xref: /freebsd/sys/kern/vfs_bio.c (revision 685dc743dc3b5645e34836464128e1c0558b404b)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/asan.h>
51 #include <sys/bio.h>
52 #include <sys/bitset.h>
53 #include <sys/boottrace.h>
54 #include <sys/buf.h>
55 #include <sys/conf.h>
56 #include <sys/counter.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/memdesc.h>
65 #include <sys/mount.h>
66 #include <sys/mutex.h>
67 #include <sys/kernel.h>
68 #include <sys/kthread.h>
69 #include <sys/pctrie.h>
70 #include <sys/proc.h>
71 #include <sys/racct.h>
72 #include <sys/refcount.h>
73 #include <sys/resourcevar.h>
74 #include <sys/rwlock.h>
75 #include <sys/sched.h>
76 #include <sys/smp.h>
77 #include <sys/sysctl.h>
78 #include <sys/syscallsubr.h>
79 #include <sys/vmem.h>
80 #include <sys/vmmeter.h>
81 #include <sys/vnode.h>
82 #include <sys/watchdog.h>
83 #include <geom/geom.h>
84 #include <vm/vm.h>
85 #include <vm/vm_param.h>
86 #include <vm/vm_kern.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vm_extern.h>
92 #include <vm/vm_map.h>
93 #include <vm/swap_pager.h>
94 
95 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
96 
97 struct	bio_ops bioops;		/* I/O operation notification */
98 
99 struct	buf_ops buf_ops_bio = {
100 	.bop_name	=	"buf_ops_bio",
101 	.bop_write	=	bufwrite,
102 	.bop_strategy	=	bufstrategy,
103 	.bop_sync	=	bufsync,
104 	.bop_bdflush	=	bufbdflush,
105 };
106 
107 struct bufqueue {
108 	struct mtx_padalign	bq_lock;
109 	TAILQ_HEAD(, buf)	bq_queue;
110 	uint8_t			bq_index;
111 	uint16_t		bq_subqueue;
112 	int			bq_len;
113 } __aligned(CACHE_LINE_SIZE);
114 
115 #define	BQ_LOCKPTR(bq)		(&(bq)->bq_lock)
116 #define	BQ_LOCK(bq)		mtx_lock(BQ_LOCKPTR((bq)))
117 #define	BQ_UNLOCK(bq)		mtx_unlock(BQ_LOCKPTR((bq)))
118 #define	BQ_ASSERT_LOCKED(bq)	mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
119 
120 struct bufdomain {
121 	struct bufqueue	*bd_subq;
122 	struct bufqueue bd_dirtyq;
123 	struct bufqueue	*bd_cleanq;
124 	struct mtx_padalign bd_run_lock;
125 	/* Constants */
126 	long		bd_maxbufspace;
127 	long		bd_hibufspace;
128 	long 		bd_lobufspace;
129 	long 		bd_bufspacethresh;
130 	int		bd_hifreebuffers;
131 	int		bd_lofreebuffers;
132 	int		bd_hidirtybuffers;
133 	int		bd_lodirtybuffers;
134 	int		bd_dirtybufthresh;
135 	int		bd_lim;
136 	/* atomics */
137 	int		bd_wanted;
138 	bool		bd_shutdown;
139 	int __aligned(CACHE_LINE_SIZE)	bd_numdirtybuffers;
140 	int __aligned(CACHE_LINE_SIZE)	bd_running;
141 	long __aligned(CACHE_LINE_SIZE) bd_bufspace;
142 	int __aligned(CACHE_LINE_SIZE)	bd_freebuffers;
143 } __aligned(CACHE_LINE_SIZE);
144 
145 #define	BD_LOCKPTR(bd)		(&(bd)->bd_cleanq->bq_lock)
146 #define	BD_LOCK(bd)		mtx_lock(BD_LOCKPTR((bd)))
147 #define	BD_UNLOCK(bd)		mtx_unlock(BD_LOCKPTR((bd)))
148 #define	BD_ASSERT_LOCKED(bd)	mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
149 #define	BD_RUN_LOCKPTR(bd)	(&(bd)->bd_run_lock)
150 #define	BD_RUN_LOCK(bd)		mtx_lock(BD_RUN_LOCKPTR((bd)))
151 #define	BD_RUN_UNLOCK(bd)	mtx_unlock(BD_RUN_LOCKPTR((bd)))
152 #define	BD_DOMAIN(bd)		(bd - bdomain)
153 
154 static char *buf;		/* buffer header pool */
155 static struct buf *
156 nbufp(unsigned i)
157 {
158 	return ((struct buf *)(buf + (sizeof(struct buf) +
159 	    sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
160 }
161 
162 caddr_t __read_mostly unmapped_buf;
163 
164 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
165 struct proc *bufdaemonproc;
166 
167 static void vm_hold_free_pages(struct buf *bp, int newbsize);
168 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
169 		vm_offset_t to);
170 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
171 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
172 		vm_page_t m);
173 static void vfs_clean_pages_dirty_buf(struct buf *bp);
174 static void vfs_setdirty_range(struct buf *bp);
175 static void vfs_vmio_invalidate(struct buf *bp);
176 static void vfs_vmio_truncate(struct buf *bp, int npages);
177 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
178 static int vfs_bio_clcheck(struct vnode *vp, int size,
179 		daddr_t lblkno, daddr_t blkno);
180 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
181 		void (*)(struct buf *));
182 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
183 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
184 static void buf_daemon(void);
185 static __inline void bd_wakeup(void);
186 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
187 static void bufkva_reclaim(vmem_t *, int);
188 static void bufkva_free(struct buf *);
189 static int buf_import(void *, void **, int, int, int);
190 static void buf_release(void *, void **, int);
191 static void maxbcachebuf_adjust(void);
192 static inline struct bufdomain *bufdomain(struct buf *);
193 static void bq_remove(struct bufqueue *bq, struct buf *bp);
194 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
195 static int buf_recycle(struct bufdomain *, bool kva);
196 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
197 	    const char *lockname);
198 static void bd_init(struct bufdomain *bd);
199 static int bd_flushall(struct bufdomain *bd);
200 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
201 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
202 
203 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
204 int vmiodirenable = TRUE;
205 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
206     "Use the VM system for directory writes");
207 long runningbufspace;
208 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
209     "Amount of presently outstanding async buffer io");
210 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
211     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
212 static counter_u64_t bufkvaspace;
213 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
214     "Kernel virtual memory used for buffers");
215 static long maxbufspace;
216 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
217     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
218     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
219     "Maximum allowed value of bufspace (including metadata)");
220 static long bufmallocspace;
221 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
222     "Amount of malloced memory for buffers");
223 static long maxbufmallocspace;
224 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
225     0, "Maximum amount of malloced memory for buffers");
226 static long lobufspace;
227 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
228     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
229     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
230     "Minimum amount of buffers we want to have");
231 long hibufspace;
232 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
233     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
234     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
235     "Maximum allowed value of bufspace (excluding metadata)");
236 long bufspacethresh;
237 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
238     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
239     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
240     "Bufspace consumed before waking the daemon to free some");
241 static counter_u64_t buffreekvacnt;
242 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
243     "Number of times we have freed the KVA space from some buffer");
244 static counter_u64_t bufdefragcnt;
245 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
246     "Number of times we have had to repeat buffer allocation to defragment");
247 static long lorunningspace;
248 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
249     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
250     "Minimum preferred space used for in-progress I/O");
251 static long hirunningspace;
252 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
253     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
254     "Maximum amount of space to use for in-progress I/O");
255 int dirtybufferflushes;
256 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
257     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
258 int bdwriteskip;
259 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
260     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
261 int altbufferflushes;
262 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
263     &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
264 static int recursiveflushes;
265 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
266     &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
267 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
268 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
269     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
270     "Number of buffers that are dirty (has unwritten changes) at the moment");
271 static int lodirtybuffers;
272 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
273     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
274     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
275     "How many buffers we want to have free before bufdaemon can sleep");
276 static int hidirtybuffers;
277 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
278     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
279     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
280     "When the number of dirty buffers is considered severe");
281 int dirtybufthresh;
282 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
283     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
284     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
285     "Number of bdwrite to bawrite conversions to clear dirty buffers");
286 static int numfreebuffers;
287 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
288     "Number of free buffers");
289 static int lofreebuffers;
290 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
291     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
292     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
293    "Target number of free buffers");
294 static int hifreebuffers;
295 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
296     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
297     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
298    "Threshold for clean buffer recycling");
299 static counter_u64_t getnewbufcalls;
300 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
301    &getnewbufcalls, "Number of calls to getnewbuf");
302 static counter_u64_t getnewbufrestarts;
303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
304     &getnewbufrestarts,
305     "Number of times getnewbuf has had to restart a buffer acquisition");
306 static counter_u64_t mappingrestarts;
307 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
308     &mappingrestarts,
309     "Number of times getblk has had to restart a buffer mapping for "
310     "unmapped buffer");
311 static counter_u64_t numbufallocfails;
312 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
313     &numbufallocfails, "Number of times buffer allocations failed");
314 static int flushbufqtarget = 100;
315 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
316     "Amount of work to do in flushbufqueues when helping bufdaemon");
317 static counter_u64_t notbufdflushes;
318 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
319     "Number of dirty buffer flushes done by the bufdaemon helpers");
320 static long barrierwrites;
321 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
322     &barrierwrites, 0, "Number of barrier writes");
323 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
324     &unmapped_buf_allowed, 0,
325     "Permit the use of the unmapped i/o");
326 int maxbcachebuf = MAXBCACHEBUF;
327 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
328     "Maximum size of a buffer cache block");
329 
330 /*
331  * This lock synchronizes access to bd_request.
332  */
333 static struct mtx_padalign __exclusive_cache_line bdlock;
334 
335 /*
336  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
337  * waitrunningbufspace().
338  */
339 static struct mtx_padalign __exclusive_cache_line rbreqlock;
340 
341 /*
342  * Lock that protects bdirtywait.
343  */
344 static struct mtx_padalign __exclusive_cache_line bdirtylock;
345 
346 /*
347  * bufdaemon shutdown request and sleep channel.
348  */
349 static bool bd_shutdown;
350 
351 /*
352  * Wakeup point for bufdaemon, as well as indicator of whether it is already
353  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
354  * is idling.
355  */
356 static int bd_request;
357 
358 /*
359  * Request for the buf daemon to write more buffers than is indicated by
360  * lodirtybuf.  This may be necessary to push out excess dependencies or
361  * defragment the address space where a simple count of the number of dirty
362  * buffers is insufficient to characterize the demand for flushing them.
363  */
364 static int bd_speedupreq;
365 
366 /*
367  * Synchronization (sleep/wakeup) variable for active buffer space requests.
368  * Set when wait starts, cleared prior to wakeup().
369  * Used in runningbufwakeup() and waitrunningbufspace().
370  */
371 static int runningbufreq;
372 
373 /*
374  * Synchronization for bwillwrite() waiters.
375  */
376 static int bdirtywait;
377 
378 /*
379  * Definitions for the buffer free lists.
380  */
381 #define QUEUE_NONE	0	/* on no queue */
382 #define QUEUE_EMPTY	1	/* empty buffer headers */
383 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
384 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
385 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
386 
387 /* Maximum number of buffer domains. */
388 #define	BUF_DOMAINS	8
389 
390 struct bufdomainset bdlodirty;		/* Domains > lodirty */
391 struct bufdomainset bdhidirty;		/* Domains > hidirty */
392 
393 /* Configured number of clean queues. */
394 static int __read_mostly buf_domains;
395 
396 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
397 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
398 struct bufqueue __exclusive_cache_line bqempty;
399 
400 /*
401  * per-cpu empty buffer cache.
402  */
403 uma_zone_t buf_zone;
404 
405 static int
406 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
407 {
408 	long value;
409 	int error;
410 
411 	value = *(long *)arg1;
412 	error = sysctl_handle_long(oidp, &value, 0, req);
413 	if (error != 0 || req->newptr == NULL)
414 		return (error);
415 	mtx_lock(&rbreqlock);
416 	if (arg1 == &hirunningspace) {
417 		if (value < lorunningspace)
418 			error = EINVAL;
419 		else
420 			hirunningspace = value;
421 	} else {
422 		KASSERT(arg1 == &lorunningspace,
423 		    ("%s: unknown arg1", __func__));
424 		if (value > hirunningspace)
425 			error = EINVAL;
426 		else
427 			lorunningspace = value;
428 	}
429 	mtx_unlock(&rbreqlock);
430 	return (error);
431 }
432 
433 static int
434 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
435 {
436 	int error;
437 	int value;
438 	int i;
439 
440 	value = *(int *)arg1;
441 	error = sysctl_handle_int(oidp, &value, 0, req);
442 	if (error != 0 || req->newptr == NULL)
443 		return (error);
444 	*(int *)arg1 = value;
445 	for (i = 0; i < buf_domains; i++)
446 		*(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
447 		    value / buf_domains;
448 
449 	return (error);
450 }
451 
452 static int
453 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
454 {
455 	long value;
456 	int error;
457 	int i;
458 
459 	value = *(long *)arg1;
460 	error = sysctl_handle_long(oidp, &value, 0, req);
461 	if (error != 0 || req->newptr == NULL)
462 		return (error);
463 	*(long *)arg1 = value;
464 	for (i = 0; i < buf_domains; i++)
465 		*(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
466 		    value / buf_domains;
467 
468 	return (error);
469 }
470 
471 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
472     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
473 static int
474 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
475 {
476 	long lvalue;
477 	int ivalue;
478 	int i;
479 
480 	lvalue = 0;
481 	for (i = 0; i < buf_domains; i++)
482 		lvalue += bdomain[i].bd_bufspace;
483 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
484 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
485 	if (lvalue > INT_MAX)
486 		/* On overflow, still write out a long to trigger ENOMEM. */
487 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
488 	ivalue = lvalue;
489 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
490 }
491 #else
492 static int
493 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
494 {
495 	long lvalue;
496 	int i;
497 
498 	lvalue = 0;
499 	for (i = 0; i < buf_domains; i++)
500 		lvalue += bdomain[i].bd_bufspace;
501 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
502 }
503 #endif
504 
505 static int
506 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
507 {
508 	int value;
509 	int i;
510 
511 	value = 0;
512 	for (i = 0; i < buf_domains; i++)
513 		value += bdomain[i].bd_numdirtybuffers;
514 	return (sysctl_handle_int(oidp, &value, 0, req));
515 }
516 
517 /*
518  *	bdirtywakeup:
519  *
520  *	Wakeup any bwillwrite() waiters.
521  */
522 static void
523 bdirtywakeup(void)
524 {
525 	mtx_lock(&bdirtylock);
526 	if (bdirtywait) {
527 		bdirtywait = 0;
528 		wakeup(&bdirtywait);
529 	}
530 	mtx_unlock(&bdirtylock);
531 }
532 
533 /*
534  *	bd_clear:
535  *
536  *	Clear a domain from the appropriate bitsets when dirtybuffers
537  *	is decremented.
538  */
539 static void
540 bd_clear(struct bufdomain *bd)
541 {
542 
543 	mtx_lock(&bdirtylock);
544 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
545 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
546 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
547 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
548 	mtx_unlock(&bdirtylock);
549 }
550 
551 /*
552  *	bd_set:
553  *
554  *	Set a domain in the appropriate bitsets when dirtybuffers
555  *	is incremented.
556  */
557 static void
558 bd_set(struct bufdomain *bd)
559 {
560 
561 	mtx_lock(&bdirtylock);
562 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
563 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
564 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
565 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
566 	mtx_unlock(&bdirtylock);
567 }
568 
569 /*
570  *	bdirtysub:
571  *
572  *	Decrement the numdirtybuffers count by one and wakeup any
573  *	threads blocked in bwillwrite().
574  */
575 static void
576 bdirtysub(struct buf *bp)
577 {
578 	struct bufdomain *bd;
579 	int num;
580 
581 	bd = bufdomain(bp);
582 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
583 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
584 		bdirtywakeup();
585 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
586 		bd_clear(bd);
587 }
588 
589 /*
590  *	bdirtyadd:
591  *
592  *	Increment the numdirtybuffers count by one and wakeup the buf
593  *	daemon if needed.
594  */
595 static void
596 bdirtyadd(struct buf *bp)
597 {
598 	struct bufdomain *bd;
599 	int num;
600 
601 	/*
602 	 * Only do the wakeup once as we cross the boundary.  The
603 	 * buf daemon will keep running until the condition clears.
604 	 */
605 	bd = bufdomain(bp);
606 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
607 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
608 		bd_wakeup();
609 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
610 		bd_set(bd);
611 }
612 
613 /*
614  *	bufspace_daemon_wakeup:
615  *
616  *	Wakeup the daemons responsible for freeing clean bufs.
617  */
618 static void
619 bufspace_daemon_wakeup(struct bufdomain *bd)
620 {
621 
622 	/*
623 	 * avoid the lock if the daemon is running.
624 	 */
625 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
626 		BD_RUN_LOCK(bd);
627 		atomic_store_int(&bd->bd_running, 1);
628 		wakeup(&bd->bd_running);
629 		BD_RUN_UNLOCK(bd);
630 	}
631 }
632 
633 /*
634  *	bufspace_adjust:
635  *
636  *	Adjust the reported bufspace for a KVA managed buffer, possibly
637  * 	waking any waiters.
638  */
639 static void
640 bufspace_adjust(struct buf *bp, int bufsize)
641 {
642 	struct bufdomain *bd;
643 	long space;
644 	int diff;
645 
646 	KASSERT((bp->b_flags & B_MALLOC) == 0,
647 	    ("bufspace_adjust: malloc buf %p", bp));
648 	bd = bufdomain(bp);
649 	diff = bufsize - bp->b_bufsize;
650 	if (diff < 0) {
651 		atomic_subtract_long(&bd->bd_bufspace, -diff);
652 	} else if (diff > 0) {
653 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
654 		/* Wake up the daemon on the transition. */
655 		if (space < bd->bd_bufspacethresh &&
656 		    space + diff >= bd->bd_bufspacethresh)
657 			bufspace_daemon_wakeup(bd);
658 	}
659 	bp->b_bufsize = bufsize;
660 }
661 
662 /*
663  *	bufspace_reserve:
664  *
665  *	Reserve bufspace before calling allocbuf().  metadata has a
666  *	different space limit than data.
667  */
668 static int
669 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
670 {
671 	long limit, new;
672 	long space;
673 
674 	if (metadata)
675 		limit = bd->bd_maxbufspace;
676 	else
677 		limit = bd->bd_hibufspace;
678 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
679 	new = space + size;
680 	if (new > limit) {
681 		atomic_subtract_long(&bd->bd_bufspace, size);
682 		return (ENOSPC);
683 	}
684 
685 	/* Wake up the daemon on the transition. */
686 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
687 		bufspace_daemon_wakeup(bd);
688 
689 	return (0);
690 }
691 
692 /*
693  *	bufspace_release:
694  *
695  *	Release reserved bufspace after bufspace_adjust() has consumed it.
696  */
697 static void
698 bufspace_release(struct bufdomain *bd, int size)
699 {
700 
701 	atomic_subtract_long(&bd->bd_bufspace, size);
702 }
703 
704 /*
705  *	bufspace_wait:
706  *
707  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
708  *	supplied.  bd_wanted must be set prior to polling for space.  The
709  *	operation must be re-tried on return.
710  */
711 static void
712 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
713     int slpflag, int slptimeo)
714 {
715 	struct thread *td;
716 	int error, fl, norunbuf;
717 
718 	if ((gbflags & GB_NOWAIT_BD) != 0)
719 		return;
720 
721 	td = curthread;
722 	BD_LOCK(bd);
723 	while (bd->bd_wanted) {
724 		if (vp != NULL && vp->v_type != VCHR &&
725 		    (td->td_pflags & TDP_BUFNEED) == 0) {
726 			BD_UNLOCK(bd);
727 			/*
728 			 * getblk() is called with a vnode locked, and
729 			 * some majority of the dirty buffers may as
730 			 * well belong to the vnode.  Flushing the
731 			 * buffers there would make a progress that
732 			 * cannot be achieved by the buf_daemon, that
733 			 * cannot lock the vnode.
734 			 */
735 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
736 			    (td->td_pflags & TDP_NORUNNINGBUF);
737 
738 			/*
739 			 * Play bufdaemon.  The getnewbuf() function
740 			 * may be called while the thread owns lock
741 			 * for another dirty buffer for the same
742 			 * vnode, which makes it impossible to use
743 			 * VOP_FSYNC() there, due to the buffer lock
744 			 * recursion.
745 			 */
746 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
747 			fl = buf_flush(vp, bd, flushbufqtarget);
748 			td->td_pflags &= norunbuf;
749 			BD_LOCK(bd);
750 			if (fl != 0)
751 				continue;
752 			if (bd->bd_wanted == 0)
753 				break;
754 		}
755 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
756 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
757 		if (error != 0)
758 			break;
759 	}
760 	BD_UNLOCK(bd);
761 }
762 
763 static void
764 bufspace_daemon_shutdown(void *arg, int howto __unused)
765 {
766 	struct bufdomain *bd = arg;
767 	int error;
768 
769 	if (KERNEL_PANICKED())
770 		return;
771 
772 	BD_RUN_LOCK(bd);
773 	bd->bd_shutdown = true;
774 	wakeup(&bd->bd_running);
775 	error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0,
776 	    "bufspace_shutdown", 60 * hz);
777 	BD_RUN_UNLOCK(bd);
778 	if (error != 0)
779 		printf("bufspacedaemon wait error: %d\n", error);
780 }
781 
782 /*
783  *	bufspace_daemon:
784  *
785  *	buffer space management daemon.  Tries to maintain some marginal
786  *	amount of free buffer space so that requesting processes neither
787  *	block nor work to reclaim buffers.
788  */
789 static void
790 bufspace_daemon(void *arg)
791 {
792 	struct bufdomain *bd = arg;
793 
794 	EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd,
795 	    SHUTDOWN_PRI_LAST + 100);
796 
797 	BD_RUN_LOCK(bd);
798 	while (!bd->bd_shutdown) {
799 		BD_RUN_UNLOCK(bd);
800 
801 		/*
802 		 * Free buffers from the clean queue until we meet our
803 		 * targets.
