xref: /freebsd/sys/kern/vfs_bio.c (revision 2c3f47a727372086b41ef9ce06eb1f1eb83a67d3)
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, &altbufferflushes,
254     0, "Number of fsync flushes to limit dirty buffers");
255 static int recursiveflushes;
256 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
257     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, &barrierwrites, 0,
313     "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 	if (bp->b_flags & B_CACHE) {
960 		int base = (foff + off) & PAGE_MASK;
961 		if (vm_page_is_valid(m, base, size) == 0)
962 			bp->b_flags &= ~B_CACHE;
963 	}
964 }
965 
966 /* Wake up the buffer daemon if necessary */
967 static void
968 bd_wakeup(void)
969 {
970 
971 	mtx_lock(&bdlock);
972 	if (bd_request == 0) {
973 		bd_request = 1;
974 		wakeup(&bd_request);
975 	}
976 	mtx_unlock(&bdlock);
977 }
978 
979 /*
980  * Adjust the maxbcachbuf tunable.
981  */
982 static void
983 maxbcachebuf_adjust(void)
984 {
985 	int i;
986 
987 	/*
988 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
989 	 */
990 	i = 2;
991 	while (i * 2 <= maxbcachebuf)
992 		i *= 2;
993 	maxbcachebuf = i;
994 	if (maxbcachebuf < MAXBSIZE)
995 		maxbcachebuf = MAXBSIZE;
996 	if (maxbcachebuf > MAXPHYS)
997 		maxbcachebuf = MAXPHYS;
998 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
999 		printf("maxbcachebuf=%d\n", maxbcachebuf);
1000 }
1001 
1002 /*
1003  * bd_speedup - speedup the buffer cache flushing code
1004  */
1005 void
1006 bd_speedup(void)
1007 {
1008 	int needwake;
1009 
1010 	mtx_lock(&bdlock);
1011 	needwake = 0;
1012 	if (bd_speedupreq == 0 || bd_request == 0)
1013 		needwake = 1;
1014 	bd_speedupreq = 1;
1015 	bd_request = 1;
1016 	if (needwake)
1017 		wakeup(&bd_request);
1018 	mtx_unlock(&bdlock);
1019 }
1020 
1021 #ifdef __i386__
1022 #define	TRANSIENT_DENOM	5
1023 #else
1024 #define	TRANSIENT_DENOM 10
1025 #endif
1026 
1027 /*
1028  * Calculating buffer cache scaling values and reserve space for buffer
1029  * headers.  This is called during low level kernel initialization and
1030  * may be called more then once.  We CANNOT write to the memory area
1031  * being reserved at this time.
1032  */
1033 caddr_t
1034 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1035 {
1036 	int tuned_nbuf;
1037 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1038 
1039 	/*
1040 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1041 	 * PAGE_SIZE is >= 1K)
1042 	 */
1043 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1044 
1045 	maxbcachebuf_adjust();
1046 	/*
1047 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1048 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1049 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1050 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1051 	 * the buffer cache we limit the eventual kva reservation to
1052 	 * maxbcache bytes.
1053 	 *
1054 	 * factor represents the 1/4 x ram conversion.
1055 	 */
1056 	if (nbuf == 0) {
1057 		int factor = 4 * BKVASIZE / 1024;
1058 
1059 		nbuf = 50;
1060 		if (physmem_est > 4096)
1061 			nbuf += min((physmem_est - 4096) / factor,
1062 			    65536 / factor);
1063 		if (physmem_est > 65536)
1064 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1065 			    32 * 1024 * 1024 / (factor * 5));
1066 
1067 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1068 			nbuf = maxbcache / BKVASIZE;
1069 		tuned_nbuf = 1;
1070 	} else
1071 		tuned_nbuf = 0;
1072 
1073 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1074 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1075 	if (nbuf > maxbuf) {
1076 		if (!tuned_nbuf)
1077 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1078 			    maxbuf);
1079 		nbuf = maxbuf;
1080 	}
1081 
1082 	/*
1083 	 * Ideal allocation size for the transient bio submap is 10%
1084 	 * of the maximal space buffer map.  This roughly corresponds
1085 	 * to the amount of the buffer mapped for typical UFS load.
1086 	 *
1087 	 * Clip the buffer map to reserve space for the transient
1088 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1089 	 * maximum buffer map extent on the platform.
1090 	 *
1091 	 * The fall-back to the maxbuf in case of maxbcache unset,
1092 	 * allows to not trim the buffer KVA for the architectures
1093 	 * with ample KVA space.
1094 	 */
1095 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1096 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1097 		buf_sz = (long)nbuf * BKVASIZE;
1098 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1099 		    (TRANSIENT_DENOM - 1)) {
1100 			/*
1101 			 * There is more KVA than memory.  Do not
1102 			 * adjust buffer map size, and assign the rest
1103 			 * of maxbuf to transient map.
1104 			 */
1105 			biotmap_sz = maxbuf_sz - buf_sz;
1106 		} else {
1107 			/*
1108 			 * Buffer map spans all KVA we could afford on
1109 			 * this platform.  Give 10% (20% on i386) of
1110 			 * the buffer map to the transient bio map.
1111 			 */
1112 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1113 			buf_sz -= biotmap_sz;
1114 		}
1115 		if (biotmap_sz / INT_MAX > MAXPHYS)
1116 			bio_transient_maxcnt = INT_MAX;
1117 		else
1118 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1119 		/*
1120 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1121 		 * using the transient mapping.
1122 		 */
1123 		if (bio_transient_maxcnt > 1024)
1124 			bio_transient_maxcnt = 1024;
1125 		if (tuned_nbuf)
1126 			nbuf = buf_sz / BKVASIZE;
1127 	}
1128 
1129 	if (nswbuf == 0) {
1130 		nswbuf = min(nbuf / 4, 256);
1131 		if (nswbuf < NSWBUF_MIN)
1132 			nswbuf = NSWBUF_MIN;
1133 	}
1134 
1135 	/*
1136 	 * Reserve space for the buffer cache buffers
1137 	 */
1138 	buf = (void *)v;
1139 	v = (caddr_t)(buf + nbuf);
1140 
1141 	return(v);
1142 }
1143 
1144 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1145 void
1146 bufinit(void)
1147 {
1148 	struct buf *bp;
1149 	int i;
1150 
1151 	KASSERT(maxbcachebuf >= MAXBSIZE,
1152 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1153 	    MAXBSIZE));
1154 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1155 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1156 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1157 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1158 
1159 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1160 
1161 	/* finally, initialize each buffer header and stick on empty q */
1162 	for (i = 0; i < nbuf; i++) {
1163 		bp = &buf[i];
1164 		bzero(bp, sizeof *bp);
1165 		bp->b_flags = B_INVAL;
1166 		bp->b_rcred = NOCRED;
1167 		bp->b_wcred = NOCRED;
1168 		bp->b_qindex = QUEUE_NONE;
1169 		bp->b_domain = -1;
1170 		bp->b_subqueue = mp_maxid + 1;
1171 		bp->b_xflags = 0;
1172 		bp->b_data = bp->b_kvabase = unmapped_buf;
1173 		LIST_INIT(&bp->b_dep);
1174 		BUF_LOCKINIT(bp);
1175 		bq_insert(&bqempty, bp, false);
1176 	}
1177 
1178 	/*
1179 	 * maxbufspace is the absolute maximum amount of buffer space we are
1180 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1181 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1182 	 * used by most other requests.  The differential is required to
1183 	 * ensure that metadata deadlocks don't occur.
1184 	 *
1185 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1186 	 * this may result in KVM fragmentation which is not handled optimally
1187 	 * by the system. XXX This is less true with vmem.  We could use
1188 	 * PAGE_SIZE.
1189 	 */
1190 	maxbufspace = (long)nbuf * BKVASIZE;
1191 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1192 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1193 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1194 
1195 	/*
1196 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1197 	 * arbitrarily and may need further tuning. It corresponds to
1198 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1199 	 * which fits with many RAID controllers' tagged queuing limits.
1200 	 * The lower 1 MiB limit is the historical upper limit for
1201 	 * hirunningspace.
1202 	 */
1203 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1204 	    16 * 1024 * 1024), 1024 * 1024);
1205 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1206 
1207 	/*
1208 	 * Limit the amount of malloc memory since it is wired permanently into
1209 	 * the kernel space.  Even though this is accounted for in the buffer
1210 	 * allocation, we don't want the malloced region to grow uncontrolled.
1211 	 * The malloc scheme improves memory utilization significantly on
1212 	 * average (small) directories.
1213 	 */
1214 	maxbufmallocspace = hibufspace / 20;
1215 
1216 	/*
1217 	 * Reduce the chance of a deadlock occurring by limiting the number
1218 	 * of delayed-write dirty buffers we allow to stack up.
1219 	 */
1220 	hidirtybuffers = nbuf / 4 + 20;
1221 	dirtybufthresh = hidirtybuffers * 9 / 10;
1222 	/*
1223 	 * To support extreme low-memory systems, make sure hidirtybuffers
1224 	 * cannot eat up all available buffer space.  This occurs when our
1225 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1226 	 * buffer space assuming BKVASIZE'd buffers.
1227 	 */
1228 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1229 		hidirtybuffers >>= 1;
1230 	}
1231 	lodirtybuffers = hidirtybuffers / 2;
1232 
1233 	/*
1234 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1235 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1236 	 * are counted as free but will be unavailable to threads executing
1237 	 * on other cpus.
1238 	 *
1239 	 * hifreebuffers is the free target for the bufspace daemon.  This
1240 	 * should be set appropriately to limit work per-iteration.
1241 	 */
1242 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1243 	hifreebuffers = (3 * lofreebuffers) / 2;
1244 	numfreebuffers = nbuf;
1245 
1246 	/* Setup the kva and free list allocators. */
1247 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1248 	buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1249 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1250 
1251 	/*
1252 	 * Size the clean queue according to the amount of buffer space.
1253 	 * One queue per-256mb up to the max.  More queues gives better
1254 	 * concurrency but less accurate LRU.
1255 	 */
1256 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1257 	for (i = 0 ; i < buf_domains; i++) {
1258 		struct bufdomain *bd;
1259 
1260 		bd = &bdomain[i];
1261 		bd_init(bd);
1262 		bd->bd_freebuffers = nbuf / buf_domains;
1263 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1264 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1265 		bd->bd_bufspace = 0;
1266 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1267 		bd->bd_hibufspace = hibufspace / buf_domains;
1268 		bd->bd_lobufspace = lobufspace / buf_domains;
1269 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1270 		bd->bd_numdirtybuffers = 0;
1271 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1272 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1273 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1274 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1275 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1276 	}
1277 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1278 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1279 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1280 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1281 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1282 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1283 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1284 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1285 }
1286 
1287 #ifdef INVARIANTS
1288 static inline void
1289 vfs_buf_check_mapped(struct buf *bp)
1290 {
1291 
1292 	KASSERT(bp->b_kvabase != unmapped_buf,
1293 	    ("mapped buf: b_kvabase was not updated %p", bp));
1294 	KASSERT(bp->b_data != unmapped_buf,
1295 	    ("mapped buf: b_data was not updated %p", bp));
1296 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1297 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1298 }
1299 
1300 static inline void
1301 vfs_buf_check_unmapped(struct buf *bp)
1302 {
1303 
1304 	KASSERT(bp->b_data == unmapped_buf,
1305 	    ("unmapped buf: corrupted b_data %p", bp));
1306 }
1307 
1308 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1309 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1310 #else
1311 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1312 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1313 #endif
1314 
1315 static int
1316 isbufbusy(struct buf *bp)
1317 {
1318 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1319 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1320 		return (1);
1321 	return (0);
1322 }
1323 
1324 /*
1325  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1326  */
1327 void
1328 bufshutdown(int show_busybufs)
1329 {
1330 	static int first_buf_printf = 1;
1331 	struct buf *bp;
1332 	int iter, nbusy, pbusy;
1333 #ifndef PREEMPTION
1334 	int subiter;
1335 #endif
1336 
1337 	/*
1338 	 * Sync filesystems for shutdown
1339 	 */
1340 	wdog_kern_pat(WD_LASTVAL);
1341 	sys_sync(curthread, NULL);
1342 
1343 	/*
1344 	 * With soft updates, some buffers that are
1345 	 * written will be remarked as dirty until other
1346 	 * buffers are written.
1347 	 */
1348 	for (iter = pbusy = 0; iter < 20; iter++) {
1349 		nbusy = 0;
1350 		for (bp = &buf[nbuf]; --bp >= buf; )
1351 			if (isbufbusy(bp))
1352 				nbusy++;
1353 		if (nbusy == 0) {
1354 			if (first_buf_printf)
1355 				printf("All buffers synced.");
1356 			break;
1357 		}
1358 		if (first_buf_printf) {
1359 			printf("Syncing disks, buffers remaining... ");
1360 			first_buf_printf = 0;
1361 		}
1362 		printf("%d ", nbusy);
1363 		if (nbusy < pbusy)
1364 			iter = 0;
1365 		pbusy = nbusy;
1366 
1367 		wdog_kern_pat(WD_LASTVAL);
1368 		sys_sync(curthread, NULL);
1369 
1370 #ifdef PREEMPTION
1371 		/*
1372 		 * Spin for a while to allow interrupt threads to run.
1373 		 */
1374 		DELAY(50000 * iter);
1375 #else
1376 		/*
1377 		 * Context switch several times to allow interrupt
1378 		 * threads to run.
1379 		 */
1380 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1381 			thread_lock(curthread);
1382 			mi_switch(SW_VOL, NULL);
1383 			thread_unlock(curthread);
1384 			DELAY(1000);
1385 		}
1386 #endif
1387 	}
1388 	printf("\n");
1389 	/*
1390 	 * Count only busy local buffers to prevent forcing
1391 	 * a fsck if we're just a client of a wedged NFS server
1392 	 */
1393 	nbusy = 0;
1394 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1395 		if (isbufbusy(bp)) {
1396 #if 0
1397 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1398 			if (bp->b_dev == NULL) {
1399 				TAILQ_REMOVE(&mountlist,
1400 				    bp->b_vp->v_mount, mnt_list);
1401 				continue;
1402 			}
1403 #endif
1404 			nbusy++;
1405 			if (show_busybufs > 0) {
1406 				printf(
1407 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1408 				    nbusy, bp, bp->b_vp, bp->b_flags,
1409 				    (intmax_t)bp->b_blkno,
1410 				    (intmax_t)bp->b_lblkno);
1411 				BUF_LOCKPRINTINFO(bp);
1412 				if (show_busybufs > 1)
1413 					vn_printf(bp->b_vp,
1414 					    "vnode content: ");
1415 			}
1416 		}
1417 	}
1418 	if (nbusy) {
1419 		/*
1420 		 * Failed to sync all blocks. Indicate this and don't
1421 		 * unmount filesystems (thus forcing an fsck on reboot).
1422 		 */
1423 		printf("Giving up on %d buffers\n", nbusy);
1424 		DELAY(5000000);	/* 5 seconds */
1425 	} else {
1426 		if (!first_buf_printf)
1427 			printf("Final sync complete\n");
1428 		/*
1429 		 * Unmount filesystems
1430 		 */
1431 		if (panicstr == NULL)
1432 			vfs_unmountall();
1433 	}
1434 	swapoff_all();
1435 	DELAY(100000);		/* wait for console output to finish */
1436 }
1437 
1438 static void
1439 bpmap_qenter(struct buf *bp)
1440 {
1441 
1442 	BUF_CHECK_MAPPED(bp);
1443 
1444 	/*
1445 	 * bp->b_data is relative to bp->b_offset, but
1446 	 * bp->b_offset may be offset into the first page.
1447 	 */
1448 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1449 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1450 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1451 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1452 }
1453 
1454 static inline struct bufdomain *
1455 bufdomain(struct buf *bp)
1456 {
1457 
1458 	return (&bdomain[bp->b_domain]);
1459 }
1460 
1461 static struct bufqueue *
1462 bufqueue(struct buf *bp)
1463 {
1464 
1465 	switch (bp->b_qindex) {
1466 	case QUEUE_NONE:
1467 		/* FALLTHROUGH */
1468 	case QUEUE_SENTINEL:
1469 		return (NULL);
1470 	case QUEUE_EMPTY:
1471 		return (&bqempty);
1472 	case QUEUE_DIRTY:
1473 		return (&bufdomain(bp)->bd_dirtyq);
1474 	case QUEUE_CLEAN:
1475 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1476 	default:
1477 		break;
1478 	}
1479 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1480 }
1481 
1482 /*
1483  * Return the locked bufqueue that bp is a member of.
1484  */
1485 static struct bufqueue *
1486 bufqueue_acquire(struct buf *bp)
1487 {
1488 	struct bufqueue *bq, *nbq;
1489 
1490 	/*
1491 	 * bp can be pushed from a per-cpu queue to the
1492 	 * cleanq while we're waiting on the lock.  Retry
1493 	 * if the queues don't match.
1494 	 */
1495 	bq = bufqueue(bp);
1496 	BQ_LOCK(bq);
1497 	for (;;) {
1498 		nbq = bufqueue(bp);
1499 		if (bq == nbq)
1500 			break;
1501 		BQ_UNLOCK(bq);
1502 		BQ_LOCK(nbq);
1503 		bq = nbq;
1504 	}
1505 	return (bq);
1506 }
1507 
1508 /*
1509  *	binsfree:
1510  *
1511  *	Insert the buffer into the appropriate free list.  Requires a
1512  *	locked buffer on entry and buffer is unlocked before return.
1513  */
1514 static void
1515 binsfree(struct buf *bp, int qindex)
1516 {
1517 	struct bufdomain *bd;
1518 	struct bufqueue *bq;
1519 
1520 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1521 	    ("binsfree: Invalid qindex %d", qindex));
1522 	BUF_ASSERT_XLOCKED(bp);
1523 
1524 	/*
1525 	 * Handle delayed bremfree() processing.
