xref: /freebsd/sys/kern/vfs_bio.c (revision b9f654b163bce26de79705e77b872427c9f2afa1)
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/resourcevar.h>
71 #include <sys/rwlock.h>
72 #include <sys/smp.h>
73 #include <sys/sysctl.h>
74 #include <sys/sysproto.h>
75 #include <sys/vmem.h>
76 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
78 #include <sys/watchdog.h>
79 #include <geom/geom.h>
80 #include <vm/vm.h>
81 #include <vm/vm_param.h>
82 #include <vm/vm_kern.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/vm_extern.h>
88 #include <vm/vm_map.h>
89 #include <vm/swap_pager.h>
90 
91 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
92 
93 struct	bio_ops bioops;		/* I/O operation notification */
94 
95 struct	buf_ops buf_ops_bio = {
96 	.bop_name	=	"buf_ops_bio",
97 	.bop_write	=	bufwrite,
98 	.bop_strategy	=	bufstrategy,
99 	.bop_sync	=	bufsync,
100 	.bop_bdflush	=	bufbdflush,
101 };
102 
103 struct bufqueue {
104 	struct mtx_padalign	bq_lock;
105 	TAILQ_HEAD(, buf)	bq_queue;
106 	uint8_t			bq_index;
107 	uint16_t		bq_subqueue;
108 	int			bq_len;
109 } __aligned(CACHE_LINE_SIZE);
110 
111 #define	BQ_LOCKPTR(bq)		(&(bq)->bq_lock)
112 #define	BQ_LOCK(bq)		mtx_lock(BQ_LOCKPTR((bq)))
113 #define	BQ_UNLOCK(bq)		mtx_unlock(BQ_LOCKPTR((bq)))
114 #define	BQ_ASSERT_LOCKED(bq)	mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
115 
116 struct bufdomain {
117 	struct bufqueue	bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
118 	struct bufqueue bd_dirtyq;
119 	struct bufqueue	*bd_cleanq;
120 	struct mtx_padalign bd_run_lock;
121 	/* Constants */
122 	long		bd_maxbufspace;
123 	long		bd_hibufspace;
124 	long 		bd_lobufspace;
125 	long 		bd_bufspacethresh;
126 	int		bd_hifreebuffers;
127 	int		bd_lofreebuffers;
128 	int		bd_hidirtybuffers;
129 	int		bd_lodirtybuffers;
130 	int		bd_dirtybufthresh;
131 	int		bd_lim;
132 	/* atomics */
133 	int		bd_wanted;
134 	int __aligned(CACHE_LINE_SIZE)	bd_numdirtybuffers;
135 	int __aligned(CACHE_LINE_SIZE)	bd_running;
136 	long __aligned(CACHE_LINE_SIZE) bd_bufspace;
137 	int __aligned(CACHE_LINE_SIZE)	bd_freebuffers;
138 } __aligned(CACHE_LINE_SIZE);
139 
140 #define	BD_LOCKPTR(bd)		(&(bd)->bd_cleanq->bq_lock)
141 #define	BD_LOCK(bd)		mtx_lock(BD_LOCKPTR((bd)))
142 #define	BD_UNLOCK(bd)		mtx_unlock(BD_LOCKPTR((bd)))
143 #define	BD_ASSERT_LOCKED(bd)	mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
144 #define	BD_RUN_LOCKPTR(bd)	(&(bd)->bd_run_lock)
145 #define	BD_RUN_LOCK(bd)		mtx_lock(BD_RUN_LOCKPTR((bd)))
146 #define	BD_RUN_UNLOCK(bd)	mtx_unlock(BD_RUN_LOCKPTR((bd)))
147 #define	BD_DOMAIN(bd)		(bd - bdomain)
148 
149 static struct buf *buf;		/* buffer header pool */
150 extern struct buf *swbuf;	/* Swap buffer header pool. */
151 caddr_t unmapped_buf;
152 
153 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
154 struct proc *bufdaemonproc;
155 
156 static int inmem(struct vnode *vp, daddr_t blkno);
157 static void vm_hold_free_pages(struct buf *bp, int newbsize);
158 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
159 		vm_offset_t to);
160 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
161 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
162 		vm_page_t m);
163 static void vfs_clean_pages_dirty_buf(struct buf *bp);
164 static void vfs_setdirty_locked_object(struct buf *bp);
165 static void vfs_vmio_invalidate(struct buf *bp);
166 static void vfs_vmio_truncate(struct buf *bp, int npages);
167 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
168 static int vfs_bio_clcheck(struct vnode *vp, int size,
169 		daddr_t lblkno, daddr_t blkno);
170 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
171 		void (*)(struct buf *));
172 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
173 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
174 static void buf_daemon(void);
175 static __inline void bd_wakeup(void);
176 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
177 static void bufkva_reclaim(vmem_t *, int);
178 static void bufkva_free(struct buf *);
179 static int buf_import(void *, void **, int, int, int);
180 static void buf_release(void *, void **, int);
181 static void maxbcachebuf_adjust(void);
182 static inline struct bufdomain *bufdomain(struct buf *);
183 static void bq_remove(struct bufqueue *bq, struct buf *bp);
184 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
185 static int buf_recycle(struct bufdomain *, bool kva);
186 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
187 	    const char *lockname);
188 static void bd_init(struct bufdomain *bd);
189 static int bd_flushall(struct bufdomain *bd);
190 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
191 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
192 
193 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
194 int vmiodirenable = TRUE;
195 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
196     "Use the VM system for directory writes");
197 long runningbufspace;
198 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
199     "Amount of presently outstanding async buffer io");
200 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
201     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
202 static counter_u64_t bufkvaspace;
203 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
204     "Kernel virtual memory used for buffers");
205 static long maxbufspace;
206 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
207     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
208     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
209     "Maximum allowed value of bufspace (including metadata)");
210 static long bufmallocspace;
211 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
212     "Amount of malloced memory for buffers");
213 static long maxbufmallocspace;
214 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
215     0, "Maximum amount of malloced memory for buffers");
216 static long lobufspace;
217 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
218     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
219     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
220     "Minimum amount of buffers we want to have");
221 long hibufspace;
222 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
223     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
224     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
225     "Maximum allowed value of bufspace (excluding metadata)");
226 long bufspacethresh;
227 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
228     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
229     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
230     "Bufspace consumed before waking the daemon to free some");
231 static counter_u64_t buffreekvacnt;
232 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
233     "Number of times we have freed the KVA space from some buffer");
234 static counter_u64_t bufdefragcnt;
235 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
236     "Number of times we have had to repeat buffer allocation to defragment");
237 static long lorunningspace;
238 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
239     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
240     "Minimum preferred space used for in-progress I/O");
241 static long hirunningspace;
242 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
243     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
244     "Maximum amount of space to use for in-progress I/O");
245 int dirtybufferflushes;
246 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
247     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
248 int bdwriteskip;
249 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
250     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
251 int altbufferflushes;
252 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
253     0, "Number of fsync flushes to limit dirty buffers");
254 static int recursiveflushes;
255 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
256     0, "Number of flushes skipped due to being recursive");
257 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
258 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
259     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
260     "Number of buffers that are dirty (has unwritten changes) at the moment");
261 static int lodirtybuffers;
262 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
263     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
264     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
265     "How many buffers we want to have free before bufdaemon can sleep");
266 static int hidirtybuffers;
267 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
268     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
269     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
270     "When the number of dirty buffers is considered severe");
271 int dirtybufthresh;
272 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
273     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
274     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
275     "Number of bdwrite to bawrite conversions to clear dirty buffers");
276 static int numfreebuffers;
277 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
278     "Number of free buffers");
279 static int lofreebuffers;
280 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
281     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
282     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
283    "Target number of free buffers");
284 static int hifreebuffers;
285 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
286     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
287     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
288    "Threshold for clean buffer recycling");
289 static counter_u64_t getnewbufcalls;
290 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
291    &getnewbufcalls, "Number of calls to getnewbuf");
292 static counter_u64_t getnewbufrestarts;
293 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
294     &getnewbufrestarts,
295     "Number of times getnewbuf has had to restart a buffer acquisition");
296 static counter_u64_t mappingrestarts;
297 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
298     &mappingrestarts,
299     "Number of times getblk has had to restart a buffer mapping for "
300     "unmapped buffer");
301 static counter_u64_t numbufallocfails;
302 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
303     &numbufallocfails, "Number of times buffer allocations failed");
304 static int flushbufqtarget = 100;
305 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
306     "Amount of work to do in flushbufqueues when helping bufdaemon");
307 static counter_u64_t notbufdflushes;
308 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
309     "Number of dirty buffer flushes done by the bufdaemon helpers");
310 static long barrierwrites;
311 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
312     "Number of barrier writes");
313 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
314     &unmapped_buf_allowed, 0,
315     "Permit the use of the unmapped i/o");
316 int maxbcachebuf = MAXBCACHEBUF;
317 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
318     "Maximum size of a buffer cache block");
319 
320 /*
321  * This lock synchronizes access to bd_request.
322  */
323 static struct mtx_padalign __exclusive_cache_line bdlock;
324 
325 /*
326  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
327  * waitrunningbufspace().
328  */
329 static struct mtx_padalign __exclusive_cache_line rbreqlock;
330 
331 /*
332  * Lock that protects bdirtywait.
333  */
334 static struct mtx_padalign __exclusive_cache_line bdirtylock;
335 
336 /*
337  * Wakeup point for bufdaemon, as well as indicator of whether it is already
338  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
339  * is idling.
340  */
341 static int bd_request;
342 
343 /*
344  * Request for the buf daemon to write more buffers than is indicated by
345  * lodirtybuf.  This may be necessary to push out excess dependencies or
346  * defragment the address space where a simple count of the number of dirty
347  * buffers is insufficient to characterize the demand for flushing them.
348  */
349 static int bd_speedupreq;
350 
351 /*
352  * Synchronization (sleep/wakeup) variable for active buffer space requests.
353  * Set when wait starts, cleared prior to wakeup().
354  * Used in runningbufwakeup() and waitrunningbufspace().
355  */
356 static int runningbufreq;
357 
358 /*
359  * Synchronization for bwillwrite() waiters.
360  */
361 static int bdirtywait;
362 
363 /*
364  * Definitions for the buffer free lists.
365  */
366 #define QUEUE_NONE	0	/* on no queue */
367 #define QUEUE_EMPTY	1	/* empty buffer headers */
368 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
369 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
370 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
371 
372 /* Maximum number of buffer domains. */
373 #define	BUF_DOMAINS	8
374 
375 struct bufdomainset bdlodirty;		/* Domains > lodirty */
376 struct bufdomainset bdhidirty;		/* Domains > hidirty */
377 
378 /* Configured number of clean queues. */
379 static int __read_mostly buf_domains;
380 
381 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
382 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
383 struct bufqueue __exclusive_cache_line bqempty;
384 
385 /*
386  * per-cpu empty buffer cache.
387  */
388 uma_zone_t buf_zone;
389 
390 /*
391  * Single global constant for BUF_WMESG, to avoid getting multiple references.
392  * buf_wmesg is referred from macros.
393  */
394 const char *buf_wmesg = BUF_WMESG;
395 
396 static int
397 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
398 {
399 	long value;
400 	int error;
401 
402 	value = *(long *)arg1;
403 	error = sysctl_handle_long(oidp, &value, 0, req);
404 	if (error != 0 || req->newptr == NULL)
405 		return (error);
406 	mtx_lock(&rbreqlock);
407 	if (arg1 == &hirunningspace) {
408 		if (value < lorunningspace)
409 			error = EINVAL;
410 		else
411 			hirunningspace = value;
412 	} else {
413 		KASSERT(arg1 == &lorunningspace,
414 		    ("%s: unknown arg1", __func__));
415 		if (value > hirunningspace)
416 			error = EINVAL;
417 		else
418 			lorunningspace = value;
419 	}
420 	mtx_unlock(&rbreqlock);
421 	return (error);
422 }
423 
424 static int
425 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
426 {
427 	int error;
428 	int value;
429 	int i;
430 
431 	value = *(int *)arg1;
432 	error = sysctl_handle_int(oidp, &value, 0, req);
433 	if (error != 0 || req->newptr == NULL)
434 		return (error);
435 	*(int *)arg1 = value;
436 	for (i = 0; i < buf_domains; i++)
437 		*(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
438 		    value / buf_domains;
439 
440 	return (error);
441 }
442 
443 static int
444 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
445 {
446 	long value;
447 	int error;
448 	int i;
449 
450 	value = *(long *)arg1;
451 	error = sysctl_handle_long(oidp, &value, 0, req);
452 	if (error != 0 || req->newptr == NULL)
453 		return (error);
454 	*(long *)arg1 = value;
455 	for (i = 0; i < buf_domains; i++)
456 		*(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
457 		    value / buf_domains;
458 
459 	return (error);
460 }
461 
462 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
463     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
464 static int
465 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
466 {
467 	long lvalue;
468 	int ivalue;
469 	int i;
470 
471 	lvalue = 0;
472 	for (i = 0; i < buf_domains; i++)
473 		lvalue += bdomain[i].bd_bufspace;
474 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
475 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
476 	if (lvalue > INT_MAX)
477 		/* On overflow, still write out a long to trigger ENOMEM. */
478 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
479 	ivalue = lvalue;
480 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
481 }
482 #else
483 static int
484 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
485 {
486 	long lvalue;
487 	int i;
488 
489 	lvalue = 0;
490 	for (i = 0; i < buf_domains; i++)
491 		lvalue += bdomain[i].bd_bufspace;
492 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
493 }
494 #endif
495 
496 static int
497 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
498 {
499 	int value;
500 	int i;
501 
502 	value = 0;
503 	for (i = 0; i < buf_domains; i++)
504 		value += bdomain[i].bd_numdirtybuffers;
505 	return (sysctl_handle_int(oidp, &value, 0, req));
506 }
507 
508 /*
509  *	bdirtywakeup:
510  *
511  *	Wakeup any bwillwrite() waiters.
512  */
513 static void
514 bdirtywakeup(void)
515 {
516 	mtx_lock(&bdirtylock);
517 	if (bdirtywait) {
518 		bdirtywait = 0;
519 		wakeup(&bdirtywait);
520 	}
521 	mtx_unlock(&bdirtylock);
522 }
523 
524 /*
525  *	bd_clear:
526  *
527  *	Clear a domain from the appropriate bitsets when dirtybuffers
528  *	is decremented.
529  */
530 static void
531 bd_clear(struct bufdomain *bd)
532 {
533 
534 	mtx_lock(&bdirtylock);
535 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
536 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
537 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
538 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
539 	mtx_unlock(&bdirtylock);
540 }
541 
542 /*
543  *	bd_set:
544  *
545  *	Set a domain in the appropriate bitsets when dirtybuffers
546  *	is incremented.
547  */
548 static void
549 bd_set(struct bufdomain *bd)
550 {
551 
552 	mtx_lock(&bdirtylock);
553 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
554 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
555 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
556 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
557 	mtx_unlock(&bdirtylock);
558 }
559 
560 /*
561  *	bdirtysub:
562  *
563  *	Decrement the numdirtybuffers count by one and wakeup any
564  *	threads blocked in bwillwrite().
565  */
566 static void
567 bdirtysub(struct buf *bp)
568 {
569 	struct bufdomain *bd;
570 	int num;
571 
572 	bd = bufdomain(bp);
573 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
574 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
575 		bdirtywakeup();
576 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
577 		bd_clear(bd);
578 }
579 
580 /*
581  *	bdirtyadd:
582  *
583  *	Increment the numdirtybuffers count by one and wakeup the buf
584  *	daemon if needed.
585  */
586 static void
587 bdirtyadd(struct buf *bp)
588 {
589 	struct bufdomain *bd;
590 	int num;
591 
592 	/*
593 	 * Only do the wakeup once as we cross the boundary.  The
594 	 * buf daemon will keep running until the condition clears.
595 	 */
596 	bd = bufdomain(bp);
597 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
598 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
599 		bd_wakeup();
600 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
601 		bd_set(bd);
602 }
603 
604 /*
605  *	bufspace_daemon_wakeup:
606  *
607  *	Wakeup the daemons responsible for freeing clean bufs.
608  */
609 static void
610 bufspace_daemon_wakeup(struct bufdomain *bd)
611 {
612 
613 	/*
614 	 * avoid the lock if the daemon is running.
615 	 */
616 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
617 		BD_RUN_LOCK(bd);
618 		atomic_store_int(&bd->bd_running, 1);
619 		wakeup(&bd->bd_running);
620 		BD_RUN_UNLOCK(bd);
621 	}
622 }
623 
624 /*
625  *	bufspace_daemon_wait:
626  *
627  *	Sleep until the domain falls below a limit or one second passes.
628  */
629 static void
630 bufspace_daemon_wait(struct bufdomain *bd)
631 {
632 	/*
633 	 * Re-check our limits and sleep.  bd_running must be
634 	 * cleared prior to checking the limits to avoid missed
635 	 * wakeups.  The waker will adjust one of bufspace or
636 	 * freebuffers prior to checking bd_running.
637 	 */
638 	BD_RUN_LOCK(bd);
639 	atomic_store_int(&bd->bd_running, 0);
640 	if (bd->bd_bufspace < bd->bd_bufspacethresh &&
641 	    bd->bd_freebuffers > bd->bd_lofreebuffers) {
642 		msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
643 		    "-", hz);
644 	} else {
645 		/* Avoid spurious wakeups while running. */
646 		atomic_store_int(&bd->bd_running, 1);
647 		BD_RUN_UNLOCK(bd);
648 	}
649 }
650 
651 /*
652  *	bufspace_adjust:
653  *
654  *	Adjust the reported bufspace for a KVA managed buffer, possibly
655  * 	waking any waiters.
656  */
657 static void
658 bufspace_adjust(struct buf *bp, int bufsize)
659 {
660 	struct bufdomain *bd;
661 	long space;
662 	int diff;
663 
664 	KASSERT((bp->b_flags & B_MALLOC) == 0,
665 	    ("bufspace_adjust: malloc buf %p", bp));
666 	bd = bufdomain(bp);
667 	diff = bufsize - bp->b_bufsize;
668 	if (diff < 0) {
669 		atomic_subtract_long(&bd->bd_bufspace, -diff);
670 	} else if (diff > 0) {
671 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
672 		/* Wake up the daemon on the transition. */
673 		if (space < bd->bd_bufspacethresh &&
674 		    space + diff >= bd->bd_bufspacethresh)
675 			bufspace_daemon_wakeup(bd);
676 	}
677 	bp->b_bufsize = bufsize;
678 }
679 
680 /*
681  *	bufspace_reserve:
682  *
683  *	Reserve bufspace before calling allocbuf().  metadata has a
684  *	different space limit than data.
685  */
686 static int
687 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
688 {
689 	long limit, new;
690 	long space;
691 
692 	if (metadata)
693 		limit = bd->bd_maxbufspace;
694 	else
695 		limit = bd->bd_hibufspace;
696 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
697 	new = space + size;
698 	if (new > limit) {
699 		atomic_subtract_long(&bd->bd_bufspace, size);
700 		return (ENOSPC);
701 	}
702 
703 	/* Wake up the daemon on the transition. */
704 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
705 		bufspace_daemon_wakeup(bd);
706 
707 	return (0);
708 }
709 
710 /*
711  *	bufspace_release:
712  *
713  *	Release reserved bufspace after bufspace_adjust() has consumed it.
714  */
715 static void
716 bufspace_release(struct bufdomain *bd, int size)
717 {
718 
719 	atomic_subtract_long(&bd->bd_bufspace, size);
720 }
721 
722 /*
723  *	bufspace_wait:
724  *
725  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
726  *	supplied.  bd_wanted must be set prior to polling for space.  The
727  *	operation must be re-tried on return.
728  */
729 static void
730 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
731     int slpflag, int slptimeo)
732 {
733 	struct thread *td;
734 	int error, fl, norunbuf;
735 
736 	if ((gbflags & GB_NOWAIT_BD) != 0)
737 		return;
738 
739 	td = curthread;
740 	BD_LOCK(bd);
741 	while (bd->bd_wanted) {
742 		if (vp != NULL && vp->v_type != VCHR &&
743 		    (td->td_pflags & TDP_BUFNEED) == 0) {
744 			BD_UNLOCK(bd);
745 			/*
746 			 * getblk() is called with a vnode locked, and
747 			 * some majority of the dirty buffers may as
748 			 * well belong to the vnode.  Flushing the
749 			 * buffers there would make a progress that
750 			 * cannot be achieved by the buf_daemon, that
751 			 * cannot lock the vnode.
752 			 */
753 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
754 			    (td->td_pflags & TDP_NORUNNINGBUF);
755 
756 			/*
757 			 * Play bufdaemon.  The getnewbuf() function
758 			 * may be called while the thread owns lock
759 			 * for another dirty buffer for the same
760 			 * vnode, which makes it impossible to use
761 			 * VOP_FSYNC() there, due to the buffer lock
762 			 * recursion.
763 			 */
764 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
765 			fl = buf_flush(vp, bd, flushbufqtarget);
766 			td->td_pflags &= norunbuf;
767 			BD_LOCK(bd);
768 			if (fl != 0)
769 				continue;
770 			if (bd->bd_wanted == 0)
771 				break;
772 		}
773 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
774 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
775 		if (error != 0)
776 			break;
777 	}
778 	BD_UNLOCK(bd);
779 }
780 
781 
782 /*
783  *	bufspace_daemon:
784  *
785  *	buffer space management daemon.  Tries to maintain some marginal
786  *	amount of free buffer space so that requesting processes neither
787  *	block nor work to reclaim buffers.
