xref: /freebsd/sys/kern/vfs_bio.c (revision 25fb30bd9abc492359ad1f66901a06cb8cd08370)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2004 Poul-Henning Kamp
5  * Copyright (c) 1994,1997 John S. Dyson
6  * Copyright (c) 2013 The FreeBSD Foundation
7  * All rights reserved.
8  *
9  * Portions of this software were developed by Konstantin Belousov
10  * under sponsorship from the FreeBSD Foundation.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * this file contains a new buffer I/O scheme implementing a coherent
36  * VM object and buffer cache scheme.  Pains have been taken to make
37  * sure that the performance degradation associated with schemes such
38  * as this is not realized.
39  *
40  * Author:  John S. Dyson
41  * Significant help during the development and debugging phases
42  * had been provided by David Greenman, also of the FreeBSD core team.
43  *
44  * see man buf(9) for more info.
45  */
46 
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/bio.h>
53 #include <sys/bitset.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/buf.h>
57 #include <sys/devicestat.h>
58 #include <sys/eventhandler.h>
59 #include <sys/fail.h>
60 #include <sys/ktr.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/malloc.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/proc.h>
69 #include <sys/racct.h>
70 #include <sys/refcount.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/smp.h>
74 #include <sys/sysctl.h>
75 #include <sys/syscallsubr.h>
76 #include <sys/vmem.h>
77 #include <sys/vmmeter.h>
78 #include <sys/vnode.h>
79 #include <sys/watchdog.h>
80 #include <geom/geom.h>
81 #include <vm/vm.h>
82 #include <vm/vm_param.h>
83 #include <vm/vm_kern.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_pager.h>
88 #include <vm/vm_extern.h>
89 #include <vm/vm_map.h>
90 #include <vm/swap_pager.h>
91 
92 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
93 
94 struct	bio_ops bioops;		/* I/O operation notification */
95 
96 struct	buf_ops buf_ops_bio = {
97 	.bop_name	=	"buf_ops_bio",
98 	.bop_write	=	bufwrite,
99 	.bop_strategy	=	bufstrategy,
100 	.bop_sync	=	bufsync,
101 	.bop_bdflush	=	bufbdflush,
102 };
103 
104 struct bufqueue {
105 	struct mtx_padalign	bq_lock;
106 	TAILQ_HEAD(, buf)	bq_queue;
107 	uint8_t			bq_index;
108 	uint16_t		bq_subqueue;
109 	int			bq_len;
110 } __aligned(CACHE_LINE_SIZE);
111 
112 #define	BQ_LOCKPTR(bq)		(&(bq)->bq_lock)
113 #define	BQ_LOCK(bq)		mtx_lock(BQ_LOCKPTR((bq)))
114 #define	BQ_UNLOCK(bq)		mtx_unlock(BQ_LOCKPTR((bq)))
115 #define	BQ_ASSERT_LOCKED(bq)	mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
116 
117 struct bufdomain {
118 	struct bufqueue	bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
119 	struct bufqueue bd_dirtyq;
120 	struct bufqueue	*bd_cleanq;
121 	struct mtx_padalign bd_run_lock;
122 	/* Constants */
123 	long		bd_maxbufspace;
124 	long		bd_hibufspace;
125 	long 		bd_lobufspace;
126 	long 		bd_bufspacethresh;
127 	int		bd_hifreebuffers;
128 	int		bd_lofreebuffers;
129 	int		bd_hidirtybuffers;
130 	int		bd_lodirtybuffers;
131 	int		bd_dirtybufthresh;
132 	int		bd_lim;
133 	/* atomics */
134 	int		bd_wanted;
135 	int __aligned(CACHE_LINE_SIZE)	bd_numdirtybuffers;
136 	int __aligned(CACHE_LINE_SIZE)	bd_running;
137 	long __aligned(CACHE_LINE_SIZE) bd_bufspace;
138 	int __aligned(CACHE_LINE_SIZE)	bd_freebuffers;
139 } __aligned(CACHE_LINE_SIZE);
140 
141 #define	BD_LOCKPTR(bd)		(&(bd)->bd_cleanq->bq_lock)
142 #define	BD_LOCK(bd)		mtx_lock(BD_LOCKPTR((bd)))
143 #define	BD_UNLOCK(bd)		mtx_unlock(BD_LOCKPTR((bd)))
144 #define	BD_ASSERT_LOCKED(bd)	mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
145 #define	BD_RUN_LOCKPTR(bd)	(&(bd)->bd_run_lock)
146 #define	BD_RUN_LOCK(bd)		mtx_lock(BD_RUN_LOCKPTR((bd)))
147 #define	BD_RUN_UNLOCK(bd)	mtx_unlock(BD_RUN_LOCKPTR((bd)))
148 #define	BD_DOMAIN(bd)		(bd - bdomain)
149 
150 static struct buf *buf;		/* buffer header pool */
151 extern struct buf *swbuf;	/* Swap buffer header pool. */
152 caddr_t __read_mostly unmapped_buf;
153 
154 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
155 struct proc *bufdaemonproc;
156 
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_range(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 | CTLFLAG_STATS,
253     &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
254 static int recursiveflushes;
255 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
256     &recursiveflushes, 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 | CTLFLAG_STATS,
312     &barrierwrites, 0, "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  *	bufspace_daemon:
783  *
784  *	buffer space management daemon.  Tries to maintain some marginal
785  *	amount of free buffer space so that requesting processes neither
786  *	block nor work to reclaim buffers.
787  */
788 static void
789 bufspace_daemon(void *arg)
790 {
791 	struct bufdomain *bd;
792 
793 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
794 	    SHUTDOWN_PRI_LAST + 100);
795 
796 	bd = arg;
797 	for (;;) {
798 		kthread_suspend_check();
799 
800 		/*
801 		 * Free buffers from the clean queue until we meet our
802 		 * targets.
803 		 *
804 		 * Theory of operation:  The buffer cache is most efficient
805 		 * when some free buffer headers and space are always
806 		 * available to getnewbuf().  This daemon attempts to prevent
807 		 * the excessive blocking and synchronization associated
808 		 * with shortfall.  It goes through three phases according
809 		 * demand:
810 		 *
811 		 * 1)	The daemon wakes up voluntarily once per-second
812 		 *	during idle periods when the counters are below
813 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
814 		 *
815 		 * 2)	The daemon wakes up as we cross the thresholds
816 		 *	ahead of any potential blocking.  This may bounce
817 		 *	slightly according to the rate of consumption and
818 		 *	release.
819 		 *
820 		 * 3)	The daemon and consumers are starved for working
821 		 *	clean buffers.  This is the 'bufspace' sleep below
822 		 *	which will inefficiently trade bufs with bqrelse
823 		 *	until we return to condition 2.
824 		 */
825 		while (bd->bd_bufspace > bd->bd_lobufspace ||
826 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
827 			if (buf_recycle(bd, false) != 0) {
828 				if (bd_flushall(bd))
829 					continue;
830 				/*
831 				 * Speedup dirty if we've run out of clean
832 				 * buffers.  This is possible in particular
833 				 * because softdep may held many bufs locked
834 				 * pending writes to other bufs which are
835 				 * marked for delayed write, exhausting
836 				 * clean space until they are written.
837 				 */
838 				bd_speedup();
839 				BD_LOCK(bd);
840 				if (bd->bd_wanted) {
841 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
842 					    PRIBIO|PDROP, "bufspace", hz/10);
843 				} else
844 					BD_UNLOCK(bd);
845 			}
846 			maybe_yield();
847 		}
848 		bufspace_daemon_wait(bd);
849 	}
850 }
851 
852 /*
853  *	bufmallocadjust:
854  *
855  *	Adjust the reported bufspace for a malloc managed buffer, possibly
856  *	waking any waiters.
857  */
858 static void
859 bufmallocadjust(struct buf *bp, int bufsize)
860 {
861 	int diff;
862 
863 	KASSERT((bp->b_flags & B_MALLOC) != 0,
864 	    ("bufmallocadjust: non-malloc buf %p", bp));
865 	diff = bufsize - bp->b_bufsize;
866 	if (diff < 0)
867 		atomic_subtract_long(&bufmallocspace, -diff);
868 	else
869 		atomic_add_long(&bufmallocspace, diff);
870 	bp->b_bufsize = bufsize;
871 }
872 
873 /*
874  *	runningwakeup:
875  *
876  *	Wake up processes that are waiting on asynchronous writes to fall
877  *	below lorunningspace.
878  */
879 static void
880 runningwakeup(void)
881 {
882 
883 	mtx_lock(&rbreqlock);
884 	if (runningbufreq) {
885 		runningbufreq = 0;
886 		wakeup(&runningbufreq);
887 	}
888 	mtx_unlock(&rbreqlock);
889 }
890 
891 /*
892  *	runningbufwakeup:
893  *
894  *	Decrement the outstanding write count according.
895  */
896 void
897 runningbufwakeup(struct buf *bp)
898 {
899 	long space, bspace;
900 
901 	bspace = bp->b_runningbufspace;
902 	if (bspace == 0)
903 		return;
904 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
905 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
906 	    space, bspace));
907 	bp->b_runningbufspace = 0;
908 	/*
909 	 * Only acquire the lock and wakeup on the transition from exceeding
910 	 * the threshold to falling below it.
911 	 */
912 	if (space < lorunningspace)
913 		return;
914 	if (space - bspace > lorunningspace)
915 		return;
916 	runningwakeup();
917 }
918 
919 /*
920  *	waitrunningbufspace()
921  *
922  *	runningbufspace is a measure of the amount of I/O currently
923  *	running.  This routine is used in async-write situations to
924  *	prevent creating huge backups of pending writes to a device.
925  *	Only asynchronous writes are governed by this function.
926  *
927  *	This does NOT turn an async write into a sync write.  It waits
928  *	for earlier writes to complete and generally returns before the
929  *	caller's write has reached the device.
930  */
931 void
932 waitrunningbufspace(void)
933 {
934 
935 	mtx_lock(&rbreqlock);
936 	while (runningbufspace > hirunningspace) {
937 		runningbufreq = 1;
938 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
939 	}
940 	mtx_unlock(&rbreqlock);
941 }
942 
943 /*
944  *	vfs_buf_test_cache:
945  *
946  *	Called when a buffer is extended.  This function clears the B_CACHE
947  *	bit if the newly extended portion of the buffer does not contain
948  *	valid data.
949  */
950 static __inline void
951 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
952     vm_offset_t size, vm_page_t m)
953 {
954 
955 	/*
956 	 * This function and its results are protected by higher level
957 	 * synchronization requiring vnode and buf locks to page in and
958 	 * validate pages.
959 	 */
960 	if (bp->b_flags & B_CACHE) {
961 		int base = (foff + off) & PAGE_MASK;
962 		if (vm_page_is_valid(m, base, size) == 0)
963 			bp->b_flags &= ~B_CACHE;
964 	}
965 }
966 
967 /* Wake up the buffer daemon if necessary */
968 static void
969 bd_wakeup(void)
970 {
971 
972 	mtx_lock(&bdlock);
973 	if (bd_request == 0) {
974 		bd_request = 1;
975 		wakeup(&bd_request);
976 	}
977 	mtx_unlock(&bdlock);
978 }
979 
980 /*
981  * Adjust the maxbcachbuf tunable.
982  */
983 static void
984 maxbcachebuf_adjust(void)
985 {
986 	int i;
987 
988 	/*
989 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
990 	 */
991 	i = 2;
992 	while (i * 2 <= maxbcachebuf)
993 		i *= 2;
994 	maxbcachebuf = i;
995 	if (maxbcachebuf < MAXBSIZE)
996 		maxbcachebuf = MAXBSIZE;
997 	if (maxbcachebuf > MAXPHYS)
998 		maxbcachebuf = MAXPHYS;
999 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1000 		printf("maxbcachebuf=%d\n", maxbcachebuf);
1001 }
1002 
1003 /*
1004  * bd_speedup - speedup the buffer cache flushing code
1005  */
1006 void
1007 bd_speedup(void)
1008 {
1009 	int needwake;
1010 
1011 	mtx_lock(&bdlock);
1012 	needwake = 0;
1013 	if (bd_speedupreq == 0 || bd_request == 0)
1014 		needwake = 1;
1015 	bd_speedupreq = 1;
1016 	bd_request = 1;
1017 	if (needwake)
1018 		wakeup(&bd_request);
1019 	mtx_unlock(&bdlock);
1020 }
1021 
1022 #ifdef __i386__
1023 #define	TRANSIENT_DENOM	5
1024 #else
1025 #define	TRANSIENT_DENOM 10
1026 #endif
1027 
1028 /*
1029  * Calculating buffer cache scaling values and reserve space for buffer
1030  * headers.  This is called during low level kernel initialization and
1031  * may be called more then once.  We CANNOT write to the memory area
1032  * being reserved at this time.
1033  */
1034 caddr_t
1035 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1036 {
1037 	int tuned_nbuf;
1038 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1039 
1040 	/*
1041 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1042 	 * PAGE_SIZE is >= 1K)
1043 	 */
1044 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1045 
1046 	maxbcachebuf_adjust();
1047 	/*
1048 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1049 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1050 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1051 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1052 	 * the buffer cache we limit the eventual kva reservation to
1053 	 * maxbcache bytes.
1054 	 *
1055 	 * factor represents the 1/4 x ram conversion.
1056 	 */
1057 	if (nbuf == 0) {
1058 		int factor = 4 * BKVASIZE / 1024;
1059 
1060 		nbuf = 50;
1061 		if (physmem_est > 4096)
1062 			nbuf += min((physmem_est - 4096) / factor,
1063 			    65536 / factor);
1064 		if (physmem_est > 65536)
1065 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1066 			    32 * 1024 * 1024 / (factor * 5));
1067 
1068 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1069 			nbuf = maxbcache / BKVASIZE;
1070 		tuned_nbuf = 1;
1071 	} else
1072 		tuned_nbuf = 0;
1073 
1074 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1075 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1076 	if (nbuf > maxbuf) {
1077 		if (!tuned_nbuf)
1078 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1079 			    maxbuf);
1080 		nbuf = maxbuf;
1081 	}
1082 
1083 	/*
1084 	 * Ideal allocation size for the transient bio submap is 10%
1085 	 * of the maximal space buffer map.  This roughly corresponds
1086 	 * to the amount of the buffer mapped for typical UFS load.
1087 	 *
1088 	 * Clip the buffer map to reserve space for the transient
1089 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1090 	 * maximum buffer map extent on the platform.
1091 	 *
1092 	 * The fall-back to the maxbuf in case of maxbcache unset,
1093 	 * allows to not trim the buffer KVA for the architectures
1094 	 * with ample KVA space.
1095 	 */
1096 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1097 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1098 		buf_sz = (long)nbuf * BKVASIZE;
1099 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1100 		    (TRANSIENT_DENOM - 1)) {
1101 			/*
1102 			 * There is more KVA than memory.  Do not
1103 			 * adjust buffer map size, and assign the rest
1104 			 * of maxbuf to transient map.
1105 			 */
1106 			biotmap_sz = maxbuf_sz - buf_sz;
1107 		} else {
1108 			/*
1109 			 * Buffer map spans all KVA we could afford on
1110 			 * this platform.  Give 10% (20% on i386) of
1111 			 * the buffer map to the transient bio map.
1112 			 */
1113 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1114 			buf_sz -= biotmap_sz;
1115 		}
1116 		if (biotmap_sz / INT_MAX > MAXPHYS)
1117 			bio_transient_maxcnt = INT_MAX;
1118 		else
1119 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1120 		/*
1121 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1122 		 * using the transient mapping.
1123 		 */
1124 		if (bio_transient_maxcnt > 1024)
1125 			bio_transient_maxcnt = 1024;
1126 		if (tuned_nbuf)
1127 			nbuf = buf_sz / BKVASIZE;
1128 	}
1129 
1130 	if (nswbuf == 0) {
1131 		nswbuf = min(nbuf / 4, 256);
1132 		if (nswbuf < NSWBUF_MIN)
1133 			nswbuf = NSWBUF_MIN;
1134 	}
1135 
1136 	/*
1137 	 * Reserve space for the buffer cache buffers
1138 	 */
1139 	buf = (void *)v;
1140 	v = (caddr_t)(buf + nbuf);
1141 
1142 	return(v);
1143 }
1144 
1145 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1146 void
1147 bufinit(void)
1148 {
1149 	struct buf *bp;
1150 	int i;
1151 
1152 	KASSERT(maxbcachebuf >= MAXBSIZE,
1153 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1154 	    MAXBSIZE));
1155 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1156 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1157 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1158 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1159 
1160 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1161 
1162 	/* finally, initialize each buffer header and stick on empty q */
1163 	for (i = 0; i < nbuf; i++) {
1164 		bp = &buf[i];
1165 		bzero(bp, sizeof *bp);
1166 		bp->b_flags = B_INVAL;
1167 		bp->b_rcred = NOCRED;
1168 		bp->b_wcred = NOCRED;
1169 		bp->b_qindex = QUEUE_NONE;
1170 		bp->b_domain = -1;
1171 		bp->b_subqueue = mp_maxid + 1;
1172 		bp->b_xflags = 0;
1173 		bp->b_data = bp->b_kvabase = unmapped_buf;
1174 		LIST_INIT(&bp->b_dep);
1175 		BUF_LOCKINIT(bp);
1176 		bq_insert(&bqempty, bp, false);
1177 	}
1178 
1179 	/*
1180 	 * maxbufspace is the absolute maximum amount of buffer space we are
1181 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1182 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1183 	 * used by most other requests.  The differential is required to
1184 	 * ensure that metadata deadlocks don't occur.
1185 	 *
1186 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1187 	 * this may result in KVM fragmentation which is not handled optimally
1188 	 * by the system. XXX This is less true with vmem.  We could use
1189 	 * PAGE_SIZE.
1190 	 */
1191 	maxbufspace = (long)nbuf * BKVASIZE;
1192 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1193 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1194 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1195 
1196 	/*
1197 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1198 	 * arbitrarily and may need further tuning. It corresponds to
1199 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1200 	 * which fits with many RAID controllers' tagged queuing limits.
1201 	 * The lower 1 MiB limit is the historical upper limit for
1202 	 * hirunningspace.
1203 	 */
1204 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1205 	    16 * 1024 * 1024), 1024 * 1024);
1206 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1207 
1208 	/*
1209 	 * Limit the amount of malloc memory since it is wired permanently into
1210 	 * the kernel space.  Even though this is accounted for in the buffer
1211 	 * allocation, we don't want the malloced region to grow uncontrolled.
1212 	 * The malloc scheme improves memory utilization significantly on
1213 	 * average (small) directories.
1214 	 */
1215 	maxbufmallocspace = hibufspace / 20;
1216 
1217 	/*
1218 	 * Reduce the chance of a deadlock occurring by limiting the number
1219 	 * of delayed-write dirty buffers we allow to stack up.
1220 	 */
1221 	hidirtybuffers = nbuf / 4 + 20;
1222 	dirtybufthresh = hidirtybuffers * 9 / 10;
1223 	/*
1224 	 * To support extreme low-memory systems, make sure hidirtybuffers
1225 	 * cannot eat up all available buffer space.  This occurs when our
1226 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1227 	 * buffer space assuming BKVASIZE'd buffers.