1526 	 */
1527 	if (bp->b_flags & B_REMFREE) {
1528 		if (bp->b_qindex == qindex) {
1529 			bp->b_flags |= B_REUSE;
1530 			bp->b_flags &= ~B_REMFREE;
1531 			BUF_UNLOCK(bp);
1532 			return;
1533 		}
1534 		bq = bufqueue_acquire(bp);
1535 		bq_remove(bq, bp);
1536 		BQ_UNLOCK(bq);
1537 	}
1538 	bd = bufdomain(bp);
1539 	if (qindex == QUEUE_CLEAN) {
1540 		if (bd->bd_lim != 0)
1541 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1542 		else
1543 			bq = bd->bd_cleanq;
1544 	} else
1545 		bq = &bd->bd_dirtyq;
1546 	bq_insert(bq, bp, true);
1547 }
1548 
1549 /*
1550  * buf_free:
1551  *
1552  *	Free a buffer to the buf zone once it no longer has valid contents.
1553  */
1554 static void
1555 buf_free(struct buf *bp)
1556 {
1557 
1558 	if (bp->b_flags & B_REMFREE)
1559 		bremfreef(bp);
1560 	if (bp->b_vflags & BV_BKGRDINPROG)
1561 		panic("losing buffer 1");
1562 	if (bp->b_rcred != NOCRED) {
1563 		crfree(bp->b_rcred);
1564 		bp->b_rcred = NOCRED;
1565 	}
1566 	if (bp->b_wcred != NOCRED) {
1567 		crfree(bp->b_wcred);
1568 		bp->b_wcred = NOCRED;
1569 	}
1570 	if (!LIST_EMPTY(&bp->b_dep))
1571 		buf_deallocate(bp);
1572 	bufkva_free(bp);
1573 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1574 	BUF_UNLOCK(bp);
1575 	uma_zfree(buf_zone, bp);
1576 }
1577 
1578 /*
1579  * buf_import:
1580  *
1581  *	Import bufs into the uma cache from the buf list.  The system still
1582  *	expects a static array of bufs and much of the synchronization
1583  *	around bufs assumes type stable storage.  As a result, UMA is used
1584  *	only as a per-cpu cache of bufs still maintained on a global list.
1585  */
1586 static int
1587 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1588 {
1589 	struct buf *bp;
1590 	int i;
1591 
1592 	BQ_LOCK(&bqempty);
1593 	for (i = 0; i < cnt; i++) {
1594 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1595 		if (bp == NULL)
1596 			break;
1597 		bq_remove(&bqempty, bp);
1598 		store[i] = bp;
1599 	}
1600 	BQ_UNLOCK(&bqempty);
1601 
1602 	return (i);
1603 }
1604 
1605 /*
1606  * buf_release:
1607  *
1608  *	Release bufs from the uma cache back to the buffer queues.
1609  */
1610 static void
1611 buf_release(void *arg, void **store, int cnt)
1612 {
1613 	struct bufqueue *bq;
1614 	struct buf *bp;
1615         int i;
1616 
1617 	bq = &bqempty;
1618 	BQ_LOCK(bq);
1619         for (i = 0; i < cnt; i++) {
1620 		bp = store[i];
1621 		/* Inline bq_insert() to batch locking. */
1622 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1623 		bp->b_flags &= ~(B_AGE | B_REUSE);
1624 		bq->bq_len++;
1625 		bp->b_qindex = bq->bq_index;
1626 	}
1627 	BQ_UNLOCK(bq);
1628 }
1629 
1630 /*
1631  * buf_alloc:
1632  *
1633  *	Allocate an empty buffer header.
1634  */
1635 static struct buf *
1636 buf_alloc(struct bufdomain *bd)
1637 {
1638 	struct buf *bp;
1639 	int freebufs;
1640 
1641 	/*
1642 	 * We can only run out of bufs in the buf zone if the average buf
1643 	 * is less than BKVASIZE.  In this case the actual wait/block will
1644 	 * come from buf_reycle() failing to flush one of these small bufs.
1645 	 */
1646 	bp = NULL;
1647 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1648 	if (freebufs > 0)
1649 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1650 	if (bp == NULL) {
1651 		atomic_add_int(&bd->bd_freebuffers, 1);
1652 		bufspace_daemon_wakeup(bd);
1653 		counter_u64_add(numbufallocfails, 1);
1654 		return (NULL);
1655 	}
1656 	/*
1657 	 * Wake-up the bufspace daemon on transition below threshold.
1658 	 */
1659 	if (freebufs == bd->bd_lofreebuffers)
1660 		bufspace_daemon_wakeup(bd);
1661 
1662 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1663 		panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1664 
1665 	KASSERT(bp->b_vp == NULL,
1666 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1667 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1668 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1669 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1670 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1671 	KASSERT(bp->b_npages == 0,
1672 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1673 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1674 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1675 
1676 	bp->b_domain = BD_DOMAIN(bd);
1677 	bp->b_flags = 0;
1678 	bp->b_ioflags = 0;
1679 	bp->b_xflags = 0;
1680 	bp->b_vflags = 0;
1681 	bp->b_vp = NULL;
1682 	bp->b_blkno = bp->b_lblkno = 0;
1683 	bp->b_offset = NOOFFSET;
1684 	bp->b_iodone = 0;
1685 	bp->b_error = 0;
1686 	bp->b_resid = 0;
1687 	bp->b_bcount = 0;
1688 	bp->b_npages = 0;
1689 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1690 	bp->b_bufobj = NULL;
1691 	bp->b_data = bp->b_kvabase = unmapped_buf;
1692 	bp->b_fsprivate1 = NULL;
1693 	bp->b_fsprivate2 = NULL;
1694 	bp->b_fsprivate3 = NULL;
1695 	LIST_INIT(&bp->b_dep);
1696 
1697 	return (bp);
1698 }
1699 
1700 /*
1701  *	buf_recycle:
1702  *
1703  *	Free a buffer from the given bufqueue.  kva controls whether the
1704  *	freed buf must own some kva resources.  This is used for
1705  *	defragmenting.
1706  */
1707 static int
1708 buf_recycle(struct bufdomain *bd, bool kva)
1709 {
1710 	struct bufqueue *bq;
1711 	struct buf *bp, *nbp;
1712 
1713 	if (kva)
1714 		counter_u64_add(bufdefragcnt, 1);
1715 	nbp = NULL;
1716 	bq = bd->bd_cleanq;
1717 	BQ_LOCK(bq);
1718 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1719 	    ("buf_recycle: Locks don't match"));
1720 	nbp = TAILQ_FIRST(&bq->bq_queue);
1721 
1722 	/*
1723 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1724 	 * depending.
1725 	 */
1726 	while ((bp = nbp) != NULL) {
1727 		/*
1728 		 * Calculate next bp (we can only use it if we do not
1729 		 * release the bqlock).
1730 		 */
1731 		nbp = TAILQ_NEXT(bp, b_freelist);
1732 
1733 		/*
1734 		 * If we are defragging then we need a buffer with
1735 		 * some kva to reclaim.
1736 		 */
1737 		if (kva && bp->b_kvasize == 0)
1738 			continue;
1739 
1740 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1741 			continue;
1742 
1743 		/*
1744 		 * Implement a second chance algorithm for frequently
1745 		 * accessed buffers.
1746 		 */
1747 		if ((bp->b_flags & B_REUSE) != 0) {
1748 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1749 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1750 			bp->b_flags &= ~B_REUSE;
1751 			BUF_UNLOCK(bp);
1752 			continue;
1753 		}
1754 
1755 		/*
1756 		 * Skip buffers with background writes in progress.
1757 		 */
1758 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1759 			BUF_UNLOCK(bp);
1760 			continue;
1761 		}
1762 
1763 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1764 		    ("buf_recycle: inconsistent queue %d bp %p",
1765 		    bp->b_qindex, bp));
1766 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1767 		    ("getnewbuf: queue domain %d doesn't match request %d",
1768 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1769 		/*
1770 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1771 		 * the scan from this point on.
1772 		 */
1773 		bq_remove(bq, bp);
1774 		BQ_UNLOCK(bq);
1775 
1776 		/*
1777 		 * Requeue the background write buffer with error and
1778 		 * restart the scan.
1779 		 */
1780 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1781 			bqrelse(bp);
1782 			BQ_LOCK(bq);
1783 			nbp = TAILQ_FIRST(&bq->bq_queue);
1784 			continue;
1785 		}
1786 		bp->b_flags |= B_INVAL;
1787 		brelse(bp);
1788 		return (0);
1789 	}
1790 	bd->bd_wanted = 1;
1791 	BQ_UNLOCK(bq);
1792 
1793 	return (ENOBUFS);
1794 }
1795 
1796 /*
1797  *	bremfree:
1798  *
1799  *	Mark the buffer for removal from the appropriate free list.
1800  *
1801  */
1802 void
1803 bremfree(struct buf *bp)
1804 {
1805 
1806 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1807 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1808 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1809 	KASSERT(bp->b_qindex != QUEUE_NONE,
1810 	    ("bremfree: buffer %p not on a queue.", bp));
1811 	BUF_ASSERT_XLOCKED(bp);
1812 
1813 	bp->b_flags |= B_REMFREE;
1814 }
1815 
1816 /*
1817  *	bremfreef:
1818  *
1819  *	Force an immediate removal from a free list.  Used only in nfs when
1820  *	it abuses the b_freelist pointer.
1821  */
1822 void
1823 bremfreef(struct buf *bp)
1824 {
1825 	struct bufqueue *bq;
1826 
1827 	bq = bufqueue_acquire(bp);
1828 	bq_remove(bq, bp);
1829 	BQ_UNLOCK(bq);
1830 }
1831 
1832 static void
1833 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1834 {
1835 
1836 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1837 	TAILQ_INIT(&bq->bq_queue);
1838 	bq->bq_len = 0;
1839 	bq->bq_index = qindex;
1840 	bq->bq_subqueue = subqueue;
1841 }
1842 
1843 static void
1844 bd_init(struct bufdomain *bd)
1845 {
1846 	int i;
1847 
1848 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1849 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1850 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1851 	for (i = 0; i <= mp_maxid; i++)
1852 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1853 		    "bufq clean subqueue lock");
1854 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1855 }
1856 
1857 /*
1858  *	bq_remove:
1859  *
1860  *	Removes a buffer from the free list, must be called with the
1861  *	correct qlock held.
1862  */
1863 static void
1864 bq_remove(struct bufqueue *bq, struct buf *bp)
1865 {
1866 
1867 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1868 	    bp, bp->b_vp, bp->b_flags);
1869 	KASSERT(bp->b_qindex != QUEUE_NONE,
1870 	    ("bq_remove: buffer %p not on a queue.", bp));
1871 	KASSERT(bufqueue(bp) == bq,
1872 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1873 
1874 	BQ_ASSERT_LOCKED(bq);
1875 	if (bp->b_qindex != QUEUE_EMPTY) {
1876 		BUF_ASSERT_XLOCKED(bp);
1877 	}
1878 	KASSERT(bq->bq_len >= 1,
1879 	    ("queue %d underflow", bp->b_qindex));
1880 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1881 	bq->bq_len--;
1882 	bp->b_qindex = QUEUE_NONE;
1883 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1884 }
1885 
1886 static void
1887 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1888 {
1889 	struct buf *bp;
1890 
1891 	BQ_ASSERT_LOCKED(bq);
1892 	if (bq != bd->bd_cleanq) {
1893 		BD_LOCK(bd);
1894 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1895 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1896 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1897 			    b_freelist);
1898 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1899 		}
1900 		bd->bd_cleanq->bq_len += bq->bq_len;
1901 		bq->bq_len = 0;
1902 	}
1903 	if (bd->bd_wanted) {
1904 		bd->bd_wanted = 0;
1905 		wakeup(&bd->bd_wanted);
1906 	}
1907 	if (bq != bd->bd_cleanq)
1908 		BD_UNLOCK(bd);
1909 }
1910 
1911 static int
1912 bd_flushall(struct bufdomain *bd)
1913 {
1914 	struct bufqueue *bq;
1915 	int flushed;
1916 	int i;
1917 
1918 	if (bd->bd_lim == 0)
1919 		return (0);
1920 	flushed = 0;
1921 	for (i = 0; i <= mp_maxid; i++) {
1922 		bq = &bd->bd_subq[i];
1923 		if (bq->bq_len == 0)
1924 			continue;
1925 		BQ_LOCK(bq);
1926 		bd_flush(bd, bq);
1927 		BQ_UNLOCK(bq);
1928 		flushed++;
1929 	}
1930 
1931 	return (flushed);
1932 }
1933 
1934 static void
1935 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1936 {
1937 	struct bufdomain *bd;
1938 
1939 	if (bp->b_qindex != QUEUE_NONE)
1940 		panic("bq_insert: free buffer %p onto another queue?", bp);
1941 
1942 	bd = bufdomain(bp);
1943 	if (bp->b_flags & B_AGE) {
1944 		/* Place this buf directly on the real queue. */
1945 		if (bq->bq_index == QUEUE_CLEAN)
1946 			bq = bd->bd_cleanq;
1947 		BQ_LOCK(bq);
1948 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1949 	} else {
1950 		BQ_LOCK(bq);
1951 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1952 	}
1953 	bp->b_flags &= ~(B_AGE | B_REUSE);
1954 	bq->bq_len++;
1955 	bp->b_qindex = bq->bq_index;
1956 	bp->b_subqueue = bq->bq_subqueue;
1957 
1958 	/*
1959 	 * Unlock before we notify so that we don't wakeup a waiter that
1960 	 * fails a trylock on the buf and sleeps again.
1961 	 */
1962 	if (unlock)
1963 		BUF_UNLOCK(bp);
1964 
1965 	if (bp->b_qindex == QUEUE_CLEAN) {
1966 		/*
1967 		 * Flush the per-cpu queue and notify any waiters.
1968 		 */
1969 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1970 		    bq->bq_len >= bd->bd_lim))
1971 			bd_flush(bd, bq);
1972 	}
1973 	BQ_UNLOCK(bq);
1974 }
1975 
1976 /*
1977  *	bufkva_free:
1978  *
1979  *	Free the kva allocation for a buffer.
1980  *
1981  */
1982 static void
1983 bufkva_free(struct buf *bp)
1984 {
1985 
1986 #ifdef INVARIANTS
1987 	if (bp->b_kvasize == 0) {
1988 		KASSERT(bp->b_kvabase == unmapped_buf &&
1989 		    bp->b_data == unmapped_buf,
1990 		    ("Leaked KVA space on %p", bp));
1991 	} else if (buf_mapped(bp))
1992 		BUF_CHECK_MAPPED(bp);
1993 	else
1994 		BUF_CHECK_UNMAPPED(bp);
1995 #endif
1996 	if (bp->b_kvasize == 0)
1997 		return;
1998 
1999 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2000 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2001 	counter_u64_add(buffreekvacnt, 1);
2002 	bp->b_data = bp->b_kvabase = unmapped_buf;
2003 	bp->b_kvasize = 0;
2004 }
2005 
2006 /*
2007  *	bufkva_alloc:
2008  *
2009  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2010  */
2011 static int
2012 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2013 {
2014 	vm_offset_t addr;
2015 	int error;
2016 
2017 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2018 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2019 
2020 	bufkva_free(bp);
2021 
2022 	addr = 0;
2023 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2024 	if (error != 0) {
2025 		/*
2026 		 * Buffer map is too fragmented.  Request the caller
2027 		 * to defragment the map.
2028 		 */
2029 		return (error);
2030 	}
2031 	bp->b_kvabase = (caddr_t)addr;
2032 	bp->b_kvasize = maxsize;
2033 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2034 	if ((gbflags & GB_UNMAPPED) != 0) {
2035 		bp->b_data = unmapped_buf;
2036 		BUF_CHECK_UNMAPPED(bp);
2037 	} else {
2038 		bp->b_data = bp->b_kvabase;
2039 		BUF_CHECK_MAPPED(bp);
2040 	}
2041 	return (0);
2042 }
2043 
2044 /*
2045  *	bufkva_reclaim:
2046  *
2047  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2048  *	callback that fires to avoid returning failure.
2049  */
2050 static void
2051 bufkva_reclaim(vmem_t *vmem, int flags)
2052 {
2053 	bool done;
2054 	int q;
2055 	int i;
2056 
2057 	done = false;
2058 	for (i = 0; i < 5; i++) {
2059 		for (q = 0; q < buf_domains; q++)
2060 			if (buf_recycle(&bdomain[q], true) != 0)
2061 				done = true;
2062 		if (done)
2063 			break;
2064 	}
2065 	return;
2066 }
2067 
2068 /*
2069  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2070  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2071  * the buffer is valid and we do not have to do anything.
2072  */
2073 static void
2074 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2075     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2076 {
2077 	struct buf *rabp;
2078 	struct thread *td;
2079 	int i;
2080 
2081 	td = curthread;
2082 
2083 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2084 		if (inmem(vp, *rablkno))
2085 			continue;
2086 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2087 		if ((rabp->b_flags & B_CACHE) != 0) {
2088 			brelse(rabp);
2089 			continue;
2090 		}
2091 #ifdef RACCT
2092 		if (racct_enable) {
2093 			PROC_LOCK(curproc);
2094 			racct_add_buf(curproc, rabp, 0);
2095 			PROC_UNLOCK(curproc);
2096 		}
2097 #endif /* RACCT */
2098 		td->td_ru.ru_inblock++;
2099 		rabp->b_flags |= B_ASYNC;
2100 		rabp->b_flags &= ~B_INVAL;
2101 		if ((flags & GB_CKHASH) != 0) {
2102 			rabp->b_flags |= B_CKHASH;
2103 			rabp->b_ckhashcalc = ckhashfunc;
2104 		}
2105 		rabp->b_ioflags &= ~BIO_ERROR;
2106 		rabp->b_iocmd = BIO_READ;
2107 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2108 			rabp->b_rcred = crhold(cred);
2109 		vfs_busy_pages(rabp, 0);
2110 		BUF_KERNPROC(rabp);
2111 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2112 		bstrategy(rabp);
2113 	}
2114 }
2115 
2116 /*
2117  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2118  *
2119  * Get a buffer with the specified data.  Look in the cache first.  We
2120  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2121  * is set, the buffer is valid and we do not have to do anything, see
2122  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2123  *
2124  * Always return a NULL buffer pointer (in bpp) when returning an error.
