xref: /freebsd/sys/kern/vfs_bio.c (revision 0012f373e43db2341c20329163ed2d5ad3b0f341)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2004 Poul-Henning Kamp
5  * Copyright (c) 1994,1997 John S. Dyson
6  * Copyright (c) 2013 The FreeBSD Foundation
7  * All rights reserved.
8  *
9  * Portions of this software were developed by Konstantin Belousov
10  * under sponsorship from the FreeBSD Foundation.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * this file contains a new buffer I/O scheme implementing a coherent
36  * VM object and buffer cache scheme.  Pains have been taken to make
37  * sure that the performance degradation associated with schemes such
38  * as this is not realized.
39  *
40  * Author:  John S. Dyson
41  * Significant help during the development and debugging phases
42  * had been provided by David Greenman, also of the FreeBSD core team.
43  *
44  * see man buf(9) for more info.
45  */
46 
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/bio.h>
53 #include <sys/bitset.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/buf.h>
57 #include <sys/devicestat.h>
58 #include <sys/eventhandler.h>
59 #include <sys/fail.h>
60 #include <sys/ktr.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/malloc.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/proc.h>
69 #include <sys/racct.h>
70 #include <sys/refcount.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/smp.h>
74 #include <sys/sysctl.h>
75 #include <sys/sysproto.h>
76 #include <sys/vmem.h>
77 #include <sys/vmmeter.h>
78 #include <sys/vnode.h>
79 #include <sys/watchdog.h>
80 #include <geom/geom.h>
81 #include <vm/vm.h>
82 #include <vm/vm_param.h>
83 #include <vm/vm_kern.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_pager.h>
88 #include <vm/vm_extern.h>
89 #include <vm/vm_map.h>
90 #include <vm/swap_pager.h>
91 
92 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
93 
94 struct	bio_ops bioops;		/* I/O operation notification */
95 
96 struct	buf_ops buf_ops_bio = {
97 	.bop_name	=	"buf_ops_bio",
98 	.bop_write	=	bufwrite,
99 	.bop_strategy	=	bufstrategy,
100 	.bop_sync	=	bufsync,
101 	.bop_bdflush	=	bufbdflush,
102 };
103 
104 struct bufqueue {
105 	struct mtx_padalign	bq_lock;
106 	TAILQ_HEAD(, buf)	bq_queue;
107 	uint8_t			bq_index;
108 	uint16_t		bq_subqueue;
109 	int			bq_len;
110 } __aligned(CACHE_LINE_SIZE);
111 
112 #define	BQ_LOCKPTR(bq)		(&(bq)->bq_lock)
113 #define	BQ_LOCK(bq)		mtx_lock(BQ_LOCKPTR((bq)))
114 #define	BQ_UNLOCK(bq)		mtx_unlock(BQ_LOCKPTR((bq)))
115 #define	BQ_ASSERT_LOCKED(bq)	mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
116 
117 struct bufdomain {
118 	struct bufqueue	bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
119 	struct bufqueue bd_dirtyq;
120 	struct bufqueue	*bd_cleanq;
121 	struct mtx_padalign bd_run_lock;
122 	/* Constants */
123 	long		bd_maxbufspace;
124 	long		bd_hibufspace;
125 	long 		bd_lobufspace;
126 	long 		bd_bufspacethresh;
127 	int		bd_hifreebuffers;
128 	int		bd_lofreebuffers;
129 	int		bd_hidirtybuffers;
130 	int		bd_lodirtybuffers;
131 	int		bd_dirtybufthresh;
132 	int		bd_lim;
133 	/* atomics */
134 	int		bd_wanted;
135 	int __aligned(CACHE_LINE_SIZE)	bd_numdirtybuffers;
136 	int __aligned(CACHE_LINE_SIZE)	bd_running;
137 	long __aligned(CACHE_LINE_SIZE) bd_bufspace;
138 	int __aligned(CACHE_LINE_SIZE)	bd_freebuffers;
139 } __aligned(CACHE_LINE_SIZE);
140 
141 #define	BD_LOCKPTR(bd)		(&(bd)->bd_cleanq->bq_lock)
142 #define	BD_LOCK(bd)		mtx_lock(BD_LOCKPTR((bd)))
143 #define	BD_UNLOCK(bd)		mtx_unlock(BD_LOCKPTR((bd)))
144 #define	BD_ASSERT_LOCKED(bd)	mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
145 #define	BD_RUN_LOCKPTR(bd)	(&(bd)->bd_run_lock)
146 #define	BD_RUN_LOCK(bd)		mtx_lock(BD_RUN_LOCKPTR((bd)))
147 #define	BD_RUN_UNLOCK(bd)	mtx_unlock(BD_RUN_LOCKPTR((bd)))
148 #define	BD_DOMAIN(bd)		(bd - bdomain)
149 
150 static struct buf *buf;		/* buffer header pool */
151 extern struct buf *swbuf;	/* Swap buffer header pool. */
152 caddr_t unmapped_buf;
153 
154 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
155 struct proc *bufdaemonproc;
156 
157 static int inmem(struct vnode *vp, daddr_t blkno);
158 static void vm_hold_free_pages(struct buf *bp, int newbsize);
159 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
160 		vm_offset_t to);
161 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
162 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
163 		vm_page_t m);
164 static void vfs_clean_pages_dirty_buf(struct buf *bp);
165 static void vfs_setdirty_locked_object(struct buf *bp);
166 static void vfs_vmio_invalidate(struct buf *bp);
167 static void vfs_vmio_truncate(struct buf *bp, int npages);
168 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
169 static int vfs_bio_clcheck(struct vnode *vp, int size,
170 		daddr_t lblkno, daddr_t blkno);
171 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
172 		void (*)(struct buf *));
173 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
174 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
175 static void buf_daemon(void);
176 static __inline void bd_wakeup(void);
177 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
178 static void bufkva_reclaim(vmem_t *, int);
179 static void bufkva_free(struct buf *);
180 static int buf_import(void *, void **, int, int, int);
181 static void buf_release(void *, void **, int);
182 static void maxbcachebuf_adjust(void);
183 static inline struct bufdomain *bufdomain(struct buf *);
184 static void bq_remove(struct bufqueue *bq, struct buf *bp);
185 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
186 static int buf_recycle(struct bufdomain *, bool kva);
187 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
188 	    const char *lockname);
189 static void bd_init(struct bufdomain *bd);
190 static int bd_flushall(struct bufdomain *bd);
191 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
192 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
193 
194 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
195 int vmiodirenable = TRUE;
196 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
197     "Use the VM system for directory writes");
198 long runningbufspace;
199 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
200     "Amount of presently outstanding async buffer io");
201 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
202     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
203 static counter_u64_t bufkvaspace;
204 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
205     "Kernel virtual memory used for buffers");
206 static long maxbufspace;
207 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
208     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
209     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
210     "Maximum allowed value of bufspace (including metadata)");
211 static long bufmallocspace;
212 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
213     "Amount of malloced memory for buffers");
214 static long maxbufmallocspace;
215 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
216     0, "Maximum amount of malloced memory for buffers");
217 static long lobufspace;
218 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
219     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
220     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
221     "Minimum amount of buffers we want to have");
222 long hibufspace;
223 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
224     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
225     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
226     "Maximum allowed value of bufspace (excluding metadata)");
227 long bufspacethresh;
228 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
229     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
230     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
231     "Bufspace consumed before waking the daemon to free some");
232 static counter_u64_t buffreekvacnt;
233 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
234     "Number of times we have freed the KVA space from some buffer");
235 static counter_u64_t bufdefragcnt;
236 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
237     "Number of times we have had to repeat buffer allocation to defragment");
238 static long lorunningspace;
239 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
240     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
241     "Minimum preferred space used for in-progress I/O");
242 static long hirunningspace;
243 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
244     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
245     "Maximum amount of space to use for in-progress I/O");
246 int dirtybufferflushes;
247 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
248     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
249 int bdwriteskip;
250 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
251     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
252 int altbufferflushes;
253 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
254     &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
255 static int recursiveflushes;
256 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
257     &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
258 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
259 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
260     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
261     "Number of buffers that are dirty (has unwritten changes) at the moment");
262 static int lodirtybuffers;
263 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
264     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
265     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
266     "How many buffers we want to have free before bufdaemon can sleep");
267 static int hidirtybuffers;
268 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
269     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
270     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
271     "When the number of dirty buffers is considered severe");
272 int dirtybufthresh;
273 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
274     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
275     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
276     "Number of bdwrite to bawrite conversions to clear dirty buffers");
277 static int numfreebuffers;
278 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
279     "Number of free buffers");
280 static int lofreebuffers;
281 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
282     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
283     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
284    "Target number of free buffers");
285 static int hifreebuffers;
286 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
287     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
288     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
289    "Threshold for clean buffer recycling");
290 static counter_u64_t getnewbufcalls;
291 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
292    &getnewbufcalls, "Number of calls to getnewbuf");
293 static counter_u64_t getnewbufrestarts;
294 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
295     &getnewbufrestarts,
296     "Number of times getnewbuf has had to restart a buffer acquisition");
297 static counter_u64_t mappingrestarts;
298 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
299     &mappingrestarts,
300     "Number of times getblk has had to restart a buffer mapping for "
301     "unmapped buffer");
302 static counter_u64_t numbufallocfails;
303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
304     &numbufallocfails, "Number of times buffer allocations failed");
305 static int flushbufqtarget = 100;
306 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
307     "Amount of work to do in flushbufqueues when helping bufdaemon");
308 static counter_u64_t notbufdflushes;
309 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
310     "Number of dirty buffer flushes done by the bufdaemon helpers");
311 static long barrierwrites;
312 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
313     &barrierwrites, 0, "Number of barrier writes");
314 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
315     &unmapped_buf_allowed, 0,
316     "Permit the use of the unmapped i/o");
317 int maxbcachebuf = MAXBCACHEBUF;
318 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
319     "Maximum size of a buffer cache block");
320 
321 /*
322  * This lock synchronizes access to bd_request.
323  */
324 static struct mtx_padalign __exclusive_cache_line bdlock;
325 
326 /*
327  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
328  * waitrunningbufspace().
329  */
330 static struct mtx_padalign __exclusive_cache_line rbreqlock;
331 
332 /*
333  * Lock that protects bdirtywait.
334  */
335 static struct mtx_padalign __exclusive_cache_line bdirtylock;
336 
337 /*
338  * Wakeup point for bufdaemon, as well as indicator of whether it is already
339  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
340  * is idling.
341  */
342 static int bd_request;
343 
344 /*
345  * Request for the buf daemon to write more buffers than is indicated by
346  * lodirtybuf.  This may be necessary to push out excess dependencies or
347  * defragment the address space where a simple count of the number of dirty
348  * buffers is insufficient to characterize the demand for flushing them.
349  */
350 static int bd_speedupreq;
351 
352 /*
353  * Synchronization (sleep/wakeup) variable for active buffer space requests.
354  * Set when wait starts, cleared prior to wakeup().
355  * Used in runningbufwakeup() and waitrunningbufspace().
356  */
357 static int runningbufreq;
358 
359 /*
360  * Synchronization for bwillwrite() waiters.
361  */
362 static int bdirtywait;
363 
364 /*
365  * Definitions for the buffer free lists.
366  */
367 #define QUEUE_NONE	0	/* on no queue */
368 #define QUEUE_EMPTY	1	/* empty buffer headers */
369 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
370 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
371 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
372 
373 /* Maximum number of buffer domains. */
374 #define	BUF_DOMAINS	8
375 
376 struct bufdomainset bdlodirty;		/* Domains > lodirty */
377 struct bufdomainset bdhidirty;		/* Domains > hidirty */
378 
379 /* Configured number of clean queues. */
380 static int __read_mostly buf_domains;
381 
382 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
383 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
384 struct bufqueue __exclusive_cache_line bqempty;
385 
386 /*
387  * per-cpu empty buffer cache.
388  */
389 uma_zone_t buf_zone;
390 
391 /*
392  * Single global constant for BUF_WMESG, to avoid getting multiple references.
393  * buf_wmesg is referred from macros.
394  */
395 const char *buf_wmesg = BUF_WMESG;
396 
397 static int
398 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
399 {
400 	long value;
401 	int error;
402 
403 	value = *(long *)arg1;
404 	error = sysctl_handle_long(oidp, &value, 0, req);
405 	if (error != 0 || req->newptr == NULL)
406 		return (error);
407 	mtx_lock(&rbreqlock);
408 	if (arg1 == &hirunningspace) {
409 		if (value < lorunningspace)
410 			error = EINVAL;
411 		else
412 			hirunningspace = value;
413 	} else {
414 		KASSERT(arg1 == &lorunningspace,
415 		    ("%s: unknown arg1", __func__));
416 		if (value > hirunningspace)
417 			error = EINVAL;
418 		else
419 			lorunningspace = value;
420 	}
421 	mtx_unlock(&rbreqlock);
422 	return (error);
423 }
424 
425 static int
426 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
427 {
428 	int error;
429 	int value;
430 	int i;
431 
432 	value = *(int *)arg1;
433 	error = sysctl_handle_int(oidp, &value, 0, req);
434 	if (error != 0 || req->newptr == NULL)
435 		return (error);
436 	*(int *)arg1 = value;
437 	for (i = 0; i < buf_domains; i++)
438 		*(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
439 		    value / buf_domains;
440 
441 	return (error);
442 }
443 
444 static int
445 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
446 {
447 	long value;
448 	int error;
449 	int i;
450 
451 	value = *(long *)arg1;
452 	error = sysctl_handle_long(oidp, &value, 0, req);
453 	if (error != 0 || req->newptr == NULL)
454 		return (error);
455 	*(long *)arg1 = value;
456 	for (i = 0; i < buf_domains; i++)
457 		*(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
458 		    value / buf_domains;
459 
460 	return (error);
461 }
462 
463 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
464     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
465 static int
466 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
467 {
468 	long lvalue;
469 	int ivalue;
470 	int i;
471 
472 	lvalue = 0;
473 	for (i = 0; i < buf_domains; i++)
474 		lvalue += bdomain[i].bd_bufspace;
475 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
476 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
477 	if (lvalue > INT_MAX)
478 		/* On overflow, still write out a long to trigger ENOMEM. */
479 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
480 	ivalue = lvalue;
481 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
482 }
483 #else
484 static int
485 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
486 {
487 	long lvalue;
488 	int i;
489 
490 	lvalue = 0;
491 	for (i = 0; i < buf_domains; i++)
492 		lvalue += bdomain[i].bd_bufspace;
493 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
494 }
495 #endif
496 
497 static int
498 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
499 {
500 	int value;
501 	int i;
502 
503 	value = 0;
504 	for (i = 0; i < buf_domains; i++)
505 		value += bdomain[i].bd_numdirtybuffers;
506 	return (sysctl_handle_int(oidp, &value, 0, req));
507 }
508 
509 /*
510  *	bdirtywakeup:
511  *
512  *	Wakeup any bwillwrite() waiters.
513  */
514 static void
515 bdirtywakeup(void)
516 {
517 	mtx_lock(&bdirtylock);
518 	if (bdirtywait) {
519 		bdirtywait = 0;
520 		wakeup(&bdirtywait);
521 	}
522 	mtx_unlock(&bdirtylock);
523 }
524 
525 /*
526  *	bd_clear:
527  *
528  *	Clear a domain from the appropriate bitsets when dirtybuffers
529  *	is decremented.
530  */
531 static void
532 bd_clear(struct bufdomain *bd)
533 {
534 
535 	mtx_lock(&bdirtylock);
536 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
537 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
538 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
539 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
540 	mtx_unlock(&bdirtylock);
541 }
542 
543 /*
544  *	bd_set:
545  *
546  *	Set a domain in the appropriate bitsets when dirtybuffers
547  *	is incremented.
548  */
549 static void
550 bd_set(struct bufdomain *bd)
551 {
552 
553 	mtx_lock(&bdirtylock);
554 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
555 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
556 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
557 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
558 	mtx_unlock(&bdirtylock);
559 }
560 
561 /*
562  *	bdirtysub:
563  *
564  *	Decrement the numdirtybuffers count by one and wakeup any
565  *	threads blocked in bwillwrite().
566  */
567 static void
568 bdirtysub(struct buf *bp)
569 {
570 	struct bufdomain *bd;
571 	int num;
572 
573 	bd = bufdomain(bp);
574 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
575 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
576 		bdirtywakeup();
577 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
578 		bd_clear(bd);
579 }
580 
581 /*
582  *	bdirtyadd:
583  *
584  *	Increment the numdirtybuffers count by one and wakeup the buf
585  *	daemon if needed.
586  */
587 static void
588 bdirtyadd(struct buf *bp)
589 {
590 	struct bufdomain *bd;
591 	int num;
592 
593 	/*
594 	 * Only do the wakeup once as we cross the boundary.  The
595 	 * buf daemon will keep running until the condition clears.
596 	 */
597 	bd = bufdomain(bp);
598 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
599 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
600 		bd_wakeup();
601 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
602 		bd_set(bd);
603 }
604 
605 /*
606  *	bufspace_daemon_wakeup:
607  *
608  *	Wakeup the daemons responsible for freeing clean bufs.
609  */
610 static void
611 bufspace_daemon_wakeup(struct bufdomain *bd)
612 {
613 
614 	/*
615 	 * avoid the lock if the daemon is running.
616 	 */
617 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
618 		BD_RUN_LOCK(bd);
619 		atomic_store_int(&bd->bd_running, 1);
620 		wakeup(&bd->bd_running);
621 		BD_RUN_UNLOCK(bd);
622 	}
623 }
624 
625 /*
626  *	bufspace_daemon_wait:
627  *
628  *	Sleep until the domain falls below a limit or one second passes.
629  */
630 static void
631 bufspace_daemon_wait(struct bufdomain *bd)
632 {
633 	/*
634 	 * Re-check our limits and sleep.  bd_running must be
635 	 * cleared prior to checking the limits to avoid missed
636 	 * wakeups.  The waker will adjust one of bufspace or
637 	 * freebuffers prior to checking bd_running.
638 	 */
639 	BD_RUN_LOCK(bd);
640 	atomic_store_int(&bd->bd_running, 0);
641 	if (bd->bd_bufspace < bd->bd_bufspacethresh &&
642 	    bd->bd_freebuffers > bd->bd_lofreebuffers) {
643 		msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
644 		    "-", hz);
645 	} else {
646 		/* Avoid spurious wakeups while running. */
647 		atomic_store_int(&bd->bd_running, 1);
648 		BD_RUN_UNLOCK(bd);
649 	}
650 }
651 
652 /*
653  *	bufspace_adjust:
654  *
655  *	Adjust the reported bufspace for a KVA managed buffer, possibly
656  * 	waking any waiters.
657  */
658 static void
659 bufspace_adjust(struct buf *bp, int bufsize)
660 {
661 	struct bufdomain *bd;
662 	long space;
663 	int diff;
664 
665 	KASSERT((bp->b_flags & B_MALLOC) == 0,
666 	    ("bufspace_adjust: malloc buf %p", bp));
667 	bd = bufdomain(bp);
668 	diff = bufsize - bp->b_bufsize;
669 	if (diff < 0) {
670 		atomic_subtract_long(&bd->bd_bufspace, -diff);
671 	} else if (diff > 0) {
672 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
673 		/* Wake up the daemon on the transition. */
674 		if (space < bd->bd_bufspacethresh &&
675 		    space + diff >= bd->bd_bufspacethresh)
676 			bufspace_daemon_wakeup(bd);
677 	}
678 	bp->b_bufsize = bufsize;
679 }
680 
681 /*
682  *	bufspace_reserve:
683  *
684  *	Reserve bufspace before calling allocbuf().  metadata has a
685  *	different space limit than data.
686  */
687 static int
688 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
689 {
690 	long limit, new;
691 	long space;
692 
693 	if (metadata)
694 		limit = bd->bd_maxbufspace;
695 	else
696 		limit = bd->bd_hibufspace;
697 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
698 	new = space + size;
699 	if (new > limit) {
700 		atomic_subtract_long(&bd->bd_bufspace, size);
701 		return (ENOSPC);
702 	}
703 
704 	/* Wake up the daemon on the transition. */
705 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
706 		bufspace_daemon_wakeup(bd);
707 
708 	return (0);
709 }
710 
711 /*
712  *	bufspace_release:
713  *
714  *	Release reserved bufspace after bufspace_adjust() has consumed it.
715  */
716 static void
717 bufspace_release(struct bufdomain *bd, int size)
718 {
719 
720 	atomic_subtract_long(&bd->bd_bufspace, size);
721 }
722 
723 /*
724  *	bufspace_wait:
725  *
726  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
727  *	supplied.  bd_wanted must be set prior to polling for space.  The
728  *	operation must be re-tried on return.
729  */
730 static void
731 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
732     int slpflag, int slptimeo)
733 {
734 	struct thread *td;
735 	int error, fl, norunbuf;
736 
737 	if ((gbflags & GB_NOWAIT_BD) != 0)
738 		return;
739 
740 	td = curthread;
741 	BD_LOCK(bd);
742 	while (bd->bd_wanted) {
743 		if (vp != NULL && vp->v_type != VCHR &&
744 		    (td->td_pflags & TDP_BUFNEED) == 0) {
745 			BD_UNLOCK(bd);
746 			/*
747 			 * getblk() is called with a vnode locked, and
748 			 * some majority of the dirty buffers may as
749 			 * well belong to the vnode.  Flushing the
750 			 * buffers there would make a progress that
751 			 * cannot be achieved by the buf_daemon, that
752 			 * cannot lock the vnode.
