	 */
4739 	base += (bp->b_offset & PAGE_MASK);
4740 	n = PAGE_SIZE - (base & PAGE_MASK);
4741 
4742 	/*
4743 	 * Busy may not be strictly necessary here because the pages are
4744 	 * unlikely to be fully valid and the vnode lock will synchronize
4745 	 * their access via getpages.  It is grabbed for consistency with
4746 	 * other page validation.
4747 	 */
4748 	vfs_busy_pages_acquire(bp);
4749 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4750 		m = bp->b_pages[i];
4751 		if (n > size)
4752 			n = size;
4753 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4754 		base += n;
4755 		size -= n;
4756 		n = PAGE_SIZE;
4757 	}
4758 	vfs_busy_pages_release(bp);
4759 }
4760 
4761 /*
4762  *	vfs_bio_clrbuf:
4763  *
4764  *	If the specified buffer is a non-VMIO buffer, clear the entire
4765  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4766  *	validate only the previously invalid portions of the buffer.
4767  *	This routine essentially fakes an I/O, so we need to clear
4768  *	BIO_ERROR and B_INVAL.
4769  *
4770  *	Note that while we only theoretically need to clear through b_bcount,
4771  *	we go ahead and clear through b_bufsize.
4772  */
4773 void
4774 vfs_bio_clrbuf(struct buf *bp)
4775 {
4776 	int i, j, mask, sa, ea, slide;
4777 
4778 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4779 		clrbuf(bp);
4780 		return;
4781 	}
4782 	bp->b_flags &= ~B_INVAL;
4783 	bp->b_ioflags &= ~BIO_ERROR;
4784 	vfs_busy_pages_acquire(bp);
4785 	sa = bp->b_offset & PAGE_MASK;
4786 	slide = 0;
4787 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4788 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4789 		ea = slide & PAGE_MASK;
4790 		if (ea == 0)
4791 			ea = PAGE_SIZE;
4792 		if (bp->b_pages[i] == bogus_page)
4793 			continue;
4794 		j = sa / DEV_BSIZE;
4795 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4796 		if ((bp->b_pages[i]->valid & mask) == mask)
4797 			continue;
4798 		if ((bp->b_pages[i]->valid & mask) == 0)
4799 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4800 		else {
4801 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4802 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4803 					pmap_zero_page_area(bp->b_pages[i],
4804 					    sa, DEV_BSIZE);
4805 				}
4806 			}
4807 		}
4808 		vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4809 		    roundup2(ea - sa, DEV_BSIZE));
4810 	}
4811 	vfs_busy_pages_release(bp);
4812 	bp->b_resid = 0;
4813 }
4814 
4815 void
4816 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4817 {
4818 	vm_page_t m;
4819 	int i, n;
4820 
4821 	if (buf_mapped(bp)) {
4822 		BUF_CHECK_MAPPED(bp);
4823 		bzero(bp->b_data + base, size);
4824 	} else {
4825 		BUF_CHECK_UNMAPPED(bp);
4826 		n = PAGE_SIZE - (base & PAGE_MASK);
4827 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4828 			m = bp->b_pages[i];
4829 			if (n > size)
4830 				n = size;
4831 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4832 			base += n;
4833 			size -= n;
4834 			n = PAGE_SIZE;
4835 		}
4836 	}
4837 }
4838 
4839 /*
4840  * Update buffer flags based on I/O request parameters, optionally releasing the
4841  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4842  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4843  * I/O).  Otherwise the buffer is released to the cache.