2125  */
2126 int
2127 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
2128     int *rabsize, int cnt, struct ucred *cred, int flags,
2129     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2130 {
2131 	struct buf *bp;
2132 	struct thread *td;
2133 	int error, readwait, rv;
2134 
2135 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2136 	td = curthread;
2137 	/*
2138 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2139 	 * are specified.
2140 	 */
2141 	error = getblkx(vp, blkno, size, 0, 0, flags, &bp);
2142 	if (error != 0) {
2143 		*bpp = NULL;
2144 		return (error);
2145 	}
2146 	flags &= ~GB_NOSPARSE;
2147 	*bpp = bp;
2148 
2149 	/*
2150 	 * If not found in cache, do some I/O
2151 	 */
2152 	readwait = 0;
2153 	if ((bp->b_flags & B_CACHE) == 0) {
2154 #ifdef RACCT
2155 		if (racct_enable) {
2156 			PROC_LOCK(td->td_proc);
2157 			racct_add_buf(td->td_proc, bp, 0);
2158 			PROC_UNLOCK(td->td_proc);
2159 		}
2160 #endif /* RACCT */
2161 		td->td_ru.ru_inblock++;
2162 		bp->b_iocmd = BIO_READ;
2163 		bp->b_flags &= ~B_INVAL;
2164 		if ((flags & GB_CKHASH) != 0) {
2165 			bp->b_flags |= B_CKHASH;
2166 			bp->b_ckhashcalc = ckhashfunc;
2167 		}
2168 		bp->b_ioflags &= ~BIO_ERROR;
2169 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2170 			bp->b_rcred = crhold(cred);
2171 		vfs_busy_pages(bp, 0);
2172 		bp->b_iooffset = dbtob(bp->b_blkno);
2173 		bstrategy(bp);
2174 		++readwait;
2175 	}
2176 
2177 	/*
2178 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2179 	 */
2180 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2181 
2182 	rv = 0;
2183 	if (readwait) {
2184 		rv = bufwait(bp);
2185 		if (rv != 0) {
2186 			brelse(bp);
2187 			*bpp = NULL;
2188 		}
2189 	}
2190 	return (rv);
2191 }
2192 
2193 /*
2194  * Write, release buffer on completion.  (Done by iodone
2195  * if async).  Do not bother writing anything if the buffer
2196  * is invalid.
2197  *
2198  * Note that we set B_CACHE here, indicating that buffer is
2199  * fully valid and thus cacheable.  This is true even of NFS
2200  * now so we set it generally.  This could be set either here
2201  * or in biodone() since the I/O is synchronous.  We put it
2202  * here.
2203  */
2204 int
2205 bufwrite(struct buf *bp)
2206 {
2207 	int oldflags;
2208 	struct vnode *vp;
2209 	long space;
2210 	int vp_md;
2211 
2212 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2213 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2214 		bp->b_flags |= B_INVAL | B_RELBUF;
2215 		bp->b_flags &= ~B_CACHE;
2216 		brelse(bp);
2217 		return (ENXIO);
2218 	}
2219 	if (bp->b_flags & B_INVAL) {
2220 		brelse(bp);
2221 		return (0);
2222 	}
2223 
2224 	if (bp->b_flags & B_BARRIER)
2225 		atomic_add_long(&barrierwrites, 1);
2226 
2227 	oldflags = bp->b_flags;
2228 
2229 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2230 	    ("FFS background buffer should not get here %p", bp));
2231 
2232 	vp = bp->b_vp;
2233 	if (vp)
2234 		vp_md = vp->v_vflag & VV_MD;
2235 	else
2236 		vp_md = 0;
2237 
2238 	/*
2239 	 * Mark the buffer clean.  Increment the bufobj write count
2240 	 * before bundirty() call, to prevent other thread from seeing
2241 	 * empty dirty list and zero counter for writes in progress,
2242 	 * falsely indicating that the bufobj is clean.
2243 	 */
2244 	bufobj_wref(bp->b_bufobj);
2245 	bundirty(bp);
2246 
2247 	bp->b_flags &= ~B_DONE;
2248 	bp->b_ioflags &= ~BIO_ERROR;
2249 	bp->b_flags |= B_CACHE;
2250 	bp->b_iocmd = BIO_WRITE;
2251 
2252 	vfs_busy_pages(bp, 1);
2253 
2254 	/*
2255 	 * Normal bwrites pipeline writes
2256 	 */
2257 	bp->b_runningbufspace = bp->b_bufsize;
2258 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2259 
2260 #ifdef RACCT
2261 	if (racct_enable) {
2262 		PROC_LOCK(curproc);
2263 		racct_add_buf(curproc, bp, 1);
2264 		PROC_UNLOCK(curproc);
2265 	}
2266 #endif /* RACCT */
2267 	curthread->td_ru.ru_oublock++;
2268 	if (oldflags & B_ASYNC)
2269 		BUF_KERNPROC(bp);
2270 	bp->b_iooffset = dbtob(bp->b_blkno);
2271 	buf_track(bp, __func__);
2272 	bstrategy(bp);
2273 
2274 	if ((oldflags & B_ASYNC) == 0) {
2275 		int rtval = bufwait(bp);
2276 		brelse(bp);
2277 		return (rtval);
2278 	} else if (space > hirunningspace) {
2279 		/*
2280 		 * don't allow the async write to saturate the I/O
2281 		 * system.  We will not deadlock here because
2282 		 * we are blocking waiting for I/O that is already in-progress
2283 		 * to complete. We do not block here if it is the update
2284 		 * or syncer daemon trying to clean up as that can lead
2285 		 * to deadlock.
2286 		 */
2287 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2288 			waitrunningbufspace();
2289 	}
2290 
2291 	return (0);
2292 }
2293 
2294 void
2295 bufbdflush(struct bufobj *bo, struct buf *bp)
2296 {
2297 	struct buf *nbp;
2298 
2299 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2300 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2301 		altbufferflushes++;
2302 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2303 		BO_LOCK(bo);
2304 		/*
2305 		 * Try to find a buffer to flush.
2306 		 */
2307 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2308 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2309 			    BUF_LOCK(nbp,
2310 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2311 				continue;
2312 			if (bp == nbp)
2313 				panic("bdwrite: found ourselves");
2314 			BO_UNLOCK(bo);
2315 			/* Don't countdeps with the bo lock held. */
2316 			if (buf_countdeps(nbp, 0)) {
2317 				BO_LOCK(bo);
2318 				BUF_UNLOCK(nbp);
2319 				continue;
2320 			}
2321 			if (nbp->b_flags & B_CLUSTEROK) {
2322 				vfs_bio_awrite(nbp);
2323 			} else {
2324 				bremfree(nbp);
2325 				bawrite(nbp);
2326 			}
2327 			dirtybufferflushes++;
2328 			break;
2329 		}
2330 		if (nbp == NULL)
2331 			BO_UNLOCK(bo);
2332 	}
2333 }
2334 
2335 /*
2336  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2337  * anything if the buffer is marked invalid.
2338  *
2339  * Note that since the buffer must be completely valid, we can safely
2340  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2341  * biodone() in order to prevent getblk from writing the buffer
2342  * out synchronously.
2343  */
2344 void
2345 bdwrite(struct buf *bp)
2346 {
2347 	struct thread *td = curthread;
2348 	struct vnode *vp;
2349 	struct bufobj *bo;
2350 
2351 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2352 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2353 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2354 	    ("Barrier request in delayed write %p", bp));
2355 
2356 	if (bp->b_flags & B_INVAL) {
2357 		brelse(bp);
2358 		return;
2359 	}
2360 
2361 	/*
2362 	 * If we have too many dirty buffers, don't create any more.
2363 	 * If we are wildly over our limit, then force a complete
2364 	 * cleanup. Otherwise, just keep the situation from getting
2365 	 * out of control. Note that we have to avoid a recursive
2366 	 * disaster and not try to clean up after our own cleanup!
2367 	 */
2368 	vp = bp->b_vp;
2369 	bo = bp->b_bufobj;
2370 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2371 		td->td_pflags |= TDP_INBDFLUSH;
2372 		BO_BDFLUSH(bo, bp);
2373 		td->td_pflags &= ~TDP_INBDFLUSH;
2374 	} else
2375 		recursiveflushes++;
2376 
2377 	bdirty(bp);
2378 	/*
2379 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2380 	 * true even of NFS now.
2381 	 */
2382 	bp->b_flags |= B_CACHE;
2383 
2384 	/*
2385 	 * This bmap keeps the system from needing to do the bmap later,
2386 	 * perhaps when the system is attempting to do a sync.  Since it
2387 	 * is likely that the indirect block -- or whatever other datastructure
2388 	 * that the filesystem needs is still in memory now, it is a good
2389 	 * thing to do this.  Note also, that if the pageout daemon is
2390 	 * requesting a sync -- there might not be enough memory to do
2391 	 * the bmap then...  So, this is important to do.
2392 	 */
2393 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2394 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2395 	}
2396 
2397 	buf_track(bp, __func__);
2398 
2399 	/*
2400 	 * Set the *dirty* buffer range based upon the VM system dirty
2401 	 * pages.
2402 	 *
2403 	 * Mark the buffer pages as clean.  We need to do this here to
2404 	 * satisfy the vnode_pager and the pageout daemon, so that it
2405 	 * thinks that the pages have been "cleaned".  Note that since
2406 	 * the pages are in a delayed write buffer -- the VFS layer
2407 	 * "will" see that the pages get written out on the next sync,
2408 	 * or perhaps the cluster will be completed.
2409 	 */
2410 	vfs_clean_pages_dirty_buf(bp);
2411 	bqrelse(bp);
2412 
2413 	/*
2414 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2415 	 * due to the softdep code.
2416 	 */
2417 }
2418 
2419 /*
2420  *	bdirty:
2421  *
2422  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2423  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2424  *	itself to properly update it in the dirty/clean lists.  We mark it
2425  *	B_DONE to ensure that any asynchronization of the buffer properly
2426  *	clears B_DONE ( else a panic will occur later ).
2427  *
2428  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2429  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2430  *	should only be called if the buffer is known-good.
2431  *
2432  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2433  *	count.
2434  *
2435  *	The buffer must be on QUEUE_NONE.
2436  */
2437 void
2438 bdirty(struct buf *bp)
2439 {
2440 
2441 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2442 	    bp, bp->b_vp, bp->b_flags);
2443 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2444 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2445 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2446 	bp->b_flags &= ~(B_RELBUF);
2447 	bp->b_iocmd = BIO_WRITE;
2448 
2449 	if ((bp->b_flags & B_DELWRI) == 0) {
2450 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2451 		reassignbuf(bp);
2452 		bdirtyadd(bp);
2453 	}
2454 }
2455 
2456 /*
2457  *	bundirty:
2458  *
2459  *	Clear B_DELWRI for buffer.
2460  *
2461  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2462  *	count.
2463  *
2464  *	The buffer must be on QUEUE_NONE.
2465  */
2466 
2467 void
2468 bundirty(struct buf *bp)
2469 {
2470 
2471 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2472 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2473 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2474 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2475 
2476 	if (bp->b_flags & B_DELWRI) {
2477 		bp->b_flags &= ~B_DELWRI;
2478 		reassignbuf(bp);
2479 		bdirtysub(bp);
2480 	}
2481 	/*
2482 	 * Since it is now being written, we can clear its deferred write flag.
2483 	 */
2484 	bp->b_flags &= ~B_DEFERRED;
2485 }
2486 
2487 /*
2488  *	bawrite:
2489  *
2490  *	Asynchronous write.  Start output on a buffer, but do not wait for
2491  *	it to complete.  The buffer is released when the output completes.
2492  *
2493  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2494  *	B_INVAL buffers.  Not us.
2495  */
2496 void
2497 bawrite(struct buf *bp)
2498 {
2499 
2500 	bp->b_flags |= B_ASYNC;
2501 	(void) bwrite(bp);
2502 }
2503 
2504 /*
2505  *	babarrierwrite:
2506  *
2507  *	Asynchronous barrier write.  Start output on a buffer, but do not
2508  *	wait for it to complete.  Place a write barrier after this write so
2509  *	that this buffer and all buffers written before it are committed to
2510  *	the disk before any buffers written after this write are committed
2511  *	to the disk.  The buffer is released when the output completes.
2512  */
2513 void
2514 babarrierwrite(struct buf *bp)
2515 {
2516 
2517 	bp->b_flags |= B_ASYNC | B_BARRIER;
2518 	(void) bwrite(bp);
2519 }
2520 
2521 /*
2522  *	bbarrierwrite:
2523  *
2524  *	Synchronous barrier write.  Start output on a buffer and wait for
2525  *	it to complete.  Place a write barrier after this write so that
2526  *	this buffer and all buffers written before it are committed to
2527  *	the disk before any buffers written after this write are committed
2528  *	to the disk.  The buffer is released when the output completes.
2529  */
2530 int
2531 bbarrierwrite(struct buf *bp)
2532 {
2533 
2534 	bp->b_flags |= B_BARRIER;
2535 	return (bwrite(bp));
2536 }
2537 
2538 /*
2539  *	bwillwrite:
2540  *
2541  *	Called prior to the locking of any vnodes when we are expecting to
2542  *	write.  We do not want to starve the buffer cache with too many
2543  *	dirty buffers so we block here.  By blocking prior to the locking
2544  *	of any vnodes we attempt to avoid the situation where a locked vnode
2545  *	prevents the various system daemons from flushing related buffers.
2546  */
2547 void
2548 bwillwrite(void)
2549 {
2550 
2551 	if (buf_dirty_count_severe()) {
2552 		mtx_lock(&bdirtylock);
2553 		while (buf_dirty_count_severe()) {
2554 			bdirtywait = 1;
2555 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2556 			    "flswai", 0);
2557 		}
2558 		mtx_unlock(&bdirtylock);
2559 	}
2560 }
2561 
2562 /*
2563  * Return true if we have too many dirty buffers.
2564  */
2565 int
2566 buf_dirty_count_severe(void)
2567 {
2568 
2569 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2570 }
2571 
2572 /*
2573  *	brelse:
2574  *
2575  *	Release a busy buffer and, if requested, free its resources.  The
2576  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2577  *	to be accessed later as a cache entity or reused for other purposes.
2578  */
2579 void
2580 brelse(struct buf *bp)
2581 {
2582 	struct mount *v_mnt;
2583 	int qindex;
2584 
2585 	/*
2586 	 * Many functions erroneously call brelse with a NULL bp under rare
2587 	 * error conditions. Simply return when called with a NULL bp.
2588 	 */
2589 	if (bp == NULL)
2590 		return;
2591 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2592 	    bp, bp->b_vp, bp->b_flags);
2593 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2594 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2595 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2596 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2597 
2598 	if (BUF_LOCKRECURSED(bp)) {
2599 		/*
2600 		 * Do not process, in particular, do not handle the
2601 		 * B_INVAL/B_RELBUF and do not release to free list.
2602 		 */
2603 		BUF_UNLOCK(bp);
2604 		return;
2605 	}
2606 
2607 	if (bp->b_flags & B_MANAGED) {
2608 		bqrelse(bp);
2609 		return;
2610 	}
2611 
2612 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2613 		BO_LOCK(bp->b_bufobj);
2614 		bp->b_vflags &= ~BV_BKGRDERR;
2615 		BO_UNLOCK(bp->b_bufobj);
2616 		bdirty(bp);
2617 	}
2618 
2619 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2620 	    (bp->b_flags & B_INVALONERR)) {
2621 		/*
2622 		 * Forced invalidation of dirty buffer contents, to be used
2623 		 * after a failed write in the rare case that the loss of the
2624 		 * contents is acceptable.  The buffer is invalidated and
2625 		 * freed.
2626 		 */
2627 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2628 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2629 	}
2630 
2631 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2632 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2633 	    !(bp->b_flags & B_INVAL)) {
2634 		/*
2635 		 * Failed write, redirty.  All errors except ENXIO (which
2636 		 * means the device is gone) are treated as being
2637 		 * transient.
2638 		 *
2639 		 * XXX Treating EIO as transient is not correct; the
2640 		 * contract with the local storage device drivers is that
2641 		 * they will only return EIO once the I/O is no longer
2642 		 * retriable.  Network I/O also respects this through the
2643 		 * guarantees of TCP and/or the internal retries of NFS.
2644 		 * ENOMEM might be transient, but we also have no way of
2645 		 * knowing when its ok to retry/reschedule.  In general,
2646 		 * this entire case should be made obsolete through better
2647 		 * error handling/recovery and resource scheduling.
2648 		 *
2649 		 * Do this also for buffers that failed with ENXIO, but have
2650 		 * non-empty dependencies - the soft updates code might need
2651 		 * to access the buffer to untangle them.
2652 		 *
2653 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2654 		 */
2655 		bp->b_ioflags &= ~BIO_ERROR;
2656 		bdirty(bp);
2657 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2658 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2659 		/*
2660 		 * Either a failed read I/O, or we were asked to free or not
2661 		 * cache the buffer, or we failed to write to a device that's
2662 		 * no longer present.