788  */
789 static void
790 bufspace_daemon(void *arg)
791 {
792 	struct bufdomain *bd;
793 
794 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
795 	    SHUTDOWN_PRI_LAST + 100);
796 
797 	bd = arg;
798 	for (;;) {
799 		kthread_suspend_check();
800 
801 		/*
802 		 * Free buffers from the clean queue until we meet our
803 		 * targets.
804 		 *
805 		 * Theory of operation:  The buffer cache is most efficient
806 		 * when some free buffer headers and space are always
807 		 * available to getnewbuf().  This daemon attempts to prevent
808 		 * the excessive blocking and synchronization associated
809 		 * with shortfall.  It goes through three phases according
810 		 * demand:
811 		 *
812 		 * 1)	The daemon wakes up voluntarily once per-second
813 		 *	during idle periods when the counters are below
814 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
815 		 *
816 		 * 2)	The daemon wakes up as we cross the thresholds
817 		 *	ahead of any potential blocking.  This may bounce
818 		 *	slightly according to the rate of consumption and
819 		 *	release.
820 		 *
821 		 * 3)	The daemon and consumers are starved for working
822 		 *	clean buffers.  This is the 'bufspace' sleep below
823 		 *	which will inefficiently trade bufs with bqrelse
824 		 *	until we return to condition 2.
825 		 */
826 		while (bd->bd_bufspace > bd->bd_lobufspace ||
827 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
828 			if (buf_recycle(bd, false) != 0) {
829 				if (bd_flushall(bd))
830 					continue;
831 				/*
832 				 * Speedup dirty if we've run out of clean
833 				 * buffers.  This is possible in particular
834 				 * because softdep may held many bufs locked
835 				 * pending writes to other bufs which are
836 				 * marked for delayed write, exhausting
837 				 * clean space until they are written.
838 				 */
839 				bd_speedup();
840 				BD_LOCK(bd);
841 				if (bd->bd_wanted) {
842 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
843 					    PRIBIO|PDROP, "bufspace", hz/10);
844 				} else
845 					BD_UNLOCK(bd);
846 			}
847 			maybe_yield();
848 		}
849 		bufspace_daemon_wait(bd);
850 	}
851 }
852 
853 /*
854  *	bufmallocadjust:
855  *
856  *	Adjust the reported bufspace for a malloc managed buffer, possibly
857  *	waking any waiters.
858  */
859 static void
860 bufmallocadjust(struct buf *bp, int bufsize)
861 {
862 	int diff;
863 
864 	KASSERT((bp->b_flags & B_MALLOC) != 0,
865 	    ("bufmallocadjust: non-malloc buf %p", bp));
866 	diff = bufsize - bp->b_bufsize;
867 	if (diff < 0)
868 		atomic_subtract_long(&bufmallocspace, -diff);
869 	else
870 		atomic_add_long(&bufmallocspace, diff);
871 	bp->b_bufsize = bufsize;
872 }
873 
874 /*
875  *	runningwakeup:
876  *
877  *	Wake up processes that are waiting on asynchronous writes to fall
878  *	below lorunningspace.
879  */
880 static void
881 runningwakeup(void)
882 {
883 
884 	mtx_lock(&rbreqlock);
885 	if (runningbufreq) {
886 		runningbufreq = 0;
887 		wakeup(&runningbufreq);
888 	}
889 	mtx_unlock(&rbreqlock);
890 }
891 
892 /*
893  *	runningbufwakeup:
894  *
895  *	Decrement the outstanding write count according.
896  */
897 void
898 runningbufwakeup(struct buf *bp)
899 {
900 	long space, bspace;
901 
902 	bspace = bp->b_runningbufspace;
903 	if (bspace == 0)
904 		return;
905 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
906 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
907 	    space, bspace));
908 	bp->b_runningbufspace = 0;
909 	/*
910 	 * Only acquire the lock and wakeup on the transition from exceeding
911 	 * the threshold to falling below it.
912 	 */
913 	if (space < lorunningspace)
914 		return;
915 	if (space - bspace > lorunningspace)
916 		return;
917 	runningwakeup();
918 }
919 
920 /*
921  *	waitrunningbufspace()
922  *
923  *	runningbufspace is a measure of the amount of I/O currently
924  *	running.  This routine is used in async-write situations to
925  *	prevent creating huge backups of pending writes to a device.
926  *	Only asynchronous writes are governed by this function.
927  *
928  *	This does NOT turn an async write into a sync write.  It waits
929  *	for earlier writes to complete and generally returns before the
930  *	caller's write has reached the device.
931  */
932 void
933 waitrunningbufspace(void)
934 {
935 
936 	mtx_lock(&rbreqlock);
937 	while (runningbufspace > hirunningspace) {
938 		runningbufreq = 1;
939 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
940 	}
941 	mtx_unlock(&rbreqlock);
942 }
943 
944 
945 /*
946  *	vfs_buf_test_cache:
947  *
948  *	Called when a buffer is extended.  This function clears the B_CACHE
949  *	bit if the newly extended portion of the buffer does not contain
950  *	valid data.
951  */
952 static __inline void
953 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
954     vm_offset_t size, vm_page_t m)
955 {
956 
957 	VM_OBJECT_ASSERT_LOCKED(m->object);
958 	if (bp->b_flags & B_CACHE) {
959 		int base = (foff + off) & PAGE_MASK;
960 		if (vm_page_is_valid(m, base, size) == 0)
961 			bp->b_flags &= ~B_CACHE;
962 	}
963 }
964 
965 /* Wake up the buffer daemon if necessary */
966 static void
967 bd_wakeup(void)
968 {
969 
970 	mtx_lock(&bdlock);
971 	if (bd_request == 0) {
972 		bd_request = 1;
973 		wakeup(&bd_request);
974 	}
975 	mtx_unlock(&bdlock);
976 }
977 
978 /*
979  * Adjust the maxbcachbuf tunable.
980  */
981 static void
982 maxbcachebuf_adjust(void)
983 {
984 	int i;
985 
986 	/*
987 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
988 	 */
989 	i = 2;
990 	while (i * 2 <= maxbcachebuf)
991 		i *= 2;
992 	maxbcachebuf = i;
993 	if (maxbcachebuf < MAXBSIZE)
994 		maxbcachebuf = MAXBSIZE;
995 	if (maxbcachebuf > MAXPHYS)
996 		maxbcachebuf = MAXPHYS;
997 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
998 		printf("maxbcachebuf=%d\n", maxbcachebuf);
999 }
1000 
1001 /*
1002  * bd_speedup - speedup the buffer cache flushing code
1003  */
1004 void
1005 bd_speedup(void)
1006 {
1007 	int needwake;
1008 
1009 	mtx_lock(&bdlock);
1010 	needwake = 0;
1011 	if (bd_speedupreq == 0 || bd_request == 0)
1012 		needwake = 1;
1013 	bd_speedupreq = 1;
1014 	bd_request = 1;
1015 	if (needwake)
1016 		wakeup(&bd_request);
1017 	mtx_unlock(&bdlock);
1018 }
1019 
1020 #ifdef __i386__
1021 #define	TRANSIENT_DENOM	5
1022 #else
1023 #define	TRANSIENT_DENOM 10
1024 #endif
1025 
1026 /*
1027  * Calculating buffer cache scaling values and reserve space for buffer
1028  * headers.  This is called during low level kernel initialization and
1029  * may be called more then once.  We CANNOT write to the memory area
1030  * being reserved at this time.
1031  */
1032 caddr_t
1033 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1034 {
1035 	int tuned_nbuf;
1036 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1037 
1038 	/*
1039 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1040 	 * PAGE_SIZE is >= 1K)
1041 	 */
1042 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1043 
1044 	maxbcachebuf_adjust();
1045 	/*
1046 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1047 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1048 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1049 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1050 	 * the buffer cache we limit the eventual kva reservation to
1051 	 * maxbcache bytes.
1052 	 *
1053 	 * factor represents the 1/4 x ram conversion.
1054 	 */
1055 	if (nbuf == 0) {
1056 		int factor = 4 * BKVASIZE / 1024;
1057 
1058 		nbuf = 50;
1059 		if (physmem_est > 4096)
1060 			nbuf += min((physmem_est - 4096) / factor,
1061 			    65536 / factor);
1062 		if (physmem_est > 65536)
1063 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1064 			    32 * 1024 * 1024 / (factor * 5));
1065 
1066 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1067 			nbuf = maxbcache / BKVASIZE;
1068 		tuned_nbuf = 1;
1069 	} else
1070 		tuned_nbuf = 0;
1071 
1072 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1073 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1074 	if (nbuf > maxbuf) {
1075 		if (!tuned_nbuf)
1076 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1077 			    maxbuf);
1078 		nbuf = maxbuf;
1079 	}
1080 
1081 	/*
1082 	 * Ideal allocation size for the transient bio submap is 10%
1083 	 * of the maximal space buffer map.  This roughly corresponds
1084 	 * to the amount of the buffer mapped for typical UFS load.
1085 	 *
1086 	 * Clip the buffer map to reserve space for the transient
1087 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1088 	 * maximum buffer map extent on the platform.
1089 	 *
1090 	 * The fall-back to the maxbuf in case of maxbcache unset,
1091 	 * allows to not trim the buffer KVA for the architectures
1092 	 * with ample KVA space.
1093 	 */
1094 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1095 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1096 		buf_sz = (long)nbuf * BKVASIZE;
1097 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1098 		    (TRANSIENT_DENOM - 1)) {
1099 			/*
1100 			 * There is more KVA than memory.  Do not
1101 			 * adjust buffer map size, and assign the rest
1102 			 * of maxbuf to transient map.
1103 			 */
1104 			biotmap_sz = maxbuf_sz - buf_sz;
1105 		} else {
1106 			/*
1107 			 * Buffer map spans all KVA we could afford on
1108 			 * this platform.  Give 10% (20% on i386) of
1109 			 * the buffer map to the transient bio map.
1110 			 */
1111 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1112 			buf_sz -= biotmap_sz;
1113 		}
1114 		if (biotmap_sz / INT_MAX > MAXPHYS)
1115 			bio_transient_maxcnt = INT_MAX;
1116 		else
1117 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1118 		/*
1119 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1120 		 * using the transient mapping.
1121 		 */
1122 		if (bio_transient_maxcnt > 1024)
1123 			bio_transient_maxcnt = 1024;
1124 		if (tuned_nbuf)
1125 			nbuf = buf_sz / BKVASIZE;
1126 	}
1127 
1128 	if (nswbuf == 0) {
1129 		nswbuf = min(nbuf / 4, 256);
1130 		if (nswbuf < NSWBUF_MIN)
1131 			nswbuf = NSWBUF_MIN;
1132 	}
1133 
1134 	/*
1135 	 * Reserve space for the buffer cache buffers
1136 	 */
1137 	buf = (void *)v;
1138 	v = (caddr_t)(buf + nbuf);
1139 
1140 	return(v);
1141 }
1142 
1143 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1144 void
1145 bufinit(void)
1146 {
1147 	struct buf *bp;
1148 	int i;
1149 
1150 	KASSERT(maxbcachebuf >= MAXBSIZE,
1151 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1152 	    MAXBSIZE));
1153 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1154 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1155 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1156 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1157 
1158 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1159 
1160 	/* finally, initialize each buffer header and stick on empty q */
1161 	for (i = 0; i < nbuf; i++) {
1162 		bp = &buf[i];
1163 		bzero(bp, sizeof *bp);
1164 		bp->b_flags = B_INVAL;
1165 		bp->b_rcred = NOCRED;
1166 		bp->b_wcred = NOCRED;
1167 		bp->b_qindex = QUEUE_NONE;
1168 		bp->b_domain = -1;
1169 		bp->b_subqueue = mp_maxid + 1;
1170 		bp->b_xflags = 0;
1171 		bp->b_data = bp->b_kvabase = unmapped_buf;
1172 		LIST_INIT(&bp->b_dep);
1173 		BUF_LOCKINIT(bp);
1174 		bq_insert(&bqempty, bp, false);
1175 	}
1176 
1177 	/*
1178 	 * maxbufspace is the absolute maximum amount of buffer space we are
1179 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1180 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1181 	 * used by most other requests.  The differential is required to
1182 	 * ensure that metadata deadlocks don't occur.
1183 	 *
1184 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1185 	 * this may result in KVM fragmentation which is not handled optimally
1186 	 * by the system. XXX This is less true with vmem.  We could use
1187 	 * PAGE_SIZE.
1188 	 */
1189 	maxbufspace = (long)nbuf * BKVASIZE;
1190 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1191 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1192 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1193 
1194 	/*
1195 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1196 	 * arbitrarily and may need further tuning. It corresponds to
1197 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1198 	 * which fits with many RAID controllers' tagged queuing limits.
1199 	 * The lower 1 MiB limit is the historical upper limit for
1200 	 * hirunningspace.
1201 	 */
1202 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1203 	    16 * 1024 * 1024), 1024 * 1024);
1204 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1205 
1206 	/*
1207 	 * Limit the amount of malloc memory since it is wired permanently into
1208 	 * the kernel space.  Even though this is accounted for in the buffer
1209 	 * allocation, we don't want the malloced region to grow uncontrolled.
1210 	 * The malloc scheme improves memory utilization significantly on
1211 	 * average (small) directories.
1212 	 */
1213 	maxbufmallocspace = hibufspace / 20;
1214 
1215 	/*
1216 	 * Reduce the chance of a deadlock occurring by limiting the number
1217 	 * of delayed-write dirty buffers we allow to stack up.
1218 	 */
1219 	hidirtybuffers = nbuf / 4 + 20;
1220 	dirtybufthresh = hidirtybuffers * 9 / 10;
1221 	/*
1222 	 * To support extreme low-memory systems, make sure hidirtybuffers
1223 	 * cannot eat up all available buffer space.  This occurs when our
1224 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1225 	 * buffer space assuming BKVASIZE'd buffers.
1226 	 */
1227 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1228 		hidirtybuffers >>= 1;
1229 	}
1230 	lodirtybuffers = hidirtybuffers / 2;
1231 
1232 	/*
1233 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1234 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1235 	 * are counted as free but will be unavailable to threads executing
1236 	 * on other cpus.
1237 	 *
1238 	 * hifreebuffers is the free target for the bufspace daemon.  This
1239 	 * should be set appropriately to limit work per-iteration.
1240 	 */
1241 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1242 	hifreebuffers = (3 * lofreebuffers) / 2;
1243 	numfreebuffers = nbuf;
1244 
1245 	/* Setup the kva and free list allocators. */
1246 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1247 	buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1248 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1249 
1250 	/*
1251 	 * Size the clean queue according to the amount of buffer space.
1252 	 * One queue per-256mb up to the max.  More queues gives better
1253 	 * concurrency but less accurate LRU.
1254 	 */
1255 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1256 	for (i = 0 ; i < buf_domains; i++) {
1257 		struct bufdomain *bd;
1258 
1259 		bd = &bdomain[i];
1260 		bd_init(bd);
1261 		bd->bd_freebuffers = nbuf / buf_domains;
1262 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1263 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1264 		bd->bd_bufspace = 0;
1265 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1266 		bd->bd_hibufspace = hibufspace / buf_domains;
1267 		bd->bd_lobufspace = lobufspace / buf_domains;
1268 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1269 		bd->bd_numdirtybuffers = 0;
1270 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1271 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1272 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1273 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1274 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1275 	}
1276 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1277 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1278 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1279 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1280 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1281 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1282 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1283 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1284 }
1285 
1286 #ifdef INVARIANTS
1287 static inline void
1288 vfs_buf_check_mapped(struct buf *bp)
1289 {
1290 
1291 	KASSERT(bp->b_kvabase != unmapped_buf,
1292 	    ("mapped buf: b_kvabase was not updated %p", bp));
1293 	KASSERT(bp->b_data != unmapped_buf,
1294 	    ("mapped buf: b_data was not updated %p", bp));
1295 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1296 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1297 }
1298 
1299 static inline void
1300 vfs_buf_check_unmapped(struct buf *bp)
1301 {
1302 
1303 	KASSERT(bp->b_data == unmapped_buf,
1304 	    ("unmapped buf: corrupted b_data %p", bp));
1305 }
1306 
1307 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1308 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1309 #else
1310 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1311 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1312 #endif
1313 
1314 static int
1315 isbufbusy(struct buf *bp)
1316 {
1317 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1318 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1319 		return (1);
1320 	return (0);
1321 }
1322 
1323 /*
1324  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1325  */
1326 void
1327 bufshutdown(int show_busybufs)
1328 {
1329 	static int first_buf_printf = 1;
1330 	struct buf *bp;
1331 	int iter, nbusy, pbusy;
1332 #ifndef PREEMPTION
1333 	int subiter;
1334 #endif
1335 
1336 	/*
1337 	 * Sync filesystems for shutdown
1338 	 */
1339 	wdog_kern_pat(WD_LASTVAL);
1340 	sys_sync(curthread, NULL);
1341 
1342 	/*
1343 	 * With soft updates, some buffers that are
1344 	 * written will be remarked as dirty until other
1345 	 * buffers are written.
1346 	 */
1347 	for (iter = pbusy = 0; iter < 20; iter++) {
1348 		nbusy = 0;
1349 		for (bp = &buf[nbuf]; --bp >= buf; )
1350 			if (isbufbusy(bp))
1351 				nbusy++;
1352 		if (nbusy == 0) {
1353 			if (first_buf_printf)
1354 				printf("All buffers synced.");
1355 			break;
1356 		}
1357 		if (first_buf_printf) {
1358 			printf("Syncing disks, buffers remaining... ");
1359 			first_buf_printf = 0;
1360 		}
1361 		printf("%d ", nbusy);
1362 		if (nbusy < pbusy)
1363 			iter = 0;
1364 		pbusy = nbusy;
1365 
1366 		wdog_kern_pat(WD_LASTVAL);
1367 		sys_sync(curthread, NULL);
1368 
1369 #ifdef PREEMPTION
1370 		/*
1371 		 * Spin for a while to allow interrupt threads to run.
1372 		 */
1373 		DELAY(50000 * iter);
1374 #else
1375 		/*
1376 		 * Context switch several times to allow interrupt
1377 		 * threads to run.
1378 		 */
1379 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1380 			thread_lock(curthread);
1381 			mi_switch(SW_VOL, NULL);
1382 			thread_unlock(curthread);
1383 			DELAY(1000);
1384 		}
1385 #endif
1386 	}
1387 	printf("\n");
1388 	/*
1389 	 * Count only busy local buffers to prevent forcing
1390 	 * a fsck if we're just a client of a wedged NFS server
1391 	 */
1392 	nbusy = 0;
1393 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1394 		if (isbufbusy(bp)) {
1395 #if 0
1396 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1397 			if (bp->b_dev == NULL) {
1398 				TAILQ_REMOVE(&mountlist,
1399 				    bp->b_vp->v_mount, mnt_list);
1400 				continue;
1401 			}
1402 #endif
1403 			nbusy++;
1404 			if (show_busybufs > 0) {
1405 				printf(
1406 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1407 				    nbusy, bp, bp->b_vp, bp->b_flags,
1408 				    (intmax_t)bp->b_blkno,
1409 				    (intmax_t)bp->b_lblkno);
1410 				BUF_LOCKPRINTINFO(bp);
1411 				if (show_busybufs > 1)
1412 					vn_printf(bp->b_vp,
1413 					    "vnode content: ");
1414 			}
1415 		}
1416 	}
1417 	if (nbusy) {
1418 		/*
1419 		 * Failed to sync all blocks. Indicate this and don't
1420 		 * unmount filesystems (thus forcing an fsck on reboot).
1421 		 */
1422 		printf("Giving up on %d buffers\n", nbusy);
1423 		DELAY(5000000);	/* 5 seconds */
1424 	} else {
1425 		if (!first_buf_printf)
1426 			printf("Final sync complete\n");
1427 		/*
1428 		 * Unmount filesystems
1429 		 */
1430 		if (panicstr == NULL)
1431 			vfs_unmountall();
1432 	}
1433 	swapoff_all();
1434 	DELAY(100000);		/* wait for console output to finish */
1435 }
1436 
1437 static void
1438 bpmap_qenter(struct buf *bp)
1439 {
1440 
1441 	BUF_CHECK_MAPPED(bp);
1442 
1443 	/*
1444 	 * bp->b_data is relative to bp->b_offset, but
1445 	 * bp->b_offset may be offset into the first page.
1446 	 */
1447 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1448 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1449 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1450 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1451 }
1452 
1453 static inline struct bufdomain *
1454 bufdomain(struct buf *bp)
1455 {
1456 
1457 	return (&bdomain[bp->b_domain]);
1458 }
1459 
1460 static struct bufqueue *
1461 bufqueue(struct buf *bp)
1462 {
1463 
1464 	switch (bp->b_qindex) {
1465 	case QUEUE_NONE:
1466 		/* FALLTHROUGH */
1467 	case QUEUE_SENTINEL:
1468 		return (NULL);
1469 	case QUEUE_EMPTY:
1470 		return (&bqempty);
1471 	case QUEUE_DIRTY:
1472 		return (&bufdomain(bp)->bd_dirtyq);
1473 	case QUEUE_CLEAN:
1474 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1475 	default:
1476 		break;
1477 	}
1478 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1479 }
1480 
1481 /*
1482  * Return the locked bufqueue that bp is a member of.