4844  */
4845 static void
4846 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4847 {
4848 
4849 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4850 	    ("buf %p non-VMIO noreuse", bp));
4851 
4852 	if ((ioflag & IO_DIRECT) != 0)
4853 		bp->b_flags |= B_DIRECT;
4854 	if ((ioflag & IO_EXT) != 0)
4855 		bp->b_xflags |= BX_ALTDATA;
4856 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4857 		bp->b_flags |= B_RELBUF;
4858 		if ((ioflag & IO_NOREUSE) != 0)
4859 			bp->b_flags |= B_NOREUSE;
4860 		if (release)
4861 			brelse(bp);
4862 	} else if (release)
4863 		bqrelse(bp);
4864 }
4865 
4866 void
4867 vfs_bio_brelse(struct buf *bp, int ioflag)
4868 {
4869 
4870 	b_io_dismiss(bp, ioflag, true);
4871 }
4872 
4873 void
4874 vfs_bio_set_flags(struct buf *bp, int ioflag)
4875 {
4876 
4877 	b_io_dismiss(bp, ioflag, false);
4878 }
4879 
4880 /*
4881  * vm_hold_load_pages and vm_hold_free_pages get pages into
4882  * a buffers address space.  The pages are anonymous and are
4883  * not associated with a file object.
4884  */
4885 static void
4886 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4887 {
4888 	vm_offset_t pg;
4889 	vm_page_t p;
4890 	int index;
4891 
4892 	BUF_CHECK_MAPPED(bp);
4893 
4894 	to = round_page(to);
4895 	from = round_page(from);
4896 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4897 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
4898 	KASSERT(to - from <= maxbcachebuf,
4899 	    ("vm_hold_load_pages too large %p %#jx %#jx %u",
4900 	    bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4901 
4902 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4903 		/*
4904 		 * note: must allocate system pages since blocking here
4905 		 * could interfere with paging I/O, no matter which
4906 		 * process we are.
4907 		 */
4908 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4909 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4910 		    VM_ALLOC_WAITOK);
4911 		pmap_qenter(pg, &p, 1);
4912 		bp->b_pages[index] = p;
4913 	}
4914 	bp->b_npages = index;
4915 }
4916 
4917 /* Return pages associated with this buf to the vm system */
4918 static void
4919 vm_hold_free_pages(struct buf *bp, int newbsize)
4920 {
4921 	vm_offset_t from;
4922 	vm_page_t p;
4923 	int index, newnpages;
4924 
4925 	BUF_CHECK_MAPPED(bp);
4926 
4927 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4928 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4929 	if (bp->b_npages > newnpages)
4930 		pmap_qremove(from, bp->b_npages - newnpages);
4931 	for (index = newnpages; index < bp->b_npages; index++) {
4932 		p = bp->b_pages[index];
4933 		bp->b_pages[index] = NULL;
4934 		vm_page_unwire_noq(p);
4935 		vm_page_free(p);
4936 	}
4937 	bp->b_npages = newnpages;
4938 }
4939 
4940 /*
4941  * Map an IO request into kernel virtual address space.
4942  *
4943  * All requests are (re)mapped into kernel VA space.
4944  * Notice that we use b_bufsize for the size of the buffer
4945  * to be mapped.  b_bcount might be modified by the driver.
4946  *
4947  * Note that even if the caller determines that the address space should
4948  * be valid, a race or a smaller-file mapped into a larger space may
4949  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4950  * check the return value.
4951  *
4952  * This function only works with pager buffers.
4953  */
4954 int
4955 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
4956 {
4957 	vm_prot_t prot;
4958 	int pidx;
4959 
4960 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
4961 	prot = VM_PROT_READ;
4962 	if (bp->b_iocmd == BIO_READ)
4963 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4964 	pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4965 	    (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
4966 	if (pidx < 0)
4967 		return (-1);
4968 	bp->b_bufsize = len;
4969 	bp->b_npages = pidx;
4970 	bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
4971 	if (mapbuf || !unmapped_buf_allowed) {
4972 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4973 		bp->b_data = bp->b_kvabase + bp->b_offset;
4974 	} else
4975 		bp->b_data = unmapped_buf;
4976 	return (0);
4977 }
4978 
4979 /*
4980  * Free the io map PTEs associated with this IO operation.
4981  * We also invalidate the TLB entries and restore the original b_addr.
4982  *
4983  * This function only works with pager buffers.