2663 		 */
2664 		bp->b_flags |= B_INVAL;
2665 		if (!LIST_EMPTY(&bp->b_dep))
2666 			buf_deallocate(bp);
2667 		if (bp->b_flags & B_DELWRI)
2668 			bdirtysub(bp);
2669 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2670 		if ((bp->b_flags & B_VMIO) == 0) {
2671 			allocbuf(bp, 0);
2672 			if (bp->b_vp)
2673 				brelvp(bp);
2674 		}
2675 	}
2676 
2677 	/*
2678 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2679 	 * is called with B_DELWRI set, the underlying pages may wind up
2680 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2681 	 * because pages associated with a B_DELWRI bp are marked clean.
2682 	 *
2683 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2684 	 * if B_DELWRI is set.
2685 	 */
2686 	if (bp->b_flags & B_DELWRI)
2687 		bp->b_flags &= ~B_RELBUF;
2688 
2689 	/*
2690 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2691 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2692 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2693 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2694 	 *
2695 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2696 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2697 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2698 	 *
2699 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2700 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2701 	 * the commit state and we cannot afford to lose the buffer. If the
2702 	 * buffer has a background write in progress, we need to keep it
2703 	 * around to prevent it from being reconstituted and starting a second
2704 	 * background write.
2705 	 */
2706 
2707 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2708 
2709 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2710 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2711 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2712 	    vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2713 		vfs_vmio_invalidate(bp);
2714 		allocbuf(bp, 0);
2715 	}
2716 
2717 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2718 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2719 		allocbuf(bp, 0);
2720 		bp->b_flags &= ~B_NOREUSE;
2721 		if (bp->b_vp != NULL)
2722 			brelvp(bp);
2723 	}
2724 
2725 	/*
2726 	 * If the buffer has junk contents signal it and eventually
2727 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2728 	 * doesn't find it.
2729 	 */
2730 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2731 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2732 		bp->b_flags |= B_INVAL;
2733 	if (bp->b_flags & B_INVAL) {
2734 		if (bp->b_flags & B_DELWRI)
2735 			bundirty(bp);
2736 		if (bp->b_vp)
2737 			brelvp(bp);
2738 	}
2739 
2740 	buf_track(bp, __func__);
2741 
2742 	/* buffers with no memory */
2743 	if (bp->b_bufsize == 0) {
2744 		buf_free(bp);
2745 		return;
2746 	}
2747 	/* buffers with junk contents */
2748 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2749 	    (bp->b_ioflags & BIO_ERROR)) {
2750 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2751 		if (bp->b_vflags & BV_BKGRDINPROG)
2752 			panic("losing buffer 2");
2753 		qindex = QUEUE_CLEAN;
2754 		bp->b_flags |= B_AGE;
2755 	/* remaining buffers */
2756 	} else if (bp->b_flags & B_DELWRI)
2757 		qindex = QUEUE_DIRTY;
2758 	else
2759 		qindex = QUEUE_CLEAN;
2760 
2761 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2762 		panic("brelse: not dirty");
2763 
2764 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2765 	/* binsfree unlocks bp. */
2766 	binsfree(bp, qindex);
2767 }
2768 
2769 /*
2770  * Release a buffer back to the appropriate queue but do not try to free
2771  * it.  The buffer is expected to be used again soon.
2772  *
2773  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2774  * biodone() to requeue an async I/O on completion.  It is also used when
2775  * known good buffers need to be requeued but we think we may need the data
2776  * again soon.
2777  *
2778  * XXX we should be able to leave the B_RELBUF hint set on completion.
2779  */
2780 void
2781 bqrelse(struct buf *bp)
2782 {
2783 	int qindex;
2784 
2785 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2786 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2787 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2788 
2789 	qindex = QUEUE_NONE;
2790 	if (BUF_LOCKRECURSED(bp)) {
2791 		/* do not release to free list */
2792 		BUF_UNLOCK(bp);
2793 		return;
2794 	}
2795 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2796 
2797 	if (bp->b_flags & B_MANAGED) {
2798 		if (bp->b_flags & B_REMFREE)
2799 			bremfreef(bp);
2800 		goto out;
2801 	}
2802 
2803 	/* buffers with stale but valid contents */
2804 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2805 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2806 		BO_LOCK(bp->b_bufobj);
2807 		bp->b_vflags &= ~BV_BKGRDERR;
2808 		BO_UNLOCK(bp->b_bufobj);
2809 		qindex = QUEUE_DIRTY;
2810 	} else {
2811 		if ((bp->b_flags & B_DELWRI) == 0 &&
2812 		    (bp->b_xflags & BX_VNDIRTY))
2813 			panic("bqrelse: not dirty");
2814 		if ((bp->b_flags & B_NOREUSE) != 0) {
2815 			brelse(bp);
2816 			return;
2817 		}
2818 		qindex = QUEUE_CLEAN;
2819 	}
2820 	buf_track(bp, __func__);
2821 	/* binsfree unlocks bp. */
2822 	binsfree(bp, qindex);
2823 	return;
2824 
2825 out:
2826 	buf_track(bp, __func__);
2827 	/* unlock */
2828 	BUF_UNLOCK(bp);
2829 }
2830 
2831 /*
2832  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2833  * restore bogus pages.
2834  */
2835 static void
2836 vfs_vmio_iodone(struct buf *bp)
2837 {
2838 	vm_ooffset_t foff;
2839 	vm_page_t m;
2840 	vm_object_t obj;
2841 	struct vnode *vp __unused;
2842 	int i, iosize, resid;
2843 	bool bogus;
2844 
2845 	obj = bp->b_bufobj->bo_object;
2846 	KASSERT(REFCOUNT_COUNT(obj->paging_in_progress) >= bp->b_npages,
2847 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2848 	    REFCOUNT_COUNT(obj->paging_in_progress), bp->b_npages));
2849 
2850 	vp = bp->b_vp;
2851 	KASSERT(vp->v_holdcnt > 0,
2852 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2853 	KASSERT(vp->v_object != NULL,
2854 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2855 
2856 	foff = bp->b_offset;
2857 	KASSERT(bp->b_offset != NOOFFSET,
2858 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2859 
2860 	bogus = false;
2861 	iosize = bp->b_bcount - bp->b_resid;
2862 	VM_OBJECT_WLOCK(obj);
2863 	for (i = 0; i < bp->b_npages; i++) {
2864 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2865 		if (resid > iosize)
2866 			resid = iosize;
2867 
2868 		/*
2869 		 * cleanup bogus pages, restoring the originals
2870 		 */
2871 		m = bp->b_pages[i];
2872 		if (m == bogus_page) {
2873 			bogus = true;
2874 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2875 			if (m == NULL)
2876 				panic("biodone: page disappeared!");
2877 			bp->b_pages[i] = m;
2878 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2879 			/*
2880 			 * In the write case, the valid and clean bits are
2881 			 * already changed correctly ( see bdwrite() ), so we
2882 			 * only need to do this here in the read case.
2883 			 */
2884 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2885 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2886 			    "has unexpected dirty bits", m));
2887 			vfs_page_set_valid(bp, foff, m);
2888 		}
2889 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2890 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2891 		    (intmax_t)foff, (uintmax_t)m->pindex));
2892 
2893 		vm_page_sunbusy(m);
2894 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2895 		iosize -= resid;
2896 	}
2897 	vm_object_pip_wakeupn(obj, bp->b_npages);
2898 	VM_OBJECT_WUNLOCK(obj);
2899 	if (bogus && buf_mapped(bp)) {
2900 		BUF_CHECK_MAPPED(bp);
2901 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2902 		    bp->b_pages, bp->b_npages);
2903 	}
2904 }
2905 
2906 /*
2907  * Perform page invalidation when a buffer is released.  The fully invalid
2908  * pages will be reclaimed later in vfs_vmio_truncate().
2909  */
2910 static void
2911 vfs_vmio_invalidate(struct buf *bp)
2912 {
2913 	vm_object_t obj;
2914 	vm_page_t m;
2915 	int flags, i, resid, poffset, presid;
2916 
2917 	if (buf_mapped(bp)) {
2918 		BUF_CHECK_MAPPED(bp);
2919 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2920 	} else
2921 		BUF_CHECK_UNMAPPED(bp);
2922 	/*
2923 	 * Get the base offset and length of the buffer.  Note that
2924 	 * in the VMIO case if the buffer block size is not
2925 	 * page-aligned then b_data pointer may not be page-aligned.
2926 	 * But our b_pages[] array *IS* page aligned.
2927 	 *
2928 	 * block sizes less then DEV_BSIZE (usually 512) are not
2929 	 * supported due to the page granularity bits (m->valid,
2930 	 * m->dirty, etc...).
2931 	 *
2932 	 * See man buf(9) for more information
2933 	 */
2934 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2935 	obj = bp->b_bufobj->bo_object;
2936 	resid = bp->b_bufsize;
2937 	poffset = bp->b_offset & PAGE_MASK;
2938 	VM_OBJECT_WLOCK(obj);
2939 	for (i = 0; i < bp->b_npages; i++) {
2940 		m = bp->b_pages[i];
2941 		if (m == bogus_page)
2942 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2943 		bp->b_pages[i] = NULL;
2944 
2945 		presid = resid > (PAGE_SIZE - poffset) ?
2946 		    (PAGE_SIZE - poffset) : resid;
2947 		KASSERT(presid >= 0, ("brelse: extra page"));
2948 		while (vm_page_xbusied(m))
2949 			vm_page_sleep_if_xbusy(m, "mbncsh");
2950 		if (pmap_page_wired_mappings(m) == 0)
2951 			vm_page_set_invalid(m, poffset, presid);
2952 		vm_page_release_locked(m, flags);
2953 		resid -= presid;
2954 		poffset = 0;
2955 	}
2956 	VM_OBJECT_WUNLOCK(obj);
2957 	bp->b_npages = 0;
2958 }
2959 
2960 /*
2961  * Page-granular truncation of an existing VMIO buffer.
2962  */
2963 static void
2964 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2965 {
2966 	vm_object_t obj;
2967 	vm_page_t m;
2968 	int flags, i;
2969 
2970 	if (bp->b_npages == desiredpages)
2971 		return;
2972 
2973 	if (buf_mapped(bp)) {
2974 		BUF_CHECK_MAPPED(bp);
2975 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2976 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2977 	} else
2978 		BUF_CHECK_UNMAPPED(bp);
2979 
2980 	/*
2981 	 * The object lock is needed only if we will attempt to free pages.
2982 	 */
2983 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2984 	if ((bp->b_flags & B_DIRECT) != 0) {
2985 		flags |= VPR_TRYFREE;
2986 		obj = bp->b_bufobj->bo_object;
2987 		VM_OBJECT_WLOCK(obj);
2988 	} else {
2989 		obj = NULL;
2990 	}
2991 	for (i = desiredpages; i < bp->b_npages; i++) {
2992 		m = bp->b_pages[i];
2993 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
2994 		bp->b_pages[i] = NULL;
2995 		if (obj != NULL)
2996 			vm_page_release_locked(m, flags);
2997 		else
2998 			vm_page_release(m, flags);
2999 	}
3000 	if (obj != NULL)
3001 		VM_OBJECT_WUNLOCK(obj);
3002 	bp->b_npages = desiredpages;
3003 }
3004 
3005 /*
3006  * Byte granular extension of VMIO buffers.
3007  */
3008 static void
3009 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3010 {
3011 	/*
3012 	 * We are growing the buffer, possibly in a
3013 	 * byte-granular fashion.
3014 	 */
3015 	vm_object_t obj;
3016 	vm_offset_t toff;
3017 	vm_offset_t tinc;
3018 	vm_page_t m;
3019 
3020 	/*
3021 	 * Step 1, bring in the VM pages from the object, allocating
3022 	 * them if necessary.  We must clear B_CACHE if these pages
3023 	 * are not valid for the range covered by the buffer.
3024 	 */
3025 	obj = bp->b_bufobj->bo_object;
3026 	VM_OBJECT_WLOCK(obj);
3027 	if (bp->b_npages < desiredpages) {
3028 		/*
3029 		 * We must allocate system pages since blocking
3030 		 * here could interfere with paging I/O, no
3031 		 * matter which process we are.
3032 		 *
3033 		 * Only exclusive busy can be tested here.
3034 		 * Blocking on shared busy might lead to
3035 		 * deadlocks once allocbuf() is called after
3036 		 * pages are vfs_busy_pages().
3037 		 */
3038 		(void)vm_page_grab_pages(obj,
3039 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3040 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3041 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3042 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3043 		bp->b_npages = desiredpages;
3044 	}
3045 
3046 	/*
3047 	 * Step 2.  We've loaded the pages into the buffer,
3048 	 * we have to figure out if we can still have B_CACHE
3049 	 * set.  Note that B_CACHE is set according to the
3050 	 * byte-granular range ( bcount and size ), not the
3051 	 * aligned range ( newbsize ).
3052 	 *
3053 	 * The VM test is against m->valid, which is DEV_BSIZE
3054 	 * aligned.  Needless to say, the validity of the data
3055 	 * needs to also be DEV_BSIZE aligned.  Note that this
3056 	 * fails with NFS if the server or some other client
3057 	 * extends the file's EOF.  If our buffer is resized,
3058 	 * B_CACHE may remain set! XXX
3059 	 */
3060 	toff = bp->b_bcount;
3061 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3062 	while ((bp->b_flags & B_CACHE) && toff < size) {
3063 		vm_pindex_t pi;
3064 
3065 		if (tinc > (size - toff))
3066 			tinc = size - toff;
3067 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3068 		m = bp->b_pages[pi];
3069 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3070 		toff += tinc;
3071 		tinc = PAGE_SIZE;
3072 	}
3073 	VM_OBJECT_WUNLOCK(obj);
3074 
3075 	/*
3076 	 * Step 3, fixup the KVA pmap.
3077 	 */
3078 	if (buf_mapped(bp))
3079 		bpmap_qenter(bp);
3080 	else
3081 		BUF_CHECK_UNMAPPED(bp);
3082 }
3083 
3084 /*
3085  * Check to see if a block at a particular lbn is available for a clustered
3086  * write.
3087  */
3088 static int
3089 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3090 {
3091 	struct buf *bpa;
3092 	int match;
3093 
3094 	match = 0;
3095 
3096 	/* If the buf isn't in core skip it */
3097 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3098 		return (0);
3099 
3100 	/* If the buf is busy we don't want to wait for it */
3101 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3102 		return (0);
3103 
3104 	/* Only cluster with valid clusterable delayed write buffers */
3105 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3106 	    (B_DELWRI | B_CLUSTEROK))
3107 		goto done;
3108 
3109 	if (bpa->b_bufsize != size)
3110 		goto done;
3111 
3112 	/*
3113 	 * Check to see if it is in the expected place on disk and that the
3114 	 * block has been mapped.
3115 	 */
3116 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3117 		match = 1;
3118 done:
3119 	BUF_UNLOCK(bpa);
3120 	return (match);
3121 }
3122 
3123 /*
3124  *	vfs_bio_awrite:
3125  *
3126  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3127  *	This is much better then the old way of writing only one buffer at
3128  *	a time.  Note that we may not be presented with the buffers in the
3129  *	correct order, so we search for the cluster in both directions.
3130  */
3131 int
3132 vfs_bio_awrite(struct buf *bp)
3133 {
3134 	struct bufobj *bo;
3135 	int i;
3136 	int j;
3137 	daddr_t lblkno = bp->b_lblkno;
3138 	struct vnode *vp = bp->b_vp;
3139 	int ncl;
3140 	int nwritten;
3141 	int size;
3142 	int maxcl;
3143 	int gbflags;
3144 
3145 	bo = &vp->v_bufobj;
3146 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3147 	/*
3148 	 * right now we support clustered writing only to regular files.  If
3149 	 * we find a clusterable block we could be in the middle of a cluster
3150 	 * rather then at the beginning.
3151 	 */
3152 	if ((vp->v_type == VREG) &&
3153 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3154 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3155 
3156 		size = vp->v_mount->mnt_stat.f_iosize;
3157 		maxcl = MAXPHYS / size;
3158 
3159 		BO_RLOCK(bo);
3160 		for (i = 1; i < maxcl; i++)
3161 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3162 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3163 				break;
3164 
3165 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3166 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3167 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3168 				break;
3169 		BO_RUNLOCK(bo);
3170 		--j;
3171 		ncl = i + j;
3172 		/*
3173 		 * this is a possible cluster write
3174 		 */
3175 		if (ncl != 1) {
3176 			BUF_UNLOCK(bp);
3177 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3178 			    gbflags);
3179 			return (nwritten);
3180 		}
3181 	}
3182 	bremfree(bp);
3183 	bp->b_flags |= B_ASYNC;
3184 	/*
3185 	 * default (old) behavior, writing out only one block
3186 	 *
3187 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3188 	 */
3189 	nwritten = bp->b_bufsize;
3190 	(void) bwrite(bp);
3191 
3192 	return (nwritten);
3193 }
3194 
3195 /*
3196  *	getnewbuf_kva:
3197  *
3198  *	Allocate KVA for an empty buf header according to gbflags.
3199  */
3200 static int
3201 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3202 {
3203 
3204 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3205 		/*
3206 		 * In order to keep fragmentation sane we only allocate kva
3207 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3208 		 */
3209 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3210 
3211 		if (maxsize != bp->b_kvasize &&
3212 		    bufkva_alloc(bp, maxsize, gbflags))
3213 			return (ENOSPC);
3214 	}
3215 	return (0);
3216 }
3217 
3218 /*
3219  *	getnewbuf:
3220  *
3221  *	Find and initialize a new buffer header, freeing up existing buffers
3222  *	in the bufqueues as necessary.  The new buffer is returned locked.
3223  *
3224  *	We block if:
3225  *		We have insufficient buffer headers
3226  *		We have insufficient buffer space
3227  *		buffer_arena is too fragmented ( space reservation fails )
3228  *		If we have to flush dirty buffers ( but we try to avoid this )
3229  *
3230  *	The caller is responsible for releasing the reserved bufspace after
3231  *	allocbuf() is called.