1483  */
1484 static struct bufqueue *
1485 bufqueue_acquire(struct buf *bp)
1486 {
1487 	struct bufqueue *bq, *nbq;
1488 
1489 	/*
1490 	 * bp can be pushed from a per-cpu queue to the
1491 	 * cleanq while we're waiting on the lock.  Retry
1492 	 * if the queues don't match.
1493 	 */
1494 	bq = bufqueue(bp);
1495 	BQ_LOCK(bq);
1496 	for (;;) {
1497 		nbq = bufqueue(bp);
1498 		if (bq == nbq)
1499 			break;
1500 		BQ_UNLOCK(bq);
1501 		BQ_LOCK(nbq);
1502 		bq = nbq;
1503 	}
1504 	return (bq);
1505 }
1506 
1507 /*
1508  *	binsfree:
1509  *
1510  *	Insert the buffer into the appropriate free list.  Requires a
1511  *	locked buffer on entry and buffer is unlocked before return.
1512  */
1513 static void
1514 binsfree(struct buf *bp, int qindex)
1515 {
1516 	struct bufdomain *bd;
1517 	struct bufqueue *bq;
1518 
1519 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1520 	    ("binsfree: Invalid qindex %d", qindex));
1521 	BUF_ASSERT_XLOCKED(bp);
1522 
1523 	/*
1524 	 * Handle delayed bremfree() processing.
1525 	 */
1526 	if (bp->b_flags & B_REMFREE) {
1527 		if (bp->b_qindex == qindex) {
1528 			bp->b_flags |= B_REUSE;
1529 			bp->b_flags &= ~B_REMFREE;
1530 			BUF_UNLOCK(bp);
1531 			return;
1532 		}
1533 		bq = bufqueue_acquire(bp);
1534 		bq_remove(bq, bp);
1535 		BQ_UNLOCK(bq);
1536 	}
1537 	bd = bufdomain(bp);
1538 	if (qindex == QUEUE_CLEAN) {
1539 		if (bd->bd_lim != 0)
1540 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1541 		else
1542 			bq = bd->bd_cleanq;
1543 	} else
1544 		bq = &bd->bd_dirtyq;
1545 	bq_insert(bq, bp, true);
1546 }
1547 
1548 /*
1549  * buf_free:
1550  *
1551  *	Free a buffer to the buf zone once it no longer has valid contents.
1552  */
1553 static void
1554 buf_free(struct buf *bp)
1555 {
1556 
1557 	if (bp->b_flags & B_REMFREE)
1558 		bremfreef(bp);
1559 	if (bp->b_vflags & BV_BKGRDINPROG)
1560 		panic("losing buffer 1");
1561 	if (bp->b_rcred != NOCRED) {
1562 		crfree(bp->b_rcred);
1563 		bp->b_rcred = NOCRED;
1564 	}
1565 	if (bp->b_wcred != NOCRED) {
1566 		crfree(bp->b_wcred);
1567 		bp->b_wcred = NOCRED;
1568 	}
1569 	if (!LIST_EMPTY(&bp->b_dep))
1570 		buf_deallocate(bp);
1571 	bufkva_free(bp);
1572 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1573 	BUF_UNLOCK(bp);
1574 	uma_zfree(buf_zone, bp);
1575 }
1576 
1577 /*
1578  * buf_import:
1579  *
1580  *	Import bufs into the uma cache from the buf list.  The system still
1581  *	expects a static array of bufs and much of the synchronization
1582  *	around bufs assumes type stable storage.  As a result, UMA is used
1583  *	only as a per-cpu cache of bufs still maintained on a global list.
1584  */
1585 static int
1586 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1587 {
1588 	struct buf *bp;
1589 	int i;
1590 
1591 	BQ_LOCK(&bqempty);
1592 	for (i = 0; i < cnt; i++) {
1593 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1594 		if (bp == NULL)
1595 			break;
1596 		bq_remove(&bqempty, bp);
1597 		store[i] = bp;
1598 	}
1599 	BQ_UNLOCK(&bqempty);
1600 
1601 	return (i);
1602 }
1603 
1604 /*
1605  * buf_release:
1606  *
1607  *	Release bufs from the uma cache back to the buffer queues.
1608  */
1609 static void
1610 buf_release(void *arg, void **store, int cnt)
1611 {
1612 	struct bufqueue *bq;
1613 	struct buf *bp;
1614         int i;
1615 
1616 	bq = &bqempty;
1617 	BQ_LOCK(bq);
1618         for (i = 0; i < cnt; i++) {
1619 		bp = store[i];
1620 		/* Inline bq_insert() to batch locking. */
1621 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1622 		bp->b_flags &= ~(B_AGE | B_REUSE);
1623 		bq->bq_len++;
1624 		bp->b_qindex = bq->bq_index;
1625 	}
1626 	BQ_UNLOCK(bq);
1627 }
1628 
1629 /*
1630  * buf_alloc:
1631  *
1632  *	Allocate an empty buffer header.
1633  */
1634 static struct buf *
1635 buf_alloc(struct bufdomain *bd)
1636 {
1637 	struct buf *bp;
1638 	int freebufs;
1639 
1640 	/*
1641 	 * We can only run out of bufs in the buf zone if the average buf
1642 	 * is less than BKVASIZE.  In this case the actual wait/block will
1643 	 * come from buf_reycle() failing to flush one of these small bufs.
1644 	 */
1645 	bp = NULL;
1646 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1647 	if (freebufs > 0)
1648 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1649 	if (bp == NULL) {
1650 		atomic_add_int(&bd->bd_freebuffers, 1);
1651 		bufspace_daemon_wakeup(bd);
1652 		counter_u64_add(numbufallocfails, 1);
1653 		return (NULL);
1654 	}
1655 	/*
1656 	 * Wake-up the bufspace daemon on transition below threshold.
1657 	 */
1658 	if (freebufs == bd->bd_lofreebuffers)
1659 		bufspace_daemon_wakeup(bd);
1660 
1661 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1662 		panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1663 
1664 	KASSERT(bp->b_vp == NULL,
1665 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1666 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1667 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1668 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1669 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1670 	KASSERT(bp->b_npages == 0,
1671 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1672 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1673 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1674 
1675 	bp->b_domain = BD_DOMAIN(bd);
1676 	bp->b_flags = 0;
1677 	bp->b_ioflags = 0;
1678 	bp->b_xflags = 0;
1679 	bp->b_vflags = 0;
1680 	bp->b_vp = NULL;
1681 	bp->b_blkno = bp->b_lblkno = 0;
1682 	bp->b_offset = NOOFFSET;
1683 	bp->b_iodone = 0;
1684 	bp->b_error = 0;
1685 	bp->b_resid = 0;
1686 	bp->b_bcount = 0;
1687 	bp->b_npages = 0;
1688 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1689 	bp->b_bufobj = NULL;
1690 	bp->b_data = bp->b_kvabase = unmapped_buf;
1691 	bp->b_fsprivate1 = NULL;
1692 	bp->b_fsprivate2 = NULL;
1693 	bp->b_fsprivate3 = NULL;
1694 	LIST_INIT(&bp->b_dep);
1695 
1696 	return (bp);
1697 }
1698 
1699 /*
1700  *	buf_recycle:
1701  *
1702  *	Free a buffer from the given bufqueue.  kva controls whether the
1703  *	freed buf must own some kva resources.  This is used for
1704  *	defragmenting.
1705  */
1706 static int
1707 buf_recycle(struct bufdomain *bd, bool kva)
1708 {
1709 	struct bufqueue *bq;
1710 	struct buf *bp, *nbp;
1711 
1712 	if (kva)
1713 		counter_u64_add(bufdefragcnt, 1);
1714 	nbp = NULL;
1715 	bq = bd->bd_cleanq;
1716 	BQ_LOCK(bq);
1717 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1718 	    ("buf_recycle: Locks don't match"));
1719 	nbp = TAILQ_FIRST(&bq->bq_queue);
1720 
1721 	/*
1722 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1723 	 * depending.
1724 	 */
1725 	while ((bp = nbp) != NULL) {
1726 		/*
1727 		 * Calculate next bp (we can only use it if we do not
1728 		 * release the bqlock).
1729 		 */
1730 		nbp = TAILQ_NEXT(bp, b_freelist);
1731 
1732 		/*
1733 		 * If we are defragging then we need a buffer with
1734 		 * some kva to reclaim.
1735 		 */
1736 		if (kva && bp->b_kvasize == 0)
1737 			continue;
1738 
1739 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1740 			continue;
1741 
1742 		/*
1743 		 * Implement a second chance algorithm for frequently
1744 		 * accessed buffers.
1745 		 */
1746 		if ((bp->b_flags & B_REUSE) != 0) {
1747 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1748 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1749 			bp->b_flags &= ~B_REUSE;
1750 			BUF_UNLOCK(bp);
1751 			continue;
1752 		}
1753 
1754 		/*
1755 		 * Skip buffers with background writes in progress.
1756 		 */
1757 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1758 			BUF_UNLOCK(bp);
1759 			continue;
1760 		}
1761 
1762 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1763 		    ("buf_recycle: inconsistent queue %d bp %p",
1764 		    bp->b_qindex, bp));
1765 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1766 		    ("getnewbuf: queue domain %d doesn't match request %d",
1767 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1768 		/*
1769 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1770 		 * the scan from this point on.
1771 		 */
1772 		bq_remove(bq, bp);
1773 		BQ_UNLOCK(bq);
1774 
1775 		/*
1776 		 * Requeue the background write buffer with error and
1777 		 * restart the scan.
1778 		 */
1779 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1780 			bqrelse(bp);
1781 			BQ_LOCK(bq);
1782 			nbp = TAILQ_FIRST(&bq->bq_queue);
1783 			continue;
1784 		}
1785 		bp->b_flags |= B_INVAL;
1786 		brelse(bp);
1787 		return (0);
1788 	}
1789 	bd->bd_wanted = 1;
1790 	BQ_UNLOCK(bq);
1791 
1792 	return (ENOBUFS);
1793 }
1794 
1795 /*
1796  *	bremfree:
1797  *
1798  *	Mark the buffer for removal from the appropriate free list.
1799  *
1800  */
1801 void
1802 bremfree(struct buf *bp)
1803 {
1804 
1805 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1806 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1807 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1808 	KASSERT(bp->b_qindex != QUEUE_NONE,
1809 	    ("bremfree: buffer %p not on a queue.", bp));
1810 	BUF_ASSERT_XLOCKED(bp);
1811 
1812 	bp->b_flags |= B_REMFREE;
1813 }
1814 
1815 /*
1816  *	bremfreef:
1817  *
1818  *	Force an immediate removal from a free list.  Used only in nfs when
1819  *	it abuses the b_freelist pointer.
1820  */
1821 void
1822 bremfreef(struct buf *bp)
1823 {
1824 	struct bufqueue *bq;
1825 
1826 	bq = bufqueue_acquire(bp);
1827 	bq_remove(bq, bp);
1828 	BQ_UNLOCK(bq);
1829 }
1830 
1831 static void
1832 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1833 {
1834 
1835 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1836 	TAILQ_INIT(&bq->bq_queue);
1837 	bq->bq_len = 0;
1838 	bq->bq_index = qindex;
1839 	bq->bq_subqueue = subqueue;
1840 }
1841 
1842 static void
1843 bd_init(struct bufdomain *bd)
1844 {
1845 	int i;
1846 
1847 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1848 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1849 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1850 	for (i = 0; i <= mp_maxid; i++)
1851 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1852 		    "bufq clean subqueue lock");
1853 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1854 }
1855 
1856 /*
1857  *	bq_remove:
1858  *
1859  *	Removes a buffer from the free list, must be called with the
1860  *	correct qlock held.
1861  */
1862 static void
1863 bq_remove(struct bufqueue *bq, struct buf *bp)
1864 {
1865 
1866 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1867 	    bp, bp->b_vp, bp->b_flags);
1868 	KASSERT(bp->b_qindex != QUEUE_NONE,
1869 	    ("bq_remove: buffer %p not on a queue.", bp));
1870 	KASSERT(bufqueue(bp) == bq,
1871 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1872 
1873 	BQ_ASSERT_LOCKED(bq);
1874 	if (bp->b_qindex != QUEUE_EMPTY) {
1875 		BUF_ASSERT_XLOCKED(bp);
1876 	}
1877 	KASSERT(bq->bq_len >= 1,
1878 	    ("queue %d underflow", bp->b_qindex));
1879 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1880 	bq->bq_len--;
1881 	bp->b_qindex = QUEUE_NONE;
1882 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1883 }
1884 
1885 static void
1886 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1887 {
1888 	struct buf *bp;
1889 
1890 	BQ_ASSERT_LOCKED(bq);
1891 	if (bq != bd->bd_cleanq) {
1892 		BD_LOCK(bd);
1893 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1894 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1895 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1896 			    b_freelist);
1897 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1898 		}
1899 		bd->bd_cleanq->bq_len += bq->bq_len;
1900 		bq->bq_len = 0;
1901 	}
1902 	if (bd->bd_wanted) {
1903 		bd->bd_wanted = 0;
1904 		wakeup(&bd->bd_wanted);
1905 	}
1906 	if (bq != bd->bd_cleanq)
1907 		BD_UNLOCK(bd);
1908 }
1909 
1910 static int
1911 bd_flushall(struct bufdomain *bd)
1912 {
1913 	struct bufqueue *bq;
1914 	int flushed;
1915 	int i;
1916 
1917 	if (bd->bd_lim == 0)
1918 		return (0);
1919 	flushed = 0;
1920 	for (i = 0; i <= mp_maxid; i++) {
1921 		bq = &bd->bd_subq[i];
1922 		if (bq->bq_len == 0)
1923 			continue;
1924 		BQ_LOCK(bq);
1925 		bd_flush(bd, bq);
1926 		BQ_UNLOCK(bq);
1927 		flushed++;
1928 	}
1929 
1930 	return (flushed);
1931 }
1932 
1933 static void
1934 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1935 {
1936 	struct bufdomain *bd;
1937 
1938 	if (bp->b_qindex != QUEUE_NONE)
1939 		panic("bq_insert: free buffer %p onto another queue?", bp);
1940 
1941 	bd = bufdomain(bp);
1942 	if (bp->b_flags & B_AGE) {
1943 		/* Place this buf directly on the real queue. */
1944 		if (bq->bq_index == QUEUE_CLEAN)
1945 			bq = bd->bd_cleanq;
1946 		BQ_LOCK(bq);
1947 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1948 	} else {
1949 		BQ_LOCK(bq);
1950 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1951 	}
1952 	bp->b_flags &= ~(B_AGE | B_REUSE);
1953 	bq->bq_len++;
1954 	bp->b_qindex = bq->bq_index;
1955 	bp->b_subqueue = bq->bq_subqueue;
1956 
1957 	/*
1958 	 * Unlock before we notify so that we don't wakeup a waiter that
1959 	 * fails a trylock on the buf and sleeps again.
1960 	 */
1961 	if (unlock)
1962 		BUF_UNLOCK(bp);
1963 
1964 	if (bp->b_qindex == QUEUE_CLEAN) {
1965 		/*
1966 		 * Flush the per-cpu queue and notify any waiters.
1967 		 */
1968 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1969 		    bq->bq_len >= bd->bd_lim))
1970 			bd_flush(bd, bq);
1971 	}
1972 	BQ_UNLOCK(bq);
1973 }
1974 
1975 /*
1976  *	bufkva_free:
1977  *
1978  *	Free the kva allocation for a buffer.
1979  *
1980  */
1981 static void
1982 bufkva_free(struct buf *bp)
1983 {
1984 
1985 #ifdef INVARIANTS
1986 	if (bp->b_kvasize == 0) {
1987 		KASSERT(bp->b_kvabase == unmapped_buf &&
1988 		    bp->b_data == unmapped_buf,
1989 		    ("Leaked KVA space on %p", bp));
1990 	} else if (buf_mapped(bp))
1991 		BUF_CHECK_MAPPED(bp);
1992 	else
1993 		BUF_CHECK_UNMAPPED(bp);
1994 #endif
1995 	if (bp->b_kvasize == 0)
1996 		return;
1997 
1998 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
1999 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2000 	counter_u64_add(buffreekvacnt, 1);
2001 	bp->b_data = bp->b_kvabase = unmapped_buf;
2002 	bp->b_kvasize = 0;
2003 }
2004 
2005 /*
2006  *	bufkva_alloc:
2007  *
2008  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2009  */
2010 static int
2011 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2012 {
2013 	vm_offset_t addr;
2014 	int error;
2015 
2016 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2017 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2018 
2019 	bufkva_free(bp);
2020 
2021 	addr = 0;
2022 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2023 	if (error != 0) {
2024 		/*
2025 		 * Buffer map is too fragmented.  Request the caller
2026 		 * to defragment the map.
2027 		 */
2028 		return (error);
2029 	}
2030 	bp->b_kvabase = (caddr_t)addr;
2031 	bp->b_kvasize = maxsize;
2032 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2033 	if ((gbflags & GB_UNMAPPED) != 0) {
2034 		bp->b_data = unmapped_buf;
2035 		BUF_CHECK_UNMAPPED(bp);
2036 	} else {
2037 		bp->b_data = bp->b_kvabase;
2038 		BUF_CHECK_MAPPED(bp);
2039 	}
2040 	return (0);
2041 }
2042 
2043 /*
2044  *	bufkva_reclaim:
2045  *
2046  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2047  *	callback that fires to avoid returning failure.
2048  */
2049 static void
2050 bufkva_reclaim(vmem_t *vmem, int flags)
2051 {
2052 	bool done;
2053 	int q;
2054 	int i;
2055 
2056 	done = false;
2057 	for (i = 0; i < 5; i++) {
2058 		for (q = 0; q < buf_domains; q++)
2059 			if (buf_recycle(&bdomain[q], true) != 0)
2060 				done = true;
2061 		if (done)
2062 			break;
2063 	}
2064 	return;
2065 }
2066 
2067 /*
2068  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2069  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2070  * the buffer is valid and we do not have to do anything.
2071  */
2072 static void
2073 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2074     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2075 {
2076 	struct buf *rabp;
2077 	struct thread *td;
2078 	int i;
2079 
2080 	td = curthread;
2081 
2082 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2083 		if (inmem(vp, *rablkno))
2084 			continue;
2085 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2086 		if ((rabp->b_flags & B_CACHE) != 0) {
2087 			brelse(rabp);
2088 			continue;
2089 		}
2090 #ifdef RACCT
2091 		if (racct_enable) {
2092 			PROC_LOCK(curproc);
2093 			racct_add_buf(curproc, rabp, 0);
2094 			PROC_UNLOCK(curproc);
2095 		}
2096 #endif /* RACCT */
2097 		td->td_ru.ru_inblock++;
2098 		rabp->b_flags |= B_ASYNC;
2099 		rabp->b_flags &= ~B_INVAL;
2100 		if ((flags & GB_CKHASH) != 0) {
2101 			rabp->b_flags |= B_CKHASH;
2102 			rabp->b_ckhashcalc = ckhashfunc;
2103 		}
2104 		rabp->b_ioflags &= ~BIO_ERROR;
2105 		rabp->b_iocmd = BIO_READ;
2106 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2107 			rabp->b_rcred = crhold(cred);
2108 		vfs_busy_pages(rabp, 0);
2109 		BUF_KERNPROC(rabp);
2110 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2111 		bstrategy(rabp);
2112 	}
2113 }
2114 
2115 /*
2116  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2117  *
2118  * Get a buffer with the specified data.  Look in the cache first.  We
2119  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2120  * is set, the buffer is valid and we do not have to do anything, see
2121  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2122  *
2123  * Always return a NULL buffer pointer (in bpp) when returning an error.
2124  */
2125 int
2126 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
2127     int *rabsize, int cnt, struct ucred *cred, int flags,
2128     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2129 {
2130 	struct buf *bp;
2131 	struct thread *td;
2132 	int error, readwait, rv;
2133 
2134 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2135 	td = curthread;
2136 	/*
2137 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2138 	 * are specified.
2139 	 */
2140 	error = getblkx(vp, blkno, size, 0, 0, flags, &bp);
2141 	if (error != 0) {
2142 		*bpp = NULL;
2143 		return (error);
2144 	}
2145 	flags &= ~GB_NOSPARSE;
2146 	*bpp = bp;
2147 
2148 	/*
2149 	 * If not found in cache, do some I/O
2150 	 */
2151 	readwait = 0;
2152 	if ((bp->b_flags & B_CACHE) == 0) {
2153 #ifdef RACCT
2154 		if (racct_enable) {
2155 			PROC_LOCK(td->td_proc);
2156 			racct_add_buf(td->td_proc, bp, 0);
2157 			PROC_UNLOCK(td->td_proc);
2158 		}
2159 #endif /* RACCT */
2160 		td->td_ru.ru_inblock++;
2161 		bp->b_iocmd = BIO_READ;
2162 		bp->b_flags &= ~B_INVAL;
2163 		if ((flags & GB_CKHASH) != 0) {
2164 			bp->b_flags |= B_CKHASH;
2165 			bp->b_ckhashcalc = ckhashfunc;
2166 		}
2167 		bp->b_ioflags &= ~BIO_ERROR;
2168 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2169 			bp->b_rcred = crhold(cred);
2170 		vfs_busy_pages(bp, 0);
2171 		bp->b_iooffset = dbtob(bp->b_blkno);
2172 		bstrategy(bp);
2173 		++readwait;
2174 	}
2175 
2176 	/*
2177 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2178 	 */
2179 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2180 
2181 	rv = 0;
2182 	if (readwait) {
2183 		rv = bufwait(bp);
2184 		if (rv != 0) {
2185 			brelse(bp);
2186 			*bpp = NULL;
2187 		}
2188 	}
2189 	return (rv);
2190 }
2191 
2192 /*
2193  * Write, release buffer on completion.  (Done by iodone
2194  * if async).  Do not bother writing anything if the buffer
2195  * is invalid.