804 		 *
805 		 * Theory of operation:  The buffer cache is most efficient
806 		 * when some free buffer headers and space are always
807 		 * available to getnewbuf().  This daemon attempts to prevent
808 		 * the excessive blocking and synchronization associated
809 		 * with shortfall.  It goes through three phases according
810 		 * demand:
811 		 *
812 		 * 1)	The daemon wakes up voluntarily once per-second
813 		 *	during idle periods when the counters are below
814 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
815 		 *
816 		 * 2)	The daemon wakes up as we cross the thresholds
817 		 *	ahead of any potential blocking.  This may bounce
818 		 *	slightly according to the rate of consumption and
819 		 *	release.
820 		 *
821 		 * 3)	The daemon and consumers are starved for working
822 		 *	clean buffers.  This is the 'bufspace' sleep below
823 		 *	which will inefficiently trade bufs with bqrelse
824 		 *	until we return to condition 2.
825 		 */
826 		while (bd->bd_bufspace > bd->bd_lobufspace ||
827 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
828 			if (buf_recycle(bd, false) != 0) {
829 				if (bd_flushall(bd))
830 					continue;
831 				/*
832 				 * Speedup dirty if we've run out of clean
833 				 * buffers.  This is possible in particular
834 				 * because softdep may held many bufs locked
835 				 * pending writes to other bufs which are
836 				 * marked for delayed write, exhausting
837 				 * clean space until they are written.
838 				 */
839 				bd_speedup();
840 				BD_LOCK(bd);
841 				if (bd->bd_wanted) {
842 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
843 					    PRIBIO|PDROP, "bufspace", hz/10);
844 				} else
845 					BD_UNLOCK(bd);
846 			}
847 			maybe_yield();
848 		}
849 
850 		/*
851 		 * Re-check our limits and sleep.  bd_running must be
852 		 * cleared prior to checking the limits to avoid missed
853 		 * wakeups.  The waker will adjust one of bufspace or
854 		 * freebuffers prior to checking bd_running.
855 		 */
856 		BD_RUN_LOCK(bd);
857 		if (bd->bd_shutdown)
858 			break;
859 		atomic_store_int(&bd->bd_running, 0);
860 		if (bd->bd_bufspace < bd->bd_bufspacethresh &&
861 		    bd->bd_freebuffers > bd->bd_lofreebuffers) {
862 			msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd),
863 			    PRIBIO, "-", hz);
864 		} else {
865 			/* Avoid spurious wakeups while running. */
866 			atomic_store_int(&bd->bd_running, 1);
867 		}
868 	}
869 	wakeup(&bd->bd_shutdown);
870 	BD_RUN_UNLOCK(bd);
871 	kthread_exit();
872 }
873 
874 /*
875  *	bufmallocadjust:
876  *
877  *	Adjust the reported bufspace for a malloc managed buffer, possibly
878  *	waking any waiters.
879  */
880 static void
881 bufmallocadjust(struct buf *bp, int bufsize)
882 {
883 	int diff;
884 
885 	KASSERT((bp->b_flags & B_MALLOC) != 0,
886 	    ("bufmallocadjust: non-malloc buf %p", bp));
887 	diff = bufsize - bp->b_bufsize;
888 	if (diff < 0)
889 		atomic_subtract_long(&bufmallocspace, -diff);
890 	else
891 		atomic_add_long(&bufmallocspace, diff);
892 	bp->b_bufsize = bufsize;
893 }
894 
895 /*
896  *	runningwakeup:
897  *
898  *	Wake up processes that are waiting on asynchronous writes to fall
899  *	below lorunningspace.
900  */
901 static void
902 runningwakeup(void)
903 {
904 
905 	mtx_lock(&rbreqlock);
906 	if (runningbufreq) {
907 		runningbufreq = 0;
908 		wakeup(&runningbufreq);
909 	}
910 	mtx_unlock(&rbreqlock);
911 }
912 
913 /*
914  *	runningbufwakeup:
915  *
916  *	Decrement the outstanding write count according.
917  */
918 void
919 runningbufwakeup(struct buf *bp)
920 {
921 	long space, bspace;
922 
923 	bspace = bp->b_runningbufspace;
924 	if (bspace == 0)
925 		return;
926 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
927 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
928 	    space, bspace));
929 	bp->b_runningbufspace = 0;
930 	/*
931 	 * Only acquire the lock and wakeup on the transition from exceeding
932 	 * the threshold to falling below it.
933 	 */
934 	if (space < lorunningspace)
935 		return;
936 	if (space - bspace > lorunningspace)
937 		return;
938 	runningwakeup();
939 }
940 
941 /*
942  *	waitrunningbufspace()
943  *
944  *	runningbufspace is a measure of the amount of I/O currently
945  *	running.  This routine is used in async-write situations to
946  *	prevent creating huge backups of pending writes to a device.
947  *	Only asynchronous writes are governed by this function.
948  *
949  *	This does NOT turn an async write into a sync write.  It waits
950  *	for earlier writes to complete and generally returns before the
951  *	caller's write has reached the device.
952  */
953 void
954 waitrunningbufspace(void)
955 {
956 
957 	mtx_lock(&rbreqlock);
958 	while (runningbufspace > hirunningspace) {
959 		runningbufreq = 1;
960 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
961 	}
962 	mtx_unlock(&rbreqlock);
963 }
964 
965 /*
966  *	vfs_buf_test_cache:
967  *
968  *	Called when a buffer is extended.  This function clears the B_CACHE
969  *	bit if the newly extended portion of the buffer does not contain
970  *	valid data.
971  */
972 static __inline void
973 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
974     vm_offset_t size, vm_page_t m)
975 {
976 
977 	/*
978 	 * This function and its results are protected by higher level
979 	 * synchronization requiring vnode and buf locks to page in and
980 	 * validate pages.
981 	 */
982 	if (bp->b_flags & B_CACHE) {
983 		int base = (foff + off) & PAGE_MASK;
984 		if (vm_page_is_valid(m, base, size) == 0)
985 			bp->b_flags &= ~B_CACHE;
986 	}
987 }
988 
989 /* Wake up the buffer daemon if necessary */
990 static void
991 bd_wakeup(void)
992 {
993 
994 	mtx_lock(&bdlock);
995 	if (bd_request == 0) {
996 		bd_request = 1;
997 		wakeup(&bd_request);
998 	}
999 	mtx_unlock(&bdlock);
1000 }
1001 
1002 /*
1003  * Adjust the maxbcachbuf tunable.
1004  */
1005 static void
1006 maxbcachebuf_adjust(void)
1007 {
1008 	int i;
1009 
1010 	/*
1011 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
1012 	 */
1013 	i = 2;
1014 	while (i * 2 <= maxbcachebuf)
1015 		i *= 2;
1016 	maxbcachebuf = i;
1017 	if (maxbcachebuf < MAXBSIZE)
1018 		maxbcachebuf = MAXBSIZE;
1019 	if (maxbcachebuf > maxphys)
1020 		maxbcachebuf = maxphys;
1021 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1022 		printf("maxbcachebuf=%d\n", maxbcachebuf);
1023 }
1024 
1025 /*
1026  * bd_speedup - speedup the buffer cache flushing code
1027  */
1028 void
1029 bd_speedup(void)
1030 {
1031 	int needwake;
1032 
1033 	mtx_lock(&bdlock);
1034 	needwake = 0;
1035 	if (bd_speedupreq == 0 || bd_request == 0)
1036 		needwake = 1;
1037 	bd_speedupreq = 1;
1038 	bd_request = 1;
1039 	if (needwake)
1040 		wakeup(&bd_request);
1041 	mtx_unlock(&bdlock);
1042 }
1043 
1044 #ifdef __i386__
1045 #define	TRANSIENT_DENOM	5
1046 #else
1047 #define	TRANSIENT_DENOM 10
1048 #endif
1049 
1050 /*
1051  * Calculating buffer cache scaling values and reserve space for buffer
1052  * headers.  This is called during low level kernel initialization and
1053  * may be called more then once.  We CANNOT write to the memory area
1054  * being reserved at this time.
1055  */
1056 caddr_t
1057 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1058 {
1059 	int tuned_nbuf;
1060 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1061 
1062 	/*
1063 	 * With KASAN or KMSAN enabled, the kernel map is shadowed.  Account for
1064 	 * this when sizing maps based on the amount of physical memory
1065 	 * available.
1066 	 */
1067 #if defined(KASAN)
1068 	physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1069 	    (KASAN_SHADOW_SCALE + 1);
1070 #elif defined(KMSAN)
1071 	physmem_est /= 3;
1072 
1073 	/*
1074 	 * KMSAN cannot reliably determine whether buffer data is initialized
1075 	 * unless it is updated through a KVA mapping.
1076 	 */
1077 	unmapped_buf_allowed = 0;
1078 #endif
1079 
1080 	/*
1081 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1082 	 * PAGE_SIZE is >= 1K)
1083 	 */
1084 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1085 
1086 	maxbcachebuf_adjust();
1087 	/*
1088 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1089 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1090 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1091 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1092 	 * the buffer cache we limit the eventual kva reservation to
1093 	 * maxbcache bytes.
1094 	 *
1095 	 * factor represents the 1/4 x ram conversion.
1096 	 */
1097 	if (nbuf == 0) {
1098 		int factor = 4 * BKVASIZE / 1024;
1099 
1100 		nbuf = 50;
1101 		if (physmem_est > 4096)
1102 			nbuf += min((physmem_est - 4096) / factor,
1103 			    65536 / factor);
1104 		if (physmem_est > 65536)
1105 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1106 			    32 * 1024 * 1024 / (factor * 5));
1107 
1108 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1109 			nbuf = maxbcache / BKVASIZE;
1110 		tuned_nbuf = 1;
1111 	} else
1112 		tuned_nbuf = 0;
1113 
1114 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1115 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1116 	if (nbuf > maxbuf) {
1117 		if (!tuned_nbuf)
1118 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1119 			    maxbuf);
1120 		nbuf = maxbuf;
1121 	}
1122 
1123 	/*
1124 	 * Ideal allocation size for the transient bio submap is 10%
1125 	 * of the maximal space buffer map.  This roughly corresponds
1126 	 * to the amount of the buffer mapped for typical UFS load.
1127 	 *
1128 	 * Clip the buffer map to reserve space for the transient
1129 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1130 	 * maximum buffer map extent on the platform.
1131 	 *
1132 	 * The fall-back to the maxbuf in case of maxbcache unset,
1133 	 * allows to not trim the buffer KVA for the architectures
1134 	 * with ample KVA space.
1135 	 */
1136 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1137 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1138 		buf_sz = (long)nbuf * BKVASIZE;
1139 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1140 		    (TRANSIENT_DENOM - 1)) {
1141 			/*
1142 			 * There is more KVA than memory.  Do not
1143 			 * adjust buffer map size, and assign the rest
1144 			 * of maxbuf to transient map.
1145 			 */
1146 			biotmap_sz = maxbuf_sz - buf_sz;
1147 		} else {
1148 			/*
1149 			 * Buffer map spans all KVA we could afford on
1150 			 * this platform.  Give 10% (20% on i386) of
1151 			 * the buffer map to the transient bio map.
1152 			 */
1153 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1154 			buf_sz -= biotmap_sz;
1155 		}
1156 		if (biotmap_sz / INT_MAX > maxphys)
1157 			bio_transient_maxcnt = INT_MAX;
1158 		else
1159 			bio_transient_maxcnt = biotmap_sz / maxphys;
1160 		/*
1161 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1162 		 * using the transient mapping.
1163 		 */
1164 		if (bio_transient_maxcnt > 1024)
1165 			bio_transient_maxcnt = 1024;
1166 		if (tuned_nbuf)
1167 			nbuf = buf_sz / BKVASIZE;
1168 	}
1169 
1170 	if (nswbuf == 0) {
1171 		/*
1172 		 * Pager buffers are allocated for short periods, so scale the
1173 		 * number of reserved buffers based on the number of CPUs rather
1174 		 * than amount of memory.
1175 		 */
1176 		nswbuf = min(nbuf / 4, 32 * mp_ncpus);
1177 		if (nswbuf < NSWBUF_MIN)
1178 			nswbuf = NSWBUF_MIN;
1179 	}
1180 
1181 	/*
1182 	 * Reserve space for the buffer cache buffers
1183 	 */
1184 	buf = (char *)v;
1185 	v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1186 	    atop(maxbcachebuf)) * nbuf;
1187 
1188 	return (v);
1189 }
1190 
1191 /*
1192  * Single global constant for BUF_WMESG, to avoid getting multiple
1193  * references.
1194  */
1195 static const char buf_wmesg[] = "bufwait";
1196 
1197 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1198 void
1199 bufinit(void)
1200 {
1201 	struct buf *bp;
1202 	int i;
1203 
1204 	TSENTER();
1205 	KASSERT(maxbcachebuf >= MAXBSIZE,
1206 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1207 	    MAXBSIZE));
1208 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1209 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1210 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1211 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1212 
1213 	unmapped_buf = (caddr_t)kva_alloc(maxphys);
1214 
1215 	/* finally, initialize each buffer header and stick on empty q */
1216 	for (i = 0; i < nbuf; i++) {
1217 		bp = nbufp(i);
1218 		bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1219 		bp->b_flags = B_INVAL;
1220 		bp->b_rcred = NOCRED;
1221 		bp->b_wcred = NOCRED;
1222 		bp->b_qindex = QUEUE_NONE;
1223 		bp->b_domain = -1;
1224 		bp->b_subqueue = mp_maxid + 1;
1225 		bp->b_xflags = 0;
1226 		bp->b_data = bp->b_kvabase = unmapped_buf;
1227 		LIST_INIT(&bp->b_dep);
1228 		BUF_LOCKINIT(bp, buf_wmesg);
1229 		bq_insert(&bqempty, bp, false);
1230 	}
1231 
1232 	/*
1233 	 * maxbufspace is the absolute maximum amount of buffer space we are
1234 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1235 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1236 	 * used by most other requests.  The differential is required to
1237 	 * ensure that metadata deadlocks don't occur.
1238 	 *
1239 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1240 	 * this may result in KVM fragmentation which is not handled optimally
1241 	 * by the system. XXX This is less true with vmem.  We could use
1242 	 * PAGE_SIZE.
1243 	 */
1244 	maxbufspace = (long)nbuf * BKVASIZE;
1245 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1246 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1247 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1248 
1249 	/*
1250 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1251 	 * arbitrarily and may need further tuning. It corresponds to
1252 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1253 	 * which fits with many RAID controllers' tagged queuing limits.
1254 	 * The lower 1 MiB limit is the historical upper limit for
1255 	 * hirunningspace.
1256 	 */
1257 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1258 	    16 * 1024 * 1024), 1024 * 1024);
1259 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1260 
1261 	/*
1262 	 * Limit the amount of malloc memory since it is wired permanently into
1263 	 * the kernel space.  Even though this is accounted for in the buffer
1264 	 * allocation, we don't want the malloced region to grow uncontrolled.
1265 	 * The malloc scheme improves memory utilization significantly on
1266 	 * average (small) directories.
1267 	 */
1268 	maxbufmallocspace = hibufspace / 20;
1269 
1270 	/*
1271 	 * Reduce the chance of a deadlock occurring by limiting the number
1272 	 * of delayed-write dirty buffers we allow to stack up.
1273 	 */
1274 	hidirtybuffers = nbuf / 4 + 20;
1275 	dirtybufthresh = hidirtybuffers * 9 / 10;
1276 	/*
1277 	 * To support extreme low-memory systems, make sure hidirtybuffers
1278 	 * cannot eat up all available buffer space.  This occurs when our
1279 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1280 	 * buffer space assuming BKVASIZE'd buffers.
1281 	 */
1282 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1283 		hidirtybuffers >>= 1;
1284 	}
1285 	lodirtybuffers = hidirtybuffers / 2;
1286 
1287 	/*
1288 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1289 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1290 	 * are counted as free but will be unavailable to threads executing
1291 	 * on other cpus.
1292 	 *
1293 	 * hifreebuffers is the free target for the bufspace daemon.  This
1294 	 * should be set appropriately to limit work per-iteration.
1295 	 */
1296 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1297 	hifreebuffers = (3 * lofreebuffers) / 2;
1298 	numfreebuffers = nbuf;
1299 
1300 	/* Setup the kva and free list allocators. */
1301 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1302 	buf_zone = uma_zcache_create("buf free cache",
1303 	    sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1304 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1305 
1306 	/*
1307 	 * Size the clean queue according to the amount of buffer space.
1308 	 * One queue per-256mb up to the max.  More queues gives better
1309 	 * concurrency but less accurate LRU.
1310 	 */
1311 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1312 	for (i = 0 ; i < buf_domains; i++) {
1313 		struct bufdomain *bd;
1314 
1315 		bd = &bdomain[i];
1316 		bd_init(bd);
1317 		bd->bd_freebuffers = nbuf / buf_domains;
1318 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1319 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1320 		bd->bd_bufspace = 0;
1321 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1322 		bd->bd_hibufspace = hibufspace / buf_domains;
1323 		bd->bd_lobufspace = lobufspace / buf_domains;
1324 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1325 		bd->bd_numdirtybuffers = 0;
1326 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1327 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1328 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1329 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1330 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1331 	}
1332 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1333 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1334 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1335 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1336 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1337 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1338 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1339 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1340 	TSEXIT();
1341 }
1342 
1343 #ifdef INVARIANTS
1344 static inline void
1345 vfs_buf_check_mapped(struct buf *bp)
1346 {
1347 
1348 	KASSERT(bp->b_kvabase != unmapped_buf,
1349 	    ("mapped buf: b_kvabase was not updated %p", bp));
1350 	KASSERT(bp->b_data != unmapped_buf,
1351 	    ("mapped buf: b_data was not updated %p", bp));
1352 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1353 	    maxphys, ("b_data + b_offset unmapped %p", bp));
1354 }
1355 
1356 static inline void
1357 vfs_buf_check_unmapped(struct buf *bp)
1358 {
1359 
1360 	KASSERT(bp->b_data == unmapped_buf,
1361 	    ("unmapped buf: corrupted b_data %p", bp));
1362 }
1363 
1364 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1365 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1366 #else
1367 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1368 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1369 #endif
1370 
1371 static int
1372 isbufbusy(struct buf *bp)
1373 {
1374 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1375 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1376 		return (1);
1377 	return (0);
1378 }
1379 
1380 /*
1381  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1382  */
1383 void
1384 bufshutdown(int show_busybufs)
1385 {
1386 	static int first_buf_printf = 1;
1387 	struct buf *bp;
1388 	int i, iter, nbusy, pbusy;
1389 #ifndef PREEMPTION
1390 	int subiter;
1391 #endif
1392 
1393 	/*
1394 	 * Sync filesystems for shutdown
1395 	 */
1396 	wdog_kern_pat(WD_LASTVAL);
1397 	kern_sync(curthread);
1398 
1399 	/*
1400 	 * With soft updates, some buffers that are
1401 	 * written will be remarked as dirty until other
1402 	 * buffers are written.
1403 	 */
1404 	for (iter = pbusy = 0; iter < 20; iter++) {
1405 		nbusy = 0;
1406 		for (i = nbuf - 1; i >= 0; i--) {
1407 			bp = nbufp(i);
1408 			if (isbufbusy(bp))
1409 				nbusy++;
1410 		}
1411 		if (nbusy == 0) {
1412 			if (first_buf_printf)
1413 				printf("All buffers synced.");
1414 			break;
1415 		}
1416 		if (first_buf_printf) {
1417 			printf("Syncing disks, buffers remaining... ");
1418 			first_buf_printf = 0;
1419 		}
1420 		printf("%d ", nbusy);
1421 		if (nbusy < pbusy)
1422 			iter = 0;
1423 		pbusy = nbusy;
1424 
1425 		wdog_kern_pat(WD_LASTVAL);
1426 		kern_sync(curthread);
1427 
1428 #ifdef PREEMPTION
1429 		/*
1430 		 * Spin for a while to allow interrupt threads to run.
1431 		 */
1432 		DELAY(50000 * iter);
1433 #else
1434 		/*
1435 		 * Context switch several times to allow interrupt
1436 		 * threads to run.
1437 		 */
1438 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1439 			sched_relinquish(curthread);
1440 			DELAY(1000);
1441 		}
1442 #endif
1443 	}
1444 	printf("\n");
1445 	/*
1446 	 * Count only busy local buffers to prevent forcing
1447 	 * a fsck if we're just a client of a wedged NFS server
1448 	 */
1449 	nbusy = 0;
1450 	for (i = nbuf - 1; i >= 0; i--) {
1451 		bp = nbufp(i);
1452 		if (isbufbusy(bp)) {
1453 #if 0
1454 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1455 			if (bp->b_dev == NULL) {
1456 				TAILQ_REMOVE(&mountlist,
1457 				    bp->b_vp->v_mount, mnt_list);
1458 				continue;
1459 			}
1460 #endif
1461 			nbusy++;
1462 			if (show_busybufs > 0) {
1463 				printf(
1464 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1465 				    nbusy, bp, bp->b_vp, bp->b_flags,
1466 				    (intmax_t)bp->b_blkno,
1467 				    (intmax_t)bp->b_lblkno);
1468 				BUF_LOCKPRINTINFO(bp);
1469 				if (show_busybufs > 1)
1470 					vn_printf(bp->b_vp,
1471 					    "vnode content: ");
1472 			}
1473 		}
1474 	}
1475 	if (nbusy) {
1476 		/*
1477 		 * Failed to sync all blocks. Indicate this and don't
1478 		 * unmount filesystems (thus forcing an fsck on reboot).