1228 	 */
1229 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1230 		hidirtybuffers >>= 1;
1231 	}
1232 	lodirtybuffers = hidirtybuffers / 2;
1233 
1234 	/*
1235 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1236 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1237 	 * are counted as free but will be unavailable to threads executing
1238 	 * on other cpus.
1239 	 *
1240 	 * hifreebuffers is the free target for the bufspace daemon.  This
1241 	 * should be set appropriately to limit work per-iteration.
1242 	 */
1243 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1244 	hifreebuffers = (3 * lofreebuffers) / 2;
1245 	numfreebuffers = nbuf;
1246 
1247 	/* Setup the kva and free list allocators. */
1248 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1249 	buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1250 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1251 
1252 	/*
1253 	 * Size the clean queue according to the amount of buffer space.
1254 	 * One queue per-256mb up to the max.  More queues gives better
1255 	 * concurrency but less accurate LRU.
1256 	 */
1257 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1258 	for (i = 0 ; i < buf_domains; i++) {
1259 		struct bufdomain *bd;
1260 
1261 		bd = &bdomain[i];
1262 		bd_init(bd);
1263 		bd->bd_freebuffers = nbuf / buf_domains;
1264 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1265 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1266 		bd->bd_bufspace = 0;
1267 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1268 		bd->bd_hibufspace = hibufspace / buf_domains;
1269 		bd->bd_lobufspace = lobufspace / buf_domains;
1270 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1271 		bd->bd_numdirtybuffers = 0;
1272 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1273 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1274 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1275 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1276 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1277 	}
1278 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1279 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1280 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1281 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1282 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1283 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1284 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1285 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1286 }
1287 
1288 #ifdef INVARIANTS
1289 static inline void
1290 vfs_buf_check_mapped(struct buf *bp)
1291 {
1292 
1293 	KASSERT(bp->b_kvabase != unmapped_buf,
1294 	    ("mapped buf: b_kvabase was not updated %p", bp));
1295 	KASSERT(bp->b_data != unmapped_buf,
1296 	    ("mapped buf: b_data was not updated %p", bp));
1297 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1298 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1299 }
1300 
1301 static inline void
1302 vfs_buf_check_unmapped(struct buf *bp)
1303 {
1304 
1305 	KASSERT(bp->b_data == unmapped_buf,
1306 	    ("unmapped buf: corrupted b_data %p", bp));
1307 }
1308 
1309 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1310 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1311 #else
1312 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1313 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1314 #endif
1315 
1316 static int
1317 isbufbusy(struct buf *bp)
1318 {
1319 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1320 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1321 		return (1);
1322 	return (0);
1323 }
1324 
1325 /*
1326  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1327  */
1328 void
1329 bufshutdown(int show_busybufs)
1330 {
1331 	static int first_buf_printf = 1;
1332 	struct buf *bp;
1333 	int iter, nbusy, pbusy;
1334 #ifndef PREEMPTION
1335 	int subiter;
1336 #endif
1337 
1338 	/*
1339 	 * Sync filesystems for shutdown
1340 	 */
1341 	wdog_kern_pat(WD_LASTVAL);
1342 	kern_sync(curthread);
1343 
1344 	/*
1345 	 * With soft updates, some buffers that are
1346 	 * written will be remarked as dirty until other
1347 	 * buffers are written.
1348 	 */
1349 	for (iter = pbusy = 0; iter < 20; iter++) {
1350 		nbusy = 0;
1351 		for (bp = &buf[nbuf]; --bp >= buf; )
1352 			if (isbufbusy(bp))
1353 				nbusy++;
1354 		if (nbusy == 0) {
1355 			if (first_buf_printf)
1356 				printf("All buffers synced.");
1357 			break;
1358 		}
1359 		if (first_buf_printf) {
1360 			printf("Syncing disks, buffers remaining... ");
1361 			first_buf_printf = 0;
1362 		}
1363 		printf("%d ", nbusy);
1364 		if (nbusy < pbusy)
1365 			iter = 0;
1366 		pbusy = nbusy;
1367 
1368 		wdog_kern_pat(WD_LASTVAL);
1369 		kern_sync(curthread);
1370 
1371 #ifdef PREEMPTION
1372 		/*
1373 		 * Spin for a while to allow interrupt threads to run.
1374 		 */
1375 		DELAY(50000 * iter);
1376 #else
1377 		/*
1378 		 * Context switch several times to allow interrupt
1379 		 * threads to run.
1380 		 */
1381 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1382 			thread_lock(curthread);
1383 			mi_switch(SW_VOL);
1384 			DELAY(1000);
1385 		}
1386 #endif
1387 	}
1388 	printf("\n");
1389 	/*
1390 	 * Count only busy local buffers to prevent forcing
1391 	 * a fsck if we're just a client of a wedged NFS server
1392 	 */
1393 	nbusy = 0;
1394 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1395 		if (isbufbusy(bp)) {
1396 #if 0
1397 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1398 			if (bp->b_dev == NULL) {
1399 				TAILQ_REMOVE(&mountlist,
1400 				    bp->b_vp->v_mount, mnt_list);
1401 				continue;
1402 			}
1403 #endif
1404 			nbusy++;
1405 			if (show_busybufs > 0) {
1406 				printf(
1407 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1408 				    nbusy, bp, bp->b_vp, bp->b_flags,
1409 				    (intmax_t)bp->b_blkno,
1410 				    (intmax_t)bp->b_lblkno);
1411 				BUF_LOCKPRINTINFO(bp);
1412 				if (show_busybufs > 1)
1413 					vn_printf(bp->b_vp,
1414 					    "vnode content: ");
1415 			}
1416 		}
1417 	}
1418 	if (nbusy) {
1419 		/*
1420 		 * Failed to sync all blocks. Indicate this and don't
1421 		 * unmount filesystems (thus forcing an fsck on reboot).
1422 		 */
1423 		printf("Giving up on %d buffers\n", nbusy);
1424 		DELAY(5000000);	/* 5 seconds */
1425 	} else {
1426 		if (!first_buf_printf)
1427 			printf("Final sync complete\n");
1428 		/*
1429 		 * Unmount filesystems
1430 		 */
1431 		if (!KERNEL_PANICKED())
1432 			vfs_unmountall();
1433 	}
1434 	swapoff_all();
1435 	DELAY(100000);		/* wait for console output to finish */
1436 }
1437 
1438 static void
1439 bpmap_qenter(struct buf *bp)
1440 {
1441 
1442 	BUF_CHECK_MAPPED(bp);
1443 
1444 	/*
1445 	 * bp->b_data is relative to bp->b_offset, but
1446 	 * bp->b_offset may be offset into the first page.
1447 	 */
1448 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1449 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1450 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1451 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1452 }
1453 
1454 static inline struct bufdomain *
1455 bufdomain(struct buf *bp)
1456 {
1457 
1458 	return (&bdomain[bp->b_domain]);
1459 }
1460 
1461 static struct bufqueue *
1462 bufqueue(struct buf *bp)
1463 {
1464 
1465 	switch (bp->b_qindex) {
1466 	case QUEUE_NONE:
1467 		/* FALLTHROUGH */
1468 	case QUEUE_SENTINEL:
1469 		return (NULL);
1470 	case QUEUE_EMPTY:
1471 		return (&bqempty);
1472 	case QUEUE_DIRTY:
1473 		return (&bufdomain(bp)->bd_dirtyq);
1474 	case QUEUE_CLEAN:
1475 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1476 	default:
1477 		break;
1478 	}
1479 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1480 }
1481 
1482 /*
1483  * Return the locked bufqueue that bp is a member of.
1484  */
1485 static struct bufqueue *
1486 bufqueue_acquire(struct buf *bp)
1487 {
1488 	struct bufqueue *bq, *nbq;
1489 
1490 	/*
1491 	 * bp can be pushed from a per-cpu queue to the
1492 	 * cleanq while we're waiting on the lock.  Retry
1493 	 * if the queues don't match.
1494 	 */
1495 	bq = bufqueue(bp);
1496 	BQ_LOCK(bq);
1497 	for (;;) {
1498 		nbq = bufqueue(bp);
1499 		if (bq == nbq)
1500 			break;
1501 		BQ_UNLOCK(bq);
1502 		BQ_LOCK(nbq);
1503 		bq = nbq;
1504 	}
1505 	return (bq);
1506 }
1507 
1508 /*
1509  *	binsfree:
1510  *
1511  *	Insert the buffer into the appropriate free list.  Requires a
1512  *	locked buffer on entry and buffer is unlocked before return.
1513  */
1514 static void
1515 binsfree(struct buf *bp, int qindex)
1516 {
1517 	struct bufdomain *bd;
1518 	struct bufqueue *bq;
1519 
1520 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1521 	    ("binsfree: Invalid qindex %d", qindex));
1522 	BUF_ASSERT_XLOCKED(bp);
1523 
1524 	/*
1525 	 * Handle delayed bremfree() processing.
1526 	 */
1527 	if (bp->b_flags & B_REMFREE) {
1528 		if (bp->b_qindex == qindex) {
1529 			bp->b_flags |= B_REUSE;
1530 			bp->b_flags &= ~B_REMFREE;
1531 			BUF_UNLOCK(bp);
1532 			return;
1533 		}
1534 		bq = bufqueue_acquire(bp);
1535 		bq_remove(bq, bp);
1536 		BQ_UNLOCK(bq);
1537 	}
1538 	bd = bufdomain(bp);
1539 	if (qindex == QUEUE_CLEAN) {
1540 		if (bd->bd_lim != 0)
1541 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1542 		else
1543 			bq = bd->bd_cleanq;
1544 	} else
1545 		bq = &bd->bd_dirtyq;
1546 	bq_insert(bq, bp, true);
1547 }
1548 
1549 /*
1550  * buf_free:
1551  *
1552  *	Free a buffer to the buf zone once it no longer has valid contents.
1553  */
1554 static void
1555 buf_free(struct buf *bp)
1556 {
1557 
1558 	if (bp->b_flags & B_REMFREE)
1559 		bremfreef(bp);
1560 	if (bp->b_vflags & BV_BKGRDINPROG)
1561 		panic("losing buffer 1");
1562 	if (bp->b_rcred != NOCRED) {
1563 		crfree(bp->b_rcred);
1564 		bp->b_rcred = NOCRED;
1565 	}
1566 	if (bp->b_wcred != NOCRED) {
1567 		crfree(bp->b_wcred);
1568 		bp->b_wcred = NOCRED;
1569 	}
1570 	if (!LIST_EMPTY(&bp->b_dep))
1571 		buf_deallocate(bp);
1572 	bufkva_free(bp);
1573 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1574 	BUF_UNLOCK(bp);
1575 	uma_zfree(buf_zone, bp);
1576 }
1577 
1578 /*
1579  * buf_import:
1580  *
1581  *	Import bufs into the uma cache from the buf list.  The system still
1582  *	expects a static array of bufs and much of the synchronization
1583  *	around bufs assumes type stable storage.  As a result, UMA is used
1584  *	only as a per-cpu cache of bufs still maintained on a global list.
1585  */
1586 static int
1587 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1588 {
1589 	struct buf *bp;
1590 	int i;
1591 
1592 	BQ_LOCK(&bqempty);
1593 	for (i = 0; i < cnt; i++) {
1594 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1595 		if (bp == NULL)
1596 			break;
1597 		bq_remove(&bqempty, bp);
1598 		store[i] = bp;
1599 	}
1600 	BQ_UNLOCK(&bqempty);
1601 
1602 	return (i);
1603 }
1604 
1605 /*
1606  * buf_release:
1607  *
1608  *	Release bufs from the uma cache back to the buffer queues.
1609  */
1610 static void
1611 buf_release(void *arg, void **store, int cnt)
1612 {
1613 	struct bufqueue *bq;
1614 	struct buf *bp;
1615         int i;
1616 
1617 	bq = &bqempty;
1618 	BQ_LOCK(bq);
1619         for (i = 0; i < cnt; i++) {
1620 		bp = store[i];
1621 		/* Inline bq_insert() to batch locking. */
1622 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1623 		bp->b_flags &= ~(B_AGE | B_REUSE);
1624 		bq->bq_len++;
1625 		bp->b_qindex = bq->bq_index;
1626 	}
1627 	BQ_UNLOCK(bq);
1628 }
1629 
1630 /*
1631  * buf_alloc:
1632  *
1633  *	Allocate an empty buffer header.
1634  */
1635 static struct buf *
1636 buf_alloc(struct bufdomain *bd)
1637 {
1638 	struct buf *bp;
1639 	int freebufs, error;
1640 
1641 	/*
1642 	 * We can only run out of bufs in the buf zone if the average buf
1643 	 * is less than BKVASIZE.  In this case the actual wait/block will
1644 	 * come from buf_reycle() failing to flush one of these small bufs.
1645 	 */
1646 	bp = NULL;
1647 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1648 	if (freebufs > 0)
1649 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1650 	if (bp == NULL) {
1651 		atomic_add_int(&bd->bd_freebuffers, 1);
1652 		bufspace_daemon_wakeup(bd);
1653 		counter_u64_add(numbufallocfails, 1);
1654 		return (NULL);
1655 	}
1656 	/*
1657 	 * Wake-up the bufspace daemon on transition below threshold.
1658 	 */
1659 	if (freebufs == bd->bd_lofreebuffers)
1660 		bufspace_daemon_wakeup(bd);
1661 
1662 	error = BUF_LOCK(bp, LK_EXCLUSIVE, NULL);
1663 	KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1664 	    error));
1665 	(void)error;
1666 
1667 	KASSERT(bp->b_vp == NULL,
1668 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1669 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1670 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1671 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1672 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1673 	KASSERT(bp->b_npages == 0,
1674 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1675 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1676 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1677 
1678 	bp->b_domain = BD_DOMAIN(bd);
1679 	bp->b_flags = 0;
1680 	bp->b_ioflags = 0;
1681 	bp->b_xflags = 0;
1682 	bp->b_vflags = 0;
1683 	bp->b_vp = NULL;
1684 	bp->b_blkno = bp->b_lblkno = 0;
1685 	bp->b_offset = NOOFFSET;
1686 	bp->b_iodone = 0;
1687 	bp->b_error = 0;
1688 	bp->b_resid = 0;
1689 	bp->b_bcount = 0;
1690 	bp->b_npages = 0;
1691 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1692 	bp->b_bufobj = NULL;
1693 	bp->b_data = bp->b_kvabase = unmapped_buf;
1694 	bp->b_fsprivate1 = NULL;
1695 	bp->b_fsprivate2 = NULL;
1696 	bp->b_fsprivate3 = NULL;
1697 	LIST_INIT(&bp->b_dep);
1698 
1699 	return (bp);
1700 }
1701 
1702 /*
1703  *	buf_recycle:
1704  *
1705  *	Free a buffer from the given bufqueue.  kva controls whether the
1706  *	freed buf must own some kva resources.  This is used for
1707  *	defragmenting.
1708  */
1709 static int
1710 buf_recycle(struct bufdomain *bd, bool kva)
1711 {
1712 	struct bufqueue *bq;
1713 	struct buf *bp, *nbp;
1714 
1715 	if (kva)
1716 		counter_u64_add(bufdefragcnt, 1);
1717 	nbp = NULL;
1718 	bq = bd->bd_cleanq;
1719 	BQ_LOCK(bq);
1720 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1721 	    ("buf_recycle: Locks don't match"));
1722 	nbp = TAILQ_FIRST(&bq->bq_queue);
1723 
1724 	/*
1725 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1726 	 * depending.
1727 	 */
1728 	while ((bp = nbp) != NULL) {
1729 		/*
1730 		 * Calculate next bp (we can only use it if we do not
1731 		 * release the bqlock).
1732 		 */
1733 		nbp = TAILQ_NEXT(bp, b_freelist);
1734 
1735 		/*
1736 		 * If we are defragging then we need a buffer with
1737 		 * some kva to reclaim.
1738 		 */
1739 		if (kva && bp->b_kvasize == 0)
1740 			continue;
1741 
1742 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1743 			continue;
1744 
1745 		/*
1746 		 * Implement a second chance algorithm for frequently
1747 		 * accessed buffers.
1748 		 */
1749 		if ((bp->b_flags & B_REUSE) != 0) {
1750 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1751 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1752 			bp->b_flags &= ~B_REUSE;
1753 			BUF_UNLOCK(bp);
1754 			continue;
1755 		}
1756 
1757 		/*
1758 		 * Skip buffers with background writes in progress.
1759 		 */
1760 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1761 			BUF_UNLOCK(bp);
1762 			continue;
1763 		}
1764 
1765 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1766 		    ("buf_recycle: inconsistent queue %d bp %p",
1767 		    bp->b_qindex, bp));
1768 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1769 		    ("getnewbuf: queue domain %d doesn't match request %d",
1770 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1771 		/*
1772 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1773 		 * the scan from this point on.
1774 		 */
1775 		bq_remove(bq, bp);
1776 		BQ_UNLOCK(bq);
1777 
1778 		/*
1779 		 * Requeue the background write buffer with error and
1780 		 * restart the scan.
1781 		 */
1782 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1783 			bqrelse(bp);
1784 			BQ_LOCK(bq);
1785 			nbp = TAILQ_FIRST(&bq->bq_queue);
1786 			continue;
1787 		}
1788 		bp->b_flags |= B_INVAL;
1789 		brelse(bp);
1790 		return (0);
1791 	}
1792 	bd->bd_wanted = 1;
1793 	BQ_UNLOCK(bq);
1794 
1795 	return (ENOBUFS);
1796 }
1797 
1798 /*
1799  *	bremfree:
1800  *
1801  *	Mark the buffer for removal from the appropriate free list.
1802  *
1803  */
1804 void
1805 bremfree(struct buf *bp)
1806 {
1807 
1808 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1809 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1810 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1811 	KASSERT(bp->b_qindex != QUEUE_NONE,
1812 	    ("bremfree: buffer %p not on a queue.", bp));
1813 	BUF_ASSERT_XLOCKED(bp);
1814 
1815 	bp->b_flags |= B_REMFREE;
1816 }
1817 
1818 /*
1819  *	bremfreef:
1820  *
1821  *	Force an immediate removal from a free list.  Used only in nfs when
1822  *	it abuses the b_freelist pointer.
1823  */
1824 void
1825 bremfreef(struct buf *bp)
1826 {
1827 	struct bufqueue *bq;
1828 
1829 	bq = bufqueue_acquire(bp);
1830 	bq_remove(bq, bp);
1831 	BQ_UNLOCK(bq);
1832 }
1833 
1834 static void
1835 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1836 {
1837 
1838 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1839 	TAILQ_INIT(&bq->bq_queue);
1840 	bq->bq_len = 0;
1841 	bq->bq_index = qindex;
1842 	bq->bq_subqueue = subqueue;
1843 }
1844 
1845 static void
1846 bd_init(struct bufdomain *bd)
1847 {
1848 	int i;
1849 
1850 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1851 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1852 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1853 	for (i = 0; i <= mp_maxid; i++)
1854 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1855 		    "bufq clean subqueue lock");
1856 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1857 }
1858 
1859 /*
1860  *	bq_remove:
1861  *
1862  *	Removes a buffer from the free list, must be called with the
1863  *	correct qlock held.