2196  *
2197  * Note that we set B_CACHE here, indicating that buffer is
2198  * fully valid and thus cacheable.  This is true even of NFS
2199  * now so we set it generally.  This could be set either here
2200  * or in biodone() since the I/O is synchronous.  We put it
2201  * here.
2202  */
2203 int
2204 bufwrite(struct buf *bp)
2205 {
2206 	int oldflags;
2207 	struct vnode *vp;
2208 	long space;
2209 	int vp_md;
2210 
2211 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2212 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2213 		bp->b_flags |= B_INVAL | B_RELBUF;
2214 		bp->b_flags &= ~B_CACHE;
2215 		brelse(bp);
2216 		return (ENXIO);
2217 	}
2218 	if (bp->b_flags & B_INVAL) {
2219 		brelse(bp);
2220 		return (0);
2221 	}
2222 
2223 	if (bp->b_flags & B_BARRIER)
2224 		atomic_add_long(&barrierwrites, 1);
2225 
2226 	oldflags = bp->b_flags;
2227 
2228 	BUF_ASSERT_HELD(bp);
2229 
2230 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2231 	    ("FFS background buffer should not get here %p", bp));
2232 
2233 	vp = bp->b_vp;
2234 	if (vp)
2235 		vp_md = vp->v_vflag & VV_MD;
2236 	else
2237 		vp_md = 0;
2238 
2239 	/*
2240 	 * Mark the buffer clean.  Increment the bufobj write count
2241 	 * before bundirty() call, to prevent other thread from seeing
2242 	 * empty dirty list and zero counter for writes in progress,
2243 	 * falsely indicating that the bufobj is clean.
2244 	 */
2245 	bufobj_wref(bp->b_bufobj);
2246 	bundirty(bp);
2247 
2248 	bp->b_flags &= ~B_DONE;
2249 	bp->b_ioflags &= ~BIO_ERROR;
2250 	bp->b_flags |= B_CACHE;
2251 	bp->b_iocmd = BIO_WRITE;
2252 
2253 	vfs_busy_pages(bp, 1);
2254 
2255 	/*
2256 	 * Normal bwrites pipeline writes
2257 	 */
2258 	bp->b_runningbufspace = bp->b_bufsize;
2259 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2260 
2261 #ifdef RACCT
2262 	if (racct_enable) {
2263 		PROC_LOCK(curproc);
2264 		racct_add_buf(curproc, bp, 1);
2265 		PROC_UNLOCK(curproc);
2266 	}
2267 #endif /* RACCT */
2268 	curthread->td_ru.ru_oublock++;
2269 	if (oldflags & B_ASYNC)
2270 		BUF_KERNPROC(bp);
2271 	bp->b_iooffset = dbtob(bp->b_blkno);
2272 	buf_track(bp, __func__);
2273 	bstrategy(bp);
2274 
2275 	if ((oldflags & B_ASYNC) == 0) {
2276 		int rtval = bufwait(bp);
2277 		brelse(bp);
2278 		return (rtval);
2279 	} else if (space > hirunningspace) {
2280 		/*
2281 		 * don't allow the async write to saturate the I/O
2282 		 * system.  We will not deadlock here because
2283 		 * we are blocking waiting for I/O that is already in-progress
2284 		 * to complete. We do not block here if it is the update
2285 		 * or syncer daemon trying to clean up as that can lead
2286 		 * to deadlock.
2287 		 */
2288 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2289 			waitrunningbufspace();
2290 	}
2291 
2292 	return (0);
2293 }
2294 
2295 void
2296 bufbdflush(struct bufobj *bo, struct buf *bp)
2297 {
2298 	struct buf *nbp;
2299 
2300 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2301 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2302 		altbufferflushes++;
2303 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2304 		BO_LOCK(bo);
2305 		/*
2306 		 * Try to find a buffer to flush.
2307 		 */
2308 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2309 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2310 			    BUF_LOCK(nbp,
2311 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2312 				continue;
2313 			if (bp == nbp)
2314 				panic("bdwrite: found ourselves");
2315 			BO_UNLOCK(bo);
2316 			/* Don't countdeps with the bo lock held. */
2317 			if (buf_countdeps(nbp, 0)) {
2318 				BO_LOCK(bo);
2319 				BUF_UNLOCK(nbp);
2320 				continue;
2321 			}
2322 			if (nbp->b_flags & B_CLUSTEROK) {
2323 				vfs_bio_awrite(nbp);
2324 			} else {
2325 				bremfree(nbp);
2326 				bawrite(nbp);
2327 			}
2328 			dirtybufferflushes++;
2329 			break;
2330 		}
2331 		if (nbp == NULL)
2332 			BO_UNLOCK(bo);
2333 	}
2334 }
2335 
2336 /*
2337  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2338  * anything if the buffer is marked invalid.
2339  *
2340  * Note that since the buffer must be completely valid, we can safely
2341  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2342  * biodone() in order to prevent getblk from writing the buffer
2343  * out synchronously.
2344  */
2345 void
2346 bdwrite(struct buf *bp)
2347 {
2348 	struct thread *td = curthread;
2349 	struct vnode *vp;
2350 	struct bufobj *bo;
2351 
2352 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2353 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2354 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2355 	    ("Barrier request in delayed write %p", bp));
2356 	BUF_ASSERT_HELD(bp);
2357 
2358 	if (bp->b_flags & B_INVAL) {
2359 		brelse(bp);
2360 		return;
2361 	}
2362 
2363 	/*
2364 	 * If we have too many dirty buffers, don't create any more.
2365 	 * If we are wildly over our limit, then force a complete
2366 	 * cleanup. Otherwise, just keep the situation from getting
2367 	 * out of control. Note that we have to avoid a recursive
2368 	 * disaster and not try to clean up after our own cleanup!
2369 	 */
2370 	vp = bp->b_vp;
2371 	bo = bp->b_bufobj;
2372 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2373 		td->td_pflags |= TDP_INBDFLUSH;
2374 		BO_BDFLUSH(bo, bp);
2375 		td->td_pflags &= ~TDP_INBDFLUSH;
2376 	} else
2377 		recursiveflushes++;
2378 
2379 	bdirty(bp);
2380 	/*
2381 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2382 	 * true even of NFS now.
2383 	 */
2384 	bp->b_flags |= B_CACHE;
2385 
2386 	/*
2387 	 * This bmap keeps the system from needing to do the bmap later,
2388 	 * perhaps when the system is attempting to do a sync.  Since it
2389 	 * is likely that the indirect block -- or whatever other datastructure
2390 	 * that the filesystem needs is still in memory now, it is a good
2391 	 * thing to do this.  Note also, that if the pageout daemon is
2392 	 * requesting a sync -- there might not be enough memory to do
2393 	 * the bmap then...  So, this is important to do.
2394 	 */
2395 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2396 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2397 	}
2398 
2399 	buf_track(bp, __func__);
2400 
2401 	/*
2402 	 * Set the *dirty* buffer range based upon the VM system dirty
2403 	 * pages.
2404 	 *
2405 	 * Mark the buffer pages as clean.  We need to do this here to
2406 	 * satisfy the vnode_pager and the pageout daemon, so that it
2407 	 * thinks that the pages have been "cleaned".  Note that since
2408 	 * the pages are in a delayed write buffer -- the VFS layer
2409 	 * "will" see that the pages get written out on the next sync,
2410 	 * or perhaps the cluster will be completed.
2411 	 */
2412 	vfs_clean_pages_dirty_buf(bp);
2413 	bqrelse(bp);
2414 
2415 	/*
2416 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2417 	 * due to the softdep code.
2418 	 */
2419 }
2420 
2421 /*
2422  *	bdirty:
2423  *
2424  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2425  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2426  *	itself to properly update it in the dirty/clean lists.  We mark it
2427  *	B_DONE to ensure that any asynchronization of the buffer properly
2428  *	clears B_DONE ( else a panic will occur later ).
2429  *
2430  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2431  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2432  *	should only be called if the buffer is known-good.
2433  *
2434  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2435  *	count.
2436  *
2437  *	The buffer must be on QUEUE_NONE.
2438  */
2439 void
2440 bdirty(struct buf *bp)
2441 {
2442 
2443 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2444 	    bp, bp->b_vp, bp->b_flags);
2445 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2446 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2447 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2448 	BUF_ASSERT_HELD(bp);
2449 	bp->b_flags &= ~(B_RELBUF);
2450 	bp->b_iocmd = BIO_WRITE;
2451 
2452 	if ((bp->b_flags & B_DELWRI) == 0) {
2453 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2454 		reassignbuf(bp);
2455 		bdirtyadd(bp);
2456 	}
2457 }
2458 
2459 /*
2460  *	bundirty:
2461  *
2462  *	Clear B_DELWRI for buffer.
2463  *
2464  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2465  *	count.
2466  *
2467  *	The buffer must be on QUEUE_NONE.
2468  */
2469 
2470 void
2471 bundirty(struct buf *bp)
2472 {
2473 
2474 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2475 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2476 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2477 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2478 	BUF_ASSERT_HELD(bp);
2479 
2480 	if (bp->b_flags & B_DELWRI) {
2481 		bp->b_flags &= ~B_DELWRI;
2482 		reassignbuf(bp);
2483 		bdirtysub(bp);
2484 	}
2485 	/*
2486 	 * Since it is now being written, we can clear its deferred write flag.
2487 	 */
2488 	bp->b_flags &= ~B_DEFERRED;
2489 }
2490 
2491 /*
2492  *	bawrite:
2493  *
2494  *	Asynchronous write.  Start output on a buffer, but do not wait for
2495  *	it to complete.  The buffer is released when the output completes.
2496  *
2497  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2498  *	B_INVAL buffers.  Not us.
2499  */
2500 void
2501 bawrite(struct buf *bp)
2502 {
2503 
2504 	bp->b_flags |= B_ASYNC;
2505 	(void) bwrite(bp);
2506 }
2507 
2508 /*
2509  *	babarrierwrite:
2510  *
2511  *	Asynchronous barrier write.  Start output on a buffer, but do not
2512  *	wait for it to complete.  Place a write barrier after this write so
2513  *	that this buffer and all buffers written before it are committed to
2514  *	the disk before any buffers written after this write are committed
2515  *	to the disk.  The buffer is released when the output completes.
2516  */
2517 void
2518 babarrierwrite(struct buf *bp)
2519 {
2520 
2521 	bp->b_flags |= B_ASYNC | B_BARRIER;
2522 	(void) bwrite(bp);
2523 }
2524 
2525 /*
2526  *	bbarrierwrite:
2527  *
2528  *	Synchronous barrier write.  Start output on a buffer and wait for
2529  *	it to complete.  Place a write barrier after this write so that
2530  *	this buffer and all buffers written before it are committed to
2531  *	the disk before any buffers written after this write are committed
2532  *	to the disk.  The buffer is released when the output completes.
2533  */
2534 int
2535 bbarrierwrite(struct buf *bp)
2536 {
2537 
2538 	bp->b_flags |= B_BARRIER;
2539 	return (bwrite(bp));
2540 }
2541 
2542 /*
2543  *	bwillwrite:
2544  *
2545  *	Called prior to the locking of any vnodes when we are expecting to
2546  *	write.  We do not want to starve the buffer cache with too many
2547  *	dirty buffers so we block here.  By blocking prior to the locking
2548  *	of any vnodes we attempt to avoid the situation where a locked vnode
2549  *	prevents the various system daemons from flushing related buffers.
2550  */
2551 void
2552 bwillwrite(void)
2553 {
2554 
2555 	if (buf_dirty_count_severe()) {
2556 		mtx_lock(&bdirtylock);
2557 		while (buf_dirty_count_severe()) {
2558 			bdirtywait = 1;
2559 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2560 			    "flswai", 0);
2561 		}
2562 		mtx_unlock(&bdirtylock);
2563 	}
2564 }
2565 
2566 /*
2567  * Return true if we have too many dirty buffers.
2568  */
2569 int
2570 buf_dirty_count_severe(void)
2571 {
2572 
2573 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2574 }
2575 
2576 /*
2577  *	brelse:
2578  *
2579  *	Release a busy buffer and, if requested, free its resources.  The
2580  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2581  *	to be accessed later as a cache entity or reused for other purposes.
2582  */
2583 void
2584 brelse(struct buf *bp)
2585 {
2586 	struct mount *v_mnt;
2587 	int qindex;
2588 
2589 	/*
2590 	 * Many functions erroneously call brelse with a NULL bp under rare
2591 	 * error conditions. Simply return when called with a NULL bp.
2592 	 */
2593 	if (bp == NULL)
2594 		return;
2595 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2596 	    bp, bp->b_vp, bp->b_flags);
2597 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2598 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2599 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2600 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2601 
2602 	if (BUF_LOCKRECURSED(bp)) {
2603 		/*
2604 		 * Do not process, in particular, do not handle the
2605 		 * B_INVAL/B_RELBUF and do not release to free list.
2606 		 */
2607 		BUF_UNLOCK(bp);
2608 		return;
2609 	}
2610 
2611 	if (bp->b_flags & B_MANAGED) {
2612 		bqrelse(bp);
2613 		return;
2614 	}
2615 
2616 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2617 		BO_LOCK(bp->b_bufobj);
2618 		bp->b_vflags &= ~BV_BKGRDERR;
2619 		BO_UNLOCK(bp->b_bufobj);
2620 		bdirty(bp);
2621 	}
2622 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2623 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2624 	    !(bp->b_flags & B_INVAL)) {
2625 		/*
2626 		 * Failed write, redirty.  All errors except ENXIO (which
2627 		 * means the device is gone) are treated as being
2628 		 * transient.
2629 		 *
2630 		 * XXX Treating EIO as transient is not correct; the
2631 		 * contract with the local storage device drivers is that
2632 		 * they will only return EIO once the I/O is no longer
2633 		 * retriable.  Network I/O also respects this through the
2634 		 * guarantees of TCP and/or the internal retries of NFS.
2635 		 * ENOMEM might be transient, but we also have no way of
2636 		 * knowing when its ok to retry/reschedule.  In general,
2637 		 * this entire case should be made obsolete through better
2638 		 * error handling/recovery and resource scheduling.
2639 		 *
2640 		 * Do this also for buffers that failed with ENXIO, but have
2641 		 * non-empty dependencies - the soft updates code might need
2642 		 * to access the buffer to untangle them.
2643 		 *
2644 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2645 		 */
2646 		bp->b_ioflags &= ~BIO_ERROR;
2647 		bdirty(bp);
2648 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2649 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2650 		/*
2651 		 * Either a failed read I/O, or we were asked to free or not
2652 		 * cache the buffer, or we failed to write to a device that's
2653 		 * no longer present.
2654 		 */
2655 		bp->b_flags |= B_INVAL;
2656 		if (!LIST_EMPTY(&bp->b_dep))
2657 			buf_deallocate(bp);
2658 		if (bp->b_flags & B_DELWRI)
2659 			bdirtysub(bp);
2660 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2661 		if ((bp->b_flags & B_VMIO) == 0) {
2662 			allocbuf(bp, 0);
2663 			if (bp->b_vp)
2664 				brelvp(bp);
2665 		}
2666 	}
2667 
2668 	/*
2669 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2670 	 * is called with B_DELWRI set, the underlying pages may wind up
2671 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2672 	 * because pages associated with a B_DELWRI bp are marked clean.
2673 	 *
2674 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2675 	 * if B_DELWRI is set.
2676 	 */
2677 	if (bp->b_flags & B_DELWRI)
2678 		bp->b_flags &= ~B_RELBUF;
2679 
2680 	/*
2681 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2682 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2683 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2684 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2685 	 *
2686 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2687 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2688 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2689 	 *
2690 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2691 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2692 	 * the commit state and we cannot afford to lose the buffer. If the
2693 	 * buffer has a background write in progress, we need to keep it
2694 	 * around to prevent it from being reconstituted and starting a second
2695 	 * background write.
2696 	 */
2697 
2698 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2699 
2700 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2701 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2702 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2703 	    vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2704 		vfs_vmio_invalidate(bp);
2705 		allocbuf(bp, 0);
2706 	}
2707 
2708 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2709 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2710 		allocbuf(bp, 0);
2711 		bp->b_flags &= ~B_NOREUSE;
2712 		if (bp->b_vp != NULL)
2713 			brelvp(bp);
2714 	}
2715 
2716 	/*
2717 	 * If the buffer has junk contents signal it and eventually
2718 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2719 	 * doesn't find it.
2720 	 */
2721 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2722 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2723 		bp->b_flags |= B_INVAL;
2724 	if (bp->b_flags & B_INVAL) {
2725 		if (bp->b_flags & B_DELWRI)
2726 			bundirty(bp);
2727 		if (bp->b_vp)
2728 			brelvp(bp);
2729 	}
2730 
2731 	buf_track(bp, __func__);
2732 
2733 	/* buffers with no memory */
2734 	if (bp->b_bufsize == 0) {
2735 		buf_free(bp);
2736 		return;
2737 	}
2738 	/* buffers with junk contents */
2739 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2740 	    (bp->b_ioflags & BIO_ERROR)) {
2741 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2742 		if (bp->b_vflags & BV_BKGRDINPROG)
2743 			panic("losing buffer 2");
2744 		qindex = QUEUE_CLEAN;
2745 		bp->b_flags |= B_AGE;
2746 	/* remaining buffers */
2747 	} else if (bp->b_flags & B_DELWRI)
2748 		qindex = QUEUE_DIRTY;
2749 	else
2750 		qindex = QUEUE_CLEAN;
2751 
2752 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2753 		panic("brelse: not dirty");
2754 
2755 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2756 	/* binsfree unlocks bp. */
2757 	binsfree(bp, qindex);
2758 }
2759 
2760 /*
2761  * Release a buffer back to the appropriate queue but do not try to free
2762  * it.  The buffer is expected to be used again soon.
2763  *
2764  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2765  * biodone() to requeue an async I/O on completion.  It is also used when
2766  * known good buffers need to be requeued but we think we may need the data
2767  * again soon.
2768  *
2769  * XXX we should be able to leave the B_RELBUF hint set on completion.
2770  */
2771 void
2772 bqrelse(struct buf *bp)
2773 {
2774 	int qindex;
2775 
2776 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2777 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2778 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2779 
2780 	qindex = QUEUE_NONE;
2781 	if (BUF_LOCKRECURSED(bp)) {
2782 		/* do not release to free list */
2783 		BUF_UNLOCK(bp);
2784 		return;
2785 	}
2786 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2787 
2788 	if (bp->b_flags & B_MANAGED) {
2789 		if (bp->b_flags & B_REMFREE)
2790 			bremfreef(bp);
2791 		goto out;
2792 	}
2793 
2794 	/* buffers with stale but valid contents */
2795 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2796 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2797 		BO_LOCK(bp->b_bufobj);
2798 		bp->b_vflags &= ~BV_BKGRDERR;
2799 		BO_UNLOCK(bp->b_bufobj);
2800 		qindex = QUEUE_DIRTY;
2801 	} else {
2802 		if ((bp->b_flags & B_DELWRI) == 0 &&
2803 		    (bp->b_xflags & BX_VNDIRTY))
2804 			panic("bqrelse: not dirty");
2805 		if ((bp->b_flags & B_NOREUSE) != 0) {
2806 			brelse(bp);
2807 			return;
2808 		}
2809 		qindex = QUEUE_CLEAN;
2810 	}
2811 	buf_track(bp, __func__);
2812 	/* binsfree unlocks bp. */
2813 	binsfree(bp, qindex);
2814 	return;
2815 
2816 out:
2817 	buf_track(bp, __func__);
2818 	/* unlock */
2819 	BUF_UNLOCK(bp);
2820 }
2821 
2822 /*
2823  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2824  * restore bogus pages.