3232  */
3233 static struct buf *
3234 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3235 {
3236 	struct bufdomain *bd;
3237 	struct buf *bp;
3238 	bool metadata, reserved;
3239 
3240 	bp = NULL;
3241 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3242 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3243 	if (!unmapped_buf_allowed)
3244 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3245 
3246 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3247 	    vp->v_type == VCHR)
3248 		metadata = true;
3249 	else
3250 		metadata = false;
3251 	if (vp == NULL)
3252 		bd = &bdomain[0];
3253 	else
3254 		bd = &bdomain[vp->v_bufobj.bo_domain];
3255 
3256 	counter_u64_add(getnewbufcalls, 1);
3257 	reserved = false;
3258 	do {
3259 		if (reserved == false &&
3260 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3261 			counter_u64_add(getnewbufrestarts, 1);
3262 			continue;
3263 		}
3264 		reserved = true;
3265 		if ((bp = buf_alloc(bd)) == NULL) {
3266 			counter_u64_add(getnewbufrestarts, 1);
3267 			continue;
3268 		}
3269 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3270 			return (bp);
3271 		break;
3272 	} while (buf_recycle(bd, false) == 0);
3273 
3274 	if (reserved)
3275 		bufspace_release(bd, maxsize);
3276 	if (bp != NULL) {
3277 		bp->b_flags |= B_INVAL;
3278 		brelse(bp);
3279 	}
3280 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3281 
3282 	return (NULL);
3283 }
3284 
3285 /*
3286  *	buf_daemon:
3287  *
3288  *	buffer flushing daemon.  Buffers are normally flushed by the
3289  *	update daemon but if it cannot keep up this process starts to
3290  *	take the load in an attempt to prevent getnewbuf() from blocking.
3291  */
3292 static struct kproc_desc buf_kp = {
3293 	"bufdaemon",
3294 	buf_daemon,
3295 	&bufdaemonproc
3296 };
3297 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3298 
3299 static int
3300 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3301 {
3302 	int flushed;
3303 
3304 	flushed = flushbufqueues(vp, bd, target, 0);
3305 	if (flushed == 0) {
3306 		/*
3307 		 * Could not find any buffers without rollback
3308 		 * dependencies, so just write the first one
3309 		 * in the hopes of eventually making progress.
3310 		 */
3311 		if (vp != NULL && target > 2)
3312 			target /= 2;
3313 		flushbufqueues(vp, bd, target, 1);
3314 	}
3315 	return (flushed);
3316 }
3317 
3318 static void
3319 buf_daemon()
3320 {
3321 	struct bufdomain *bd;
3322 	int speedupreq;
3323 	int lodirty;
3324 	int i;
3325 
3326 	/*
3327 	 * This process needs to be suspended prior to shutdown sync.
3328 	 */
3329 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3330 	    SHUTDOWN_PRI_LAST + 100);
3331 
3332 	/*
3333 	 * Start the buf clean daemons as children threads.
3334 	 */
3335 	for (i = 0 ; i < buf_domains; i++) {
3336 		int error;
3337 
3338 		error = kthread_add((void (*)(void *))bufspace_daemon,
3339 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3340 		if (error)
3341 			panic("error %d spawning bufspace daemon", error);
3342 	}
3343 
3344 	/*
3345 	 * This process is allowed to take the buffer cache to the limit
3346 	 */
3347 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3348 	mtx_lock(&bdlock);
3349 	for (;;) {
3350 		bd_request = 0;
3351 		mtx_unlock(&bdlock);
3352 
3353 		kthread_suspend_check();
3354 
3355 		/*
3356 		 * Save speedupreq for this pass and reset to capture new
3357 		 * requests.
3358 		 */
3359 		speedupreq = bd_speedupreq;
3360 		bd_speedupreq = 0;
3361 
3362 		/*
3363 		 * Flush each domain sequentially according to its level and
3364 		 * the speedup request.
3365 		 */
3366 		for (i = 0; i < buf_domains; i++) {
3367 			bd = &bdomain[i];
3368 			if (speedupreq)
3369 				lodirty = bd->bd_numdirtybuffers / 2;
3370 			else
3371 				lodirty = bd->bd_lodirtybuffers;
3372 			while (bd->bd_numdirtybuffers > lodirty) {
3373 				if (buf_flush(NULL, bd,
3374 				    bd->bd_numdirtybuffers - lodirty) == 0)
3375 					break;
3376 				kern_yield(PRI_USER);
3377 			}
3378 		}
3379 
3380 		/*
3381 		 * Only clear bd_request if we have reached our low water
3382 		 * mark.  The buf_daemon normally waits 1 second and
3383 		 * then incrementally flushes any dirty buffers that have
3384 		 * built up, within reason.
3385 		 *
3386 		 * If we were unable to hit our low water mark and couldn't
3387 		 * find any flushable buffers, we sleep for a short period
3388 		 * to avoid endless loops on unlockable buffers.
3389 		 */
3390 		mtx_lock(&bdlock);
3391 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3392 			/*
3393 			 * We reached our low water mark, reset the
3394 			 * request and sleep until we are needed again.
3395 			 * The sleep is just so the suspend code works.
3396 			 */
3397 			bd_request = 0;
3398 			/*
3399 			 * Do an extra wakeup in case dirty threshold
3400 			 * changed via sysctl and the explicit transition
3401 			 * out of shortfall was missed.
3402 			 */
3403 			bdirtywakeup();
3404 			if (runningbufspace <= lorunningspace)
3405 				runningwakeup();
3406 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3407 		} else {
3408 			/*
3409 			 * We couldn't find any flushable dirty buffers but
3410 			 * still have too many dirty buffers, we
3411 			 * have to sleep and try again.  (rare)
3412 			 */
3413 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3414 		}
3415 	}
3416 }
3417 
3418 /*
3419  *	flushbufqueues:
3420  *
3421  *	Try to flush a buffer in the dirty queue.  We must be careful to
3422  *	free up B_INVAL buffers instead of write them, which NFS is
3423  *	particularly sensitive to.
3424  */
3425 static int flushwithdeps = 0;
3426 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
3427     0, "Number of buffers flushed with dependecies that require rollbacks");
3428 
3429 static int
3430 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3431     int flushdeps)
3432 {
3433 	struct bufqueue *bq;
3434 	struct buf *sentinel;
3435 	struct vnode *vp;
3436 	struct mount *mp;
3437 	struct buf *bp;
3438 	int hasdeps;
3439 	int flushed;
3440 	int error;
3441 	bool unlock;
3442 
3443 	flushed = 0;
3444 	bq = &bd->bd_dirtyq;
3445 	bp = NULL;
3446 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3447 	sentinel->b_qindex = QUEUE_SENTINEL;
3448 	BQ_LOCK(bq);
3449 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3450 	BQ_UNLOCK(bq);
3451 	while (flushed != target) {
3452 		maybe_yield();
3453 		BQ_LOCK(bq);
3454 		bp = TAILQ_NEXT(sentinel, b_freelist);
3455 		if (bp != NULL) {
3456 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3457 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3458 			    b_freelist);
3459 		} else {
3460 			BQ_UNLOCK(bq);
3461 			break;
3462 		}
3463 		/*
3464 		 * Skip sentinels inserted by other invocations of the
3465 		 * flushbufqueues(), taking care to not reorder them.
3466 		 *
3467 		 * Only flush the buffers that belong to the
3468 		 * vnode locked by the curthread.
3469 		 */
3470 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3471 		    bp->b_vp != lvp)) {
3472 			BQ_UNLOCK(bq);
3473 			continue;
3474 		}
3475 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3476 		BQ_UNLOCK(bq);
3477 		if (error != 0)
3478 			continue;
3479 
3480 		/*
3481 		 * BKGRDINPROG can only be set with the buf and bufobj
3482 		 * locks both held.  We tolerate a race to clear it here.
3483 		 */
3484 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3485 		    (bp->b_flags & B_DELWRI) == 0) {
3486 			BUF_UNLOCK(bp);
3487 			continue;
3488 		}
3489 		if (bp->b_flags & B_INVAL) {
3490 			bremfreef(bp);
3491 			brelse(bp);
3492 			flushed++;
3493 			continue;
3494 		}
3495 
3496 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3497 			if (flushdeps == 0) {
3498 				BUF_UNLOCK(bp);
3499 				continue;
3500 			}
3501 			hasdeps = 1;
3502 		} else
3503 			hasdeps = 0;
3504 		/*
3505 		 * We must hold the lock on a vnode before writing
3506 		 * one of its buffers. Otherwise we may confuse, or
3507 		 * in the case of a snapshot vnode, deadlock the
3508 		 * system.
3509 		 *
3510 		 * The lock order here is the reverse of the normal
3511 		 * of vnode followed by buf lock.  This is ok because
3512 		 * the NOWAIT will prevent deadlock.
3513 		 */
3514 		vp = bp->b_vp;
3515 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3516 			BUF_UNLOCK(bp);
3517 			continue;
3518 		}
3519 		if (lvp == NULL) {
3520 			unlock = true;
3521 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3522 		} else {
3523 			ASSERT_VOP_LOCKED(vp, "getbuf");
3524 			unlock = false;
3525 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3526 			    vn_lock(vp, LK_TRYUPGRADE);
3527 		}
3528 		if (error == 0) {
3529 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3530 			    bp, bp->b_vp, bp->b_flags);
3531 			if (curproc == bufdaemonproc) {
3532 				vfs_bio_awrite(bp);
3533 			} else {
3534 				bremfree(bp);
3535 				bwrite(bp);
3536 				counter_u64_add(notbufdflushes, 1);
3537 			}
3538 			vn_finished_write(mp);
3539 			if (unlock)
3540 				VOP_UNLOCK(vp, 0);
3541 			flushwithdeps += hasdeps;
3542 			flushed++;
3543 
3544 			/*
3545 			 * Sleeping on runningbufspace while holding
3546 			 * vnode lock leads to deadlock.
3547 			 */
3548 			if (curproc == bufdaemonproc &&
3549 			    runningbufspace > hirunningspace)
3550 				waitrunningbufspace();
3551 			continue;
3552 		}
3553 		vn_finished_write(mp);
3554 		BUF_UNLOCK(bp);
3555 	}
3556 	BQ_LOCK(bq);
3557 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3558 	BQ_UNLOCK(bq);
3559 	free(sentinel, M_TEMP);
3560 	return (flushed);
3561 }
3562 
3563 /*
3564  * Check to see if a block is currently memory resident.
3565  */
3566 struct buf *
3567 incore(struct bufobj *bo, daddr_t blkno)
3568 {
3569 	struct buf *bp;
3570 
3571 	BO_RLOCK(bo);
3572 	bp = gbincore(bo, blkno);
3573 	BO_RUNLOCK(bo);
3574 	return (bp);
3575 }
3576 
3577 /*
3578  * Returns true if no I/O is needed to access the
3579  * associated VM object.  This is like incore except
3580  * it also hunts around in the VM system for the data.
3581  */
3582 
3583 static int
3584 inmem(struct vnode * vp, daddr_t blkno)
3585 {
3586 	vm_object_t obj;
3587 	vm_offset_t toff, tinc, size;
3588 	vm_page_t m;
3589 	vm_ooffset_t off;
3590 
3591 	ASSERT_VOP_LOCKED(vp, "inmem");
3592 
3593 	if (incore(&vp->v_bufobj, blkno))
3594 		return 1;
3595 	if (vp->v_mount == NULL)
3596 		return 0;
3597 	obj = vp->v_object;
3598 	if (obj == NULL)
3599 		return (0);
3600 
3601 	size = PAGE_SIZE;
3602 	if (size > vp->v_mount->mnt_stat.f_iosize)
3603 		size = vp->v_mount->mnt_stat.f_iosize;
3604 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3605 
3606 	VM_OBJECT_RLOCK(obj);
3607 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3608 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3609 		if (!m)
3610 			goto notinmem;
3611 		tinc = size;
3612 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3613 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3614 		if (vm_page_is_valid(m,
3615 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3616 			goto notinmem;
3617 	}
3618 	VM_OBJECT_RUNLOCK(obj);
3619 	return 1;
3620 
3621 notinmem:
3622 	VM_OBJECT_RUNLOCK(obj);
3623 	return (0);
3624 }
3625 
3626 /*
3627  * Set the dirty range for a buffer based on the status of the dirty
3628  * bits in the pages comprising the buffer.  The range is limited
3629  * to the size of the buffer.
3630  *
3631  * Tell the VM system that the pages associated with this buffer
3632  * are clean.  This is used for delayed writes where the data is
3633  * going to go to disk eventually without additional VM intevention.
3634  *
3635  * Note that while we only really need to clean through to b_bcount, we
3636  * just go ahead and clean through to b_bufsize.
3637  */
3638 static void
3639 vfs_clean_pages_dirty_buf(struct buf *bp)
3640 {
3641 	vm_ooffset_t foff, noff, eoff;
3642 	vm_page_t m;
3643 	int i;
3644 
3645 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3646 		return;
3647 
3648 	foff = bp->b_offset;
3649 	KASSERT(bp->b_offset != NOOFFSET,
3650 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3651 
3652 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3653 	vfs_drain_busy_pages(bp);
3654 	vfs_setdirty_locked_object(bp);
3655 	for (i = 0; i < bp->b_npages; i++) {
3656 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3657 		eoff = noff;
3658 		if (eoff > bp->b_offset + bp->b_bufsize)
3659 			eoff = bp->b_offset + bp->b_bufsize;
3660 		m = bp->b_pages[i];
3661 		vfs_page_set_validclean(bp, foff, m);
3662 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3663 		foff = noff;
3664 	}
3665 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3666 }
3667 
3668 static void
3669 vfs_setdirty_locked_object(struct buf *bp)
3670 {
3671 	vm_object_t object;
3672 	int i;
3673 
3674 	object = bp->b_bufobj->bo_object;
3675 	VM_OBJECT_ASSERT_WLOCKED(object);
3676 
3677 	/*
3678 	 * We qualify the scan for modified pages on whether the
3679 	 * object has been flushed yet.
3680 	 */
3681 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3682 		vm_offset_t boffset;
3683 		vm_offset_t eoffset;
3684 
3685 		/*
3686 		 * test the pages to see if they have been modified directly
3687 		 * by users through the VM system.
3688 		 */
3689 		for (i = 0; i < bp->b_npages; i++)
3690 			vm_page_test_dirty(bp->b_pages[i]);
3691 
3692 		/*
3693 		 * Calculate the encompassing dirty range, boffset and eoffset,
3694 		 * (eoffset - boffset) bytes.
3695 		 */
3696 
3697 		for (i = 0; i < bp->b_npages; i++) {
3698 			if (bp->b_pages[i]->dirty)
3699 				break;
3700 		}
3701 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3702 
3703 		for (i = bp->b_npages - 1; i >= 0; --i) {
3704 			if (bp->b_pages[i]->dirty) {
3705 				break;
3706 			}
3707 		}
3708 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3709 
3710 		/*
3711 		 * Fit it to the buffer.
3712 		 */
3713 
3714 		if (eoffset > bp->b_bcount)
3715 			eoffset = bp->b_bcount;
3716 
3717 		/*
3718 		 * If we have a good dirty range, merge with the existing
3719 		 * dirty range.
3720 		 */
3721 
3722 		if (boffset < eoffset) {
3723 			if (bp->b_dirtyoff > boffset)
3724 				bp->b_dirtyoff = boffset;
3725 			if (bp->b_dirtyend < eoffset)
3726 				bp->b_dirtyend = eoffset;
3727 		}
3728 	}
3729 }
3730 
3731 /*
3732  * Allocate the KVA mapping for an existing buffer.
3733  * If an unmapped buffer is provided but a mapped buffer is requested, take
3734  * also care to properly setup mappings between pages and KVA.
3735  */
3736 static void
3737 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3738 {
3739 	int bsize, maxsize, need_mapping, need_kva;
3740 	off_t offset;
3741 
3742 	need_mapping = bp->b_data == unmapped_buf &&
3743 	    (gbflags & GB_UNMAPPED) == 0;
3744 	need_kva = bp->b_kvabase == unmapped_buf &&
3745 	    bp->b_data == unmapped_buf &&
3746 	    (gbflags & GB_KVAALLOC) != 0;
3747 	if (!need_mapping && !need_kva)
3748 		return;
3749 
3750 	BUF_CHECK_UNMAPPED(bp);
3751 
3752 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3753 		/*
3754 		 * Buffer is not mapped, but the KVA was already
3755 		 * reserved at the time of the instantiation.  Use the
3756 		 * allocated space.
3757 		 */
3758 		goto has_addr;
3759 	}
3760 
3761 	/*
3762 	 * Calculate the amount of the address space we would reserve
3763 	 * if the buffer was mapped.
3764 	 */
3765 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3766 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3767 	offset = blkno * bsize;
3768 	maxsize = size + (offset & PAGE_MASK);
3769 	maxsize = imax(maxsize, bsize);
3770 
3771 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3772 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3773 			/*
3774 			 * XXXKIB: defragmentation cannot
3775 			 * succeed, not sure what else to do.
3776 			 */
3777 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3778 		}
3779 		counter_u64_add(mappingrestarts, 1);
3780 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3781 	}
3782 has_addr:
3783 	if (need_mapping) {
3784 		/* b_offset is handled by bpmap_qenter. */
3785 		bp->b_data = bp->b_kvabase;
3786 		BUF_CHECK_MAPPED(bp);
3787 		bpmap_qenter(bp);
3788 	}
3789 }
3790 
3791 struct buf *
3792 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3793     int flags)
3794 {
3795 	struct buf *bp;
3796 	int error;
3797 
3798 	error = getblkx(vp, blkno, size, slpflag, slptimeo, flags, &bp);
3799 	if (error != 0)
3800 		return (NULL);
3801 	return (bp);
3802 }
3803 
3804 /*
3805  *	getblkx:
3806  *
3807  *	Get a block given a specified block and offset into a file/device.