1864  */
1865 static void
1866 bq_remove(struct bufqueue *bq, struct buf *bp)
1867 {
1868 
1869 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1870 	    bp, bp->b_vp, bp->b_flags);
1871 	KASSERT(bp->b_qindex != QUEUE_NONE,
1872 	    ("bq_remove: buffer %p not on a queue.", bp));
1873 	KASSERT(bufqueue(bp) == bq,
1874 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1875 
1876 	BQ_ASSERT_LOCKED(bq);
1877 	if (bp->b_qindex != QUEUE_EMPTY) {
1878 		BUF_ASSERT_XLOCKED(bp);
1879 	}
1880 	KASSERT(bq->bq_len >= 1,
1881 	    ("queue %d underflow", bp->b_qindex));
1882 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1883 	bq->bq_len--;
1884 	bp->b_qindex = QUEUE_NONE;
1885 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1886 }
1887 
1888 static void
1889 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1890 {
1891 	struct buf *bp;
1892 
1893 	BQ_ASSERT_LOCKED(bq);
1894 	if (bq != bd->bd_cleanq) {
1895 		BD_LOCK(bd);
1896 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1897 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1898 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1899 			    b_freelist);
1900 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1901 		}
1902 		bd->bd_cleanq->bq_len += bq->bq_len;
1903 		bq->bq_len = 0;
1904 	}
1905 	if (bd->bd_wanted) {
1906 		bd->bd_wanted = 0;
1907 		wakeup(&bd->bd_wanted);
1908 	}
1909 	if (bq != bd->bd_cleanq)
1910 		BD_UNLOCK(bd);
1911 }
1912 
1913 static int
1914 bd_flushall(struct bufdomain *bd)
1915 {
1916 	struct bufqueue *bq;
1917 	int flushed;
1918 	int i;
1919 
1920 	if (bd->bd_lim == 0)
1921 		return (0);
1922 	flushed = 0;
1923 	for (i = 0; i <= mp_maxid; i++) {
1924 		bq = &bd->bd_subq[i];
1925 		if (bq->bq_len == 0)
1926 			continue;
1927 		BQ_LOCK(bq);
1928 		bd_flush(bd, bq);
1929 		BQ_UNLOCK(bq);
1930 		flushed++;
1931 	}
1932 
1933 	return (flushed);
1934 }
1935 
1936 static void
1937 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1938 {
1939 	struct bufdomain *bd;
1940 
1941 	if (bp->b_qindex != QUEUE_NONE)
1942 		panic("bq_insert: free buffer %p onto another queue?", bp);
1943 
1944 	bd = bufdomain(bp);
1945 	if (bp->b_flags & B_AGE) {
1946 		/* Place this buf directly on the real queue. */
1947 		if (bq->bq_index == QUEUE_CLEAN)
1948 			bq = bd->bd_cleanq;
1949 		BQ_LOCK(bq);
1950 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1951 	} else {
1952 		BQ_LOCK(bq);
1953 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1954 	}
1955 	bp->b_flags &= ~(B_AGE | B_REUSE);
1956 	bq->bq_len++;
1957 	bp->b_qindex = bq->bq_index;
1958 	bp->b_subqueue = bq->bq_subqueue;
1959 
1960 	/*
1961 	 * Unlock before we notify so that we don't wakeup a waiter that
1962 	 * fails a trylock on the buf and sleeps again.
1963 	 */
1964 	if (unlock)
1965 		BUF_UNLOCK(bp);
1966 
1967 	if (bp->b_qindex == QUEUE_CLEAN) {
1968 		/*
1969 		 * Flush the per-cpu queue and notify any waiters.
1970 		 */
1971 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1972 		    bq->bq_len >= bd->bd_lim))
1973 			bd_flush(bd, bq);
1974 	}
1975 	BQ_UNLOCK(bq);
1976 }
1977 
1978 /*
1979  *	bufkva_free:
1980  *
1981  *	Free the kva allocation for a buffer.
1982  *
1983  */
1984 static void
1985 bufkva_free(struct buf *bp)
1986 {
1987 
1988 #ifdef INVARIANTS
1989 	if (bp->b_kvasize == 0) {
1990 		KASSERT(bp->b_kvabase == unmapped_buf &&
1991 		    bp->b_data == unmapped_buf,
1992 		    ("Leaked KVA space on %p", bp));
1993 	} else if (buf_mapped(bp))
1994 		BUF_CHECK_MAPPED(bp);
1995 	else
1996 		BUF_CHECK_UNMAPPED(bp);
1997 #endif
1998 	if (bp->b_kvasize == 0)
1999 		return;
2000 
2001 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2002 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2003 	counter_u64_add(buffreekvacnt, 1);
2004 	bp->b_data = bp->b_kvabase = unmapped_buf;
2005 	bp->b_kvasize = 0;
2006 }
2007 
2008 /*
2009  *	bufkva_alloc:
2010  *
2011  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2012  */
2013 static int
2014 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2015 {
2016 	vm_offset_t addr;
2017 	int error;
2018 
2019 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2020 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2021 
2022 	bufkva_free(bp);
2023 
2024 	addr = 0;
2025 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2026 	if (error != 0) {
2027 		/*
2028 		 * Buffer map is too fragmented.  Request the caller
2029 		 * to defragment the map.
2030 		 */
2031 		return (error);
2032 	}
2033 	bp->b_kvabase = (caddr_t)addr;
2034 	bp->b_kvasize = maxsize;
2035 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2036 	if ((gbflags & GB_UNMAPPED) != 0) {
2037 		bp->b_data = unmapped_buf;
2038 		BUF_CHECK_UNMAPPED(bp);
2039 	} else {
2040 		bp->b_data = bp->b_kvabase;
2041 		BUF_CHECK_MAPPED(bp);
2042 	}
2043 	return (0);
2044 }
2045 
2046 /*
2047  *	bufkva_reclaim:
2048  *
2049  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2050  *	callback that fires to avoid returning failure.
2051  */
2052 static void
2053 bufkva_reclaim(vmem_t *vmem, int flags)
2054 {
2055 	bool done;
2056 	int q;
2057 	int i;
2058 
2059 	done = false;
2060 	for (i = 0; i < 5; i++) {
2061 		for (q = 0; q < buf_domains; q++)
2062 			if (buf_recycle(&bdomain[q], true) != 0)
2063 				done = true;
2064 		if (done)
2065 			break;
2066 	}
2067 	return;
2068 }
2069 
2070 /*
2071  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2072  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2073  * the buffer is valid and we do not have to do anything.
2074  */
2075 static void
2076 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2077     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2078 {
2079 	struct buf *rabp;
2080 	struct thread *td;
2081 	int i;
2082 
2083 	td = curthread;
2084 
2085 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2086 		if (inmem(vp, *rablkno))
2087 			continue;
2088 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2089 		if ((rabp->b_flags & B_CACHE) != 0) {
2090 			brelse(rabp);
2091 			continue;
2092 		}
2093 #ifdef RACCT
2094 		if (racct_enable) {
2095 			PROC_LOCK(curproc);
2096 			racct_add_buf(curproc, rabp, 0);
2097 			PROC_UNLOCK(curproc);
2098 		}
2099 #endif /* RACCT */
2100 		td->td_ru.ru_inblock++;
2101 		rabp->b_flags |= B_ASYNC;
2102 		rabp->b_flags &= ~B_INVAL;
2103 		if ((flags & GB_CKHASH) != 0) {
2104 			rabp->b_flags |= B_CKHASH;
2105 			rabp->b_ckhashcalc = ckhashfunc;
2106 		}
2107 		rabp->b_ioflags &= ~BIO_ERROR;
2108 		rabp->b_iocmd = BIO_READ;
2109 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2110 			rabp->b_rcred = crhold(cred);
2111 		vfs_busy_pages(rabp, 0);
2112 		BUF_KERNPROC(rabp);
2113 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2114 		bstrategy(rabp);
2115 	}
2116 }
2117 
2118 /*
2119  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2120  *
2121  * Get a buffer with the specified data.  Look in the cache first.  We
2122  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2123  * is set, the buffer is valid and we do not have to do anything, see
2124  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2125  *
2126  * Always return a NULL buffer pointer (in bpp) when returning an error.
2127  *
2128  * The blkno parameter is the logical block being requested. Normally
2129  * the mapping of logical block number to disk block address is done
2130  * by calling VOP_BMAP(). However, if the mapping is already known, the
2131  * disk block address can be passed using the dblkno parameter. If the
2132  * disk block address is not known, then the same value should be passed
2133  * for blkno and dblkno.
2134  */
2135 int
2136 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2137     daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2138     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2139 {
2140 	struct buf *bp;
2141 	struct thread *td;
2142 	int error, readwait, rv;
2143 
2144 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2145 	td = curthread;
2146 	/*
2147 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2148 	 * are specified.
2149 	 */
2150 	error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2151 	if (error != 0) {
2152 		*bpp = NULL;
2153 		return (error);
2154 	}
2155 	KASSERT(blkno == bp->b_lblkno,
2156 	    ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2157 	    (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2158 	flags &= ~GB_NOSPARSE;
2159 	*bpp = bp;
2160 
2161 	/*
2162 	 * If not found in cache, do some I/O
2163 	 */
2164 	readwait = 0;
2165 	if ((bp->b_flags & B_CACHE) == 0) {
2166 #ifdef RACCT
2167 		if (racct_enable) {
2168 			PROC_LOCK(td->td_proc);
2169 			racct_add_buf(td->td_proc, bp, 0);
2170 			PROC_UNLOCK(td->td_proc);
2171 		}
2172 #endif /* RACCT */
2173 		td->td_ru.ru_inblock++;
2174 		bp->b_iocmd = BIO_READ;
2175 		bp->b_flags &= ~B_INVAL;
2176 		if ((flags & GB_CKHASH) != 0) {
2177 			bp->b_flags |= B_CKHASH;
2178 			bp->b_ckhashcalc = ckhashfunc;
2179 		}
2180 		if ((flags & GB_CVTENXIO) != 0)
2181 			bp->b_xflags |= BX_CVTENXIO;
2182 		bp->b_ioflags &= ~BIO_ERROR;
2183 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2184 			bp->b_rcred = crhold(cred);
2185 		vfs_busy_pages(bp, 0);
2186 		bp->b_iooffset = dbtob(bp->b_blkno);
2187 		bstrategy(bp);
2188 		++readwait;
2189 	}
2190 
2191 	/*
2192 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2193 	 */
2194 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2195 
2196 	rv = 0;
2197 	if (readwait) {
2198 		rv = bufwait(bp);
2199 		if (rv != 0) {
2200 			brelse(bp);
2201 			*bpp = NULL;
2202 		}
2203 	}
2204 	return (rv);
2205 }
2206 
2207 /*
2208  * Write, release buffer on completion.  (Done by iodone
2209  * if async).  Do not bother writing anything if the buffer
2210  * is invalid.
2211  *
2212  * Note that we set B_CACHE here, indicating that buffer is
2213  * fully valid and thus cacheable.  This is true even of NFS
2214  * now so we set it generally.  This could be set either here
2215  * or in biodone() since the I/O is synchronous.  We put it
2216  * here.
2217  */
2218 int
2219 bufwrite(struct buf *bp)
2220 {
2221 	int oldflags;
2222 	struct vnode *vp;
2223 	long space;
2224 	int vp_md;
2225 
2226 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2227 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2228 		bp->b_flags |= B_INVAL | B_RELBUF;
2229 		bp->b_flags &= ~B_CACHE;
2230 		brelse(bp);
2231 		return (ENXIO);
2232 	}
2233 	if (bp->b_flags & B_INVAL) {
2234 		brelse(bp);
2235 		return (0);
2236 	}
2237 
2238 	if (bp->b_flags & B_BARRIER)
2239 		atomic_add_long(&barrierwrites, 1);
2240 
2241 	oldflags = bp->b_flags;
2242 
2243 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2244 	    ("FFS background buffer should not get here %p", bp));
2245 
2246 	vp = bp->b_vp;
2247 	if (vp)
2248 		vp_md = vp->v_vflag & VV_MD;
2249 	else
2250 		vp_md = 0;
2251 
2252 	/*
2253 	 * Mark the buffer clean.  Increment the bufobj write count
2254 	 * before bundirty() call, to prevent other thread from seeing
2255 	 * empty dirty list and zero counter for writes in progress,
2256 	 * falsely indicating that the bufobj is clean.
2257 	 */
2258 	bufobj_wref(bp->b_bufobj);
2259 	bundirty(bp);
2260 
2261 	bp->b_flags &= ~B_DONE;
2262 	bp->b_ioflags &= ~BIO_ERROR;
2263 	bp->b_flags |= B_CACHE;
2264 	bp->b_iocmd = BIO_WRITE;
2265 
2266 	vfs_busy_pages(bp, 1);
2267 
2268 	/*
2269 	 * Normal bwrites pipeline writes
2270 	 */
2271 	bp->b_runningbufspace = bp->b_bufsize;
2272 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2273 
2274 #ifdef RACCT
2275 	if (racct_enable) {
2276 		PROC_LOCK(curproc);
2277 		racct_add_buf(curproc, bp, 1);
2278 		PROC_UNLOCK(curproc);
2279 	}
2280 #endif /* RACCT */
2281 	curthread->td_ru.ru_oublock++;
2282 	if (oldflags & B_ASYNC)
2283 		BUF_KERNPROC(bp);
2284 	bp->b_iooffset = dbtob(bp->b_blkno);
2285 	buf_track(bp, __func__);
2286 	bstrategy(bp);
2287 
2288 	if ((oldflags & B_ASYNC) == 0) {
2289 		int rtval = bufwait(bp);
2290 		brelse(bp);
2291 		return (rtval);
2292 	} else if (space > hirunningspace) {
2293 		/*
2294 		 * don't allow the async write to saturate the I/O
2295 		 * system.  We will not deadlock here because
2296 		 * we are blocking waiting for I/O that is already in-progress
2297 		 * to complete. We do not block here if it is the update
2298 		 * or syncer daemon trying to clean up as that can lead
2299 		 * to deadlock.
2300 		 */
2301 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2302 			waitrunningbufspace();
2303 	}
2304 
2305 	return (0);
2306 }
2307 
2308 void
2309 bufbdflush(struct bufobj *bo, struct buf *bp)
2310 {
2311 	struct buf *nbp;
2312 
2313 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2314 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2315 		altbufferflushes++;
2316 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2317 		BO_LOCK(bo);
2318 		/*
2319 		 * Try to find a buffer to flush.
2320 		 */
2321 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2322 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2323 			    BUF_LOCK(nbp,
2324 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2325 				continue;
2326 			if (bp == nbp)
2327 				panic("bdwrite: found ourselves");
2328 			BO_UNLOCK(bo);
2329 			/* Don't countdeps with the bo lock held. */
2330 			if (buf_countdeps(nbp, 0)) {
2331 				BO_LOCK(bo);
2332 				BUF_UNLOCK(nbp);
2333 				continue;
2334 			}
2335 			if (nbp->b_flags & B_CLUSTEROK) {
2336 				vfs_bio_awrite(nbp);
2337 			} else {
2338 				bremfree(nbp);
2339 				bawrite(nbp);
2340 			}
2341 			dirtybufferflushes++;
2342 			break;
2343 		}
2344 		if (nbp == NULL)
2345 			BO_UNLOCK(bo);
2346 	}
2347 }
2348 
2349 /*
2350  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2351  * anything if the buffer is marked invalid.
2352  *
2353  * Note that since the buffer must be completely valid, we can safely
2354  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2355  * biodone() in order to prevent getblk from writing the buffer
2356  * out synchronously.
2357  */
2358 void
2359 bdwrite(struct buf *bp)
2360 {
2361 	struct thread *td = curthread;
2362 	struct vnode *vp;
2363 	struct bufobj *bo;
2364 
2365 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2366 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2367 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2368 	    ("Barrier request in delayed write %p", bp));
2369 
2370 	if (bp->b_flags & B_INVAL) {
2371 		brelse(bp);
2372 		return;
2373 	}
2374 
2375 	/*
2376 	 * If we have too many dirty buffers, don't create any more.
2377 	 * If we are wildly over our limit, then force a complete
2378 	 * cleanup. Otherwise, just keep the situation from getting
2379 	 * out of control. Note that we have to avoid a recursive
2380 	 * disaster and not try to clean up after our own cleanup!
2381 	 */
2382 	vp = bp->b_vp;
2383 	bo = bp->b_bufobj;
2384 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2385 		td->td_pflags |= TDP_INBDFLUSH;
2386 		BO_BDFLUSH(bo, bp);
2387 		td->td_pflags &= ~TDP_INBDFLUSH;
2388 	} else
2389 		recursiveflushes++;
2390 
2391 	bdirty(bp);
2392 	/*
2393 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2394 	 * true even of NFS now.
2395 	 */
2396 	bp->b_flags |= B_CACHE;
2397 
2398 	/*
2399 	 * This bmap keeps the system from needing to do the bmap later,
2400 	 * perhaps when the system is attempting to do a sync.  Since it
2401 	 * is likely that the indirect block -- or whatever other datastructure
2402 	 * that the filesystem needs is still in memory now, it is a good
2403 	 * thing to do this.  Note also, that if the pageout daemon is
2404 	 * requesting a sync -- there might not be enough memory to do
2405 	 * the bmap then...  So, this is important to do.
2406 	 */
2407 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2408 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2409 	}
2410 
2411 	buf_track(bp, __func__);
2412 
2413 	/*
2414 	 * Set the *dirty* buffer range based upon the VM system dirty
2415 	 * pages.
2416 	 *
2417 	 * Mark the buffer pages as clean.  We need to do this here to
2418 	 * satisfy the vnode_pager and the pageout daemon, so that it
2419 	 * thinks that the pages have been "cleaned".  Note that since
2420 	 * the pages are in a delayed write buffer -- the VFS layer
2421 	 * "will" see that the pages get written out on the next sync,
2422 	 * or perhaps the cluster will be completed.
2423 	 */
2424 	vfs_clean_pages_dirty_buf(bp);
2425 	bqrelse(bp);
2426 
2427 	/*
2428 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2429 	 * due to the softdep code.
2430 	 */
2431 }
2432 
2433 /*
2434  *	bdirty:
2435  *
2436  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2437  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2438  *	itself to properly update it in the dirty/clean lists.  We mark it
2439  *	B_DONE to ensure that any asynchronization of the buffer properly
2440  *	clears B_DONE ( else a panic will occur later ).
2441  *
2442  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2443  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2444  *	should only be called if the buffer is known-good.
2445  *
2446  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2447  *	count.
2448  *
2449  *	The buffer must be on QUEUE_NONE.
2450  */
2451 void
2452 bdirty(struct buf *bp)
2453 {
2454 
2455 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2456 	    bp, bp->b_vp, bp->b_flags);
2457 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2458 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2459 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2460 	bp->b_flags &= ~(B_RELBUF);
2461 	bp->b_iocmd = BIO_WRITE;
2462 
2463 	if ((bp->b_flags & B_DELWRI) == 0) {
2464 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2465 		reassignbuf(bp);
2466 		bdirtyadd(bp);
2467 	}
2468 }
2469 
2470 /*
2471  *	bundirty:
2472  *
2473  *	Clear B_DELWRI for buffer.