1479 		 */
1480 		BOOTTRACE("shutdown failed to sync buffers");
1481 		printf("Giving up on %d buffers\n", nbusy);
1482 		DELAY(5000000);	/* 5 seconds */
1483 		swapoff_all();
1484 	} else {
1485 		BOOTTRACE("shutdown sync complete");
1486 		if (!first_buf_printf)
1487 			printf("Final sync complete\n");
1488 
1489 		/*
1490 		 * Unmount filesystems and perform swapoff, to quiesce
1491 		 * the system as much as possible.  In particular, no
1492 		 * I/O should be initiated from top levels since it
1493 		 * might be abruptly terminated by reset, or otherwise
1494 		 * erronously handled because other parts of the
1495 		 * system are disabled.
1496 		 *
1497 		 * Swapoff before unmount, because file-backed swap is
1498 		 * non-operational after unmount of the underlying
1499 		 * filesystem.
1500 		 */
1501 		if (!KERNEL_PANICKED()) {
1502 			swapoff_all();
1503 			vfs_unmountall();
1504 		}
1505 		BOOTTRACE("shutdown unmounted all filesystems");
1506 	}
1507 	DELAY(100000);		/* wait for console output to finish */
1508 }
1509 
1510 static void
1511 bpmap_qenter(struct buf *bp)
1512 {
1513 
1514 	BUF_CHECK_MAPPED(bp);
1515 
1516 	/*
1517 	 * bp->b_data is relative to bp->b_offset, but
1518 	 * bp->b_offset may be offset into the first page.
1519 	 */
1520 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1521 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1522 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1523 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1524 }
1525 
1526 static inline struct bufdomain *
1527 bufdomain(struct buf *bp)
1528 {
1529 
1530 	return (&bdomain[bp->b_domain]);
1531 }
1532 
1533 static struct bufqueue *
1534 bufqueue(struct buf *bp)
1535 {
1536 
1537 	switch (bp->b_qindex) {
1538 	case QUEUE_NONE:
1539 		/* FALLTHROUGH */
1540 	case QUEUE_SENTINEL:
1541 		return (NULL);
1542 	case QUEUE_EMPTY:
1543 		return (&bqempty);
1544 	case QUEUE_DIRTY:
1545 		return (&bufdomain(bp)->bd_dirtyq);
1546 	case QUEUE_CLEAN:
1547 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1548 	default:
1549 		break;
1550 	}
1551 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1552 }
1553 
1554 /*
1555  * Return the locked bufqueue that bp is a member of.
1556  */
1557 static struct bufqueue *
1558 bufqueue_acquire(struct buf *bp)
1559 {
1560 	struct bufqueue *bq, *nbq;
1561 
1562 	/*
1563 	 * bp can be pushed from a per-cpu queue to the
1564 	 * cleanq while we're waiting on the lock.  Retry
1565 	 * if the queues don't match.
1566 	 */
1567 	bq = bufqueue(bp);
1568 	BQ_LOCK(bq);
1569 	for (;;) {
1570 		nbq = bufqueue(bp);
1571 		if (bq == nbq)
1572 			break;
1573 		BQ_UNLOCK(bq);
1574 		BQ_LOCK(nbq);
1575 		bq = nbq;
1576 	}
1577 	return (bq);
1578 }
1579 
1580 /*
1581  *	binsfree:
1582  *
1583  *	Insert the buffer into the appropriate free list.  Requires a
1584  *	locked buffer on entry and buffer is unlocked before return.
1585  */
1586 static void
1587 binsfree(struct buf *bp, int qindex)
1588 {
1589 	struct bufdomain *bd;
1590 	struct bufqueue *bq;
1591 
1592 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1593 	    ("binsfree: Invalid qindex %d", qindex));
1594 	BUF_ASSERT_XLOCKED(bp);
1595 
1596 	/*
1597 	 * Handle delayed bremfree() processing.
1598 	 */
1599 	if (bp->b_flags & B_REMFREE) {
1600 		if (bp->b_qindex == qindex) {
1601 			bp->b_flags |= B_REUSE;
1602 			bp->b_flags &= ~B_REMFREE;
1603 			BUF_UNLOCK(bp);
1604 			return;
1605 		}
1606 		bq = bufqueue_acquire(bp);
1607 		bq_remove(bq, bp);
1608 		BQ_UNLOCK(bq);
1609 	}
1610 	bd = bufdomain(bp);
1611 	if (qindex == QUEUE_CLEAN) {
1612 		if (bd->bd_lim != 0)
1613 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1614 		else
1615 			bq = bd->bd_cleanq;
1616 	} else
1617 		bq = &bd->bd_dirtyq;
1618 	bq_insert(bq, bp, true);
1619 }
1620 
1621 /*
1622  * buf_free:
1623  *
1624  *	Free a buffer to the buf zone once it no longer has valid contents.
1625  */
1626 static void
1627 buf_free(struct buf *bp)
1628 {
1629 
1630 	if (bp->b_flags & B_REMFREE)
1631 		bremfreef(bp);
1632 	if (bp->b_vflags & BV_BKGRDINPROG)
1633 		panic("losing buffer 1");
1634 	if (bp->b_rcred != NOCRED) {
1635 		crfree(bp->b_rcred);
1636 		bp->b_rcred = NOCRED;
1637 	}
1638 	if (bp->b_wcred != NOCRED) {
1639 		crfree(bp->b_wcred);
1640 		bp->b_wcred = NOCRED;
1641 	}
1642 	if (!LIST_EMPTY(&bp->b_dep))
1643 		buf_deallocate(bp);
1644 	bufkva_free(bp);
1645 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1646 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
1647 	BUF_UNLOCK(bp);
1648 	uma_zfree(buf_zone, bp);
1649 }
1650 
1651 /*
1652  * buf_import:
1653  *
1654  *	Import bufs into the uma cache from the buf list.  The system still
1655  *	expects a static array of bufs and much of the synchronization
1656  *	around bufs assumes type stable storage.  As a result, UMA is used
1657  *	only as a per-cpu cache of bufs still maintained on a global list.
1658  */
1659 static int
1660 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1661 {
1662 	struct buf *bp;
1663 	int i;
1664 
1665 	BQ_LOCK(&bqempty);
1666 	for (i = 0; i < cnt; i++) {
1667 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1668 		if (bp == NULL)
1669 			break;
1670 		bq_remove(&bqempty, bp);
1671 		store[i] = bp;
1672 	}
1673 	BQ_UNLOCK(&bqempty);
1674 
1675 	return (i);
1676 }
1677 
1678 /*
1679  * buf_release:
1680  *
1681  *	Release bufs from the uma cache back to the buffer queues.
1682  */
1683 static void
1684 buf_release(void *arg, void **store, int cnt)
1685 {
1686 	struct bufqueue *bq;
1687 	struct buf *bp;
1688         int i;
1689 
1690 	bq = &bqempty;
1691 	BQ_LOCK(bq);
1692         for (i = 0; i < cnt; i++) {
1693 		bp = store[i];
1694 		/* Inline bq_insert() to batch locking. */
1695 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1696 		bp->b_flags &= ~(B_AGE | B_REUSE);
1697 		bq->bq_len++;
1698 		bp->b_qindex = bq->bq_index;
1699 	}
1700 	BQ_UNLOCK(bq);
1701 }
1702 
1703 /*
1704  * buf_alloc:
1705  *
1706  *	Allocate an empty buffer header.
1707  */
1708 static struct buf *
1709 buf_alloc(struct bufdomain *bd)
1710 {
1711 	struct buf *bp;
1712 	int freebufs, error;
1713 
1714 	/*
1715 	 * We can only run out of bufs in the buf zone if the average buf
1716 	 * is less than BKVASIZE.  In this case the actual wait/block will
1717 	 * come from buf_reycle() failing to flush one of these small bufs.
1718 	 */
1719 	bp = NULL;
1720 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1721 	if (freebufs > 0)
1722 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1723 	if (bp == NULL) {
1724 		atomic_add_int(&bd->bd_freebuffers, 1);
1725 		bufspace_daemon_wakeup(bd);
1726 		counter_u64_add(numbufallocfails, 1);
1727 		return (NULL);
1728 	}
1729 	/*
1730 	 * Wake-up the bufspace daemon on transition below threshold.
1731 	 */
1732 	if (freebufs == bd->bd_lofreebuffers)
1733 		bufspace_daemon_wakeup(bd);
1734 
1735 	error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWITNESS, NULL);
1736 	KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1737 	    error));
1738 	(void)error;
1739 
1740 	KASSERT(bp->b_vp == NULL,
1741 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1742 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1743 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1744 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1745 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1746 	KASSERT(bp->b_npages == 0,
1747 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1748 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1749 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1750 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
1751 
1752 	bp->b_domain = BD_DOMAIN(bd);
1753 	bp->b_flags = 0;
1754 	bp->b_ioflags = 0;
1755 	bp->b_xflags = 0;
1756 	bp->b_vflags = 0;
1757 	bp->b_vp = NULL;
1758 	bp->b_blkno = bp->b_lblkno = 0;
1759 	bp->b_offset = NOOFFSET;
1760 	bp->b_iodone = 0;
1761 	bp->b_error = 0;
1762 	bp->b_resid = 0;
1763 	bp->b_bcount = 0;
1764 	bp->b_npages = 0;
1765 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1766 	bp->b_bufobj = NULL;
1767 	bp->b_data = bp->b_kvabase = unmapped_buf;
1768 	bp->b_fsprivate1 = NULL;
1769 	bp->b_fsprivate2 = NULL;
1770 	bp->b_fsprivate3 = NULL;
1771 	LIST_INIT(&bp->b_dep);
1772 
1773 	return (bp);
1774 }
1775 
1776 /*
1777  *	buf_recycle:
1778  *
1779  *	Free a buffer from the given bufqueue.  kva controls whether the
1780  *	freed buf must own some kva resources.  This is used for
1781  *	defragmenting.
1782  */
1783 static int
1784 buf_recycle(struct bufdomain *bd, bool kva)
1785 {
1786 	struct bufqueue *bq;
1787 	struct buf *bp, *nbp;
1788 
1789 	if (kva)
1790 		counter_u64_add(bufdefragcnt, 1);
1791 	nbp = NULL;
1792 	bq = bd->bd_cleanq;
1793 	BQ_LOCK(bq);
1794 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1795 	    ("buf_recycle: Locks don't match"));
1796 	nbp = TAILQ_FIRST(&bq->bq_queue);
1797 
1798 	/*
1799 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1800 	 * depending.
1801 	 */
1802 	while ((bp = nbp) != NULL) {
1803 		/*
1804 		 * Calculate next bp (we can only use it if we do not
1805 		 * release the bqlock).
1806 		 */
1807 		nbp = TAILQ_NEXT(bp, b_freelist);
1808 
1809 		/*
1810 		 * If we are defragging then we need a buffer with
1811 		 * some kva to reclaim.
1812 		 */
1813 		if (kva && bp->b_kvasize == 0)
1814 			continue;
1815 
1816 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1817 			continue;
1818 
1819 		/*
1820 		 * Implement a second chance algorithm for frequently
1821 		 * accessed buffers.
1822 		 */
1823 		if ((bp->b_flags & B_REUSE) != 0) {
1824 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1825 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1826 			bp->b_flags &= ~B_REUSE;
1827 			BUF_UNLOCK(bp);
1828 			continue;
1829 		}
1830 
1831 		/*
1832 		 * Skip buffers with background writes in progress.
1833 		 */
1834 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1835 			BUF_UNLOCK(bp);
1836 			continue;
1837 		}
1838 
1839 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1840 		    ("buf_recycle: inconsistent queue %d bp %p",
1841 		    bp->b_qindex, bp));
1842 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1843 		    ("getnewbuf: queue domain %d doesn't match request %d",
1844 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1845 		/*
1846 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1847 		 * the scan from this point on.
1848 		 */
1849 		bq_remove(bq, bp);
1850 		BQ_UNLOCK(bq);
1851 
1852 		/*
1853 		 * Requeue the background write buffer with error and
1854 		 * restart the scan.
1855 		 */
1856 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1857 			bqrelse(bp);
1858 			BQ_LOCK(bq);
1859 			nbp = TAILQ_FIRST(&bq->bq_queue);
1860 			continue;
1861 		}
1862 		bp->b_flags |= B_INVAL;
1863 		brelse(bp);
1864 		return (0);
1865 	}
1866 	bd->bd_wanted = 1;
1867 	BQ_UNLOCK(bq);
1868 
1869 	return (ENOBUFS);
1870 }
1871 
1872 /*
1873  *	bremfree:
1874  *
1875  *	Mark the buffer for removal from the appropriate free list.
1876  *
1877  */
1878 void
1879 bremfree(struct buf *bp)
1880 {
1881 
1882 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1883 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1884 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1885 	KASSERT(bp->b_qindex != QUEUE_NONE,
1886 	    ("bremfree: buffer %p not on a queue.", bp));
1887 	BUF_ASSERT_XLOCKED(bp);
1888 
1889 	bp->b_flags |= B_REMFREE;
1890 }
1891 
1892 /*
1893  *	bremfreef:
1894  *
1895  *	Force an immediate removal from a free list.  Used only in nfs when
1896  *	it abuses the b_freelist pointer.
1897  */
1898 void
1899 bremfreef(struct buf *bp)
1900 {
1901 	struct bufqueue *bq;
1902 
1903 	bq = bufqueue_acquire(bp);
1904 	bq_remove(bq, bp);
1905 	BQ_UNLOCK(bq);
1906 }
1907 
1908 static void
1909 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1910 {
1911 
1912 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1913 	TAILQ_INIT(&bq->bq_queue);
1914 	bq->bq_len = 0;
1915 	bq->bq_index = qindex;
1916 	bq->bq_subqueue = subqueue;
1917 }
1918 
1919 static void
1920 bd_init(struct bufdomain *bd)
1921 {
1922 	int i;
1923 
1924 	/* Per-CPU clean buf queues, plus one global queue. */
1925 	bd->bd_subq = mallocarray(mp_maxid + 2, sizeof(struct bufqueue),
1926 	    M_BIOBUF, M_WAITOK | M_ZERO);
1927 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1928 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1929 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1930 	for (i = 0; i <= mp_maxid; i++)
1931 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1932 		    "bufq clean subqueue lock");
1933 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1934 }
1935 
1936 /*
1937  *	bq_remove:
1938  *
1939  *	Removes a buffer from the free list, must be called with the
1940  *	correct qlock held.
1941  */
1942 static void
1943 bq_remove(struct bufqueue *bq, struct buf *bp)
1944 {
1945 
1946 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1947 	    bp, bp->b_vp, bp->b_flags);
1948 	KASSERT(bp->b_qindex != QUEUE_NONE,
1949 	    ("bq_remove: buffer %p not on a queue.", bp));
1950 	KASSERT(bufqueue(bp) == bq,
1951 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1952 
1953 	BQ_ASSERT_LOCKED(bq);
1954 	if (bp->b_qindex != QUEUE_EMPTY) {
1955 		BUF_ASSERT_XLOCKED(bp);
1956 	}
1957 	KASSERT(bq->bq_len >= 1,
1958 	    ("queue %d underflow", bp->b_qindex));
1959 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1960 	bq->bq_len--;
1961 	bp->b_qindex = QUEUE_NONE;
1962 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1963 }
1964 
1965 static void
1966 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1967 {
1968 	struct buf *bp;
1969 
1970 	BQ_ASSERT_LOCKED(bq);
1971 	if (bq != bd->bd_cleanq) {
1972 		BD_LOCK(bd);
1973 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1974 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1975 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1976 			    b_freelist);
1977 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1978 		}
1979 		bd->bd_cleanq->bq_len += bq->bq_len;
1980 		bq->bq_len = 0;
1981 	}
1982 	if (bd->bd_wanted) {
1983 		bd->bd_wanted = 0;
1984 		wakeup(&bd->bd_wanted);
1985 	}
1986 	if (bq != bd->bd_cleanq)
1987 		BD_UNLOCK(bd);
1988 }
1989 
1990 static int
1991 bd_flushall(struct bufdomain *bd)
1992 {
1993 	struct bufqueue *bq;
1994 	int flushed;
1995 	int i;
1996 
1997 	if (bd->bd_lim == 0)
1998 		return (0);
1999 	flushed = 0;
2000 	for (i = 0; i <= mp_maxid; i++) {
2001 		bq = &bd->bd_subq[i];
2002 		if (bq->bq_len == 0)
2003 			continue;
2004 		BQ_LOCK(bq);
2005 		bd_flush(bd, bq);
2006 		BQ_UNLOCK(bq);
2007 		flushed++;
2008 	}
2009 
2010 	return (flushed);
2011 }
2012 
2013 static void
2014 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
2015 {
2016 	struct bufdomain *bd;
2017 
2018 	if (bp->b_qindex != QUEUE_NONE)
2019 		panic("bq_insert: free buffer %p onto another queue?", bp);
2020 
2021 	bd = bufdomain(bp);
2022 	if (bp->b_flags & B_AGE) {
2023 		/* Place this buf directly on the real queue. */
2024 		if (bq->bq_index == QUEUE_CLEAN)
2025 			bq = bd->bd_cleanq;
2026 		BQ_LOCK(bq);
2027 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
2028 	} else {
2029 		BQ_LOCK(bq);
2030 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
2031 	}
2032 	bp->b_flags &= ~(B_AGE | B_REUSE);
2033 	bq->bq_len++;
2034 	bp->b_qindex = bq->bq_index;
2035 	bp->b_subqueue = bq->bq_subqueue;
2036 
2037 	/*
2038 	 * Unlock before we notify so that we don't wakeup a waiter that
2039 	 * fails a trylock on the buf and sleeps again.
2040 	 */
2041 	if (unlock)
2042 		BUF_UNLOCK(bp);
2043 
2044 	if (bp->b_qindex == QUEUE_CLEAN) {
2045 		/*
2046 		 * Flush the per-cpu queue and notify any waiters.
2047 		 */
2048 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2049 		    bq->bq_len >= bd->bd_lim))
2050 			bd_flush(bd, bq);
2051 	}
2052 	BQ_UNLOCK(bq);
2053 }
2054 
2055 /*
2056  *	bufkva_free:
2057  *
2058  *	Free the kva allocation for a buffer.
2059  *
2060  */
2061 static void
2062 bufkva_free(struct buf *bp)
2063 {
2064 
2065 #ifdef INVARIANTS
2066 	if (bp->b_kvasize == 0) {
2067 		KASSERT(bp->b_kvabase == unmapped_buf &&
2068 		    bp->b_data == unmapped_buf,
2069 		    ("Leaked KVA space on %p", bp));
2070 	} else if (buf_mapped(bp))
2071 		BUF_CHECK_MAPPED(bp);
2072 	else
2073 		BUF_CHECK_UNMAPPED(bp);
2074 #endif
2075 	if (bp->b_kvasize == 0)
2076 		return;
2077 
2078 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2079 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2080 	counter_u64_add(buffreekvacnt, 1);
2081 	bp->b_data = bp->b_kvabase = unmapped_buf;
2082 	bp->b_kvasize = 0;
2083 }
2084 
2085 /*
2086  *	bufkva_alloc:
2087  *
2088  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2089  */
2090 static int
2091 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2092 {
2093 	vm_offset_t addr;
2094 	int error;
2095 
2096 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2097 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2098 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
2099 	KASSERT(maxsize <= maxbcachebuf,
2100 	    ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2101 
2102 	bufkva_free(bp);
2103 
2104 	addr = 0;
2105 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2106 	if (error != 0) {
2107 		/*
2108 		 * Buffer map is too fragmented.  Request the caller
2109 		 * to defragment the map.
2110 		 */
2111 		return (error);
2112 	}
2113 	bp->b_kvabase = (caddr_t)addr;
2114 	bp->b_kvasize = maxsize;
2115 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2116 	if ((gbflags & GB_UNMAPPED) != 0) {
2117 		bp->b_data = unmapped_buf;
2118 		BUF_CHECK_UNMAPPED(bp);
2119 	} else {
2120 		bp->b_data = bp->b_kvabase;
2121 		BUF_CHECK_MAPPED(bp);
2122 	}
2123 	return (0);
2124 }
2125 
2126 /*
2127  *	bufkva_reclaim:
2128  *
2129  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2130  *	callback that fires to avoid returning failure.
2131  */
2132 static void
2133 bufkva_reclaim(vmem_t *vmem, int flags)
2134 {
2135 	bool done;
2136 	int q;
2137 	int i;
2138 
2139 	done = false;
2140 	for (i = 0; i < 5; i++) {
2141 		for (q = 0; q < buf_domains; q++)
2142 			if (buf_recycle(&bdomain[q], true) != 0)
2143 				done = true;
2144 		if (done)
2145 			break;
2146 	}
2147 	return;
2148 }
2149 
2150 /*
2151  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2152  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2153  * the buffer is valid and we do not have to do anything.
2154  */
2155 static void
2156 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2157     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2158 {
2159 	struct buf *rabp;
2160 	struct thread *td;
2161 	int i;
2162 
2163 	td = curthread;
2164 
2165 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2166 		if (inmem(vp, *rablkno))
2167 			continue;
2168 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2169 		if ((rabp->b_flags & B_CACHE) != 0) {
2170 			brelse(rabp);
2171 			continue;
2172 		}
2173 #ifdef RACCT
2174 		if (racct_enable) {
2175 			PROC_LOCK(curproc);
2176 			racct_add_buf(curproc, rabp, 0);
2177 			PROC_UNLOCK(curproc);
2178 		}
2179 #endif /* RACCT */
2180 		td->td_ru.ru_inblock++;
2181 		rabp->b_flags |= B_ASYNC;
2182 		rabp->b_flags &= ~B_INVAL;
2183 		if ((flags & GB_CKHASH) != 0) {
2184 			rabp->b_flags |= B_CKHASH;
2185 			rabp->b_ckhashcalc = ckhashfunc;
2186 		}
2187 		rabp->b_ioflags &= ~BIO_ERROR;
2188 		rabp->b_iocmd = BIO_READ;
2189 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2190 			rabp->b_rcred = crhold(cred);
2191 		vfs_busy_pages(rabp, 0);
2192 		BUF_KERNPROC(rabp);
2193 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2194 		bstrategy(rabp);
2195 	}
2196 }
2197 
2198 /*
2199  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2200  *
2201  * Get a buffer with the specified data.  Look in the cache first.  We
2202  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2203  * is set, the buffer is valid and we do not have to do anything, see
2204  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2205  *
2206  * Always return a NULL buffer pointer (in bpp) when returning an error.