3808  *	The buffers B_DONE bit will be cleared on return, making it almost
3809  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3810  *	return.  The caller should clear B_INVAL prior to initiating a
3811  *	READ.
3812  *
3813  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3814  *	an existing buffer.
3815  *
3816  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3817  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3818  *	and then cleared based on the backing VM.  If the previous buffer is
3819  *	non-0-sized but invalid, B_CACHE will be cleared.
3820  *
3821  *	If getblk() must create a new buffer, the new buffer is returned with
3822  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3823  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3824  *	backing VM.
3825  *
3826  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3827  *	B_CACHE bit is clear.
3828  *
3829  *	What this means, basically, is that the caller should use B_CACHE to
3830  *	determine whether the buffer is fully valid or not and should clear
3831  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3832  *	the buffer by loading its data area with something, the caller needs
3833  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3834  *	the caller should set B_CACHE ( as an optimization ), else the caller
3835  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3836  *	a write attempt or if it was a successful read.  If the caller
3837  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3838  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3839  */
3840 int
3841 getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3842     int flags, struct buf **bpp)
3843 {
3844 	struct buf *bp;
3845 	struct bufobj *bo;
3846 	daddr_t d_blkno;
3847 	int bsize, error, maxsize, vmio;
3848 	off_t offset;
3849 
3850 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3851 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3852 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3853 	ASSERT_VOP_LOCKED(vp, "getblk");
3854 	if (size > maxbcachebuf)
3855 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3856 		    maxbcachebuf);
3857 	if (!unmapped_buf_allowed)
3858 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3859 
3860 	bo = &vp->v_bufobj;
3861 	d_blkno = blkno;
3862 loop:
3863 	BO_RLOCK(bo);
3864 	bp = gbincore(bo, blkno);
3865 	if (bp != NULL) {
3866 		int lockflags;
3867 		/*
3868 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3869 		 * it must be on a queue.
3870 		 */
3871 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3872 
3873 		if ((flags & GB_LOCK_NOWAIT) != 0)
3874 			lockflags |= LK_NOWAIT;
3875 
3876 		error = BUF_TIMELOCK(bp, lockflags,
3877 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3878 
3879 		/*
3880 		 * If we slept and got the lock we have to restart in case
3881 		 * the buffer changed identities.
3882 		 */
3883 		if (error == ENOLCK)
3884 			goto loop;
3885 		/* We timed out or were interrupted. */
3886 		else if (error != 0)
3887 			return (error);
3888 		/* If recursed, assume caller knows the rules. */
3889 		else if (BUF_LOCKRECURSED(bp))
3890 			goto end;
3891 
3892 		/*
3893 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3894 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3895 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3896 		 * backing VM cache.
3897 		 */
3898 		if (bp->b_flags & B_INVAL)
3899 			bp->b_flags &= ~B_CACHE;
3900 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3901 			bp->b_flags |= B_CACHE;
3902 		if (bp->b_flags & B_MANAGED)
3903 			MPASS(bp->b_qindex == QUEUE_NONE);
3904 		else
3905 			bremfree(bp);
3906 
3907 		/*
3908 		 * check for size inconsistencies for non-VMIO case.
3909 		 */
3910 		if (bp->b_bcount != size) {
3911 			if ((bp->b_flags & B_VMIO) == 0 ||
3912 			    (size > bp->b_kvasize)) {
3913 				if (bp->b_flags & B_DELWRI) {
3914 					bp->b_flags |= B_NOCACHE;
3915 					bwrite(bp);
3916 				} else {
3917 					if (LIST_EMPTY(&bp->b_dep)) {
3918 						bp->b_flags |= B_RELBUF;
3919 						brelse(bp);
3920 					} else {
3921 						bp->b_flags |= B_NOCACHE;
3922 						bwrite(bp);
3923 					}
3924 				}
3925 				goto loop;
3926 			}
3927 		}
3928 
3929 		/*
3930 		 * Handle the case of unmapped buffer which should
3931 		 * become mapped, or the buffer for which KVA
3932 		 * reservation is requested.
3933 		 */
3934 		bp_unmapped_get_kva(bp, blkno, size, flags);
3935 
3936 		/*
3937 		 * If the size is inconsistent in the VMIO case, we can resize
3938 		 * the buffer.  This might lead to B_CACHE getting set or
3939 		 * cleared.  If the size has not changed, B_CACHE remains
3940 		 * unchanged from its previous state.
3941 		 */
3942 		allocbuf(bp, size);
3943 
3944 		KASSERT(bp->b_offset != NOOFFSET,
3945 		    ("getblk: no buffer offset"));
3946 
3947 		/*
3948 		 * A buffer with B_DELWRI set and B_CACHE clear must
3949 		 * be committed before we can return the buffer in
3950 		 * order to prevent the caller from issuing a read
3951 		 * ( due to B_CACHE not being set ) and overwriting
3952 		 * it.
3953 		 *
3954 		 * Most callers, including NFS and FFS, need this to
3955 		 * operate properly either because they assume they
3956 		 * can issue a read if B_CACHE is not set, or because
3957 		 * ( for example ) an uncached B_DELWRI might loop due
3958 		 * to softupdates re-dirtying the buffer.  In the latter
3959 		 * case, B_CACHE is set after the first write completes,
3960 		 * preventing further loops.
3961 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3962 		 * above while extending the buffer, we cannot allow the
3963 		 * buffer to remain with B_CACHE set after the write
3964 		 * completes or it will represent a corrupt state.  To
3965 		 * deal with this we set B_NOCACHE to scrap the buffer
3966 		 * after the write.
3967 		 *
3968 		 * We might be able to do something fancy, like setting
3969 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3970 		 * so the below call doesn't set B_CACHE, but that gets real
3971 		 * confusing.  This is much easier.
3972 		 */
3973 
3974 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3975 			bp->b_flags |= B_NOCACHE;
3976 			bwrite(bp);
3977 			goto loop;
3978 		}
3979 		bp->b_flags &= ~B_DONE;
3980 	} else {
3981 		/*
3982 		 * Buffer is not in-core, create new buffer.  The buffer
3983 		 * returned by getnewbuf() is locked.  Note that the returned
3984 		 * buffer is also considered valid (not marked B_INVAL).
3985 		 */
3986 		BO_RUNLOCK(bo);
3987 		/*
3988 		 * If the user does not want us to create the buffer, bail out
3989 		 * here.
3990 		 */
3991 		if (flags & GB_NOCREAT)
3992 			return (EEXIST);
3993 
3994 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3995 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3996 		offset = blkno * bsize;
3997 		vmio = vp->v_object != NULL;
3998 		if (vmio) {
3999 			maxsize = size + (offset & PAGE_MASK);
4000 		} else {
4001 			maxsize = size;
4002 			/* Do not allow non-VMIO notmapped buffers. */
4003 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4004 		}
4005 		maxsize = imax(maxsize, bsize);
4006 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4007 		    !vn_isdisk(vp, NULL)) {
4008 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4009 			KASSERT(error != EOPNOTSUPP,
4010 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4011 			    vp));
4012 			if (error != 0)
4013 				return (error);
4014 			if (d_blkno == -1)
4015 				return (EJUSTRETURN);
4016 		}
4017 
4018 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4019 		if (bp == NULL) {
4020 			if (slpflag || slptimeo)
4021 				return (ETIMEDOUT);
4022 			/*
4023 			 * XXX This is here until the sleep path is diagnosed
4024 			 * enough to work under very low memory conditions.
4025 			 *
4026 			 * There's an issue on low memory, 4BSD+non-preempt
4027 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4028 			 * exhaustion occurs without sleeping for buffer
4029 			 * reclaimation.  This just sticks in a loop and
4030 			 * constantly attempts to allocate a buffer, which
4031 			 * hits exhaustion and tries to wakeup bufdaemon.
4032 			 * This never happens because we never yield.
4033 			 *
4034 			 * The real solution is to identify and fix these cases
4035 			 * so we aren't effectively busy-waiting in a loop
4036 			 * until the reclaimation path has cycles to run.
4037 			 */
4038 			kern_yield(PRI_USER);
4039 			goto loop;
4040 		}
4041 
4042 		/*
4043 		 * This code is used to make sure that a buffer is not
4044 		 * created while the getnewbuf routine is blocked.
4045 		 * This can be a problem whether the vnode is locked or not.
4046 		 * If the buffer is created out from under us, we have to
4047 		 * throw away the one we just created.
4048 		 *
4049 		 * Note: this must occur before we associate the buffer
4050 		 * with the vp especially considering limitations in
4051 		 * the splay tree implementation when dealing with duplicate
4052 		 * lblkno's.
4053 		 */
4054 		BO_LOCK(bo);
4055 		if (gbincore(bo, blkno)) {
4056 			BO_UNLOCK(bo);
4057 			bp->b_flags |= B_INVAL;
4058 			bufspace_release(bufdomain(bp), maxsize);
4059 			brelse(bp);
4060 			goto loop;
4061 		}
4062 
4063 		/*
4064 		 * Insert the buffer into the hash, so that it can
4065 		 * be found by incore.
4066 		 */
4067 		bp->b_lblkno = blkno;
4068 		bp->b_blkno = d_blkno;
4069 		bp->b_offset = offset;
4070 		bgetvp(vp, bp);
4071 		BO_UNLOCK(bo);
4072 
4073 		/*
4074 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4075 		 * buffer size starts out as 0, B_CACHE will be set by
4076 		 * allocbuf() for the VMIO case prior to it testing the
4077 		 * backing store for validity.
4078 		 */
4079 
4080 		if (vmio) {
4081 			bp->b_flags |= B_VMIO;
4082 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4083 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4084 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4085 		} else {
4086 			bp->b_flags &= ~B_VMIO;
4087 			KASSERT(bp->b_bufobj->bo_object == NULL,
4088 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4089 			    bp, bp->b_bufobj->bo_object));
4090 			BUF_CHECK_MAPPED(bp);
4091 		}
4092 
4093 		allocbuf(bp, size);
4094 		bufspace_release(bufdomain(bp), maxsize);
4095 		bp->b_flags &= ~B_DONE;
4096 	}
4097 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4098 end:
4099 	buf_track(bp, __func__);
4100 	KASSERT(bp->b_bufobj == bo,
4101 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4102 	*bpp = bp;
4103 	return (0);
4104 }
4105 
4106 /*
4107  * Get an empty, disassociated buffer of given size.  The buffer is initially
4108  * set to B_INVAL.
4109  */
4110 struct buf *
4111 geteblk(int size, int flags)
4112 {
4113 	struct buf *bp;
4114 	int maxsize;
4115 
4116 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4117 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4118 		if ((flags & GB_NOWAIT_BD) &&
4119 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4120 			return (NULL);
4121 	}
4122 	allocbuf(bp, size);
4123 	bufspace_release(bufdomain(bp), maxsize);
4124 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4125 	return (bp);
4126 }
4127 
4128 /*
4129  * Truncate the backing store for a non-vmio buffer.
4130  */
4131 static void
4132 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4133 {
4134 
4135 	if (bp->b_flags & B_MALLOC) {
4136 		/*
4137 		 * malloced buffers are not shrunk
4138 		 */
4139 		if (newbsize == 0) {
4140 			bufmallocadjust(bp, 0);
4141 			free(bp->b_data, M_BIOBUF);
4142 			bp->b_data = bp->b_kvabase;
4143 			bp->b_flags &= ~B_MALLOC;
4144 		}
4145 		return;
4146 	}
4147 	vm_hold_free_pages(bp, newbsize);
4148 	bufspace_adjust(bp, newbsize);
4149 }
4150 
4151 /*
4152  * Extend the backing for a non-VMIO buffer.
4153  */
4154 static void
4155 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4156 {
4157 	caddr_t origbuf;
4158 	int origbufsize;
4159 
4160 	/*
4161 	 * We only use malloced memory on the first allocation.
4162 	 * and revert to page-allocated memory when the buffer
4163 	 * grows.
4164 	 *
4165 	 * There is a potential smp race here that could lead
4166 	 * to bufmallocspace slightly passing the max.  It
4167 	 * is probably extremely rare and not worth worrying
4168 	 * over.
4169 	 */
4170 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4171 	    bufmallocspace < maxbufmallocspace) {
4172 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4173 		bp->b_flags |= B_MALLOC;
4174 		bufmallocadjust(bp, newbsize);
4175 		return;
4176 	}
4177 
4178 	/*
4179 	 * If the buffer is growing on its other-than-first
4180 	 * allocation then we revert to the page-allocation
4181 	 * scheme.
4182 	 */
4183 	origbuf = NULL;
4184 	origbufsize = 0;
4185 	if (bp->b_flags & B_MALLOC) {
4186 		origbuf = bp->b_data;
4187 		origbufsize = bp->b_bufsize;
4188 		bp->b_data = bp->b_kvabase;
4189 		bufmallocadjust(bp, 0);
4190 		bp->b_flags &= ~B_MALLOC;
4191 		newbsize = round_page(newbsize);
4192 	}
4193 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4194 	    (vm_offset_t) bp->b_data + newbsize);
4195 	if (origbuf != NULL) {
4196 		bcopy(origbuf, bp->b_data, origbufsize);
4197 		free(origbuf, M_BIOBUF);
4198 	}
4199 	bufspace_adjust(bp, newbsize);
4200 }
4201 
4202 /*
4203  * This code constitutes the buffer memory from either anonymous system
4204  * memory (in the case of non-VMIO operations) or from an associated
4205  * VM object (in the case of VMIO operations).  This code is able to
4206  * resize a buffer up or down.
4207  *
4208  * Note that this code is tricky, and has many complications to resolve
4209  * deadlock or inconsistent data situations.  Tread lightly!!!
4210  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4211  * the caller.  Calling this code willy nilly can result in the loss of data.
4212  *
4213  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4214  * B_CACHE for the non-VMIO case.
4215  */
4216 int
4217 allocbuf(struct buf *bp, int size)
4218 {
4219 	int newbsize;
4220 
4221 	if (bp->b_bcount == size)
4222 		return (1);
4223 
4224 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4225 		panic("allocbuf: buffer too small");
4226 
4227 	newbsize = roundup2(size, DEV_BSIZE);
4228 	if ((bp->b_flags & B_VMIO) == 0) {
4229 		if ((bp->b_flags & B_MALLOC) == 0)
4230 			newbsize = round_page(newbsize);
4231 		/*
4232 		 * Just get anonymous memory from the kernel.  Don't
4233 		 * mess with B_CACHE.
4234 		 */
4235 		if (newbsize < bp->b_bufsize)
4236 			vfs_nonvmio_truncate(bp, newbsize);
4237 		else if (newbsize > bp->b_bufsize)
4238 			vfs_nonvmio_extend(bp, newbsize);
4239 	} else {
4240 		int desiredpages;
4241 
4242 		desiredpages = (size == 0) ? 0 :
4243 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4244 
4245 		if (bp->b_flags & B_MALLOC)
4246 			panic("allocbuf: VMIO buffer can't be malloced");
4247 		/*
4248 		 * Set B_CACHE initially if buffer is 0 length or will become
4249 		 * 0-length.
4250 		 */
4251 		if (size == 0 || bp->b_bufsize == 0)
4252 			bp->b_flags |= B_CACHE;
4253 
4254 		if (newbsize < bp->b_bufsize)
4255 			vfs_vmio_truncate(bp, desiredpages);
4256 		/* XXX This looks as if it should be newbsize > b_bufsize */
4257 		else if (size > bp->b_bcount)
4258 			vfs_vmio_extend(bp, desiredpages, size);
4259 		bufspace_adjust(bp, newbsize);
4260 	}
4261 	bp->b_bcount = size;		/* requested buffer size. */
4262 	return (1);
4263 }
4264 
4265 extern int inflight_transient_maps;
4266 
4267 static struct bio_queue nondump_bios;
4268 
4269 void
4270 biodone(struct bio *bp)
4271 {
4272 	struct mtx *mtxp;
4273 	void (*done)(struct bio *);
4274 	vm_offset_t start, end;
4275 
4276 	biotrack(bp, __func__);
4277 
4278 	/*
4279 	 * Avoid completing I/O when dumping after a panic since that may
4280 	 * result in a deadlock in the filesystem or pager code.  Note that
4281 	 * this doesn't affect dumps that were started manually since we aim
4282 	 * to keep the system usable after it has been resumed.
4283 	 */
4284 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4285 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4286 		return;
4287 	}
4288 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4289 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4290 		bp->bio_flags |= BIO_UNMAPPED;
4291 		start = trunc_page((vm_offset_t)bp->bio_data);
4292 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4293 		bp->bio_data = unmapped_buf;
4294 		pmap_qremove(start, atop(end - start));
4295 		vmem_free(transient_arena, start, end - start);
4296 		atomic_add_int(&inflight_transient_maps, -1);
4297 	}
4298 	done = bp->bio_done;
4299 	if (done == NULL) {
4300 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4301 		mtx_lock(mtxp);
4302 		bp->bio_flags |= BIO_DONE;
4303 		wakeup(bp);
4304 		mtx_unlock(mtxp);
4305 	} else
4306 		done(bp);
4307 }
4308 
4309 /*
4310  * Wait for a BIO to finish.
4311  */
4312 int
4313 biowait(struct bio *bp, const char *wchan)
4314 {
4315 	struct mtx *mtxp;
4316 
4317 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4318 	mtx_lock(mtxp);
4319 	while ((bp->bio_flags & BIO_DONE) == 0)
4320 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4321 	mtx_unlock(mtxp);
4322 	if (bp->bio_error != 0)
4323 		return (bp->bio_error);
4324 	if (!(bp->bio_flags & BIO_ERROR))
4325 		return (0);
4326 	return (EIO);
4327 }
4328 
4329 void
4330 biofinish(struct bio *bp, struct devstat *stat, int error)
4331 {
4332 
4333 	if (error) {
4334 		bp->bio_error = error;
4335 		bp->bio_flags |= BIO_ERROR;
4336 	}
4337 	if (stat != NULL)
4338 		devstat_end_transaction_bio(stat, bp);
4339 	biodone(bp);
4340 }
4341 
4342 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4343 void
4344 biotrack_buf(struct bio *bp, const char *location)
4345 {
4346 
4347 	buf_track(bp->bio_track_bp, location);
4348 }
4349 #endif
4350 
4351 /*
4352  *	bufwait:
4353  *
4354  *	Wait for buffer I/O completion, returning error status.  The buffer
4355  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4356  *	error and cleared.