2474  *
2475  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2476  *	count.
2477  *
2478  *	The buffer must be on QUEUE_NONE.
2479  */
2480 
2481 void
2482 bundirty(struct buf *bp)
2483 {
2484 
2485 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2486 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2487 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2488 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2489 
2490 	if (bp->b_flags & B_DELWRI) {
2491 		bp->b_flags &= ~B_DELWRI;
2492 		reassignbuf(bp);
2493 		bdirtysub(bp);
2494 	}
2495 	/*
2496 	 * Since it is now being written, we can clear its deferred write flag.
2497 	 */
2498 	bp->b_flags &= ~B_DEFERRED;
2499 }
2500 
2501 /*
2502  *	bawrite:
2503  *
2504  *	Asynchronous write.  Start output on a buffer, but do not wait for
2505  *	it to complete.  The buffer is released when the output completes.
2506  *
2507  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2508  *	B_INVAL buffers.  Not us.
2509  */
2510 void
2511 bawrite(struct buf *bp)
2512 {
2513 
2514 	bp->b_flags |= B_ASYNC;
2515 	(void) bwrite(bp);
2516 }
2517 
2518 /*
2519  *	babarrierwrite:
2520  *
2521  *	Asynchronous barrier write.  Start output on a buffer, but do not
2522  *	wait for it to complete.  Place a write barrier after this write so
2523  *	that this buffer and all buffers written before it are committed to
2524  *	the disk before any buffers written after this write are committed
2525  *	to the disk.  The buffer is released when the output completes.
2526  */
2527 void
2528 babarrierwrite(struct buf *bp)
2529 {
2530 
2531 	bp->b_flags |= B_ASYNC | B_BARRIER;
2532 	(void) bwrite(bp);
2533 }
2534 
2535 /*
2536  *	bbarrierwrite:
2537  *
2538  *	Synchronous barrier write.  Start output on a buffer and wait for
2539  *	it to complete.  Place a write barrier after this write so that
2540  *	this buffer and all buffers written before it are committed to
2541  *	the disk before any buffers written after this write are committed
2542  *	to the disk.  The buffer is released when the output completes.
2543  */
2544 int
2545 bbarrierwrite(struct buf *bp)
2546 {
2547 
2548 	bp->b_flags |= B_BARRIER;
2549 	return (bwrite(bp));
2550 }
2551 
2552 /*
2553  *	bwillwrite:
2554  *
2555  *	Called prior to the locking of any vnodes when we are expecting to
2556  *	write.  We do not want to starve the buffer cache with too many
2557  *	dirty buffers so we block here.  By blocking prior to the locking
2558  *	of any vnodes we attempt to avoid the situation where a locked vnode
2559  *	prevents the various system daemons from flushing related buffers.
2560  */
2561 void
2562 bwillwrite(void)
2563 {
2564 
2565 	if (buf_dirty_count_severe()) {
2566 		mtx_lock(&bdirtylock);
2567 		while (buf_dirty_count_severe()) {
2568 			bdirtywait = 1;
2569 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2570 			    "flswai", 0);
2571 		}
2572 		mtx_unlock(&bdirtylock);
2573 	}
2574 }
2575 
2576 /*
2577  * Return true if we have too many dirty buffers.
2578  */
2579 int
2580 buf_dirty_count_severe(void)
2581 {
2582 
2583 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2584 }
2585 
2586 /*
2587  *	brelse:
2588  *
2589  *	Release a busy buffer and, if requested, free its resources.  The
2590  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2591  *	to be accessed later as a cache entity or reused for other purposes.
2592  */
2593 void
2594 brelse(struct buf *bp)
2595 {
2596 	struct mount *v_mnt;
2597 	int qindex;
2598 
2599 	/*
2600 	 * Many functions erroneously call brelse with a NULL bp under rare
2601 	 * error conditions. Simply return when called with a NULL bp.
2602 	 */
2603 	if (bp == NULL)
2604 		return;
2605 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2606 	    bp, bp->b_vp, bp->b_flags);
2607 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2608 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2609 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2610 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2611 
2612 	if (BUF_LOCKRECURSED(bp)) {
2613 		/*
2614 		 * Do not process, in particular, do not handle the
2615 		 * B_INVAL/B_RELBUF and do not release to free list.
2616 		 */
2617 		BUF_UNLOCK(bp);
2618 		return;
2619 	}
2620 
2621 	if (bp->b_flags & B_MANAGED) {
2622 		bqrelse(bp);
2623 		return;
2624 	}
2625 
2626 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2627 		BO_LOCK(bp->b_bufobj);
2628 		bp->b_vflags &= ~BV_BKGRDERR;
2629 		BO_UNLOCK(bp->b_bufobj);
2630 		bdirty(bp);
2631 	}
2632 
2633 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2634 	    (bp->b_flags & B_INVALONERR)) {
2635 		/*
2636 		 * Forced invalidation of dirty buffer contents, to be used
2637 		 * after a failed write in the rare case that the loss of the
2638 		 * contents is acceptable.  The buffer is invalidated and
2639 		 * freed.
2640 		 */
2641 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2642 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2643 	}
2644 
2645 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2646 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2647 	    !(bp->b_flags & B_INVAL)) {
2648 		/*
2649 		 * Failed write, redirty.  All errors except ENXIO (which
2650 		 * means the device is gone) are treated as being
2651 		 * transient.
2652 		 *
2653 		 * XXX Treating EIO as transient is not correct; the
2654 		 * contract with the local storage device drivers is that
2655 		 * they will only return EIO once the I/O is no longer
2656 		 * retriable.  Network I/O also respects this through the
2657 		 * guarantees of TCP and/or the internal retries of NFS.
2658 		 * ENOMEM might be transient, but we also have no way of
2659 		 * knowing when its ok to retry/reschedule.  In general,
2660 		 * this entire case should be made obsolete through better
2661 		 * error handling/recovery and resource scheduling.
2662 		 *
2663 		 * Do this also for buffers that failed with ENXIO, but have
2664 		 * non-empty dependencies - the soft updates code might need
2665 		 * to access the buffer to untangle them.
2666 		 *
2667 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2668 		 */
2669 		bp->b_ioflags &= ~BIO_ERROR;
2670 		bdirty(bp);
2671 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2672 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2673 		/*
2674 		 * Either a failed read I/O, or we were asked to free or not
2675 		 * cache the buffer, or we failed to write to a device that's
2676 		 * no longer present.
2677 		 */
2678 		bp->b_flags |= B_INVAL;
2679 		if (!LIST_EMPTY(&bp->b_dep))
2680 			buf_deallocate(bp);
2681 		if (bp->b_flags & B_DELWRI)
2682 			bdirtysub(bp);
2683 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2684 		if ((bp->b_flags & B_VMIO) == 0) {
2685 			allocbuf(bp, 0);
2686 			if (bp->b_vp)
2687 				brelvp(bp);
2688 		}
2689 	}
2690 
2691 	/*
2692 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2693 	 * is called with B_DELWRI set, the underlying pages may wind up
2694 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2695 	 * because pages associated with a B_DELWRI bp are marked clean.
2696 	 *
2697 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2698 	 * if B_DELWRI is set.
2699 	 */
2700 	if (bp->b_flags & B_DELWRI)
2701 		bp->b_flags &= ~B_RELBUF;
2702 
2703 	/*
2704 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2705 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2706 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2707 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2708 	 *
2709 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2710 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2711 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2712 	 *
2713 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2714 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2715 	 * the commit state and we cannot afford to lose the buffer. If the
2716 	 * buffer has a background write in progress, we need to keep it
2717 	 * around to prevent it from being reconstituted and starting a second
2718 	 * background write.
2719 	 */
2720 
2721 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2722 
2723 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2724 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2725 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2726 	    vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2727 		vfs_vmio_invalidate(bp);
2728 		allocbuf(bp, 0);
2729 	}
2730 
2731 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2732 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2733 		allocbuf(bp, 0);
2734 		bp->b_flags &= ~B_NOREUSE;
2735 		if (bp->b_vp != NULL)
2736 			brelvp(bp);
2737 	}
2738 
2739 	/*
2740 	 * If the buffer has junk contents signal it and eventually
2741 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2742 	 * doesn't find it.
2743 	 */
2744 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2745 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2746 		bp->b_flags |= B_INVAL;
2747 	if (bp->b_flags & B_INVAL) {
2748 		if (bp->b_flags & B_DELWRI)
2749 			bundirty(bp);
2750 		if (bp->b_vp)
2751 			brelvp(bp);
2752 	}
2753 
2754 	buf_track(bp, __func__);
2755 
2756 	/* buffers with no memory */
2757 	if (bp->b_bufsize == 0) {
2758 		buf_free(bp);
2759 		return;
2760 	}
2761 	/* buffers with junk contents */
2762 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2763 	    (bp->b_ioflags & BIO_ERROR)) {
2764 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2765 		if (bp->b_vflags & BV_BKGRDINPROG)
2766 			panic("losing buffer 2");
2767 		qindex = QUEUE_CLEAN;
2768 		bp->b_flags |= B_AGE;
2769 	/* remaining buffers */
2770 	} else if (bp->b_flags & B_DELWRI)
2771 		qindex = QUEUE_DIRTY;
2772 	else
2773 		qindex = QUEUE_CLEAN;
2774 
2775 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2776 		panic("brelse: not dirty");
2777 
2778 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2779 	bp->b_xflags &= ~(BX_CVTENXIO);
2780 	/* binsfree unlocks bp. */
2781 	binsfree(bp, qindex);
2782 }
2783 
2784 /*
2785  * Release a buffer back to the appropriate queue but do not try to free
2786  * it.  The buffer is expected to be used again soon.
2787  *
2788  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2789  * biodone() to requeue an async I/O on completion.  It is also used when
2790  * known good buffers need to be requeued but we think we may need the data
2791  * again soon.
2792  *
2793  * XXX we should be able to leave the B_RELBUF hint set on completion.
2794  */
2795 void
2796 bqrelse(struct buf *bp)
2797 {
2798 	int qindex;
2799 
2800 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2801 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2802 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2803 
2804 	qindex = QUEUE_NONE;
2805 	if (BUF_LOCKRECURSED(bp)) {
2806 		/* do not release to free list */
2807 		BUF_UNLOCK(bp);
2808 		return;
2809 	}
2810 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2811 	bp->b_xflags &= ~(BX_CVTENXIO);
2812 
2813 	if (bp->b_flags & B_MANAGED) {
2814 		if (bp->b_flags & B_REMFREE)
2815 			bremfreef(bp);
2816 		goto out;
2817 	}
2818 
2819 	/* buffers with stale but valid contents */
2820 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2821 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2822 		BO_LOCK(bp->b_bufobj);
2823 		bp->b_vflags &= ~BV_BKGRDERR;
2824 		BO_UNLOCK(bp->b_bufobj);
2825 		qindex = QUEUE_DIRTY;
2826 	} else {
2827 		if ((bp->b_flags & B_DELWRI) == 0 &&
2828 		    (bp->b_xflags & BX_VNDIRTY))
2829 			panic("bqrelse: not dirty");
2830 		if ((bp->b_flags & B_NOREUSE) != 0) {
2831 			brelse(bp);
2832 			return;
2833 		}
2834 		qindex = QUEUE_CLEAN;
2835 	}
2836 	buf_track(bp, __func__);
2837 	/* binsfree unlocks bp. */
2838 	binsfree(bp, qindex);
2839 	return;
2840 
2841 out:
2842 	buf_track(bp, __func__);
2843 	/* unlock */
2844 	BUF_UNLOCK(bp);
2845 }
2846 
2847 /*
2848  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2849  * restore bogus pages.
2850  */
2851 static void
2852 vfs_vmio_iodone(struct buf *bp)
2853 {
2854 	vm_ooffset_t foff;
2855 	vm_page_t m;
2856 	vm_object_t obj;
2857 	struct vnode *vp __unused;
2858 	int i, iosize, resid;
2859 	bool bogus;
2860 
2861 	obj = bp->b_bufobj->bo_object;
2862 	KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2863 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2864 	    blockcount_read(&obj->paging_in_progress), bp->b_npages));
2865 
2866 	vp = bp->b_vp;
2867 	VNPASS(vp->v_holdcnt > 0, vp);
2868 	VNPASS(vp->v_object != NULL, vp);
2869 
2870 	foff = bp->b_offset;
2871 	KASSERT(bp->b_offset != NOOFFSET,
2872 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2873 
2874 	bogus = false;
2875 	iosize = bp->b_bcount - bp->b_resid;
2876 	for (i = 0; i < bp->b_npages; i++) {
2877 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2878 		if (resid > iosize)
2879 			resid = iosize;
2880 
2881 		/*
2882 		 * cleanup bogus pages, restoring the originals
2883 		 */
2884 		m = bp->b_pages[i];
2885 		if (m == bogus_page) {
2886 			bogus = true;
2887 			m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2888 			if (m == NULL)
2889 				panic("biodone: page disappeared!");
2890 			bp->b_pages[i] = m;
2891 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2892 			/*
2893 			 * In the write case, the valid and clean bits are
2894 			 * already changed correctly ( see bdwrite() ), so we
2895 			 * only need to do this here in the read case.
2896 			 */
2897 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2898 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2899 			    "has unexpected dirty bits", m));
2900 			vfs_page_set_valid(bp, foff, m);
2901 		}
2902 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2903 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2904 		    (intmax_t)foff, (uintmax_t)m->pindex));
2905 
2906 		vm_page_sunbusy(m);
2907 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2908 		iosize -= resid;
2909 	}
2910 	vm_object_pip_wakeupn(obj, bp->b_npages);
2911 	if (bogus && buf_mapped(bp)) {
2912 		BUF_CHECK_MAPPED(bp);
2913 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2914 		    bp->b_pages, bp->b_npages);
2915 	}
2916 }
2917 
2918 /*
2919  * Perform page invalidation when a buffer is released.  The fully invalid
2920  * pages will be reclaimed later in vfs_vmio_truncate().
2921  */
2922 static void
2923 vfs_vmio_invalidate(struct buf *bp)
2924 {
2925 	vm_object_t obj;
2926 	vm_page_t m;
2927 	int flags, i, resid, poffset, presid;
2928 
2929 	if (buf_mapped(bp)) {
2930 		BUF_CHECK_MAPPED(bp);
2931 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2932 	} else
2933 		BUF_CHECK_UNMAPPED(bp);
2934 	/*
2935 	 * Get the base offset and length of the buffer.  Note that
2936 	 * in the VMIO case if the buffer block size is not
2937 	 * page-aligned then b_data pointer may not be page-aligned.
2938 	 * But our b_pages[] array *IS* page aligned.
2939 	 *
2940 	 * block sizes less then DEV_BSIZE (usually 512) are not
2941 	 * supported due to the page granularity bits (m->valid,
2942 	 * m->dirty, etc...).
2943 	 *
2944 	 * See man buf(9) for more information
2945 	 */
2946 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2947 	obj = bp->b_bufobj->bo_object;
2948 	resid = bp->b_bufsize;
2949 	poffset = bp->b_offset & PAGE_MASK;
2950 	VM_OBJECT_WLOCK(obj);
2951 	for (i = 0; i < bp->b_npages; i++) {
2952 		m = bp->b_pages[i];
2953 		if (m == bogus_page)
2954 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2955 		bp->b_pages[i] = NULL;
2956 
2957 		presid = resid > (PAGE_SIZE - poffset) ?
2958 		    (PAGE_SIZE - poffset) : resid;
2959 		KASSERT(presid >= 0, ("brelse: extra page"));
2960 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
2961 		if (pmap_page_wired_mappings(m) == 0)
2962 			vm_page_set_invalid(m, poffset, presid);
2963 		vm_page_sunbusy(m);
2964 		vm_page_release_locked(m, flags);
2965 		resid -= presid;
2966 		poffset = 0;
2967 	}
2968 	VM_OBJECT_WUNLOCK(obj);
2969 	bp->b_npages = 0;
2970 }
2971 
2972 /*
2973  * Page-granular truncation of an existing VMIO buffer.
2974  */
2975 static void
2976 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2977 {
2978 	vm_object_t obj;
2979 	vm_page_t m;
2980 	int flags, i;
2981 
2982 	if (bp->b_npages == desiredpages)
2983 		return;
2984 
2985 	if (buf_mapped(bp)) {
2986 		BUF_CHECK_MAPPED(bp);
2987 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2988 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2989 	} else
2990 		BUF_CHECK_UNMAPPED(bp);
2991 
2992 	/*
2993 	 * The object lock is needed only if we will attempt to free pages.
2994 	 */
2995 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2996 	if ((bp->b_flags & B_DIRECT) != 0) {
2997 		flags |= VPR_TRYFREE;
2998 		obj = bp->b_bufobj->bo_object;
2999 		VM_OBJECT_WLOCK(obj);
3000 	} else {
3001 		obj = NULL;
3002 	}
3003 	for (i = desiredpages; i < bp->b_npages; i++) {
3004 		m = bp->b_pages[i];
3005 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3006 		bp->b_pages[i] = NULL;
3007 		if (obj != NULL)
3008 			vm_page_release_locked(m, flags);
3009 		else
3010 			vm_page_release(m, flags);
3011 	}
3012 	if (obj != NULL)
3013 		VM_OBJECT_WUNLOCK(obj);
3014 	bp->b_npages = desiredpages;
3015 }
3016 
3017 /*
3018  * Byte granular extension of VMIO buffers.
3019  */
3020 static void
3021 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3022 {
3023 	/*
3024 	 * We are growing the buffer, possibly in a
3025 	 * byte-granular fashion.
3026 	 */
3027 	vm_object_t obj;
3028 	vm_offset_t toff;
3029 	vm_offset_t tinc;
3030 	vm_page_t m;
3031 
3032 	/*
3033 	 * Step 1, bring in the VM pages from the object, allocating
3034 	 * them if necessary.  We must clear B_CACHE if these pages
3035 	 * are not valid for the range covered by the buffer.
3036 	 */
3037 	obj = bp->b_bufobj->bo_object;
3038 	if (bp->b_npages < desiredpages) {
3039 		/*
3040 		 * We must allocate system pages since blocking
3041 		 * here could interfere with paging I/O, no
3042 		 * matter which process we are.
3043 		 *
3044 		 * Only exclusive busy can be tested here.
3045 		 * Blocking on shared busy might lead to
3046 		 * deadlocks once allocbuf() is called after
3047 		 * pages are vfs_busy_pages().
3048 		 */
3049 		(void)vm_page_grab_pages_unlocked(obj,
3050 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3051 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3052 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3053 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3054 		bp->b_npages = desiredpages;
3055 	}
3056 
3057 	/*
3058 	 * Step 2.  We've loaded the pages into the buffer,
3059 	 * we have to figure out if we can still have B_CACHE
3060 	 * set.  Note that B_CACHE is set according to the
3061 	 * byte-granular range ( bcount and size ), not the
3062 	 * aligned range ( newbsize ).