753 			 */
754 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
755 			    (td->td_pflags & TDP_NORUNNINGBUF);
756 
757 			/*
758 			 * Play bufdaemon.  The getnewbuf() function
759 			 * may be called while the thread owns lock
760 			 * for another dirty buffer for the same
761 			 * vnode, which makes it impossible to use
762 			 * VOP_FSYNC() there, due to the buffer lock
763 			 * recursion.
764 			 */
765 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
766 			fl = buf_flush(vp, bd, flushbufqtarget);
767 			td->td_pflags &= norunbuf;
768 			BD_LOCK(bd);
769 			if (fl != 0)
770 				continue;
771 			if (bd->bd_wanted == 0)
772 				break;
773 		}
774 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
775 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
776 		if (error != 0)
777 			break;
778 	}
779 	BD_UNLOCK(bd);
780 }
781 
782 
783 /*
784  *	bufspace_daemon:
785  *
786  *	buffer space management daemon.  Tries to maintain some marginal
787  *	amount of free buffer space so that requesting processes neither
788  *	block nor work to reclaim buffers.
789  */
790 static void
791 bufspace_daemon(void *arg)
792 {
793 	struct bufdomain *bd;
794 
795 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
796 	    SHUTDOWN_PRI_LAST + 100);
797 
798 	bd = arg;
799 	for (;;) {
800 		kthread_suspend_check();
801 
802 		/*
803 		 * Free buffers from the clean queue until we meet our
804 		 * targets.
805 		 *
806 		 * Theory of operation:  The buffer cache is most efficient
807 		 * when some free buffer headers and space are always
808 		 * available to getnewbuf().  This daemon attempts to prevent
809 		 * the excessive blocking and synchronization associated
810 		 * with shortfall.  It goes through three phases according
811 		 * demand:
812 		 *
813 		 * 1)	The daemon wakes up voluntarily once per-second
814 		 *	during idle periods when the counters are below
815 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
816 		 *
817 		 * 2)	The daemon wakes up as we cross the thresholds
818 		 *	ahead of any potential blocking.  This may bounce
819 		 *	slightly according to the rate of consumption and
820 		 *	release.
821 		 *
822 		 * 3)	The daemon and consumers are starved for working
823 		 *	clean buffers.  This is the 'bufspace' sleep below
824 		 *	which will inefficiently trade bufs with bqrelse
825 		 *	until we return to condition 2.
826 		 */
827 		while (bd->bd_bufspace > bd->bd_lobufspace ||
828 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
829 			if (buf_recycle(bd, false) != 0) {
830 				if (bd_flushall(bd))
831 					continue;
832 				/*
833 				 * Speedup dirty if we've run out of clean
834 				 * buffers.  This is possible in particular
835 				 * because softdep may held many bufs locked
836 				 * pending writes to other bufs which are
837 				 * marked for delayed write, exhausting
838 				 * clean space until they are written.
839 				 */
840 				bd_speedup();
841 				BD_LOCK(bd);
842 				if (bd->bd_wanted) {
843 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
844 					    PRIBIO|PDROP, "bufspace", hz/10);
845 				} else
846 					BD_UNLOCK(bd);
847 			}
848 			maybe_yield();
849 		}
850 		bufspace_daemon_wait(bd);
851 	}
852 }
853 
854 /*
855  *	bufmallocadjust:
856  *
857  *	Adjust the reported bufspace for a malloc managed buffer, possibly
858  *	waking any waiters.
859  */
860 static void
861 bufmallocadjust(struct buf *bp, int bufsize)
862 {
863 	int diff;
864 
865 	KASSERT((bp->b_flags & B_MALLOC) != 0,
866 	    ("bufmallocadjust: non-malloc buf %p", bp));
867 	diff = bufsize - bp->b_bufsize;
868 	if (diff < 0)
869 		atomic_subtract_long(&bufmallocspace, -diff);
870 	else
871 		atomic_add_long(&bufmallocspace, diff);
872 	bp->b_bufsize = bufsize;
873 }
874 
875 /*
876  *	runningwakeup:
877  *
878  *	Wake up processes that are waiting on asynchronous writes to fall
879  *	below lorunningspace.
880  */
881 static void
882 runningwakeup(void)
883 {
884 
885 	mtx_lock(&rbreqlock);
886 	if (runningbufreq) {
887 		runningbufreq = 0;
888 		wakeup(&runningbufreq);
889 	}
890 	mtx_unlock(&rbreqlock);
891 }
892 
893 /*
894  *	runningbufwakeup:
895  *
896  *	Decrement the outstanding write count according.
897  */
898 void
899 runningbufwakeup(struct buf *bp)
900 {
901 	long space, bspace;
902 
903 	bspace = bp->b_runningbufspace;
904 	if (bspace == 0)
905 		return;
906 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
907 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
908 	    space, bspace));
909 	bp->b_runningbufspace = 0;
910 	/*
911 	 * Only acquire the lock and wakeup on the transition from exceeding
912 	 * the threshold to falling below it.
913 	 */
914 	if (space < lorunningspace)
915 		return;
916 	if (space - bspace > lorunningspace)
917 		return;
918 	runningwakeup();
919 }
920 
921 /*
922  *	waitrunningbufspace()
923  *
924  *	runningbufspace is a measure of the amount of I/O currently
925  *	running.  This routine is used in async-write situations to
926  *	prevent creating huge backups of pending writes to a device.
927  *	Only asynchronous writes are governed by this function.
928  *
929  *	This does NOT turn an async write into a sync write.  It waits
930  *	for earlier writes to complete and generally returns before the
931  *	caller's write has reached the device.
932  */
933 void
934 waitrunningbufspace(void)
935 {
936 
937 	mtx_lock(&rbreqlock);
938 	while (runningbufspace > hirunningspace) {
939 		runningbufreq = 1;
940 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
941 	}
942 	mtx_unlock(&rbreqlock);
943 }
944 
945 
946 /*
947  *	vfs_buf_test_cache:
948  *
949  *	Called when a buffer is extended.  This function clears the B_CACHE
950  *	bit if the newly extended portion of the buffer does not contain
951  *	valid data.
952  */
953 static __inline void
954 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
955     vm_offset_t size, vm_page_t m)
956 {
957 
958 	VM_OBJECT_ASSERT_LOCKED(m->object);
959 
960 	/*
961 	 * This function and its results are protected by higher level
962 	 * synchronization requiring vnode and buf locks to page in and
963 	 * validate pages.
964 	 */
965 	if (bp->b_flags & B_CACHE) {
966 		int base = (foff + off) & PAGE_MASK;
967 		if (vm_page_is_valid(m, base, size) == 0)
968 			bp->b_flags &= ~B_CACHE;
969 	}
970 }
971 
972 /* Wake up the buffer daemon if necessary */
973 static void
974 bd_wakeup(void)
975 {
976 
977 	mtx_lock(&bdlock);
978 	if (bd_request == 0) {
979 		bd_request = 1;
980 		wakeup(&bd_request);
981 	}
982 	mtx_unlock(&bdlock);
983 }
984 
985 /*
986  * Adjust the maxbcachbuf tunable.
987  */
988 static void
989 maxbcachebuf_adjust(void)
990 {
991 	int i;
992 
993 	/*
994 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
995 	 */
996 	i = 2;
997 	while (i * 2 <= maxbcachebuf)
998 		i *= 2;
999 	maxbcachebuf = i;
1000 	if (maxbcachebuf < MAXBSIZE)
1001 		maxbcachebuf = MAXBSIZE;
1002 	if (maxbcachebuf > MAXPHYS)
1003 		maxbcachebuf = MAXPHYS;
1004 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1005 		printf("maxbcachebuf=%d\n", maxbcachebuf);
1006 }
1007 
1008 /*
1009  * bd_speedup - speedup the buffer cache flushing code
1010  */
1011 void
1012 bd_speedup(void)
1013 {
1014 	int needwake;
1015 
1016 	mtx_lock(&bdlock);
1017 	needwake = 0;
1018 	if (bd_speedupreq == 0 || bd_request == 0)
1019 		needwake = 1;
1020 	bd_speedupreq = 1;
1021 	bd_request = 1;
1022 	if (needwake)
1023 		wakeup(&bd_request);
1024 	mtx_unlock(&bdlock);
1025 }
1026 
1027 #ifdef __i386__
1028 #define	TRANSIENT_DENOM	5
1029 #else
1030 #define	TRANSIENT_DENOM 10
1031 #endif
1032 
1033 /*
1034  * Calculating buffer cache scaling values and reserve space for buffer
1035  * headers.  This is called during low level kernel initialization and
1036  * may be called more then once.  We CANNOT write to the memory area
1037  * being reserved at this time.
1038  */
1039 caddr_t
1040 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1041 {
1042 	int tuned_nbuf;
1043 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1044 
1045 	/*
1046 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1047 	 * PAGE_SIZE is >= 1K)
1048 	 */
1049 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1050 
1051 	maxbcachebuf_adjust();
1052 	/*
1053 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1054 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1055 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1056 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1057 	 * the buffer cache we limit the eventual kva reservation to
1058 	 * maxbcache bytes.
1059 	 *
1060 	 * factor represents the 1/4 x ram conversion.
1061 	 */
1062 	if (nbuf == 0) {
1063 		int factor = 4 * BKVASIZE / 1024;
1064 
1065 		nbuf = 50;
1066 		if (physmem_est > 4096)
1067 			nbuf += min((physmem_est - 4096) / factor,
1068 			    65536 / factor);
1069 		if (physmem_est > 65536)
1070 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1071 			    32 * 1024 * 1024 / (factor * 5));
1072 
1073 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1074 			nbuf = maxbcache / BKVASIZE;
1075 		tuned_nbuf = 1;
1076 	} else
1077 		tuned_nbuf = 0;
1078 
1079 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1080 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1081 	if (nbuf > maxbuf) {
1082 		if (!tuned_nbuf)
1083 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1084 			    maxbuf);
1085 		nbuf = maxbuf;
1086 	}
1087 
1088 	/*
1089 	 * Ideal allocation size for the transient bio submap is 10%
1090 	 * of the maximal space buffer map.  This roughly corresponds
1091 	 * to the amount of the buffer mapped for typical UFS load.
1092 	 *
1093 	 * Clip the buffer map to reserve space for the transient
1094 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1095 	 * maximum buffer map extent on the platform.
1096 	 *
1097 	 * The fall-back to the maxbuf in case of maxbcache unset,
1098 	 * allows to not trim the buffer KVA for the architectures
1099 	 * with ample KVA space.
1100 	 */
1101 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1102 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1103 		buf_sz = (long)nbuf * BKVASIZE;
1104 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1105 		    (TRANSIENT_DENOM - 1)) {
1106 			/*
1107 			 * There is more KVA than memory.  Do not
1108 			 * adjust buffer map size, and assign the rest
1109 			 * of maxbuf to transient map.
1110 			 */
1111 			biotmap_sz = maxbuf_sz - buf_sz;
1112 		} else {
1113 			/*
1114 			 * Buffer map spans all KVA we could afford on
1115 			 * this platform.  Give 10% (20% on i386) of
1116 			 * the buffer map to the transient bio map.
1117 			 */
1118 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1119 			buf_sz -= biotmap_sz;
1120 		}
1121 		if (biotmap_sz / INT_MAX > MAXPHYS)
1122 			bio_transient_maxcnt = INT_MAX;
1123 		else
1124 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1125 		/*
1126 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1127 		 * using the transient mapping.
1128 		 */
1129 		if (bio_transient_maxcnt > 1024)
1130 			bio_transient_maxcnt = 1024;
1131 		if (tuned_nbuf)
1132 			nbuf = buf_sz / BKVASIZE;
1133 	}
1134 
1135 	if (nswbuf == 0) {
1136 		nswbuf = min(nbuf / 4, 256);
1137 		if (nswbuf < NSWBUF_MIN)
1138 			nswbuf = NSWBUF_MIN;
1139 	}
1140 
1141 	/*
1142 	 * Reserve space for the buffer cache buffers
1143 	 */
1144 	buf = (void *)v;
1145 	v = (caddr_t)(buf + nbuf);
1146 
1147 	return(v);
1148 }
1149 
1150 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1151 void
1152 bufinit(void)
1153 {
1154 	struct buf *bp;
1155 	int i;
1156 
1157 	KASSERT(maxbcachebuf >= MAXBSIZE,
1158 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1159 	    MAXBSIZE));
1160 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1161 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1162 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1163 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1164 
1165 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1166 
1167 	/* finally, initialize each buffer header and stick on empty q */
1168 	for (i = 0; i < nbuf; i++) {
1169 		bp = &buf[i];
1170 		bzero(bp, sizeof *bp);
1171 		bp->b_flags = B_INVAL;
1172 		bp->b_rcred = NOCRED;
1173 		bp->b_wcred = NOCRED;
1174 		bp->b_qindex = QUEUE_NONE;
1175 		bp->b_domain = -1;
1176 		bp->b_subqueue = mp_maxid + 1;
1177 		bp->b_xflags = 0;
1178 		bp->b_data = bp->b_kvabase = unmapped_buf;
1179 		LIST_INIT(&bp->b_dep);
1180 		BUF_LOCKINIT(bp);
1181 		bq_insert(&bqempty, bp, false);
1182 	}
1183 
1184 	/*
1185 	 * maxbufspace is the absolute maximum amount of buffer space we are
1186 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1187 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1188 	 * used by most other requests.  The differential is required to
1189 	 * ensure that metadata deadlocks don't occur.
1190 	 *
1191 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1192 	 * this may result in KVM fragmentation which is not handled optimally
1193 	 * by the system. XXX This is less true with vmem.  We could use
1194 	 * PAGE_SIZE.
1195 	 */
1196 	maxbufspace = (long)nbuf * BKVASIZE;
1197 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1198 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1199 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1200 
1201 	/*
1202 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1203 	 * arbitrarily and may need further tuning. It corresponds to
1204 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1205 	 * which fits with many RAID controllers' tagged queuing limits.
1206 	 * The lower 1 MiB limit is the historical upper limit for
1207 	 * hirunningspace.
1208 	 */
1209 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1210 	    16 * 1024 * 1024), 1024 * 1024);
1211 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1212 
1213 	/*
1214 	 * Limit the amount of malloc memory since it is wired permanently into
1215 	 * the kernel space.  Even though this is accounted for in the buffer
1216 	 * allocation, we don't want the malloced region to grow uncontrolled.
1217 	 * The malloc scheme improves memory utilization significantly on
1218 	 * average (small) directories.
1219 	 */
1220 	maxbufmallocspace = hibufspace / 20;
1221 
1222 	/*
1223 	 * Reduce the chance of a deadlock occurring by limiting the number
1224 	 * of delayed-write dirty buffers we allow to stack up.
1225 	 */
1226 	hidirtybuffers = nbuf / 4 + 20;
1227 	dirtybufthresh = hidirtybuffers * 9 / 10;
1228 	/*
1229 	 * To support extreme low-memory systems, make sure hidirtybuffers
1230 	 * cannot eat up all available buffer space.  This occurs when our
1231 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1232 	 * buffer space assuming BKVASIZE'd buffers.
1233 	 */
1234 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1235 		hidirtybuffers >>= 1;
1236 	}
1237 	lodirtybuffers = hidirtybuffers / 2;
1238 
1239 	/*
1240 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1241 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1242 	 * are counted as free but will be unavailable to threads executing
1243 	 * on other cpus.
1244 	 *
1245 	 * hifreebuffers is the free target for the bufspace daemon.  This
1246 	 * should be set appropriately to limit work per-iteration.
1247 	 */
1248 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1249 	hifreebuffers = (3 * lofreebuffers) / 2;
1250 	numfreebuffers = nbuf;
1251 
1252 	/* Setup the kva and free list allocators. */
1253 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1254 	buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1255 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1256 
1257 	/*
1258 	 * Size the clean queue according to the amount of buffer space.
1259 	 * One queue per-256mb up to the max.  More queues gives better
1260 	 * concurrency but less accurate LRU.
1261 	 */
1262 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1263 	for (i = 0 ; i < buf_domains; i++) {
1264 		struct bufdomain *bd;
1265 
1266 		bd = &bdomain[i];
1267 		bd_init(bd);
1268 		bd->bd_freebuffers = nbuf / buf_domains;
1269 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1270 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1271 		bd->bd_bufspace = 0;
1272 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1273 		bd->bd_hibufspace = hibufspace / buf_domains;
1274 		bd->bd_lobufspace = lobufspace / buf_domains;
1275 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1276 		bd->bd_numdirtybuffers = 0;
1277 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1278 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1279 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1280 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1281 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1282 	}
1283 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1284 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1285 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1286 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1287 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1288 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1289 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1290 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1291 }
1292 
1293 #ifdef INVARIANTS
1294 static inline void
1295 vfs_buf_check_mapped(struct buf *bp)
1296 {
1297 
1298 	KASSERT(bp->b_kvabase != unmapped_buf,
1299 	    ("mapped buf: b_kvabase was not updated %p", bp));
1300 	KASSERT(bp->b_data != unmapped_buf,
1301 	    ("mapped buf: b_data was not updated %p", bp));
1302 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1303 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1304 }
1305 
1306 static inline void
1307 vfs_buf_check_unmapped(struct buf *bp)
1308 {
1309 
1310 	KASSERT(bp->b_data == unmapped_buf,
1311 	    ("unmapped buf: corrupted b_data %p", bp));
1312 }
1313 
1314 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1315 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1316 #else
1317 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1318 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1319 #endif
1320 
1321 static int
1322 isbufbusy(struct buf *bp)
1323 {
1324 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1325 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1326 		return (1);
1327 	return (0);
1328 }
1329 
1330 /*
1331  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1332  */
1333 void
1334 bufshutdown(int show_busybufs)
1335 {
1336 	static int first_buf_printf = 1;
1337 	struct buf *bp;
1338 	int iter, nbusy, pbusy;
1339 #ifndef PREEMPTION
1340 	int subiter;
1341 #endif
1342 
1343 	/*
1344 	 * Sync filesystems for shutdown
1345 	 */
1346 	wdog_kern_pat(WD_LASTVAL);
1347 	sys_sync(curthread, NULL);
1348 
1349 	/*
1350 	 * With soft updates, some buffers that are
1351 	 * written will be remarked as dirty until other
1352 	 * buffers are written.
1353 	 */
1354 	for (iter = pbusy = 0; iter < 20; iter++) {
1355 		nbusy = 0;
1356 		for (bp = &buf[nbuf]; --bp >= buf; )
1357 			if (isbufbusy(bp))
1358 				nbusy++;
1359 		if (nbusy == 0) {
1360 			if (first_buf_printf)
1361 				printf("All buffers synced.");
1362 			break;
1363 		}
1364 		if (first_buf_printf) {
1365 			printf("Syncing disks, buffers remaining... ");
1366 			first_buf_printf = 0;
1367 		}
1368 		printf("%d ", nbusy);
1369 		if (nbusy < pbusy)
1370 			iter = 0;
1371 		pbusy = nbusy;
1372 
1373 		wdog_kern_pat(WD_LASTVAL);
1374 		sys_sync(curthread, NULL);
1375 
1376 #ifdef PREEMPTION
1377 		/*
1378 		 * Spin for a while to allow interrupt threads to run.
1379 		 */
1380 		DELAY(50000 * iter);
1381 #else
1382 		/*
1383 		 * Context switch several times to allow interrupt
1384 		 * threads to run.
1385 		 */
1386 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1387 			thread_lock(curthread);
1388 			mi_switch(SW_VOL, NULL);
1389 			thread_unlock(curthread);
1390 			DELAY(1000);
1391 		}
1392 #endif
1393 	}
1394 	printf("\n");
1395 	/*
1396 	 * Count only busy local buffers to prevent forcing
1397 	 * a fsck if we're just a client of a wedged NFS server
1398 	 */
1399 	nbusy = 0;
1400 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1401 		if (isbufbusy(bp)) {
1402 #if 0
1403 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1404 			if (bp->b_dev == NULL) {
1405 				TAILQ_REMOVE(&mountlist,
1406 				    bp->b_vp->v_mount, mnt_list);
1407 				continue;
1408 			}
1409 #endif
1410 			nbusy++;
1411 			if (show_busybufs > 0) {
1412 				printf(
1413 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1414 				    nbusy, bp, bp->b_vp, bp->b_flags,
1415 				    (intmax_t)bp->b_blkno,
1416 				    (intmax_t)bp->b_lblkno);
1417 				BUF_LOCKPRINTINFO(bp);
1418 				if (show_busybufs > 1)
1419 					vn_printf(bp->b_vp,
1420 					    "vnode content: ");
1421 			}
1422 		}
1423 	}
1424 	if (nbusy) {
1425 		/*
1426 		 * Failed to sync all blocks. Indicate this and don't
1427 		 * unmount filesystems (thus forcing an fsck on reboot).