4357  */
4358 int
4359 bufwait(struct buf *bp)
4360 {
4361 	if (bp->b_iocmd == BIO_READ)
4362 		bwait(bp, PRIBIO, "biord");
4363 	else
4364 		bwait(bp, PRIBIO, "biowr");
4365 	if (bp->b_flags & B_EINTR) {
4366 		bp->b_flags &= ~B_EINTR;
4367 		return (EINTR);
4368 	}
4369 	if (bp->b_ioflags & BIO_ERROR) {
4370 		return (bp->b_error ? bp->b_error : EIO);
4371 	} else {
4372 		return (0);
4373 	}
4374 }
4375 
4376 /*
4377  *	bufdone:
4378  *
4379  *	Finish I/O on a buffer, optionally calling a completion function.
4380  *	This is usually called from an interrupt so process blocking is
4381  *	not allowed.
4382  *
4383  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4384  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4385  *	assuming B_INVAL is clear.
4386  *
4387  *	For the VMIO case, we set B_CACHE if the op was a read and no
4388  *	read error occurred, or if the op was a write.  B_CACHE is never
4389  *	set if the buffer is invalid or otherwise uncacheable.
4390  *
4391  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4392  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4393  *	in the biodone routine.
4394  */
4395 void
4396 bufdone(struct buf *bp)
4397 {
4398 	struct bufobj *dropobj;
4399 	void    (*biodone)(struct buf *);
4400 
4401 	buf_track(bp, __func__);
4402 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4403 	dropobj = NULL;
4404 
4405 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4406 
4407 	runningbufwakeup(bp);
4408 	if (bp->b_iocmd == BIO_WRITE)
4409 		dropobj = bp->b_bufobj;
4410 	/* call optional completion function if requested */
4411 	if (bp->b_iodone != NULL) {
4412 		biodone = bp->b_iodone;
4413 		bp->b_iodone = NULL;
4414 		(*biodone) (bp);
4415 		if (dropobj)
4416 			bufobj_wdrop(dropobj);
4417 		return;
4418 	}
4419 	if (bp->b_flags & B_VMIO) {
4420 		/*
4421 		 * Set B_CACHE if the op was a normal read and no error
4422 		 * occurred.  B_CACHE is set for writes in the b*write()
4423 		 * routines.
4424 		 */
4425 		if (bp->b_iocmd == BIO_READ &&
4426 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4427 		    !(bp->b_ioflags & BIO_ERROR))
4428 			bp->b_flags |= B_CACHE;
4429 		vfs_vmio_iodone(bp);
4430 	}
4431 	if (!LIST_EMPTY(&bp->b_dep))
4432 		buf_complete(bp);
4433 	if ((bp->b_flags & B_CKHASH) != 0) {
4434 		KASSERT(bp->b_iocmd == BIO_READ,
4435 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4436 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4437 		(*bp->b_ckhashcalc)(bp);
4438 	}
4439 	/*
4440 	 * For asynchronous completions, release the buffer now. The brelse
4441 	 * will do a wakeup there if necessary - so no need to do a wakeup
4442 	 * here in the async case. The sync case always needs to do a wakeup.
4443 	 */
4444 	if (bp->b_flags & B_ASYNC) {
4445 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4446 		    (bp->b_ioflags & BIO_ERROR))
4447 			brelse(bp);
4448 		else
4449 			bqrelse(bp);
4450 	} else
4451 		bdone(bp);
4452 	if (dropobj)
4453 		bufobj_wdrop(dropobj);
4454 }
4455 
4456 /*
4457  * This routine is called in lieu of iodone in the case of
4458  * incomplete I/O.  This keeps the busy status for pages
4459  * consistent.
4460  */
4461 void
4462 vfs_unbusy_pages(struct buf *bp)
4463 {
4464 	int i;
4465 	vm_object_t obj;
4466 	vm_page_t m;
4467 
4468 	runningbufwakeup(bp);
4469 	if (!(bp->b_flags & B_VMIO))
4470 		return;
4471 
4472 	obj = bp->b_bufobj->bo_object;
4473 	VM_OBJECT_WLOCK(obj);
4474 	for (i = 0; i < bp->b_npages; i++) {
4475 		m = bp->b_pages[i];
4476 		if (m == bogus_page) {
4477 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4478 			if (!m)
4479 				panic("vfs_unbusy_pages: page missing\n");
4480 			bp->b_pages[i] = m;
4481 			if (buf_mapped(bp)) {
4482 				BUF_CHECK_MAPPED(bp);
4483 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4484 				    bp->b_pages, bp->b_npages);
4485 			} else
4486 				BUF_CHECK_UNMAPPED(bp);
4487 		}
4488 		vm_page_sunbusy(m);
4489 	}
4490 	vm_object_pip_wakeupn(obj, bp->b_npages);
4491 	VM_OBJECT_WUNLOCK(obj);
4492 }
4493 
4494 /*
4495  * vfs_page_set_valid:
4496  *
4497  *	Set the valid bits in a page based on the supplied offset.   The
4498  *	range is restricted to the buffer's size.
4499  *
4500  *	This routine is typically called after a read completes.
4501  */
4502 static void
4503 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4504 {
4505 	vm_ooffset_t eoff;
4506 
4507 	/*
4508 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4509 	 * page boundary and eoff is not greater than the end of the buffer.
4510 	 * The end of the buffer, in this case, is our file EOF, not the
4511 	 * allocation size of the buffer.
4512 	 */
4513 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4514 	if (eoff > bp->b_offset + bp->b_bcount)
4515 		eoff = bp->b_offset + bp->b_bcount;
4516 
4517 	/*
4518 	 * Set valid range.  This is typically the entire buffer and thus the
4519 	 * entire page.
4520 	 */
4521 	if (eoff > off)
4522 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4523 }
4524 
4525 /*
4526  * vfs_page_set_validclean:
4527  *
4528  *	Set the valid bits and clear the dirty bits in a page based on the
4529  *	supplied offset.   The range is restricted to the buffer's size.
4530  */
4531 static void
4532 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4533 {
4534 	vm_ooffset_t soff, eoff;
4535 
4536 	/*
4537 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4538 	 * page boundary or cross the end of the buffer.  The end of the
4539 	 * buffer, in this case, is our file EOF, not the allocation size
4540 	 * of the buffer.
4541 	 */
4542 	soff = off;
4543 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4544 	if (eoff > bp->b_offset + bp->b_bcount)
4545 		eoff = bp->b_offset + bp->b_bcount;
4546 
4547 	/*
4548 	 * Set valid range.  This is typically the entire buffer and thus the
4549 	 * entire page.
4550 	 */
4551 	if (eoff > soff) {
4552 		vm_page_set_validclean(
4553 		    m,
4554 		   (vm_offset_t) (soff & PAGE_MASK),
4555 		   (vm_offset_t) (eoff - soff)
4556 		);
4557 	}
4558 }
4559 
4560 /*
4561  * Ensure that all buffer pages are not exclusive busied.  If any page is
4562  * exclusive busy, drain it.
4563  */
4564 void
4565 vfs_drain_busy_pages(struct buf *bp)
4566 {
4567 	vm_page_t m;
4568 	int i, last_busied;
4569 
4570 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4571 	last_busied = 0;
4572 	for (i = 0; i < bp->b_npages; i++) {
4573 		m = bp->b_pages[i];
4574 		if (vm_page_xbusied(m)) {
4575 			for (; last_busied < i; last_busied++)
4576 				vm_page_sbusy(bp->b_pages[last_busied]);
4577 			while (vm_page_xbusied(m)) {
4578 				vm_page_sleep_if_xbusy(m, "vbpage");
4579 			}
4580 		}
4581 	}
4582 	for (i = 0; i < last_busied; i++)
4583 		vm_page_sunbusy(bp->b_pages[i]);
4584 }
4585 
4586 /*
4587  * This routine is called before a device strategy routine.
4588  * It is used to tell the VM system that paging I/O is in
4589  * progress, and treat the pages associated with the buffer
4590  * almost as being exclusive busy.  Also the object paging_in_progress
4591  * flag is handled to make sure that the object doesn't become
4592  * inconsistent.
4593  *
4594  * Since I/O has not been initiated yet, certain buffer flags
4595  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4596  * and should be ignored.
4597  */
4598 void
4599 vfs_busy_pages(struct buf *bp, int clear_modify)
4600 {
4601 	vm_object_t obj;
4602 	vm_ooffset_t foff;
4603 	vm_page_t m;
4604 	int i;
4605 	bool bogus;
4606 
4607 	if (!(bp->b_flags & B_VMIO))
4608 		return;
4609 
4610 	obj = bp->b_bufobj->bo_object;
4611 	foff = bp->b_offset;
4612 	KASSERT(bp->b_offset != NOOFFSET,
4613 	    ("vfs_busy_pages: no buffer offset"));
4614 	VM_OBJECT_WLOCK(obj);
4615 	vfs_drain_busy_pages(bp);
4616 	if (bp->b_bufsize != 0)
4617 		vfs_setdirty_locked_object(bp);
4618 	bogus = false;
4619 	for (i = 0; i < bp->b_npages; i++) {
4620 		m = bp->b_pages[i];
4621 
4622 		if ((bp->b_flags & B_CLUSTER) == 0) {
4623 			vm_object_pip_add(obj, 1);
4624 			vm_page_sbusy(m);
4625 		}
4626 		/*
4627 		 * When readying a buffer for a read ( i.e
4628 		 * clear_modify == 0 ), it is important to do
4629 		 * bogus_page replacement for valid pages in
4630 		 * partially instantiated buffers.  Partially
4631 		 * instantiated buffers can, in turn, occur when
4632 		 * reconstituting a buffer from its VM backing store
4633 		 * base.  We only have to do this if B_CACHE is
4634 		 * clear ( which causes the I/O to occur in the
4635 		 * first place ).  The replacement prevents the read
4636 		 * I/O from overwriting potentially dirty VM-backed
4637 		 * pages.  XXX bogus page replacement is, uh, bogus.
4638 		 * It may not work properly with small-block devices.
4639 		 * We need to find a better way.
4640 		 */
4641 		if (clear_modify) {
4642 			pmap_remove_write(m);
4643 			vfs_page_set_validclean(bp, foff, m);
4644 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4645 		    (bp->b_flags & B_CACHE) == 0) {
4646 			bp->b_pages[i] = bogus_page;
4647 			bogus = true;
4648 		}
4649 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4650 	}
4651 	VM_OBJECT_WUNLOCK(obj);
4652 	if (bogus && buf_mapped(bp)) {
4653 		BUF_CHECK_MAPPED(bp);
4654 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4655 		    bp->b_pages, bp->b_npages);
4656 	}
4657 }
4658 
4659 /*
4660  *	vfs_bio_set_valid:
4661  *
4662  *	Set the range within the buffer to valid.  The range is
4663  *	relative to the beginning of the buffer, b_offset.  Note that
4664  *	b_offset itself may be offset from the beginning of the first
4665  *	page.
4666  */
4667 void
4668 vfs_bio_set_valid(struct buf *bp, int base, int size)
4669 {
4670 	int i, n;
4671 	vm_page_t m;
4672 
4673 	if (!(bp->b_flags & B_VMIO))
4674 		return;
4675 
4676 	/*
4677 	 * Fixup base to be relative to beginning of first page.
4678 	 * Set initial n to be the maximum number of bytes in the
4679 	 * first page that can be validated.
4680 	 */
4681 	base += (bp->b_offset & PAGE_MASK);
4682 	n = PAGE_SIZE - (base & PAGE_MASK);
4683 
4684 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4685 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4686 		m = bp->b_pages[i];
4687 		if (n > size)
4688 			n = size;
4689 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4690 		base += n;
4691 		size -= n;
4692 		n = PAGE_SIZE;
4693 	}
4694 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4695 }
4696 
4697 /*
4698  *	vfs_bio_clrbuf:
4699  *
4700  *	If the specified buffer is a non-VMIO buffer, clear the entire
4701  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4702  *	validate only the previously invalid portions of the buffer.
4703  *	This routine essentially fakes an I/O, so we need to clear
4704  *	BIO_ERROR and B_INVAL.
4705  *
4706  *	Note that while we only theoretically need to clear through b_bcount,
4707  *	we go ahead and clear through b_bufsize.
4708  */
4709 void
4710 vfs_bio_clrbuf(struct buf *bp)
4711 {
4712 	int i, j, mask, sa, ea, slide;
4713 
4714 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4715 		clrbuf(bp);
4716 		return;
4717 	}
4718 	bp->b_flags &= ~B_INVAL;
4719 	bp->b_ioflags &= ~BIO_ERROR;
4720 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4721 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4722 	    (bp->b_offset & PAGE_MASK) == 0) {
4723 		if (bp->b_pages[0] == bogus_page)
4724 			goto unlock;
4725 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4726 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4727 		if ((bp->b_pages[0]->valid & mask) == mask)
4728 			goto unlock;
4729 		if ((bp->b_pages[0]->valid & mask) == 0) {
4730 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4731 			bp->b_pages[0]->valid |= mask;
4732 			goto unlock;
4733 		}
4734 	}
4735 	sa = bp->b_offset & PAGE_MASK;
4736 	slide = 0;
4737 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4738 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4739 		ea = slide & PAGE_MASK;
4740 		if (ea == 0)
4741 			ea = PAGE_SIZE;
4742 		if (bp->b_pages[i] == bogus_page)
4743 			continue;
4744 		j = sa / DEV_BSIZE;
4745 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4746 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4747 		if ((bp->b_pages[i]->valid & mask) == mask)
4748 			continue;
4749 		if ((bp->b_pages[i]->valid & mask) == 0)
4750 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4751 		else {
4752 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4753 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4754 					pmap_zero_page_area(bp->b_pages[i],
4755 					    sa, DEV_BSIZE);
4756 				}
4757 			}
4758 		}
4759 		bp->b_pages[i]->valid |= mask;
4760 	}
4761 unlock:
4762 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4763 	bp->b_resid = 0;
4764 }
4765 
4766 void
4767 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4768 {
4769 	vm_page_t m;
4770 	int i, n;
4771 
4772 	if (buf_mapped(bp)) {
4773 		BUF_CHECK_MAPPED(bp);
4774 		bzero(bp->b_data + base, size);
4775 	} else {
4776 		BUF_CHECK_UNMAPPED(bp);
4777 		n = PAGE_SIZE - (base & PAGE_MASK);
4778 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4779 			m = bp->b_pages[i];
4780 			if (n > size)
4781 				n = size;
4782 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4783 			base += n;
4784 			size -= n;
4785 			n = PAGE_SIZE;
4786 		}
4787 	}
4788 }
4789 
4790 /*
4791  * Update buffer flags based on I/O request parameters, optionally releasing the
4792  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4793  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4794  * I/O).  Otherwise the buffer is released to the cache.
4795  */
4796 static void
4797 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4798 {
4799 
4800 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4801 	    ("buf %p non-VMIO noreuse", bp));
4802 
4803 	if ((ioflag & IO_DIRECT) != 0)
4804 		bp->b_flags |= B_DIRECT;
4805 	if ((ioflag & IO_EXT) != 0)
4806 		bp->b_xflags |= BX_ALTDATA;
4807 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4808 		bp->b_flags |= B_RELBUF;
4809 		if ((ioflag & IO_NOREUSE) != 0)
4810 			bp->b_flags |= B_NOREUSE;
4811 		if (release)
4812 			brelse(bp);
4813 	} else if (release)
4814 		bqrelse(bp);
4815 }
4816 
4817 void
4818 vfs_bio_brelse(struct buf *bp, int ioflag)
4819 {
4820 
4821 	b_io_dismiss(bp, ioflag, true);
4822 }
4823 
4824 void
4825 vfs_bio_set_flags(struct buf *bp, int ioflag)
4826 {
4827 
4828 	b_io_dismiss(bp, ioflag, false);
4829 }
4830 
4831 /*
4832  * vm_hold_load_pages and vm_hold_free_pages get pages into
4833  * a buffers address space.  The pages are anonymous and are
4834  * not associated with a file object.
4835  */
4836 static void
4837 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4838 {
4839 	vm_offset_t pg;
4840 	vm_page_t p;
4841 	int index;
4842 
4843 	BUF_CHECK_MAPPED(bp);
4844 
4845 	to = round_page(to);
4846 	from = round_page(from);
4847 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4848 
4849 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4850 		/*
4851 		 * note: must allocate system pages since blocking here
4852 		 * could interfere with paging I/O, no matter which
4853 		 * process we are.
4854 		 */
4855 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4856 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4857 		    VM_ALLOC_WAITOK);
4858 		pmap_qenter(pg, &p, 1);
4859 		bp->b_pages[index] = p;
4860 	}
4861 	bp->b_npages = index;
4862 }
4863 
4864 /* Return pages associated with this buf to the vm system */
4865 static void
4866 vm_hold_free_pages(struct buf *bp, int newbsize)
4867 {
4868 	vm_offset_t from;
4869 	vm_page_t p;
4870 	int index, newnpages;
4871 
4872 	BUF_CHECK_MAPPED(bp);
4873 
4874 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4875 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4876 	if (bp->b_npages > newnpages)
4877 		pmap_qremove(from, bp->b_npages - newnpages);
4878 	for (index = newnpages; index < bp->b_npages; index++) {
4879 		p = bp->b_pages[index];
4880 		bp->b_pages[index] = NULL;
4881 		vm_page_unwire_noq(p);
4882 		vm_page_free(p);
4883 	}
4884 	bp->b_npages = newnpages;
4885 }
4886 
4887 /*
4888  * Map an IO request into kernel virtual address space.