4984  */
4985 void
4986 vunmapbuf(struct buf *bp)
4987 {
4988 	int npages;
4989 
4990 	npages = bp->b_npages;
4991 	if (buf_mapped(bp))
4992 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4993 	vm_page_unhold_pages(bp->b_pages, npages);
4994 
4995 	bp->b_data = unmapped_buf;
4996 }
4997 
4998 void
4999 bdone(struct buf *bp)
5000 {
5001 	struct mtx *mtxp;
5002 
5003 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5004 	mtx_lock(mtxp);
5005 	bp->b_flags |= B_DONE;
5006 	wakeup(bp);
5007 	mtx_unlock(mtxp);
5008 }
5009 
5010 void
5011 bwait(struct buf *bp, u_char pri, const char *wchan)
5012 {
5013 	struct mtx *mtxp;
5014 
5015 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5016 	mtx_lock(mtxp);
5017 	while ((bp->b_flags & B_DONE) == 0)
5018 		msleep(bp, mtxp, pri, wchan, 0);
5019 	mtx_unlock(mtxp);
5020 }
5021 
5022 int
5023 bufsync(struct bufobj *bo, int waitfor)
5024 {
5025 
5026 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5027 }
5028 
5029 void
5030 bufstrategy(struct bufobj *bo, struct buf *bp)
5031 {
5032 	int i __unused;
5033 	struct vnode *vp;
5034 
5035 	vp = bp->b_vp;
5036 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5037 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5038 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5039 	i = VOP_STRATEGY(vp, bp);
5040 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5041 }
5042 
5043 /*
5044  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
5045  */
5046 void
5047 bufobj_init(struct bufobj *bo, void *private)
5048 {
5049 	static volatile int bufobj_cleanq;
5050 
5051         bo->bo_domain =
5052             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5053         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5054         bo->bo_private = private;
5055         TAILQ_INIT(&bo->bo_clean.bv_hd);
5056         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5057 }
5058 
5059 void
5060 bufobj_wrefl(struct bufobj *bo)
5061 {
5062 
5063 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5064 	ASSERT_BO_WLOCKED(bo);
5065 	bo->bo_numoutput++;
5066 }
5067 
5068 void
5069 bufobj_wref(struct bufobj *bo)
5070 {
5071 
5072 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5073 	BO_LOCK(bo);
5074 	bo->bo_numoutput++;
5075 	BO_UNLOCK(bo);
5076 }
5077 
5078 void
5079 bufobj_wdrop(struct bufobj *bo)
5080 {
5081 
5082 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5083 	BO_LOCK(bo);
5084 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5085 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5086 		bo->bo_flag &= ~BO_WWAIT;
5087 		wakeup(&bo->bo_numoutput);
5088 	}
5089 	BO_UNLOCK(bo);
5090 }
5091 
5092 int
5093 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5094 {
5095 	int error;
5096 
5097 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5098 	ASSERT_BO_WLOCKED(bo);
5099 	error = 0;
5100 	while (bo->bo_numoutput) {
5101 		bo->bo_flag |= BO_WWAIT;
5102 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5103 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5104 		if (error)
5105 			break;
5106 	}
5107 	return (error);
5108 }
5109 
5110 /*
5111  * Set bio_data or bio_ma for struct bio from the struct buf.
5112  */
5113 void
5114 bdata2bio(struct buf *bp, struct bio *bip)
5115 {
5116 
5117 	if (!buf_mapped(bp)) {
5118 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5119 		bip->bio_ma = bp->b_pages;
5120 		bip->bio_ma_n = bp->b_npages;
5121 		bip->bio_data = unmapped_buf;
5122 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5123 		bip->bio_flags |= BIO_UNMAPPED;
5124 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5125 		    PAGE_SIZE == bp->b_npages,
5126 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5127 		    (long long)bip->bio_length, bip->bio_ma_n));
5128 	} else {
5129 		bip->bio_data = bp->b_data;
5130 		bip->bio_ma = NULL;
5131 	}
5132 }
5133 
5134 /*
5135  * The MIPS pmap code currently doesn't handle aliased pages.
5136  * The VIPT caches may not handle page aliasing themselves, leading
5137  * to data corruption.
5138  *
5139  * As such, this code makes a system extremely unhappy if said
5140  * system doesn't support unaliasing the above situation in hardware.
5141  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5142  * this feature at build time, so it has to be handled in software.
5143  *
5144  * Once the MIPS pmap/cache code grows to support this function on
5145  * earlier chips, it should be flipped back off.