2207  *
2208  * The blkno parameter is the logical block being requested. Normally
2209  * the mapping of logical block number to disk block address is done
2210  * by calling VOP_BMAP(). However, if the mapping is already known, the
2211  * disk block address can be passed using the dblkno parameter. If the
2212  * disk block address is not known, then the same value should be passed
2213  * for blkno and dblkno.
2214  */
2215 int
2216 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2217     daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2218     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2219 {
2220 	struct buf *bp;
2221 	struct thread *td;
2222 	int error, readwait, rv;
2223 
2224 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2225 	td = curthread;
2226 	/*
2227 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2228 	 * are specified.
2229 	 */
2230 	error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2231 	if (error != 0) {
2232 		*bpp = NULL;
2233 		return (error);
2234 	}
2235 	KASSERT(blkno == bp->b_lblkno,
2236 	    ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2237 	    (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2238 	flags &= ~GB_NOSPARSE;
2239 	*bpp = bp;
2240 
2241 	/*
2242 	 * If not found in cache, do some I/O
2243 	 */
2244 	readwait = 0;
2245 	if ((bp->b_flags & B_CACHE) == 0) {
2246 #ifdef RACCT
2247 		if (racct_enable) {
2248 			PROC_LOCK(td->td_proc);
2249 			racct_add_buf(td->td_proc, bp, 0);
2250 			PROC_UNLOCK(td->td_proc);
2251 		}
2252 #endif /* RACCT */
2253 		td->td_ru.ru_inblock++;
2254 		bp->b_iocmd = BIO_READ;
2255 		bp->b_flags &= ~B_INVAL;
2256 		if ((flags & GB_CKHASH) != 0) {
2257 			bp->b_flags |= B_CKHASH;
2258 			bp->b_ckhashcalc = ckhashfunc;
2259 		}
2260 		if ((flags & GB_CVTENXIO) != 0)
2261 			bp->b_xflags |= BX_CVTENXIO;
2262 		bp->b_ioflags &= ~BIO_ERROR;
2263 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2264 			bp->b_rcred = crhold(cred);
2265 		vfs_busy_pages(bp, 0);
2266 		bp->b_iooffset = dbtob(bp->b_blkno);
2267 		bstrategy(bp);
2268 		++readwait;
2269 	}
2270 
2271 	/*
2272 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2273 	 */
2274 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2275 
2276 	rv = 0;
2277 	if (readwait) {
2278 		rv = bufwait(bp);
2279 		if (rv != 0) {
2280 			brelse(bp);
2281 			*bpp = NULL;
2282 		}
2283 	}
2284 	return (rv);
2285 }
2286 
2287 /*
2288  * Write, release buffer on completion.  (Done by iodone
2289  * if async).  Do not bother writing anything if the buffer
2290  * is invalid.
2291  *
2292  * Note that we set B_CACHE here, indicating that buffer is
2293  * fully valid and thus cacheable.  This is true even of NFS
2294  * now so we set it generally.  This could be set either here
2295  * or in biodone() since the I/O is synchronous.  We put it
2296  * here.
2297  */
2298 int
2299 bufwrite(struct buf *bp)
2300 {
2301 	int oldflags;
2302 	struct vnode *vp;
2303 	long space;
2304 	int vp_md;
2305 
2306 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2307 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2308 		bp->b_flags |= B_INVAL | B_RELBUF;
2309 		bp->b_flags &= ~B_CACHE;
2310 		brelse(bp);
2311 		return (ENXIO);
2312 	}
2313 	if (bp->b_flags & B_INVAL) {
2314 		brelse(bp);
2315 		return (0);
2316 	}
2317 
2318 	if (bp->b_flags & B_BARRIER)
2319 		atomic_add_long(&barrierwrites, 1);
2320 
2321 	oldflags = bp->b_flags;
2322 
2323 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2324 	    ("FFS background buffer should not get here %p", bp));
2325 
2326 	vp = bp->b_vp;
2327 	if (vp)
2328 		vp_md = vp->v_vflag & VV_MD;
2329 	else
2330 		vp_md = 0;
2331 
2332 	/*
2333 	 * Mark the buffer clean.  Increment the bufobj write count
2334 	 * before bundirty() call, to prevent other thread from seeing
2335 	 * empty dirty list and zero counter for writes in progress,
2336 	 * falsely indicating that the bufobj is clean.
2337 	 */
2338 	bufobj_wref(bp->b_bufobj);
2339 	bundirty(bp);
2340 
2341 	bp->b_flags &= ~B_DONE;
2342 	bp->b_ioflags &= ~BIO_ERROR;
2343 	bp->b_flags |= B_CACHE;
2344 	bp->b_iocmd = BIO_WRITE;
2345 
2346 	vfs_busy_pages(bp, 1);
2347 
2348 	/*
2349 	 * Normal bwrites pipeline writes
2350 	 */
2351 	bp->b_runningbufspace = bp->b_bufsize;
2352 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2353 
2354 #ifdef RACCT
2355 	if (racct_enable) {
2356 		PROC_LOCK(curproc);
2357 		racct_add_buf(curproc, bp, 1);
2358 		PROC_UNLOCK(curproc);
2359 	}
2360 #endif /* RACCT */
2361 	curthread->td_ru.ru_oublock++;
2362 	if (oldflags & B_ASYNC)
2363 		BUF_KERNPROC(bp);
2364 	bp->b_iooffset = dbtob(bp->b_blkno);
2365 	buf_track(bp, __func__);
2366 	bstrategy(bp);
2367 
2368 	if ((oldflags & B_ASYNC) == 0) {
2369 		int rtval = bufwait(bp);
2370 		brelse(bp);
2371 		return (rtval);
2372 	} else if (space > hirunningspace) {
2373 		/*
2374 		 * don't allow the async write to saturate the I/O
2375 		 * system.  We will not deadlock here because
2376 		 * we are blocking waiting for I/O that is already in-progress
2377 		 * to complete. We do not block here if it is the update
2378 		 * or syncer daemon trying to clean up as that can lead
2379 		 * to deadlock.
2380 		 */
2381 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2382 			waitrunningbufspace();
2383 	}
2384 
2385 	return (0);
2386 }
2387 
2388 void
2389 bufbdflush(struct bufobj *bo, struct buf *bp)
2390 {
2391 	struct buf *nbp;
2392 	struct bufdomain *bd;
2393 
2394 	bd = &bdomain[bo->bo_domain];
2395 	if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2396 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2397 		altbufferflushes++;
2398 	} else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2399 		BO_LOCK(bo);
2400 		/*
2401 		 * Try to find a buffer to flush.
2402 		 */
2403 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2404 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2405 			    BUF_LOCK(nbp,
2406 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2407 				continue;
2408 			if (bp == nbp)
2409 				panic("bdwrite: found ourselves");
2410 			BO_UNLOCK(bo);
2411 			/* Don't countdeps with the bo lock held. */
2412 			if (buf_countdeps(nbp, 0)) {
2413 				BO_LOCK(bo);
2414 				BUF_UNLOCK(nbp);
2415 				continue;
2416 			}
2417 			if (nbp->b_flags & B_CLUSTEROK) {
2418 				vfs_bio_awrite(nbp);
2419 			} else {
2420 				bremfree(nbp);
2421 				bawrite(nbp);
2422 			}
2423 			dirtybufferflushes++;
2424 			break;
2425 		}
2426 		if (nbp == NULL)
2427 			BO_UNLOCK(bo);
2428 	}
2429 }
2430 
2431 /*
2432  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2433  * anything if the buffer is marked invalid.
2434  *
2435  * Note that since the buffer must be completely valid, we can safely
2436  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2437  * biodone() in order to prevent getblk from writing the buffer
2438  * out synchronously.
2439  */
2440 void
2441 bdwrite(struct buf *bp)
2442 {
2443 	struct thread *td = curthread;
2444 	struct vnode *vp;
2445 	struct bufobj *bo;
2446 
2447 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2448 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2449 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2450 	    ("Barrier request in delayed write %p", bp));
2451 
2452 	if (bp->b_flags & B_INVAL) {
2453 		brelse(bp);
2454 		return;
2455 	}
2456 
2457 	/*
2458 	 * If we have too many dirty buffers, don't create any more.
2459 	 * If we are wildly over our limit, then force a complete
2460 	 * cleanup. Otherwise, just keep the situation from getting
2461 	 * out of control. Note that we have to avoid a recursive
2462 	 * disaster and not try to clean up after our own cleanup!
2463 	 */
2464 	vp = bp->b_vp;
2465 	bo = bp->b_bufobj;
2466 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2467 		td->td_pflags |= TDP_INBDFLUSH;
2468 		BO_BDFLUSH(bo, bp);
2469 		td->td_pflags &= ~TDP_INBDFLUSH;
2470 	} else
2471 		recursiveflushes++;
2472 
2473 	bdirty(bp);
2474 	/*
2475 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2476 	 * true even of NFS now.
2477 	 */
2478 	bp->b_flags |= B_CACHE;
2479 
2480 	/*
2481 	 * This bmap keeps the system from needing to do the bmap later,
2482 	 * perhaps when the system is attempting to do a sync.  Since it
2483 	 * is likely that the indirect block -- or whatever other datastructure
2484 	 * that the filesystem needs is still in memory now, it is a good
2485 	 * thing to do this.  Note also, that if the pageout daemon is
2486 	 * requesting a sync -- there might not be enough memory to do
2487 	 * the bmap then...  So, this is important to do.
2488 	 */
2489 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2490 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2491 	}
2492 
2493 	buf_track(bp, __func__);
2494 
2495 	/*
2496 	 * Set the *dirty* buffer range based upon the VM system dirty
2497 	 * pages.
2498 	 *
2499 	 * Mark the buffer pages as clean.  We need to do this here to
2500 	 * satisfy the vnode_pager and the pageout daemon, so that it
2501 	 * thinks that the pages have been "cleaned".  Note that since
2502 	 * the pages are in a delayed write buffer -- the VFS layer
2503 	 * "will" see that the pages get written out on the next sync,
2504 	 * or perhaps the cluster will be completed.
2505 	 */
2506 	vfs_clean_pages_dirty_buf(bp);
2507 	bqrelse(bp);
2508 
2509 	/*
2510 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2511 	 * due to the softdep code.
2512 	 */
2513 }
2514 
2515 /*
2516  *	bdirty:
2517  *
2518  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2519  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2520  *	itself to properly update it in the dirty/clean lists.  We mark it
2521  *	B_DONE to ensure that any asynchronization of the buffer properly
2522  *	clears B_DONE ( else a panic will occur later ).
2523  *
2524  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2525  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2526  *	should only be called if the buffer is known-good.
2527  *
2528  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2529  *	count.
2530  *
2531  *	The buffer must be on QUEUE_NONE.
2532  */
2533 void
2534 bdirty(struct buf *bp)
2535 {
2536 
2537 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2538 	    bp, bp->b_vp, bp->b_flags);
2539 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2540 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2541 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2542 	bp->b_flags &= ~(B_RELBUF);
2543 	bp->b_iocmd = BIO_WRITE;
2544 
2545 	if ((bp->b_flags & B_DELWRI) == 0) {
2546 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2547 		reassignbuf(bp);
2548 		bdirtyadd(bp);
2549 	}
2550 }
2551 
2552 /*
2553  *	bundirty:
2554  *
2555  *	Clear B_DELWRI for buffer.
2556  *
2557  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2558  *	count.
2559  *
2560  *	The buffer must be on QUEUE_NONE.
2561  */
2562 
2563 void
2564 bundirty(struct buf *bp)
2565 {
2566 
2567 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2568 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2569 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2570 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2571 
2572 	if (bp->b_flags & B_DELWRI) {
2573 		bp->b_flags &= ~B_DELWRI;
2574 		reassignbuf(bp);
2575 		bdirtysub(bp);
2576 	}
2577 	/*
2578 	 * Since it is now being written, we can clear its deferred write flag.
2579 	 */
2580 	bp->b_flags &= ~B_DEFERRED;
2581 }
2582 
2583 /*
2584  *	bawrite:
2585  *
2586  *	Asynchronous write.  Start output on a buffer, but do not wait for
2587  *	it to complete.  The buffer is released when the output completes.
2588  *
2589  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2590  *	B_INVAL buffers.  Not us.
2591  */
2592 void
2593 bawrite(struct buf *bp)
2594 {
2595 
2596 	bp->b_flags |= B_ASYNC;
2597 	(void) bwrite(bp);
2598 }
2599 
2600 /*
2601  *	babarrierwrite:
2602  *
2603  *	Asynchronous barrier write.  Start output on a buffer, but do not
2604  *	wait for it to complete.  Place a write barrier after this write so
2605  *	that this buffer and all buffers written before it are committed to
2606  *	the disk before any buffers written after this write are committed
2607  *	to the disk.  The buffer is released when the output completes.
2608  */
2609 void
2610 babarrierwrite(struct buf *bp)
2611 {
2612 
2613 	bp->b_flags |= B_ASYNC | B_BARRIER;
2614 	(void) bwrite(bp);
2615 }
2616 
2617 /*
2618  *	bbarrierwrite:
2619  *
2620  *	Synchronous barrier write.  Start output on a buffer and wait for
2621  *	it to complete.  Place a write barrier after this write so that
2622  *	this buffer and all buffers written before it are committed to
2623  *	the disk before any buffers written after this write are committed
2624  *	to the disk.  The buffer is released when the output completes.
2625  */
2626 int
2627 bbarrierwrite(struct buf *bp)
2628 {
2629 
2630 	bp->b_flags |= B_BARRIER;
2631 	return (bwrite(bp));
2632 }
2633 
2634 /*
2635  *	bwillwrite:
2636  *
2637  *	Called prior to the locking of any vnodes when we are expecting to
2638  *	write.  We do not want to starve the buffer cache with too many
2639  *	dirty buffers so we block here.  By blocking prior to the locking
2640  *	of any vnodes we attempt to avoid the situation where a locked vnode
2641  *	prevents the various system daemons from flushing related buffers.
2642  */
2643 void
2644 bwillwrite(void)
2645 {
2646 
2647 	if (buf_dirty_count_severe()) {
2648 		mtx_lock(&bdirtylock);
2649 		while (buf_dirty_count_severe()) {
2650 			bdirtywait = 1;
2651 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2652 			    "flswai", 0);
2653 		}
2654 		mtx_unlock(&bdirtylock);
2655 	}
2656 }
2657 
2658 /*
2659  * Return true if we have too many dirty buffers.
2660  */
2661 int
2662 buf_dirty_count_severe(void)
2663 {
2664 
2665 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2666 }
2667 
2668 /*
2669  *	brelse:
2670  *
2671  *	Release a busy buffer and, if requested, free its resources.  The
2672  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2673  *	to be accessed later as a cache entity or reused for other purposes.
2674  */
2675 void
2676 brelse(struct buf *bp)
2677 {
2678 	struct mount *v_mnt;
2679 	int qindex;
2680 
2681 	/*
2682 	 * Many functions erroneously call brelse with a NULL bp under rare
2683 	 * error conditions. Simply return when called with a NULL bp.
2684 	 */
2685 	if (bp == NULL)
2686 		return;
2687 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2688 	    bp, bp->b_vp, bp->b_flags);
2689 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2690 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2691 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2692 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2693 
2694 	if (BUF_LOCKRECURSED(bp)) {
2695 		/*
2696 		 * Do not process, in particular, do not handle the
2697 		 * B_INVAL/B_RELBUF and do not release to free list.
2698 		 */
2699 		BUF_UNLOCK(bp);
2700 		return;
2701 	}
2702 
2703 	if (bp->b_flags & B_MANAGED) {
2704 		bqrelse(bp);
2705 		return;
2706 	}
2707 
2708 	if (LIST_EMPTY(&bp->b_dep)) {
2709 		bp->b_flags &= ~B_IOSTARTED;
2710 	} else {
2711 		KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2712 		    ("brelse: SU io not finished bp %p", bp));
2713 	}
2714 
2715 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2716 		BO_LOCK(bp->b_bufobj);
2717 		bp->b_vflags &= ~BV_BKGRDERR;
2718 		BO_UNLOCK(bp->b_bufobj);
2719 		bdirty(bp);
2720 	}
2721 
2722 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2723 	    (bp->b_flags & B_INVALONERR)) {
2724 		/*
2725 		 * Forced invalidation of dirty buffer contents, to be used
2726 		 * after a failed write in the rare case that the loss of the
2727 		 * contents is acceptable.  The buffer is invalidated and
2728 		 * freed.
2729 		 */
2730 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2731 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2732 	}
2733 
2734 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2735 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2736 	    !(bp->b_flags & B_INVAL)) {
2737 		/*
2738 		 * Failed write, redirty.  All errors except ENXIO (which
2739 		 * means the device is gone) are treated as being
2740 		 * transient.
2741 		 *
2742 		 * XXX Treating EIO as transient is not correct; the
2743 		 * contract with the local storage device drivers is that
2744 		 * they will only return EIO once the I/O is no longer
2745 		 * retriable.  Network I/O also respects this through the
2746 		 * guarantees of TCP and/or the internal retries of NFS.
2747 		 * ENOMEM might be transient, but we also have no way of
2748 		 * knowing when its ok to retry/reschedule.  In general,
2749 		 * this entire case should be made obsolete through better
2750 		 * error handling/recovery and resource scheduling.
2751 		 *
2752 		 * Do this also for buffers that failed with ENXIO, but have
2753 		 * non-empty dependencies - the soft updates code might need
2754 		 * to access the buffer to untangle them.
2755 		 *
2756 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2757 		 */
2758 		bp->b_ioflags &= ~BIO_ERROR;
2759 		bdirty(bp);
2760 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2761 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2762 		/*
2763 		 * Either a failed read I/O, or we were asked to free or not
2764 		 * cache the buffer, or we failed to write to a device that's
2765 		 * no longer present.
2766 		 */
2767 		bp->b_flags |= B_INVAL;
2768 		if (!LIST_EMPTY(&bp->b_dep))
2769 			buf_deallocate(bp);
2770 		if (bp->b_flags & B_DELWRI)
2771 			bdirtysub(bp);
2772 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2773 		if ((bp->b_flags & B_VMIO) == 0) {
2774 			allocbuf(bp, 0);
2775 			if (bp->b_vp)
2776 				brelvp(bp);
2777 		}
2778 	}
2779 
2780 	/*
2781 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2782 	 * is called with B_DELWRI set, the underlying pages may wind up
2783 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2784 	 * because pages associated with a B_DELWRI bp are marked clean.
2785 	 *
2786 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2787 	 * if B_DELWRI is set.
2788 	 */
2789 	if (bp->b_flags & B_DELWRI)
2790 		bp->b_flags &= ~B_RELBUF;
2791 
2792 	/*
2793 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2794 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2795 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2796 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2797 	 *
2798 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2799 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2800 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2801 	 *
2802 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2803 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2804 	 * the commit state and we cannot afford to lose the buffer. If the
2805 	 * buffer has a background write in progress, we need to keep it
2806 	 * around to prevent it from being reconstituted and starting a second
2807 	 * background write.
2808 	 */
2809 
2810 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2811 
2812 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2813 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2814 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2815 	    vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2816 		vfs_vmio_invalidate(bp);
2817 		allocbuf(bp, 0);
2818 	}
2819 
2820 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2821 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2822 		allocbuf(bp, 0);
2823 		bp->b_flags &= ~B_NOREUSE;
2824 		if (bp->b_vp != NULL)
2825 			brelvp(bp);
2826 	}
2827 
2828 	/*
2829 	 * If the buffer has junk contents signal it and eventually
2830 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2831 	 * doesn't find it.
2832 	 */
2833 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2834 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2835 		bp->b_flags |= B_INVAL;
2836 	if (bp->b_flags & B_INVAL) {
2837 		if (bp->b_flags & B_DELWRI)
2838 			bundirty(bp);
2839 		if (bp->b_vp)
2840 			brelvp(bp);
2841 	}
2842 
2843 	buf_track(bp, __func__);
2844 
2845 	/* buffers with no memory */
2846 	if (bp->b_bufsize == 0) {
2847 		buf_free(bp);
2848 		return;
2849 	}
2850 	/* buffers with junk contents */
2851 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2852 	    (bp->b_ioflags & BIO_ERROR)) {
2853 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2854 		if (bp->b_vflags & BV_BKGRDINPROG)
2855 			panic("losing buffer 2");
2856 		qindex = QUEUE_CLEAN;
2857 		bp->b_flags |= B_AGE;
2858 	/* remaining buffers */
2859 	} else if (bp->b_flags & B_DELWRI)
2860 		qindex = QUEUE_DIRTY;
2861 	else
2862 		qindex = QUEUE_CLEAN;
2863 
2864 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2865 		panic("brelse: not dirty");
2866 
2867 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2868 	bp->b_xflags &= ~(BX_CVTENXIO);
2869 	/* binsfree unlocks bp. */
2870 	binsfree(bp, qindex);
2871 }
2872 
2873 /*
2874  * Release a buffer back to the appropriate queue but do not try to free
2875  * it.  The buffer is expected to be used again soon.
2876  *
2877  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2878  * biodone() to requeue an async I/O on completion.  It is also used when
2879  * known good buffers need to be requeued but we think we may need the data
2880  * again soon.
2881  *
2882  * XXX we should be able to leave the B_RELBUF hint set on completion.