1428 		 */
1429 		printf("Giving up on %d buffers\n", nbusy);
1430 		DELAY(5000000);	/* 5 seconds */
1431 	} else {
1432 		if (!first_buf_printf)
1433 			printf("Final sync complete\n");
1434 		/*
1435 		 * Unmount filesystems
1436 		 */
1437 		if (panicstr == NULL)
1438 			vfs_unmountall();
1439 	}
1440 	swapoff_all();
1441 	DELAY(100000);		/* wait for console output to finish */
1442 }
1443 
1444 static void
1445 bpmap_qenter(struct buf *bp)
1446 {
1447 
1448 	BUF_CHECK_MAPPED(bp);
1449 
1450 	/*
1451 	 * bp->b_data is relative to bp->b_offset, but
1452 	 * bp->b_offset may be offset into the first page.
1453 	 */
1454 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1455 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1456 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1457 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1458 }
1459 
1460 static inline struct bufdomain *
1461 bufdomain(struct buf *bp)
1462 {
1463 
1464 	return (&bdomain[bp->b_domain]);
1465 }
1466 
1467 static struct bufqueue *
1468 bufqueue(struct buf *bp)
1469 {
1470 
1471 	switch (bp->b_qindex) {
1472 	case QUEUE_NONE:
1473 		/* FALLTHROUGH */
1474 	case QUEUE_SENTINEL:
1475 		return (NULL);
1476 	case QUEUE_EMPTY:
1477 		return (&bqempty);
1478 	case QUEUE_DIRTY:
1479 		return (&bufdomain(bp)->bd_dirtyq);
1480 	case QUEUE_CLEAN:
1481 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1482 	default:
1483 		break;
1484 	}
1485 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1486 }
1487 
1488 /*
1489  * Return the locked bufqueue that bp is a member of.
1490  */
1491 static struct bufqueue *
1492 bufqueue_acquire(struct buf *bp)
1493 {
1494 	struct bufqueue *bq, *nbq;
1495 
1496 	/*
1497 	 * bp can be pushed from a per-cpu queue to the
1498 	 * cleanq while we're waiting on the lock.  Retry
1499 	 * if the queues don't match.
1500 	 */
1501 	bq = bufqueue(bp);
1502 	BQ_LOCK(bq);
1503 	for (;;) {
1504 		nbq = bufqueue(bp);
1505 		if (bq == nbq)
1506 			break;
1507 		BQ_UNLOCK(bq);
1508 		BQ_LOCK(nbq);
1509 		bq = nbq;
1510 	}
1511 	return (bq);
1512 }
1513 
1514 /*
1515  *	binsfree:
1516  *
1517  *	Insert the buffer into the appropriate free list.  Requires a
1518  *	locked buffer on entry and buffer is unlocked before return.
1519  */
1520 static void
1521 binsfree(struct buf *bp, int qindex)
1522 {
1523 	struct bufdomain *bd;
1524 	struct bufqueue *bq;
1525 
1526 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1527 	    ("binsfree: Invalid qindex %d", qindex));
1528 	BUF_ASSERT_XLOCKED(bp);
1529 
1530 	/*
1531 	 * Handle delayed bremfree() processing.
1532 	 */
1533 	if (bp->b_flags & B_REMFREE) {
1534 		if (bp->b_qindex == qindex) {
1535 			bp->b_flags |= B_REUSE;
1536 			bp->b_flags &= ~B_REMFREE;
1537 			BUF_UNLOCK(bp);
1538 			return;
1539 		}
1540 		bq = bufqueue_acquire(bp);
1541 		bq_remove(bq, bp);
1542 		BQ_UNLOCK(bq);
1543 	}
1544 	bd = bufdomain(bp);
1545 	if (qindex == QUEUE_CLEAN) {
1546 		if (bd->bd_lim != 0)
1547 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1548 		else
1549 			bq = bd->bd_cleanq;
1550 	} else
1551 		bq = &bd->bd_dirtyq;
1552 	bq_insert(bq, bp, true);
1553 }
1554 
1555 /*
1556  * buf_free:
1557  *
1558  *	Free a buffer to the buf zone once it no longer has valid contents.
1559  */
1560 static void
1561 buf_free(struct buf *bp)
1562 {
1563 
1564 	if (bp->b_flags & B_REMFREE)
1565 		bremfreef(bp);
1566 	if (bp->b_vflags & BV_BKGRDINPROG)
1567 		panic("losing buffer 1");
1568 	if (bp->b_rcred != NOCRED) {
1569 		crfree(bp->b_rcred);
1570 		bp->b_rcred = NOCRED;
1571 	}
1572 	if (bp->b_wcred != NOCRED) {
1573 		crfree(bp->b_wcred);
1574 		bp->b_wcred = NOCRED;
1575 	}
1576 	if (!LIST_EMPTY(&bp->b_dep))
1577 		buf_deallocate(bp);
1578 	bufkva_free(bp);
1579 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1580 	BUF_UNLOCK(bp);
1581 	uma_zfree(buf_zone, bp);
1582 }
1583 
1584 /*
1585  * buf_import:
1586  *
1587  *	Import bufs into the uma cache from the buf list.  The system still
1588  *	expects a static array of bufs and much of the synchronization
1589  *	around bufs assumes type stable storage.  As a result, UMA is used
1590  *	only as a per-cpu cache of bufs still maintained on a global list.
1591  */
1592 static int
1593 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1594 {
1595 	struct buf *bp;
1596 	int i;
1597 
1598 	BQ_LOCK(&bqempty);
1599 	for (i = 0; i < cnt; i++) {
1600 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1601 		if (bp == NULL)
1602 			break;
1603 		bq_remove(&bqempty, bp);
1604 		store[i] = bp;
1605 	}
1606 	BQ_UNLOCK(&bqempty);
1607 
1608 	return (i);
1609 }
1610 
1611 /*
1612  * buf_release:
1613  *
1614  *	Release bufs from the uma cache back to the buffer queues.
1615  */
1616 static void
1617 buf_release(void *arg, void **store, int cnt)
1618 {
1619 	struct bufqueue *bq;
1620 	struct buf *bp;
1621         int i;
1622 
1623 	bq = &bqempty;
1624 	BQ_LOCK(bq);
1625         for (i = 0; i < cnt; i++) {
1626 		bp = store[i];
1627 		/* Inline bq_insert() to batch locking. */
1628 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1629 		bp->b_flags &= ~(B_AGE | B_REUSE);
1630 		bq->bq_len++;
1631 		bp->b_qindex = bq->bq_index;
1632 	}
1633 	BQ_UNLOCK(bq);
1634 }
1635 
1636 /*
1637  * buf_alloc:
1638  *
1639  *	Allocate an empty buffer header.
1640  */
1641 static struct buf *
1642 buf_alloc(struct bufdomain *bd)
1643 {
1644 	struct buf *bp;
1645 	int freebufs;
1646 
1647 	/*
1648 	 * We can only run out of bufs in the buf zone if the average buf
1649 	 * is less than BKVASIZE.  In this case the actual wait/block will
1650 	 * come from buf_reycle() failing to flush one of these small bufs.
1651 	 */
1652 	bp = NULL;
1653 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1654 	if (freebufs > 0)
1655 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1656 	if (bp == NULL) {
1657 		atomic_add_int(&bd->bd_freebuffers, 1);
1658 		bufspace_daemon_wakeup(bd);
1659 		counter_u64_add(numbufallocfails, 1);
1660 		return (NULL);
1661 	}
1662 	/*
1663 	 * Wake-up the bufspace daemon on transition below threshold.
1664 	 */
1665 	if (freebufs == bd->bd_lofreebuffers)
1666 		bufspace_daemon_wakeup(bd);
1667 
1668 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1669 		panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1670 
1671 	KASSERT(bp->b_vp == NULL,
1672 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1673 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1674 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1675 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1676 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1677 	KASSERT(bp->b_npages == 0,
1678 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1679 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1680 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1681 
1682 	bp->b_domain = BD_DOMAIN(bd);
1683 	bp->b_flags = 0;
1684 	bp->b_ioflags = 0;
1685 	bp->b_xflags = 0;
1686 	bp->b_vflags = 0;
1687 	bp->b_vp = NULL;
1688 	bp->b_blkno = bp->b_lblkno = 0;
1689 	bp->b_offset = NOOFFSET;
1690 	bp->b_iodone = 0;
1691 	bp->b_error = 0;
1692 	bp->b_resid = 0;
1693 	bp->b_bcount = 0;
1694 	bp->b_npages = 0;
1695 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1696 	bp->b_bufobj = NULL;
1697 	bp->b_data = bp->b_kvabase = unmapped_buf;
1698 	bp->b_fsprivate1 = NULL;
1699 	bp->b_fsprivate2 = NULL;
1700 	bp->b_fsprivate3 = NULL;
1701 	LIST_INIT(&bp->b_dep);
1702 
1703 	return (bp);
1704 }
1705 
1706 /*
1707  *	buf_recycle:
1708  *
1709  *	Free a buffer from the given bufqueue.  kva controls whether the
1710  *	freed buf must own some kva resources.  This is used for
1711  *	defragmenting.
1712  */
1713 static int
1714 buf_recycle(struct bufdomain *bd, bool kva)
1715 {
1716 	struct bufqueue *bq;
1717 	struct buf *bp, *nbp;
1718 
1719 	if (kva)
1720 		counter_u64_add(bufdefragcnt, 1);
1721 	nbp = NULL;
1722 	bq = bd->bd_cleanq;
1723 	BQ_LOCK(bq);
1724 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1725 	    ("buf_recycle: Locks don't match"));
1726 	nbp = TAILQ_FIRST(&bq->bq_queue);
1727 
1728 	/*
1729 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1730 	 * depending.
1731 	 */
1732 	while ((bp = nbp) != NULL) {
1733 		/*
1734 		 * Calculate next bp (we can only use it if we do not
1735 		 * release the bqlock).
1736 		 */
1737 		nbp = TAILQ_NEXT(bp, b_freelist);
1738 
1739 		/*
1740 		 * If we are defragging then we need a buffer with
1741 		 * some kva to reclaim.
1742 		 */
1743 		if (kva && bp->b_kvasize == 0)
1744 			continue;
1745 
1746 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1747 			continue;
1748 
1749 		/*
1750 		 * Implement a second chance algorithm for frequently
1751 		 * accessed buffers.
1752 		 */
1753 		if ((bp->b_flags & B_REUSE) != 0) {
1754 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1755 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1756 			bp->b_flags &= ~B_REUSE;
1757 			BUF_UNLOCK(bp);
1758 			continue;
1759 		}
1760 
1761 		/*
1762 		 * Skip buffers with background writes in progress.
1763 		 */
1764 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1765 			BUF_UNLOCK(bp);
1766 			continue;
1767 		}
1768 
1769 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1770 		    ("buf_recycle: inconsistent queue %d bp %p",
1771 		    bp->b_qindex, bp));
1772 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1773 		    ("getnewbuf: queue domain %d doesn't match request %d",
1774 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1775 		/*
1776 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1777 		 * the scan from this point on.
1778 		 */
1779 		bq_remove(bq, bp);
1780 		BQ_UNLOCK(bq);
1781 
1782 		/*
1783 		 * Requeue the background write buffer with error and
1784 		 * restart the scan.
1785 		 */
1786 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1787 			bqrelse(bp);
1788 			BQ_LOCK(bq);
1789 			nbp = TAILQ_FIRST(&bq->bq_queue);
1790 			continue;
1791 		}
1792 		bp->b_flags |= B_INVAL;
1793 		brelse(bp);
1794 		return (0);
1795 	}
1796 	bd->bd_wanted = 1;
1797 	BQ_UNLOCK(bq);
1798 
1799 	return (ENOBUFS);
1800 }
1801 
1802 /*
1803  *	bremfree:
1804  *
1805  *	Mark the buffer for removal from the appropriate free list.
1806  *
1807  */
1808 void
1809 bremfree(struct buf *bp)
1810 {
1811 
1812 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1813 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1814 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1815 	KASSERT(bp->b_qindex != QUEUE_NONE,
1816 	    ("bremfree: buffer %p not on a queue.", bp));
1817 	BUF_ASSERT_XLOCKED(bp);
1818 
1819 	bp->b_flags |= B_REMFREE;
1820 }
1821 
1822 /*
1823  *	bremfreef:
1824  *
1825  *	Force an immediate removal from a free list.  Used only in nfs when
1826  *	it abuses the b_freelist pointer.
1827  */
1828 void
1829 bremfreef(struct buf *bp)
1830 {
1831 	struct bufqueue *bq;
1832 
1833 	bq = bufqueue_acquire(bp);
1834 	bq_remove(bq, bp);
1835 	BQ_UNLOCK(bq);
1836 }
1837 
1838 static void
1839 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1840 {
1841 
1842 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1843 	TAILQ_INIT(&bq->bq_queue);
1844 	bq->bq_len = 0;
1845 	bq->bq_index = qindex;
1846 	bq->bq_subqueue = subqueue;
1847 }
1848 
1849 static void
1850 bd_init(struct bufdomain *bd)
1851 {
1852 	int i;
1853 
1854 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1855 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1856 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1857 	for (i = 0; i <= mp_maxid; i++)
1858 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1859 		    "bufq clean subqueue lock");
1860 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1861 }
1862 
1863 /*
1864  *	bq_remove:
1865  *
1866  *	Removes a buffer from the free list, must be called with the
1867  *	correct qlock held.
1868  */
1869 static void
1870 bq_remove(struct bufqueue *bq, struct buf *bp)
1871 {
1872 
1873 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1874 	    bp, bp->b_vp, bp->b_flags);
1875 	KASSERT(bp->b_qindex != QUEUE_NONE,
1876 	    ("bq_remove: buffer %p not on a queue.", bp));
1877 	KASSERT(bufqueue(bp) == bq,
1878 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1879 
1880 	BQ_ASSERT_LOCKED(bq);
1881 	if (bp->b_qindex != QUEUE_EMPTY) {
1882 		BUF_ASSERT_XLOCKED(bp);
1883 	}
1884 	KASSERT(bq->bq_len >= 1,
1885 	    ("queue %d underflow", bp->b_qindex));
1886 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1887 	bq->bq_len--;
1888 	bp->b_qindex = QUEUE_NONE;
1889 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1890 }
1891 
1892 static void
1893 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1894 {
1895 	struct buf *bp;
1896 
1897 	BQ_ASSERT_LOCKED(bq);
1898 	if (bq != bd->bd_cleanq) {
1899 		BD_LOCK(bd);
1900 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1901 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1902 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1903 			    b_freelist);
1904 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1905 		}
1906 		bd->bd_cleanq->bq_len += bq->bq_len;
1907 		bq->bq_len = 0;
1908 	}
1909 	if (bd->bd_wanted) {
1910 		bd->bd_wanted = 0;
1911 		wakeup(&bd->bd_wanted);
1912 	}
1913 	if (bq != bd->bd_cleanq)
1914 		BD_UNLOCK(bd);
1915 }
1916 
1917 static int
1918 bd_flushall(struct bufdomain *bd)
1919 {
1920 	struct bufqueue *bq;
1921 	int flushed;
1922 	int i;
1923 
1924 	if (bd->bd_lim == 0)
1925 		return (0);
1926 	flushed = 0;
1927 	for (i = 0; i <= mp_maxid; i++) {
1928 		bq = &bd->bd_subq[i];
1929 		if (bq->bq_len == 0)
1930 			continue;
1931 		BQ_LOCK(bq);
1932 		bd_flush(bd, bq);
1933 		BQ_UNLOCK(bq);
1934 		flushed++;
1935 	}
1936 
1937 	return (flushed);
1938 }
1939 
1940 static void
1941 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1942 {
1943 	struct bufdomain *bd;
1944 
1945 	if (bp->b_qindex != QUEUE_NONE)
1946 		panic("bq_insert: free buffer %p onto another queue?", bp);
1947 
1948 	bd = bufdomain(bp);
1949 	if (bp->b_flags & B_AGE) {
1950 		/* Place this buf directly on the real queue. */
1951 		if (bq->bq_index == QUEUE_CLEAN)
1952 			bq = bd->bd_cleanq;
1953 		BQ_LOCK(bq);
1954 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1955 	} else {
1956 		BQ_LOCK(bq);
1957 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1958 	}
1959 	bp->b_flags &= ~(B_AGE | B_REUSE);
1960 	bq->bq_len++;
1961 	bp->b_qindex = bq->bq_index;
1962 	bp->b_subqueue = bq->bq_subqueue;
1963 
1964 	/*
1965 	 * Unlock before we notify so that we don't wakeup a waiter that
1966 	 * fails a trylock on the buf and sleeps again.
1967 	 */
1968 	if (unlock)
1969 		BUF_UNLOCK(bp);
1970 
1971 	if (bp->b_qindex == QUEUE_CLEAN) {
1972 		/*
1973 		 * Flush the per-cpu queue and notify any waiters.
1974 		 */
1975 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1976 		    bq->bq_len >= bd->bd_lim))
1977 			bd_flush(bd, bq);
1978 	}
1979 	BQ_UNLOCK(bq);
1980 }
1981 
1982 /*
1983  *	bufkva_free:
1984  *
1985  *	Free the kva allocation for a buffer.
1986  *
1987  */
1988 static void
1989 bufkva_free(struct buf *bp)
1990 {
1991 
1992 #ifdef INVARIANTS
1993 	if (bp->b_kvasize == 0) {
1994 		KASSERT(bp->b_kvabase == unmapped_buf &&
1995 		    bp->b_data == unmapped_buf,
1996 		    ("Leaked KVA space on %p", bp));
1997 	} else if (buf_mapped(bp))
1998 		BUF_CHECK_MAPPED(bp);
1999 	else
2000 		BUF_CHECK_UNMAPPED(bp);
2001 #endif
2002 	if (bp->b_kvasize == 0)
2003 		return;
2004 
2005 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2006 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2007 	counter_u64_add(buffreekvacnt, 1);
2008 	bp->b_data = bp->b_kvabase = unmapped_buf;
2009 	bp->b_kvasize = 0;
2010 }
2011 
2012 /*
2013  *	bufkva_alloc:
2014  *
2015  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2016  */
2017 static int
2018 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2019 {
2020 	vm_offset_t addr;
2021 	int error;
2022 
2023 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2024 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2025 
2026 	bufkva_free(bp);
2027 
2028 	addr = 0;
2029 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2030 	if (error != 0) {
2031 		/*
2032 		 * Buffer map is too fragmented.  Request the caller
2033 		 * to defragment the map.
2034 		 */
2035 		return (error);
2036 	}
2037 	bp->b_kvabase = (caddr_t)addr;
2038 	bp->b_kvasize = maxsize;
2039 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2040 	if ((gbflags & GB_UNMAPPED) != 0) {
2041 		bp->b_data = unmapped_buf;
2042 		BUF_CHECK_UNMAPPED(bp);
2043 	} else {
2044 		bp->b_data = bp->b_kvabase;
2045 		BUF_CHECK_MAPPED(bp);
2046 	}
2047 	return (0);
2048 }
2049 
2050 /*
2051  *	bufkva_reclaim:
2052  *
2053  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2054  *	callback that fires to avoid returning failure.
2055  */
2056 static void
2057 bufkva_reclaim(vmem_t *vmem, int flags)
2058 {
2059 	bool done;
2060 	int q;
2061 	int i;
2062 
2063 	done = false;
2064 	for (i = 0; i < 5; i++) {
2065 		for (q = 0; q < buf_domains; q++)
2066 			if (buf_recycle(&bdomain[q], true) != 0)
2067 				done = true;
2068 		if (done)
2069 			break;
2070 	}
2071 	return;
2072 }
2073 
2074 /*
2075  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2076  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2077  * the buffer is valid and we do not have to do anything.
2078  */
2079 static void
2080 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2081     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2082 {
2083 	struct buf *rabp;
2084 	struct thread *td;
2085 	int i;
2086 
2087 	td = curthread;
2088 
2089 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2090 		if (inmem(vp, *rablkno))
2091 			continue;
2092 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2093 		if ((rabp->b_flags & B_CACHE) != 0) {
2094 			brelse(rabp);
2095 			continue;
2096 		}
2097 #ifdef RACCT
2098 		if (racct_enable) {
2099 			PROC_LOCK(curproc);
2100 			racct_add_buf(curproc, rabp, 0);
2101 			PROC_UNLOCK(curproc);
2102 		}
2103 #endif /* RACCT */
2104 		td->td_ru.ru_inblock++;
2105 		rabp->b_flags |= B_ASYNC;
2106 		rabp->b_flags &= ~B_INVAL;
2107 		if ((flags & GB_CKHASH) != 0) {
2108 			rabp->b_flags |= B_CKHASH;
2109 			rabp->b_ckhashcalc = ckhashfunc;
2110 		}
2111 		rabp->b_ioflags &= ~BIO_ERROR;
2112 		rabp->b_iocmd = BIO_READ;
2113 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2114 			rabp->b_rcred = crhold(cred);
2115 		vfs_busy_pages(rabp, 0);
2116 		BUF_KERNPROC(rabp);
2117 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2118 		bstrategy(rabp);
2119 	}
2120 }
2121 
2122 /*
2123  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2124  *
2125  * Get a buffer with the specified data.  Look in the cache first.  We
2126  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2127  * is set, the buffer is valid and we do not have to do anything, see
2128  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2129  *
2130  * Always return a NULL buffer pointer (in bpp) when returning an error.