2825  */
2826 static void
2827 vfs_vmio_iodone(struct buf *bp)
2828 {
2829 	vm_ooffset_t foff;
2830 	vm_page_t m;
2831 	vm_object_t obj;
2832 	struct vnode *vp __unused;
2833 	int i, iosize, resid;
2834 	bool bogus;
2835 
2836 	obj = bp->b_bufobj->bo_object;
2837 	KASSERT(obj->paging_in_progress >= bp->b_npages,
2838 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2839 	    obj->paging_in_progress, bp->b_npages));
2840 
2841 	vp = bp->b_vp;
2842 	KASSERT(vp->v_holdcnt > 0,
2843 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2844 	KASSERT(vp->v_object != NULL,
2845 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2846 
2847 	foff = bp->b_offset;
2848 	KASSERT(bp->b_offset != NOOFFSET,
2849 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2850 
2851 	bogus = false;
2852 	iosize = bp->b_bcount - bp->b_resid;
2853 	VM_OBJECT_WLOCK(obj);
2854 	for (i = 0; i < bp->b_npages; i++) {
2855 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2856 		if (resid > iosize)
2857 			resid = iosize;
2858 
2859 		/*
2860 		 * cleanup bogus pages, restoring the originals
2861 		 */
2862 		m = bp->b_pages[i];
2863 		if (m == bogus_page) {
2864 			bogus = true;
2865 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2866 			if (m == NULL)
2867 				panic("biodone: page disappeared!");
2868 			bp->b_pages[i] = m;
2869 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2870 			/*
2871 			 * In the write case, the valid and clean bits are
2872 			 * already changed correctly ( see bdwrite() ), so we
2873 			 * only need to do this here in the read case.
2874 			 */
2875 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2876 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2877 			    "has unexpected dirty bits", m));
2878 			vfs_page_set_valid(bp, foff, m);
2879 		}
2880 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2881 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2882 		    (intmax_t)foff, (uintmax_t)m->pindex));
2883 
2884 		vm_page_sunbusy(m);
2885 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2886 		iosize -= resid;
2887 	}
2888 	vm_object_pip_wakeupn(obj, bp->b_npages);
2889 	VM_OBJECT_WUNLOCK(obj);
2890 	if (bogus && buf_mapped(bp)) {
2891 		BUF_CHECK_MAPPED(bp);
2892 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2893 		    bp->b_pages, bp->b_npages);
2894 	}
2895 }
2896 
2897 /*
2898  * Unwire a page held by a buf and either free it or update the page queues to
2899  * reflect its recent use.
2900  */
2901 static void
2902 vfs_vmio_unwire(struct buf *bp, vm_page_t m)
2903 {
2904 	bool freed;
2905 
2906 	vm_page_lock(m);
2907 	if (vm_page_unwire_noq(m)) {
2908 		if ((bp->b_flags & B_DIRECT) != 0)
2909 			freed = vm_page_try_to_free(m);
2910 		else
2911 			freed = false;
2912 		if (!freed) {
2913 			/*
2914 			 * Use a racy check of the valid bits to determine
2915 			 * whether we can accelerate reclamation of the page.
2916 			 * The valid bits will be stable unless the page is
2917 			 * being mapped or is referenced by multiple buffers,
2918 			 * and in those cases we expect races to be rare.  At
2919 			 * worst we will either accelerate reclamation of a
2920 			 * valid page and violate LRU, or unnecessarily defer
2921 			 * reclamation of an invalid page.
2922 			 *
2923 			 * The B_NOREUSE flag marks data that is not expected to
2924 			 * be reused, so accelerate reclamation in that case
2925 			 * too.  Otherwise, maintain LRU.
2926 			 */
2927 			if (m->valid == 0 || (bp->b_flags & B_NOREUSE) != 0)
2928 				vm_page_deactivate_noreuse(m);
2929 			else if (vm_page_active(m))
2930 				vm_page_reference(m);
2931 			else
2932 				vm_page_deactivate(m);
2933 		}
2934 	}
2935 	vm_page_unlock(m);
2936 }
2937 
2938 /*
2939  * Perform page invalidation when a buffer is released.  The fully invalid
2940  * pages will be reclaimed later in vfs_vmio_truncate().
2941  */
2942 static void
2943 vfs_vmio_invalidate(struct buf *bp)
2944 {
2945 	vm_object_t obj;
2946 	vm_page_t m;
2947 	int i, resid, poffset, presid;
2948 
2949 	if (buf_mapped(bp)) {
2950 		BUF_CHECK_MAPPED(bp);
2951 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2952 	} else
2953 		BUF_CHECK_UNMAPPED(bp);
2954 	/*
2955 	 * Get the base offset and length of the buffer.  Note that
2956 	 * in the VMIO case if the buffer block size is not
2957 	 * page-aligned then b_data pointer may not be page-aligned.
2958 	 * But our b_pages[] array *IS* page aligned.
2959 	 *
2960 	 * block sizes less then DEV_BSIZE (usually 512) are not
2961 	 * supported due to the page granularity bits (m->valid,
2962 	 * m->dirty, etc...).
2963 	 *
2964 	 * See man buf(9) for more information
2965 	 */
2966 	obj = bp->b_bufobj->bo_object;
2967 	resid = bp->b_bufsize;
2968 	poffset = bp->b_offset & PAGE_MASK;
2969 	VM_OBJECT_WLOCK(obj);
2970 	for (i = 0; i < bp->b_npages; i++) {
2971 		m = bp->b_pages[i];
2972 		if (m == bogus_page)
2973 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2974 		bp->b_pages[i] = NULL;
2975 
2976 		presid = resid > (PAGE_SIZE - poffset) ?
2977 		    (PAGE_SIZE - poffset) : resid;
2978 		KASSERT(presid >= 0, ("brelse: extra page"));
2979 		while (vm_page_xbusied(m)) {
2980 			vm_page_lock(m);
2981 			VM_OBJECT_WUNLOCK(obj);
2982 			vm_page_busy_sleep(m, "mbncsh", true);
2983 			VM_OBJECT_WLOCK(obj);
2984 		}
2985 		if (pmap_page_wired_mappings(m) == 0)
2986 			vm_page_set_invalid(m, poffset, presid);
2987 		vfs_vmio_unwire(bp, m);
2988 		resid -= presid;
2989 		poffset = 0;
2990 	}
2991 	VM_OBJECT_WUNLOCK(obj);
2992 	bp->b_npages = 0;
2993 }
2994 
2995 /*
2996  * Page-granular truncation of an existing VMIO buffer.
2997  */
2998 static void
2999 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3000 {
3001 	vm_object_t obj;
3002 	vm_page_t m;
3003 	int i;
3004 
3005 	if (bp->b_npages == desiredpages)
3006 		return;
3007 
3008 	if (buf_mapped(bp)) {
3009 		BUF_CHECK_MAPPED(bp);
3010 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3011 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3012 	} else
3013 		BUF_CHECK_UNMAPPED(bp);
3014 
3015 	/*
3016 	 * The object lock is needed only if we will attempt to free pages.
3017 	 */
3018 	obj = (bp->b_flags & B_DIRECT) != 0 ? bp->b_bufobj->bo_object : NULL;
3019 	if (obj != NULL)
3020 		VM_OBJECT_WLOCK(obj);
3021 	for (i = desiredpages; i < bp->b_npages; i++) {
3022 		m = bp->b_pages[i];
3023 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3024 		bp->b_pages[i] = NULL;
3025 		vfs_vmio_unwire(bp, m);
3026 	}
3027 	if (obj != NULL)
3028 		VM_OBJECT_WUNLOCK(obj);
3029 	bp->b_npages = desiredpages;
3030 }
3031 
3032 /*
3033  * Byte granular extension of VMIO buffers.
3034  */
3035 static void
3036 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3037 {
3038 	/*
3039 	 * We are growing the buffer, possibly in a
3040 	 * byte-granular fashion.
3041 	 */
3042 	vm_object_t obj;
3043 	vm_offset_t toff;
3044 	vm_offset_t tinc;
3045 	vm_page_t m;
3046 
3047 	/*
3048 	 * Step 1, bring in the VM pages from the object, allocating
3049 	 * them if necessary.  We must clear B_CACHE if these pages
3050 	 * are not valid for the range covered by the buffer.
3051 	 */
3052 	obj = bp->b_bufobj->bo_object;
3053 	VM_OBJECT_WLOCK(obj);
3054 	if (bp->b_npages < desiredpages) {
3055 		/*
3056 		 * We must allocate system pages since blocking
3057 		 * here could interfere with paging I/O, no
3058 		 * matter which process we are.
3059 		 *
3060 		 * Only exclusive busy can be tested here.
3061 		 * Blocking on shared busy might lead to
3062 		 * deadlocks once allocbuf() is called after
3063 		 * pages are vfs_busy_pages().
3064 		 */
3065 		(void)vm_page_grab_pages(obj,
3066 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3067 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3068 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3069 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3070 		bp->b_npages = desiredpages;
3071 	}
3072 
3073 	/*
3074 	 * Step 2.  We've loaded the pages into the buffer,
3075 	 * we have to figure out if we can still have B_CACHE
3076 	 * set.  Note that B_CACHE is set according to the
3077 	 * byte-granular range ( bcount and size ), not the
3078 	 * aligned range ( newbsize ).
3079 	 *
3080 	 * The VM test is against m->valid, which is DEV_BSIZE
3081 	 * aligned.  Needless to say, the validity of the data
3082 	 * needs to also be DEV_BSIZE aligned.  Note that this
3083 	 * fails with NFS if the server or some other client
3084 	 * extends the file's EOF.  If our buffer is resized,
3085 	 * B_CACHE may remain set! XXX
3086 	 */
3087 	toff = bp->b_bcount;
3088 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3089 	while ((bp->b_flags & B_CACHE) && toff < size) {
3090 		vm_pindex_t pi;
3091 
3092 		if (tinc > (size - toff))
3093 			tinc = size - toff;
3094 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3095 		m = bp->b_pages[pi];
3096 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3097 		toff += tinc;
3098 		tinc = PAGE_SIZE;
3099 	}
3100 	VM_OBJECT_WUNLOCK(obj);
3101 
3102 	/*
3103 	 * Step 3, fixup the KVA pmap.
3104 	 */
3105 	if (buf_mapped(bp))
3106 		bpmap_qenter(bp);
3107 	else
3108 		BUF_CHECK_UNMAPPED(bp);
3109 }
3110 
3111 /*
3112  * Check to see if a block at a particular lbn is available for a clustered
3113  * write.
3114  */
3115 static int
3116 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3117 {
3118 	struct buf *bpa;
3119 	int match;
3120 
3121 	match = 0;
3122 
3123 	/* If the buf isn't in core skip it */
3124 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3125 		return (0);
3126 
3127 	/* If the buf is busy we don't want to wait for it */
3128 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3129 		return (0);
3130 
3131 	/* Only cluster with valid clusterable delayed write buffers */
3132 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3133 	    (B_DELWRI | B_CLUSTEROK))
3134 		goto done;
3135 
3136 	if (bpa->b_bufsize != size)
3137 		goto done;
3138 
3139 	/*
3140 	 * Check to see if it is in the expected place on disk and that the
3141 	 * block has been mapped.
3142 	 */
3143 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3144 		match = 1;
3145 done:
3146 	BUF_UNLOCK(bpa);
3147 	return (match);
3148 }
3149 
3150 /*
3151  *	vfs_bio_awrite:
3152  *
3153  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3154  *	This is much better then the old way of writing only one buffer at
3155  *	a time.  Note that we may not be presented with the buffers in the
3156  *	correct order, so we search for the cluster in both directions.
3157  */
3158 int
3159 vfs_bio_awrite(struct buf *bp)
3160 {
3161 	struct bufobj *bo;
3162 	int i;
3163 	int j;
3164 	daddr_t lblkno = bp->b_lblkno;
3165 	struct vnode *vp = bp->b_vp;
3166 	int ncl;
3167 	int nwritten;
3168 	int size;
3169 	int maxcl;
3170 	int gbflags;
3171 
3172 	bo = &vp->v_bufobj;
3173 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3174 	/*
3175 	 * right now we support clustered writing only to regular files.  If
3176 	 * we find a clusterable block we could be in the middle of a cluster
3177 	 * rather then at the beginning.
3178 	 */
3179 	if ((vp->v_type == VREG) &&
3180 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3181 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3182 
3183 		size = vp->v_mount->mnt_stat.f_iosize;
3184 		maxcl = MAXPHYS / size;
3185 
3186 		BO_RLOCK(bo);
3187 		for (i = 1; i < maxcl; i++)
3188 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3189 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3190 				break;
3191 
3192 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3193 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3194 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3195 				break;
3196 		BO_RUNLOCK(bo);
3197 		--j;
3198 		ncl = i + j;
3199 		/*
3200 		 * this is a possible cluster write
3201 		 */
3202 		if (ncl != 1) {
3203 			BUF_UNLOCK(bp);
3204 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3205 			    gbflags);
3206 			return (nwritten);
3207 		}
3208 	}
3209 	bremfree(bp);
3210 	bp->b_flags |= B_ASYNC;
3211 	/*
3212 	 * default (old) behavior, writing out only one block
3213 	 *
3214 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3215 	 */
3216 	nwritten = bp->b_bufsize;
3217 	(void) bwrite(bp);
3218 
3219 	return (nwritten);
3220 }
3221 
3222 /*
3223  *	getnewbuf_kva:
3224  *
3225  *	Allocate KVA for an empty buf header according to gbflags.
3226  */
3227 static int
3228 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3229 {
3230 
3231 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3232 		/*
3233 		 * In order to keep fragmentation sane we only allocate kva
3234 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3235 		 */
3236 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3237 
3238 		if (maxsize != bp->b_kvasize &&
3239 		    bufkva_alloc(bp, maxsize, gbflags))
3240 			return (ENOSPC);
3241 	}
3242 	return (0);
3243 }
3244 
3245 /*
3246  *	getnewbuf:
3247  *
3248  *	Find and initialize a new buffer header, freeing up existing buffers
3249  *	in the bufqueues as necessary.  The new buffer is returned locked.
3250  *
3251  *	We block if:
3252  *		We have insufficient buffer headers
3253  *		We have insufficient buffer space
3254  *		buffer_arena is too fragmented ( space reservation fails )
3255  *		If we have to flush dirty buffers ( but we try to avoid this )
3256  *
3257  *	The caller is responsible for releasing the reserved bufspace after
3258  *	allocbuf() is called.
3259  */
3260 static struct buf *
3261 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3262 {
3263 	struct bufdomain *bd;
3264 	struct buf *bp;
3265 	bool metadata, reserved;
3266 
3267 	bp = NULL;
3268 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3269 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3270 	if (!unmapped_buf_allowed)
3271 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3272 
3273 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3274 	    vp->v_type == VCHR)
3275 		metadata = true;
3276 	else
3277 		metadata = false;
3278 	if (vp == NULL)
3279 		bd = &bdomain[0];
3280 	else
3281 		bd = &bdomain[vp->v_bufobj.bo_domain];
3282 
3283 	counter_u64_add(getnewbufcalls, 1);
3284 	reserved = false;
3285 	do {
3286 		if (reserved == false &&
3287 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3288 			counter_u64_add(getnewbufrestarts, 1);
3289 			continue;
3290 		}
3291 		reserved = true;
3292 		if ((bp = buf_alloc(bd)) == NULL) {
3293 			counter_u64_add(getnewbufrestarts, 1);
3294 			continue;
3295 		}
3296 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3297 			return (bp);
3298 		break;
3299 	} while (buf_recycle(bd, false) == 0);
3300 
3301 	if (reserved)
3302 		bufspace_release(bd, maxsize);
3303 	if (bp != NULL) {
3304 		bp->b_flags |= B_INVAL;
3305 		brelse(bp);
3306 	}
3307 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3308 
3309 	return (NULL);
3310 }
3311 
3312 /*
3313  *	buf_daemon:
3314  *
3315  *	buffer flushing daemon.  Buffers are normally flushed by the
3316  *	update daemon but if it cannot keep up this process starts to
3317  *	take the load in an attempt to prevent getnewbuf() from blocking.
3318  */
3319 static struct kproc_desc buf_kp = {
3320 	"bufdaemon",
3321 	buf_daemon,
3322 	&bufdaemonproc
3323 };
3324 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3325 
3326 static int
3327 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3328 {
3329 	int flushed;
3330 
3331 	flushed = flushbufqueues(vp, bd, target, 0);
3332 	if (flushed == 0) {
3333 		/*
3334 		 * Could not find any buffers without rollback
3335 		 * dependencies, so just write the first one
3336 		 * in the hopes of eventually making progress.
3337 		 */
3338 		if (vp != NULL && target > 2)
3339 			target /= 2;
3340 		flushbufqueues(vp, bd, target, 1);
3341 	}
3342 	return (flushed);
3343 }
3344 
3345 static void
3346 buf_daemon()
3347 {
3348 	struct bufdomain *bd;
3349 	int speedupreq;
3350 	int lodirty;
3351 	int i;
3352 
3353 	/*
3354 	 * This process needs to be suspended prior to shutdown sync.
3355 	 */
3356 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3357 	    SHUTDOWN_PRI_LAST + 100);
3358 
3359 	/*
3360 	 * Start the buf clean daemons as children threads.
3361 	 */
3362 	for (i = 0 ; i < buf_domains; i++) {
3363 		int error;
3364 
3365 		error = kthread_add((void (*)(void *))bufspace_daemon,
3366 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3367 		if (error)
3368 			panic("error %d spawning bufspace daemon", error);
3369 	}
3370 
3371 	/*
3372 	 * This process is allowed to take the buffer cache to the limit
3373 	 */
3374 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3375 	mtx_lock(&bdlock);
3376 	for (;;) {
3377 		bd_request = 0;
3378 		mtx_unlock(&bdlock);
3379 
3380 		kthread_suspend_check();
3381 
3382 		/*
3383 		 * Save speedupreq for this pass and reset to capture new
3384 		 * requests.
3385 		 */
3386 		speedupreq = bd_speedupreq;
3387 		bd_speedupreq = 0;
3388 
3389 		/*
3390 		 * Flush each domain sequentially according to its level and
3391 		 * the speedup request.
3392 		 */
3393 		for (i = 0; i < buf_domains; i++) {
3394 			bd = &bdomain[i];
3395 			if (speedupreq)
3396 				lodirty = bd->bd_numdirtybuffers / 2;
3397 			else
3398 				lodirty = bd->bd_lodirtybuffers;
3399 			while (bd->bd_numdirtybuffers > lodirty) {
3400 				if (buf_flush(NULL, bd,
3401 				    bd->bd_numdirtybuffers - lodirty) == 0)
3402 					break;
3403 				kern_yield(PRI_USER);
3404 			}
3405 		}
3406 
3407 		/*
3408 		 * Only clear bd_request if we have reached our low water
3409 		 * mark.  The buf_daemon normally waits 1 second and
3410 		 * then incrementally flushes any dirty buffers that have
3411 		 * built up, within reason.
3412 		 *
3413 		 * If we were unable to hit our low water mark and couldn't
3414 		 * find any flushable buffers, we sleep for a short period
3415 		 * to avoid endless loops on unlockable buffers.
3416 		 */
3417 		mtx_lock(&bdlock);
3418 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3419 			/*
3420 			 * We reached our low water mark, reset the
3421 			 * request and sleep until we are needed again.
3422 			 * The sleep is just so the suspend code works.
3423 			 */
3424 			bd_request = 0;
3425 			/*
3426 			 * Do an extra wakeup in case dirty threshold
3427 			 * changed via sysctl and the explicit transition
3428 			 * out of shortfall was missed.
3429 			 */
3430 			bdirtywakeup();
3431 			if (runningbufspace <= lorunningspace)
3432 				runningwakeup();
3433 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3434 		} else {
3435 			/*
3436 			 * We couldn't find any flushable dirty buffers but
3437 			 * still have too many dirty buffers, we
3438 			 * have to sleep and try again.  (rare)
3439 			 */
3440 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3441 		}
3442 	}
3443 }
3444 
3445 /*
3446  *	flushbufqueues:
3447  *
3448  *	Try to flush a buffer in the dirty queue.  We must be careful to
3449  *	free up B_INVAL buffers instead of write them, which NFS is
3450  *	particularly sensitive to.
3451  */
3452 static int flushwithdeps = 0;
3453 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
3454     0, "Number of buffers flushed with dependecies that require rollbacks");
3455 
3456 static int
3457 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3458     int flushdeps)
3459 {
3460 	struct bufqueue *bq;
3461 	struct buf *sentinel;
3462 	struct vnode *vp;
3463 	struct mount *mp;
3464 	struct buf *bp;
3465 	int hasdeps;
3466 	int flushed;
3467 	int error;
3468 	bool unlock;
3469 
3470 	flushed = 0;
3471 	bq = &bd->bd_dirtyq;
3472 	bp = NULL;
3473 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3474 	sentinel->b_qindex = QUEUE_SENTINEL;
3475 	BQ_LOCK(bq);
3476 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3477 	BQ_UNLOCK(bq);
3478 	while (flushed != target) {
3479 		maybe_yield();
3480 		BQ_LOCK(bq);
3481 		bp = TAILQ_NEXT(sentinel, b_freelist);
3482 		if (bp != NULL) {
3483 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3484 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3485 			    b_freelist);
3486 		} else {
3487 			BQ_UNLOCK(bq);
3488 			break;
3489 		}
3490 		/*
3491 		 * Skip sentinels inserted by other invocations of the
3492 		 * flushbufqueues(), taking care to not reorder them.
3493 		 *
3494 		 * Only flush the buffers that belong to the
3495 		 * vnode locked by the curthread.
3496 		 */
3497 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3498 		    bp->b_vp != lvp)) {
3499 			BQ_UNLOCK(bq);
3500 			continue;
3501 		}
3502 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3503 		BQ_UNLOCK(bq);
3504 		if (error != 0)
3505 			continue;
3506 
3507 		/*
3508 		 * BKGRDINPROG can only be set with the buf and bufobj
3509 		 * locks both held.  We tolerate a race to clear it here.