2883  */
2884 void
2885 bqrelse(struct buf *bp)
2886 {
2887 	int qindex;
2888 
2889 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2890 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2891 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2892 
2893 	qindex = QUEUE_NONE;
2894 	if (BUF_LOCKRECURSED(bp)) {
2895 		/* do not release to free list */
2896 		BUF_UNLOCK(bp);
2897 		return;
2898 	}
2899 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2900 	bp->b_xflags &= ~(BX_CVTENXIO);
2901 
2902 	if (LIST_EMPTY(&bp->b_dep)) {
2903 		bp->b_flags &= ~B_IOSTARTED;
2904 	} else {
2905 		KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2906 		    ("bqrelse: SU io not finished bp %p", bp));
2907 	}
2908 
2909 	if (bp->b_flags & B_MANAGED) {
2910 		if (bp->b_flags & B_REMFREE)
2911 			bremfreef(bp);
2912 		goto out;
2913 	}
2914 
2915 	/* buffers with stale but valid contents */
2916 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2917 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2918 		BO_LOCK(bp->b_bufobj);
2919 		bp->b_vflags &= ~BV_BKGRDERR;
2920 		BO_UNLOCK(bp->b_bufobj);
2921 		qindex = QUEUE_DIRTY;
2922 	} else {
2923 		if ((bp->b_flags & B_DELWRI) == 0 &&
2924 		    (bp->b_xflags & BX_VNDIRTY))
2925 			panic("bqrelse: not dirty");
2926 		if ((bp->b_flags & B_NOREUSE) != 0) {
2927 			brelse(bp);
2928 			return;
2929 		}
2930 		qindex = QUEUE_CLEAN;
2931 	}
2932 	buf_track(bp, __func__);
2933 	/* binsfree unlocks bp. */
2934 	binsfree(bp, qindex);
2935 	return;
2936 
2937 out:
2938 	buf_track(bp, __func__);
2939 	/* unlock */
2940 	BUF_UNLOCK(bp);
2941 }
2942 
2943 /*
2944  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2945  * restore bogus pages.
2946  */
2947 static void
2948 vfs_vmio_iodone(struct buf *bp)
2949 {
2950 	vm_ooffset_t foff;
2951 	vm_page_t m;
2952 	vm_object_t obj;
2953 	struct vnode *vp __unused;
2954 	int i, iosize, resid;
2955 	bool bogus;
2956 
2957 	obj = bp->b_bufobj->bo_object;
2958 	KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2959 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2960 	    blockcount_read(&obj->paging_in_progress), bp->b_npages));
2961 
2962 	vp = bp->b_vp;
2963 	VNPASS(vp->v_holdcnt > 0, vp);
2964 	VNPASS(vp->v_object != NULL, vp);
2965 
2966 	foff = bp->b_offset;
2967 	KASSERT(bp->b_offset != NOOFFSET,
2968 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2969 
2970 	bogus = false;
2971 	iosize = bp->b_bcount - bp->b_resid;
2972 	for (i = 0; i < bp->b_npages; i++) {
2973 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2974 		if (resid > iosize)
2975 			resid = iosize;
2976 
2977 		/*
2978 		 * cleanup bogus pages, restoring the originals
2979 		 */
2980 		m = bp->b_pages[i];
2981 		if (m == bogus_page) {
2982 			bogus = true;
2983 			m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2984 			if (m == NULL)
2985 				panic("biodone: page disappeared!");
2986 			bp->b_pages[i] = m;
2987 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2988 			/*
2989 			 * In the write case, the valid and clean bits are
2990 			 * already changed correctly ( see bdwrite() ), so we
2991 			 * only need to do this here in the read case.
2992 			 */
2993 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2994 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2995 			    "has unexpected dirty bits", m));
2996 			vfs_page_set_valid(bp, foff, m);
2997 		}
2998 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2999 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
3000 		    (intmax_t)foff, (uintmax_t)m->pindex));
3001 
3002 		vm_page_sunbusy(m);
3003 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3004 		iosize -= resid;
3005 	}
3006 	vm_object_pip_wakeupn(obj, bp->b_npages);
3007 	if (bogus && buf_mapped(bp)) {
3008 		BUF_CHECK_MAPPED(bp);
3009 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3010 		    bp->b_pages, bp->b_npages);
3011 	}
3012 }
3013 
3014 /*
3015  * Perform page invalidation when a buffer is released.  The fully invalid
3016  * pages will be reclaimed later in vfs_vmio_truncate().
3017  */
3018 static void
3019 vfs_vmio_invalidate(struct buf *bp)
3020 {
3021 	vm_object_t obj;
3022 	vm_page_t m;
3023 	int flags, i, resid, poffset, presid;
3024 
3025 	if (buf_mapped(bp)) {
3026 		BUF_CHECK_MAPPED(bp);
3027 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
3028 	} else
3029 		BUF_CHECK_UNMAPPED(bp);
3030 	/*
3031 	 * Get the base offset and length of the buffer.  Note that
3032 	 * in the VMIO case if the buffer block size is not
3033 	 * page-aligned then b_data pointer may not be page-aligned.
3034 	 * But our b_pages[] array *IS* page aligned.
3035 	 *
3036 	 * block sizes less then DEV_BSIZE (usually 512) are not
3037 	 * supported due to the page granularity bits (m->valid,
3038 	 * m->dirty, etc...).
3039 	 *
3040 	 * See man buf(9) for more information
3041 	 */
3042 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3043 	obj = bp->b_bufobj->bo_object;
3044 	resid = bp->b_bufsize;
3045 	poffset = bp->b_offset & PAGE_MASK;
3046 	VM_OBJECT_WLOCK(obj);
3047 	for (i = 0; i < bp->b_npages; i++) {
3048 		m = bp->b_pages[i];
3049 		if (m == bogus_page)
3050 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
3051 		bp->b_pages[i] = NULL;
3052 
3053 		presid = resid > (PAGE_SIZE - poffset) ?
3054 		    (PAGE_SIZE - poffset) : resid;
3055 		KASSERT(presid >= 0, ("brelse: extra page"));
3056 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3057 		if (pmap_page_wired_mappings(m) == 0)
3058 			vm_page_set_invalid(m, poffset, presid);
3059 		vm_page_sunbusy(m);
3060 		vm_page_release_locked(m, flags);
3061 		resid -= presid;
3062 		poffset = 0;
3063 	}
3064 	VM_OBJECT_WUNLOCK(obj);
3065 	bp->b_npages = 0;
3066 }
3067 
3068 /*
3069  * Page-granular truncation of an existing VMIO buffer.
3070  */
3071 static void
3072 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3073 {
3074 	vm_object_t obj;
3075 	vm_page_t m;
3076 	int flags, i;
3077 
3078 	if (bp->b_npages == desiredpages)
3079 		return;
3080 
3081 	if (buf_mapped(bp)) {
3082 		BUF_CHECK_MAPPED(bp);
3083 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3084 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3085 	} else
3086 		BUF_CHECK_UNMAPPED(bp);
3087 
3088 	/*
3089 	 * The object lock is needed only if we will attempt to free pages.
3090 	 */
3091 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3092 	if ((bp->b_flags & B_DIRECT) != 0) {
3093 		flags |= VPR_TRYFREE;
3094 		obj = bp->b_bufobj->bo_object;
3095 		VM_OBJECT_WLOCK(obj);
3096 	} else {
3097 		obj = NULL;
3098 	}
3099 	for (i = desiredpages; i < bp->b_npages; i++) {
3100 		m = bp->b_pages[i];
3101 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3102 		bp->b_pages[i] = NULL;
3103 		if (obj != NULL)
3104 			vm_page_release_locked(m, flags);
3105 		else
3106 			vm_page_release(m, flags);
3107 	}
3108 	if (obj != NULL)
3109 		VM_OBJECT_WUNLOCK(obj);
3110 	bp->b_npages = desiredpages;
3111 }
3112 
3113 /*
3114  * Byte granular extension of VMIO buffers.
3115  */
3116 static void
3117 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3118 {
3119 	/*
3120 	 * We are growing the buffer, possibly in a
3121 	 * byte-granular fashion.
3122 	 */
3123 	vm_object_t obj;
3124 	vm_offset_t toff;
3125 	vm_offset_t tinc;
3126 	vm_page_t m;
3127 
3128 	/*
3129 	 * Step 1, bring in the VM pages from the object, allocating
3130 	 * them if necessary.  We must clear B_CACHE if these pages
3131 	 * are not valid for the range covered by the buffer.
3132 	 */
3133 	obj = bp->b_bufobj->bo_object;
3134 	if (bp->b_npages < desiredpages) {
3135 		KASSERT(desiredpages <= atop(maxbcachebuf),
3136 		    ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3137 		    bp, desiredpages, maxbcachebuf));
3138 
3139 		/*
3140 		 * We must allocate system pages since blocking
3141 		 * here could interfere with paging I/O, no
3142 		 * matter which process we are.
3143 		 *
3144 		 * Only exclusive busy can be tested here.
3145 		 * Blocking on shared busy might lead to
3146 		 * deadlocks once allocbuf() is called after
3147 		 * pages are vfs_busy_pages().
3148 		 */
3149 		(void)vm_page_grab_pages_unlocked(obj,
3150 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3151 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3152 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3153 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3154 		bp->b_npages = desiredpages;
3155 	}
3156 
3157 	/*
3158 	 * Step 2.  We've loaded the pages into the buffer,
3159 	 * we have to figure out if we can still have B_CACHE
3160 	 * set.  Note that B_CACHE is set according to the
3161 	 * byte-granular range ( bcount and size ), not the
3162 	 * aligned range ( newbsize ).
3163 	 *
3164 	 * The VM test is against m->valid, which is DEV_BSIZE
3165 	 * aligned.  Needless to say, the validity of the data
3166 	 * needs to also be DEV_BSIZE aligned.  Note that this
3167 	 * fails with NFS if the server or some other client
3168 	 * extends the file's EOF.  If our buffer is resized,
3169 	 * B_CACHE may remain set! XXX
3170 	 */
3171 	toff = bp->b_bcount;
3172 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3173 	while ((bp->b_flags & B_CACHE) && toff < size) {
3174 		vm_pindex_t pi;
3175 
3176 		if (tinc > (size - toff))
3177 			tinc = size - toff;
3178 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3179 		m = bp->b_pages[pi];
3180 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3181 		toff += tinc;
3182 		tinc = PAGE_SIZE;
3183 	}
3184 
3185 	/*
3186 	 * Step 3, fixup the KVA pmap.
3187 	 */
3188 	if (buf_mapped(bp))
3189 		bpmap_qenter(bp);
3190 	else
3191 		BUF_CHECK_UNMAPPED(bp);
3192 }
3193 
3194 /*
3195  * Check to see if a block at a particular lbn is available for a clustered
3196  * write.
3197  */
3198 static int
3199 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3200 {
3201 	struct buf *bpa;
3202 	int match;
3203 
3204 	match = 0;
3205 
3206 	/* If the buf isn't in core skip it */
3207 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3208 		return (0);
3209 
3210 	/* If the buf is busy we don't want to wait for it */
3211 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3212 		return (0);
3213 
3214 	/* Only cluster with valid clusterable delayed write buffers */
3215 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3216 	    (B_DELWRI | B_CLUSTEROK))
3217 		goto done;
3218 
3219 	if (bpa->b_bufsize != size)
3220 		goto done;
3221 
3222 	/*
3223 	 * Check to see if it is in the expected place on disk and that the
3224 	 * block has been mapped.
3225 	 */
3226 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3227 		match = 1;
3228 done:
3229 	BUF_UNLOCK(bpa);
3230 	return (match);
3231 }
3232 
3233 /*
3234  *	vfs_bio_awrite:
3235  *
3236  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3237  *	This is much better then the old way of writing only one buffer at
3238  *	a time.  Note that we may not be presented with the buffers in the
3239  *	correct order, so we search for the cluster in both directions.
3240  */
3241 int
3242 vfs_bio_awrite(struct buf *bp)
3243 {
3244 	struct bufobj *bo;
3245 	int i;
3246 	int j;
3247 	daddr_t lblkno = bp->b_lblkno;
3248 	struct vnode *vp = bp->b_vp;
3249 	int ncl;
3250 	int nwritten;
3251 	int size;
3252 	int maxcl;
3253 	int gbflags;
3254 
3255 	bo = &vp->v_bufobj;
3256 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3257 	/*
3258 	 * right now we support clustered writing only to regular files.  If
3259 	 * we find a clusterable block we could be in the middle of a cluster
3260 	 * rather then at the beginning.
3261 	 */
3262 	if ((vp->v_type == VREG) &&
3263 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3264 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3265 		size = vp->v_mount->mnt_stat.f_iosize;
3266 		maxcl = maxphys / size;
3267 
3268 		BO_RLOCK(bo);
3269 		for (i = 1; i < maxcl; i++)
3270 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3271 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3272 				break;
3273 
3274 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3275 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3276 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3277 				break;
3278 		BO_RUNLOCK(bo);
3279 		--j;
3280 		ncl = i + j;
3281 		/*
3282 		 * this is a possible cluster write
3283 		 */
3284 		if (ncl != 1) {
3285 			BUF_UNLOCK(bp);
3286 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3287 			    gbflags);
3288 			return (nwritten);
3289 		}
3290 	}
3291 	bremfree(bp);
3292 	bp->b_flags |= B_ASYNC;
3293 	/*
3294 	 * default (old) behavior, writing out only one block
3295 	 *
3296 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3297 	 */
3298 	nwritten = bp->b_bufsize;
3299 	(void) bwrite(bp);
3300 
3301 	return (nwritten);
3302 }
3303 
3304 /*
3305  *	getnewbuf_kva:
3306  *
3307  *	Allocate KVA for an empty buf header according to gbflags.
3308  */
3309 static int
3310 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3311 {
3312 
3313 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3314 		/*
3315 		 * In order to keep fragmentation sane we only allocate kva
3316 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3317 		 */
3318 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3319 
3320 		if (maxsize != bp->b_kvasize &&
3321 		    bufkva_alloc(bp, maxsize, gbflags))
3322 			return (ENOSPC);
3323 	}
3324 	return (0);
3325 }
3326 
3327 /*
3328  *	getnewbuf:
3329  *
3330  *	Find and initialize a new buffer header, freeing up existing buffers
3331  *	in the bufqueues as necessary.  The new buffer is returned locked.
3332  *
3333  *	We block if:
3334  *		We have insufficient buffer headers
3335  *		We have insufficient buffer space
3336  *		buffer_arena is too fragmented ( space reservation fails )
3337  *		If we have to flush dirty buffers ( but we try to avoid this )
3338  *
3339  *	The caller is responsible for releasing the reserved bufspace after
3340  *	allocbuf() is called.
3341  */
3342 static struct buf *
3343 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3344 {
3345 	struct bufdomain *bd;
3346 	struct buf *bp;
3347 	bool metadata, reserved;
3348 
3349 	bp = NULL;
3350 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3351 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3352 	if (!unmapped_buf_allowed)
3353 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3354 
3355 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3356 	    vp->v_type == VCHR)
3357 		metadata = true;
3358 	else
3359 		metadata = false;
3360 	if (vp == NULL)
3361 		bd = &bdomain[0];
3362 	else
3363 		bd = &bdomain[vp->v_bufobj.bo_domain];
3364 
3365 	counter_u64_add(getnewbufcalls, 1);
3366 	reserved = false;
3367 	do {
3368 		if (reserved == false &&
3369 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3370 			counter_u64_add(getnewbufrestarts, 1);
3371 			continue;
3372 		}
3373 		reserved = true;
3374 		if ((bp = buf_alloc(bd)) == NULL) {
3375 			counter_u64_add(getnewbufrestarts, 1);
3376 			continue;
3377 		}
3378 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3379 			return (bp);
3380 		break;
3381 	} while (buf_recycle(bd, false) == 0);
3382 
3383 	if (reserved)
3384 		bufspace_release(bd, maxsize);
3385 	if (bp != NULL) {
3386 		bp->b_flags |= B_INVAL;
3387 		brelse(bp);
3388 	}
3389 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3390 
3391 	return (NULL);
3392 }
3393 
3394 /*
3395  *	buf_daemon:
3396  *
3397  *	buffer flushing daemon.  Buffers are normally flushed by the
3398  *	update daemon but if it cannot keep up this process starts to
3399  *	take the load in an attempt to prevent getnewbuf() from blocking.
3400  */
3401 static struct kproc_desc buf_kp = {
3402 	"bufdaemon",
3403 	buf_daemon,
3404 	&bufdaemonproc
3405 };
3406 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3407 
3408 static int
3409 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3410 {
3411 	int flushed;
3412 
3413 	flushed = flushbufqueues(vp, bd, target, 0);
3414 	if (flushed == 0) {
3415 		/*
3416 		 * Could not find any buffers without rollback
3417 		 * dependencies, so just write the first one
3418 		 * in the hopes of eventually making progress.
3419 		 */
3420 		if (vp != NULL && target > 2)
3421 			target /= 2;
3422 		flushbufqueues(vp, bd, target, 1);
3423 	}
3424 	return (flushed);
3425 }
3426 
3427 static void
3428 buf_daemon_shutdown(void *arg __unused, int howto __unused)
3429 {
3430 	int error;
3431 
3432 	if (KERNEL_PANICKED())
3433 		return;
3434 
3435 	mtx_lock(&bdlock);
3436 	bd_shutdown = true;
3437 	wakeup(&bd_request);
3438 	error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown",
3439 	    60 * hz);
3440 	mtx_unlock(&bdlock);
3441 	if (error != 0)
3442 		printf("bufdaemon wait error: %d\n", error);
3443 }
3444 
3445 static void
3446 buf_daemon(void)
3447 {
3448 	struct bufdomain *bd;
3449 	int speedupreq;
3450 	int lodirty;
3451 	int i;
3452 
3453 	/*
3454 	 * This process needs to be suspended prior to shutdown sync.
3455 	 */
3456 	EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL,
3457 	    SHUTDOWN_PRI_LAST + 100);
3458 
3459 	/*
3460 	 * Start the buf clean daemons as children threads.
3461 	 */
3462 	for (i = 0 ; i < buf_domains; i++) {
3463 		int error;
3464 
3465 		error = kthread_add((void (*)(void *))bufspace_daemon,
3466 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3467 		if (error)
3468 			panic("error %d spawning bufspace daemon", error);
3469 	}
3470 
3471 	/*
3472 	 * This process is allowed to take the buffer cache to the limit
3473 	 */
3474 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3475 	mtx_lock(&bdlock);
3476 	while (!bd_shutdown) {
3477 		bd_request = 0;
3478 		mtx_unlock(&bdlock);
3479 
3480 		/*
3481 		 * Save speedupreq for this pass and reset to capture new
3482 		 * requests.
3483 		 */
3484 		speedupreq = bd_speedupreq;
3485 		bd_speedupreq = 0;
3486 
3487 		/*
3488 		 * Flush each domain sequentially according to its level and
3489 		 * the speedup request.
3490 		 */
3491 		for (i = 0; i < buf_domains; i++) {
3492 			bd = &bdomain[i];
3493 			if (speedupreq)
3494 				lodirty = bd->bd_numdirtybuffers / 2;
3495 			else
3496 				lodirty = bd->bd_lodirtybuffers;
3497 			while (bd->bd_numdirtybuffers > lodirty) {
3498 				if (buf_flush(NULL, bd,
3499 				    bd->bd_numdirtybuffers - lodirty) == 0)
3500 					break;
3501 				kern_yield(PRI_USER);
3502 			}
3503 		}
3504 
3505 		/*
3506 		 * Only clear bd_request if we have reached our low water
3507 		 * mark.  The buf_daemon normally waits 1 second and
3508 		 * then incrementally flushes any dirty buffers that have
3509 		 * built up, within reason.
3510 		 *
3511 		 * If we were unable to hit our low water mark and couldn't
3512 		 * find any flushable buffers, we sleep for a short period
3513 		 * to avoid endless loops on unlockable buffers.
3514 		 */
3515 		mtx_lock(&bdlock);
3516 		if (bd_shutdown)
3517 			break;
3518 		if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3519 			/*
3520 			 * We reached our low water mark, reset the
3521 			 * request and sleep until we are needed again.
3522 			 * The sleep is just so the suspend code works.
3523 			 */
3524 			bd_request = 0;
3525 			/*
3526 			 * Do an extra wakeup in case dirty threshold
3527 			 * changed via sysctl and the explicit transition
3528 			 * out of shortfall was missed.
3529 			 */
3530 			bdirtywakeup();
3531 			if (runningbufspace <= lorunningspace)
3532 				runningwakeup();
3533 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3534 		} else {
3535 			/*
3536 			 * We couldn't find any flushable dirty buffers but
3537 			 * still have too many dirty buffers, we
3538 			 * have to sleep and try again.  (rare)
3539 			 */
3540 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3541 		}
3542 	}
3543 	wakeup(&bd_shutdown);
3544 	mtx_unlock(&bdlock);
3545 	kthread_exit();
3546 }
3547 
3548 /*
3549  *	flushbufqueues:
3550  *
3551  *	Try to flush a buffer in the dirty queue.  We must be careful to
3552  *	free up B_INVAL buffers instead of write them, which NFS is
3553  *	particularly sensitive to.
3554  */
3555 static int flushwithdeps = 0;
3556 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3557     &flushwithdeps, 0,
3558     "Number of buffers flushed with dependencies that require rollbacks");
3559 
3560 static int
3561 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3562     int flushdeps)
3563 {
3564 	struct bufqueue *bq;
3565 	struct buf *sentinel;
3566 	struct vnode *vp;
3567 	struct mount *mp;
3568 	struct buf *bp;
3569 	int hasdeps;
3570 	int flushed;
3571 	int error;
3572 	bool unlock;
3573 
3574 	flushed = 0;
3575 	bq = &bd->bd_dirtyq;
3576 	bp = NULL;
3577 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3578 	sentinel->b_qindex = QUEUE_SENTINEL;
3579 	BQ_LOCK(bq);
3580 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3581 	BQ_UNLOCK(bq);
3582 	while (flushed != target) {
3583 		maybe_yield();
3584 		BQ_LOCK(bq);
3585 		bp = TAILQ_NEXT(sentinel, b_freelist);
3586 		if (bp != NULL) {
3587 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3588 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3589 			    b_freelist);
3590 		} else {
3591 			BQ_UNLOCK(bq);
3592 			break;
3593 		}
3594 		/*
3595 		 * Skip sentinels inserted by other invocations of the
3596 		 * flushbufqueues(), taking care to not reorder them.