2131  */
2132 int
2133 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
2134     int *rabsize, int cnt, struct ucred *cred, int flags,
2135     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2136 {
2137 	struct buf *bp;
2138 	struct thread *td;
2139 	int error, readwait, rv;
2140 
2141 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2142 	td = curthread;
2143 	/*
2144 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2145 	 * are specified.
2146 	 */
2147 	error = getblkx(vp, blkno, size, 0, 0, flags, &bp);
2148 	if (error != 0) {
2149 		*bpp = NULL;
2150 		return (error);
2151 	}
2152 	flags &= ~GB_NOSPARSE;
2153 	*bpp = bp;
2154 
2155 	/*
2156 	 * If not found in cache, do some I/O
2157 	 */
2158 	readwait = 0;
2159 	if ((bp->b_flags & B_CACHE) == 0) {
2160 #ifdef RACCT
2161 		if (racct_enable) {
2162 			PROC_LOCK(td->td_proc);
2163 			racct_add_buf(td->td_proc, bp, 0);
2164 			PROC_UNLOCK(td->td_proc);
2165 		}
2166 #endif /* RACCT */
2167 		td->td_ru.ru_inblock++;
2168 		bp->b_iocmd = BIO_READ;
2169 		bp->b_flags &= ~B_INVAL;
2170 		if ((flags & GB_CKHASH) != 0) {
2171 			bp->b_flags |= B_CKHASH;
2172 			bp->b_ckhashcalc = ckhashfunc;
2173 		}
2174 		bp->b_ioflags &= ~BIO_ERROR;
2175 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2176 			bp->b_rcred = crhold(cred);
2177 		vfs_busy_pages(bp, 0);
2178 		bp->b_iooffset = dbtob(bp->b_blkno);
2179 		bstrategy(bp);
2180 		++readwait;
2181 	}
2182 
2183 	/*
2184 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2185 	 */
2186 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2187 
2188 	rv = 0;
2189 	if (readwait) {
2190 		rv = bufwait(bp);
2191 		if (rv != 0) {
2192 			brelse(bp);
2193 			*bpp = NULL;
2194 		}
2195 	}
2196 	return (rv);
2197 }
2198 
2199 /*
2200  * Write, release buffer on completion.  (Done by iodone
2201  * if async).  Do not bother writing anything if the buffer
2202  * is invalid.
2203  *
2204  * Note that we set B_CACHE here, indicating that buffer is
2205  * fully valid and thus cacheable.  This is true even of NFS
2206  * now so we set it generally.  This could be set either here
2207  * or in biodone() since the I/O is synchronous.  We put it
2208  * here.
2209  */
2210 int
2211 bufwrite(struct buf *bp)
2212 {
2213 	int oldflags;
2214 	struct vnode *vp;
2215 	long space;
2216 	int vp_md;
2217 
2218 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2219 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2220 		bp->b_flags |= B_INVAL | B_RELBUF;
2221 		bp->b_flags &= ~B_CACHE;
2222 		brelse(bp);
2223 		return (ENXIO);
2224 	}
2225 	if (bp->b_flags & B_INVAL) {
2226 		brelse(bp);
2227 		return (0);
2228 	}
2229 
2230 	if (bp->b_flags & B_BARRIER)
2231 		atomic_add_long(&barrierwrites, 1);
2232 
2233 	oldflags = bp->b_flags;
2234 
2235 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2236 	    ("FFS background buffer should not get here %p", bp));
2237 
2238 	vp = bp->b_vp;
2239 	if (vp)
2240 		vp_md = vp->v_vflag & VV_MD;
2241 	else
2242 		vp_md = 0;
2243 
2244 	/*
2245 	 * Mark the buffer clean.  Increment the bufobj write count
2246 	 * before bundirty() call, to prevent other thread from seeing
2247 	 * empty dirty list and zero counter for writes in progress,
2248 	 * falsely indicating that the bufobj is clean.
2249 	 */
2250 	bufobj_wref(bp->b_bufobj);
2251 	bundirty(bp);
2252 
2253 	bp->b_flags &= ~B_DONE;
2254 	bp->b_ioflags &= ~BIO_ERROR;
2255 	bp->b_flags |= B_CACHE;
2256 	bp->b_iocmd = BIO_WRITE;
2257 
2258 	vfs_busy_pages(bp, 1);
2259 
2260 	/*
2261 	 * Normal bwrites pipeline writes
2262 	 */
2263 	bp->b_runningbufspace = bp->b_bufsize;
2264 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2265 
2266 #ifdef RACCT
2267 	if (racct_enable) {
2268 		PROC_LOCK(curproc);
2269 		racct_add_buf(curproc, bp, 1);
2270 		PROC_UNLOCK(curproc);
2271 	}
2272 #endif /* RACCT */
2273 	curthread->td_ru.ru_oublock++;
2274 	if (oldflags & B_ASYNC)
2275 		BUF_KERNPROC(bp);
2276 	bp->b_iooffset = dbtob(bp->b_blkno);
2277 	buf_track(bp, __func__);
2278 	bstrategy(bp);
2279 
2280 	if ((oldflags & B_ASYNC) == 0) {
2281 		int rtval = bufwait(bp);
2282 		brelse(bp);
2283 		return (rtval);
2284 	} else if (space > hirunningspace) {
2285 		/*
2286 		 * don't allow the async write to saturate the I/O
2287 		 * system.  We will not deadlock here because
2288 		 * we are blocking waiting for I/O that is already in-progress
2289 		 * to complete. We do not block here if it is the update
2290 		 * or syncer daemon trying to clean up as that can lead
2291 		 * to deadlock.
2292 		 */
2293 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2294 			waitrunningbufspace();
2295 	}
2296 
2297 	return (0);
2298 }
2299 
2300 void
2301 bufbdflush(struct bufobj *bo, struct buf *bp)
2302 {
2303 	struct buf *nbp;
2304 
2305 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2306 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2307 		altbufferflushes++;
2308 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2309 		BO_LOCK(bo);
2310 		/*
2311 		 * Try to find a buffer to flush.
2312 		 */
2313 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2314 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2315 			    BUF_LOCK(nbp,
2316 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2317 				continue;
2318 			if (bp == nbp)
2319 				panic("bdwrite: found ourselves");
2320 			BO_UNLOCK(bo);
2321 			/* Don't countdeps with the bo lock held. */
2322 			if (buf_countdeps(nbp, 0)) {
2323 				BO_LOCK(bo);
2324 				BUF_UNLOCK(nbp);
2325 				continue;
2326 			}
2327 			if (nbp->b_flags & B_CLUSTEROK) {
2328 				vfs_bio_awrite(nbp);
2329 			} else {
2330 				bremfree(nbp);
2331 				bawrite(nbp);
2332 			}
2333 			dirtybufferflushes++;
2334 			break;
2335 		}
2336 		if (nbp == NULL)
2337 			BO_UNLOCK(bo);
2338 	}
2339 }
2340 
2341 /*
2342  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2343  * anything if the buffer is marked invalid.
2344  *
2345  * Note that since the buffer must be completely valid, we can safely
2346  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2347  * biodone() in order to prevent getblk from writing the buffer
2348  * out synchronously.
2349  */
2350 void
2351 bdwrite(struct buf *bp)
2352 {
2353 	struct thread *td = curthread;
2354 	struct vnode *vp;
2355 	struct bufobj *bo;
2356 
2357 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2358 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2359 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2360 	    ("Barrier request in delayed write %p", bp));
2361 
2362 	if (bp->b_flags & B_INVAL) {
2363 		brelse(bp);
2364 		return;
2365 	}
2366 
2367 	/*
2368 	 * If we have too many dirty buffers, don't create any more.
2369 	 * If we are wildly over our limit, then force a complete
2370 	 * cleanup. Otherwise, just keep the situation from getting
2371 	 * out of control. Note that we have to avoid a recursive
2372 	 * disaster and not try to clean up after our own cleanup!
2373 	 */
2374 	vp = bp->b_vp;
2375 	bo = bp->b_bufobj;
2376 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2377 		td->td_pflags |= TDP_INBDFLUSH;
2378 		BO_BDFLUSH(bo, bp);
2379 		td->td_pflags &= ~TDP_INBDFLUSH;
2380 	} else
2381 		recursiveflushes++;
2382 
2383 	bdirty(bp);
2384 	/*
2385 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2386 	 * true even of NFS now.
2387 	 */
2388 	bp->b_flags |= B_CACHE;
2389 
2390 	/*
2391 	 * This bmap keeps the system from needing to do the bmap later,
2392 	 * perhaps when the system is attempting to do a sync.  Since it
2393 	 * is likely that the indirect block -- or whatever other datastructure
2394 	 * that the filesystem needs is still in memory now, it is a good
2395 	 * thing to do this.  Note also, that if the pageout daemon is
2396 	 * requesting a sync -- there might not be enough memory to do
2397 	 * the bmap then...  So, this is important to do.
2398 	 */
2399 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2400 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2401 	}
2402 
2403 	buf_track(bp, __func__);
2404 
2405 	/*
2406 	 * Set the *dirty* buffer range based upon the VM system dirty
2407 	 * pages.
2408 	 *
2409 	 * Mark the buffer pages as clean.  We need to do this here to
2410 	 * satisfy the vnode_pager and the pageout daemon, so that it
2411 	 * thinks that the pages have been "cleaned".  Note that since
2412 	 * the pages are in a delayed write buffer -- the VFS layer
2413 	 * "will" see that the pages get written out on the next sync,
2414 	 * or perhaps the cluster will be completed.
2415 	 */
2416 	vfs_clean_pages_dirty_buf(bp);
2417 	bqrelse(bp);
2418 
2419 	/*
2420 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2421 	 * due to the softdep code.
2422 	 */
2423 }
2424 
2425 /*
2426  *	bdirty:
2427  *
2428  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2429  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2430  *	itself to properly update it in the dirty/clean lists.  We mark it
2431  *	B_DONE to ensure that any asynchronization of the buffer properly
2432  *	clears B_DONE ( else a panic will occur later ).
2433  *
2434  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2435  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2436  *	should only be called if the buffer is known-good.
2437  *
2438  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2439  *	count.
2440  *
2441  *	The buffer must be on QUEUE_NONE.
2442  */
2443 void
2444 bdirty(struct buf *bp)
2445 {
2446 
2447 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2448 	    bp, bp->b_vp, bp->b_flags);
2449 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2450 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2451 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2452 	bp->b_flags &= ~(B_RELBUF);
2453 	bp->b_iocmd = BIO_WRITE;
2454 
2455 	if ((bp->b_flags & B_DELWRI) == 0) {
2456 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2457 		reassignbuf(bp);
2458 		bdirtyadd(bp);
2459 	}
2460 }
2461 
2462 /*
2463  *	bundirty:
2464  *
2465  *	Clear B_DELWRI for buffer.
2466  *
2467  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2468  *	count.
2469  *
2470  *	The buffer must be on QUEUE_NONE.
2471  */
2472 
2473 void
2474 bundirty(struct buf *bp)
2475 {
2476 
2477 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2478 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2479 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2480 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2481 
2482 	if (bp->b_flags & B_DELWRI) {
2483 		bp->b_flags &= ~B_DELWRI;
2484 		reassignbuf(bp);
2485 		bdirtysub(bp);
2486 	}
2487 	/*
2488 	 * Since it is now being written, we can clear its deferred write flag.
2489 	 */
2490 	bp->b_flags &= ~B_DEFERRED;
2491 }
2492 
2493 /*
2494  *	bawrite:
2495  *
2496  *	Asynchronous write.  Start output on a buffer, but do not wait for
2497  *	it to complete.  The buffer is released when the output completes.
2498  *
2499  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2500  *	B_INVAL buffers.  Not us.
2501  */
2502 void
2503 bawrite(struct buf *bp)
2504 {
2505 
2506 	bp->b_flags |= B_ASYNC;
2507 	(void) bwrite(bp);
2508 }
2509 
2510 /*
2511  *	babarrierwrite:
2512  *
2513  *	Asynchronous barrier write.  Start output on a buffer, but do not
2514  *	wait for it to complete.  Place a write barrier after this write so
2515  *	that this buffer and all buffers written before it are committed to
2516  *	the disk before any buffers written after this write are committed
2517  *	to the disk.  The buffer is released when the output completes.
2518  */
2519 void
2520 babarrierwrite(struct buf *bp)
2521 {
2522 
2523 	bp->b_flags |= B_ASYNC | B_BARRIER;
2524 	(void) bwrite(bp);
2525 }
2526 
2527 /*
2528  *	bbarrierwrite:
2529  *
2530  *	Synchronous barrier write.  Start output on a buffer and wait for
2531  *	it to complete.  Place a write barrier after this write so that
2532  *	this buffer and all buffers written before it are committed to
2533  *	the disk before any buffers written after this write are committed
2534  *	to the disk.  The buffer is released when the output completes.
2535  */
2536 int
2537 bbarrierwrite(struct buf *bp)
2538 {
2539 
2540 	bp->b_flags |= B_BARRIER;
2541 	return (bwrite(bp));
2542 }
2543 
2544 /*
2545  *	bwillwrite:
2546  *
2547  *	Called prior to the locking of any vnodes when we are expecting to
2548  *	write.  We do not want to starve the buffer cache with too many
2549  *	dirty buffers so we block here.  By blocking prior to the locking
2550  *	of any vnodes we attempt to avoid the situation where a locked vnode
2551  *	prevents the various system daemons from flushing related buffers.
2552  */
2553 void
2554 bwillwrite(void)
2555 {
2556 
2557 	if (buf_dirty_count_severe()) {
2558 		mtx_lock(&bdirtylock);
2559 		while (buf_dirty_count_severe()) {
2560 			bdirtywait = 1;
2561 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2562 			    "flswai", 0);
2563 		}
2564 		mtx_unlock(&bdirtylock);
2565 	}
2566 }
2567 
2568 /*
2569  * Return true if we have too many dirty buffers.
2570  */
2571 int
2572 buf_dirty_count_severe(void)
2573 {
2574 
2575 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2576 }
2577 
2578 /*
2579  *	brelse:
2580  *
2581  *	Release a busy buffer and, if requested, free its resources.  The
2582  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2583  *	to be accessed later as a cache entity or reused for other purposes.
2584  */
2585 void
2586 brelse(struct buf *bp)
2587 {
2588 	struct mount *v_mnt;
2589 	int qindex;
2590 
2591 	/*
2592 	 * Many functions erroneously call brelse with a NULL bp under rare
2593 	 * error conditions. Simply return when called with a NULL bp.
2594 	 */
2595 	if (bp == NULL)
2596 		return;
2597 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2598 	    bp, bp->b_vp, bp->b_flags);
2599 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2600 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2601 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2602 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2603 
2604 	if (BUF_LOCKRECURSED(bp)) {
2605 		/*
2606 		 * Do not process, in particular, do not handle the
2607 		 * B_INVAL/B_RELBUF and do not release to free list.
2608 		 */
2609 		BUF_UNLOCK(bp);
2610 		return;
2611 	}
2612 
2613 	if (bp->b_flags & B_MANAGED) {
2614 		bqrelse(bp);
2615 		return;
2616 	}
2617 
2618 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2619 		BO_LOCK(bp->b_bufobj);
2620 		bp->b_vflags &= ~BV_BKGRDERR;
2621 		BO_UNLOCK(bp->b_bufobj);
2622 		bdirty(bp);
2623 	}
2624 
2625 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2626 	    (bp->b_flags & B_INVALONERR)) {
2627 		/*
2628 		 * Forced invalidation of dirty buffer contents, to be used
2629 		 * after a failed write in the rare case that the loss of the
2630 		 * contents is acceptable.  The buffer is invalidated and
2631 		 * freed.
2632 		 */
2633 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2634 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2635 	}
2636 
2637 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2638 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2639 	    !(bp->b_flags & B_INVAL)) {
2640 		/*
2641 		 * Failed write, redirty.  All errors except ENXIO (which
2642 		 * means the device is gone) are treated as being
2643 		 * transient.
2644 		 *
2645 		 * XXX Treating EIO as transient is not correct; the
2646 		 * contract with the local storage device drivers is that
2647 		 * they will only return EIO once the I/O is no longer
2648 		 * retriable.  Network I/O also respects this through the
2649 		 * guarantees of TCP and/or the internal retries of NFS.
2650 		 * ENOMEM might be transient, but we also have no way of
2651 		 * knowing when its ok to retry/reschedule.  In general,
2652 		 * this entire case should be made obsolete through better
2653 		 * error handling/recovery and resource scheduling.
2654 		 *
2655 		 * Do this also for buffers that failed with ENXIO, but have
2656 		 * non-empty dependencies - the soft updates code might need
2657 		 * to access the buffer to untangle them.
2658 		 *
2659 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2660 		 */
2661 		bp->b_ioflags &= ~BIO_ERROR;
2662 		bdirty(bp);
2663 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2664 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2665 		/*
2666 		 * Either a failed read I/O, or we were asked to free or not
2667 		 * cache the buffer, or we failed to write to a device that's
2668 		 * no longer present.
2669 		 */
2670 		bp->b_flags |= B_INVAL;
2671 		if (!LIST_EMPTY(&bp->b_dep))
2672 			buf_deallocate(bp);
2673 		if (bp->b_flags & B_DELWRI)
2674 			bdirtysub(bp);
2675 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2676 		if ((bp->b_flags & B_VMIO) == 0) {
2677 			allocbuf(bp, 0);
2678 			if (bp->b_vp)
2679 				brelvp(bp);
2680 		}
2681 	}
2682 
2683 	/*
2684 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2685 	 * is called with B_DELWRI set, the underlying pages may wind up
2686 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2687 	 * because pages associated with a B_DELWRI bp are marked clean.
2688 	 *
2689 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2690 	 * if B_DELWRI is set.
2691 	 */
2692 	if (bp->b_flags & B_DELWRI)
2693 		bp->b_flags &= ~B_RELBUF;
2694 
2695 	/*
2696 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2697 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2698 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2699 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2700 	 *
2701 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2702 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2703 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2704 	 *
2705 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2706 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2707 	 * the commit state and we cannot afford to lose the buffer. If the
2708 	 * buffer has a background write in progress, we need to keep it
2709 	 * around to prevent it from being reconstituted and starting a second
2710 	 * background write.
2711 	 */
2712 
2713 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2714 
2715 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2716 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2717 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2718 	    vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2719 		vfs_vmio_invalidate(bp);
2720 		allocbuf(bp, 0);
2721 	}
2722 
2723 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2724 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2725 		allocbuf(bp, 0);
2726 		bp->b_flags &= ~B_NOREUSE;
2727 		if (bp->b_vp != NULL)
2728 			brelvp(bp);
2729 	}
2730 
2731 	/*
2732 	 * If the buffer has junk contents signal it and eventually
2733 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2734 	 * doesn't find it.
2735 	 */
2736 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2737 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2738 		bp->b_flags |= B_INVAL;
2739 	if (bp->b_flags & B_INVAL) {
2740 		if (bp->b_flags & B_DELWRI)
2741 			bundirty(bp);
2742 		if (bp->b_vp)
2743 			brelvp(bp);
2744 	}
2745 
2746 	buf_track(bp, __func__);
2747 
2748 	/* buffers with no memory */
2749 	if (bp->b_bufsize == 0) {
2750 		buf_free(bp);
2751 		return;
2752 	}
2753 	/* buffers with junk contents */
2754 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2755 	    (bp->b_ioflags & BIO_ERROR)) {
2756 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2757 		if (bp->b_vflags & BV_BKGRDINPROG)
2758 			panic("losing buffer 2");
2759 		qindex = QUEUE_CLEAN;
2760 		bp->b_flags |= B_AGE;
2761 	/* remaining buffers */
2762 	} else if (bp->b_flags & B_DELWRI)
2763 		qindex = QUEUE_DIRTY;
2764 	else
2765 		qindex = QUEUE_CLEAN;
2766 
2767 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2768 		panic("brelse: not dirty");
2769 
2770 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2771 	/* binsfree unlocks bp. */
2772 	binsfree(bp, qindex);
2773 }
2774 
2775 /*
2776  * Release a buffer back to the appropriate queue but do not try to free
2777  * it.  The buffer is expected to be used again soon.
2778  *
2779  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2780  * biodone() to requeue an async I/O on completion.  It is also used when
2781  * known good buffers need to be requeued but we think we may need the data
2782  * again soon.
2783  *
2784  * XXX we should be able to leave the B_RELBUF hint set on completion.
2785  */
2786 void
2787 bqrelse(struct buf *bp)
2788 {
2789 	int qindex;
2790 
2791 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2792 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2793 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2794 
2795 	qindex = QUEUE_NONE;
2796 	if (BUF_LOCKRECURSED(bp)) {
2797 		/* do not release to free list */
2798 		BUF_UNLOCK(bp);
2799 		return;
2800 	}
2801 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2802 
2803 	if (bp->b_flags & B_MANAGED) {
2804 		if (bp->b_flags & B_REMFREE)
2805 			bremfreef(bp);
2806 		goto out;
2807 	}
2808 
2809 	/* buffers with stale but valid contents */
2810 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2811 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2812 		BO_LOCK(bp->b_bufobj);
2813 		bp->b_vflags &= ~BV_BKGRDERR;
2814 		BO_UNLOCK(bp->b_bufobj);
2815 		qindex = QUEUE_DIRTY;
2816 	} else {
2817 		if ((bp->b_flags & B_DELWRI) == 0 &&
2818 		    (bp->b_xflags & BX_VNDIRTY))
2819 			panic("bqrelse: not dirty");
2820 		if ((bp->b_flags & B_NOREUSE) != 0) {
2821 			brelse(bp);
2822 			return;
2823 		}
2824 		qindex = QUEUE_CLEAN;
2825 	}
2826 	buf_track(bp, __func__);
2827 	/* binsfree unlocks bp. */
2828 	binsfree(bp, qindex);
2829 	return;
2830 
2831 out:
2832 	buf_track(bp, __func__);
2833 	/* unlock */
2834 	BUF_UNLOCK(bp);
2835 }
2836 
2837 /*
2838  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2839  * restore bogus pages.