3510 		 */
3511 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3512 		    (bp->b_flags & B_DELWRI) == 0) {
3513 			BUF_UNLOCK(bp);
3514 			continue;
3515 		}
3516 		if (bp->b_flags & B_INVAL) {
3517 			bremfreef(bp);
3518 			brelse(bp);
3519 			flushed++;
3520 			continue;
3521 		}
3522 
3523 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3524 			if (flushdeps == 0) {
3525 				BUF_UNLOCK(bp);
3526 				continue;
3527 			}
3528 			hasdeps = 1;
3529 		} else
3530 			hasdeps = 0;
3531 		/*
3532 		 * We must hold the lock on a vnode before writing
3533 		 * one of its buffers. Otherwise we may confuse, or
3534 		 * in the case of a snapshot vnode, deadlock the
3535 		 * system.
3536 		 *
3537 		 * The lock order here is the reverse of the normal
3538 		 * of vnode followed by buf lock.  This is ok because
3539 		 * the NOWAIT will prevent deadlock.
3540 		 */
3541 		vp = bp->b_vp;
3542 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3543 			BUF_UNLOCK(bp);
3544 			continue;
3545 		}
3546 		if (lvp == NULL) {
3547 			unlock = true;
3548 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3549 		} else {
3550 			ASSERT_VOP_LOCKED(vp, "getbuf");
3551 			unlock = false;
3552 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3553 			    vn_lock(vp, LK_TRYUPGRADE);
3554 		}
3555 		if (error == 0) {
3556 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3557 			    bp, bp->b_vp, bp->b_flags);
3558 			if (curproc == bufdaemonproc) {
3559 				vfs_bio_awrite(bp);
3560 			} else {
3561 				bremfree(bp);
3562 				bwrite(bp);
3563 				counter_u64_add(notbufdflushes, 1);
3564 			}
3565 			vn_finished_write(mp);
3566 			if (unlock)
3567 				VOP_UNLOCK(vp, 0);
3568 			flushwithdeps += hasdeps;
3569 			flushed++;
3570 
3571 			/*
3572 			 * Sleeping on runningbufspace while holding
3573 			 * vnode lock leads to deadlock.
3574 			 */
3575 			if (curproc == bufdaemonproc &&
3576 			    runningbufspace > hirunningspace)
3577 				waitrunningbufspace();
3578 			continue;
3579 		}
3580 		vn_finished_write(mp);
3581 		BUF_UNLOCK(bp);
3582 	}
3583 	BQ_LOCK(bq);
3584 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3585 	BQ_UNLOCK(bq);
3586 	free(sentinel, M_TEMP);
3587 	return (flushed);
3588 }
3589 
3590 /*
3591  * Check to see if a block is currently memory resident.
3592  */
3593 struct buf *
3594 incore(struct bufobj *bo, daddr_t blkno)
3595 {
3596 	struct buf *bp;
3597 
3598 	BO_RLOCK(bo);
3599 	bp = gbincore(bo, blkno);
3600 	BO_RUNLOCK(bo);
3601 	return (bp);
3602 }
3603 
3604 /*
3605  * Returns true if no I/O is needed to access the
3606  * associated VM object.  This is like incore except
3607  * it also hunts around in the VM system for the data.
3608  */
3609 
3610 static int
3611 inmem(struct vnode * vp, daddr_t blkno)
3612 {
3613 	vm_object_t obj;
3614 	vm_offset_t toff, tinc, size;
3615 	vm_page_t m;
3616 	vm_ooffset_t off;
3617 
3618 	ASSERT_VOP_LOCKED(vp, "inmem");
3619 
3620 	if (incore(&vp->v_bufobj, blkno))
3621 		return 1;
3622 	if (vp->v_mount == NULL)
3623 		return 0;
3624 	obj = vp->v_object;
3625 	if (obj == NULL)
3626 		return (0);
3627 
3628 	size = PAGE_SIZE;
3629 	if (size > vp->v_mount->mnt_stat.f_iosize)
3630 		size = vp->v_mount->mnt_stat.f_iosize;
3631 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3632 
3633 	VM_OBJECT_RLOCK(obj);
3634 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3635 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3636 		if (!m)
3637 			goto notinmem;
3638 		tinc = size;
3639 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3640 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3641 		if (vm_page_is_valid(m,
3642 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3643 			goto notinmem;
3644 	}
3645 	VM_OBJECT_RUNLOCK(obj);
3646 	return 1;
3647 
3648 notinmem:
3649 	VM_OBJECT_RUNLOCK(obj);
3650 	return (0);
3651 }
3652 
3653 /*
3654  * Set the dirty range for a buffer based on the status of the dirty
3655  * bits in the pages comprising the buffer.  The range is limited
3656  * to the size of the buffer.
3657  *
3658  * Tell the VM system that the pages associated with this buffer
3659  * are clean.  This is used for delayed writes where the data is
3660  * going to go to disk eventually without additional VM intevention.
3661  *
3662  * Note that while we only really need to clean through to b_bcount, we
3663  * just go ahead and clean through to b_bufsize.
3664  */
3665 static void
3666 vfs_clean_pages_dirty_buf(struct buf *bp)
3667 {
3668 	vm_ooffset_t foff, noff, eoff;
3669 	vm_page_t m;
3670 	int i;
3671 
3672 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3673 		return;
3674 
3675 	foff = bp->b_offset;
3676 	KASSERT(bp->b_offset != NOOFFSET,
3677 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3678 
3679 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3680 	vfs_drain_busy_pages(bp);
3681 	vfs_setdirty_locked_object(bp);
3682 	for (i = 0; i < bp->b_npages; i++) {
3683 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3684 		eoff = noff;
3685 		if (eoff > bp->b_offset + bp->b_bufsize)
3686 			eoff = bp->b_offset + bp->b_bufsize;
3687 		m = bp->b_pages[i];
3688 		vfs_page_set_validclean(bp, foff, m);
3689 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3690 		foff = noff;
3691 	}
3692 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3693 }
3694 
3695 static void
3696 vfs_setdirty_locked_object(struct buf *bp)
3697 {
3698 	vm_object_t object;
3699 	int i;
3700 
3701 	object = bp->b_bufobj->bo_object;
3702 	VM_OBJECT_ASSERT_WLOCKED(object);
3703 
3704 	/*
3705 	 * We qualify the scan for modified pages on whether the
3706 	 * object has been flushed yet.
3707 	 */
3708 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3709 		vm_offset_t boffset;
3710 		vm_offset_t eoffset;
3711 
3712 		/*
3713 		 * test the pages to see if they have been modified directly
3714 		 * by users through the VM system.
3715 		 */
3716 		for (i = 0; i < bp->b_npages; i++)
3717 			vm_page_test_dirty(bp->b_pages[i]);
3718 
3719 		/*
3720 		 * Calculate the encompassing dirty range, boffset and eoffset,
3721 		 * (eoffset - boffset) bytes.
3722 		 */
3723 
3724 		for (i = 0; i < bp->b_npages; i++) {
3725 			if (bp->b_pages[i]->dirty)
3726 				break;
3727 		}
3728 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3729 
3730 		for (i = bp->b_npages - 1; i >= 0; --i) {
3731 			if (bp->b_pages[i]->dirty) {
3732 				break;
3733 			}
3734 		}
3735 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3736 
3737 		/*
3738 		 * Fit it to the buffer.
3739 		 */
3740 
3741 		if (eoffset > bp->b_bcount)
3742 			eoffset = bp->b_bcount;
3743 
3744 		/*
3745 		 * If we have a good dirty range, merge with the existing
3746 		 * dirty range.
3747 		 */
3748 
3749 		if (boffset < eoffset) {
3750 			if (bp->b_dirtyoff > boffset)
3751 				bp->b_dirtyoff = boffset;
3752 			if (bp->b_dirtyend < eoffset)
3753 				bp->b_dirtyend = eoffset;
3754 		}
3755 	}
3756 }
3757 
3758 /*
3759  * Allocate the KVA mapping for an existing buffer.
3760  * If an unmapped buffer is provided but a mapped buffer is requested, take
3761  * also care to properly setup mappings between pages and KVA.
3762  */
3763 static void
3764 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3765 {
3766 	int bsize, maxsize, need_mapping, need_kva;
3767 	off_t offset;
3768 
3769 	need_mapping = bp->b_data == unmapped_buf &&
3770 	    (gbflags & GB_UNMAPPED) == 0;
3771 	need_kva = bp->b_kvabase == unmapped_buf &&
3772 	    bp->b_data == unmapped_buf &&
3773 	    (gbflags & GB_KVAALLOC) != 0;
3774 	if (!need_mapping && !need_kva)
3775 		return;
3776 
3777 	BUF_CHECK_UNMAPPED(bp);
3778 
3779 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3780 		/*
3781 		 * Buffer is not mapped, but the KVA was already
3782 		 * reserved at the time of the instantiation.  Use the
3783 		 * allocated space.
3784 		 */
3785 		goto has_addr;
3786 	}
3787 
3788 	/*
3789 	 * Calculate the amount of the address space we would reserve
3790 	 * if the buffer was mapped.
3791 	 */
3792 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3793 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3794 	offset = blkno * bsize;
3795 	maxsize = size + (offset & PAGE_MASK);
3796 	maxsize = imax(maxsize, bsize);
3797 
3798 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3799 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3800 			/*
3801 			 * XXXKIB: defragmentation cannot
3802 			 * succeed, not sure what else to do.
3803 			 */
3804 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3805 		}
3806 		counter_u64_add(mappingrestarts, 1);
3807 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3808 	}
3809 has_addr:
3810 	if (need_mapping) {
3811 		/* b_offset is handled by bpmap_qenter. */
3812 		bp->b_data = bp->b_kvabase;
3813 		BUF_CHECK_MAPPED(bp);
3814 		bpmap_qenter(bp);
3815 	}
3816 }
3817 
3818 struct buf *
3819 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3820     int flags)
3821 {
3822 	struct buf *bp;
3823 	int error;
3824 
3825 	error = getblkx(vp, blkno, size, slpflag, slptimeo, flags, &bp);
3826 	if (error != 0)
3827 		return (NULL);
3828 	return (bp);
3829 }
3830 
3831 /*
3832  *	getblkx:
3833  *
3834  *	Get a block given a specified block and offset into a file/device.
3835  *	The buffers B_DONE bit will be cleared on return, making it almost
3836  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3837  *	return.  The caller should clear B_INVAL prior to initiating a
3838  *	READ.
3839  *
3840  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3841  *	an existing buffer.
3842  *
3843  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3844  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3845  *	and then cleared based on the backing VM.  If the previous buffer is
3846  *	non-0-sized but invalid, B_CACHE will be cleared.
3847  *
3848  *	If getblk() must create a new buffer, the new buffer is returned with
3849  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3850  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3851  *	backing VM.
3852  *
3853  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3854  *	B_CACHE bit is clear.
3855  *
3856  *	What this means, basically, is that the caller should use B_CACHE to
3857  *	determine whether the buffer is fully valid or not and should clear
3858  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3859  *	the buffer by loading its data area with something, the caller needs
3860  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3861  *	the caller should set B_CACHE ( as an optimization ), else the caller
3862  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3863  *	a write attempt or if it was a successful read.  If the caller
3864  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3865  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3866  */
3867 int
3868 getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3869     int flags, struct buf **bpp)
3870 {
3871 	struct buf *bp;
3872 	struct bufobj *bo;
3873 	daddr_t d_blkno;
3874 	int bsize, error, maxsize, vmio;
3875 	off_t offset;
3876 
3877 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3878 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3879 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3880 	ASSERT_VOP_LOCKED(vp, "getblk");
3881 	if (size > maxbcachebuf)
3882 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3883 		    maxbcachebuf);
3884 	if (!unmapped_buf_allowed)
3885 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3886 
3887 	bo = &vp->v_bufobj;
3888 	d_blkno = blkno;
3889 loop:
3890 	BO_RLOCK(bo);
3891 	bp = gbincore(bo, blkno);
3892 	if (bp != NULL) {
3893 		int lockflags;
3894 		/*
3895 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3896 		 * it must be on a queue.
3897 		 */
3898 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3899 
3900 		if ((flags & GB_LOCK_NOWAIT) != 0)
3901 			lockflags |= LK_NOWAIT;
3902 
3903 		error = BUF_TIMELOCK(bp, lockflags,
3904 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3905 
3906 		/*
3907 		 * If we slept and got the lock we have to restart in case
3908 		 * the buffer changed identities.
3909 		 */
3910 		if (error == ENOLCK)
3911 			goto loop;
3912 		/* We timed out or were interrupted. */
3913 		else if (error != 0)
3914 			return (error);
3915 		/* If recursed, assume caller knows the rules. */
3916 		else if (BUF_LOCKRECURSED(bp))
3917 			goto end;
3918 
3919 		/*
3920 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3921 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3922 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3923 		 * backing VM cache.
3924 		 */
3925 		if (bp->b_flags & B_INVAL)
3926 			bp->b_flags &= ~B_CACHE;
3927 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3928 			bp->b_flags |= B_CACHE;
3929 		if (bp->b_flags & B_MANAGED)
3930 			MPASS(bp->b_qindex == QUEUE_NONE);
3931 		else
3932 			bremfree(bp);
3933 
3934 		/*
3935 		 * check for size inconsistencies for non-VMIO case.
3936 		 */
3937 		if (bp->b_bcount != size) {
3938 			if ((bp->b_flags & B_VMIO) == 0 ||
3939 			    (size > bp->b_kvasize)) {
3940 				if (bp->b_flags & B_DELWRI) {
3941 					bp->b_flags |= B_NOCACHE;
3942 					bwrite(bp);
3943 				} else {
3944 					if (LIST_EMPTY(&bp->b_dep)) {
3945 						bp->b_flags |= B_RELBUF;
3946 						brelse(bp);
3947 					} else {
3948 						bp->b_flags |= B_NOCACHE;
3949 						bwrite(bp);
3950 					}
3951 				}
3952 				goto loop;
3953 			}
3954 		}
3955 
3956 		/*
3957 		 * Handle the case of unmapped buffer which should
3958 		 * become mapped, or the buffer for which KVA
3959 		 * reservation is requested.
3960 		 */
3961 		bp_unmapped_get_kva(bp, blkno, size, flags);
3962 
3963 		/*
3964 		 * If the size is inconsistent in the VMIO case, we can resize
3965 		 * the buffer.  This might lead to B_CACHE getting set or
3966 		 * cleared.  If the size has not changed, B_CACHE remains
3967 		 * unchanged from its previous state.
3968 		 */
3969 		allocbuf(bp, size);
3970 
3971 		KASSERT(bp->b_offset != NOOFFSET,
3972 		    ("getblk: no buffer offset"));
3973 
3974 		/*
3975 		 * A buffer with B_DELWRI set and B_CACHE clear must
3976 		 * be committed before we can return the buffer in
3977 		 * order to prevent the caller from issuing a read
3978 		 * ( due to B_CACHE not being set ) and overwriting
3979 		 * it.
3980 		 *
3981 		 * Most callers, including NFS and FFS, need this to
3982 		 * operate properly either because they assume they
3983 		 * can issue a read if B_CACHE is not set, or because
3984 		 * ( for example ) an uncached B_DELWRI might loop due
3985 		 * to softupdates re-dirtying the buffer.  In the latter
3986 		 * case, B_CACHE is set after the first write completes,
3987 		 * preventing further loops.
3988 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3989 		 * above while extending the buffer, we cannot allow the
3990 		 * buffer to remain with B_CACHE set after the write
3991 		 * completes or it will represent a corrupt state.  To
3992 		 * deal with this we set B_NOCACHE to scrap the buffer
3993 		 * after the write.
3994 		 *
3995 		 * We might be able to do something fancy, like setting
3996 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3997 		 * so the below call doesn't set B_CACHE, but that gets real
3998 		 * confusing.  This is much easier.
3999 		 */
4000 
4001 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4002 			bp->b_flags |= B_NOCACHE;
4003 			bwrite(bp);
4004 			goto loop;
4005 		}
4006 		bp->b_flags &= ~B_DONE;
4007 	} else {
4008 		/*
4009 		 * Buffer is not in-core, create new buffer.  The buffer
4010 		 * returned by getnewbuf() is locked.  Note that the returned
4011 		 * buffer is also considered valid (not marked B_INVAL).
4012 		 */
4013 		BO_RUNLOCK(bo);
4014 		/*
4015 		 * If the user does not want us to create the buffer, bail out
4016 		 * here.
4017 		 */
4018 		if (flags & GB_NOCREAT)
4019 			return (EEXIST);
4020 
4021 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
4022 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4023 		offset = blkno * bsize;
4024 		vmio = vp->v_object != NULL;
4025 		if (vmio) {
4026 			maxsize = size + (offset & PAGE_MASK);
4027 		} else {
4028 			maxsize = size;
4029 			/* Do not allow non-VMIO notmapped buffers. */
4030 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4031 		}
4032 		maxsize = imax(maxsize, bsize);
4033 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4034 		    !vn_isdisk(vp, NULL)) {
4035 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4036 			KASSERT(error != EOPNOTSUPP,
4037 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4038 			    vp));
4039 			if (error != 0)
4040 				return (error);
4041 			if (d_blkno == -1)
4042 				return (EJUSTRETURN);
4043 		}
4044 
4045 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4046 		if (bp == NULL) {
4047 			if (slpflag || slptimeo)
4048 				return (ETIMEDOUT);
4049 			/*
4050 			 * XXX This is here until the sleep path is diagnosed
4051 			 * enough to work under very low memory conditions.
4052 			 *
4053 			 * There's an issue on low memory, 4BSD+non-preempt
4054 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4055 			 * exhaustion occurs without sleeping for buffer
4056 			 * reclaimation.  This just sticks in a loop and
4057 			 * constantly attempts to allocate a buffer, which
4058 			 * hits exhaustion and tries to wakeup bufdaemon.
4059 			 * This never happens because we never yield.
4060 			 *
4061 			 * The real solution is to identify and fix these cases
4062 			 * so we aren't effectively busy-waiting in a loop
4063 			 * until the reclaimation path has cycles to run.
4064 			 */
4065 			kern_yield(PRI_USER);
4066 			goto loop;
4067 		}
4068 
4069 		/*
4070 		 * This code is used to make sure that a buffer is not
4071 		 * created while the getnewbuf routine is blocked.
4072 		 * This can be a problem whether the vnode is locked or not.
4073 		 * If the buffer is created out from under us, we have to
4074 		 * throw away the one we just created.
4075 		 *
4076 		 * Note: this must occur before we associate the buffer
4077 		 * with the vp especially considering limitations in
4078 		 * the splay tree implementation when dealing with duplicate
4079 		 * lblkno's.
4080 		 */
4081 		BO_LOCK(bo);
4082 		if (gbincore(bo, blkno)) {
4083 			BO_UNLOCK(bo);
4084 			bp->b_flags |= B_INVAL;
4085 			bufspace_release(bufdomain(bp), maxsize);
4086 			brelse(bp);
4087 			goto loop;
4088 		}
4089 
4090 		/*
4091 		 * Insert the buffer into the hash, so that it can
4092 		 * be found by incore.
4093 		 */
4094 		bp->b_lblkno = blkno;
4095 		bp->b_blkno = d_blkno;
4096 		bp->b_offset = offset;
4097 		bgetvp(vp, bp);
4098 		BO_UNLOCK(bo);
4099 
4100 		/*
4101 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4102 		 * buffer size starts out as 0, B_CACHE will be set by
4103 		 * allocbuf() for the VMIO case prior to it testing the
4104 		 * backing store for validity.
4105 		 */
4106 
4107 		if (vmio) {
4108 			bp->b_flags |= B_VMIO;
4109 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4110 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4111 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4112 		} else {
4113 			bp->b_flags &= ~B_VMIO;
4114 			KASSERT(bp->b_bufobj->bo_object == NULL,
4115 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4116 			    bp, bp->b_bufobj->bo_object));
4117 			BUF_CHECK_MAPPED(bp);
4118 		}
4119 
4120 		allocbuf(bp, size);
4121 		bufspace_release(bufdomain(bp), maxsize);
4122 		bp->b_flags &= ~B_DONE;
4123 	}
4124 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4125 	BUF_ASSERT_HELD(bp);
4126 end:
4127 	buf_track(bp, __func__);
4128 	KASSERT(bp->b_bufobj == bo,
4129 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4130 	*bpp = bp;
4131 	return (0);
4132 }
4133 
4134 /*
4135  * Get an empty, disassociated buffer of given size.  The buffer is initially
4136  * set to B_INVAL.
4137  */
4138 struct buf *
4139 geteblk(int size, int flags)
4140 {
4141 	struct buf *bp;
4142 	int maxsize;
4143 
4144 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4145 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4146 		if ((flags & GB_NOWAIT_BD) &&
4147 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4148 			return (NULL);
4149 	}
4150 	allocbuf(bp, size);
4151 	bufspace_release(bufdomain(bp), maxsize);
4152 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4153 	BUF_ASSERT_HELD(bp);
4154 	return (bp);
4155 }
4156 
4157 /*
4158  * Truncate the backing store for a non-vmio buffer.