3063 	 *
3064 	 * The VM test is against m->valid, which is DEV_BSIZE
3065 	 * aligned.  Needless to say, the validity of the data
3066 	 * needs to also be DEV_BSIZE aligned.  Note that this
3067 	 * fails with NFS if the server or some other client
3068 	 * extends the file's EOF.  If our buffer is resized,
3069 	 * B_CACHE may remain set! XXX
3070 	 */
3071 	toff = bp->b_bcount;
3072 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3073 	while ((bp->b_flags & B_CACHE) && toff < size) {
3074 		vm_pindex_t pi;
3075 
3076 		if (tinc > (size - toff))
3077 			tinc = size - toff;
3078 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3079 		m = bp->b_pages[pi];
3080 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3081 		toff += tinc;
3082 		tinc = PAGE_SIZE;
3083 	}
3084 
3085 	/*
3086 	 * Step 3, fixup the KVA pmap.
3087 	 */
3088 	if (buf_mapped(bp))
3089 		bpmap_qenter(bp);
3090 	else
3091 		BUF_CHECK_UNMAPPED(bp);
3092 }
3093 
3094 /*
3095  * Check to see if a block at a particular lbn is available for a clustered
3096  * write.
3097  */
3098 static int
3099 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3100 {
3101 	struct buf *bpa;
3102 	int match;
3103 
3104 	match = 0;
3105 
3106 	/* If the buf isn't in core skip it */
3107 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3108 		return (0);
3109 
3110 	/* If the buf is busy we don't want to wait for it */
3111 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3112 		return (0);
3113 
3114 	/* Only cluster with valid clusterable delayed write buffers */
3115 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3116 	    (B_DELWRI | B_CLUSTEROK))
3117 		goto done;
3118 
3119 	if (bpa->b_bufsize != size)
3120 		goto done;
3121 
3122 	/*
3123 	 * Check to see if it is in the expected place on disk and that the
3124 	 * block has been mapped.
3125 	 */
3126 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3127 		match = 1;
3128 done:
3129 	BUF_UNLOCK(bpa);
3130 	return (match);
3131 }
3132 
3133 /*
3134  *	vfs_bio_awrite:
3135  *
3136  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3137  *	This is much better then the old way of writing only one buffer at
3138  *	a time.  Note that we may not be presented with the buffers in the
3139  *	correct order, so we search for the cluster in both directions.
3140  */
3141 int
3142 vfs_bio_awrite(struct buf *bp)
3143 {
3144 	struct bufobj *bo;
3145 	int i;
3146 	int j;
3147 	daddr_t lblkno = bp->b_lblkno;
3148 	struct vnode *vp = bp->b_vp;
3149 	int ncl;
3150 	int nwritten;
3151 	int size;
3152 	int maxcl;
3153 	int gbflags;
3154 
3155 	bo = &vp->v_bufobj;
3156 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3157 	/*
3158 	 * right now we support clustered writing only to regular files.  If
3159 	 * we find a clusterable block we could be in the middle of a cluster
3160 	 * rather then at the beginning.
3161 	 */
3162 	if ((vp->v_type == VREG) &&
3163 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3164 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3165 		size = vp->v_mount->mnt_stat.f_iosize;
3166 		maxcl = MAXPHYS / size;
3167 
3168 		BO_RLOCK(bo);
3169 		for (i = 1; i < maxcl; i++)
3170 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3171 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3172 				break;
3173 
3174 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3175 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3176 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3177 				break;
3178 		BO_RUNLOCK(bo);
3179 		--j;
3180 		ncl = i + j;
3181 		/*
3182 		 * this is a possible cluster write
3183 		 */
3184 		if (ncl != 1) {
3185 			BUF_UNLOCK(bp);
3186 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3187 			    gbflags);
3188 			return (nwritten);
3189 		}
3190 	}
3191 	bremfree(bp);
3192 	bp->b_flags |= B_ASYNC;
3193 	/*
3194 	 * default (old) behavior, writing out only one block
3195 	 *
3196 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3197 	 */
3198 	nwritten = bp->b_bufsize;
3199 	(void) bwrite(bp);
3200 
3201 	return (nwritten);
3202 }
3203 
3204 /*
3205  *	getnewbuf_kva:
3206  *
3207  *	Allocate KVA for an empty buf header according to gbflags.
3208  */
3209 static int
3210 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3211 {
3212 
3213 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3214 		/*
3215 		 * In order to keep fragmentation sane we only allocate kva
3216 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3217 		 */
3218 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3219 
3220 		if (maxsize != bp->b_kvasize &&
3221 		    bufkva_alloc(bp, maxsize, gbflags))
3222 			return (ENOSPC);
3223 	}
3224 	return (0);
3225 }
3226 
3227 /*
3228  *	getnewbuf:
3229  *
3230  *	Find and initialize a new buffer header, freeing up existing buffers
3231  *	in the bufqueues as necessary.  The new buffer is returned locked.
3232  *
3233  *	We block if:
3234  *		We have insufficient buffer headers
3235  *		We have insufficient buffer space
3236  *		buffer_arena is too fragmented ( space reservation fails )
3237  *		If we have to flush dirty buffers ( but we try to avoid this )
3238  *
3239  *	The caller is responsible for releasing the reserved bufspace after
3240  *	allocbuf() is called.
3241  */
3242 static struct buf *
3243 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3244 {
3245 	struct bufdomain *bd;
3246 	struct buf *bp;
3247 	bool metadata, reserved;
3248 
3249 	bp = NULL;
3250 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3251 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3252 	if (!unmapped_buf_allowed)
3253 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3254 
3255 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3256 	    vp->v_type == VCHR)
3257 		metadata = true;
3258 	else
3259 		metadata = false;
3260 	if (vp == NULL)
3261 		bd = &bdomain[0];
3262 	else
3263 		bd = &bdomain[vp->v_bufobj.bo_domain];
3264 
3265 	counter_u64_add(getnewbufcalls, 1);
3266 	reserved = false;
3267 	do {
3268 		if (reserved == false &&
3269 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3270 			counter_u64_add(getnewbufrestarts, 1);
3271 			continue;
3272 		}
3273 		reserved = true;
3274 		if ((bp = buf_alloc(bd)) == NULL) {
3275 			counter_u64_add(getnewbufrestarts, 1);
3276 			continue;
3277 		}
3278 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3279 			return (bp);
3280 		break;
3281 	} while (buf_recycle(bd, false) == 0);
3282 
3283 	if (reserved)
3284 		bufspace_release(bd, maxsize);
3285 	if (bp != NULL) {
3286 		bp->b_flags |= B_INVAL;
3287 		brelse(bp);
3288 	}
3289 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3290 
3291 	return (NULL);
3292 }
3293 
3294 /*
3295  *	buf_daemon:
3296  *
3297  *	buffer flushing daemon.  Buffers are normally flushed by the
3298  *	update daemon but if it cannot keep up this process starts to
3299  *	take the load in an attempt to prevent getnewbuf() from blocking.
3300  */
3301 static struct kproc_desc buf_kp = {
3302 	"bufdaemon",
3303 	buf_daemon,
3304 	&bufdaemonproc
3305 };
3306 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3307 
3308 static int
3309 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3310 {
3311 	int flushed;
3312 
3313 	flushed = flushbufqueues(vp, bd, target, 0);
3314 	if (flushed == 0) {
3315 		/*
3316 		 * Could not find any buffers without rollback
3317 		 * dependencies, so just write the first one
3318 		 * in the hopes of eventually making progress.
3319 		 */
3320 		if (vp != NULL && target > 2)
3321 			target /= 2;
3322 		flushbufqueues(vp, bd, target, 1);
3323 	}
3324 	return (flushed);
3325 }
3326 
3327 static void
3328 buf_daemon()
3329 {
3330 	struct bufdomain *bd;
3331 	int speedupreq;
3332 	int lodirty;
3333 	int i;
3334 
3335 	/*
3336 	 * This process needs to be suspended prior to shutdown sync.
3337 	 */
3338 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3339 	    SHUTDOWN_PRI_LAST + 100);
3340 
3341 	/*
3342 	 * Start the buf clean daemons as children threads.
3343 	 */
3344 	for (i = 0 ; i < buf_domains; i++) {
3345 		int error;
3346 
3347 		error = kthread_add((void (*)(void *))bufspace_daemon,
3348 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3349 		if (error)
3350 			panic("error %d spawning bufspace daemon", error);
3351 	}
3352 
3353 	/*
3354 	 * This process is allowed to take the buffer cache to the limit
3355 	 */
3356 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3357 	mtx_lock(&bdlock);
3358 	for (;;) {
3359 		bd_request = 0;
3360 		mtx_unlock(&bdlock);
3361 
3362 		kthread_suspend_check();
3363 
3364 		/*
3365 		 * Save speedupreq for this pass and reset to capture new
3366 		 * requests.
3367 		 */
3368 		speedupreq = bd_speedupreq;
3369 		bd_speedupreq = 0;
3370 
3371 		/*
3372 		 * Flush each domain sequentially according to its level and
3373 		 * the speedup request.
3374 		 */
3375 		for (i = 0; i < buf_domains; i++) {
3376 			bd = &bdomain[i];
3377 			if (speedupreq)
3378 				lodirty = bd->bd_numdirtybuffers / 2;
3379 			else
3380 				lodirty = bd->bd_lodirtybuffers;
3381 			while (bd->bd_numdirtybuffers > lodirty) {
3382 				if (buf_flush(NULL, bd,
3383 				    bd->bd_numdirtybuffers - lodirty) == 0)
3384 					break;
3385 				kern_yield(PRI_USER);
3386 			}
3387 		}
3388 
3389 		/*
3390 		 * Only clear bd_request if we have reached our low water
3391 		 * mark.  The buf_daemon normally waits 1 second and
3392 		 * then incrementally flushes any dirty buffers that have
3393 		 * built up, within reason.
3394 		 *
3395 		 * If we were unable to hit our low water mark and couldn't
3396 		 * find any flushable buffers, we sleep for a short period
3397 		 * to avoid endless loops on unlockable buffers.
3398 		 */
3399 		mtx_lock(&bdlock);
3400 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3401 			/*
3402 			 * We reached our low water mark, reset the
3403 			 * request and sleep until we are needed again.
3404 			 * The sleep is just so the suspend code works.
3405 			 */
3406 			bd_request = 0;
3407 			/*
3408 			 * Do an extra wakeup in case dirty threshold
3409 			 * changed via sysctl and the explicit transition
3410 			 * out of shortfall was missed.
3411 			 */
3412 			bdirtywakeup();
3413 			if (runningbufspace <= lorunningspace)
3414 				runningwakeup();
3415 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3416 		} else {
3417 			/*
3418 			 * We couldn't find any flushable dirty buffers but
3419 			 * still have too many dirty buffers, we
3420 			 * have to sleep and try again.  (rare)
3421 			 */
3422 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3423 		}
3424 	}
3425 }
3426 
3427 /*
3428  *	flushbufqueues:
3429  *
3430  *	Try to flush a buffer in the dirty queue.  We must be careful to
3431  *	free up B_INVAL buffers instead of write them, which NFS is
3432  *	particularly sensitive to.
3433  */
3434 static int flushwithdeps = 0;
3435 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3436     &flushwithdeps, 0,
3437     "Number of buffers flushed with dependecies that require rollbacks");
3438 
3439 static int
3440 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3441     int flushdeps)
3442 {
3443 	struct bufqueue *bq;
3444 	struct buf *sentinel;
3445 	struct vnode *vp;
3446 	struct mount *mp;
3447 	struct buf *bp;
3448 	int hasdeps;
3449 	int flushed;
3450 	int error;
3451 	bool unlock;
3452 
3453 	flushed = 0;
3454 	bq = &bd->bd_dirtyq;
3455 	bp = NULL;
3456 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3457 	sentinel->b_qindex = QUEUE_SENTINEL;
3458 	BQ_LOCK(bq);
3459 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3460 	BQ_UNLOCK(bq);
3461 	while (flushed != target) {
3462 		maybe_yield();
3463 		BQ_LOCK(bq);
3464 		bp = TAILQ_NEXT(sentinel, b_freelist);
3465 		if (bp != NULL) {
3466 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3467 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3468 			    b_freelist);
3469 		} else {
3470 			BQ_UNLOCK(bq);
3471 			break;
3472 		}
3473 		/*
3474 		 * Skip sentinels inserted by other invocations of the
3475 		 * flushbufqueues(), taking care to not reorder them.
3476 		 *
3477 		 * Only flush the buffers that belong to the
3478 		 * vnode locked by the curthread.
3479 		 */
3480 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3481 		    bp->b_vp != lvp)) {
3482 			BQ_UNLOCK(bq);
3483 			continue;
3484 		}
3485 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3486 		BQ_UNLOCK(bq);
3487 		if (error != 0)
3488 			continue;
3489 
3490 		/*
3491 		 * BKGRDINPROG can only be set with the buf and bufobj
3492 		 * locks both held.  We tolerate a race to clear it here.
3493 		 */
3494 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3495 		    (bp->b_flags & B_DELWRI) == 0) {
3496 			BUF_UNLOCK(bp);
3497 			continue;
3498 		}
3499 		if (bp->b_flags & B_INVAL) {
3500 			bremfreef(bp);
3501 			brelse(bp);
3502 			flushed++;
3503 			continue;
3504 		}
3505 
3506 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3507 			if (flushdeps == 0) {
3508 				BUF_UNLOCK(bp);
3509 				continue;
3510 			}
3511 			hasdeps = 1;
3512 		} else
3513 			hasdeps = 0;
3514 		/*
3515 		 * We must hold the lock on a vnode before writing
3516 		 * one of its buffers. Otherwise we may confuse, or
3517 		 * in the case of a snapshot vnode, deadlock the
3518 		 * system.
3519 		 *
3520 		 * The lock order here is the reverse of the normal
3521 		 * of vnode followed by buf lock.  This is ok because
3522 		 * the NOWAIT will prevent deadlock.
3523 		 */
3524 		vp = bp->b_vp;
3525 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3526 			BUF_UNLOCK(bp);
3527 			continue;
3528 		}
3529 		if (lvp == NULL) {
3530 			unlock = true;
3531 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3532 		} else {
3533 			ASSERT_VOP_LOCKED(vp, "getbuf");
3534 			unlock = false;
3535 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3536 			    vn_lock(vp, LK_TRYUPGRADE);
3537 		}
3538 		if (error == 0) {
3539 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3540 			    bp, bp->b_vp, bp->b_flags);
3541 			if (curproc == bufdaemonproc) {
3542 				vfs_bio_awrite(bp);
3543 			} else {
3544 				bremfree(bp);
3545 				bwrite(bp);
3546 				counter_u64_add(notbufdflushes, 1);
3547 			}
3548 			vn_finished_write(mp);
3549 			if (unlock)
3550 				VOP_UNLOCK(vp);
3551 			flushwithdeps += hasdeps;
3552 			flushed++;
3553 
3554 			/*
3555 			 * Sleeping on runningbufspace while holding
3556 			 * vnode lock leads to deadlock.
3557 			 */
3558 			if (curproc == bufdaemonproc &&
3559 			    runningbufspace > hirunningspace)
3560 				waitrunningbufspace();
3561 			continue;
3562 		}
3563 		vn_finished_write(mp);
3564 		BUF_UNLOCK(bp);
3565 	}
3566 	BQ_LOCK(bq);
3567 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3568 	BQ_UNLOCK(bq);
3569 	free(sentinel, M_TEMP);
3570 	return (flushed);
3571 }
3572 
3573 /*
3574  * Check to see if a block is currently memory resident.
3575  */
3576 struct buf *
3577 incore(struct bufobj *bo, daddr_t blkno)
3578 {
3579 	return (gbincore_unlocked(bo, blkno));
3580 }
3581 
3582 /*
3583  * Returns true if no I/O is needed to access the
3584  * associated VM object.  This is like incore except
3585  * it also hunts around in the VM system for the data.
3586  */
3587 bool
3588 inmem(struct vnode * vp, daddr_t blkno)
3589 {
3590 	vm_object_t obj;
3591 	vm_offset_t toff, tinc, size;
3592 	vm_page_t m, n;
3593 	vm_ooffset_t off;
3594 	int valid;
3595 
3596 	ASSERT_VOP_LOCKED(vp, "inmem");
3597 
3598 	if (incore(&vp->v_bufobj, blkno))
3599 		return (true);
3600 	if (vp->v_mount == NULL)
3601 		return (false);
3602 	obj = vp->v_object;
3603 	if (obj == NULL)
3604 		return (false);
3605 
3606 	size = PAGE_SIZE;
3607 	if (size > vp->v_mount->mnt_stat.f_iosize)
3608 		size = vp->v_mount->mnt_stat.f_iosize;
3609 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3610 
3611 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3612 		m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3613 recheck:
3614 		if (m == NULL)
3615 			return (false);
3616 
3617 		tinc = size;
3618 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3619 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3620 		/*
3621 		 * Consider page validity only if page mapping didn't change
3622 		 * during the check.
3623 		 */
3624 		valid = vm_page_is_valid(m,
3625 		    (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3626 		n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3627 		if (m != n) {
3628 			m = n;
3629 			goto recheck;
3630 		}
3631 		if (!valid)
3632 			return (false);
3633 	}
3634 	return (true);
3635 }
3636 
3637 /*
3638  * Set the dirty range for a buffer based on the status of the dirty
3639  * bits in the pages comprising the buffer.  The range is limited
3640  * to the size of the buffer.
3641  *
3642  * Tell the VM system that the pages associated with this buffer
3643  * are clean.  This is used for delayed writes where the data is
3644  * going to go to disk eventually without additional VM intevention.
3645  *
3646  * Note that while we only really need to clean through to b_bcount, we
3647  * just go ahead and clean through to b_bufsize.
3648  */
3649 static void
3650 vfs_clean_pages_dirty_buf(struct buf *bp)
3651 {
3652 	vm_ooffset_t foff, noff, eoff;
3653 	vm_page_t m;
3654 	int i;
3655 
3656 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3657 		return;
3658 
3659 	foff = bp->b_offset;
3660 	KASSERT(bp->b_offset != NOOFFSET,
3661 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3662 
3663 	vfs_busy_pages_acquire(bp);
3664 	vfs_setdirty_range(bp);
3665 	for (i = 0; i < bp->b_npages; i++) {
3666 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3667 		eoff = noff;
3668 		if (eoff > bp->b_offset + bp->b_bufsize)
3669 			eoff = bp->b_offset + bp->b_bufsize;
3670 		m = bp->b_pages[i];
3671 		vfs_page_set_validclean(bp, foff, m);
3672 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3673 		foff = noff;
3674 	}
3675 	vfs_busy_pages_release(bp);
3676 }
3677 
3678 static void
3679 vfs_setdirty_range(struct buf *bp)
3680 {
3681 	vm_offset_t boffset;
3682 	vm_offset_t eoffset;
3683 	int i;
3684 
3685 	/*
3686 	 * test the pages to see if they have been modified directly
3687 	 * by users through the VM system.
3688 	 */
3689 	for (i = 0; i < bp->b_npages; i++)
3690 		vm_page_test_dirty(bp->b_pages[i]);
3691 
3692 	/*
3693 	 * Calculate the encompassing dirty range, boffset and eoffset,
3694 	 * (eoffset - boffset) bytes.