3597 		 *
3598 		 * Only flush the buffers that belong to the
3599 		 * vnode locked by the curthread.
3600 		 */
3601 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3602 		    bp->b_vp != lvp)) {
3603 			BQ_UNLOCK(bq);
3604 			continue;
3605 		}
3606 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3607 		BQ_UNLOCK(bq);
3608 		if (error != 0)
3609 			continue;
3610 
3611 		/*
3612 		 * BKGRDINPROG can only be set with the buf and bufobj
3613 		 * locks both held.  We tolerate a race to clear it here.
3614 		 */
3615 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3616 		    (bp->b_flags & B_DELWRI) == 0) {
3617 			BUF_UNLOCK(bp);
3618 			continue;
3619 		}
3620 		if (bp->b_flags & B_INVAL) {
3621 			bremfreef(bp);
3622 			brelse(bp);
3623 			flushed++;
3624 			continue;
3625 		}
3626 
3627 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3628 			if (flushdeps == 0) {
3629 				BUF_UNLOCK(bp);
3630 				continue;
3631 			}
3632 			hasdeps = 1;
3633 		} else
3634 			hasdeps = 0;
3635 		/*
3636 		 * We must hold the lock on a vnode before writing
3637 		 * one of its buffers. Otherwise we may confuse, or
3638 		 * in the case of a snapshot vnode, deadlock the
3639 		 * system.
3640 		 *
3641 		 * The lock order here is the reverse of the normal
3642 		 * of vnode followed by buf lock.  This is ok because
3643 		 * the NOWAIT will prevent deadlock.
3644 		 */
3645 		vp = bp->b_vp;
3646 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3647 			BUF_UNLOCK(bp);
3648 			continue;
3649 		}
3650 		if (lvp == NULL) {
3651 			unlock = true;
3652 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3653 		} else {
3654 			ASSERT_VOP_LOCKED(vp, "getbuf");
3655 			unlock = false;
3656 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3657 			    vn_lock(vp, LK_TRYUPGRADE);
3658 		}
3659 		if (error == 0) {
3660 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3661 			    bp, bp->b_vp, bp->b_flags);
3662 			if (curproc == bufdaemonproc) {
3663 				vfs_bio_awrite(bp);
3664 			} else {
3665 				bremfree(bp);
3666 				bwrite(bp);
3667 				counter_u64_add(notbufdflushes, 1);
3668 			}
3669 			vn_finished_write(mp);
3670 			if (unlock)
3671 				VOP_UNLOCK(vp);
3672 			flushwithdeps += hasdeps;
3673 			flushed++;
3674 
3675 			/*
3676 			 * Sleeping on runningbufspace while holding
3677 			 * vnode lock leads to deadlock.
3678 			 */
3679 			if (curproc == bufdaemonproc &&
3680 			    runningbufspace > hirunningspace)
3681 				waitrunningbufspace();
3682 			continue;
3683 		}
3684 		vn_finished_write(mp);
3685 		BUF_UNLOCK(bp);
3686 	}
3687 	BQ_LOCK(bq);
3688 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3689 	BQ_UNLOCK(bq);
3690 	free(sentinel, M_TEMP);
3691 	return (flushed);
3692 }
3693 
3694 /*
3695  * Check to see if a block is currently memory resident.
3696  */
3697 struct buf *
3698 incore(struct bufobj *bo, daddr_t blkno)
3699 {
3700 	return (gbincore_unlocked(bo, blkno));
3701 }
3702 
3703 /*
3704  * Returns true if no I/O is needed to access the
3705  * associated VM object.  This is like incore except
3706  * it also hunts around in the VM system for the data.
3707  */
3708 bool
3709 inmem(struct vnode * vp, daddr_t blkno)
3710 {
3711 	vm_object_t obj;
3712 	vm_offset_t toff, tinc, size;
3713 	vm_page_t m, n;
3714 	vm_ooffset_t off;
3715 	int valid;
3716 
3717 	ASSERT_VOP_LOCKED(vp, "inmem");
3718 
3719 	if (incore(&vp->v_bufobj, blkno))
3720 		return (true);
3721 	if (vp->v_mount == NULL)
3722 		return (false);
3723 	obj = vp->v_object;
3724 	if (obj == NULL)
3725 		return (false);
3726 
3727 	size = PAGE_SIZE;
3728 	if (size > vp->v_mount->mnt_stat.f_iosize)
3729 		size = vp->v_mount->mnt_stat.f_iosize;
3730 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3731 
3732 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3733 		m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3734 recheck:
3735 		if (m == NULL)
3736 			return (false);
3737 
3738 		tinc = size;
3739 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3740 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3741 		/*
3742 		 * Consider page validity only if page mapping didn't change
3743 		 * during the check.
3744 		 */
3745 		valid = vm_page_is_valid(m,
3746 		    (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3747 		n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3748 		if (m != n) {
3749 			m = n;
3750 			goto recheck;
3751 		}
3752 		if (!valid)
3753 			return (false);
3754 	}
3755 	return (true);
3756 }
3757 
3758 /*
3759  * Set the dirty range for a buffer based on the status of the dirty
3760  * bits in the pages comprising the buffer.  The range is limited
3761  * to the size of the buffer.
3762  *
3763  * Tell the VM system that the pages associated with this buffer
3764  * are clean.  This is used for delayed writes where the data is
3765  * going to go to disk eventually without additional VM intevention.
3766  *
3767  * Note that while we only really need to clean through to b_bcount, we
3768  * just go ahead and clean through to b_bufsize.
3769  */
3770 static void
3771 vfs_clean_pages_dirty_buf(struct buf *bp)
3772 {
3773 	vm_ooffset_t foff, noff, eoff;
3774 	vm_page_t m;
3775 	int i;
3776 
3777 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3778 		return;
3779 
3780 	foff = bp->b_offset;
3781 	KASSERT(bp->b_offset != NOOFFSET,
3782 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3783 
3784 	vfs_busy_pages_acquire(bp);
3785 	vfs_setdirty_range(bp);
3786 	for (i = 0; i < bp->b_npages; i++) {
3787 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3788 		eoff = noff;
3789 		if (eoff > bp->b_offset + bp->b_bufsize)
3790 			eoff = bp->b_offset + bp->b_bufsize;
3791 		m = bp->b_pages[i];
3792 		vfs_page_set_validclean(bp, foff, m);
3793 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3794 		foff = noff;
3795 	}
3796 	vfs_busy_pages_release(bp);
3797 }
3798 
3799 static void
3800 vfs_setdirty_range(struct buf *bp)
3801 {
3802 	vm_offset_t boffset;
3803 	vm_offset_t eoffset;
3804 	int i;
3805 
3806 	/*
3807 	 * test the pages to see if they have been modified directly
3808 	 * by users through the VM system.
3809 	 */
3810 	for (i = 0; i < bp->b_npages; i++)
3811 		vm_page_test_dirty(bp->b_pages[i]);
3812 
3813 	/*
3814 	 * Calculate the encompassing dirty range, boffset and eoffset,
3815 	 * (eoffset - boffset) bytes.
3816 	 */
3817 
3818 	for (i = 0; i < bp->b_npages; i++) {
3819 		if (bp->b_pages[i]->dirty)
3820 			break;
3821 	}
3822 	boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3823 
3824 	for (i = bp->b_npages - 1; i >= 0; --i) {
3825 		if (bp->b_pages[i]->dirty) {
3826 			break;
3827 		}
3828 	}
3829 	eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3830 
3831 	/*
3832 	 * Fit it to the buffer.
3833 	 */
3834 
3835 	if (eoffset > bp->b_bcount)
3836 		eoffset = bp->b_bcount;
3837 
3838 	/*
3839 	 * If we have a good dirty range, merge with the existing
3840 	 * dirty range.
3841 	 */
3842 
3843 	if (boffset < eoffset) {
3844 		if (bp->b_dirtyoff > boffset)
3845 			bp->b_dirtyoff = boffset;
3846 		if (bp->b_dirtyend < eoffset)
3847 			bp->b_dirtyend = eoffset;
3848 	}
3849 }
3850 
3851 /*
3852  * Allocate the KVA mapping for an existing buffer.
3853  * If an unmapped buffer is provided but a mapped buffer is requested, take
3854  * also care to properly setup mappings between pages and KVA.
3855  */
3856 static void
3857 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3858 {
3859 	int bsize, maxsize, need_mapping, need_kva;
3860 	off_t offset;
3861 
3862 	need_mapping = bp->b_data == unmapped_buf &&
3863 	    (gbflags & GB_UNMAPPED) == 0;
3864 	need_kva = bp->b_kvabase == unmapped_buf &&
3865 	    bp->b_data == unmapped_buf &&
3866 	    (gbflags & GB_KVAALLOC) != 0;
3867 	if (!need_mapping && !need_kva)
3868 		return;
3869 
3870 	BUF_CHECK_UNMAPPED(bp);
3871 
3872 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3873 		/*
3874 		 * Buffer is not mapped, but the KVA was already
3875 		 * reserved at the time of the instantiation.  Use the
3876 		 * allocated space.
3877 		 */
3878 		goto has_addr;
3879 	}
3880 
3881 	/*
3882 	 * Calculate the amount of the address space we would reserve
3883 	 * if the buffer was mapped.
3884 	 */
3885 	bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3886 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3887 	offset = blkno * bsize;
3888 	maxsize = size + (offset & PAGE_MASK);
3889 	maxsize = imax(maxsize, bsize);
3890 
3891 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3892 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3893 			/*
3894 			 * XXXKIB: defragmentation cannot
3895 			 * succeed, not sure what else to do.
3896 			 */
3897 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3898 		}
3899 		counter_u64_add(mappingrestarts, 1);
3900 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3901 	}
3902 has_addr:
3903 	if (need_mapping) {
3904 		/* b_offset is handled by bpmap_qenter. */
3905 		bp->b_data = bp->b_kvabase;
3906 		BUF_CHECK_MAPPED(bp);
3907 		bpmap_qenter(bp);
3908 	}
3909 }
3910 
3911 struct buf *
3912 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3913     int flags)
3914 {
3915 	struct buf *bp;
3916 	int error;
3917 
3918 	error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3919 	if (error != 0)
3920 		return (NULL);
3921 	return (bp);
3922 }
3923 
3924 /*
3925  *	getblkx:
3926  *
3927  *	Get a block given a specified block and offset into a file/device.
3928  *	The buffers B_DONE bit will be cleared on return, making it almost
3929  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3930  *	return.  The caller should clear B_INVAL prior to initiating a
3931  *	READ.
3932  *
3933  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3934  *	an existing buffer.
3935  *
3936  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3937  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3938  *	and then cleared based on the backing VM.  If the previous buffer is
3939  *	non-0-sized but invalid, B_CACHE will be cleared.
3940  *
3941  *	If getblk() must create a new buffer, the new buffer is returned with
3942  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3943  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3944  *	backing VM.
3945  *
3946  *	getblk() also forces a bwrite() for any B_DELWRI buffer whose
3947  *	B_CACHE bit is clear.
3948  *
3949  *	What this means, basically, is that the caller should use B_CACHE to
3950  *	determine whether the buffer is fully valid or not and should clear
3951  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3952  *	the buffer by loading its data area with something, the caller needs
3953  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3954  *	the caller should set B_CACHE ( as an optimization ), else the caller
3955  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3956  *	a write attempt or if it was a successful read.  If the caller
3957  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3958  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3959  *
3960  *	The blkno parameter is the logical block being requested. Normally
3961  *	the mapping of logical block number to disk block address is done
3962  *	by calling VOP_BMAP(). However, if the mapping is already known, the
3963  *	disk block address can be passed using the dblkno parameter. If the
3964  *	disk block address is not known, then the same value should be passed
3965  *	for blkno and dblkno.
3966  */
3967 int
3968 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3969     int slptimeo, int flags, struct buf **bpp)
3970 {
3971 	struct buf *bp;
3972 	struct bufobj *bo;
3973 	daddr_t d_blkno;
3974 	int bsize, error, maxsize, vmio;
3975 	off_t offset;
3976 
3977 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3978 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3979 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3980 	if (vp->v_type != VCHR)
3981 		ASSERT_VOP_LOCKED(vp, "getblk");
3982 	if (size > maxbcachebuf)
3983 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3984 		    maxbcachebuf);
3985 	if (!unmapped_buf_allowed)
3986 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3987 
3988 	bo = &vp->v_bufobj;
3989 	d_blkno = dblkno;
3990 
3991 	/* Attempt lockless lookup first. */
3992 	bp = gbincore_unlocked(bo, blkno);
3993 	if (bp == NULL) {
3994 		/*
3995 		 * With GB_NOCREAT we must be sure about not finding the buffer
3996 		 * as it may have been reassigned during unlocked lookup.
3997 		 */
3998 		if ((flags & GB_NOCREAT) != 0)
3999 			goto loop;
4000 		goto newbuf_unlocked;
4001 	}
4002 
4003 	error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
4004 	    0);
4005 	if (error != 0)
4006 		goto loop;
4007 
4008 	/* Verify buf identify has not changed since lookup. */
4009 	if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
4010 		goto foundbuf_fastpath;
4011 
4012 	/* It changed, fallback to locked lookup. */
4013 	BUF_UNLOCK_RAW(bp);
4014 
4015 loop:
4016 	BO_RLOCK(bo);
4017 	bp = gbincore(bo, blkno);
4018 	if (bp != NULL) {
4019 		int lockflags;
4020 
4021 		/*
4022 		 * Buffer is in-core.  If the buffer is not busy nor managed,
4023 		 * it must be on a queue.
4024 		 */
4025 		lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
4026 		    ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL);
4027 #ifdef WITNESS
4028 		lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0;
4029 #endif
4030 
4031 		error = BUF_TIMELOCK(bp, lockflags,
4032 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
4033 
4034 		/*
4035 		 * If we slept and got the lock we have to restart in case
4036 		 * the buffer changed identities.
4037 		 */
4038 		if (error == ENOLCK)
4039 			goto loop;
4040 		/* We timed out or were interrupted. */
4041 		else if (error != 0)
4042 			return (error);
4043 
4044 foundbuf_fastpath:
4045 		/* If recursed, assume caller knows the rules. */
4046 		if (BUF_LOCKRECURSED(bp))
4047 			goto end;
4048 
4049 		/*
4050 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
4051 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
4052 		 * and for a VMIO buffer B_CACHE is adjusted according to the
4053 		 * backing VM cache.
4054 		 */
4055 		if (bp->b_flags & B_INVAL)
4056 			bp->b_flags &= ~B_CACHE;
4057 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4058 			bp->b_flags |= B_CACHE;
4059 		if (bp->b_flags & B_MANAGED)
4060 			MPASS(bp->b_qindex == QUEUE_NONE);
4061 		else
4062 			bremfree(bp);
4063 
4064 		/*
4065 		 * check for size inconsistencies for non-VMIO case.
4066 		 */
4067 		if (bp->b_bcount != size) {
4068 			if ((bp->b_flags & B_VMIO) == 0 ||
4069 			    (size > bp->b_kvasize)) {
4070 				if (bp->b_flags & B_DELWRI) {
4071 					bp->b_flags |= B_NOCACHE;
4072 					bwrite(bp);
4073 				} else {
4074 					if (LIST_EMPTY(&bp->b_dep)) {
4075 						bp->b_flags |= B_RELBUF;
4076 						brelse(bp);
4077 					} else {
4078 						bp->b_flags |= B_NOCACHE;
4079 						bwrite(bp);
4080 					}
4081 				}
4082 				goto loop;
4083 			}
4084 		}
4085 
4086 		/*
4087 		 * Handle the case of unmapped buffer which should
4088 		 * become mapped, or the buffer for which KVA
4089 		 * reservation is requested.
4090 		 */
4091 		bp_unmapped_get_kva(bp, blkno, size, flags);
4092 
4093 		/*
4094 		 * If the size is inconsistent in the VMIO case, we can resize
4095 		 * the buffer.  This might lead to B_CACHE getting set or
4096 		 * cleared.  If the size has not changed, B_CACHE remains
4097 		 * unchanged from its previous state.
4098 		 */
4099 		allocbuf(bp, size);
4100 
4101 		KASSERT(bp->b_offset != NOOFFSET,
4102 		    ("getblk: no buffer offset"));
4103 
4104 		/*
4105 		 * A buffer with B_DELWRI set and B_CACHE clear must
4106 		 * be committed before we can return the buffer in
4107 		 * order to prevent the caller from issuing a read
4108 		 * ( due to B_CACHE not being set ) and overwriting
4109 		 * it.
4110 		 *
4111 		 * Most callers, including NFS and FFS, need this to
4112 		 * operate properly either because they assume they
4113 		 * can issue a read if B_CACHE is not set, or because
4114 		 * ( for example ) an uncached B_DELWRI might loop due
4115 		 * to softupdates re-dirtying the buffer.  In the latter
4116 		 * case, B_CACHE is set after the first write completes,
4117 		 * preventing further loops.
4118 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
4119 		 * above while extending the buffer, we cannot allow the
4120 		 * buffer to remain with B_CACHE set after the write
4121 		 * completes or it will represent a corrupt state.  To
4122 		 * deal with this we set B_NOCACHE to scrap the buffer
4123 		 * after the write.
4124 		 *
4125 		 * We might be able to do something fancy, like setting
4126 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
4127 		 * so the below call doesn't set B_CACHE, but that gets real
4128 		 * confusing.  This is much easier.
4129 		 */
4130 
4131 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4132 			bp->b_flags |= B_NOCACHE;
4133 			bwrite(bp);
4134 			goto loop;
4135 		}
4136 		bp->b_flags &= ~B_DONE;
4137 	} else {
4138 		/*
4139 		 * Buffer is not in-core, create new buffer.  The buffer
4140 		 * returned by getnewbuf() is locked.  Note that the returned
4141 		 * buffer is also considered valid (not marked B_INVAL).
4142 		 */
4143 		BO_RUNLOCK(bo);
4144 newbuf_unlocked:
4145 		/*
4146 		 * If the user does not want us to create the buffer, bail out
4147 		 * here.
4148 		 */
4149 		if (flags & GB_NOCREAT)
4150 			return (EEXIST);
4151 
4152 		bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4153 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4154 		offset = blkno * bsize;
4155 		vmio = vp->v_object != NULL;
4156 		if (vmio) {
4157 			maxsize = size + (offset & PAGE_MASK);
4158 		} else {
4159 			maxsize = size;
4160 			/* Do not allow non-VMIO notmapped buffers. */
4161 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4162 		}
4163 		maxsize = imax(maxsize, bsize);
4164 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4165 		    !vn_isdisk(vp)) {
4166 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4167 			KASSERT(error != EOPNOTSUPP,
4168 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4169 			    vp));
4170 			if (error != 0)
4171 				return (error);
4172 			if (d_blkno == -1)
4173 				return (EJUSTRETURN);
4174 		}
4175 
4176 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4177 		if (bp == NULL) {
4178 			if (slpflag || slptimeo)
4179 				return (ETIMEDOUT);
4180 			/*
4181 			 * XXX This is here until the sleep path is diagnosed
4182 			 * enough to work under very low memory conditions.
4183 			 *
4184 			 * There's an issue on low memory, 4BSD+non-preempt
4185 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4186 			 * exhaustion occurs without sleeping for buffer
4187 			 * reclaimation.  This just sticks in a loop and
4188 			 * constantly attempts to allocate a buffer, which
4189 			 * hits exhaustion and tries to wakeup bufdaemon.
4190 			 * This never happens because we never yield.
4191 			 *
4192 			 * The real solution is to identify and fix these cases
4193 			 * so we aren't effectively busy-waiting in a loop
4194 			 * until the reclaimation path has cycles to run.
4195 			 */
4196 			kern_yield(PRI_USER);
4197 			goto loop;
4198 		}
4199 
4200 		/*
4201 		 * This code is used to make sure that a buffer is not
4202 		 * created while the getnewbuf routine is blocked.
4203 		 * This can be a problem whether the vnode is locked or not.
4204 		 * If the buffer is created out from under us, we have to
4205 		 * throw away the one we just created.
4206 		 *
4207 		 * Note: this must occur before we associate the buffer
4208 		 * with the vp especially considering limitations in
4209 		 * the splay tree implementation when dealing with duplicate
4210 		 * lblkno's.
4211 		 */
4212 		BO_LOCK(bo);
4213 		if (gbincore(bo, blkno)) {
4214 			BO_UNLOCK(bo);
4215 			bp->b_flags |= B_INVAL;
4216 			bufspace_release(bufdomain(bp), maxsize);
4217 			brelse(bp);
4218 			goto loop;
4219 		}
4220 
4221 		/*
4222 		 * Insert the buffer into the hash, so that it can
4223 		 * be found by incore.
4224 		 */
4225 		bp->b_lblkno = blkno;
4226 		bp->b_blkno = d_blkno;
4227 		bp->b_offset = offset;
4228 		bgetvp(vp, bp);
4229 		BO_UNLOCK(bo);
4230 
4231 		/*
4232 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4233 		 * buffer size starts out as 0, B_CACHE will be set by
4234 		 * allocbuf() for the VMIO case prior to it testing the
4235 		 * backing store for validity.
4236 		 */
4237 
4238 		if (vmio) {
4239 			bp->b_flags |= B_VMIO;
4240 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4241 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4242 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4243 		} else {
4244 			bp->b_flags &= ~B_VMIO;
4245 			KASSERT(bp->b_bufobj->bo_object == NULL,
4246 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4247 			    bp, bp->b_bufobj->bo_object));
4248 			BUF_CHECK_MAPPED(bp);
4249 		}
4250 
4251 		allocbuf(bp, size);
4252 		bufspace_release(bufdomain(bp), maxsize);
4253 		bp->b_flags &= ~B_DONE;
4254 	}
4255 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4256 end:
4257 	buf_track(bp, __func__);
4258 	KASSERT(bp->b_bufobj == bo,
4259 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4260 	*bpp = bp;
4261 	return (0);
4262 }
4263 
4264 /*
4265  * Get an empty, disassociated buffer of given size.  The buffer is initially
4266  * set to B_INVAL.