4159  */
4160 static void
4161 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4162 {
4163 
4164 	if (bp->b_flags & B_MALLOC) {
4165 		/*
4166 		 * malloced buffers are not shrunk
4167 		 */
4168 		if (newbsize == 0) {
4169 			bufmallocadjust(bp, 0);
4170 			free(bp->b_data, M_BIOBUF);
4171 			bp->b_data = bp->b_kvabase;
4172 			bp->b_flags &= ~B_MALLOC;
4173 		}
4174 		return;
4175 	}
4176 	vm_hold_free_pages(bp, newbsize);
4177 	bufspace_adjust(bp, newbsize);
4178 }
4179 
4180 /*
4181  * Extend the backing for a non-VMIO buffer.
4182  */
4183 static void
4184 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4185 {
4186 	caddr_t origbuf;
4187 	int origbufsize;
4188 
4189 	/*
4190 	 * We only use malloced memory on the first allocation.
4191 	 * and revert to page-allocated memory when the buffer
4192 	 * grows.
4193 	 *
4194 	 * There is a potential smp race here that could lead
4195 	 * to bufmallocspace slightly passing the max.  It
4196 	 * is probably extremely rare and not worth worrying
4197 	 * over.
4198 	 */
4199 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4200 	    bufmallocspace < maxbufmallocspace) {
4201 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4202 		bp->b_flags |= B_MALLOC;
4203 		bufmallocadjust(bp, newbsize);
4204 		return;
4205 	}
4206 
4207 	/*
4208 	 * If the buffer is growing on its other-than-first
4209 	 * allocation then we revert to the page-allocation
4210 	 * scheme.
4211 	 */
4212 	origbuf = NULL;
4213 	origbufsize = 0;
4214 	if (bp->b_flags & B_MALLOC) {
4215 		origbuf = bp->b_data;
4216 		origbufsize = bp->b_bufsize;
4217 		bp->b_data = bp->b_kvabase;
4218 		bufmallocadjust(bp, 0);
4219 		bp->b_flags &= ~B_MALLOC;
4220 		newbsize = round_page(newbsize);
4221 	}
4222 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4223 	    (vm_offset_t) bp->b_data + newbsize);
4224 	if (origbuf != NULL) {
4225 		bcopy(origbuf, bp->b_data, origbufsize);
4226 		free(origbuf, M_BIOBUF);
4227 	}
4228 	bufspace_adjust(bp, newbsize);
4229 }
4230 
4231 /*
4232  * This code constitutes the buffer memory from either anonymous system
4233  * memory (in the case of non-VMIO operations) or from an associated
4234  * VM object (in the case of VMIO operations).  This code is able to
4235  * resize a buffer up or down.
4236  *
4237  * Note that this code is tricky, and has many complications to resolve
4238  * deadlock or inconsistent data situations.  Tread lightly!!!
4239  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4240  * the caller.  Calling this code willy nilly can result in the loss of data.
4241  *
4242  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4243  * B_CACHE for the non-VMIO case.
4244  */
4245 int
4246 allocbuf(struct buf *bp, int size)
4247 {
4248 	int newbsize;
4249 
4250 	BUF_ASSERT_HELD(bp);
4251 
4252 	if (bp->b_bcount == size)
4253 		return (1);
4254 
4255 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4256 		panic("allocbuf: buffer too small");
4257 
4258 	newbsize = roundup2(size, DEV_BSIZE);
4259 	if ((bp->b_flags & B_VMIO) == 0) {
4260 		if ((bp->b_flags & B_MALLOC) == 0)
4261 			newbsize = round_page(newbsize);
4262 		/*
4263 		 * Just get anonymous memory from the kernel.  Don't
4264 		 * mess with B_CACHE.
4265 		 */
4266 		if (newbsize < bp->b_bufsize)
4267 			vfs_nonvmio_truncate(bp, newbsize);
4268 		else if (newbsize > bp->b_bufsize)
4269 			vfs_nonvmio_extend(bp, newbsize);
4270 	} else {
4271 		int desiredpages;
4272 
4273 		desiredpages = (size == 0) ? 0 :
4274 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4275 
4276 		if (bp->b_flags & B_MALLOC)
4277 			panic("allocbuf: VMIO buffer can't be malloced");
4278 		/*
4279 		 * Set B_CACHE initially if buffer is 0 length or will become
4280 		 * 0-length.
4281 		 */
4282 		if (size == 0 || bp->b_bufsize == 0)
4283 			bp->b_flags |= B_CACHE;
4284 
4285 		if (newbsize < bp->b_bufsize)
4286 			vfs_vmio_truncate(bp, desiredpages);
4287 		/* XXX This looks as if it should be newbsize > b_bufsize */
4288 		else if (size > bp->b_bcount)
4289 			vfs_vmio_extend(bp, desiredpages, size);
4290 		bufspace_adjust(bp, newbsize);
4291 	}
4292 	bp->b_bcount = size;		/* requested buffer size. */
4293 	return (1);
4294 }
4295 
4296 extern int inflight_transient_maps;
4297 
4298 static struct bio_queue nondump_bios;
4299 
4300 void
4301 biodone(struct bio *bp)
4302 {
4303 	struct mtx *mtxp;
4304 	void (*done)(struct bio *);
4305 	vm_offset_t start, end;
4306 
4307 	biotrack(bp, __func__);
4308 
4309 	/*
4310 	 * Avoid completing I/O when dumping after a panic since that may
4311 	 * result in a deadlock in the filesystem or pager code.  Note that
4312 	 * this doesn't affect dumps that were started manually since we aim
4313 	 * to keep the system usable after it has been resumed.
4314 	 */
4315 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4316 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4317 		return;
4318 	}
4319 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4320 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4321 		bp->bio_flags |= BIO_UNMAPPED;
4322 		start = trunc_page((vm_offset_t)bp->bio_data);
4323 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4324 		bp->bio_data = unmapped_buf;
4325 		pmap_qremove(start, atop(end - start));
4326 		vmem_free(transient_arena, start, end - start);
4327 		atomic_add_int(&inflight_transient_maps, -1);
4328 	}
4329 	done = bp->bio_done;
4330 	if (done == NULL) {
4331 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4332 		mtx_lock(mtxp);
4333 		bp->bio_flags |= BIO_DONE;
4334 		wakeup(bp);
4335 		mtx_unlock(mtxp);
4336 	} else
4337 		done(bp);
4338 }
4339 
4340 /*
4341  * Wait for a BIO to finish.
4342  */
4343 int
4344 biowait(struct bio *bp, const char *wchan)
4345 {
4346 	struct mtx *mtxp;
4347 
4348 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4349 	mtx_lock(mtxp);
4350 	while ((bp->bio_flags & BIO_DONE) == 0)
4351 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4352 	mtx_unlock(mtxp);
4353 	if (bp->bio_error != 0)
4354 		return (bp->bio_error);
4355 	if (!(bp->bio_flags & BIO_ERROR))
4356 		return (0);
4357 	return (EIO);
4358 }
4359 
4360 void
4361 biofinish(struct bio *bp, struct devstat *stat, int error)
4362 {
4363 
4364 	if (error) {
4365 		bp->bio_error = error;
4366 		bp->bio_flags |= BIO_ERROR;
4367 	}
4368 	if (stat != NULL)
4369 		devstat_end_transaction_bio(stat, bp);
4370 	biodone(bp);
4371 }
4372 
4373 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4374 void
4375 biotrack_buf(struct bio *bp, const char *location)
4376 {
4377 
4378 	buf_track(bp->bio_track_bp, location);
4379 }
4380 #endif
4381 
4382 /*
4383  *	bufwait:
4384  *
4385  *	Wait for buffer I/O completion, returning error status.  The buffer
4386  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4387  *	error and cleared.
4388  */
4389 int
4390 bufwait(struct buf *bp)
4391 {
4392 	if (bp->b_iocmd == BIO_READ)
4393 		bwait(bp, PRIBIO, "biord");
4394 	else
4395 		bwait(bp, PRIBIO, "biowr");
4396 	if (bp->b_flags & B_EINTR) {
4397 		bp->b_flags &= ~B_EINTR;
4398 		return (EINTR);
4399 	}
4400 	if (bp->b_ioflags & BIO_ERROR) {
4401 		return (bp->b_error ? bp->b_error : EIO);
4402 	} else {
4403 		return (0);
4404 	}
4405 }
4406 
4407 /*
4408  *	bufdone:
4409  *
4410  *	Finish I/O on a buffer, optionally calling a completion function.
4411  *	This is usually called from an interrupt so process blocking is
4412  *	not allowed.
4413  *
4414  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4415  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4416  *	assuming B_INVAL is clear.
4417  *
4418  *	For the VMIO case, we set B_CACHE if the op was a read and no
4419  *	read error occurred, or if the op was a write.  B_CACHE is never
4420  *	set if the buffer is invalid or otherwise uncacheable.
4421  *
4422  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4423  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4424  *	in the biodone routine.
4425  */
4426 void
4427 bufdone(struct buf *bp)
4428 {
4429 	struct bufobj *dropobj;
4430 	void    (*biodone)(struct buf *);
4431 
4432 	buf_track(bp, __func__);
4433 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4434 	dropobj = NULL;
4435 
4436 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4437 	BUF_ASSERT_HELD(bp);
4438 
4439 	runningbufwakeup(bp);
4440 	if (bp->b_iocmd == BIO_WRITE)
4441 		dropobj = bp->b_bufobj;
4442 	/* call optional completion function if requested */
4443 	if (bp->b_iodone != NULL) {
4444 		biodone = bp->b_iodone;
4445 		bp->b_iodone = NULL;
4446 		(*biodone) (bp);
4447 		if (dropobj)
4448 			bufobj_wdrop(dropobj);
4449 		return;
4450 	}
4451 	if (bp->b_flags & B_VMIO) {
4452 		/*
4453 		 * Set B_CACHE if the op was a normal read and no error
4454 		 * occurred.  B_CACHE is set for writes in the b*write()
4455 		 * routines.
4456 		 */
4457 		if (bp->b_iocmd == BIO_READ &&
4458 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4459 		    !(bp->b_ioflags & BIO_ERROR))
4460 			bp->b_flags |= B_CACHE;
4461 		vfs_vmio_iodone(bp);
4462 	}
4463 	if (!LIST_EMPTY(&bp->b_dep))
4464 		buf_complete(bp);
4465 	if ((bp->b_flags & B_CKHASH) != 0) {
4466 		KASSERT(bp->b_iocmd == BIO_READ,
4467 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4468 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4469 		(*bp->b_ckhashcalc)(bp);
4470 	}
4471 	/*
4472 	 * For asynchronous completions, release the buffer now. The brelse
4473 	 * will do a wakeup there if necessary - so no need to do a wakeup
4474 	 * here in the async case. The sync case always needs to do a wakeup.
4475 	 */
4476 	if (bp->b_flags & B_ASYNC) {
4477 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4478 		    (bp->b_ioflags & BIO_ERROR))
4479 			brelse(bp);
4480 		else
4481 			bqrelse(bp);
4482 	} else
4483 		bdone(bp);
4484 	if (dropobj)
4485 		bufobj_wdrop(dropobj);
4486 }
4487 
4488 /*
4489  * This routine is called in lieu of iodone in the case of
4490  * incomplete I/O.  This keeps the busy status for pages
4491  * consistent.
4492  */
4493 void
4494 vfs_unbusy_pages(struct buf *bp)
4495 {
4496 	int i;
4497 	vm_object_t obj;
4498 	vm_page_t m;
4499 
4500 	runningbufwakeup(bp);
4501 	if (!(bp->b_flags & B_VMIO))
4502 		return;
4503 
4504 	obj = bp->b_bufobj->bo_object;
4505 	VM_OBJECT_WLOCK(obj);
4506 	for (i = 0; i < bp->b_npages; i++) {
4507 		m = bp->b_pages[i];
4508 		if (m == bogus_page) {
4509 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4510 			if (!m)
4511 				panic("vfs_unbusy_pages: page missing\n");
4512 			bp->b_pages[i] = m;
4513 			if (buf_mapped(bp)) {
4514 				BUF_CHECK_MAPPED(bp);
4515 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4516 				    bp->b_pages, bp->b_npages);
4517 			} else
4518 				BUF_CHECK_UNMAPPED(bp);
4519 		}
4520 		vm_page_sunbusy(m);
4521 	}
4522 	vm_object_pip_wakeupn(obj, bp->b_npages);
4523 	VM_OBJECT_WUNLOCK(obj);
4524 }
4525 
4526 /*
4527  * vfs_page_set_valid:
4528  *
4529  *	Set the valid bits in a page based on the supplied offset.   The
4530  *	range is restricted to the buffer's size.
4531  *
4532  *	This routine is typically called after a read completes.
4533  */
4534 static void
4535 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4536 {
4537 	vm_ooffset_t eoff;
4538 
4539 	/*
4540 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4541 	 * page boundary and eoff is not greater than the end of the buffer.
4542 	 * The end of the buffer, in this case, is our file EOF, not the
4543 	 * allocation size of the buffer.
4544 	 */
4545 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4546 	if (eoff > bp->b_offset + bp->b_bcount)
4547 		eoff = bp->b_offset + bp->b_bcount;
4548 
4549 	/*
4550 	 * Set valid range.  This is typically the entire buffer and thus the
4551 	 * entire page.
4552 	 */
4553 	if (eoff > off)
4554 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4555 }
4556 
4557 /*
4558  * vfs_page_set_validclean:
4559  *
4560  *	Set the valid bits and clear the dirty bits in a page based on the
4561  *	supplied offset.   The range is restricted to the buffer's size.
4562  */
4563 static void
4564 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4565 {
4566 	vm_ooffset_t soff, eoff;
4567 
4568 	/*
4569 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4570 	 * page boundary or cross the end of the buffer.  The end of the
4571 	 * buffer, in this case, is our file EOF, not the allocation size
4572 	 * of the buffer.
4573 	 */
4574 	soff = off;
4575 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4576 	if (eoff > bp->b_offset + bp->b_bcount)
4577 		eoff = bp->b_offset + bp->b_bcount;
4578 
4579 	/*
4580 	 * Set valid range.  This is typically the entire buffer and thus the
4581 	 * entire page.
4582 	 */
4583 	if (eoff > soff) {
4584 		vm_page_set_validclean(
4585 		    m,
4586 		   (vm_offset_t) (soff & PAGE_MASK),
4587 		   (vm_offset_t) (eoff - soff)
4588 		);
4589 	}
4590 }
4591 
4592 /*
4593  * Ensure that all buffer pages are not exclusive busied.  If any page is
4594  * exclusive busy, drain it.
4595  */
4596 void
4597 vfs_drain_busy_pages(struct buf *bp)
4598 {
4599 	vm_page_t m;
4600 	int i, last_busied;
4601 
4602 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4603 	last_busied = 0;
4604 	for (i = 0; i < bp->b_npages; i++) {
4605 		m = bp->b_pages[i];
4606 		if (vm_page_xbusied(m)) {
4607 			for (; last_busied < i; last_busied++)
4608 				vm_page_sbusy(bp->b_pages[last_busied]);
4609 			while (vm_page_xbusied(m)) {
4610 				vm_page_lock(m);
4611 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4612 				vm_page_busy_sleep(m, "vbpage", true);
4613 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4614 			}
4615 		}
4616 	}
4617 	for (i = 0; i < last_busied; i++)
4618 		vm_page_sunbusy(bp->b_pages[i]);
4619 }
4620 
4621 /*
4622  * This routine is called before a device strategy routine.
4623  * It is used to tell the VM system that paging I/O is in
4624  * progress, and treat the pages associated with the buffer
4625  * almost as being exclusive busy.  Also the object paging_in_progress
4626  * flag is handled to make sure that the object doesn't become
4627  * inconsistent.
4628  *
4629  * Since I/O has not been initiated yet, certain buffer flags
4630  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4631  * and should be ignored.
4632  */
4633 void
4634 vfs_busy_pages(struct buf *bp, int clear_modify)
4635 {
4636 	vm_object_t obj;
4637 	vm_ooffset_t foff;
4638 	vm_page_t m;
4639 	int i;
4640 	bool bogus;
4641 
4642 	if (!(bp->b_flags & B_VMIO))
4643 		return;
4644 
4645 	obj = bp->b_bufobj->bo_object;
4646 	foff = bp->b_offset;
4647 	KASSERT(bp->b_offset != NOOFFSET,
4648 	    ("vfs_busy_pages: no buffer offset"));
4649 	VM_OBJECT_WLOCK(obj);
4650 	vfs_drain_busy_pages(bp);
4651 	if (bp->b_bufsize != 0)
4652 		vfs_setdirty_locked_object(bp);
4653 	bogus = false;
4654 	for (i = 0; i < bp->b_npages; i++) {
4655 		m = bp->b_pages[i];
4656 
4657 		if ((bp->b_flags & B_CLUSTER) == 0) {
4658 			vm_object_pip_add(obj, 1);
4659 			vm_page_sbusy(m);
4660 		}
4661 		/*
4662 		 * When readying a buffer for a read ( i.e
4663 		 * clear_modify == 0 ), it is important to do
4664 		 * bogus_page replacement for valid pages in
4665 		 * partially instantiated buffers.  Partially
4666 		 * instantiated buffers can, in turn, occur when
4667 		 * reconstituting a buffer from its VM backing store
4668 		 * base.  We only have to do this if B_CACHE is
4669 		 * clear ( which causes the I/O to occur in the
4670 		 * first place ).  The replacement prevents the read
4671 		 * I/O from overwriting potentially dirty VM-backed
4672 		 * pages.  XXX bogus page replacement is, uh, bogus.
4673 		 * It may not work properly with small-block devices.
4674 		 * We need to find a better way.
4675 		 */
4676 		if (clear_modify) {
4677 			pmap_remove_write(m);
4678 			vfs_page_set_validclean(bp, foff, m);
4679 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4680 		    (bp->b_flags & B_CACHE) == 0) {
4681 			bp->b_pages[i] = bogus_page;
4682 			bogus = true;
4683 		}
4684 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4685 	}
4686 	VM_OBJECT_WUNLOCK(obj);
4687 	if (bogus && buf_mapped(bp)) {
4688 		BUF_CHECK_MAPPED(bp);
4689 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4690 		    bp->b_pages, bp->b_npages);
4691 	}
4692 }
4693 
4694 /*
4695  *	vfs_bio_set_valid:
4696  *
4697  *	Set the range within the buffer to valid.  The range is
4698  *	relative to the beginning of the buffer, b_offset.  Note that
4699  *	b_offset itself may be offset from the beginning of the first
4700  *	page.
4701  */
4702 void
4703 vfs_bio_set_valid(struct buf *bp, int base, int size)
4704 {
4705 	int i, n;
4706 	vm_page_t m;
4707 
4708 	if (!(bp->b_flags & B_VMIO))
4709 		return;
4710 
4711 	/*
4712 	 * Fixup base to be relative to beginning of first page.
4713 	 * Set initial n to be the maximum number of bytes in the
4714 	 * first page that can be validated.
4715 	 */
4716 	base += (bp->b_offset & PAGE_MASK);
4717 	n = PAGE_SIZE - (base & PAGE_MASK);
4718 
4719 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4720 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4721 		m = bp->b_pages[i];
4722 		if (n > size)
4723 			n = size;
4724 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4725 		base += n;
4726 		size -= n;
4727 		n = PAGE_SIZE;
4728 	}
4729 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4730 }
4731 
4732 /*
4733  *	vfs_bio_clrbuf:
4734  *
4735  *	If the specified buffer is a non-VMIO buffer, clear the entire
4736  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4737  *	validate only the previously invalid portions of the buffer.
4738  *	This routine essentially fakes an I/O, so we need to clear
4739  *	BIO_ERROR and B_INVAL.
4740  *
4741  *	Note that while we only theoretically need to clear through b_bcount,
4742  *	we go ahead and clear through b_bufsize.
4743  */
4744 void
4745 vfs_bio_clrbuf(struct buf *bp)
4746 {
4747 	int i, j, mask, sa, ea, slide;
4748 
4749 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4750 		clrbuf(bp);
4751 		return;
4752 	}
4753 	bp->b_flags &= ~B_INVAL;
4754 	bp->b_ioflags &= ~BIO_ERROR;
4755 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4756 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4757 	    (bp->b_offset & PAGE_MASK) == 0) {
4758 		if (bp->b_pages[0] == bogus_page)
4759 			goto unlock;
4760 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4761 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4762 		if ((bp->b_pages[0]->valid & mask) == mask)
4763 			goto unlock;
4764 		if ((bp->b_pages[0]->valid & mask) == 0) {
4765 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4766 			bp->b_pages[0]->valid |= mask;
4767 			goto unlock;
4768 		}
4769 	}
4770 	sa = bp->b_offset & PAGE_MASK;
4771 	slide = 0;
4772 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4773 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4774 		ea = slide & PAGE_MASK;
4775 		if (ea == 0)
4776 			ea = PAGE_SIZE;
4777 		if (bp->b_pages[i] == bogus_page)
4778 			continue;
4779 		j = sa / DEV_BSIZE;
4780 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4781 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4782 		if ((bp->b_pages[i]->valid & mask) == mask)
4783 			continue;
4784 		if ((bp->b_pages[i]->valid & mask) == 0)
4785 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4786 		else {
4787 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4788 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4789 					pmap_zero_page_area(bp->b_pages[i],
4790 					    sa, DEV_BSIZE);
4791 				}
4792 			}
4793 		}
4794 		bp->b_pages[i]->valid |= mask;
4795 	}
4796 unlock:
4797 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4798 	bp->b_resid = 0;
4799 }
4800 
4801 void
4802 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4803 {
4804 	vm_page_t m;
4805 	int i, n;
4806 
4807 	if (buf_mapped(bp)) {
4808 		BUF_CHECK_MAPPED(bp);
4809 		bzero(bp->b_data + base, size);
4810 	} else {
4811 		BUF_CHECK_UNMAPPED(bp);
4812 		n = PAGE_SIZE - (base & PAGE_MASK);
4813 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4814 			m = bp->b_pages[i];
4815 			if (n > size)
4816 				n = size;
4817 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4818 			base += n;
4819 			size -= n;
4820 			n = PAGE_SIZE;
4821 		}
4822 	}
4823 }
4824 
4825 /*
4826  * Update buffer flags based on I/O request parameters, optionally releasing the
4827  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4828  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4829  * I/O).  Otherwise the buffer is released to the cache.