3695 	 */
3696 
3697 	for (i = 0; i < bp->b_npages; i++) {
3698 		if (bp->b_pages[i]->dirty)
3699 			break;
3700 	}
3701 	boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3702 
3703 	for (i = bp->b_npages - 1; i >= 0; --i) {
3704 		if (bp->b_pages[i]->dirty) {
3705 			break;
3706 		}
3707 	}
3708 	eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3709 
3710 	/*
3711 	 * Fit it to the buffer.
3712 	 */
3713 
3714 	if (eoffset > bp->b_bcount)
3715 		eoffset = bp->b_bcount;
3716 
3717 	/*
3718 	 * If we have a good dirty range, merge with the existing
3719 	 * dirty range.
3720 	 */
3721 
3722 	if (boffset < eoffset) {
3723 		if (bp->b_dirtyoff > boffset)
3724 			bp->b_dirtyoff = boffset;
3725 		if (bp->b_dirtyend < eoffset)
3726 			bp->b_dirtyend = eoffset;
3727 	}
3728 }
3729 
3730 /*
3731  * Allocate the KVA mapping for an existing buffer.
3732  * If an unmapped buffer is provided but a mapped buffer is requested, take
3733  * also care to properly setup mappings between pages and KVA.
3734  */
3735 static void
3736 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3737 {
3738 	int bsize, maxsize, need_mapping, need_kva;
3739 	off_t offset;
3740 
3741 	need_mapping = bp->b_data == unmapped_buf &&
3742 	    (gbflags & GB_UNMAPPED) == 0;
3743 	need_kva = bp->b_kvabase == unmapped_buf &&
3744 	    bp->b_data == unmapped_buf &&
3745 	    (gbflags & GB_KVAALLOC) != 0;
3746 	if (!need_mapping && !need_kva)
3747 		return;
3748 
3749 	BUF_CHECK_UNMAPPED(bp);
3750 
3751 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3752 		/*
3753 		 * Buffer is not mapped, but the KVA was already
3754 		 * reserved at the time of the instantiation.  Use the
3755 		 * allocated space.
3756 		 */
3757 		goto has_addr;
3758 	}
3759 
3760 	/*
3761 	 * Calculate the amount of the address space we would reserve
3762 	 * if the buffer was mapped.
3763 	 */
3764 	bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3765 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3766 	offset = blkno * bsize;
3767 	maxsize = size + (offset & PAGE_MASK);
3768 	maxsize = imax(maxsize, bsize);
3769 
3770 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3771 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3772 			/*
3773 			 * XXXKIB: defragmentation cannot
3774 			 * succeed, not sure what else to do.
3775 			 */
3776 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3777 		}
3778 		counter_u64_add(mappingrestarts, 1);
3779 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3780 	}
3781 has_addr:
3782 	if (need_mapping) {
3783 		/* b_offset is handled by bpmap_qenter. */
3784 		bp->b_data = bp->b_kvabase;
3785 		BUF_CHECK_MAPPED(bp);
3786 		bpmap_qenter(bp);
3787 	}
3788 }
3789 
3790 struct buf *
3791 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3792     int flags)
3793 {
3794 	struct buf *bp;
3795 	int error;
3796 
3797 	error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3798 	if (error != 0)
3799 		return (NULL);
3800 	return (bp);
3801 }
3802 
3803 /*
3804  *	getblkx:
3805  *
3806  *	Get a block given a specified block and offset into a file/device.
3807  *	The buffers B_DONE bit will be cleared on return, making it almost
3808  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3809  *	return.  The caller should clear B_INVAL prior to initiating a
3810  *	READ.
3811  *
3812  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3813  *	an existing buffer.
3814  *
3815  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3816  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3817  *	and then cleared based on the backing VM.  If the previous buffer is
3818  *	non-0-sized but invalid, B_CACHE will be cleared.
3819  *
3820  *	If getblk() must create a new buffer, the new buffer is returned with
3821  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3822  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3823  *	backing VM.
3824  *
3825  *	getblk() also forces a bwrite() for any B_DELWRI buffer whose
3826  *	B_CACHE bit is clear.
3827  *
3828  *	What this means, basically, is that the caller should use B_CACHE to
3829  *	determine whether the buffer is fully valid or not and should clear
3830  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3831  *	the buffer by loading its data area with something, the caller needs
3832  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3833  *	the caller should set B_CACHE ( as an optimization ), else the caller
3834  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3835  *	a write attempt or if it was a successful read.  If the caller
3836  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3837  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3838  *
3839  *	The blkno parameter is the logical block being requested. Normally
3840  *	the mapping of logical block number to disk block address is done
3841  *	by calling VOP_BMAP(). However, if the mapping is already known, the
3842  *	disk block address can be passed using the dblkno parameter. If the
3843  *	disk block address is not known, then the same value should be passed
3844  *	for blkno and dblkno.
3845  */
3846 int
3847 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3848     int slptimeo, int flags, struct buf **bpp)
3849 {
3850 	struct buf *bp;
3851 	struct bufobj *bo;
3852 	daddr_t d_blkno;
3853 	int bsize, error, maxsize, vmio;
3854 	off_t offset;
3855 
3856 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3857 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3858 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3859 	ASSERT_VOP_LOCKED(vp, "getblk");
3860 	if (size > maxbcachebuf)
3861 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3862 		    maxbcachebuf);
3863 	if (!unmapped_buf_allowed)
3864 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3865 
3866 	bo = &vp->v_bufobj;
3867 	d_blkno = dblkno;
3868 
3869 	/* Attempt lockless lookup first. */
3870 	bp = gbincore_unlocked(bo, blkno);
3871 	if (bp == NULL)
3872 		goto newbuf_unlocked;
3873 
3874 	error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3875 	    0);
3876 	if (error != 0)
3877 		goto loop;
3878 
3879 	/* Verify buf identify has not changed since lookup. */
3880 	if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3881 		goto foundbuf_fastpath;
3882 
3883 	/* It changed, fallback to locked lookup. */
3884 	BUF_UNLOCK_RAW(bp);
3885 
3886 loop:
3887 	BO_RLOCK(bo);
3888 	bp = gbincore(bo, blkno);
3889 	if (bp != NULL) {
3890 		int lockflags;
3891 
3892 		/*
3893 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3894 		 * it must be on a queue.
3895 		 */
3896 		lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3897 		    ((flags & GB_LOCK_NOWAIT) ? LK_NOWAIT : LK_SLEEPFAIL);
3898 
3899 		error = BUF_TIMELOCK(bp, lockflags,
3900 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3901 
3902 		/*
3903 		 * If we slept and got the lock we have to restart in case
3904 		 * the buffer changed identities.
3905 		 */
3906 		if (error == ENOLCK)
3907 			goto loop;
3908 		/* We timed out or were interrupted. */
3909 		else if (error != 0)
3910 			return (error);
3911 
3912 foundbuf_fastpath:
3913 		/* If recursed, assume caller knows the rules. */
3914 		if (BUF_LOCKRECURSED(bp))
3915 			goto end;
3916 
3917 		/*
3918 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3919 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3920 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3921 		 * backing VM cache.
3922 		 */
3923 		if (bp->b_flags & B_INVAL)
3924 			bp->b_flags &= ~B_CACHE;
3925 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3926 			bp->b_flags |= B_CACHE;
3927 		if (bp->b_flags & B_MANAGED)
3928 			MPASS(bp->b_qindex == QUEUE_NONE);
3929 		else
3930 			bremfree(bp);
3931 
3932 		/*
3933 		 * check for size inconsistencies for non-VMIO case.
3934 		 */
3935 		if (bp->b_bcount != size) {
3936 			if ((bp->b_flags & B_VMIO) == 0 ||
3937 			    (size > bp->b_kvasize)) {
3938 				if (bp->b_flags & B_DELWRI) {
3939 					bp->b_flags |= B_NOCACHE;
3940 					bwrite(bp);
3941 				} else {
3942 					if (LIST_EMPTY(&bp->b_dep)) {
3943 						bp->b_flags |= B_RELBUF;
3944 						brelse(bp);
3945 					} else {
3946 						bp->b_flags |= B_NOCACHE;
3947 						bwrite(bp);
3948 					}
3949 				}
3950 				goto loop;
3951 			}
3952 		}
3953 
3954 		/*
3955 		 * Handle the case of unmapped buffer which should
3956 		 * become mapped, or the buffer for which KVA
3957 		 * reservation is requested.
3958 		 */
3959 		bp_unmapped_get_kva(bp, blkno, size, flags);
3960 
3961 		/*
3962 		 * If the size is inconsistent in the VMIO case, we can resize
3963 		 * the buffer.  This might lead to B_CACHE getting set or
3964 		 * cleared.  If the size has not changed, B_CACHE remains
3965 		 * unchanged from its previous state.
3966 		 */
3967 		allocbuf(bp, size);
3968 
3969 		KASSERT(bp->b_offset != NOOFFSET,
3970 		    ("getblk: no buffer offset"));
3971 
3972 		/*
3973 		 * A buffer with B_DELWRI set and B_CACHE clear must
3974 		 * be committed before we can return the buffer in
3975 		 * order to prevent the caller from issuing a read
3976 		 * ( due to B_CACHE not being set ) and overwriting
3977 		 * it.
3978 		 *
3979 		 * Most callers, including NFS and FFS, need this to
3980 		 * operate properly either because they assume they
3981 		 * can issue a read if B_CACHE is not set, or because
3982 		 * ( for example ) an uncached B_DELWRI might loop due
3983 		 * to softupdates re-dirtying the buffer.  In the latter
3984 		 * case, B_CACHE is set after the first write completes,
3985 		 * preventing further loops.
3986 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3987 		 * above while extending the buffer, we cannot allow the
3988 		 * buffer to remain with B_CACHE set after the write
3989 		 * completes or it will represent a corrupt state.  To
3990 		 * deal with this we set B_NOCACHE to scrap the buffer
3991 		 * after the write.
3992 		 *
3993 		 * We might be able to do something fancy, like setting
3994 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3995 		 * so the below call doesn't set B_CACHE, but that gets real
3996 		 * confusing.  This is much easier.
3997 		 */
3998 
3999 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4000 			bp->b_flags |= B_NOCACHE;
4001 			bwrite(bp);
4002 			goto loop;
4003 		}
4004 		bp->b_flags &= ~B_DONE;
4005 	} else {
4006 		/*
4007 		 * Buffer is not in-core, create new buffer.  The buffer
4008 		 * returned by getnewbuf() is locked.  Note that the returned
4009 		 * buffer is also considered valid (not marked B_INVAL).
4010 		 */
4011 		BO_RUNLOCK(bo);
4012 newbuf_unlocked:
4013 		/*
4014 		 * If the user does not want us to create the buffer, bail out
4015 		 * here.
4016 		 */
4017 		if (flags & GB_NOCREAT)
4018 			return (EEXIST);
4019 
4020 		bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4021 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4022 		offset = blkno * bsize;
4023 		vmio = vp->v_object != NULL;
4024 		if (vmio) {
4025 			maxsize = size + (offset & PAGE_MASK);
4026 		} else {
4027 			maxsize = size;
4028 			/* Do not allow non-VMIO notmapped buffers. */
4029 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4030 		}
4031 		maxsize = imax(maxsize, bsize);
4032 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4033 		    !vn_isdisk(vp)) {
4034 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4035 			KASSERT(error != EOPNOTSUPP,
4036 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4037 			    vp));
4038 			if (error != 0)
4039 				return (error);
4040 			if (d_blkno == -1)
4041 				return (EJUSTRETURN);
4042 		}
4043 
4044 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4045 		if (bp == NULL) {
4046 			if (slpflag || slptimeo)
4047 				return (ETIMEDOUT);
4048 			/*
4049 			 * XXX This is here until the sleep path is diagnosed
4050 			 * enough to work under very low memory conditions.
4051 			 *
4052 			 * There's an issue on low memory, 4BSD+non-preempt
4053 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4054 			 * exhaustion occurs without sleeping for buffer
4055 			 * reclaimation.  This just sticks in a loop and
4056 			 * constantly attempts to allocate a buffer, which
4057 			 * hits exhaustion and tries to wakeup bufdaemon.
4058 			 * This never happens because we never yield.
4059 			 *
4060 			 * The real solution is to identify and fix these cases
4061 			 * so we aren't effectively busy-waiting in a loop
4062 			 * until the reclaimation path has cycles to run.
4063 			 */
4064 			kern_yield(PRI_USER);
4065 			goto loop;
4066 		}
4067 
4068 		/*
4069 		 * This code is used to make sure that a buffer is not
4070 		 * created while the getnewbuf routine is blocked.
4071 		 * This can be a problem whether the vnode is locked or not.
4072 		 * If the buffer is created out from under us, we have to
4073 		 * throw away the one we just created.
4074 		 *
4075 		 * Note: this must occur before we associate the buffer
4076 		 * with the vp especially considering limitations in
4077 		 * the splay tree implementation when dealing with duplicate
4078 		 * lblkno's.
4079 		 */
4080 		BO_LOCK(bo);
4081 		if (gbincore(bo, blkno)) {
4082 			BO_UNLOCK(bo);
4083 			bp->b_flags |= B_INVAL;
4084 			bufspace_release(bufdomain(bp), maxsize);
4085 			brelse(bp);
4086 			goto loop;
4087 		}
4088 
4089 		/*
4090 		 * Insert the buffer into the hash, so that it can
4091 		 * be found by incore.
4092 		 */
4093 		bp->b_lblkno = blkno;
4094 		bp->b_blkno = d_blkno;
4095 		bp->b_offset = offset;
4096 		bgetvp(vp, bp);
4097 		BO_UNLOCK(bo);
4098 
4099 		/*
4100 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4101 		 * buffer size starts out as 0, B_CACHE will be set by
4102 		 * allocbuf() for the VMIO case prior to it testing the
4103 		 * backing store for validity.
4104 		 */
4105 
4106 		if (vmio) {
4107 			bp->b_flags |= B_VMIO;
4108 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4109 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4110 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4111 		} else {
4112 			bp->b_flags &= ~B_VMIO;
4113 			KASSERT(bp->b_bufobj->bo_object == NULL,
4114 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4115 			    bp, bp->b_bufobj->bo_object));
4116 			BUF_CHECK_MAPPED(bp);
4117 		}
4118 
4119 		allocbuf(bp, size);
4120 		bufspace_release(bufdomain(bp), maxsize);
4121 		bp->b_flags &= ~B_DONE;
4122 	}
4123 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4124 end:
4125 	buf_track(bp, __func__);
4126 	KASSERT(bp->b_bufobj == bo,
4127 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4128 	*bpp = bp;
4129 	return (0);
4130 }
4131 
4132 /*
4133  * Get an empty, disassociated buffer of given size.  The buffer is initially
4134  * set to B_INVAL.
4135  */
4136 struct buf *
4137 geteblk(int size, int flags)
4138 {
4139 	struct buf *bp;
4140 	int maxsize;
4141 
4142 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4143 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4144 		if ((flags & GB_NOWAIT_BD) &&
4145 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4146 			return (NULL);
4147 	}
4148 	allocbuf(bp, size);
4149 	bufspace_release(bufdomain(bp), maxsize);
4150 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4151 	return (bp);
4152 }
4153 
4154 /*
4155  * Truncate the backing store for a non-vmio buffer.
4156  */
4157 static void
4158 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4159 {
4160 
4161 	if (bp->b_flags & B_MALLOC) {
4162 		/*
4163 		 * malloced buffers are not shrunk
4164 		 */
4165 		if (newbsize == 0) {
4166 			bufmallocadjust(bp, 0);
4167 			free(bp->b_data, M_BIOBUF);
4168 			bp->b_data = bp->b_kvabase;
4169 			bp->b_flags &= ~B_MALLOC;
4170 		}
4171 		return;
4172 	}
4173 	vm_hold_free_pages(bp, newbsize);
4174 	bufspace_adjust(bp, newbsize);
4175 }
4176 
4177 /*
4178  * Extend the backing for a non-VMIO buffer.
4179  */
4180 static void
4181 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4182 {
4183 	caddr_t origbuf;
4184 	int origbufsize;
4185 
4186 	/*
4187 	 * We only use malloced memory on the first allocation.
4188 	 * and revert to page-allocated memory when the buffer
4189 	 * grows.
4190 	 *
4191 	 * There is a potential smp race here that could lead
4192 	 * to bufmallocspace slightly passing the max.  It
4193 	 * is probably extremely rare and not worth worrying
4194 	 * over.
4195 	 */
4196 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4197 	    bufmallocspace < maxbufmallocspace) {
4198 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4199 		bp->b_flags |= B_MALLOC;
4200 		bufmallocadjust(bp, newbsize);
4201 		return;
4202 	}
4203 
4204 	/*
4205 	 * If the buffer is growing on its other-than-first
4206 	 * allocation then we revert to the page-allocation
4207 	 * scheme.
4208 	 */
4209 	origbuf = NULL;
4210 	origbufsize = 0;
4211 	if (bp->b_flags & B_MALLOC) {
4212 		origbuf = bp->b_data;
4213 		origbufsize = bp->b_bufsize;
4214 		bp->b_data = bp->b_kvabase;
4215 		bufmallocadjust(bp, 0);
4216 		bp->b_flags &= ~B_MALLOC;
4217 		newbsize = round_page(newbsize);
4218 	}
4219 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4220 	    (vm_offset_t) bp->b_data + newbsize);
4221 	if (origbuf != NULL) {
4222 		bcopy(origbuf, bp->b_data, origbufsize);
4223 		free(origbuf, M_BIOBUF);
4224 	}
4225 	bufspace_adjust(bp, newbsize);
4226 }
4227 
4228 /*
4229  * This code constitutes the buffer memory from either anonymous system
4230  * memory (in the case of non-VMIO operations) or from an associated
4231  * VM object (in the case of VMIO operations).  This code is able to
4232  * resize a buffer up or down.
4233  *
4234  * Note that this code is tricky, and has many complications to resolve
4235  * deadlock or inconsistent data situations.  Tread lightly!!!
4236  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4237  * the caller.  Calling this code willy nilly can result in the loss of data.
4238  *
4239  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4240  * B_CACHE for the non-VMIO case.
4241  */
4242 int
4243 allocbuf(struct buf *bp, int size)
4244 {
4245 	int newbsize;
4246 
4247 	if (bp->b_bcount == size)
4248 		return (1);
4249 
4250 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4251 		panic("allocbuf: buffer too small");
4252 
4253 	newbsize = roundup2(size, DEV_BSIZE);
4254 	if ((bp->b_flags & B_VMIO) == 0) {
4255 		if ((bp->b_flags & B_MALLOC) == 0)
4256 			newbsize = round_page(newbsize);
4257 		/*
4258 		 * Just get anonymous memory from the kernel.  Don't
4259 		 * mess with B_CACHE.