5146  */
5147 #ifdef	__mips__
5148 static int buf_pager_relbuf = 1;
5149 #else
5150 static int buf_pager_relbuf = 0;
5151 #endif
5152 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5153     &buf_pager_relbuf, 0,
5154     "Make buffer pager release buffers after reading");
5155 
5156 /*
5157  * The buffer pager.  It uses buffer reads to validate pages.
5158  *
5159  * In contrast to the generic local pager from vm/vnode_pager.c, this
5160  * pager correctly and easily handles volumes where the underlying
5161  * device block size is greater than the machine page size.  The
5162  * buffer cache transparently extends the requested page run to be
5163  * aligned at the block boundary, and does the necessary bogus page
5164  * replacements in the addends to avoid obliterating already valid
5165  * pages.
5166  *
5167  * The only non-trivial issue is that the exclusive busy state for
5168  * pages, which is assumed by the vm_pager_getpages() interface, is
5169  * incompatible with the VMIO buffer cache's desire to share-busy the
5170  * pages.  This function performs a trivial downgrade of the pages'
5171  * state before reading buffers, and a less trivial upgrade from the
5172  * shared-busy to excl-busy state after the read.
5173  */
5174 int
5175 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5176     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5177     vbg_get_blksize_t get_blksize)
5178 {
5179 	vm_page_t m;
5180 	vm_object_t object;
5181 	struct buf *bp;
5182 	struct mount *mp;
5183 	daddr_t lbn, lbnp;
5184 	vm_ooffset_t la, lb, poff, poffe;
5185 	long bsize;
5186 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5187 	bool redo, lpart;
5188 
5189 	object = vp->v_object;
5190 	mp = vp->v_mount;
5191 	error = 0;
5192 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5193 	if (la >= object->un_pager.vnp.vnp_size)
5194 		return (VM_PAGER_BAD);
5195 
5196 	/*
5197 	 * Change the meaning of la from where the last requested page starts
5198 	 * to where it ends, because that's the end of the requested region
5199 	 * and the start of the potential read-ahead region.
5200 	 */
5201 	la += PAGE_SIZE;
5202 	lpart = la > object->un_pager.vnp.vnp_size;
5203 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5204 
5205 	/*
5206 	 * Calculate read-ahead, behind and total pages.
5207 	 */
5208 	pgsin = count;
5209 	lb = IDX_TO_OFF(ma[0]->pindex);
5210 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5211 	pgsin += pgsin_b;
5212 	if (rbehind != NULL)
5213 		*rbehind = pgsin_b;
5214 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5215 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5216 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5217 		    PAGE_SIZE) - la);
5218 	pgsin += pgsin_a;
5219 	if (rahead != NULL)
5220 		*rahead = pgsin_a;
5221 	VM_CNT_INC(v_vnodein);
5222 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5223 
5224 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5225 	    != 0) ? GB_UNMAPPED : 0;
5226 again:
5227 	for (i = 0; i < count; i++) {
5228 		if (ma[i] != bogus_page)
5229 			vm_page_busy_downgrade(ma[i]);
5230 	}
5231 
5232 	lbnp = -1;
5233 	for (i = 0; i < count; i++) {
5234 		m = ma[i];
5235 		if (m == bogus_page)
5236 			continue;
5237 
5238 		/*
5239 		 * Pages are shared busy and the object lock is not
5240 		 * owned, which together allow for the pages'
5241 		 * invalidation.  The racy test for validity avoids
5242 		 * useless creation of the buffer for the most typical
5243 		 * case when invalidation is not used in redo or for
5244 		 * parallel read.  The shared->excl upgrade loop at
5245 		 * the end of the function catches the race in a
5246 		 * reliable way (protected by the object lock).