4830  */
4831 static void
4832 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4833 {
4834 
4835 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4836 	    ("buf %p non-VMIO noreuse", bp));
4837 
4838 	if ((ioflag & IO_DIRECT) != 0)
4839 		bp->b_flags |= B_DIRECT;
4840 	if ((ioflag & IO_EXT) != 0)
4841 		bp->b_xflags |= BX_ALTDATA;
4842 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4843 		bp->b_flags |= B_RELBUF;
4844 		if ((ioflag & IO_NOREUSE) != 0)
4845 			bp->b_flags |= B_NOREUSE;
4846 		if (release)
4847 			brelse(bp);
4848 	} else if (release)
4849 		bqrelse(bp);
4850 }
4851 
4852 void
4853 vfs_bio_brelse(struct buf *bp, int ioflag)
4854 {
4855 
4856 	b_io_dismiss(bp, ioflag, true);
4857 }
4858 
4859 void
4860 vfs_bio_set_flags(struct buf *bp, int ioflag)
4861 {
4862 
4863 	b_io_dismiss(bp, ioflag, false);
4864 }
4865 
4866 /*
4867  * vm_hold_load_pages and vm_hold_free_pages get pages into
4868  * a buffers address space.  The pages are anonymous and are
4869  * not associated with a file object.
4870  */
4871 static void
4872 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4873 {
4874 	vm_offset_t pg;
4875 	vm_page_t p;
4876 	int index;
4877 
4878 	BUF_CHECK_MAPPED(bp);
4879 
4880 	to = round_page(to);
4881 	from = round_page(from);
4882 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4883 
4884 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4885 		/*
4886 		 * note: must allocate system pages since blocking here
4887 		 * could interfere with paging I/O, no matter which
4888 		 * process we are.
4889 		 */
4890 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4891 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4892 		    VM_ALLOC_WAITOK);
4893 		pmap_qenter(pg, &p, 1);
4894 		bp->b_pages[index] = p;
4895 	}
4896 	bp->b_npages = index;
4897 }
4898 
4899 /* Return pages associated with this buf to the vm system */
4900 static void
4901 vm_hold_free_pages(struct buf *bp, int newbsize)
4902 {
4903 	vm_offset_t from;
4904 	vm_page_t p;
4905 	int index, newnpages;
4906 
4907 	BUF_CHECK_MAPPED(bp);
4908 
4909 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4910 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4911 	if (bp->b_npages > newnpages)
4912 		pmap_qremove(from, bp->b_npages - newnpages);
4913 	for (index = newnpages; index < bp->b_npages; index++) {
4914 		p = bp->b_pages[index];
4915 		bp->b_pages[index] = NULL;
4916 		p->wire_count--;
4917 		vm_page_free(p);
4918 	}
4919 	vm_wire_sub(bp->b_npages - newnpages);
4920 	bp->b_npages = newnpages;
4921 }
4922 
4923 /*
4924  * Map an IO request into kernel virtual address space.
4925  *
4926  * All requests are (re)mapped into kernel VA space.
4927  * Notice that we use b_bufsize for the size of the buffer
4928  * to be mapped.  b_bcount might be modified by the driver.
4929  *
4930  * Note that even if the caller determines that the address space should
4931  * be valid, a race or a smaller-file mapped into a larger space may
4932  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4933  * check the return value.
4934  *
4935  * This function only works with pager buffers.
4936  */
4937 int
4938 vmapbuf(struct buf *bp, int mapbuf)
4939 {
4940 	vm_prot_t prot;
4941 	int pidx;
4942 
4943 	if (bp->b_bufsize < 0)
4944 		return (-1);
4945 	prot = VM_PROT_READ;
4946 	if (bp->b_iocmd == BIO_READ)
4947 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4948 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4949 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4950 	    btoc(MAXPHYS))) < 0)
4951 		return (-1);
4952 	bp->b_npages = pidx;
4953 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4954 	if (mapbuf || !unmapped_buf_allowed) {
4955 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4956 		bp->b_data = bp->b_kvabase + bp->b_offset;
4957 	} else
4958 		bp->b_data = unmapped_buf;
4959 	return(0);
4960 }
4961 
4962 /*
4963  * Free the io map PTEs associated with this IO operation.
4964  * We also invalidate the TLB entries and restore the original b_addr.
4965  *
4966  * This function only works with pager buffers.
4967  */
4968 void
4969 vunmapbuf(struct buf *bp)
4970 {
4971 	int npages;
4972 
4973 	npages = bp->b_npages;
4974 	if (buf_mapped(bp))
4975 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4976 	vm_page_unhold_pages(bp->b_pages, npages);
4977 
4978 	bp->b_data = unmapped_buf;
4979 }
4980 
4981 void
4982 bdone(struct buf *bp)
4983 {
4984 	struct mtx *mtxp;
4985 
4986 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4987 	mtx_lock(mtxp);
4988 	bp->b_flags |= B_DONE;
4989 	wakeup(bp);
4990 	mtx_unlock(mtxp);
4991 }
4992 
4993 void
4994 bwait(struct buf *bp, u_char pri, const char *wchan)
4995 {
4996 	struct mtx *mtxp;
4997 
4998 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4999 	mtx_lock(mtxp);
5000 	while ((bp->b_flags & B_DONE) == 0)
5001 		msleep(bp, mtxp, pri, wchan, 0);
5002 	mtx_unlock(mtxp);
5003 }
5004 
5005 int
5006 bufsync(struct bufobj *bo, int waitfor)
5007 {
5008 
5009 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5010 }
5011 
5012 void
5013 bufstrategy(struct bufobj *bo, struct buf *bp)
5014 {
5015 	int i __unused;
5016 	struct vnode *vp;
5017 
5018 	vp = bp->b_vp;
5019 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5020 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5021 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5022 	i = VOP_STRATEGY(vp, bp);
5023 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5024 }
5025 
5026 /*
5027  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
5028  */
5029 void
5030 bufobj_init(struct bufobj *bo, void *private)
5031 {
5032 	static volatile int bufobj_cleanq;
5033 
5034         bo->bo_domain =
5035             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5036         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5037         bo->bo_private = private;
5038         TAILQ_INIT(&bo->bo_clean.bv_hd);
5039         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5040 }
5041 
5042 void
5043 bufobj_wrefl(struct bufobj *bo)
5044 {
5045 
5046 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5047 	ASSERT_BO_WLOCKED(bo);
5048 	bo->bo_numoutput++;
5049 }
5050 
5051 void
5052 bufobj_wref(struct bufobj *bo)
5053 {
5054 
5055 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5056 	BO_LOCK(bo);
5057 	bo->bo_numoutput++;
5058 	BO_UNLOCK(bo);
5059 }
5060 
5061 void
5062 bufobj_wdrop(struct bufobj *bo)
5063 {
5064 
5065 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5066 	BO_LOCK(bo);
5067 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5068 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5069 		bo->bo_flag &= ~BO_WWAIT;
5070 		wakeup(&bo->bo_numoutput);
5071 	}
5072 	BO_UNLOCK(bo);
5073 }
5074 
5075 int
5076 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5077 {
5078 	int error;
5079 
5080 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5081 	ASSERT_BO_WLOCKED(bo);
5082 	error = 0;
5083 	while (bo->bo_numoutput) {
5084 		bo->bo_flag |= BO_WWAIT;
5085 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5086 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5087 		if (error)
5088 			break;
5089 	}
5090 	return (error);
5091 }
5092 
5093 /*
5094  * Set bio_data or bio_ma for struct bio from the struct buf.
5095  */
5096 void
5097 bdata2bio(struct buf *bp, struct bio *bip)
5098 {
5099 
5100 	if (!buf_mapped(bp)) {
5101 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5102 		bip->bio_ma = bp->b_pages;
5103 		bip->bio_ma_n = bp->b_npages;
5104 		bip->bio_data = unmapped_buf;
5105 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5106 		bip->bio_flags |= BIO_UNMAPPED;
5107 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5108 		    PAGE_SIZE == bp->b_npages,
5109 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5110 		    (long long)bip->bio_length, bip->bio_ma_n));
5111 	} else {
5112 		bip->bio_data = bp->b_data;
5113 		bip->bio_ma = NULL;
5114 	}
5115 }
5116 
5117 /*
5118  * The MIPS pmap code currently doesn't handle aliased pages.
5119  * The VIPT caches may not handle page aliasing themselves, leading
5120  * to data corruption.
5121  *
5122  * As such, this code makes a system extremely unhappy if said
5123  * system doesn't support unaliasing the above situation in hardware.
5124  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5125  * this feature at build time, so it has to be handled in software.
5126  *
5127  * Once the MIPS pmap/cache code grows to support this function on
5128  * earlier chips, it should be flipped back off.
5129  */
5130 #ifdef	__mips__
5131 static int buf_pager_relbuf = 1;
5132 #else
5133 static int buf_pager_relbuf = 0;
5134 #endif
5135 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5136     &buf_pager_relbuf, 0,
5137     "Make buffer pager release buffers after reading");
5138 
5139 /*
5140  * The buffer pager.  It uses buffer reads to validate pages.
5141  *
5142  * In contrast to the generic local pager from vm/vnode_pager.c, this
5143  * pager correctly and easily handles volumes where the underlying
5144  * device block size is greater than the machine page size.  The
5145  * buffer cache transparently extends the requested page run to be
5146  * aligned at the block boundary, and does the necessary bogus page
5147  * replacements in the addends to avoid obliterating already valid
5148  * pages.
5149  *
5150  * The only non-trivial issue is that the exclusive busy state for
5151  * pages, which is assumed by the vm_pager_getpages() interface, is
5152  * incompatible with the VMIO buffer cache's desire to share-busy the
5153  * pages.  This function performs a trivial downgrade of the pages'
5154  * state before reading buffers, and a less trivial upgrade from the
5155  * shared-busy to excl-busy state after the read.
5156  */
5157 int
5158 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5159     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5160     vbg_get_blksize_t get_blksize)
5161 {
5162 	vm_page_t m;
5163 	vm_object_t object;
5164 	struct buf *bp;
5165 	struct mount *mp;
5166 	daddr_t lbn, lbnp;
5167 	vm_ooffset_t la, lb, poff, poffe;
5168 	long bsize;
5169 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5170 	bool redo, lpart;
5171 
5172 	object = vp->v_object;
5173 	mp = vp->v_mount;
5174 	error = 0;
5175 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5176 	if (la >= object->un_pager.vnp.vnp_size)
5177 		return (VM_PAGER_BAD);
5178 
5179 	/*
5180 	 * Change the meaning of la from where the last requested page starts
5181 	 * to where it ends, because that's the end of the requested region
5182 	 * and the start of the potential read-ahead region.
5183 	 */
5184 	la += PAGE_SIZE;
5185 	lpart = la > object->un_pager.vnp.vnp_size;
5186 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5187 
5188 	/*
5189 	 * Calculate read-ahead, behind and total pages.
5190 	 */
5191 	pgsin = count;
5192 	lb = IDX_TO_OFF(ma[0]->pindex);
5193 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5194 	pgsin += pgsin_b;
5195 	if (rbehind != NULL)
5196 		*rbehind = pgsin_b;
5197 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5198 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5199 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5200 		    PAGE_SIZE) - la);
5201 	pgsin += pgsin_a;
5202 	if (rahead != NULL)
5203 		*rahead = pgsin_a;
5204 	VM_CNT_INC(v_vnodein);
5205 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5206 
5207 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5208 	    != 0) ? GB_UNMAPPED : 0;
5209 	VM_OBJECT_WLOCK(object);
5210 again:
5211 	for (i = 0; i < count; i++)
5212 		vm_page_busy_downgrade(ma[i]);
5213 	VM_OBJECT_WUNLOCK(object);
5214 
5215 	lbnp = -1;
5216 	for (i = 0; i < count; i++) {
5217 		m = ma[i];
5218 
5219 		/*
5220 		 * Pages are shared busy and the object lock is not
5221 		 * owned, which together allow for the pages'
5222 		 * invalidation.  The racy test for validity avoids
5223 		 * useless creation of the buffer for the most typical
5224 		 * case when invalidation is not used in redo or for
5225 		 * parallel read.  The shared->excl upgrade loop at
5226 		 * the end of the function catches the race in a
5227 		 * reliable way (protected by the object lock).
5228 		 */
5229 		if (m->valid == VM_PAGE_BITS_ALL)
5230 			continue;
5231 
5232 		poff = IDX_TO_OFF(m->pindex);
5233 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5234 		for (; poff < poffe; poff += bsize) {
5235 			lbn = get_lblkno(vp, poff);
5236 			if (lbn == lbnp)
5237 				goto next_page;
5238 			lbnp = lbn;
5239 
5240 			bsize = get_blksize(vp, lbn);
5241 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5242 			    br_flags, &bp);
5243 			if (error != 0)
5244 				goto end_pages;
5245 			if (LIST_EMPTY(&bp->b_dep)) {
5246 				/*
5247 				 * Invalidation clears m->valid, but
5248 				 * may leave B_CACHE flag if the
5249 				 * buffer existed at the invalidation
5250 				 * time.  In this case, recycle the
5251 				 * buffer to do real read on next
5252 				 * bread() after redo.
5253 				 *
5254 				 * Otherwise B_RELBUF is not strictly
5255 				 * necessary, enable to reduce buf
5256 				 * cache pressure.
5257 				 */
5258 				if (buf_pager_relbuf ||
5259 				    m->valid != VM_PAGE_BITS_ALL)
5260 					bp->b_flags |= B_RELBUF;
5261 
5262 				bp->b_flags &= ~B_NOCACHE;
5263 				brelse(bp);
5264 			} else {
5265 				bqrelse(bp);
5266 			}
5267 		}
5268 		KASSERT(1 /* racy, enable for debugging */ ||
5269 		    m->valid == VM_PAGE_BITS_ALL || i == count - 1,
5270 		    ("buf %d %p invalid", i, m));
5271 		if (i == count - 1 && lpart) {
5272 			VM_OBJECT_WLOCK(object);
5273 			if (m->valid != 0 &&
5274 			    m->valid != VM_PAGE_BITS_ALL)
5275 				vm_page_zero_invalid(m, TRUE);
5276 			VM_OBJECT_WUNLOCK(object);
5277 		}
5278 next_page:;
5279 	}
5280 end_pages:
5281 
5282 	VM_OBJECT_WLOCK(object);
5283 	redo = false;
5284 	for (i = 0; i < count; i++) {
5285 		vm_page_sunbusy(ma[i]);
5286 		ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
5287 
5288 		/*
5289 		 * Since the pages were only sbusy while neither the
5290 		 * buffer nor the object lock was held by us, or
5291 		 * reallocated while vm_page_grab() slept for busy
5292 		 * relinguish, they could have been invalidated.
5293 		 * Recheck the valid bits and re-read as needed.
5294 		 *
5295 		 * Note that the last page is made fully valid in the
5296 		 * read loop, and partial validity for the page at
5297 		 * index count - 1 could mean that the page was
5298 		 * invalidated or removed, so we must restart for
5299 		 * safety as well.
5300 		 */
5301 		if (ma[i]->valid != VM_PAGE_BITS_ALL)
5302 			redo = true;
5303 	}
5304 	if (redo && error == 0)
5305 		goto again;
5306 	VM_OBJECT_WUNLOCK(object);
5307 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5308 }
5309 
5310 #include "opt_ddb.h"
5311 #ifdef DDB
5312 #include <ddb/ddb.h>
5313 
5314 /* DDB command to show buffer data */
5315 DB_SHOW_COMMAND(buffer, db_show_buffer)
5316 {
5317 	/* get args */
5318 	struct buf *bp = (struct buf *)addr;
5319 #ifdef FULL_BUF_TRACKING
5320 	uint32_t i, j;
5321 #endif
5322 
5323 	if (!have_addr) {
5324 		db_printf("usage: show buffer <addr>\n");
5325 		return;
5326 	}
5327 
5328 	db_printf("buf at %p\n", bp);
5329 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5330 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5331 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5332 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5333 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5334 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5335 	db_printf(
5336 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5337 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5338 	    "b_vp = %p, b_dep = %p\n",
5339 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5340 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5341 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5342 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5343 	    bp->b_kvabase, bp->b_kvasize);
5344 	if (bp->b_npages) {
5345 		int i;
5346 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5347 		for (i = 0; i < bp->b_npages; i++) {
5348 			vm_page_t m;
5349 			m = bp->b_pages[i];
5350 			if (m != NULL)
5351 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5352 				    (u_long)m->pindex,
5353 				    (u_long)VM_PAGE_TO_PHYS(m));
5354 			else
5355 				db_printf("( ??? )");
5356 			if ((i + 1) < bp->b_npages)
5357 				db_printf(",");
5358 		}
5359 		db_printf("\n");
5360 	}
5361 	BUF_LOCKPRINTINFO(bp);
5362 #if defined(FULL_BUF_TRACKING)
5363 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5364 
5365 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5366 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5367 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5368 			continue;
5369 		db_printf(" %2u: %s\n", j,
5370 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5371 	}
5372 #elif defined(BUF_TRACKING)
5373 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5374 #endif
5375 	db_printf(" ");
5376 }
5377 
5378 DB_SHOW_COMMAND(bufqueues, bufqueues)
5379 {
5380 	struct bufdomain *bd;
5381 	struct buf *bp;
5382 	long total;
5383 	int i, j, cnt;
5384 
5385 	db_printf("bqempty: %d\n", bqempty.bq_len);
5386 
5387 	for (i = 0; i < buf_domains; i++) {
5388 		bd = &bdomain[i];
5389 		db_printf("Buf domain %d\n", i);
5390 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5391 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5392 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5393 		db_printf("\n");
5394 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5395 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5396 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5397 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5398 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5399 		db_printf("\n");
5400 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5401 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5402 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5403 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5404 		db_printf("\n");
5405 		total = 0;
5406 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5407 			total += bp->b_bufsize;
5408 		db_printf("\tcleanq count\t%d (%ld)\n",
5409 		    bd->bd_cleanq->bq_len, total);
5410 		total = 0;
5411 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5412 			total += bp->b_bufsize;
5413 		db_printf("\tdirtyq count\t%d (%ld)\n",
5414 		    bd->bd_dirtyq.bq_len, total);
5415 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5416 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5417 		db_printf("\tCPU ");
5418 		for (j = 0; j <= mp_maxid; j++)
5419 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5420 		db_printf("\n");
5421 		cnt = 0;
5422 		total = 0;
5423 		for (j = 0; j < nbuf; j++)
5424 			if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5425 				cnt++;
5426 				total += buf[j].b_bufsize;
5427 			}
5428 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5429 		cnt = 0;
5430 		total = 0;
5431 		for (j = 0; j < nbuf; j++)
5432 			if (buf[j].b_domain == i) {
5433 				cnt++;
5434 				total += buf[j].b_bufsize;
5435 			}
5436 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5437 	}
5438 }
5439 
5440 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5441 {
5442 	struct buf *bp;
5443 	int i;
5444 
5445 	for (i = 0; i < nbuf; i++) {
5446 		bp = &buf[i];
5447 		if (BUF_ISLOCKED(bp)) {
5448 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5449 			db_printf("\n");
5450 			if (db_pager_quit)
5451 				break;
5452 		}
5453 	}
5454 }
5455 
5456 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5457 {
5458 	struct vnode *vp;
5459 	struct buf *bp;
5460 
5461 	if (!have_addr) {
5462 		db_printf("usage: show vnodebufs <addr>\n");
5463 		return;
5464 	}
5465 	vp = (struct vnode *)addr;
5466 	db_printf("Clean buffers:\n");
5467 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5468 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5469 		db_printf("\n");
5470 	}
5471 	db_printf("Dirty buffers:\n");
5472 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5473 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5474 		db_printf("\n");
5475 	}
5476 }
5477 
5478 DB_COMMAND(countfreebufs, db_coundfreebufs)
5479 {
5480 	struct buf *bp;
5481 	int i, used = 0, nfree = 0;
5482 
5483 	if (have_addr) {
5484 		db_printf("usage: countfreebufs\n");
5485 		return;
5486 	}
5487 
5488 	for (i = 0; i < nbuf; i++) {
5489 		bp = &buf[i];
5490 		if (bp->b_qindex == QUEUE_EMPTY)
5491 			nfree++;
5492 		else
5493 			used++;
5494 	}
5495 
5496 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5497 	    nfree + used);
5498 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5499 }
5500 #endif /* DDB */
5501