4267  */
4268 struct buf *
4269 geteblk(int size, int flags)
4270 {
4271 	struct buf *bp;
4272 	int maxsize;
4273 
4274 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4275 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4276 		if ((flags & GB_NOWAIT_BD) &&
4277 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4278 			return (NULL);
4279 	}
4280 	allocbuf(bp, size);
4281 	bufspace_release(bufdomain(bp), maxsize);
4282 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4283 	return (bp);
4284 }
4285 
4286 /*
4287  * Truncate the backing store for a non-vmio buffer.
4288  */
4289 static void
4290 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4291 {
4292 
4293 	if (bp->b_flags & B_MALLOC) {
4294 		/*
4295 		 * malloced buffers are not shrunk
4296 		 */
4297 		if (newbsize == 0) {
4298 			bufmallocadjust(bp, 0);
4299 			free(bp->b_data, M_BIOBUF);
4300 			bp->b_data = bp->b_kvabase;
4301 			bp->b_flags &= ~B_MALLOC;
4302 		}
4303 		return;
4304 	}
4305 	vm_hold_free_pages(bp, newbsize);
4306 	bufspace_adjust(bp, newbsize);
4307 }
4308 
4309 /*
4310  * Extend the backing for a non-VMIO buffer.
4311  */
4312 static void
4313 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4314 {
4315 	caddr_t origbuf;
4316 	int origbufsize;
4317 
4318 	/*
4319 	 * We only use malloced memory on the first allocation.
4320 	 * and revert to page-allocated memory when the buffer
4321 	 * grows.
4322 	 *
4323 	 * There is a potential smp race here that could lead
4324 	 * to bufmallocspace slightly passing the max.  It
4325 	 * is probably extremely rare and not worth worrying
4326 	 * over.
4327 	 */
4328 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4329 	    bufmallocspace < maxbufmallocspace) {
4330 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4331 		bp->b_flags |= B_MALLOC;
4332 		bufmallocadjust(bp, newbsize);
4333 		return;
4334 	}
4335 
4336 	/*
4337 	 * If the buffer is growing on its other-than-first
4338 	 * allocation then we revert to the page-allocation
4339 	 * scheme.
4340 	 */
4341 	origbuf = NULL;
4342 	origbufsize = 0;
4343 	if (bp->b_flags & B_MALLOC) {
4344 		origbuf = bp->b_data;
4345 		origbufsize = bp->b_bufsize;
4346 		bp->b_data = bp->b_kvabase;
4347 		bufmallocadjust(bp, 0);
4348 		bp->b_flags &= ~B_MALLOC;
4349 		newbsize = round_page(newbsize);
4350 	}
4351 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4352 	    (vm_offset_t) bp->b_data + newbsize);
4353 	if (origbuf != NULL) {
4354 		bcopy(origbuf, bp->b_data, origbufsize);
4355 		free(origbuf, M_BIOBUF);
4356 	}
4357 	bufspace_adjust(bp, newbsize);
4358 }
4359 
4360 /*
4361  * This code constitutes the buffer memory from either anonymous system
4362  * memory (in the case of non-VMIO operations) or from an associated
4363  * VM object (in the case of VMIO operations).  This code is able to
4364  * resize a buffer up or down.
4365  *
4366  * Note that this code is tricky, and has many complications to resolve
4367  * deadlock or inconsistent data situations.  Tread lightly!!!
4368  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4369  * the caller.  Calling this code willy nilly can result in the loss of data.
4370  *
4371  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4372  * B_CACHE for the non-VMIO case.
4373  */
4374 int
4375 allocbuf(struct buf *bp, int size)
4376 {
4377 	int newbsize;
4378 
4379 	if (bp->b_bcount == size)
4380 		return (1);
4381 
4382 	KASSERT(bp->b_kvasize == 0 || bp->b_kvasize >= size,
4383 	    ("allocbuf: buffer too small %p %#x %#x",
4384 	    bp, bp->b_kvasize, size));
4385 
4386 	newbsize = roundup2(size, DEV_BSIZE);
4387 	if ((bp->b_flags & B_VMIO) == 0) {
4388 		if ((bp->b_flags & B_MALLOC) == 0)
4389 			newbsize = round_page(newbsize);
4390 		/*
4391 		 * Just get anonymous memory from the kernel.  Don't
4392 		 * mess with B_CACHE.
4393 		 */
4394 		if (newbsize < bp->b_bufsize)
4395 			vfs_nonvmio_truncate(bp, newbsize);
4396 		else if (newbsize > bp->b_bufsize)
4397 			vfs_nonvmio_extend(bp, newbsize);
4398 	} else {
4399 		int desiredpages;
4400 
4401 		desiredpages = size == 0 ? 0 :
4402 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4403 
4404 		KASSERT((bp->b_flags & B_MALLOC) == 0,
4405 		    ("allocbuf: VMIO buffer can't be malloced %p", bp));
4406 
4407 		/*
4408 		 * Set B_CACHE initially if buffer is 0 length or will become
4409 		 * 0-length.
4410 		 */
4411 		if (size == 0 || bp->b_bufsize == 0)
4412 			bp->b_flags |= B_CACHE;
4413 
4414 		if (newbsize < bp->b_bufsize)
4415 			vfs_vmio_truncate(bp, desiredpages);
4416 		/* XXX This looks as if it should be newbsize > b_bufsize */
4417 		else if (size > bp->b_bcount)
4418 			vfs_vmio_extend(bp, desiredpages, size);
4419 		bufspace_adjust(bp, newbsize);
4420 	}
4421 	bp->b_bcount = size;		/* requested buffer size. */
4422 	return (1);
4423 }
4424 
4425 extern int inflight_transient_maps;
4426 
4427 static struct bio_queue nondump_bios;
4428 
4429 void
4430 biodone(struct bio *bp)
4431 {
4432 	struct mtx *mtxp;
4433 	void (*done)(struct bio *);
4434 	vm_offset_t start, end;
4435 
4436 	biotrack(bp, __func__);
4437 
4438 	/*
4439 	 * Avoid completing I/O when dumping after a panic since that may
4440 	 * result in a deadlock in the filesystem or pager code.  Note that
4441 	 * this doesn't affect dumps that were started manually since we aim
4442 	 * to keep the system usable after it has been resumed.
4443 	 */
4444 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4445 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4446 		return;
4447 	}
4448 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4449 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4450 		bp->bio_flags |= BIO_UNMAPPED;
4451 		start = trunc_page((vm_offset_t)bp->bio_data);
4452 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4453 		bp->bio_data = unmapped_buf;
4454 		pmap_qremove(start, atop(end - start));
4455 		vmem_free(transient_arena, start, end - start);
4456 		atomic_add_int(&inflight_transient_maps, -1);
4457 	}
4458 	done = bp->bio_done;
4459 	/*
4460 	 * The check for done == biodone is to allow biodone to be
4461 	 * used as a bio_done routine.
4462 	 */
4463 	if (done == NULL || done == biodone) {
4464 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4465 		mtx_lock(mtxp);
4466 		bp->bio_flags |= BIO_DONE;
4467 		wakeup(bp);
4468 		mtx_unlock(mtxp);
4469 	} else
4470 		done(bp);
4471 }
4472 
4473 /*
4474  * Wait for a BIO to finish.
4475  */
4476 int
4477 biowait(struct bio *bp, const char *wmesg)
4478 {
4479 	struct mtx *mtxp;
4480 
4481 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4482 	mtx_lock(mtxp);
4483 	while ((bp->bio_flags & BIO_DONE) == 0)
4484 		msleep(bp, mtxp, PRIBIO, wmesg, 0);
4485 	mtx_unlock(mtxp);
4486 	if (bp->bio_error != 0)
4487 		return (bp->bio_error);
4488 	if (!(bp->bio_flags & BIO_ERROR))
4489 		return (0);
4490 	return (EIO);
4491 }
4492 
4493 void
4494 biofinish(struct bio *bp, struct devstat *stat, int error)
4495 {
4496 
4497 	if (error) {
4498 		bp->bio_error = error;
4499 		bp->bio_flags |= BIO_ERROR;
4500 	}
4501 	if (stat != NULL)
4502 		devstat_end_transaction_bio(stat, bp);
4503 	biodone(bp);
4504 }
4505 
4506 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4507 void
4508 biotrack_buf(struct bio *bp, const char *location)
4509 {
4510 
4511 	buf_track(bp->bio_track_bp, location);
4512 }
4513 #endif
4514 
4515 /*
4516  *	bufwait:
4517  *
4518  *	Wait for buffer I/O completion, returning error status.  The buffer
4519  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4520  *	error and cleared.
4521  */
4522 int
4523 bufwait(struct buf *bp)
4524 {
4525 	if (bp->b_iocmd == BIO_READ)
4526 		bwait(bp, PRIBIO, "biord");
4527 	else
4528 		bwait(bp, PRIBIO, "biowr");
4529 	if (bp->b_flags & B_EINTR) {
4530 		bp->b_flags &= ~B_EINTR;
4531 		return (EINTR);
4532 	}
4533 	if (bp->b_ioflags & BIO_ERROR) {
4534 		return (bp->b_error ? bp->b_error : EIO);
4535 	} else {
4536 		return (0);
4537 	}
4538 }
4539 
4540 /*
4541  *	bufdone:
4542  *
4543  *	Finish I/O on a buffer, optionally calling a completion function.
4544  *	This is usually called from an interrupt so process blocking is
4545  *	not allowed.
4546  *
4547  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4548  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4549  *	assuming B_INVAL is clear.
4550  *
4551  *	For the VMIO case, we set B_CACHE if the op was a read and no
4552  *	read error occurred, or if the op was a write.  B_CACHE is never
4553  *	set if the buffer is invalid or otherwise uncacheable.
4554  *
4555  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4556  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4557  *	in the biodone routine.
4558  */
4559 void
4560 bufdone(struct buf *bp)
4561 {
4562 	struct bufobj *dropobj;
4563 	void    (*biodone)(struct buf *);
4564 
4565 	buf_track(bp, __func__);
4566 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4567 	dropobj = NULL;
4568 
4569 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4570 
4571 	runningbufwakeup(bp);
4572 	if (bp->b_iocmd == BIO_WRITE)
4573 		dropobj = bp->b_bufobj;
4574 	/* call optional completion function if requested */
4575 	if (bp->b_iodone != NULL) {
4576 		biodone = bp->b_iodone;
4577 		bp->b_iodone = NULL;
4578 		(*biodone) (bp);
4579 		if (dropobj)
4580 			bufobj_wdrop(dropobj);
4581 		return;
4582 	}
4583 	if (bp->b_flags & B_VMIO) {
4584 		/*
4585 		 * Set B_CACHE if the op was a normal read and no error
4586 		 * occurred.  B_CACHE is set for writes in the b*write()
4587 		 * routines.
4588 		 */
4589 		if (bp->b_iocmd == BIO_READ &&
4590 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4591 		    !(bp->b_ioflags & BIO_ERROR))
4592 			bp->b_flags |= B_CACHE;
4593 		vfs_vmio_iodone(bp);
4594 	}
4595 	if (!LIST_EMPTY(&bp->b_dep))
4596 		buf_complete(bp);
4597 	if ((bp->b_flags & B_CKHASH) != 0) {
4598 		KASSERT(bp->b_iocmd == BIO_READ,
4599 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4600 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4601 		(*bp->b_ckhashcalc)(bp);
4602 	}
4603 	/*
4604 	 * For asynchronous completions, release the buffer now. The brelse
4605 	 * will do a wakeup there if necessary - so no need to do a wakeup
4606 	 * here in the async case. The sync case always needs to do a wakeup.
4607 	 */
4608 	if (bp->b_flags & B_ASYNC) {
4609 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4610 		    (bp->b_ioflags & BIO_ERROR))
4611 			brelse(bp);
4612 		else
4613 			bqrelse(bp);
4614 	} else
4615 		bdone(bp);
4616 	if (dropobj)
4617 		bufobj_wdrop(dropobj);
4618 }
4619 
4620 /*
4621  * This routine is called in lieu of iodone in the case of
4622  * incomplete I/O.  This keeps the busy status for pages
4623  * consistent.
4624  */
4625 void
4626 vfs_unbusy_pages(struct buf *bp)
4627 {
4628 	int i;
4629 	vm_object_t obj;
4630 	vm_page_t m;
4631 
4632 	runningbufwakeup(bp);
4633 	if (!(bp->b_flags & B_VMIO))
4634 		return;
4635 
4636 	obj = bp->b_bufobj->bo_object;
4637 	for (i = 0; i < bp->b_npages; i++) {
4638 		m = bp->b_pages[i];
4639 		if (m == bogus_page) {
4640 			m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4641 			if (!m)
4642 				panic("vfs_unbusy_pages: page missing\n");
4643 			bp->b_pages[i] = m;
4644 			if (buf_mapped(bp)) {
4645 				BUF_CHECK_MAPPED(bp);
4646 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4647 				    bp->b_pages, bp->b_npages);
4648 			} else
4649 				BUF_CHECK_UNMAPPED(bp);
4650 		}
4651 		vm_page_sunbusy(m);
4652 	}
4653 	vm_object_pip_wakeupn(obj, bp->b_npages);
4654 }
4655 
4656 /*
4657  * vfs_page_set_valid:
4658  *
4659  *	Set the valid bits in a page based on the supplied offset.   The
4660  *	range is restricted to the buffer's size.
4661  *
4662  *	This routine is typically called after a read completes.
4663  */
4664 static void
4665 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4666 {
4667 	vm_ooffset_t eoff;
4668 
4669 	/*
4670 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4671 	 * page boundary and eoff is not greater than the end of the buffer.
4672 	 * The end of the buffer, in this case, is our file EOF, not the
4673 	 * allocation size of the buffer.
4674 	 */
4675 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4676 	if (eoff > bp->b_offset + bp->b_bcount)
4677 		eoff = bp->b_offset + bp->b_bcount;
4678 
4679 	/*
4680 	 * Set valid range.  This is typically the entire buffer and thus the
4681 	 * entire page.
4682 	 */
4683 	if (eoff > off)
4684 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4685 }
4686 
4687 /*
4688  * vfs_page_set_validclean:
4689  *
4690  *	Set the valid bits and clear the dirty bits in a page based on the
4691  *	supplied offset.   The range is restricted to the buffer's size.
4692  */
4693 static void
4694 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4695 {
4696 	vm_ooffset_t soff, eoff;
4697 
4698 	/*
4699 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4700 	 * page boundary or cross the end of the buffer.  The end of the
4701 	 * buffer, in this case, is our file EOF, not the allocation size
4702 	 * of the buffer.
4703 	 */
4704 	soff = off;
4705 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4706 	if (eoff > bp->b_offset + bp->b_bcount)
4707 		eoff = bp->b_offset + bp->b_bcount;
4708 
4709 	/*
4710 	 * Set valid range.  This is typically the entire buffer and thus the
4711 	 * entire page.
4712 	 */
4713 	if (eoff > soff) {
4714 		vm_page_set_validclean(
4715 		    m,
4716 		   (vm_offset_t) (soff & PAGE_MASK),
4717 		   (vm_offset_t) (eoff - soff)
4718 		);
4719 	}
4720 }
4721 
4722 /*
4723  * Acquire a shared busy on all pages in the buf.
4724  */
4725 void
4726 vfs_busy_pages_acquire(struct buf *bp)
4727 {
4728 	int i;
4729 
4730 	for (i = 0; i < bp->b_npages; i++)
4731 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4732 }
4733 
4734 void
4735 vfs_busy_pages_release(struct buf *bp)
4736 {
4737 	int i;
4738 
4739 	for (i = 0; i < bp->b_npages; i++)
4740 		vm_page_sunbusy(bp->b_pages[i]);
4741 }
4742 
4743 /*
4744  * This routine is called before a device strategy routine.
4745  * It is used to tell the VM system that paging I/O is in
4746  * progress, and treat the pages associated with the buffer
4747  * almost as being exclusive busy.  Also the object paging_in_progress
4748  * flag is handled to make sure that the object doesn't become
4749  * inconsistent.
4750  *
4751  * Since I/O has not been initiated yet, certain buffer flags
4752  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4753  * and should be ignored.
4754  */
4755 void
4756 vfs_busy_pages(struct buf *bp, int clear_modify)
4757 {
4758 	vm_object_t obj;
4759 	vm_ooffset_t foff;
4760 	vm_page_t m;
4761 	int i;
4762 	bool bogus;
4763 
4764 	if (!(bp->b_flags & B_VMIO))
4765 		return;
4766 
4767 	obj = bp->b_bufobj->bo_object;
4768 	foff = bp->b_offset;
4769 	KASSERT(bp->b_offset != NOOFFSET,
4770 	    ("vfs_busy_pages: no buffer offset"));
4771 	if ((bp->b_flags & B_CLUSTER) == 0) {
4772 		vm_object_pip_add(obj, bp->b_npages);
4773 		vfs_busy_pages_acquire(bp);
4774 	}
4775 	if (bp->b_bufsize != 0)
4776 		vfs_setdirty_range(bp);
4777 	bogus = false;
4778 	for (i = 0; i < bp->b_npages; i++) {
4779 		m = bp->b_pages[i];
4780 		vm_page_assert_sbusied(m);
4781 
4782 		/*
4783 		 * When readying a buffer for a read ( i.e
4784 		 * clear_modify == 0 ), it is important to do
4785 		 * bogus_page replacement for valid pages in
4786 		 * partially instantiated buffers.  Partially
4787 		 * instantiated buffers can, in turn, occur when
4788 		 * reconstituting a buffer from its VM backing store
4789 		 * base.  We only have to do this if B_CACHE is
4790 		 * clear ( which causes the I/O to occur in the
4791 		 * first place ).  The replacement prevents the read
4792 		 * I/O from overwriting potentially dirty VM-backed
4793 		 * pages.  XXX bogus page replacement is, uh, bogus.
4794 		 * It may not work properly with small-block devices.
4795 		 * We need to find a better way.
4796 		 */
4797 		if (clear_modify) {
4798 			pmap_remove_write(m);
4799 			vfs_page_set_validclean(bp, foff, m);
4800 		} else if (vm_page_all_valid(m) &&
4801 		    (bp->b_flags & B_CACHE) == 0) {
4802 			bp->b_pages[i] = bogus_page;
4803 			bogus = true;
4804 		}
4805 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4806 	}
4807 	if (bogus && buf_mapped(bp)) {
4808 		BUF_CHECK_MAPPED(bp);
4809 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4810 		    bp->b_pages, bp->b_npages);
4811 	}
4812 }
4813 
4814 /*
4815  *	vfs_bio_set_valid:
4816  *
4817  *	Set the range within the buffer to valid.  The range is
4818  *	relative to the beginning of the buffer, b_offset.  Note that
4819  *	b_offset itself may be offset from the beginning of the first
4820  *	page.
4821  */
4822 void
4823 vfs_bio_set_valid(struct buf *bp, int base, int size)
4824 {
4825 	int i, n;
4826 	vm_page_t m;
4827 
4828 	if (!(bp->b_flags & B_VMIO))
4829 		return;
4830 
4831 	/*
4832 	 * Fixup base to be relative to beginning of first page.
4833 	 * Set initial n to be the maximum number of bytes in the
4834 	 * first page that can be validated.
4835 	 */
4836 	base += (bp->b_offset & PAGE_MASK);
4837 	n = PAGE_SIZE - (base & PAGE_MASK);
4838 
4839 	/*
4840 	 * Busy may not be strictly necessary here because the pages are
4841 	 * unlikely to be fully valid and the vnode lock will synchronize
4842 	 * their access via getpages.  It is grabbed for consistency with
4843 	 * other page validation.
4844 	 */
4845 	vfs_busy_pages_acquire(bp);
4846 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4847 		m = bp->b_pages[i];
4848 		if (n > size)
4849 			n = size;
4850 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4851 		base += n;
4852 		size -= n;
4853 		n = PAGE_SIZE;
4854 	}
4855 	vfs_busy_pages_release(bp);
4856 }
4857 
4858 /*
4859  *	vfs_bio_clrbuf:
4860  *
4861  *	If the specified buffer is a non-VMIO buffer, clear the entire
4862  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4863  *	validate only the previously invalid portions of the buffer.
4864  *	This routine essentially fakes an I/O, so we need to clear
4865  *	BIO_ERROR and B_INVAL.
4866  *
4867  *	Note that while we only theoretically need to clear through b_bcount,
4868  *	we go ahead and clear through b_bufsize.
4869  */
4870 void
4871 vfs_bio_clrbuf(struct buf *bp)
4872 {
4873 	int i, j, sa, ea, slide, zbits;
4874 	vm_page_bits_t mask;
4875 
4876 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4877 		clrbuf(bp);
4878 		return;
4879 	}
4880 	bp->b_flags &= ~B_INVAL;
4881 	bp->b_ioflags &= ~BIO_ERROR;
4882 	vfs_busy_pages_acquire(bp);
4883 	sa = bp->b_offset & PAGE_MASK;
4884 	slide = 0;
4885 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4886 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4887 		ea = slide & PAGE_MASK;
4888 		if (ea == 0)
4889 			ea = PAGE_SIZE;
4890 		if (bp->b_pages[i] == bogus_page)
4891 			continue;
4892 		j = sa / DEV_BSIZE;
4893 		zbits = (sizeof(vm_page_bits_t) * NBBY) -
4894 		    (ea - sa) / DEV_BSIZE;
4895 		mask = (VM_PAGE_BITS_ALL >> zbits) << j;
4896 		if ((bp->b_pages[i]->valid & mask) == mask)
4897 			continue;
4898 		if ((bp->b_pages[i]->valid & mask) == 0)
4899 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4900 		else {
4901 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4902 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4903 					pmap_zero_page_area(bp->b_pages[i],
4904 					    sa, DEV_BSIZE);
4905 				}
4906 			}
4907 		}
4908 		vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4909 		    roundup2(ea - sa, DEV_BSIZE));
4910 	}
4911 	vfs_busy_pages_release(bp);
4912 	bp->b_resid = 0;
4913 }
4914 
4915 void
4916 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4917 {
4918 	vm_page_t m;
4919 	int i, n;
4920 
4921 	if (buf_mapped(bp)) {
4922 		BUF_CHECK_MAPPED(bp);
4923 		bzero(bp->b_data + base, size);
4924 	} else {
4925 		BUF_CHECK_UNMAPPED(bp);
4926 		n = PAGE_SIZE - (base & PAGE_MASK);
4927 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4928 			m = bp->b_pages[i];
4929 			if (n > size)
4930 				n = size;
4931 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4932 			base += n;
4933 			size -= n;
4934 			n = PAGE_SIZE;
4935 		}
4936 	}
4937 }
4938 
4939 /*
4940  * Update buffer flags based on I/O request parameters, optionally releasing the
4941  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4942  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4943  * I/O).  Otherwise the buffer is released to the cache.