5247 		 */
5248 		if (vm_page_all_valid(m))
5249 			continue;
5250 
5251 		poff = IDX_TO_OFF(m->pindex);
5252 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5253 		for (; poff < poffe; poff += bsize) {
5254 			lbn = get_lblkno(vp, poff);
5255 			if (lbn == lbnp)
5256 				goto next_page;
5257 			lbnp = lbn;
5258 
5259 			bsize = get_blksize(vp, lbn);
5260 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5261 			    br_flags, &bp);
5262 			if (error != 0)
5263 				goto end_pages;
5264 			if (bp->b_rcred == curthread->td_ucred) {
5265 				crfree(bp->b_rcred);
5266 				bp->b_rcred = NOCRED;
5267 			}
5268 			if (LIST_EMPTY(&bp->b_dep)) {
5269 				/*
5270 				 * Invalidation clears m->valid, but
5271 				 * may leave B_CACHE flag if the
5272 				 * buffer existed at the invalidation
5273 				 * time.  In this case, recycle the
5274 				 * buffer to do real read on next
5275 				 * bread() after redo.
5276 				 *
5277 				 * Otherwise B_RELBUF is not strictly
5278 				 * necessary, enable to reduce buf
5279 				 * cache pressure.
5280 				 */
5281 				if (buf_pager_relbuf ||
5282 				    !vm_page_all_valid(m))
5283 					bp->b_flags |= B_RELBUF;
5284 
5285 				bp->b_flags &= ~B_NOCACHE;
5286 				brelse(bp);
5287 			} else {
5288 				bqrelse(bp);
5289 			}
5290 		}
5291 		KASSERT(1 /* racy, enable for debugging */ ||
5292 		    vm_page_all_valid(m) || i == count - 1,
5293 		    ("buf %d %p invalid", i, m));
5294 		if (i == count - 1 && lpart) {
5295 			if (!vm_page_none_valid(m) &&
5296 			    !vm_page_all_valid(m))
5297 				vm_page_zero_invalid(m, TRUE);
5298 		}
5299 next_page:;
5300 	}
5301 end_pages:
5302 
5303 	redo = false;
5304 	for (i = 0; i < count; i++) {
5305 		if (ma[i] == bogus_page)
5306 			continue;
5307 		if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5308 			vm_page_sunbusy(ma[i]);
5309 			ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5310 			    VM_ALLOC_NORMAL);
5311 		}
5312 
5313 		/*
5314 		 * Since the pages were only sbusy while neither the
5315 		 * buffer nor the object lock was held by us, or
5316 		 * reallocated while vm_page_grab() slept for busy
5317 		 * relinguish, they could have been invalidated.
5318 		 * Recheck the valid bits and re-read as needed.
5319 		 *
5320 		 * Note that the last page is made fully valid in the
5321 		 * read loop, and partial validity for the page at
5322 		 * index count - 1 could mean that the page was
5323 		 * invalidated or removed, so we must restart for
5324 		 * safety as well.
5325 		 */
5326 		if (!vm_page_all_valid(ma[i]))
5327 			redo = true;
5328 	}
5329 	if (redo && error == 0)
5330 		goto again;
5331 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5332 }
5333 
5334 #include "opt_ddb.h"
5335 #ifdef DDB
5336 #include <ddb/ddb.h>
5337 
5338 /* DDB command to show buffer data */
5339 DB_SHOW_COMMAND(buffer, db_show_buffer)
5340 {
5341 	/* get args */
5342 	struct buf *bp = (struct buf *)addr;
5343 #ifdef FULL_BUF_TRACKING
5344 	uint32_t i, j;
5345 #endif
5346 
5347 	if (!have_addr) {
5348 		db_printf("usage: show buffer <addr>\n");
5349 		return;
5350 	}
5351 
5352 	db_printf("buf at %p\n", bp);
5353 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5354 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5355 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5356 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5357 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5358 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5359 	db_printf(
5360 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5361 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5362 	    "b_vp = %p, b_dep = %p\n",
5363 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5364 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5365 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5366 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5367 	    bp->b_kvabase, bp->b_kvasize);
5368 	if (bp->b_npages) {
5369 		int i;
5370 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5371 		for (i = 0; i < bp->b_npages; i++) {
5372 			vm_page_t m;
5373 			m = bp->b_pages[i];
5374 			if (m != NULL)
5375 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5376 				    (u_long)m->pindex,
5377 				    (u_long)VM_PAGE_TO_PHYS(m));
5378 			else
5379 				db_printf("( ??? )");
5380 			if ((i + 1) < bp->b_npages)
5381 				db_printf(",");
5382 		}