4260 		 */
4261 		if (newbsize < bp->b_bufsize)
4262 			vfs_nonvmio_truncate(bp, newbsize);
4263 		else if (newbsize > bp->b_bufsize)
4264 			vfs_nonvmio_extend(bp, newbsize);
4265 	} else {
4266 		int desiredpages;
4267 
4268 		desiredpages = (size == 0) ? 0 :
4269 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4270 
4271 		if (bp->b_flags & B_MALLOC)
4272 			panic("allocbuf: VMIO buffer can't be malloced");
4273 		/*
4274 		 * Set B_CACHE initially if buffer is 0 length or will become
4275 		 * 0-length.
4276 		 */
4277 		if (size == 0 || bp->b_bufsize == 0)
4278 			bp->b_flags |= B_CACHE;
4279 
4280 		if (newbsize < bp->b_bufsize)
4281 			vfs_vmio_truncate(bp, desiredpages);
4282 		/* XXX This looks as if it should be newbsize > b_bufsize */
4283 		else if (size > bp->b_bcount)
4284 			vfs_vmio_extend(bp, desiredpages, size);
4285 		bufspace_adjust(bp, newbsize);
4286 	}
4287 	bp->b_bcount = size;		/* requested buffer size. */
4288 	return (1);
4289 }
4290 
4291 extern int inflight_transient_maps;
4292 
4293 static struct bio_queue nondump_bios;
4294 
4295 void
4296 biodone(struct bio *bp)
4297 {
4298 	struct mtx *mtxp;
4299 	void (*done)(struct bio *);
4300 	vm_offset_t start, end;
4301 
4302 	biotrack(bp, __func__);
4303 
4304 	/*
4305 	 * Avoid completing I/O when dumping after a panic since that may
4306 	 * result in a deadlock in the filesystem or pager code.  Note that
4307 	 * this doesn't affect dumps that were started manually since we aim
4308 	 * to keep the system usable after it has been resumed.
4309 	 */
4310 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4311 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4312 		return;
4313 	}
4314 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4315 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4316 		bp->bio_flags |= BIO_UNMAPPED;
4317 		start = trunc_page((vm_offset_t)bp->bio_data);
4318 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4319 		bp->bio_data = unmapped_buf;
4320 		pmap_qremove(start, atop(end - start));
4321 		vmem_free(transient_arena, start, end - start);
4322 		atomic_add_int(&inflight_transient_maps, -1);
4323 	}
4324 	done = bp->bio_done;
4325 	if (done == NULL) {
4326 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4327 		mtx_lock(mtxp);
4328 		bp->bio_flags |= BIO_DONE;
4329 		wakeup(bp);
4330 		mtx_unlock(mtxp);
4331 	} else
4332 		done(bp);
4333 }
4334 
4335 /*
4336  * Wait for a BIO to finish.
4337  */
4338 int
4339 biowait(struct bio *bp, const char *wchan)
4340 {
4341 	struct mtx *mtxp;
4342 
4343 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4344 	mtx_lock(mtxp);
4345 	while ((bp->bio_flags & BIO_DONE) == 0)
4346 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4347 	mtx_unlock(mtxp);
4348 	if (bp->bio_error != 0)
4349 		return (bp->bio_error);
4350 	if (!(bp->bio_flags & BIO_ERROR))
4351 		return (0);
4352 	return (EIO);
4353 }
4354 
4355 void
4356 biofinish(struct bio *bp, struct devstat *stat, int error)
4357 {
4358 
4359 	if (error) {
4360 		bp->bio_error = error;
4361 		bp->bio_flags |= BIO_ERROR;
4362 	}
4363 	if (stat != NULL)
4364 		devstat_end_transaction_bio(stat, bp);
4365 	biodone(bp);
4366 }
4367 
4368 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4369 void
4370 biotrack_buf(struct bio *bp, const char *location)
4371 {
4372 
4373 	buf_track(bp->bio_track_bp, location);
4374 }
4375 #endif
4376 
4377 /*
4378  *	bufwait:
4379  *
4380  *	Wait for buffer I/O completion, returning error status.  The buffer
4381  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4382  *	error and cleared.
4383  */
4384 int
4385 bufwait(struct buf *bp)
4386 {
4387 	if (bp->b_iocmd == BIO_READ)
4388 		bwait(bp, PRIBIO, "biord");
4389 	else
4390 		bwait(bp, PRIBIO, "biowr");
4391 	if (bp->b_flags & B_EINTR) {
4392 		bp->b_flags &= ~B_EINTR;
4393 		return (EINTR);
4394 	}
4395 	if (bp->b_ioflags & BIO_ERROR) {
4396 		return (bp->b_error ? bp->b_error : EIO);
4397 	} else {
4398 		return (0);
4399 	}
4400 }
4401 
4402 /*
4403  *	bufdone:
4404  *
4405  *	Finish I/O on a buffer, optionally calling a completion function.
4406  *	This is usually called from an interrupt so process blocking is
4407  *	not allowed.
4408  *
4409  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4410  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4411  *	assuming B_INVAL is clear.
4412  *
4413  *	For the VMIO case, we set B_CACHE if the op was a read and no
4414  *	read error occurred, or if the op was a write.  B_CACHE is never
4415  *	set if the buffer is invalid or otherwise uncacheable.
4416  *
4417  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4418  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4419  *	in the biodone routine.
4420  */
4421 void
4422 bufdone(struct buf *bp)
4423 {
4424 	struct bufobj *dropobj;
4425 	void    (*biodone)(struct buf *);
4426 
4427 	buf_track(bp, __func__);
4428 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4429 	dropobj = NULL;
4430 
4431 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4432 
4433 	runningbufwakeup(bp);
4434 	if (bp->b_iocmd == BIO_WRITE)
4435 		dropobj = bp->b_bufobj;
4436 	/* call optional completion function if requested */
4437 	if (bp->b_iodone != NULL) {
4438 		biodone = bp->b_iodone;
4439 		bp->b_iodone = NULL;
4440 		(*biodone) (bp);
4441 		if (dropobj)
4442 			bufobj_wdrop(dropobj);
4443 		return;
4444 	}
4445 	if (bp->b_flags & B_VMIO) {
4446 		/*
4447 		 * Set B_CACHE if the op was a normal read and no error
4448 		 * occurred.  B_CACHE is set for writes in the b*write()
4449 		 * routines.
4450 		 */
4451 		if (bp->b_iocmd == BIO_READ &&
4452 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4453 		    !(bp->b_ioflags & BIO_ERROR))
4454 			bp->b_flags |= B_CACHE;
4455 		vfs_vmio_iodone(bp);
4456 	}
4457 	if (!LIST_EMPTY(&bp->b_dep))
4458 		buf_complete(bp);
4459 	if ((bp->b_flags & B_CKHASH) != 0) {
4460 		KASSERT(bp->b_iocmd == BIO_READ,
4461 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4462 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4463 		(*bp->b_ckhashcalc)(bp);
4464 	}
4465 	/*
4466 	 * For asynchronous completions, release the buffer now. The brelse
4467 	 * will do a wakeup there if necessary - so no need to do a wakeup
4468 	 * here in the async case. The sync case always needs to do a wakeup.
4469 	 */
4470 	if (bp->b_flags & B_ASYNC) {
4471 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4472 		    (bp->b_ioflags & BIO_ERROR))
4473 			brelse(bp);
4474 		else
4475 			bqrelse(bp);
4476 	} else
4477 		bdone(bp);
4478 	if (dropobj)
4479 		bufobj_wdrop(dropobj);
4480 }
4481 
4482 /*
4483  * This routine is called in lieu of iodone in the case of
4484  * incomplete I/O.  This keeps the busy status for pages
4485  * consistent.
4486  */
4487 void
4488 vfs_unbusy_pages(struct buf *bp)
4489 {
4490 	int i;
4491 	vm_object_t obj;
4492 	vm_page_t m;
4493 
4494 	runningbufwakeup(bp);
4495 	if (!(bp->b_flags & B_VMIO))
4496 		return;
4497 
4498 	obj = bp->b_bufobj->bo_object;
4499 	for (i = 0; i < bp->b_npages; i++) {
4500 		m = bp->b_pages[i];
4501 		if (m == bogus_page) {
4502 			m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4503 			if (!m)
4504 				panic("vfs_unbusy_pages: page missing\n");
4505 			bp->b_pages[i] = m;
4506 			if (buf_mapped(bp)) {
4507 				BUF_CHECK_MAPPED(bp);
4508 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4509 				    bp->b_pages, bp->b_npages);
4510 			} else
4511 				BUF_CHECK_UNMAPPED(bp);
4512 		}
4513 		vm_page_sunbusy(m);
4514 	}
4515 	vm_object_pip_wakeupn(obj, bp->b_npages);
4516 }
4517 
4518 /*
4519  * vfs_page_set_valid:
4520  *
4521  *	Set the valid bits in a page based on the supplied offset.   The
4522  *	range is restricted to the buffer's size.
4523  *
4524  *	This routine is typically called after a read completes.
4525  */
4526 static void
4527 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4528 {
4529 	vm_ooffset_t eoff;
4530 
4531 	/*
4532 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4533 	 * page boundary and eoff is not greater than the end of the buffer.
4534 	 * The end of the buffer, in this case, is our file EOF, not the
4535 	 * allocation size of the buffer.
4536 	 */
4537 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4538 	if (eoff > bp->b_offset + bp->b_bcount)
4539 		eoff = bp->b_offset + bp->b_bcount;
4540 
4541 	/*
4542 	 * Set valid range.  This is typically the entire buffer and thus the
4543 	 * entire page.
4544 	 */
4545 	if (eoff > off)
4546 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4547 }
4548 
4549 /*
4550  * vfs_page_set_validclean:
4551  *
4552  *	Set the valid bits and clear the dirty bits in a page based on the
4553  *	supplied offset.   The range is restricted to the buffer's size.
4554  */
4555 static void
4556 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4557 {
4558 	vm_ooffset_t soff, eoff;
4559 
4560 	/*
4561 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4562 	 * page boundary or cross the end of the buffer.  The end of the
4563 	 * buffer, in this case, is our file EOF, not the allocation size
4564 	 * of the buffer.
4565 	 */
4566 	soff = off;
4567 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4568 	if (eoff > bp->b_offset + bp->b_bcount)
4569 		eoff = bp->b_offset + bp->b_bcount;
4570 
4571 	/*
4572 	 * Set valid range.  This is typically the entire buffer and thus the
4573 	 * entire page.
4574 	 */
4575 	if (eoff > soff) {
4576 		vm_page_set_validclean(
4577 		    m,
4578 		   (vm_offset_t) (soff & PAGE_MASK),
4579 		   (vm_offset_t) (eoff - soff)
4580 		);
4581 	}
4582 }
4583 
4584 /*
4585  * Acquire a shared busy on all pages in the buf.
4586  */
4587 void
4588 vfs_busy_pages_acquire(struct buf *bp)
4589 {
4590 	int i;
4591 
4592 	for (i = 0; i < bp->b_npages; i++)
4593 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4594 }
4595 
4596 void
4597 vfs_busy_pages_release(struct buf *bp)
4598 {
4599 	int i;
4600 
4601 	for (i = 0; i < bp->b_npages; i++)
4602 		vm_page_sunbusy(bp->b_pages[i]);
4603 }
4604 
4605 /*
4606  * This routine is called before a device strategy routine.
4607  * It is used to tell the VM system that paging I/O is in
4608  * progress, and treat the pages associated with the buffer
4609  * almost as being exclusive busy.  Also the object paging_in_progress
4610  * flag is handled to make sure that the object doesn't become
4611  * inconsistent.
4612  *
4613  * Since I/O has not been initiated yet, certain buffer flags
4614  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4615  * and should be ignored.
4616  */
4617 void
4618 vfs_busy_pages(struct buf *bp, int clear_modify)
4619 {
4620 	vm_object_t obj;
4621 	vm_ooffset_t foff;
4622 	vm_page_t m;
4623 	int i;
4624 	bool bogus;
4625 
4626 	if (!(bp->b_flags & B_VMIO))
4627 		return;
4628 
4629 	obj = bp->b_bufobj->bo_object;
4630 	foff = bp->b_offset;
4631 	KASSERT(bp->b_offset != NOOFFSET,
4632 	    ("vfs_busy_pages: no buffer offset"));
4633 	if ((bp->b_flags & B_CLUSTER) == 0) {
4634 		vm_object_pip_add(obj, bp->b_npages);
4635 		vfs_busy_pages_acquire(bp);
4636 	}
4637 	if (bp->b_bufsize != 0)
4638 		vfs_setdirty_range(bp);
4639 	bogus = false;
4640 	for (i = 0; i < bp->b_npages; i++) {
4641 		m = bp->b_pages[i];
4642 		vm_page_assert_sbusied(m);
4643 
4644 		/*
4645 		 * When readying a buffer for a read ( i.e
4646 		 * clear_modify == 0 ), it is important to do
4647 		 * bogus_page replacement for valid pages in
4648 		 * partially instantiated buffers.  Partially
4649 		 * instantiated buffers can, in turn, occur when
4650 		 * reconstituting a buffer from its VM backing store
4651 		 * base.  We only have to do this if B_CACHE is
4652 		 * clear ( which causes the I/O to occur in the
4653 		 * first place ).  The replacement prevents the read
4654 		 * I/O from overwriting potentially dirty VM-backed
4655 		 * pages.  XXX bogus page replacement is, uh, bogus.
4656 		 * It may not work properly with small-block devices.
4657 		 * We need to find a better way.
4658 		 */
4659 		if (clear_modify) {
4660 			pmap_remove_write(m);
4661 			vfs_page_set_validclean(bp, foff, m);
4662 		} else if (vm_page_all_valid(m) &&
4663 		    (bp->b_flags & B_CACHE) == 0) {
4664 			bp->b_pages[i] = bogus_page;
4665 			bogus = true;
4666 		}
4667 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4668 	}
4669 	if (bogus && buf_mapped(bp)) {
4670 		BUF_CHECK_MAPPED(bp);
4671 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4672 		    bp->b_pages, bp->b_npages);
4673 	}
4674 }
4675 
4676 /*
4677  *	vfs_bio_set_valid:
4678  *
4679  *	Set the range within the buffer to valid.  The range is
4680  *	relative to the beginning of the buffer, b_offset.  Note that
4681  *	b_offset itself may be offset from the beginning of the first
4682  *	page.
4683  */
4684 void
4685 vfs_bio_set_valid(struct buf *bp, int base, int size)
4686 {
4687 	int i, n;
4688 	vm_page_t m;
4689 
4690 	if (!(bp->b_flags & B_VMIO))
4691 		return;
4692 
4693 	/*
4694 	 * Fixup base to be relative to beginning of first page.
4695 	 * Set initial n to be the maximum number of bytes in the
4696 	 * first page that can be validated.
4697 	 */
4698 	base += (bp->b_offset & PAGE_MASK);
4699 	n = PAGE_SIZE - (base & PAGE_MASK);
4700 
4701 	/*
4702 	 * Busy may not be strictly necessary here because the pages are
4703 	 * unlikely to be fully valid and the vnode lock will synchronize
4704 	 * their access via getpages.  It is grabbed for consistency with
4705 	 * other page validation.
4706 	 */
4707 	vfs_busy_pages_acquire(bp);
4708 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4709 		m = bp->b_pages[i];
4710 		if (n > size)
4711 			n = size;
4712 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4713 		base += n;
4714 		size -= n;
4715 		n = PAGE_SIZE;
4716 	}
4717 	vfs_busy_pages_release(bp);
4718 }
4719 
4720 /*
4721  *	vfs_bio_clrbuf:
4722  *
4723  *	If the specified buffer is a non-VMIO buffer, clear the entire
4724  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4725  *	validate only the previously invalid portions of the buffer.
4726  *	This routine essentially fakes an I/O, so we need to clear
4727  *	BIO_ERROR and B_INVAL.
4728  *
4729  *	Note that while we only theoretically need to clear through b_bcount,
4730  *	we go ahead and clear through b_bufsize.
4731  */
4732 void
4733 vfs_bio_clrbuf(struct buf *bp)
4734 {
4735 	int i, j, mask, sa, ea, slide;
4736 
4737 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4738 		clrbuf(bp);
4739 		return;
4740 	}
4741 	bp->b_flags &= ~B_INVAL;
4742 	bp->b_ioflags &= ~BIO_ERROR;
4743 	vfs_busy_pages_acquire(bp);
4744 	sa = bp->b_offset & PAGE_MASK;
4745 	slide = 0;
4746 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4747 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4748 		ea = slide & PAGE_MASK;
4749 		if (ea == 0)
4750 			ea = PAGE_SIZE;
4751 		if (bp->b_pages[i] == bogus_page)
4752 			continue;
4753 		j = sa / DEV_BSIZE;
4754 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4755 		if ((bp->b_pages[i]->valid & mask) == mask)
4756 			continue;
4757 		if ((bp->b_pages[i]->valid & mask) == 0)
4758 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4759 		else {
4760 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4761 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4762 					pmap_zero_page_area(bp->b_pages[i],
4763 					    sa, DEV_BSIZE);
4764 				}
4765 			}
4766 		}
4767 		vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4768 		    roundup2(ea - sa, DEV_BSIZE));
4769 	}
4770 	vfs_busy_pages_release(bp);
4771 	bp->b_resid = 0;
4772 }
4773 
4774 void
4775 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4776 {
4777 	vm_page_t m;
4778 	int i, n;
4779 
4780 	if (buf_mapped(bp)) {
4781 		BUF_CHECK_MAPPED(bp);
4782 		bzero(bp->b_data + base, size);
4783 	} else {
4784 		BUF_CHECK_UNMAPPED(bp);
4785 		n = PAGE_SIZE - (base & PAGE_MASK);
4786 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4787 			m = bp->b_pages[i];
4788 			if (n > size)
4789 				n = size;
4790 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4791 			base += n;
4792 			size -= n;
4793 			n = PAGE_SIZE;
4794 		}
4795 	}
4796 }
4797 
4798 /*
4799  * Update buffer flags based on I/O request parameters, optionally releasing the
4800  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4801  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4802  * I/O).  Otherwise the buffer is released to the cache.
4803  */
4804 static void
4805 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4806 {
4807 
4808 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4809 	    ("buf %p non-VMIO noreuse", bp));
4810 
4811 	if ((ioflag & IO_DIRECT) != 0)
4812 		bp->b_flags |= B_DIRECT;
4813 	if ((ioflag & IO_EXT) != 0)
4814 		bp->b_xflags |= BX_ALTDATA;
4815 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4816 		bp->b_flags |= B_RELBUF;
4817 		if ((ioflag & IO_NOREUSE) != 0)
4818 			bp->b_flags |= B_NOREUSE;
4819 		if (release)
4820 			brelse(bp);
4821 	} else if (release)
4822 		bqrelse(bp);
4823 }
4824 
4825 void
4826 vfs_bio_brelse(struct buf *bp, int ioflag)
4827 {
4828 
4829 	b_io_dismiss(bp, ioflag, true);
4830 }
4831 
4832 void
4833 vfs_bio_set_flags(struct buf *bp, int ioflag)
4834 {
4835 
4836 	b_io_dismiss(bp, ioflag, false);
4837 }
4838 
4839 /*
4840  * vm_hold_load_pages and vm_hold_free_pages get pages into
4841  * a buffers address space.  The pages are anonymous and are
4842  * not associated with a file object.