4889  *
4890  * All requests are (re)mapped into kernel VA space.
4891  * Notice that we use b_bufsize for the size of the buffer
4892  * to be mapped.  b_bcount might be modified by the driver.
4893  *
4894  * Note that even if the caller determines that the address space should
4895  * be valid, a race or a smaller-file mapped into a larger space may
4896  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4897  * check the return value.
4898  *
4899  * This function only works with pager buffers.
4900  */
4901 int
4902 vmapbuf(struct buf *bp, int mapbuf)
4903 {
4904 	vm_prot_t prot;
4905 	int pidx;
4906 
4907 	if (bp->b_bufsize < 0)
4908 		return (-1);
4909 	prot = VM_PROT_READ;
4910 	if (bp->b_iocmd == BIO_READ)
4911 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4912 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4913 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4914 	    btoc(MAXPHYS))) < 0)
4915 		return (-1);
4916 	bp->b_npages = pidx;
4917 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4918 	if (mapbuf || !unmapped_buf_allowed) {
4919 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4920 		bp->b_data = bp->b_kvabase + bp->b_offset;
4921 	} else
4922 		bp->b_data = unmapped_buf;
4923 	return(0);
4924 }
4925 
4926 /*
4927  * Free the io map PTEs associated with this IO operation.
4928  * We also invalidate the TLB entries and restore the original b_addr.
4929  *
4930  * This function only works with pager buffers.
4931  */
4932 void
4933 vunmapbuf(struct buf *bp)
4934 {
4935 	int npages;
4936 
4937 	npages = bp->b_npages;
4938 	if (buf_mapped(bp))
4939 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4940 	vm_page_unhold_pages(bp->b_pages, npages);
4941 
4942 	bp->b_data = unmapped_buf;
4943 }
4944 
4945 void
4946 bdone(struct buf *bp)
4947 {
4948 	struct mtx *mtxp;
4949 
4950 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4951 	mtx_lock(mtxp);
4952 	bp->b_flags |= B_DONE;
4953 	wakeup(bp);
4954 	mtx_unlock(mtxp);
4955 }
4956 
4957 void
4958 bwait(struct buf *bp, u_char pri, const char *wchan)
4959 {
4960 	struct mtx *mtxp;
4961 
4962 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4963 	mtx_lock(mtxp);
4964 	while ((bp->b_flags & B_DONE) == 0)
4965 		msleep(bp, mtxp, pri, wchan, 0);
4966 	mtx_unlock(mtxp);
4967 }
4968 
4969 int
4970 bufsync(struct bufobj *bo, int waitfor)
4971 {
4972 
4973 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
4974 }
4975 
4976 void
4977 bufstrategy(struct bufobj *bo, struct buf *bp)
4978 {
4979 	int i __unused;
4980 	struct vnode *vp;
4981 
4982 	vp = bp->b_vp;
4983 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4984 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4985 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4986 	i = VOP_STRATEGY(vp, bp);
4987 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4988 }
4989 
4990 /*
4991  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
4992  */
4993 void
4994 bufobj_init(struct bufobj *bo, void *private)
4995 {
4996 	static volatile int bufobj_cleanq;
4997 
4998         bo->bo_domain =
4999             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5000         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5001         bo->bo_private = private;
5002         TAILQ_INIT(&bo->bo_clean.bv_hd);
5003         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5004 }
5005 
5006 void
5007 bufobj_wrefl(struct bufobj *bo)
5008 {
5009 
5010 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5011 	ASSERT_BO_WLOCKED(bo);
5012 	bo->bo_numoutput++;
5013 }
5014 
5015 void
5016 bufobj_wref(struct bufobj *bo)
5017 {
5018 
5019 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5020 	BO_LOCK(bo);
5021 	bo->bo_numoutput++;
5022 	BO_UNLOCK(bo);
5023 }
5024 
5025 void
5026 bufobj_wdrop(struct bufobj *bo)
5027 {
5028 
5029 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5030 	BO_LOCK(bo);
5031 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5032 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5033 		bo->bo_flag &= ~BO_WWAIT;
5034 		wakeup(&bo->bo_numoutput);
5035 	}
5036 	BO_UNLOCK(bo);
5037 }
5038 
5039 int
5040 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5041 {
5042 	int error;
5043 
5044 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5045 	ASSERT_BO_WLOCKED(bo);
5046 	error = 0;
5047 	while (bo->bo_numoutput) {
5048 		bo->bo_flag |= BO_WWAIT;
5049 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5050 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5051 		if (error)
5052 			break;
5053 	}
5054 	return (error);
5055 }
5056 
5057 /*
5058  * Set bio_data or bio_ma for struct bio from the struct buf.
5059  */
5060 void
5061 bdata2bio(struct buf *bp, struct bio *bip)
5062 {
5063 
5064 	if (!buf_mapped(bp)) {
5065 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5066 		bip->bio_ma = bp->b_pages;
5067 		bip->bio_ma_n = bp->b_npages;
5068 		bip->bio_data = unmapped_buf;
5069 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5070 		bip->bio_flags |= BIO_UNMAPPED;
5071 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5072 		    PAGE_SIZE == bp->b_npages,
5073 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5074 		    (long long)bip->bio_length, bip->bio_ma_n));
5075 	} else {
5076 		bip->bio_data = bp->b_data;
5077 		bip->bio_ma = NULL;
5078 	}
5079 }
5080 
5081 /*
5082  * The MIPS pmap code currently doesn't handle aliased pages.
5083  * The VIPT caches may not handle page aliasing themselves, leading
5084  * to data corruption.
5085  *
5086  * As such, this code makes a system extremely unhappy if said
5087  * system doesn't support unaliasing the above situation in hardware.
5088  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5089  * this feature at build time, so it has to be handled in software.
5090  *
5091  * Once the MIPS pmap/cache code grows to support this function on
5092  * earlier chips, it should be flipped back off.
5093  */
5094 #ifdef	__mips__
5095 static int buf_pager_relbuf = 1;
5096 #else
5097 static int buf_pager_relbuf = 0;
5098 #endif
5099 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5100     &buf_pager_relbuf, 0,
5101     "Make buffer pager release buffers after reading");
5102 
5103 /*
5104  * The buffer pager.  It uses buffer reads to validate pages.
5105  *
5106  * In contrast to the generic local pager from vm/vnode_pager.c, this
5107  * pager correctly and easily handles volumes where the underlying
5108  * device block size is greater than the machine page size.  The
5109  * buffer cache transparently extends the requested page run to be
5110  * aligned at the block boundary, and does the necessary bogus page
5111  * replacements in the addends to avoid obliterating already valid
5112  * pages.
5113  *
5114  * The only non-trivial issue is that the exclusive busy state for
5115  * pages, which is assumed by the vm_pager_getpages() interface, is
5116  * incompatible with the VMIO buffer cache's desire to share-busy the
5117  * pages.  This function performs a trivial downgrade of the pages'
5118  * state before reading buffers, and a less trivial upgrade from the
5119  * shared-busy to excl-busy state after the read.
5120  */
5121 int
5122 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5123     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5124     vbg_get_blksize_t get_blksize)
5125 {
5126 	vm_page_t m;
5127 	vm_object_t object;
5128 	struct buf *bp;
5129 	struct mount *mp;
5130 	daddr_t lbn, lbnp;
5131 	vm_ooffset_t la, lb, poff, poffe;
5132 	long bsize;
5133 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5134 	bool redo, lpart;
5135 
5136 	object = vp->v_object;
5137 	mp = vp->v_mount;
5138 	error = 0;
5139 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5140 	if (la >= object->un_pager.vnp.vnp_size)
5141 		return (VM_PAGER_BAD);
5142 
5143 	/*
5144 	 * Change the meaning of la from where the last requested page starts
5145 	 * to where it ends, because that's the end of the requested region
5146 	 * and the start of the potential read-ahead region.
5147 	 */
5148 	la += PAGE_SIZE;
5149 	lpart = la > object->un_pager.vnp.vnp_size;
5150 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5151 
5152 	/*
5153 	 * Calculate read-ahead, behind and total pages.
5154 	 */
5155 	pgsin = count;
5156 	lb = IDX_TO_OFF(ma[0]->pindex);
5157 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5158 	pgsin += pgsin_b;
5159 	if (rbehind != NULL)
5160 		*rbehind = pgsin_b;
5161 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5162 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5163 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5164 		    PAGE_SIZE) - la);
5165 	pgsin += pgsin_a;
5166 	if (rahead != NULL)
5167 		*rahead = pgsin_a;
5168 	VM_CNT_INC(v_vnodein);
5169 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5170 
5171 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5172 	    != 0) ? GB_UNMAPPED : 0;
5173 	VM_OBJECT_WLOCK(object);
5174 again:
5175 	for (i = 0; i < count; i++)
5176 		vm_page_busy_downgrade(ma[i]);
5177 	VM_OBJECT_WUNLOCK(object);
5178 
5179 	lbnp = -1;
5180 	for (i = 0; i < count; i++) {
5181 		m = ma[i];
5182 
5183 		/*
5184 		 * Pages are shared busy and the object lock is not
5185 		 * owned, which together allow for the pages'
5186 		 * invalidation.  The racy test for validity avoids
5187 		 * useless creation of the buffer for the most typical
5188 		 * case when invalidation is not used in redo or for
5189 		 * parallel read.  The shared->excl upgrade loop at
5190 		 * the end of the function catches the race in a
5191 		 * reliable way (protected by the object lock).
5192 		 */
5193 		if (m->valid == VM_PAGE_BITS_ALL)
5194 			continue;
5195 
5196 		poff = IDX_TO_OFF(m->pindex);
5197 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5198 		for (; poff < poffe; poff += bsize) {
5199 			lbn = get_lblkno(vp, poff);
5200 			if (lbn == lbnp)
5201 				goto next_page;
5202 			lbnp = lbn;
5203 
5204 			bsize = get_blksize(vp, lbn);
5205 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5206 			    br_flags, &bp);
5207 			if (error != 0)
5208 				goto end_pages;
5209 			if (LIST_EMPTY(&bp->b_dep)) {
5210 				/*
5211 				 * Invalidation clears m->valid, but
5212 				 * may leave B_CACHE flag if the
5213 				 * buffer existed at the invalidation
5214 				 * time.  In this case, recycle the
5215 				 * buffer to do real read on next
5216 				 * bread() after redo.
5217 				 *
5218 				 * Otherwise B_RELBUF is not strictly
5219 				 * necessary, enable to reduce buf
5220 				 * cache pressure.
5221 				 */
5222 				if (buf_pager_relbuf ||
5223 				    m->valid != VM_PAGE_BITS_ALL)
5224 					bp->b_flags |= B_RELBUF;
5225 
5226 				bp->b_flags &= ~B_NOCACHE;
5227 				brelse(bp);
5228 			} else {
5229 				bqrelse(bp);
5230 			}
5231 		}
5232 		KASSERT(1 /* racy, enable for debugging */ ||
5233 		    m->valid == VM_PAGE_BITS_ALL || i == count - 1,
5234 		    ("buf %d %p invalid", i, m));
5235 		if (i == count - 1 && lpart) {
5236 			VM_OBJECT_WLOCK(object);
5237 			if (m->valid != 0 &&
5238 			    m->valid != VM_PAGE_BITS_ALL)
5239 				vm_page_zero_invalid(m, TRUE);
5240 			VM_OBJECT_WUNLOCK(object);
5241 		}
5242 next_page:;
5243 	}
5244 end_pages:
5245 
5246 	VM_OBJECT_WLOCK(object);
5247 	redo = false;
5248 	for (i = 0; i < count; i++) {
5249 		vm_page_sunbusy(ma[i]);
5250 		ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
5251 
5252 		/*
5253 		 * Since the pages were only sbusy while neither the
5254 		 * buffer nor the object lock was held by us, or
5255 		 * reallocated while vm_page_grab() slept for busy
5256 		 * relinguish, they could have been invalidated.
5257 		 * Recheck the valid bits and re-read as needed.
5258 		 *
5259 		 * Note that the last page is made fully valid in the
5260 		 * read loop, and partial validity for the page at
5261 		 * index count - 1 could mean that the page was
5262 		 * invalidated or removed, so we must restart for
5263 		 * safety as well.
5264 		 */
5265 		if (ma[i]->valid != VM_PAGE_BITS_ALL)
5266 			redo = true;
5267 	}
5268 	if (redo && error == 0)
5269 		goto again;
5270 	VM_OBJECT_WUNLOCK(object);
5271 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5272 }
5273 
5274 #include "opt_ddb.h"
5275 #ifdef DDB
5276 #include <ddb/ddb.h>
5277 
5278 /* DDB command to show buffer data */
5279 DB_SHOW_COMMAND(buffer, db_show_buffer)
5280 {
5281 	/* get args */
5282 	struct buf *bp = (struct buf *)addr;
5283 #ifdef FULL_BUF_TRACKING
5284 	uint32_t i, j;
5285 #endif
5286 
5287 	if (!have_addr) {
5288 		db_printf("usage: show buffer <addr>\n");
5289 		return;
5290 	}
5291 
5292 	db_printf("buf at %p\n", bp);
5293 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5294 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5295 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5296 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5297 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5298 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5299 	db_printf(
5300 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5301 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5302 	    "b_vp = %p, b_dep = %p\n",
5303 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5304 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5305 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5306 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5307 	    bp->b_kvabase, bp->b_kvasize);
5308 	if (bp->b_npages) {
5309 		int i;
5310 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5311 		for (i = 0; i < bp->b_npages; i++) {
5312 			vm_page_t m;
5313 			m = bp->b_pages[i];
5314 			if (m != NULL)
5315 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5316 				    (u_long)m->pindex,
5317 				    (u_long)VM_PAGE_TO_PHYS(m));
5318 			else
5319 				db_printf("( ??? )");
5320 			if ((i + 1) < bp->b_npages)
5321 				db_printf(",");
5322 		}
5323 		db_printf("\n");
5324 	}
5325 	BUF_LOCKPRINTINFO(bp);
5326 #if defined(FULL_BUF_TRACKING)
5327 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5328 
5329 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5330 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5331 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5332 			continue;
5333 		db_printf(" %2u: %s\n", j,
5334 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5335 	}
5336 #elif defined(BUF_TRACKING)
5337 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5338 #endif
5339 	db_printf(" ");
5340 }
5341 
5342 DB_SHOW_COMMAND(bufqueues, bufqueues)
5343 {
5344 	struct bufdomain *bd;
5345 	struct buf *bp;
5346 	long total;
5347 	int i, j, cnt;
5348 
5349 	db_printf("bqempty: %d\n", bqempty.bq_len);
5350 
5351 	for (i = 0; i < buf_domains; i++) {
5352 		bd = &bdomain[i];
5353 		db_printf("Buf domain %d\n", i);
5354 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5355 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5356 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5357 		db_printf("\n");
5358 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5359 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5360 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5361 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5362 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5363 		db_printf("\n");
5364 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5365 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5366 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5367 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5368 		db_printf("\n");
5369 		total = 0;
5370 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5371 			total += bp->b_bufsize;
5372 		db_printf("\tcleanq count\t%d (%ld)\n",
5373 		    bd->bd_cleanq->bq_len, total);
5374 		total = 0;
5375 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5376 			total += bp->b_bufsize;
5377 		db_printf("\tdirtyq count\t%d (%ld)\n",
5378 		    bd->bd_dirtyq.bq_len, total);
5379 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5380 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5381 		db_printf("\tCPU ");
5382 		for (j = 0; j <= mp_maxid; j++)
5383 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5384 		db_printf("\n");
5385 		cnt = 0;
5386 		total = 0;
5387 		for (j = 0; j < nbuf; j++)
5388 			if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5389 				cnt++;
5390 				total += buf[j].b_bufsize;
5391 			}
5392 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5393 		cnt = 0;
5394 		total = 0;
5395 		for (j = 0; j < nbuf; j++)
5396 			if (buf[j].b_domain == i) {
5397 				cnt++;
5398 				total += buf[j].b_bufsize;
5399 			}
5400 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5401 	}
5402 }
5403 
5404 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5405 {
5406 	struct buf *bp;
5407 	int i;
5408 
5409 	for (i = 0; i < nbuf; i++) {
5410 		bp = &buf[i];
5411 		if (BUF_ISLOCKED(bp)) {
5412 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5413 			db_printf("\n");
5414 			if (db_pager_quit)
5415 				break;
5416 		}
5417 	}
5418 }
5419 
5420 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5421 {
5422 	struct vnode *vp;
5423 	struct buf *bp;
5424 
5425 	if (!have_addr) {
5426 		db_printf("usage: show vnodebufs <addr>\n");
5427 		return;
5428 	}
5429 	vp = (struct vnode *)addr;
5430 	db_printf("Clean buffers:\n");
5431 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5432 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5433 		db_printf("\n");
5434 	}
5435 	db_printf("Dirty buffers:\n");
5436 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5437 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5438 		db_printf("\n");
5439 	}
5440 }
5441 
5442 DB_COMMAND(countfreebufs, db_coundfreebufs)
5443 {
5444 	struct buf *bp;
5445 	int i, used = 0, nfree = 0;
5446 
5447 	if (have_addr) {
5448 		db_printf("usage: countfreebufs\n");
5449 		return;
5450 	}
5451 
5452 	for (i = 0; i < nbuf; i++) {
5453 		bp = &buf[i];
5454 		if (bp->b_qindex == QUEUE_EMPTY)
5455 			nfree++;
5456 		else
5457 			used++;
5458 	}
5459 
5460 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5461 	    nfree + used);
5462 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5463 }
5464 #endif /* DDB */
5465