4944  */
4945 static void
4946 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4947 {
4948 
4949 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4950 	    ("buf %p non-VMIO noreuse", bp));
4951 
4952 	if ((ioflag & IO_DIRECT) != 0)
4953 		bp->b_flags |= B_DIRECT;
4954 	if ((ioflag & IO_EXT) != 0)
4955 		bp->b_xflags |= BX_ALTDATA;
4956 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4957 		bp->b_flags |= B_RELBUF;
4958 		if ((ioflag & IO_NOREUSE) != 0)
4959 			bp->b_flags |= B_NOREUSE;
4960 		if (release)
4961 			brelse(bp);
4962 	} else if (release)
4963 		bqrelse(bp);
4964 }
4965 
4966 void
4967 vfs_bio_brelse(struct buf *bp, int ioflag)
4968 {
4969 
4970 	b_io_dismiss(bp, ioflag, true);
4971 }
4972 
4973 void
4974 vfs_bio_set_flags(struct buf *bp, int ioflag)
4975 {
4976 
4977 	b_io_dismiss(bp, ioflag, false);
4978 }
4979 
4980 /*
4981  * vm_hold_load_pages and vm_hold_free_pages get pages into
4982  * a buffers address space.  The pages are anonymous and are
4983  * not associated with a file object.
4984  */
4985 static void
4986 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4987 {
4988 	vm_offset_t pg;
4989 	vm_page_t p;
4990 	int index;
4991 
4992 	BUF_CHECK_MAPPED(bp);
4993 
4994 	to = round_page(to);
4995 	from = round_page(from);
4996 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4997 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
4998 	KASSERT(to - from <= maxbcachebuf,
4999 	    ("vm_hold_load_pages too large %p %#jx %#jx %u",
5000 	    bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
5001 
5002 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
5003 		/*
5004 		 * note: must allocate system pages since blocking here
5005 		 * could interfere with paging I/O, no matter which
5006 		 * process we are.
5007 		 */
5008 		p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
5009 		    VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
5010 		pmap_qenter(pg, &p, 1);
5011 		bp->b_pages[index] = p;
5012 	}
5013 	bp->b_npages = index;
5014 }
5015 
5016 /* Return pages associated with this buf to the vm system */
5017 static void
5018 vm_hold_free_pages(struct buf *bp, int newbsize)
5019 {
5020 	vm_offset_t from;
5021 	vm_page_t p;
5022 	int index, newnpages;
5023 
5024 	BUF_CHECK_MAPPED(bp);
5025 
5026 	from = round_page((vm_offset_t)bp->b_data + newbsize);
5027 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
5028 	if (bp->b_npages > newnpages)
5029 		pmap_qremove(from, bp->b_npages - newnpages);
5030 	for (index = newnpages; index < bp->b_npages; index++) {
5031 		p = bp->b_pages[index];
5032 		bp->b_pages[index] = NULL;
5033 		vm_page_unwire_noq(p);
5034 		vm_page_free(p);
5035 	}
5036 	bp->b_npages = newnpages;
5037 }
5038 
5039 /*
5040  * Map an IO request into kernel virtual address space.
5041  *
5042  * All requests are (re)mapped into kernel VA space.
5043  * Notice that we use b_bufsize for the size of the buffer
5044  * to be mapped.  b_bcount might be modified by the driver.
5045  *
5046  * Note that even if the caller determines that the address space should
5047  * be valid, a race or a smaller-file mapped into a larger space may
5048  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
5049  * check the return value.
5050  *
5051  * This function only works with pager buffers.
5052  */
5053 int
5054 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
5055 {
5056 	vm_prot_t prot;
5057 	int pidx;
5058 
5059 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
5060 	prot = VM_PROT_READ;
5061 	if (bp->b_iocmd == BIO_READ)
5062 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
5063 	pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5064 	    (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5065 	if (pidx < 0)
5066 		return (-1);
5067 	bp->b_bufsize = len;
5068 	bp->b_npages = pidx;
5069 	bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5070 	if (mapbuf || !unmapped_buf_allowed) {
5071 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5072 		bp->b_data = bp->b_kvabase + bp->b_offset;
5073 	} else
5074 		bp->b_data = unmapped_buf;
5075 	return (0);
5076 }
5077 
5078 /*
5079  * Free the io map PTEs associated with this IO operation.
5080  * We also invalidate the TLB entries and restore the original b_addr.
5081  *
5082  * This function only works with pager buffers.
5083  */
5084 void
5085 vunmapbuf(struct buf *bp)
5086 {
5087 	int npages;
5088 
5089 	npages = bp->b_npages;
5090 	if (buf_mapped(bp))
5091 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5092 	vm_page_unhold_pages(bp->b_pages, npages);
5093 
5094 	bp->b_data = unmapped_buf;
5095 }
5096 
5097 void
5098 bdone(struct buf *bp)
5099 {
5100 	struct mtx *mtxp;
5101 
5102 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5103 	mtx_lock(mtxp);
5104 	bp->b_flags |= B_DONE;
5105 	wakeup(bp);
5106 	mtx_unlock(mtxp);
5107 }
5108 
5109 void
5110 bwait(struct buf *bp, u_char pri, const char *wchan)
5111 {
5112 	struct mtx *mtxp;
5113 
5114 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5115 	mtx_lock(mtxp);
5116 	while ((bp->b_flags & B_DONE) == 0)
5117 		msleep(bp, mtxp, pri, wchan, 0);
5118 	mtx_unlock(mtxp);
5119 }
5120 
5121 int
5122 bufsync(struct bufobj *bo, int waitfor)
5123 {
5124 
5125 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5126 }
5127 
5128 void
5129 bufstrategy(struct bufobj *bo, struct buf *bp)
5130 {
5131 	int i __unused;
5132 	struct vnode *vp;
5133 
5134 	vp = bp->b_vp;
5135 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5136 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5137 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5138 	i = VOP_STRATEGY(vp, bp);
5139 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5140 }
5141 
5142 /*
5143  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
5144  */
5145 void
5146 bufobj_init(struct bufobj *bo, void *private)
5147 {
5148 	static volatile int bufobj_cleanq;
5149 
5150         bo->bo_domain =
5151             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5152         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5153         bo->bo_private = private;
5154         TAILQ_INIT(&bo->bo_clean.bv_hd);
5155 	pctrie_init(&bo->bo_clean.bv_root);
5156         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5157 	pctrie_init(&bo->bo_dirty.bv_root);
5158 }
5159 
5160 void
5161 bufobj_wrefl(struct bufobj *bo)
5162 {
5163 
5164 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5165 	ASSERT_BO_WLOCKED(bo);
5166 	bo->bo_numoutput++;
5167 }
5168 
5169 void
5170 bufobj_wref(struct bufobj *bo)
5171 {
5172 
5173 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5174 	BO_LOCK(bo);
5175 	bo->bo_numoutput++;
5176 	BO_UNLOCK(bo);
5177 }
5178 
5179 void
5180 bufobj_wdrop(struct bufobj *bo)
5181 {
5182 
5183 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5184 	BO_LOCK(bo);
5185 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5186 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5187 		bo->bo_flag &= ~BO_WWAIT;
5188 		wakeup(&bo->bo_numoutput);
5189 	}
5190 	BO_UNLOCK(bo);
5191 }
5192 
5193 int
5194 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5195 {
5196 	int error;
5197 
5198 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5199 	ASSERT_BO_WLOCKED(bo);
5200 	error = 0;
5201 	while (bo->bo_numoutput) {
5202 		bo->bo_flag |= BO_WWAIT;
5203 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5204 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5205 		if (error)
5206 			break;
5207 	}
5208 	return (error);
5209 }
5210 
5211 /*
5212  * Set bio_data or bio_ma for struct bio from the struct buf.
5213  */
5214 void
5215 bdata2bio(struct buf *bp, struct bio *bip)
5216 {
5217 
5218 	if (!buf_mapped(bp)) {
5219 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5220 		bip->bio_ma = bp->b_pages;
5221 		bip->bio_ma_n = bp->b_npages;
5222 		bip->bio_data = unmapped_buf;
5223 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5224 		bip->bio_flags |= BIO_UNMAPPED;
5225 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5226 		    PAGE_SIZE == bp->b_npages,
5227 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5228 		    (long long)bip->bio_length, bip->bio_ma_n));
5229 	} else {
5230 		bip->bio_data = bp->b_data;
5231 		bip->bio_ma = NULL;
5232 	}
5233 }
5234 
5235 struct memdesc
5236 memdesc_bio(struct bio *bio)
5237 {
5238 	if ((bio->bio_flags & BIO_VLIST) != 0)
5239 		return (memdesc_vlist((struct bus_dma_segment *)bio->bio_data,
5240 		    bio->bio_ma_n));
5241 
5242 	if ((bio->bio_flags & BIO_UNMAPPED) != 0)
5243 		return (memdesc_vmpages(bio->bio_ma, bio->bio_bcount,
5244 		    bio->bio_ma_offset));
5245 
5246 	return (memdesc_vaddr(bio->bio_data, bio->bio_bcount));
5247 }
5248 
5249 static int buf_pager_relbuf;
5250 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5251     &buf_pager_relbuf, 0,
5252     "Make buffer pager release buffers after reading");
5253 
5254 /*
5255  * The buffer pager.  It uses buffer reads to validate pages.
5256  *
5257  * In contrast to the generic local pager from vm/vnode_pager.c, this
5258  * pager correctly and easily handles volumes where the underlying
5259  * device block size is greater than the machine page size.  The
5260  * buffer cache transparently extends the requested page run to be
5261  * aligned at the block boundary, and does the necessary bogus page
5262  * replacements in the addends to avoid obliterating already valid
5263  * pages.
5264  *
5265  * The only non-trivial issue is that the exclusive busy state for
5266  * pages, which is assumed by the vm_pager_getpages() interface, is
5267  * incompatible with the VMIO buffer cache's desire to share-busy the
5268  * pages.  This function performs a trivial downgrade of the pages'
5269  * state before reading buffers, and a less trivial upgrade from the
5270  * shared-busy to excl-busy state after the read.
5271  */
5272 int
5273 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5274     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5275     vbg_get_blksize_t get_blksize)
5276 {
5277 	vm_page_t m;
5278 	vm_object_t object;
5279 	struct buf *bp;
5280 	struct mount *mp;
5281 	daddr_t lbn, lbnp;
5282 	vm_ooffset_t la, lb, poff, poffe;
5283 	long bo_bs, bsize;
5284 	int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5285 	bool redo, lpart;
5286 
5287 	object = vp->v_object;
5288 	mp = vp->v_mount;
5289 	error = 0;
5290 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5291 	if (la >= object->un_pager.vnp.vnp_size)
5292 		return (VM_PAGER_BAD);
5293 
5294 	/*
5295 	 * Change the meaning of la from where the last requested page starts
5296 	 * to where it ends, because that's the end of the requested region
5297 	 * and the start of the potential read-ahead region.
5298 	 */
5299 	la += PAGE_SIZE;
5300 	lpart = la > object->un_pager.vnp.vnp_size;
5301 	error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5302 	    &bo_bs);
5303 	if (error != 0)
5304 		return (VM_PAGER_ERROR);
5305 
5306 	/*
5307 	 * Calculate read-ahead, behind and total pages.
5308 	 */
5309 	pgsin = count;
5310 	lb = IDX_TO_OFF(ma[0]->pindex);
5311 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5312 	pgsin += pgsin_b;
5313 	if (rbehind != NULL)
5314 		*rbehind = pgsin_b;
5315 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5316 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5317 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5318 		    PAGE_SIZE) - la);
5319 	pgsin += pgsin_a;
5320 	if (rahead != NULL)
5321 		*rahead = pgsin_a;
5322 	VM_CNT_INC(v_vnodein);
5323 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5324 
5325 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5326 	    != 0) ? GB_UNMAPPED : 0;
5327 again:
5328 	for (i = 0; i < count; i++) {
5329 		if (ma[i] != bogus_page)
5330 			vm_page_busy_downgrade(ma[i]);
5331 	}
5332 
5333 	lbnp = -1;
5334 	for (i = 0; i < count; i++) {
5335 		m = ma[i];
5336 		if (m == bogus_page)
5337 			continue;
5338 
5339 		/*
5340 		 * Pages are shared busy and the object lock is not
5341 		 * owned, which together allow for the pages'
5342 		 * invalidation.  The racy test for validity avoids
5343 		 * useless creation of the buffer for the most typical
5344 		 * case when invalidation is not used in redo or for
5345 		 * parallel read.  The shared->excl upgrade loop at
5346 		 * the end of the function catches the race in a
5347 		 * reliable way (protected by the object lock).
5348 		 */
5349 		if (vm_page_all_valid(m))
5350 			continue;
5351 
5352 		poff = IDX_TO_OFF(m->pindex);
5353 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5354 		for (; poff < poffe; poff += bsize) {
5355 			lbn = get_lblkno(vp, poff);
5356 			if (lbn == lbnp)
5357 				goto next_page;
5358 			lbnp = lbn;
5359 
5360 			error = get_blksize(vp, lbn, &bsize);
5361 			if (error == 0)
5362 				error = bread_gb(vp, lbn, bsize,
5363 				    curthread->td_ucred, br_flags, &bp);
5364 			if (error != 0)
5365 				goto end_pages;
5366 			if (bp->b_rcred == curthread->td_ucred) {
5367 				crfree(bp->b_rcred);
5368 				bp->b_rcred = NOCRED;
5369 			}
5370 			if (LIST_EMPTY(&bp->b_dep)) {
5371 				/*
5372 				 * Invalidation clears m->valid, but
5373 				 * may leave B_CACHE flag if the
5374 				 * buffer existed at the invalidation
5375 				 * time.  In this case, recycle the
5376 				 * buffer to do real read on next
5377 				 * bread() after redo.
5378 				 *
5379 				 * Otherwise B_RELBUF is not strictly
5380 				 * necessary, enable to reduce buf
5381 				 * cache pressure.
5382 				 */
5383 				if (buf_pager_relbuf ||
5384 				    !vm_page_all_valid(m))
5385 					bp->b_flags |= B_RELBUF;
5386 
5387 				bp->b_flags &= ~B_NOCACHE;
5388 				brelse(bp);
5389 			} else {
5390 				bqrelse(bp);
5391 			}
5392 		}
5393 		KASSERT(1 /* racy, enable for debugging */ ||
5394 		    vm_page_all_valid(m) || i == count - 1,
5395 		    ("buf %d %p invalid", i, m));
5396 		if (i == count - 1 && lpart) {
5397 			if (!vm_page_none_valid(m) &&
5398 			    !vm_page_all_valid(m))
5399 				vm_page_zero_invalid(m, TRUE);
5400 		}
5401 next_page:;
5402 	}
5403 end_pages:
5404 
5405 	redo = false;
5406 	for (i = 0; i < count; i++) {
5407 		if (ma[i] == bogus_page)
5408 			continue;
5409 		if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5410 			vm_page_sunbusy(ma[i]);
5411 			ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5412 			    VM_ALLOC_NORMAL);
5413 		}
5414 
5415 		/*
5416 		 * Since the pages were only sbusy while neither the
5417 		 * buffer nor the object lock was held by us, or
5418 		 * reallocated while vm_page_grab() slept for busy
5419 		 * relinguish, they could have been invalidated.
5420 		 * Recheck the valid bits and re-read as needed.
5421 		 *
5422 		 * Note that the last page is made fully valid in the
5423 		 * read loop, and partial validity for the page at
5424 		 * index count - 1 could mean that the page was
5425 		 * invalidated or removed, so we must restart for
5426 		 * safety as well.
5427 		 */
5428 		if (!vm_page_all_valid(ma[i]))
5429 			redo = true;
5430 	}
5431 	if (redo && error == 0)
5432 		goto again;
5433 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5434 }
5435 
5436 #include "opt_ddb.h"
5437 #ifdef DDB
5438 #include <ddb/ddb.h>
5439 
5440 /* DDB command to show buffer data */
5441 DB_SHOW_COMMAND(buffer, db_show_buffer)
5442 {
5443 	/* get args */
5444 	struct buf *bp = (struct buf *)addr;
5445 #ifdef FULL_BUF_TRACKING
5446 	uint32_t i, j;
5447 #endif
5448 
5449 	if (!have_addr) {
5450 		db_printf("usage: show buffer <addr>\n");
5451 		return;
5452 	}
5453 
5454 	db_printf("buf at %p\n", bp);
5455 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5456 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5457 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5458 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5459 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5460 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5461 	db_printf(
5462 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5463 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5464 	    "b_vp = %p, b_dep = %p\n",
5465 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5466 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5467 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5468 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5469 	    bp->b_kvabase, bp->b_kvasize);
5470 	if (bp->b_npages) {
5471 		int i;
5472 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5473 		for (i = 0; i < bp->b_npages; i++) {
5474 			vm_page_t m;
5475 			m = bp->b_pages[i];
5476 			if (m != NULL)
5477 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5478 				    (u_long)m->pindex,
5479 				    (u_long)VM_PAGE_TO_PHYS(m));
5480 			else
5481 				db_printf("( ??? )");
5482 			if ((i + 1) < bp->b_npages)
5483 				db_printf(",");
5484 		}
5485 		db_printf("\n");
5486 	}
5487 	BUF_LOCKPRINTINFO(bp);
5488 #if defined(FULL_BUF_TRACKING)
5489 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5490 
5491 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5492 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5493 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5494 			continue;
5495 		db_printf(" %2u: %s\n", j,
5496 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5497 	}
5498 #elif defined(BUF_TRACKING)
5499 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5500 #endif
5501 	db_printf(" ");
5502 }
5503 
5504 DB_SHOW_COMMAND_FLAGS(bufqueues, bufqueues, DB_CMD_MEMSAFE)
5505 {
5506 	struct bufdomain *bd;
5507 	struct buf *bp;
5508 	long total;
5509 	int i, j, cnt;
5510 
5511 	db_printf("bqempty: %d\n", bqempty.bq_len);
5512 
5513 	for (i = 0; i < buf_domains; i++) {
5514 		bd = &bdomain[i];
5515 		db_printf("Buf domain %d\n", i);
5516 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5517 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5518 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5519 		db_printf("\n");
5520 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5521 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5522 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5523 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5524 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5525 		db_printf("\n");
5526 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5527 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5528 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5529 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5530 		db_printf("\n");
5531 		total = 0;
5532 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5533 			total += bp->b_bufsize;
5534 		db_printf("\tcleanq count\t%d (%ld)\n",
5535 		    bd->bd_cleanq->bq_len, total);
5536 		total = 0;
5537 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5538 			total += bp->b_bufsize;
5539 		db_printf("\tdirtyq count\t%d (%ld)\n",
5540 		    bd->bd_dirtyq.bq_len, total);
5541 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5542 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5543 		db_printf("\tCPU ");
5544 		for (j = 0; j <= mp_maxid; j++)
5545 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5546 		db_printf("\n");
5547 		cnt = 0;
5548 		total = 0;
5549 		for (j = 0; j < nbuf; j++) {
5550 			bp = nbufp(j);
5551 			if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5552 				cnt++;
5553 				total += bp->b_bufsize;
5554 			}
5555 		}
5556 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5557 		cnt = 0;
5558 		total = 0;
5559 		for (j = 0; j < nbuf; j++) {
5560 			bp = nbufp(j);
5561 			if (bp->b_domain == i) {
5562 				cnt++;
5563 				total += bp->b_bufsize;
5564 			}
5565 		}
5566 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5567 	}
5568 }
5569 
5570 DB_SHOW_COMMAND_FLAGS(lockedbufs, lockedbufs, DB_CMD_MEMSAFE)
5571 {
5572 	struct buf *bp;
5573 	int i;
5574 
5575 	for (i = 0; i < nbuf; i++) {
5576 		bp = nbufp(i);
5577 		if (BUF_ISLOCKED(bp)) {
5578 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5579 			db_printf("\n");
5580 			if (db_pager_quit)
5581 				break;
5582 		}
5583 	}
5584 }
5585 
5586 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5587 {
5588 	struct vnode *vp;
5589 	struct buf *bp;
5590 
5591 	if (!have_addr) {
5592 		db_printf("usage: show vnodebufs <addr>\n");
5593 		return;
5594 	}
5595 	vp = (struct vnode *)addr;
5596 	db_printf("Clean buffers:\n");
5597 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5598 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5599 		db_printf("\n");
5600 	}
5601 	db_printf("Dirty buffers:\n");
5602 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5603 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5604 		db_printf("\n");
5605 	}
5606 }
5607 
5608 DB_COMMAND_FLAGS(countfreebufs, db_coundfreebufs, DB_CMD_MEMSAFE)
5609 {
5610 	struct buf *bp;
5611 	int i, used = 0, nfree = 0;
5612 
5613 	if (have_addr) {
5614 		db_printf("usage: countfreebufs\n");
5615 		return;
5616 	}
5617 
5618 	for (i = 0; i < nbuf; i++) {
5619 		bp = nbufp(i);
5620 		if (bp->b_qindex == QUEUE_EMPTY)
5621 			nfree++;
5622 		else
5623 			used++;
5624 	}
5625 
5626 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5627 	    nfree + used);
5628 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5629 }
5630 #endif /* DDB */
5631