2840  */
2841 static void
2842 vfs_vmio_iodone(struct buf *bp)
2843 {
2844 	vm_ooffset_t foff;
2845 	vm_page_t m;
2846 	vm_object_t obj;
2847 	struct vnode *vp __unused;
2848 	int i, iosize, resid;
2849 	bool bogus;
2850 
2851 	obj = bp->b_bufobj->bo_object;
2852 	KASSERT(REFCOUNT_COUNT(obj->paging_in_progress) >= bp->b_npages,
2853 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2854 	    REFCOUNT_COUNT(obj->paging_in_progress), bp->b_npages));
2855 
2856 	vp = bp->b_vp;
2857 	KASSERT(vp->v_holdcnt > 0,
2858 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2859 	KASSERT(vp->v_object != NULL,
2860 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2861 
2862 	foff = bp->b_offset;
2863 	KASSERT(bp->b_offset != NOOFFSET,
2864 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2865 
2866 	bogus = false;
2867 	iosize = bp->b_bcount - bp->b_resid;
2868 	VM_OBJECT_WLOCK(obj);
2869 	for (i = 0; i < bp->b_npages; i++) {
2870 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2871 		if (resid > iosize)
2872 			resid = iosize;
2873 
2874 		/*
2875 		 * cleanup bogus pages, restoring the originals
2876 		 */
2877 		m = bp->b_pages[i];
2878 		if (m == bogus_page) {
2879 			bogus = true;
2880 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2881 			if (m == NULL)
2882 				panic("biodone: page disappeared!");
2883 			bp->b_pages[i] = m;
2884 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2885 			/*
2886 			 * In the write case, the valid and clean bits are
2887 			 * already changed correctly ( see bdwrite() ), so we
2888 			 * only need to do this here in the read case.
2889 			 */
2890 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2891 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2892 			    "has unexpected dirty bits", m));
2893 			vfs_page_set_valid(bp, foff, m);
2894 		}
2895 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2896 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2897 		    (intmax_t)foff, (uintmax_t)m->pindex));
2898 
2899 		vm_page_sunbusy(m);
2900 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2901 		iosize -= resid;
2902 	}
2903 	vm_object_pip_wakeupn(obj, bp->b_npages);
2904 	VM_OBJECT_WUNLOCK(obj);
2905 	if (bogus && buf_mapped(bp)) {
2906 		BUF_CHECK_MAPPED(bp);
2907 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2908 		    bp->b_pages, bp->b_npages);
2909 	}
2910 }
2911 
2912 /*
2913  * Perform page invalidation when a buffer is released.  The fully invalid
2914  * pages will be reclaimed later in vfs_vmio_truncate().
2915  */
2916 static void
2917 vfs_vmio_invalidate(struct buf *bp)
2918 {
2919 	vm_object_t obj;
2920 	vm_page_t m;
2921 	int flags, i, resid, poffset, presid;
2922 
2923 	if (buf_mapped(bp)) {
2924 		BUF_CHECK_MAPPED(bp);
2925 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2926 	} else
2927 		BUF_CHECK_UNMAPPED(bp);
2928 	/*
2929 	 * Get the base offset and length of the buffer.  Note that
2930 	 * in the VMIO case if the buffer block size is not
2931 	 * page-aligned then b_data pointer may not be page-aligned.
2932 	 * But our b_pages[] array *IS* page aligned.
2933 	 *
2934 	 * block sizes less then DEV_BSIZE (usually 512) are not
2935 	 * supported due to the page granularity bits (m->valid,
2936 	 * m->dirty, etc...).
2937 	 *
2938 	 * See man buf(9) for more information
2939 	 */
2940 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2941 	obj = bp->b_bufobj->bo_object;
2942 	resid = bp->b_bufsize;
2943 	poffset = bp->b_offset & PAGE_MASK;
2944 	VM_OBJECT_WLOCK(obj);
2945 	for (i = 0; i < bp->b_npages; i++) {
2946 		m = bp->b_pages[i];
2947 		if (m == bogus_page)
2948 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2949 		bp->b_pages[i] = NULL;
2950 
2951 		presid = resid > (PAGE_SIZE - poffset) ?
2952 		    (PAGE_SIZE - poffset) : resid;
2953 		KASSERT(presid >= 0, ("brelse: extra page"));
2954 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
2955 		if (pmap_page_wired_mappings(m) == 0)
2956 			vm_page_set_invalid(m, poffset, presid);
2957 		vm_page_sunbusy(m);
2958 		vm_page_release_locked(m, flags);
2959 		resid -= presid;
2960 		poffset = 0;
2961 	}
2962 	VM_OBJECT_WUNLOCK(obj);
2963 	bp->b_npages = 0;
2964 }
2965 
2966 /*
2967  * Page-granular truncation of an existing VMIO buffer.
2968  */
2969 static void
2970 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2971 {
2972 	vm_object_t obj;
2973 	vm_page_t m;
2974 	int flags, i;
2975 
2976 	if (bp->b_npages == desiredpages)
2977 		return;
2978 
2979 	if (buf_mapped(bp)) {
2980 		BUF_CHECK_MAPPED(bp);
2981 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2982 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2983 	} else
2984 		BUF_CHECK_UNMAPPED(bp);
2985 
2986 	/*
2987 	 * The object lock is needed only if we will attempt to free pages.
2988 	 */
2989 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2990 	if ((bp->b_flags & B_DIRECT) != 0) {
2991 		flags |= VPR_TRYFREE;
2992 		obj = bp->b_bufobj->bo_object;
2993 		VM_OBJECT_WLOCK(obj);
2994 	} else {
2995 		obj = NULL;
2996 	}
2997 	for (i = desiredpages; i < bp->b_npages; i++) {
2998 		m = bp->b_pages[i];
2999 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3000 		bp->b_pages[i] = NULL;
3001 		if (obj != NULL)
3002 			vm_page_release_locked(m, flags);
3003 		else
3004 			vm_page_release(m, flags);
3005 	}
3006 	if (obj != NULL)
3007 		VM_OBJECT_WUNLOCK(obj);
3008 	bp->b_npages = desiredpages;
3009 }
3010 
3011 /*
3012  * Byte granular extension of VMIO buffers.
3013  */
3014 static void
3015 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3016 {
3017 	/*
3018 	 * We are growing the buffer, possibly in a
3019 	 * byte-granular fashion.
3020 	 */
3021 	vm_object_t obj;
3022 	vm_offset_t toff;
3023 	vm_offset_t tinc;
3024 	vm_page_t m;
3025 
3026 	/*
3027 	 * Step 1, bring in the VM pages from the object, allocating
3028 	 * them if necessary.  We must clear B_CACHE if these pages
3029 	 * are not valid for the range covered by the buffer.
3030 	 */
3031 	obj = bp->b_bufobj->bo_object;
3032 	VM_OBJECT_WLOCK(obj);
3033 	if (bp->b_npages < desiredpages) {
3034 		/*
3035 		 * We must allocate system pages since blocking
3036 		 * here could interfere with paging I/O, no
3037 		 * matter which process we are.
3038 		 *
3039 		 * Only exclusive busy can be tested here.
3040 		 * Blocking on shared busy might lead to
3041 		 * deadlocks once allocbuf() is called after
3042 		 * pages are vfs_busy_pages().
3043 		 */
3044 		(void)vm_page_grab_pages(obj,
3045 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3046 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3047 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3048 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3049 		bp->b_npages = desiredpages;
3050 	}
3051 
3052 	/*
3053 	 * Step 2.  We've loaded the pages into the buffer,
3054 	 * we have to figure out if we can still have B_CACHE
3055 	 * set.  Note that B_CACHE is set according to the
3056 	 * byte-granular range ( bcount and size ), not the
3057 	 * aligned range ( newbsize ).
3058 	 *
3059 	 * The VM test is against m->valid, which is DEV_BSIZE
3060 	 * aligned.  Needless to say, the validity of the data
3061 	 * needs to also be DEV_BSIZE aligned.  Note that this
3062 	 * fails with NFS if the server or some other client
3063 	 * extends the file's EOF.  If our buffer is resized,
3064 	 * B_CACHE may remain set! XXX
3065 	 */
3066 	toff = bp->b_bcount;
3067 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3068 	while ((bp->b_flags & B_CACHE) && toff < size) {
3069 		vm_pindex_t pi;
3070 
3071 		if (tinc > (size - toff))
3072 			tinc = size - toff;
3073 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3074 		m = bp->b_pages[pi];
3075 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3076 		toff += tinc;
3077 		tinc = PAGE_SIZE;
3078 	}
3079 	VM_OBJECT_WUNLOCK(obj);
3080 
3081 	/*
3082 	 * Step 3, fixup the KVA pmap.
3083 	 */
3084 	if (buf_mapped(bp))
3085 		bpmap_qenter(bp);
3086 	else
3087 		BUF_CHECK_UNMAPPED(bp);
3088 }
3089 
3090 /*
3091  * Check to see if a block at a particular lbn is available for a clustered
3092  * write.
3093  */
3094 static int
3095 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3096 {
3097 	struct buf *bpa;
3098 	int match;
3099 
3100 	match = 0;
3101 
3102 	/* If the buf isn't in core skip it */
3103 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3104 		return (0);
3105 
3106 	/* If the buf is busy we don't want to wait for it */
3107 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3108 		return (0);
3109 
3110 	/* Only cluster with valid clusterable delayed write buffers */
3111 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3112 	    (B_DELWRI | B_CLUSTEROK))
3113 		goto done;
3114 
3115 	if (bpa->b_bufsize != size)
3116 		goto done;
3117 
3118 	/*
3119 	 * Check to see if it is in the expected place on disk and that the
3120 	 * block has been mapped.
3121 	 */
3122 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3123 		match = 1;
3124 done:
3125 	BUF_UNLOCK(bpa);
3126 	return (match);
3127 }
3128 
3129 /*
3130  *	vfs_bio_awrite:
3131  *
3132  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3133  *	This is much better then the old way of writing only one buffer at
3134  *	a time.  Note that we may not be presented with the buffers in the
3135  *	correct order, so we search for the cluster in both directions.
3136  */
3137 int
3138 vfs_bio_awrite(struct buf *bp)
3139 {
3140 	struct bufobj *bo;
3141 	int i;
3142 	int j;
3143 	daddr_t lblkno = bp->b_lblkno;
3144 	struct vnode *vp = bp->b_vp;
3145 	int ncl;
3146 	int nwritten;
3147 	int size;
3148 	int maxcl;
3149 	int gbflags;
3150 
3151 	bo = &vp->v_bufobj;
3152 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3153 	/*
3154 	 * right now we support clustered writing only to regular files.  If
3155 	 * we find a clusterable block we could be in the middle of a cluster
3156 	 * rather then at the beginning.
3157 	 */
3158 	if ((vp->v_type == VREG) &&
3159 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3160 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3161 
3162 		size = vp->v_mount->mnt_stat.f_iosize;
3163 		maxcl = MAXPHYS / size;
3164 
3165 		BO_RLOCK(bo);
3166 		for (i = 1; i < maxcl; i++)
3167 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3168 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3169 				break;
3170 
3171 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3172 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3173 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3174 				break;
3175 		BO_RUNLOCK(bo);
3176 		--j;
3177 		ncl = i + j;
3178 		/*
3179 		 * this is a possible cluster write
3180 		 */
3181 		if (ncl != 1) {
3182 			BUF_UNLOCK(bp);
3183 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3184 			    gbflags);
3185 			return (nwritten);
3186 		}
3187 	}
3188 	bremfree(bp);
3189 	bp->b_flags |= B_ASYNC;
3190 	/*
3191 	 * default (old) behavior, writing out only one block
3192 	 *
3193 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3194 	 */
3195 	nwritten = bp->b_bufsize;
3196 	(void) bwrite(bp);
3197 
3198 	return (nwritten);
3199 }
3200 
3201 /*
3202  *	getnewbuf_kva:
3203  *
3204  *	Allocate KVA for an empty buf header according to gbflags.
3205  */
3206 static int
3207 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3208 {
3209 
3210 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3211 		/*
3212 		 * In order to keep fragmentation sane we only allocate kva
3213 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3214 		 */
3215 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3216 
3217 		if (maxsize != bp->b_kvasize &&
3218 		    bufkva_alloc(bp, maxsize, gbflags))
3219 			return (ENOSPC);
3220 	}
3221 	return (0);
3222 }
3223 
3224 /*
3225  *	getnewbuf:
3226  *
3227  *	Find and initialize a new buffer header, freeing up existing buffers
3228  *	in the bufqueues as necessary.  The new buffer is returned locked.
3229  *
3230  *	We block if:
3231  *		We have insufficient buffer headers
3232  *		We have insufficient buffer space
3233  *		buffer_arena is too fragmented ( space reservation fails )
3234  *		If we have to flush dirty buffers ( but we try to avoid this )
3235  *
3236  *	The caller is responsible for releasing the reserved bufspace after
3237  *	allocbuf() is called.
3238  */
3239 static struct buf *
3240 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3241 {
3242 	struct bufdomain *bd;
3243 	struct buf *bp;
3244 	bool metadata, reserved;
3245 
3246 	bp = NULL;
3247 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3248 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3249 	if (!unmapped_buf_allowed)
3250 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3251 
3252 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3253 	    vp->v_type == VCHR)
3254 		metadata = true;
3255 	else
3256 		metadata = false;
3257 	if (vp == NULL)
3258 		bd = &bdomain[0];
3259 	else
3260 		bd = &bdomain[vp->v_bufobj.bo_domain];
3261 
3262 	counter_u64_add(getnewbufcalls, 1);
3263 	reserved = false;
3264 	do {
3265 		if (reserved == false &&
3266 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3267 			counter_u64_add(getnewbufrestarts, 1);
3268 			continue;
3269 		}
3270 		reserved = true;
3271 		if ((bp = buf_alloc(bd)) == NULL) {
3272 			counter_u64_add(getnewbufrestarts, 1);
3273 			continue;
3274 		}
3275 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3276 			return (bp);
3277 		break;
3278 	} while (buf_recycle(bd, false) == 0);
3279 
3280 	if (reserved)
3281 		bufspace_release(bd, maxsize);
3282 	if (bp != NULL) {
3283 		bp->b_flags |= B_INVAL;
3284 		brelse(bp);
3285 	}
3286 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3287 
3288 	return (NULL);
3289 }
3290 
3291 /*
3292  *	buf_daemon:
3293  *
3294  *	buffer flushing daemon.  Buffers are normally flushed by the
3295  *	update daemon but if it cannot keep up this process starts to
3296  *	take the load in an attempt to prevent getnewbuf() from blocking.
3297  */
3298 static struct kproc_desc buf_kp = {
3299 	"bufdaemon",
3300 	buf_daemon,
3301 	&bufdaemonproc
3302 };
3303 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3304 
3305 static int
3306 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3307 {
3308 	int flushed;
3309 
3310 	flushed = flushbufqueues(vp, bd, target, 0);
3311 	if (flushed == 0) {
3312 		/*
3313 		 * Could not find any buffers without rollback
3314 		 * dependencies, so just write the first one
3315 		 * in the hopes of eventually making progress.
3316 		 */
3317 		if (vp != NULL && target > 2)
3318 			target /= 2;
3319 		flushbufqueues(vp, bd, target, 1);
3320 	}
3321 	return (flushed);
3322 }
3323 
3324 static void
3325 buf_daemon()
3326 {
3327 	struct bufdomain *bd;
3328 	int speedupreq;
3329 	int lodirty;
3330 	int i;
3331 
3332 	/*
3333 	 * This process needs to be suspended prior to shutdown sync.
3334 	 */
3335 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3336 	    SHUTDOWN_PRI_LAST + 100);
3337 
3338 	/*
3339 	 * Start the buf clean daemons as children threads.
3340 	 */
3341 	for (i = 0 ; i < buf_domains; i++) {
3342 		int error;
3343 
3344 		error = kthread_add((void (*)(void *))bufspace_daemon,
3345 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3346 		if (error)
3347 			panic("error %d spawning bufspace daemon", error);
3348 	}
3349 
3350 	/*
3351 	 * This process is allowed to take the buffer cache to the limit
3352 	 */
3353 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3354 	mtx_lock(&bdlock);
3355 	for (;;) {
3356 		bd_request = 0;
3357 		mtx_unlock(&bdlock);
3358 
3359 		kthread_suspend_check();
3360 
3361 		/*
3362 		 * Save speedupreq for this pass and reset to capture new
3363 		 * requests.
3364 		 */
3365 		speedupreq = bd_speedupreq;
3366 		bd_speedupreq = 0;
3367 
3368 		/*
3369 		 * Flush each domain sequentially according to its level and
3370 		 * the speedup request.
3371 		 */
3372 		for (i = 0; i < buf_domains; i++) {
3373 			bd = &bdomain[i];
3374 			if (speedupreq)
3375 				lodirty = bd->bd_numdirtybuffers / 2;
3376 			else
3377 				lodirty = bd->bd_lodirtybuffers;
3378 			while (bd->bd_numdirtybuffers > lodirty) {
3379 				if (buf_flush(NULL, bd,
3380 				    bd->bd_numdirtybuffers - lodirty) == 0)
3381 					break;
3382 				kern_yield(PRI_USER);
3383 			}
3384 		}
3385 
3386 		/*
3387 		 * Only clear bd_request if we have reached our low water
3388 		 * mark.  The buf_daemon normally waits 1 second and
3389 		 * then incrementally flushes any dirty buffers that have
3390 		 * built up, within reason.
3391 		 *
3392 		 * If we were unable to hit our low water mark and couldn't
3393 		 * find any flushable buffers, we sleep for a short period
3394 		 * to avoid endless loops on unlockable buffers.
3395 		 */
3396 		mtx_lock(&bdlock);
3397 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3398 			/*
3399 			 * We reached our low water mark, reset the
3400 			 * request and sleep until we are needed again.
3401 			 * The sleep is just so the suspend code works.
3402 			 */
3403 			bd_request = 0;
3404 			/*
3405 			 * Do an extra wakeup in case dirty threshold
3406 			 * changed via sysctl and the explicit transition
3407 			 * out of shortfall was missed.
3408 			 */
3409 			bdirtywakeup();
3410 			if (runningbufspace <= lorunningspace)
3411 				runningwakeup();
3412 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3413 		} else {
3414 			/*
3415 			 * We couldn't find any flushable dirty buffers but
3416 			 * still have too many dirty buffers, we
3417 			 * have to sleep and try again.  (rare)
3418 			 */
3419 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3420 		}
3421 	}
3422 }
3423 
3424 /*
3425  *	flushbufqueues:
3426  *
3427  *	Try to flush a buffer in the dirty queue.  We must be careful to
3428  *	free up B_INVAL buffers instead of write them, which NFS is
3429  *	particularly sensitive to.
3430  */
3431 static int flushwithdeps = 0;
3432 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3433     &flushwithdeps, 0,
3434     "Number of buffers flushed with dependecies that require rollbacks");
3435 
3436 static int
3437 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3438     int flushdeps)
3439 {
3440 	struct bufqueue *bq;
3441 	struct buf *sentinel;
3442 	struct vnode *vp;
3443 	struct mount *mp;
3444 	struct buf *bp;
3445 	int hasdeps;
3446 	int flushed;
3447 	int error;
3448 	bool unlock;
3449 
3450 	flushed = 0;
3451 	bq = &bd->bd_dirtyq;
3452 	bp = NULL;
3453 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3454 	sentinel->b_qindex = QUEUE_SENTINEL;
3455 	BQ_LOCK(bq);
3456 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3457 	BQ_UNLOCK(bq);
3458 	while (flushed != target) {
3459 		maybe_yield();
3460 		BQ_LOCK(bq);
3461 		bp = TAILQ_NEXT(sentinel, b_freelist);
3462 		if (bp != NULL) {
3463 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3464 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3465 			    b_freelist);
3466 		} else {
3467 			BQ_UNLOCK(bq);
3468 			break;
3469 		}
3470 		/*
3471 		 * Skip sentinels inserted by other invocations of the
3472 		 * flushbufqueues(), taking care to not reorder them.
3473 		 *
3474 		 * Only flush the buffers that belong to the
3475 		 * vnode locked by the curthread.
3476 		 */
3477 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3478 		    bp->b_vp != lvp)) {
3479 			BQ_UNLOCK(bq);
3480 			continue;
3481 		}
3482 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3483 		BQ_UNLOCK(bq);
3484 		if (error != 0)
3485 			continue;
3486 
3487 		/*
3488 		 * BKGRDINPROG can only be set with the buf and bufobj
3489 		 * locks both held.  We tolerate a race to clear it here.