5383 		db_printf("\n");
5384 	}
5385 	BUF_LOCKPRINTINFO(bp);
5386 #if defined(FULL_BUF_TRACKING)
5387 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5388 
5389 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5390 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5391 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5392 			continue;
5393 		db_printf(" %2u: %s\n", j,
5394 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5395 	}
5396 #elif defined(BUF_TRACKING)
5397 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5398 #endif
5399 	db_printf(" ");
5400 }
5401 
5402 DB_SHOW_COMMAND(bufqueues, bufqueues)
5403 {
5404 	struct bufdomain *bd;
5405 	struct buf *bp;
5406 	long total;
5407 	int i, j, cnt;
5408 
5409 	db_printf("bqempty: %d\n", bqempty.bq_len);
5410 
5411 	for (i = 0; i < buf_domains; i++) {
5412 		bd = &bdomain[i];
5413 		db_printf("Buf domain %d\n", i);
5414 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5415 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5416 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5417 		db_printf("\n");
5418 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5419 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5420 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5421 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5422 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5423 		db_printf("\n");
5424 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5425 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5426 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5427 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5428 		db_printf("\n");
5429 		total = 0;
5430 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5431 			total += bp->b_bufsize;
5432 		db_printf("\tcleanq count\t%d (%ld)\n",
5433 		    bd->bd_cleanq->bq_len, total);
5434 		total = 0;
5435 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5436 			total += bp->b_bufsize;
5437 		db_printf("\tdirtyq count\t%d (%ld)\n",
5438 		    bd->bd_dirtyq.bq_len, total);
5439 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5440 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5441 		db_printf("\tCPU ");
5442 		for (j = 0; j <= mp_maxid; j++)
5443 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5444 		db_printf("\n");
5445 		cnt = 0;
5446 		total = 0;
5447 		for (j = 0; j < nbuf; j++) {
5448 			bp = nbufp(j);
5449 			if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5450 				cnt++;
5451 				total += bp->b_bufsize;
5452 			}
5453 		}
5454 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5455 		cnt = 0;
5456 		total = 0;
5457 		for (j = 0; j < nbuf; j++) {
5458 			bp = nbufp(j);
5459 			if (bp->b_domain == i) {
5460 				cnt++;
5461 				total += bp->b_bufsize;
5462 			}
5463 		}
5464 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5465 	}
5466 }
5467 
5468 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5469 {
5470 	struct buf *bp;
5471 	int i;
5472 
5473 	for (i = 0; i < nbuf; i++) {
5474 		bp = nbufp(i);
5475 		if (BUF_ISLOCKED(bp)) {
5476 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5477 			db_printf("\n");
5478 			if (db_pager_quit)
5479 				break;
5480 		}
5481 	}
5482 }
5483 
5484 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5485 {
5486 	struct vnode *vp;
5487 	struct buf *bp;
5488 
5489 	if (!have_addr) {
5490 		db_printf("usage: show vnodebufs <addr>\n");
5491 		return;
5492 	}
5493 	vp = (struct vnode *)addr;
5494 	db_printf("Clean buffers:\n");
5495 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5496 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5497 		db_printf("\n");
5498 	}
5499 	db_printf("Dirty buffers:\n");
5500 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5501 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5502 		db_printf("\n");
5503 	}
5504 }
5505 
5506 DB_COMMAND(countfreebufs, db_coundfreebufs)
5507 {
5508 	struct buf *bp;
5509 	int i, used = 0, nfree = 0;
5510 
5511 	if (have_addr) {
5512 		db_printf("usage: countfreebufs\n");
5513 		return;
5514 	}
5515 
5516 	for (i = 0; i < nbuf; i++) {
5517 		bp = nbufp(i);
5518 		if (bp->b_qindex == QUEUE_EMPTY)
5519 			nfree++;
5520 		else
5521 			used++;
5522 	}
5523 
5524 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5525 	    nfree + used);
5526 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5527 }
5528 #endif /* DDB */
5529