4843  */
4844 static void
4845 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4846 {
4847 	vm_offset_t pg;
4848 	vm_page_t p;
4849 	int index;
4850 
4851 	BUF_CHECK_MAPPED(bp);
4852 
4853 	to = round_page(to);
4854 	from = round_page(from);
4855 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4856 
4857 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4858 		/*
4859 		 * note: must allocate system pages since blocking here
4860 		 * could interfere with paging I/O, no matter which
4861 		 * process we are.
4862 		 */
4863 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4864 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4865 		    VM_ALLOC_WAITOK);
4866 		pmap_qenter(pg, &p, 1);
4867 		bp->b_pages[index] = p;
4868 	}
4869 	bp->b_npages = index;
4870 }
4871 
4872 /* Return pages associated with this buf to the vm system */
4873 static void
4874 vm_hold_free_pages(struct buf *bp, int newbsize)
4875 {
4876 	vm_offset_t from;
4877 	vm_page_t p;
4878 	int index, newnpages;
4879 
4880 	BUF_CHECK_MAPPED(bp);
4881 
4882 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4883 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4884 	if (bp->b_npages > newnpages)
4885 		pmap_qremove(from, bp->b_npages - newnpages);
4886 	for (index = newnpages; index < bp->b_npages; index++) {
4887 		p = bp->b_pages[index];
4888 		bp->b_pages[index] = NULL;
4889 		vm_page_unwire_noq(p);
4890 		vm_page_free(p);
4891 	}
4892 	bp->b_npages = newnpages;
4893 }
4894 
4895 /*
4896  * Map an IO request into kernel virtual address space.
4897  *
4898  * All requests are (re)mapped into kernel VA space.
4899  * Notice that we use b_bufsize for the size of the buffer
4900  * to be mapped.  b_bcount might be modified by the driver.
4901  *
4902  * Note that even if the caller determines that the address space should
4903  * be valid, a race or a smaller-file mapped into a larger space may
4904  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4905  * check the return value.
4906  *
4907  * This function only works with pager buffers.
4908  */
4909 int
4910 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
4911 {
4912 	vm_prot_t prot;
4913 	int pidx;
4914 
4915 	prot = VM_PROT_READ;
4916 	if (bp->b_iocmd == BIO_READ)
4917 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4918 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4919 	    (vm_offset_t)uaddr, len, prot, bp->b_pages,
4920 	    btoc(MAXPHYS))) < 0)
4921 		return (-1);
4922 	bp->b_bufsize = len;
4923 	bp->b_npages = pidx;
4924 	bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
4925 	if (mapbuf || !unmapped_buf_allowed) {
4926 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4927 		bp->b_data = bp->b_kvabase + bp->b_offset;
4928 	} else
4929 		bp->b_data = unmapped_buf;
4930 	return(0);
4931 }
4932 
4933 /*
4934  * Free the io map PTEs associated with this IO operation.
4935  * We also invalidate the TLB entries and restore the original b_addr.
4936  *
4937  * This function only works with pager buffers.
4938  */
4939 void
4940 vunmapbuf(struct buf *bp)
4941 {
4942 	int npages;
4943 
4944 	npages = bp->b_npages;
4945 	if (buf_mapped(bp))
4946 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4947 	vm_page_unhold_pages(bp->b_pages, npages);
4948 
4949 	bp->b_data = unmapped_buf;
4950 }
4951 
4952 void
4953 bdone(struct buf *bp)
4954 {
4955 	struct mtx *mtxp;
4956 
4957 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4958 	mtx_lock(mtxp);
4959 	bp->b_flags |= B_DONE;
4960 	wakeup(bp);
4961 	mtx_unlock(mtxp);
4962 }
4963 
4964 void
4965 bwait(struct buf *bp, u_char pri, const char *wchan)
4966 {
4967 	struct mtx *mtxp;
4968 
4969 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4970 	mtx_lock(mtxp);
4971 	while ((bp->b_flags & B_DONE) == 0)
4972 		msleep(bp, mtxp, pri, wchan, 0);
4973 	mtx_unlock(mtxp);
4974 }
4975 
4976 int
4977 bufsync(struct bufobj *bo, int waitfor)
4978 {
4979 
4980 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
4981 }
4982 
4983 void
4984 bufstrategy(struct bufobj *bo, struct buf *bp)
4985 {
4986 	int i __unused;
4987 	struct vnode *vp;
4988 
4989 	vp = bp->b_vp;
4990 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4991 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4992 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4993 	i = VOP_STRATEGY(vp, bp);
4994 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4995 }
4996 
4997 /*
4998  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
4999  */
5000 void
5001 bufobj_init(struct bufobj *bo, void *private)
5002 {
5003 	static volatile int bufobj_cleanq;
5004 
5005         bo->bo_domain =
5006             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5007         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5008         bo->bo_private = private;
5009         TAILQ_INIT(&bo->bo_clean.bv_hd);
5010         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5011 }
5012 
5013 void
5014 bufobj_wrefl(struct bufobj *bo)
5015 {
5016 
5017 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5018 	ASSERT_BO_WLOCKED(bo);
5019 	bo->bo_numoutput++;
5020 }
5021 
5022 void
5023 bufobj_wref(struct bufobj *bo)
5024 {
5025 
5026 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5027 	BO_LOCK(bo);
5028 	bo->bo_numoutput++;
5029 	BO_UNLOCK(bo);
5030 }
5031 
5032 void
5033 bufobj_wdrop(struct bufobj *bo)
5034 {
5035 
5036 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5037 	BO_LOCK(bo);
5038 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5039 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5040 		bo->bo_flag &= ~BO_WWAIT;
5041 		wakeup(&bo->bo_numoutput);
5042 	}
5043 	BO_UNLOCK(bo);
5044 }
5045 
5046 int
5047 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5048 {
5049 	int error;
5050 
5051 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5052 	ASSERT_BO_WLOCKED(bo);
5053 	error = 0;
5054 	while (bo->bo_numoutput) {
5055 		bo->bo_flag |= BO_WWAIT;
5056 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5057 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5058 		if (error)
5059 			break;
5060 	}
5061 	return (error);
5062 }
5063 
5064 /*
5065  * Set bio_data or bio_ma for struct bio from the struct buf.
5066  */
5067 void
5068 bdata2bio(struct buf *bp, struct bio *bip)
5069 {
5070 
5071 	if (!buf_mapped(bp)) {
5072 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5073 		bip->bio_ma = bp->b_pages;
5074 		bip->bio_ma_n = bp->b_npages;
5075 		bip->bio_data = unmapped_buf;
5076 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5077 		bip->bio_flags |= BIO_UNMAPPED;
5078 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5079 		    PAGE_SIZE == bp->b_npages,
5080 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5081 		    (long long)bip->bio_length, bip->bio_ma_n));
5082 	} else {
5083 		bip->bio_data = bp->b_data;
5084 		bip->bio_ma = NULL;
5085 	}
5086 }
5087 
5088 /*
5089  * The MIPS pmap code currently doesn't handle aliased pages.
5090  * The VIPT caches may not handle page aliasing themselves, leading
5091  * to data corruption.
5092  *
5093  * As such, this code makes a system extremely unhappy if said
5094  * system doesn't support unaliasing the above situation in hardware.
5095  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5096  * this feature at build time, so it has to be handled in software.
5097  *
5098  * Once the MIPS pmap/cache code grows to support this function on
5099  * earlier chips, it should be flipped back off.
5100  */
5101 #ifdef	__mips__
5102 static int buf_pager_relbuf = 1;
5103 #else
5104 static int buf_pager_relbuf = 0;
5105 #endif
5106 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5107     &buf_pager_relbuf, 0,
5108     "Make buffer pager release buffers after reading");
5109 
5110 /*
5111  * The buffer pager.  It uses buffer reads to validate pages.
5112  *
5113  * In contrast to the generic local pager from vm/vnode_pager.c, this
5114  * pager correctly and easily handles volumes where the underlying
5115  * device block size is greater than the machine page size.  The
5116  * buffer cache transparently extends the requested page run to be
5117  * aligned at the block boundary, and does the necessary bogus page
5118  * replacements in the addends to avoid obliterating already valid
5119  * pages.
5120  *
5121  * The only non-trivial issue is that the exclusive busy state for
5122  * pages, which is assumed by the vm_pager_getpages() interface, is
5123  * incompatible with the VMIO buffer cache's desire to share-busy the
5124  * pages.  This function performs a trivial downgrade of the pages'
5125  * state before reading buffers, and a less trivial upgrade from the
5126  * shared-busy to excl-busy state after the read.
5127  */
5128 int
5129 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5130     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5131     vbg_get_blksize_t get_blksize)
5132 {
5133 	vm_page_t m;
5134 	vm_object_t object;
5135 	struct buf *bp;
5136 	struct mount *mp;
5137 	daddr_t lbn, lbnp;
5138 	vm_ooffset_t la, lb, poff, poffe;
5139 	long bsize;
5140 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5141 	bool redo, lpart;
5142 
5143 	object = vp->v_object;
5144 	mp = vp->v_mount;
5145 	error = 0;
5146 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5147 	if (la >= object->un_pager.vnp.vnp_size)
5148 		return (VM_PAGER_BAD);
5149 
5150 	/*
5151 	 * Change the meaning of la from where the last requested page starts
5152 	 * to where it ends, because that's the end of the requested region
5153 	 * and the start of the potential read-ahead region.
5154 	 */
5155 	la += PAGE_SIZE;
5156 	lpart = la > object->un_pager.vnp.vnp_size;
5157 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5158 
5159 	/*
5160 	 * Calculate read-ahead, behind and total pages.
5161 	 */
5162 	pgsin = count;
5163 	lb = IDX_TO_OFF(ma[0]->pindex);
5164 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5165 	pgsin += pgsin_b;
5166 	if (rbehind != NULL)
5167 		*rbehind = pgsin_b;
5168 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5169 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5170 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5171 		    PAGE_SIZE) - la);
5172 	pgsin += pgsin_a;
5173 	if (rahead != NULL)
5174 		*rahead = pgsin_a;
5175 	VM_CNT_INC(v_vnodein);
5176 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5177 
5178 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5179 	    != 0) ? GB_UNMAPPED : 0;
5180 again:
5181 	for (i = 0; i < count; i++) {
5182 		if (ma[i] != bogus_page)
5183 			vm_page_busy_downgrade(ma[i]);
5184 	}
5185 
5186 	lbnp = -1;
5187 	for (i = 0; i < count; i++) {
5188 		m = ma[i];
5189 		if (m == bogus_page)
5190 			continue;
5191 
5192 		/*
5193 		 * Pages are shared busy and the object lock is not
5194 		 * owned, which together allow for the pages'
5195 		 * invalidation.  The racy test for validity avoids
5196 		 * useless creation of the buffer for the most typical
5197 		 * case when invalidation is not used in redo or for
5198 		 * parallel read.  The shared->excl upgrade loop at
5199 		 * the end of the function catches the race in a
5200 		 * reliable way (protected by the object lock).
5201 		 */
5202 		if (vm_page_all_valid(m))
5203 			continue;
5204 
5205 		poff = IDX_TO_OFF(m->pindex);
5206 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5207 		for (; poff < poffe; poff += bsize) {
5208 			lbn = get_lblkno(vp, poff);
5209 			if (lbn == lbnp)
5210 				goto next_page;
5211 			lbnp = lbn;
5212 
5213 			bsize = get_blksize(vp, lbn);
5214 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5215 			    br_flags, &bp);
5216 			if (error != 0)
5217 				goto end_pages;
5218 			if (bp->b_rcred == curthread->td_ucred) {
5219 				crfree(bp->b_rcred);
5220 				bp->b_rcred = NOCRED;
5221 			}
5222 			if (LIST_EMPTY(&bp->b_dep)) {
5223 				/*
5224 				 * Invalidation clears m->valid, but
5225 				 * may leave B_CACHE flag if the
5226 				 * buffer existed at the invalidation
5227 				 * time.  In this case, recycle the
5228 				 * buffer to do real read on next
5229 				 * bread() after redo.
5230 				 *
5231 				 * Otherwise B_RELBUF is not strictly
5232 				 * necessary, enable to reduce buf
5233 				 * cache pressure.
5234 				 */
5235 				if (buf_pager_relbuf ||
5236 				    !vm_page_all_valid(m))
5237 					bp->b_flags |= B_RELBUF;
5238 
5239 				bp->b_flags &= ~B_NOCACHE;
5240 				brelse(bp);
5241 			} else {
5242 				bqrelse(bp);
5243 			}
5244 		}
5245 		KASSERT(1 /* racy, enable for debugging */ ||
5246 		    vm_page_all_valid(m) || i == count - 1,
5247 		    ("buf %d %p invalid", i, m));
5248 		if (i == count - 1 && lpart) {
5249 			if (!vm_page_none_valid(m) &&
5250 			    !vm_page_all_valid(m))
5251 				vm_page_zero_invalid(m, TRUE);
5252 		}
5253 next_page:;
5254 	}
5255 end_pages:
5256 
5257 	redo = false;
5258 	for (i = 0; i < count; i++) {
5259 		if (ma[i] == bogus_page)
5260 			continue;
5261 		if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5262 			vm_page_sunbusy(ma[i]);
5263 			ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5264 			    VM_ALLOC_NORMAL);
5265 		}
5266 
5267 		/*
5268 		 * Since the pages were only sbusy while neither the
5269 		 * buffer nor the object lock was held by us, or
5270 		 * reallocated while vm_page_grab() slept for busy
5271 		 * relinguish, they could have been invalidated.
5272 		 * Recheck the valid bits and re-read as needed.
5273 		 *
5274 		 * Note that the last page is made fully valid in the
5275 		 * read loop, and partial validity for the page at
5276 		 * index count - 1 could mean that the page was
5277 		 * invalidated or removed, so we must restart for
5278 		 * safety as well.
5279 		 */
5280 		if (!vm_page_all_valid(ma[i]))
5281 			redo = true;
5282 	}
5283 	if (redo && error == 0)
5284 		goto again;
5285 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5286 }
5287 
5288 #include "opt_ddb.h"
5289 #ifdef DDB
5290 #include <ddb/ddb.h>
5291 
5292 /* DDB command to show buffer data */
5293 DB_SHOW_COMMAND(buffer, db_show_buffer)
5294 {
5295 	/* get args */
5296 	struct buf *bp = (struct buf *)addr;
5297 #ifdef FULL_BUF_TRACKING
5298 	uint32_t i, j;
5299 #endif
5300 
5301 	if (!have_addr) {
5302 		db_printf("usage: show buffer <addr>\n");
5303 		return;
5304 	}
5305 
5306 	db_printf("buf at %p\n", bp);
5307 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5308 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5309 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5310 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5311 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5312 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5313 	db_printf(
5314 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5315 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5316 	    "b_vp = %p, b_dep = %p\n",
5317 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5318 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5319 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5320 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5321 	    bp->b_kvabase, bp->b_kvasize);
5322 	if (bp->b_npages) {
5323 		int i;
5324 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5325 		for (i = 0; i < bp->b_npages; i++) {
5326 			vm_page_t m;
5327 			m = bp->b_pages[i];
5328 			if (m != NULL)
5329 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5330 				    (u_long)m->pindex,
5331 				    (u_long)VM_PAGE_TO_PHYS(m));
5332 			else
5333 				db_printf("( ??? )");
5334 			if ((i + 1) < bp->b_npages)
5335 				db_printf(",");
5336 		}
5337 		db_printf("\n");
5338 	}
5339 	BUF_LOCKPRINTINFO(bp);
5340 #if defined(FULL_BUF_TRACKING)
5341 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5342 
5343 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5344 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5345 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5346 			continue;
5347 		db_printf(" %2u: %s\n", j,
5348 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5349 	}
5350 #elif defined(BUF_TRACKING)
5351 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5352 #endif
5353 	db_printf(" ");
5354 }
5355 
5356 DB_SHOW_COMMAND(bufqueues, bufqueues)
5357 {
5358 	struct bufdomain *bd;
5359 	struct buf *bp;
5360 	long total;
5361 	int i, j, cnt;
5362 
5363 	db_printf("bqempty: %d\n", bqempty.bq_len);
5364 
5365 	for (i = 0; i < buf_domains; i++) {
5366 		bd = &bdomain[i];
5367 		db_printf("Buf domain %d\n", i);
5368 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5369 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5370 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5371 		db_printf("\n");
5372 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5373 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5374 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5375 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5376 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5377 		db_printf("\n");
5378 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5379 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5380 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5381 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5382 		db_printf("\n");
5383 		total = 0;
5384 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5385 			total += bp->b_bufsize;
5386 		db_printf("\tcleanq count\t%d (%ld)\n",
5387 		    bd->bd_cleanq->bq_len, total);
5388 		total = 0;
5389 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5390 			total += bp->b_bufsize;
5391 		db_printf("\tdirtyq count\t%d (%ld)\n",
5392 		    bd->bd_dirtyq.bq_len, total);
5393 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5394 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5395 		db_printf("\tCPU ");
5396 		for (j = 0; j <= mp_maxid; j++)
5397 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5398 		db_printf("\n");
5399 		cnt = 0;
5400 		total = 0;
5401 		for (j = 0; j < nbuf; j++)
5402 			if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5403 				cnt++;
5404 				total += buf[j].b_bufsize;
5405 			}
5406 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5407 		cnt = 0;
5408 		total = 0;
5409 		for (j = 0; j < nbuf; j++)
5410 			if (buf[j].b_domain == i) {
5411 				cnt++;
5412 				total += buf[j].b_bufsize;
5413 			}
5414 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5415 	}
5416 }
5417 
5418 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5419 {
5420 	struct buf *bp;
5421 	int i;
5422 
5423 	for (i = 0; i < nbuf; i++) {
5424 		bp = &buf[i];
5425 		if (BUF_ISLOCKED(bp)) {
5426 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5427 			db_printf("\n");
5428 			if (db_pager_quit)
5429 				break;
5430 		}
5431 	}
5432 }
5433 
5434 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5435 {
5436 	struct vnode *vp;
5437 	struct buf *bp;
5438 
5439 	if (!have_addr) {
5440 		db_printf("usage: show vnodebufs <addr>\n");
5441 		return;
5442 	}
5443 	vp = (struct vnode *)addr;
5444 	db_printf("Clean buffers:\n");
5445 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5446 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5447 		db_printf("\n");
5448 	}
5449 	db_printf("Dirty buffers:\n");
5450 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5451 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5452 		db_printf("\n");
5453 	}
5454 }
5455 
5456 DB_COMMAND(countfreebufs, db_coundfreebufs)
5457 {
5458 	struct buf *bp;
5459 	int i, used = 0, nfree = 0;
5460 
5461 	if (have_addr) {
5462 		db_printf("usage: countfreebufs\n");
5463 		return;
5464 	}
5465 
5466 	for (i = 0; i < nbuf; i++) {
5467 		bp = &buf[i];
5468 		if (bp->b_qindex == QUEUE_EMPTY)
5469 			nfree++;
5470 		else
5471 			used++;
5472 	}
5473 
5474 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5475 	    nfree + used);
5476 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5477 }
5478 #endif /* DDB */
5479