3490 		 */
3491 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3492 		    (bp->b_flags & B_DELWRI) == 0) {
3493 			BUF_UNLOCK(bp);
3494 			continue;
3495 		}
3496 		if (bp->b_flags & B_INVAL) {
3497 			bremfreef(bp);
3498 			brelse(bp);
3499 			flushed++;
3500 			continue;
3501 		}
3502 
3503 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3504 			if (flushdeps == 0) {
3505 				BUF_UNLOCK(bp);
3506 				continue;
3507 			}
3508 			hasdeps = 1;
3509 		} else
3510 			hasdeps = 0;
3511 		/*
3512 		 * We must hold the lock on a vnode before writing
3513 		 * one of its buffers. Otherwise we may confuse, or
3514 		 * in the case of a snapshot vnode, deadlock the
3515 		 * system.
3516 		 *
3517 		 * The lock order here is the reverse of the normal
3518 		 * of vnode followed by buf lock.  This is ok because
3519 		 * the NOWAIT will prevent deadlock.
3520 		 */
3521 		vp = bp->b_vp;
3522 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3523 			BUF_UNLOCK(bp);
3524 			continue;
3525 		}
3526 		if (lvp == NULL) {
3527 			unlock = true;
3528 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3529 		} else {
3530 			ASSERT_VOP_LOCKED(vp, "getbuf");
3531 			unlock = false;
3532 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3533 			    vn_lock(vp, LK_TRYUPGRADE);
3534 		}
3535 		if (error == 0) {
3536 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3537 			    bp, bp->b_vp, bp->b_flags);
3538 			if (curproc == bufdaemonproc) {
3539 				vfs_bio_awrite(bp);
3540 			} else {
3541 				bremfree(bp);
3542 				bwrite(bp);
3543 				counter_u64_add(notbufdflushes, 1);
3544 			}
3545 			vn_finished_write(mp);
3546 			if (unlock)
3547 				VOP_UNLOCK(vp, 0);
3548 			flushwithdeps += hasdeps;
3549 			flushed++;
3550 
3551 			/*
3552 			 * Sleeping on runningbufspace while holding
3553 			 * vnode lock leads to deadlock.
3554 			 */
3555 			if (curproc == bufdaemonproc &&
3556 			    runningbufspace > hirunningspace)
3557 				waitrunningbufspace();
3558 			continue;
3559 		}
3560 		vn_finished_write(mp);
3561 		BUF_UNLOCK(bp);
3562 	}
3563 	BQ_LOCK(bq);
3564 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3565 	BQ_UNLOCK(bq);
3566 	free(sentinel, M_TEMP);
3567 	return (flushed);
3568 }
3569 
3570 /*
3571  * Check to see if a block is currently memory resident.
3572  */
3573 struct buf *
3574 incore(struct bufobj *bo, daddr_t blkno)
3575 {
3576 	struct buf *bp;
3577 
3578 	BO_RLOCK(bo);
3579 	bp = gbincore(bo, blkno);
3580 	BO_RUNLOCK(bo);
3581 	return (bp);
3582 }
3583 
3584 /*
3585  * Returns true if no I/O is needed to access the
3586  * associated VM object.  This is like incore except
3587  * it also hunts around in the VM system for the data.
3588  */
3589 
3590 static int
3591 inmem(struct vnode * vp, daddr_t blkno)
3592 {
3593 	vm_object_t obj;
3594 	vm_offset_t toff, tinc, size;
3595 	vm_page_t m;
3596 	vm_ooffset_t off;
3597 
3598 	ASSERT_VOP_LOCKED(vp, "inmem");
3599 
3600 	if (incore(&vp->v_bufobj, blkno))
3601 		return 1;
3602 	if (vp->v_mount == NULL)
3603 		return 0;
3604 	obj = vp->v_object;
3605 	if (obj == NULL)
3606 		return (0);
3607 
3608 	size = PAGE_SIZE;
3609 	if (size > vp->v_mount->mnt_stat.f_iosize)
3610 		size = vp->v_mount->mnt_stat.f_iosize;
3611 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3612 
3613 	VM_OBJECT_RLOCK(obj);
3614 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3615 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3616 		if (!m)
3617 			goto notinmem;
3618 		tinc = size;
3619 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3620 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3621 		if (vm_page_is_valid(m,
3622 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3623 			goto notinmem;
3624 	}
3625 	VM_OBJECT_RUNLOCK(obj);
3626 	return 1;
3627 
3628 notinmem:
3629 	VM_OBJECT_RUNLOCK(obj);
3630 	return (0);
3631 }
3632 
3633 /*
3634  * Set the dirty range for a buffer based on the status of the dirty
3635  * bits in the pages comprising the buffer.  The range is limited
3636  * to the size of the buffer.
3637  *
3638  * Tell the VM system that the pages associated with this buffer
3639  * are clean.  This is used for delayed writes where the data is
3640  * going to go to disk eventually without additional VM intevention.
3641  *
3642  * Note that while we only really need to clean through to b_bcount, we
3643  * just go ahead and clean through to b_bufsize.
3644  */
3645 static void
3646 vfs_clean_pages_dirty_buf(struct buf *bp)
3647 {
3648 	vm_ooffset_t foff, noff, eoff;
3649 	vm_page_t m;
3650 	int i;
3651 
3652 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3653 		return;
3654 
3655 	foff = bp->b_offset;
3656 	KASSERT(bp->b_offset != NOOFFSET,
3657 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3658 
3659 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3660 	vfs_busy_pages_acquire(bp);
3661 	vfs_setdirty_locked_object(bp);
3662 	for (i = 0; i < bp->b_npages; i++) {
3663 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3664 		eoff = noff;
3665 		if (eoff > bp->b_offset + bp->b_bufsize)
3666 			eoff = bp->b_offset + bp->b_bufsize;
3667 		m = bp->b_pages[i];
3668 		vfs_page_set_validclean(bp, foff, m);
3669 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3670 		foff = noff;
3671 	}
3672 	vfs_busy_pages_release(bp);
3673 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3674 }
3675 
3676 static void
3677 vfs_setdirty_locked_object(struct buf *bp)
3678 {
3679 	vm_object_t object;
3680 	int i;
3681 
3682 	object = bp->b_bufobj->bo_object;
3683 	VM_OBJECT_ASSERT_WLOCKED(object);
3684 
3685 	/*
3686 	 * We qualify the scan for modified pages on whether the
3687 	 * object has been flushed yet.
3688 	 */
3689 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3690 		vm_offset_t boffset;
3691 		vm_offset_t eoffset;
3692 
3693 		/*
3694 		 * test the pages to see if they have been modified directly
3695 		 * by users through the VM system.
3696 		 */
3697 		for (i = 0; i < bp->b_npages; i++)
3698 			vm_page_test_dirty(bp->b_pages[i]);
3699 
3700 		/*
3701 		 * Calculate the encompassing dirty range, boffset and eoffset,
3702 		 * (eoffset - boffset) bytes.
3703 		 */
3704 
3705 		for (i = 0; i < bp->b_npages; i++) {
3706 			if (bp->b_pages[i]->dirty)
3707 				break;
3708 		}
3709 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3710 
3711 		for (i = bp->b_npages - 1; i >= 0; --i) {
3712 			if (bp->b_pages[i]->dirty) {
3713 				break;
3714 			}
3715 		}
3716 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3717 
3718 		/*
3719 		 * Fit it to the buffer.
3720 		 */
3721 
3722 		if (eoffset > bp->b_bcount)
3723 			eoffset = bp->b_bcount;
3724 
3725 		/*
3726 		 * If we have a good dirty range, merge with the existing
3727 		 * dirty range.
3728 		 */
3729 
3730 		if (boffset < eoffset) {
3731 			if (bp->b_dirtyoff > boffset)
3732 				bp->b_dirtyoff = boffset;
3733 			if (bp->b_dirtyend < eoffset)
3734 				bp->b_dirtyend = eoffset;
3735 		}
3736 	}
3737 }
3738 
3739 /*
3740  * Allocate the KVA mapping for an existing buffer.
3741  * If an unmapped buffer is provided but a mapped buffer is requested, take
3742  * also care to properly setup mappings between pages and KVA.
3743  */
3744 static void
3745 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3746 {
3747 	int bsize, maxsize, need_mapping, need_kva;
3748 	off_t offset;
3749 
3750 	need_mapping = bp->b_data == unmapped_buf &&
3751 	    (gbflags & GB_UNMAPPED) == 0;
3752 	need_kva = bp->b_kvabase == unmapped_buf &&
3753 	    bp->b_data == unmapped_buf &&
3754 	    (gbflags & GB_KVAALLOC) != 0;
3755 	if (!need_mapping && !need_kva)
3756 		return;
3757 
3758 	BUF_CHECK_UNMAPPED(bp);
3759 
3760 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3761 		/*
3762 		 * Buffer is not mapped, but the KVA was already
3763 		 * reserved at the time of the instantiation.  Use the
3764 		 * allocated space.
3765 		 */
3766 		goto has_addr;
3767 	}
3768 
3769 	/*
3770 	 * Calculate the amount of the address space we would reserve
3771 	 * if the buffer was mapped.
3772 	 */
3773 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3774 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3775 	offset = blkno * bsize;
3776 	maxsize = size + (offset & PAGE_MASK);
3777 	maxsize = imax(maxsize, bsize);
3778 
3779 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3780 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3781 			/*
3782 			 * XXXKIB: defragmentation cannot
3783 			 * succeed, not sure what else to do.
3784 			 */
3785 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3786 		}
3787 		counter_u64_add(mappingrestarts, 1);
3788 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3789 	}
3790 has_addr:
3791 	if (need_mapping) {
3792 		/* b_offset is handled by bpmap_qenter. */
3793 		bp->b_data = bp->b_kvabase;
3794 		BUF_CHECK_MAPPED(bp);
3795 		bpmap_qenter(bp);
3796 	}
3797 }
3798 
3799 struct buf *
3800 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3801     int flags)
3802 {
3803 	struct buf *bp;
3804 	int error;
3805 
3806 	error = getblkx(vp, blkno, size, slpflag, slptimeo, flags, &bp);
3807 	if (error != 0)
3808 		return (NULL);
3809 	return (bp);
3810 }
3811 
3812 /*
3813  *	getblkx:
3814  *
3815  *	Get a block given a specified block and offset into a file/device.
3816  *	The buffers B_DONE bit will be cleared on return, making it almost
3817  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3818  *	return.  The caller should clear B_INVAL prior to initiating a
3819  *	READ.
3820  *
3821  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3822  *	an existing buffer.
3823  *
3824  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3825  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3826  *	and then cleared based on the backing VM.  If the previous buffer is
3827  *	non-0-sized but invalid, B_CACHE will be cleared.
3828  *
3829  *	If getblk() must create a new buffer, the new buffer is returned with
3830  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3831  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3832  *	backing VM.
3833  *
3834  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3835  *	B_CACHE bit is clear.
3836  *
3837  *	What this means, basically, is that the caller should use B_CACHE to
3838  *	determine whether the buffer is fully valid or not and should clear
3839  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3840  *	the buffer by loading its data area with something, the caller needs
3841  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3842  *	the caller should set B_CACHE ( as an optimization ), else the caller
3843  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3844  *	a write attempt or if it was a successful read.  If the caller
3845  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3846  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3847  */
3848 int
3849 getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3850     int flags, struct buf **bpp)
3851 {
3852 	struct buf *bp;
3853 	struct bufobj *bo;
3854 	daddr_t d_blkno;
3855 	int bsize, error, maxsize, vmio;
3856 	off_t offset;
3857 
3858 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3859 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3860 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3861 	ASSERT_VOP_LOCKED(vp, "getblk");
3862 	if (size > maxbcachebuf)
3863 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3864 		    maxbcachebuf);
3865 	if (!unmapped_buf_allowed)
3866 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3867 
3868 	bo = &vp->v_bufobj;
3869 	d_blkno = blkno;
3870 loop:
3871 	BO_RLOCK(bo);
3872 	bp = gbincore(bo, blkno);
3873 	if (bp != NULL) {
3874 		int lockflags;
3875 		/*
3876 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3877 		 * it must be on a queue.
3878 		 */
3879 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3880 
3881 		if ((flags & GB_LOCK_NOWAIT) != 0)
3882 			lockflags |= LK_NOWAIT;
3883 
3884 		error = BUF_TIMELOCK(bp, lockflags,
3885 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3886 
3887 		/*
3888 		 * If we slept and got the lock we have to restart in case
3889 		 * the buffer changed identities.
3890 		 */
3891 		if (error == ENOLCK)
3892 			goto loop;
3893 		/* We timed out or were interrupted. */
3894 		else if (error != 0)
3895 			return (error);
3896 		/* If recursed, assume caller knows the rules. */
3897 		else if (BUF_LOCKRECURSED(bp))
3898 			goto end;
3899 
3900 		/*
3901 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3902 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3903 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3904 		 * backing VM cache.
3905 		 */
3906 		if (bp->b_flags & B_INVAL)
3907 			bp->b_flags &= ~B_CACHE;
3908 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3909 			bp->b_flags |= B_CACHE;
3910 		if (bp->b_flags & B_MANAGED)
3911 			MPASS(bp->b_qindex == QUEUE_NONE);
3912 		else
3913 			bremfree(bp);
3914 
3915 		/*
3916 		 * check for size inconsistencies for non-VMIO case.
3917 		 */
3918 		if (bp->b_bcount != size) {
3919 			if ((bp->b_flags & B_VMIO) == 0 ||
3920 			    (size > bp->b_kvasize)) {
3921 				if (bp->b_flags & B_DELWRI) {
3922 					bp->b_flags |= B_NOCACHE;
3923 					bwrite(bp);
3924 				} else {
3925 					if (LIST_EMPTY(&bp->b_dep)) {
3926 						bp->b_flags |= B_RELBUF;
3927 						brelse(bp);
3928 					} else {
3929 						bp->b_flags |= B_NOCACHE;
3930 						bwrite(bp);
3931 					}
3932 				}
3933 				goto loop;
3934 			}
3935 		}
3936 
3937 		/*
3938 		 * Handle the case of unmapped buffer which should
3939 		 * become mapped, or the buffer for which KVA
3940 		 * reservation is requested.
3941 		 */
3942 		bp_unmapped_get_kva(bp, blkno, size, flags);
3943 
3944 		/*
3945 		 * If the size is inconsistent in the VMIO case, we can resize
3946 		 * the buffer.  This might lead to B_CACHE getting set or
3947 		 * cleared.  If the size has not changed, B_CACHE remains
3948 		 * unchanged from its previous state.
3949 		 */
3950 		allocbuf(bp, size);
3951 
3952 		KASSERT(bp->b_offset != NOOFFSET,
3953 		    ("getblk: no buffer offset"));
3954 
3955 		/*
3956 		 * A buffer with B_DELWRI set and B_CACHE clear must
3957 		 * be committed before we can return the buffer in
3958 		 * order to prevent the caller from issuing a read
3959 		 * ( due to B_CACHE not being set ) and overwriting
3960 		 * it.
3961 		 *
3962 		 * Most callers, including NFS and FFS, need this to
3963 		 * operate properly either because they assume they
3964 		 * can issue a read if B_CACHE is not set, or because
3965 		 * ( for example ) an uncached B_DELWRI might loop due
3966 		 * to softupdates re-dirtying the buffer.  In the latter
3967 		 * case, B_CACHE is set after the first write completes,
3968 		 * preventing further loops.
3969 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3970 		 * above while extending the buffer, we cannot allow the
3971 		 * buffer to remain with B_CACHE set after the write
3972 		 * completes or it will represent a corrupt state.  To
3973 		 * deal with this we set B_NOCACHE to scrap the buffer
3974 		 * after the write.
3975 		 *
3976 		 * We might be able to do something fancy, like setting
3977 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3978 		 * so the below call doesn't set B_CACHE, but that gets real
3979 		 * confusing.  This is much easier.
3980 		 */
3981 
3982 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3983 			bp->b_flags |= B_NOCACHE;
3984 			bwrite(bp);
3985 			goto loop;
3986 		}
3987 		bp->b_flags &= ~B_DONE;
3988 	} else {
3989 		/*
3990 		 * Buffer is not in-core, create new buffer.  The buffer
3991 		 * returned by getnewbuf() is locked.  Note that the returned
3992 		 * buffer is also considered valid (not marked B_INVAL).
3993 		 */
3994 		BO_RUNLOCK(bo);
3995 		/*
3996 		 * If the user does not want us to create the buffer, bail out
3997 		 * here.
3998 		 */
3999 		if (flags & GB_NOCREAT)
4000 			return (EEXIST);
4001 
4002 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
4003 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4004 		offset = blkno * bsize;
4005 		vmio = vp->v_object != NULL;
4006 		if (vmio) {
4007 			maxsize = size + (offset & PAGE_MASK);
4008 		} else {
4009 			maxsize = size;
4010 			/* Do not allow non-VMIO notmapped buffers. */
4011 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4012 		}
4013 		maxsize = imax(maxsize, bsize);
4014 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4015 		    !vn_isdisk(vp, NULL)) {
4016 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4017 			KASSERT(error != EOPNOTSUPP,
4018 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4019 			    vp));
4020 			if (error != 0)
4021 				return (error);
4022 			if (d_blkno == -1)
4023 				return (EJUSTRETURN);
4024 		}
4025 
4026 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4027 		if (bp == NULL) {
4028 			if (slpflag || slptimeo)
4029 				return (ETIMEDOUT);
4030 			/*
4031 			 * XXX This is here until the sleep path is diagnosed
4032 			 * enough to work under very low memory conditions.
4033 			 *
4034 			 * There's an issue on low memory, 4BSD+non-preempt
4035 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4036 			 * exhaustion occurs without sleeping for buffer
4037 			 * reclaimation.  This just sticks in a loop and
4038 			 * constantly attempts to allocate a buffer, which
4039 			 * hits exhaustion and tries to wakeup bufdaemon.
4040 			 * This never happens because we never yield.
4041 			 *
4042 			 * The real solution is to identify and fix these cases
4043 			 * so we aren't effectively busy-waiting in a loop
4044 			 * until the reclaimation path has cycles to run.
4045 			 */
4046 			kern_yield(PRI_USER);
4047 			goto loop;
4048 		}
4049 
4050 		/*
4051 		 * This code is used to make sure that a buffer is not
4052 		 * created while the getnewbuf routine is blocked.
4053 		 * This can be a problem whether the vnode is locked or not.
4054 		 * If the buffer is created out from under us, we have to
4055 		 * throw away the one we just created.
4056 		 *
4057 		 * Note: this must occur before we associate the buffer
4058 		 * with the vp especially considering limitations in
4059 		 * the splay tree implementation when dealing with duplicate
4060 		 * lblkno's.
4061 		 */
4062 		BO_LOCK(bo);
4063 		if (gbincore(bo, blkno)) {
4064 			BO_UNLOCK(bo);
4065 			bp->b_flags |= B_INVAL;
4066 			bufspace_release(bufdomain(bp), maxsize);
4067 			brelse(bp);
4068 			goto loop;
4069 		}
4070 
4071 		/*
4072 		 * Insert the buffer into the hash, so that it can
4073 		 * be found by incore.
4074 		 */
4075 		bp->b_lblkno = blkno;
4076 		bp->b_blkno = d_blkno;
4077 		bp->b_offset = offset;
4078 		bgetvp(vp, bp);
4079 		BO_UNLOCK(bo);
4080 
4081 		/*
4082 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4083 		 * buffer size starts out as 0, B_CACHE will be set by
4084 		 * allocbuf() for the VMIO case prior to it testing the
4085 		 * backing store for validity.
4086 		 */
4087 
4088 		if (vmio) {
4089 			bp->b_flags |= B_VMIO;
4090 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4091 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4092 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4093 		} else {
4094 			bp->b_flags &= ~B_VMIO;
4095 			KASSERT(bp->b_bufobj->bo_object == NULL,
4096 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4097 			    bp, bp->b_bufobj->bo_object));
4098 			BUF_CHECK_MAPPED(bp);
4099 		}
4100 
4101 		allocbuf(bp, size);
4102 		bufspace_release(bufdomain(bp), maxsize);
4103 		bp->b_flags &= ~B_DONE;
4104 	}
4105 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4106 end:
4107 	buf_track(bp, __func__);
4108 	KASSERT(bp->b_bufobj == bo,
4109 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4110 	*bpp = bp;
4111 	return (0);
4112 }
4113 
4114 /*
4115  * Get an empty, disassociated buffer of given size.  The buffer is initially
4116  * set to B_INVAL.
4117  */
4118 struct buf *
4119 geteblk(int size, int flags)
4120 {
4121 	struct buf *bp;
4122 	int maxsize;
4123 
4124 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4125 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4126 		if ((flags & GB_NOWAIT_BD) &&
4127 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4128 			return (NULL);
4129 	}
4130 	allocbuf(bp, size);
4131 	bufspace_release(bufdomain(bp), maxsize);
4132 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4133 	return (bp);
4134 }
4135 
4136 /*
4137  * Truncate the backing store for a non-vmio buffer.
4138  */
4139 static void
4140 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4141 {
4142 
4143 	if (bp->b_flags & B_MALLOC) {
4144 		/*
4145 		 * malloced buffers are not shrunk
4146 		 */
4147 		if (newbsize == 0) {
4148 			bufmallocadjust(bp, 0);
4149 			free(bp->b_data, M_BIOBUF);
4150 			bp->b_data = bp->b_kvabase;
4151 			bp->b_flags &= ~B_MALLOC;
4152 		}
4153 		return;
4154 	}
4155 	vm_hold_free_pages(bp, newbsize);
4156 	bufspace_adjust(bp, newbsize);
4157 }
4158 
4159 /*
4160  * Extend the backing for a non-VMIO buffer.
4161  */
4162 static void
4163 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4164 {
4165 	caddr_t origbuf;
4166 	int origbufsize;
4167 
4168 	/*
4169 	 * We only use malloced memory on the first allocation.
4170 	 * and revert to page-allocated memory when the buffer
4171 	 * grows.
4172 	 *
4173 	 * There is a potential smp race here that could lead
4174 	 * to bufmallocspace slightly passing the max.  It
4175 	 * is probably extremely rare and not worth worrying
4176 	 * over.
4177 	 */
4178 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4179 	    bufmallocspace < maxbufmallocspace) {
4180 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4181 		bp->b_flags |= B_MALLOC;
4182 		bufmallocadjust(bp, newbsize);
4183 		return;
4184 	}
4185 
4186 	/*
4187 	 * If the buffer is growing on its other-than-first
4188 	 * allocation then we revert to the page-allocation
4189 	 * scheme.
4190 	 */
4191 	origbuf = NULL;
4192 	origbufsize = 0;
4193 	if (bp->b_flags & B_MALLOC) {
4194 		origbuf = bp->b_data;
4195 		origbufsize = bp->b_bufsize;
4196 		bp->b_data = bp->b_kvabase;
4197 		bufmallocadjust(bp, 0);
4198 		bp->b_flags &= ~B_MALLOC;
4199 		newbsize = round_page(newbsize);
4200 	}
4201 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4202 	    (vm_offset_t) bp->b_data + newbsize);
4203 	if (origbuf != NULL) {
4204 		bcopy(origbuf, bp->b_data, origbufsize);
4205 		free(origbuf, M_BIOBUF);
4206 	}
4207 	bufspace_adjust(bp, newbsize);
4208 }
4209 
4210 /*
4211  * This code constitutes the buffer memory from either anonymous system
4212  * memory (in the case of non-VMIO operations) or from an associated
4213  * VM object (in the case of VMIO operations).  This code is able to
4214  * resize a buffer up or down.
4215  *
4216  * Note that this code is tricky, and has many complications to resolve
4217  * deadlock or inconsistent data situations.  Tread lightly!!!
4218  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4219  * the caller.  Calling this code willy nilly can result in the loss of data.
4220  *
4221  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4222  * B_CACHE for the non-VMIO case.
4223  */
4224 int
4225 allocbuf(struct buf *bp, int size)
4226 {
4227 	int newbsize;
4228 
4229 	if (bp->b_bcount == size)
4230 		return (1);
4231 
4232 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4233 		panic("allocbuf: buffer too small");
4234 
4235 	newbsize = roundup2(size, DEV_BSIZE);
4236 	if ((bp->b_flags & B_VMIO) == 0) {
4237 		if ((bp->b_flags & B_MALLOC) == 0)
4238 			newbsize = round_page(newbsize);
4239 		/*
4240 		 * Just get anonymous memory from the kernel.  Don't
4241 		 * mess with B_CACHE.
4242 		 */
4243 		if (newbsize < bp->b_bufsize)
4244 			vfs_nonvmio_truncate(bp, newbsize);
4245 		else if (newbsize > bp->b_bufsize)
4246 			vfs_nonvmio_extend(bp, newbsize);
4247 	} else {
4248 		int desiredpages;
4249 
4250 		desiredpages = (size == 0) ? 0 :
4251 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4252 
4253 		if (bp->b_flags & B_MALLOC)
4254 			panic("allocbuf: VMIO buffer can't be malloced");
4255 		/*
4256 		 * Set B_CACHE initially if buffer is 0 length or will become
4257 		 * 0-length.
4258 		 */
4259 		if (size == 0 || bp->b_bufsize == 0)
4260 			bp->b_flags |= B_CACHE;
4261 
4262 		if (newbsize < bp->b_bufsize)
4263 			vfs_vmio_truncate(bp, desiredpages);
4264 		/* XXX This looks as if it should be newbsize > b_bufsize */
4265 		else if (size > bp->b_bcount)
4266 			vfs_vmio_extend(bp, desiredpages, size);
4267 		bufspace_adjust(bp, newbsize);
4268 	}
4269 	bp->b_bcount = size;		/* requested buffer size. */
4270 	return (1);
4271 }
4272 
4273 extern int inflight_transient_maps;
4274 
4275 static struct bio_queue nondump_bios;
4276 
4277 void
4278 biodone(struct bio *bp)
4279 {
4280 	struct mtx *mtxp;
4281 	void (*done)(struct bio *);
4282 	vm_offset_t start, end;
4283 
4284 	biotrack(bp, __func__);
4285 
4286 	/*
4287 	 * Avoid completing I/O when dumping after a panic since that may
4288 	 * result in a deadlock in the filesystem or pager code.  Note that
4289 	 * this doesn't affect dumps that were started manually since we aim
4290 	 * to keep the system usable after it has been resumed.
4291 	 */
4292 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4293 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4294 		return;
4295 	}
4296 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4297 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4298 		bp->bio_flags |= BIO_UNMAPPED;
4299 		start = trunc_page((vm_offset_t)bp->bio_data);
4300 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4301 		bp->bio_data = unmapped_buf;
4302 		pmap_qremove(start, atop(end - start));
4303 		vmem_free(transient_arena, start, end - start);
4304 		atomic_add_int(&inflight_transient_maps, -1);
4305 	}
4306 	done = bp->bio_done;
4307 	if (done == NULL) {
4308 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4309 		mtx_lock(mtxp);
4310 		bp->bio_flags |= BIO_DONE;
4311 		wakeup(bp);
4312 		mtx_unlock(mtxp);
4313 	} else
4314 		done(bp);
4315 }
4316 
4317 /*
4318  * Wait for a BIO to finish.
4319  */
4320 int
4321 biowait(struct bio *bp, const char *wchan)
4322 {
4323 	struct mtx *mtxp;
4324 
4325 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4326 	mtx_lock(mtxp);
4327 	while ((bp->bio_flags & BIO_DONE) == 0)
4328 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4329 	mtx_unlock(mtxp);
4330 	if (bp->bio_error != 0)
4331 		return (bp->bio_error);
4332 	if (!(bp->bio_flags & BIO_ERROR))
4333 		return (0);
4334 	return (EIO);
4335 }
4336 
4337 void
4338 biofinish(struct bio *bp, struct devstat *stat, int error)
4339 {
4340 
4341 	if (error) {
4342 		bp->bio_error = error;
4343 		bp->bio_flags |= BIO_ERROR;
4344 	}
4345 	if (stat != NULL)
4346 		devstat_end_transaction_bio(stat, bp);
4347 	biodone(bp);
4348 }
4349 
4350 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4351 void
4352 biotrack_buf(struct bio *bp, const char *location)
4353 {
4354 
4355 	buf_track(bp->bio_track_bp, location);
4356 }
4357 #endif
4358 
4359 /*
4360  *	bufwait:
4361  *
4362  *	Wait for buffer I/O completion, returning error status.  The buffer
4363  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4364  *	error and cleared.
4365  */
4366 int
4367 bufwait(struct buf *bp)
4368 {
4369 	if (bp->b_iocmd == BIO_READ)
4370 		bwait(bp, PRIBIO, "biord");
4371 	else
4372 		bwait(bp, PRIBIO, "biowr");
4373 	if (bp->b_flags & B_EINTR) {
4374 		bp->b_flags &= ~B_EINTR;
4375 		return (EINTR);
4376 	}
4377 	if (bp->b_ioflags & BIO_ERROR) {
4378 		return (bp->b_error ? bp->b_error : EIO);
4379 	} else {
4380 		return (0);
4381 	}
4382 }
4383 
4384 /*
4385  *	bufdone:
4386  *
4387  *	Finish I/O on a buffer, optionally calling a completion function.
4388  *	This is usually called from an interrupt so process blocking is
4389  *	not allowed.
4390  *
4391  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4392  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4393  *	assuming B_INVAL is clear.
4394  *
4395  *	For the VMIO case, we set B_CACHE if the op was a read and no
4396  *	read error occurred, or if the op was a write.  B_CACHE is never
4397  *	set if the buffer is invalid or otherwise uncacheable.
4398  *
4399  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4400  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4401  *	in the biodone routine.
4402  */
4403 void
4404 bufdone(struct buf *bp)
4405 {
4406 	struct bufobj *dropobj;
4407 	void    (*biodone)(struct buf *);
4408 
4409 	buf_track(bp, __func__);
4410 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4411 	dropobj = NULL;
4412 
4413 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4414 
4415 	runningbufwakeup(bp);
4416 	if (bp->b_iocmd == BIO_WRITE)
4417 		dropobj = bp->b_bufobj;
4418 	/* call optional completion function if requested */
4419 	if (bp->b_iodone != NULL) {
4420 		biodone = bp->b_iodone;
4421 		bp->b_iodone = NULL;
4422 		(*biodone) (bp);
4423 		if (dropobj)
4424 			bufobj_wdrop(dropobj);
4425 		return;
4426 	}
4427 	if (bp->b_flags & B_VMIO) {
4428 		/*
4429 		 * Set B_CACHE if the op was a normal read and no error
4430 		 * occurred.  B_CACHE is set for writes in the b*write()
4431 		 * routines.
4432 		 */
4433 		if (bp->b_iocmd == BIO_READ &&
4434 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4435 		    !(bp->b_ioflags & BIO_ERROR))
4436 			bp->b_flags |= B_CACHE;
4437 		vfs_vmio_iodone(bp);
4438 	}
4439 	if (!LIST_EMPTY(&bp->b_dep))
4440 		buf_complete(bp);
4441 	if ((bp->b_flags & B_CKHASH) != 0) {
4442 		KASSERT(bp->b_iocmd == BIO_READ,
4443 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4444 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4445 		(*bp->b_ckhashcalc)(bp);
4446 	}
4447 	/*
4448 	 * For asynchronous completions, release the buffer now. The brelse
4449 	 * will do a wakeup there if necessary - so no need to do a wakeup
4450 	 * here in the async case. The sync case always needs to do a wakeup.
4451 	 */
4452 	if (bp->b_flags & B_ASYNC) {
4453 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4454 		    (bp->b_ioflags & BIO_ERROR))
4455 			brelse(bp);
4456 		else
4457 			bqrelse(bp);
4458 	} else
4459 		bdone(bp);
4460 	if (dropobj)
4461 		bufobj_wdrop(dropobj);
4462 }
4463 
4464 /*
4465  * This routine is called in lieu of iodone in the case of
4466  * incomplete I/O.  This keeps the busy status for pages
4467  * consistent.
4468  */
4469 void
4470 vfs_unbusy_pages(struct buf *bp)
4471 {
4472 	int i;
4473 	vm_object_t obj;
4474 	vm_page_t m;
4475 
4476 	runningbufwakeup(bp);
4477 	if (!(bp->b_flags & B_VMIO))
4478 		return;
4479 
4480 	obj = bp->b_bufobj->bo_object;
4481 	VM_OBJECT_WLOCK(obj);
4482 	for (i = 0; i < bp->b_npages; i++) {
4483 		m = bp->b_pages[i];
4484 		if (m == bogus_page) {
4485 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4486 			if (!m)
4487 				panic("vfs_unbusy_pages: page missing\n");
4488 			bp->b_pages[i] = m;
4489 			if (buf_mapped(bp)) {
4490 				BUF_CHECK_MAPPED(bp);
4491 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4492 				    bp->b_pages, bp->b_npages);
4493 			} else
4494 				BUF_CHECK_UNMAPPED(bp);
4495 		}
4496 		vm_page_sunbusy(m);
4497 	}
4498 	vm_object_pip_wakeupn(obj, bp->b_npages);
4499 	VM_OBJECT_WUNLOCK(obj);
4500 }
4501 
4502 /*
4503  * vfs_page_set_valid:
4504  *
4505  *	Set the valid bits in a page based on the supplied offset.   The
4506  *	range is restricted to the buffer's size.
4507  *
4508  *	This routine is typically called after a read completes.
4509  */
4510 static void
4511 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4512 {
4513 	vm_ooffset_t eoff;
4514 
4515 	/*
4516 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4517 	 * page boundary and eoff is not greater than the end of the buffer.
4518 	 * The end of the buffer, in this case, is our file EOF, not the
4519 	 * allocation size of the buffer.
4520 	 */
4521 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4522 	if (eoff > bp->b_offset + bp->b_bcount)
4523 		eoff = bp->b_offset + bp->b_bcount;
4524 
4525 	/*
4526 	 * Set valid range.  This is typically the entire buffer and thus the
4527 	 * entire page.
4528 	 */
4529 	if (eoff > off)
4530 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4531 }
4532 
4533 /*
4534  * vfs_page_set_validclean:
4535  *
4536  *	Set the valid bits and clear the dirty bits in a page based on the
4537  *	supplied offset.   The range is restricted to the buffer's size.
4538  */
4539 static void
4540 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4541 {
4542 	vm_ooffset_t soff, eoff;
4543 
4544 	/*
4545 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4546 	 * page boundary or cross the end of the buffer.  The end of the
4547 	 * buffer, in this case, is our file EOF, not the allocation size
4548 	 * of the buffer.
4549 	 */
4550 	soff = off;
4551 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4552 	if (eoff > bp->b_offset + bp->b_bcount)
4553 		eoff = bp->b_offset + bp->b_bcount;
4554 
4555 	/*
4556 	 * Set valid range.  This is typically the entire buffer and thus the
4557 	 * entire page.
4558 	 */
4559 	if (eoff > soff) {
4560 		vm_page_set_validclean(
4561 		    m,
4562 		   (vm_offset_t) (soff & PAGE_MASK),
4563 		   (vm_offset_t) (eoff - soff)
4564 		);
4565 	}
4566 }
4567 
4568 /*
4569  * Acquire a shared busy on all pages in the buf.
4570  */
4571 void
4572 vfs_busy_pages_acquire(struct buf *bp)
4573 {
4574 	int i;
4575 
4576 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4577 	for (i = 0; i < bp->b_npages; i++)
4578 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4579 }
4580 
4581 void
4582 vfs_busy_pages_release(struct buf *bp)
4583 {
4584 	int i;
4585 
4586 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4587 	for (i = 0; i < bp->b_npages; i++)
4588 		vm_page_sunbusy(bp->b_pages[i]);
4589 }
4590 
4591 /*
4592  * This routine is called before a device strategy routine.
4593  * It is used to tell the VM system that paging I/O is in
4594  * progress, and treat the pages associated with the buffer
4595  * almost as being exclusive busy.  Also the object paging_in_progress
4596  * flag is handled to make sure that the object doesn't become
4597  * inconsistent.
4598  *
4599  * Since I/O has not been initiated yet, certain buffer flags
4600  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4601  * and should be ignored.
4602  */
4603 void
4604 vfs_busy_pages(struct buf *bp, int clear_modify)
4605 {
4606 	vm_object_t obj;
4607 	vm_ooffset_t foff;
4608 	vm_page_t m;
4609 	int i;
4610 	bool bogus;
4611 
4612 	if (!(bp->b_flags & B_VMIO))
4613 		return;
4614 
4615 	obj = bp->b_bufobj->bo_object;
4616 	foff = bp->b_offset;
4617 	KASSERT(bp->b_offset != NOOFFSET,
4618 	    ("vfs_busy_pages: no buffer offset"));
4619 	VM_OBJECT_WLOCK(obj);
4620 	if ((bp->b_flags & B_CLUSTER) == 0) {
4621 		vm_object_pip_add(obj, bp->b_npages);
4622 		vfs_busy_pages_acquire(bp);
4623 	}
4624 	if (bp->b_bufsize != 0)
4625 		vfs_setdirty_locked_object(bp);
4626 	bogus = false;
4627 	for (i = 0; i < bp->b_npages; i++) {
4628 		m = bp->b_pages[i];
4629 		vm_page_assert_sbusied(m);
4630 
4631 		/*
4632 		 * When readying a buffer for a read ( i.e
4633 		 * clear_modify == 0 ), it is important to do
4634 		 * bogus_page replacement for valid pages in
4635 		 * partially instantiated buffers.  Partially
4636 		 * instantiated buffers can, in turn, occur when
4637 		 * reconstituting a buffer from its VM backing store
4638 		 * base.  We only have to do this if B_CACHE is
4639 		 * clear ( which causes the I/O to occur in the
4640 		 * first place ).  The replacement prevents the read
4641 		 * I/O from overwriting potentially dirty VM-backed
4642 		 * pages.  XXX bogus page replacement is, uh, bogus.
4643 		 * It may not work properly with small-block devices.
4644 		 * We need to find a better way.
4645 		 */
4646 		if (clear_modify) {
4647 			pmap_remove_write(m);
4648 			vfs_page_set_validclean(bp, foff, m);
4649 		} else if (vm_page_all_valid(m) &&
4650 		    (bp->b_flags & B_CACHE) == 0) {
4651 			bp->b_pages[i] = bogus_page;
4652 			bogus = true;
4653 		}
4654 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4655 	}
4656 	VM_OBJECT_WUNLOCK(obj);
4657 	if (bogus && buf_mapped(bp)) {
4658 		BUF_CHECK_MAPPED(bp);
4659 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4660 		    bp->b_pages, bp->b_npages);
4661 	}
4662 }
4663 
4664 /*
4665  *	vfs_bio_set_valid:
4666  *
4667  *	Set the range within the buffer to valid.  The range is
4668  *	relative to the beginning of the buffer, b_offset.  Note that
4669  *	b_offset itself may be offset from the beginning of the first
4670  *	page.
4671  */
4672 void
4673 vfs_bio_set_valid(struct buf *bp, int base, int size)
4674 {
4675 	int i, n;
4676 	vm_page_t m;
4677 
4678 	if (!(bp->b_flags & B_VMIO))
4679 		return;
4680 
4681 	/*
4682 	 * Fixup base to be relative to beginning of first page.
4683 	 * Set initial n to be the maximum number of bytes in the
4684 	 * first page that can be validated.
4685 	 */
4686 	base += (bp->b_offset & PAGE_MASK);
4687 	n = PAGE_SIZE - (base & PAGE_MASK);
4688 
4689 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4690 
4691 	/*
4692 	 * Busy may not be strictly necessary here because the pages are
4693 	 * unlikely to be fully valid and the vnode lock will synchronize
4694 	 * their access via getpages.  It is grabbed for consistency with
4695 	 * other page validation.
4696 	 */
4697 	vfs_busy_pages_acquire(bp);
4698 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4699 		m = bp->b_pages[i];
4700 		if (n > size)
4701 			n = size;
4702 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4703 		base += n;
4704 		size -= n;
4705 		n = PAGE_SIZE;
4706 	}
4707 	vfs_busy_pages_release(bp);
4708 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4709 }
4710 
4711 /*
4712  *	vfs_bio_clrbuf:
4713  *
4714  *	If the specified buffer is a non-VMIO buffer, clear the entire
4715  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4716  *	validate only the previously invalid portions of the buffer.
4717  *	This routine essentially fakes an I/O, so we need to clear
4718  *	BIO_ERROR and B_INVAL.
4719  *
4720  *	Note that while we only theoretically need to clear through b_bcount,
4721  *	we go ahead and clear through b_bufsize.
4722  */
4723 void
4724 vfs_bio_clrbuf(struct buf *bp)
4725 {
4726 	int i, j, mask, sa, ea, slide;
4727 
4728 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4729 		clrbuf(bp);
4730 		return;
4731 	}
4732 	bp->b_flags &= ~B_INVAL;
4733 	bp->b_ioflags &= ~BIO_ERROR;
4734 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4735 	vfs_busy_pages_acquire(bp);
4736 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4737 	    (bp->b_offset & PAGE_MASK) == 0) {
4738 		if (bp->b_pages[0] == bogus_page)
4739 			goto unlock;
4740 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4741 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4742 		if ((bp->b_pages[0]->valid & mask) == mask)
4743 			goto unlock;
4744 		if ((bp->b_pages[0]->valid & mask) == 0) {
4745 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4746 			bp->b_pages[0]->valid |= mask;
4747 			goto unlock;
4748 		}
4749 	}
4750 	sa = bp->b_offset & PAGE_MASK;
4751 	slide = 0;
4752 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4753 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4754 		ea = slide & PAGE_MASK;
4755 		if (ea == 0)
4756 			ea = PAGE_SIZE;
4757 		if (bp->b_pages[i] == bogus_page)
4758 			continue;
4759 		j = sa / DEV_BSIZE;
4760 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4761 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4762 		if ((bp->b_pages[i]->valid & mask) == mask)
4763 			continue;
4764 		if ((bp->b_pages[i]->valid & mask) == 0)
4765 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4766 		else {
4767 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4768 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4769 					pmap_zero_page_area(bp->b_pages[i],
4770 					    sa, DEV_BSIZE);
4771 				}
4772 			}
4773 		}
4774 		bp->b_pages[i]->valid |= mask;
4775 	}
4776 unlock:
4777 	vfs_busy_pages_release(bp);
4778 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4779 	bp->b_resid = 0;
4780 }
4781 
4782 void
4783 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4784 {
4785 	vm_page_t m;
4786 	int i, n;
4787 
4788 	if (buf_mapped(bp)) {
4789 		BUF_CHECK_MAPPED(bp);
4790 		bzero(bp->b_data + base, size);
4791 	} else {
4792 		BUF_CHECK_UNMAPPED(bp);
4793 		n = PAGE_SIZE - (base & PAGE_MASK);
4794 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4795 			m = bp->b_pages[i];
4796 			if (n > size)
4797 				n = size;
4798 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4799 			base += n;
4800 			size -= n;
4801 			n = PAGE_SIZE;
4802 		}
4803 	}
4804 }
4805 
4806 /*
4807  * Update buffer flags based on I/O request parameters, optionally releasing the
4808  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4809  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4810  * I/O).  Otherwise the buffer is released to the cache.
4811  */
4812 static void
4813 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4814 {
4815 
4816 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4817 	    ("buf %p non-VMIO noreuse", bp));
4818 
4819 	if ((ioflag & IO_DIRECT) != 0)
4820 		bp->b_flags |= B_DIRECT;
4821 	if ((ioflag & IO_EXT) != 0)
4822 		bp->b_xflags |= BX_ALTDATA;
4823 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4824 		bp->b_flags |= B_RELBUF;
4825 		if ((ioflag & IO_NOREUSE) != 0)
4826 			bp->b_flags |= B_NOREUSE;
4827 		if (release)
4828 			brelse(bp);
4829 	} else if (release)
4830 		bqrelse(bp);
4831 }
4832 
4833 void
4834 vfs_bio_brelse(struct buf *bp, int ioflag)
4835 {
4836 
4837 	b_io_dismiss(bp, ioflag, true);
4838 }
4839 
4840 void
4841 vfs_bio_set_flags(struct buf *bp, int ioflag)
4842 {
4843 
4844 	b_io_dismiss(bp, ioflag, false);
4845 }
4846 
4847 /*
4848  * vm_hold_load_pages and vm_hold_free_pages get pages into
4849  * a buffers address space.  The pages are anonymous and are
4850  * not associated with a file object.
4851  */
4852 static void
4853 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4854 {
4855 	vm_offset_t pg;
4856 	vm_page_t p;
4857 	int index;
4858 
4859 	BUF_CHECK_MAPPED(bp);
4860 
4861 	to = round_page(to);
4862 	from = round_page(from);
4863 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4864 
4865 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4866 		/*
4867 		 * note: must allocate system pages since blocking here
4868 		 * could interfere with paging I/O, no matter which
4869 		 * process we are.
4870 		 */
4871 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4872 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4873 		    VM_ALLOC_WAITOK);
4874 		pmap_qenter(pg, &p, 1);
4875 		bp->b_pages[index] = p;
4876 	}
4877 	bp->b_npages = index;
4878 }
4879 
4880 /* Return pages associated with this buf to the vm system */
4881 static void
4882 vm_hold_free_pages(struct buf *bp, int newbsize)
4883 {
4884 	vm_offset_t from;
4885 	vm_page_t p;
4886 	int index, newnpages;
4887 
4888 	BUF_CHECK_MAPPED(bp);
4889 
4890 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4891 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4892 	if (bp->b_npages > newnpages)
4893 		pmap_qremove(from, bp->b_npages - newnpages);
4894 	for (index = newnpages; index < bp->b_npages; index++) {
4895 		p = bp->b_pages[index];
4896 		bp->b_pages[index] = NULL;
4897 		vm_page_unwire_noq(p);
4898 		vm_page_free(p);
4899 	}
4900 	bp->b_npages = newnpages;
4901 }
4902 
4903 /*
4904  * Map an IO request into kernel virtual address space.
4905  *
4906  * All requests are (re)mapped into kernel VA space.
4907  * Notice that we use b_bufsize for the size of the buffer
4908  * to be mapped.  b_bcount might be modified by the driver.
4909  *
4910  * Note that even if the caller determines that the address space should
4911  * be valid, a race or a smaller-file mapped into a larger space may
4912  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4913  * check the return value.
4914  *
4915  * This function only works with pager buffers.
4916  */
4917 int
4918 vmapbuf(struct buf *bp, int mapbuf)
4919 {
4920 	vm_prot_t prot;
4921 	int pidx;
4922 
4923 	if (bp->b_bufsize < 0)
4924 		return (-1);
4925 	prot = VM_PROT_READ;
4926 	if (bp->b_iocmd == BIO_READ)
4927 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4928 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4929 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4930 	    btoc(MAXPHYS))) < 0)
4931 		return (-1);
4932 	bp->b_npages = pidx;
4933 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4934 	if (mapbuf || !unmapped_buf_allowed) {
4935 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4936 		bp->b_data = bp->b_kvabase + bp->b_offset;
4937 	} else
4938 		bp->b_data = unmapped_buf;
4939 	return(0);
4940 }
4941 
4942 /*
4943  * Free the io map PTEs associated with this IO operation.
4944  * We also invalidate the TLB entries and restore the original b_addr.
4945  *
4946  * This function only works with pager buffers.
4947  */
4948 void
4949 vunmapbuf(struct buf *bp)
4950 {
4951 	int npages;
4952 
4953 	npages = bp->b_npages;
4954 	if (buf_mapped(bp))
4955 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4956 	vm_page_unhold_pages(bp->b_pages, npages);
4957 
4958 	bp->b_data = unmapped_buf;
4959 }
4960 
4961 void
4962 bdone(struct buf *bp)
4963 {
4964 	struct mtx *mtxp;
4965 
4966 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4967 	mtx_lock(mtxp);
4968 	bp->b_flags |= B_DONE;
4969 	wakeup(bp);
4970 	mtx_unlock(mtxp);
4971 }
4972 
4973 void
4974 bwait(struct buf *bp, u_char pri, const char *wchan)
4975 {
4976 	struct mtx *mtxp;
4977 
4978 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4979 	mtx_lock(mtxp);
4980 	while ((bp->b_flags & B_DONE) == 0)
4981 		msleep(bp, mtxp, pri, wchan, 0);
4982 	mtx_unlock(mtxp);
4983 }
4984 
4985 int
4986 bufsync(struct bufobj *bo, int waitfor)
4987 {
4988 
4989 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
4990 }
4991 
4992 void
4993 bufstrategy(struct bufobj *bo, struct buf *bp)
4994 {
4995 	int i __unused;
4996 	struct vnode *vp;
4997 
4998 	vp = bp->b_vp;
4999 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5000 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5001 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5002 	i = VOP_STRATEGY(vp, bp);
5003 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5004 }
5005 
5006 /*
5007  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
5008  */
5009 void
5010 bufobj_init(struct bufobj *bo, void *private)
5011 {
5012 	static volatile int bufobj_cleanq;
5013 
5014         bo->bo_domain =
5015             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5016         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5017         bo->bo_private = private;
5018         TAILQ_INIT(&bo->bo_clean.bv_hd);
5019         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5020 }
5021 
5022 void
5023 bufobj_wrefl(struct bufobj *bo)
5024 {
5025 
5026 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5027 	ASSERT_BO_WLOCKED(bo);
5028 	bo->bo_numoutput++;
5029 }
5030 
5031 void
5032 bufobj_wref(struct bufobj *bo)
5033 {
5034 
5035 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5036 	BO_LOCK(bo);
5037 	bo->bo_numoutput++;
5038 	BO_UNLOCK(bo);
5039 }
5040 
5041 void
5042 bufobj_wdrop(struct bufobj *bo)
5043 {
5044 
5045 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5046 	BO_LOCK(bo);
5047 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5048 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5049 		bo->bo_flag &= ~BO_WWAIT;
5050 		wakeup(&bo->bo_numoutput);
5051 	}
5052 	BO_UNLOCK(bo);
5053 }
5054 
5055 int
5056 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5057 {
5058 	int error;
5059 
5060 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5061 	ASSERT_BO_WLOCKED(bo);
5062 	error = 0;
5063 	while (bo->bo_numoutput) {
5064 		bo->bo_flag |= BO_WWAIT;
5065 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5066 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5067 		if (error)
5068 			break;
5069 	}
5070 	return (error);
5071 }
5072 
5073 /*
5074  * Set bio_data or bio_ma for struct bio from the struct buf.
5075  */
5076 void
5077 bdata2bio(struct buf *bp, struct bio *bip)
5078 {
5079 
5080 	if (!buf_mapped(bp)) {
5081 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5082 		bip->bio_ma = bp->b_pages;
5083 		bip->bio_ma_n = bp->b_npages;
5084 		bip->bio_data = unmapped_buf;
5085 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5086 		bip->bio_flags |= BIO_UNMAPPED;
5087 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5088 		    PAGE_SIZE == bp->b_npages,
5089 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5090 		    (long long)bip->bio_length, bip->bio_ma_n));
5091 	} else {
5092 		bip->bio_data = bp->b_data;
5093 		bip->bio_ma = NULL;
5094 	}
5095 }
5096 
5097 /*
5098  * The MIPS pmap code currently doesn't handle aliased pages.
5099  * The VIPT caches may not handle page aliasing themselves, leading
5100  * to data corruption.
5101  *
5102  * As such, this code makes a system extremely unhappy if said
5103  * system doesn't support unaliasing the above situation in hardware.
5104  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5105  * this feature at build time, so it has to be handled in software.
5106  *
5107  * Once the MIPS pmap/cache code grows to support this function on
5108  * earlier chips, it should be flipped back off.
5109  */
5110 #ifdef	__mips__
5111 static int buf_pager_relbuf = 1;
5112 #else
5113 static int buf_pager_relbuf = 0;
5114 #endif
5115 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5116     &buf_pager_relbuf, 0,
5117     "Make buffer pager release buffers after reading");
5118 
5119 /*
5120  * The buffer pager.  It uses buffer reads to validate pages.
5121  *
5122  * In contrast to the generic local pager from vm/vnode_pager.c, this
5123  * pager correctly and easily handles volumes where the underlying
5124  * device block size is greater than the machine page size.  The
5125  * buffer cache transparently extends the requested page run to be
5126  * aligned at the block boundary, and does the necessary bogus page
5127  * replacements in the addends to avoid obliterating already valid
5128  * pages.
5129  *
5130  * The only non-trivial issue is that the exclusive busy state for
5131  * pages, which is assumed by the vm_pager_getpages() interface, is
5132  * incompatible with the VMIO buffer cache's desire to share-busy the
5133  * pages.  This function performs a trivial downgrade of the pages'
5134  * state before reading buffers, and a less trivial upgrade from the
5135  * shared-busy to excl-busy state after the read.
5136  */
5137 int
5138 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5139     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5140     vbg_get_blksize_t get_blksize)
5141 {
5142 	vm_page_t m;
5143 	vm_object_t object;
5144 	struct buf *bp;
5145 	struct mount *mp;
5146 	daddr_t lbn, lbnp;
5147 	vm_ooffset_t la, lb, poff, poffe;
5148 	long bsize;
5149 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5150 	bool redo, lpart;
5151 
5152 	object = vp->v_object;
5153 	mp = vp->v_mount;
5154 	error = 0;
5155 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5156 	if (la >= object->un_pager.vnp.vnp_size)
5157 		return (VM_PAGER_BAD);
5158 
5159 	/*
5160 	 * Change the meaning of la from where the last requested page starts
5161 	 * to where it ends, because that's the end of the requested region
5162 	 * and the start of the potential read-ahead region.
5163 	 */
5164 	la += PAGE_SIZE;
5165 	lpart = la > object->un_pager.vnp.vnp_size;
5166 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5167 
5168 	/*
5169 	 * Calculate read-ahead, behind and total pages.
5170 	 */
5171 	pgsin = count;
5172 	lb = IDX_TO_OFF(ma[0]->pindex);
5173 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5174 	pgsin += pgsin_b;
5175 	if (rbehind != NULL)
5176 		*rbehind = pgsin_b;
5177 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5178 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5179 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5180 		    PAGE_SIZE) - la);
5181 	pgsin += pgsin_a;
5182 	if (rahead != NULL)
5183 		*rahead = pgsin_a;
5184 	VM_CNT_INC(v_vnodein);
5185 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5186 
5187 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5188 	    != 0) ? GB_UNMAPPED : 0;
5189 	VM_OBJECT_WLOCK(object);
5190 again:
5191 	for (i = 0; i < count; i++)
5192 		vm_page_busy_downgrade(ma[i]);
5193 	VM_OBJECT_WUNLOCK(object);
5194 
5195 	lbnp = -1;
5196 	for (i = 0; i < count; i++) {
5197 		m = ma[i];
5198 
5199 		/*
5200 		 * Pages are shared busy and the object lock is not
5201 		 * owned, which together allow for the pages'
5202 		 * invalidation.  The racy test for validity avoids
5203 		 * useless creation of the buffer for the most typical
5204 		 * case when invalidation is not used in redo or for
5205 		 * parallel read.  The shared->excl upgrade loop at
5206 		 * the end of the function catches the race in a
5207 		 * reliable way (protected by the object lock).
5208 		 */
5209 		if (vm_page_all_valid(m))
5210 			continue;
5211 
5212 		poff = IDX_TO_OFF(m->pindex);
5213 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5214 		for (; poff < poffe; poff += bsize) {
5215 			lbn = get_lblkno(vp, poff);
5216 			if (lbn == lbnp)
5217 				goto next_page;
5218 			lbnp = lbn;
5219 
5220 			bsize = get_blksize(vp, lbn);
5221 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5222 			    br_flags, &bp);
5223 			if (error != 0)
5224 				goto end_pages;
5225 			if (LIST_EMPTY(&bp->b_dep)) {
5226 				/*
5227 				 * Invalidation clears m->valid, but
5228 				 * may leave B_CACHE flag if the
5229 				 * buffer existed at the invalidation
5230 				 * time.  In this case, recycle the
5231 				 * buffer to do real read on next
5232 				 * bread() after redo.
5233 				 *
5234 				 * Otherwise B_RELBUF is not strictly
5235 				 * necessary, enable to reduce buf
5236 				 * cache pressure.
5237 				 */
5238 				if (buf_pager_relbuf ||
5239 				    !vm_page_all_valid(m))
5240 					bp->b_flags |= B_RELBUF;
5241 
5242 				bp->b_flags &= ~B_NOCACHE;
5243 				brelse(bp);
5244 			} else {
5245 				bqrelse(bp);
5246 			}
5247 		}
5248 		KASSERT(1 /* racy, enable for debugging */ ||
5249 		    vm_page_all_valid(m) || i == count - 1,
5250 		    ("buf %d %p invalid", i, m));
5251 		if (i == count - 1 && lpart) {
5252 			VM_OBJECT_WLOCK(object);
5253 			if (!vm_page_none_valid(m) &&
5254 			    !vm_page_all_valid(m))
5255 				vm_page_zero_invalid(m, TRUE);
5256 			VM_OBJECT_WUNLOCK(object);
5257 		}
5258 next_page:;
5259 	}
5260 end_pages:
5261 
5262 	VM_OBJECT_WLOCK(object);
5263 	redo = false;
5264 	for (i = 0; i < count; i++) {
5265 		vm_page_sunbusy(ma[i]);
5266 		ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
5267 
5268 		/*
5269 		 * Since the pages were only sbusy while neither the
5270 		 * buffer nor the object lock was held by us, or
5271 		 * reallocated while vm_page_grab() slept for busy
5272 		 * relinguish, they could have been invalidated.
5273 		 * Recheck the valid bits and re-read as needed.
5274 		 *
5275 		 * Note that the last page is made fully valid in the
5276 		 * read loop, and partial validity for the page at
5277 		 * index count - 1 could mean that the page was
5278 		 * invalidated or removed, so we must restart for
5279 		 * safety as well.
5280 		 */
5281 		if (!vm_page_all_valid(ma[i]))
5282 			redo = true;
5283 	}
5284 	if (redo && error == 0)
5285 		goto again;
5286 	VM_OBJECT_WUNLOCK(object);
5287 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5288 }
5289 
5290 #include "opt_ddb.h"
5291 #ifdef DDB
5292 #include <ddb/ddb.h>
5293 
5294 /* DDB command to show buffer data */
5295 DB_SHOW_COMMAND(buffer, db_show_buffer)
5296 {
5297 	/* get args */
5298 	struct buf *bp = (struct buf *)addr;
5299 #ifdef FULL_BUF_TRACKING
5300 	uint32_t i, j;
5301 #endif
5302 
5303 	if (!have_addr) {
5304 		db_printf("usage: show buffer <addr>\n");
5305 		return;
5306 	}
5307 
5308 	db_printf("buf at %p\n", bp);
5309 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5310 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5311 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5312 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5313 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5314 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5315 	db_printf(
5316 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5317 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5318 	    "b_vp = %p, b_dep = %p\n",
5319 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5320 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5321 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5322 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5323 	    bp->b_kvabase, bp->b_kvasize);
5324 	if (bp->b_npages) {
5325 		int i;
5326 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5327 		for (i = 0; i < bp->b_npages; i++) {
5328 			vm_page_t m;
5329 			m = bp->b_pages[i];
5330 			if (m != NULL)
5331 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5332 				    (u_long)m->pindex,
5333 				    (u_long)VM_PAGE_TO_PHYS(m));
5334 			else
5335 				db_printf("( ??? )");
5336 			if ((i + 1) < bp->b_npages)
5337 				db_printf(",");
5338 		}
5339 		db_printf("\n");
5340 	}
5341 	BUF_LOCKPRINTINFO(bp);
5342 #if defined(FULL_BUF_TRACKING)
5343 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5344 
5345 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5346 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5347 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5348 			continue;
5349 		db_printf(" %2u: %s\n", j,
5350 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5351 	}
5352 #elif defined(BUF_TRACKING)
5353 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5354 #endif
5355 	db_printf(" ");
5356 }
5357 
5358 DB_SHOW_COMMAND(bufqueues, bufqueues)
5359 {
5360 	struct bufdomain *bd;
5361 	struct buf *bp;
5362 	long total;
5363 	int i, j, cnt;
5364 
5365 	db_printf("bqempty: %d\n", bqempty.bq_len);
5366 
5367 	for (i = 0; i < buf_domains; i++) {
5368 		bd = &bdomain[i];
5369 		db_printf("Buf domain %d\n", i);
5370 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5371 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5372 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5373 		db_printf("\n");
5374 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5375 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5376 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5377 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5378 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5379 		db_printf("\n");
5380 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5381 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5382 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5383 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5384 		db_printf("\n");
5385 		total = 0;
5386 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5387 			total += bp->b_bufsize;
5388 		db_printf("\tcleanq count\t%d (%ld)\n",
5389 		    bd->bd_cleanq->bq_len, total);
5390 		total = 0;
5391 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5392 			total += bp->b_bufsize;
5393 		db_printf("\tdirtyq count\t%d (%ld)\n",
5394 		    bd->bd_dirtyq.bq_len, total);
5395 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5396 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5397 		db_printf("\tCPU ");
5398 		for (j = 0; j <= mp_maxid; j++)
5399 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5400 		db_printf("\n");
5401 		cnt = 0;
5402 		total = 0;
5403 		for (j = 0; j < nbuf; j++)
5404 			if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5405 				cnt++;
5406 				total += buf[j].b_bufsize;
5407 			}
5408 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5409 		cnt = 0;
5410 		total = 0;
5411 		for (j = 0; j < nbuf; j++)
5412 			if (buf[j].b_domain == i) {
5413 				cnt++;
5414 				total += buf[j].b_bufsize;
5415 			}
5416 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5417 	}
5418 }
5419 
5420 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5421 {
5422 	struct buf *bp;
5423 	int i;
5424 
5425 	for (i = 0; i < nbuf; i++) {
5426 		bp = &buf[i];
5427 		if (BUF_ISLOCKED(bp)) {
5428 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5429 			db_printf("\n");
5430 			if (db_pager_quit)
5431 				break;
5432 		}
5433 	}
5434 }
5435 
5436 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5437 {
5438 	struct vnode *vp;
5439 	struct buf *bp;
5440 
5441 	if (!have_addr) {
5442 		db_printf("usage: show vnodebufs <addr>\n");
5443 		return;
5444 	}
5445 	vp = (struct vnode *)addr;
5446 	db_printf("Clean buffers:\n");
5447 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5448 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5449 		db_printf("\n");
5450 	}
5451 	db_printf("Dirty buffers:\n");
5452 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5453 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5454 		db_printf("\n");
5455 	}
5456 }
5457 
5458 DB_COMMAND(countfreebufs, db_coundfreebufs)
5459 {
5460 	struct buf *bp;
5461 	int i, used = 0, nfree = 0;
5462 
5463 	if (have_addr) {
5464 		db_printf("usage: countfreebufs\n");
5465 		return;
5466 	}
5467 
5468 	for (i = 0; i < nbuf; i++) {
5469 		bp = &buf[i];
5470 		if (bp->b_qindex == QUEUE_EMPTY)
5471 			nfree++;
5472 		else
5473 			used++;
5474 	}
5475 
5476 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5477 	    nfree + used);
5478 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5479 }
5480 #endif /* DDB */
5481