xref: /freebsd/sys/kern/vfs_bio.c (revision 732a02b4e77866604a120a275c082bb6221bd2ff)
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 int inmem(struct vnode *vp, daddr_t blkno);
158 static void vm_hold_free_pages(struct buf *bp, int newbsize);
159 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
160 		vm_offset_t to);
161 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
162 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
163 		vm_page_t m);
164 static void vfs_clean_pages_dirty_buf(struct buf *bp);
165 static void vfs_setdirty_range(struct buf *bp);
166 static void vfs_vmio_invalidate(struct buf *bp);
167 static void vfs_vmio_truncate(struct buf *bp, int npages);
168 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
169 static int vfs_bio_clcheck(struct vnode *vp, int size,
170 		daddr_t lblkno, daddr_t blkno);
171 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
172 		void (*)(struct buf *));
173 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
174 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
175 static void buf_daemon(void);
176 static __inline void bd_wakeup(void);
177 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
178 static void bufkva_reclaim(vmem_t *, int);
179 static void bufkva_free(struct buf *);
180 static int buf_import(void *, void **, int, int, int);
181 static void buf_release(void *, void **, int);
182 static void maxbcachebuf_adjust(void);
183 static inline struct bufdomain *bufdomain(struct buf *);
184 static void bq_remove(struct bufqueue *bq, struct buf *bp);
185 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
186 static int buf_recycle(struct bufdomain *, bool kva);
187 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
188 	    const char *lockname);
189 static void bd_init(struct bufdomain *bd);
190 static int bd_flushall(struct bufdomain *bd);
191 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
192 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
193 
194 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
195 int vmiodirenable = TRUE;
196 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
197     "Use the VM system for directory writes");
198 long runningbufspace;
199 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
200     "Amount of presently outstanding async buffer io");
201 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
202     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
203 static counter_u64_t bufkvaspace;
204 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
205     "Kernel virtual memory used for buffers");
206 static long maxbufspace;
207 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
208     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
209     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
210     "Maximum allowed value of bufspace (including metadata)");
211 static long bufmallocspace;
212 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
213     "Amount of malloced memory for buffers");
214 static long maxbufmallocspace;
215 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
216     0, "Maximum amount of malloced memory for buffers");
217 static long lobufspace;
218 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
219     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
220     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
221     "Minimum amount of buffers we want to have");
222 long hibufspace;
223 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
224     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
225     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
226     "Maximum allowed value of bufspace (excluding metadata)");
227 long bufspacethresh;
228 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
229     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
230     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
231     "Bufspace consumed before waking the daemon to free some");
232 static counter_u64_t buffreekvacnt;
233 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
234     "Number of times we have freed the KVA space from some buffer");
235 static counter_u64_t bufdefragcnt;
236 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
237     "Number of times we have had to repeat buffer allocation to defragment");
238 static long lorunningspace;
239 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
240     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
241     "Minimum preferred space used for in-progress I/O");
242 static long hirunningspace;
243 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
244     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
245     "Maximum amount of space to use for in-progress I/O");
246 int dirtybufferflushes;
247 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
248     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
249 int bdwriteskip;
250 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
251     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
252 int altbufferflushes;
253 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
254     &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
255 static int recursiveflushes;
256 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
257     &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
258 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
259 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
260     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
261     "Number of buffers that are dirty (has unwritten changes) at the moment");
262 static int lodirtybuffers;
263 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
264     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
265     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
266     "How many buffers we want to have free before bufdaemon can sleep");
267 static int hidirtybuffers;
268 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
269     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
270     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
271     "When the number of dirty buffers is considered severe");
272 int dirtybufthresh;
273 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
274     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
275     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
276     "Number of bdwrite to bawrite conversions to clear dirty buffers");
277 static int numfreebuffers;
278 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
279     "Number of free buffers");
280 static int lofreebuffers;
281 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
282     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
283     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
284    "Target number of free buffers");
285 static int hifreebuffers;
286 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
287     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
288     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
289    "Threshold for clean buffer recycling");
290 static counter_u64_t getnewbufcalls;
291 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
292    &getnewbufcalls, "Number of calls to getnewbuf");
293 static counter_u64_t getnewbufrestarts;
294 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
295     &getnewbufrestarts,
296     "Number of times getnewbuf has had to restart a buffer acquisition");
297 static counter_u64_t mappingrestarts;
298 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
299     &mappingrestarts,
300     "Number of times getblk has had to restart a buffer mapping for "
301     "unmapped buffer");
302 static counter_u64_t numbufallocfails;
303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
304     &numbufallocfails, "Number of times buffer allocations failed");
305 static int flushbufqtarget = 100;
306 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
307     "Amount of work to do in flushbufqueues when helping bufdaemon");
308 static counter_u64_t notbufdflushes;
309 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
310     "Number of dirty buffer flushes done by the bufdaemon helpers");
311 static long barrierwrites;
312 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
313     &barrierwrites, 0, "Number of barrier writes");
314 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
315     &unmapped_buf_allowed, 0,
316     "Permit the use of the unmapped i/o");
317 int maxbcachebuf = MAXBCACHEBUF;
318 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
319     "Maximum size of a buffer cache block");
320 
321 /*
322  * This lock synchronizes access to bd_request.
323  */
324 static struct mtx_padalign __exclusive_cache_line bdlock;
325 
326 /*
327  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
328  * waitrunningbufspace().
329  */
330 static struct mtx_padalign __exclusive_cache_line rbreqlock;
331 
332 /*
333  * Lock that protects bdirtywait.
334  */
335 static struct mtx_padalign __exclusive_cache_line bdirtylock;
336 
337 /*
338  * Wakeup point for bufdaemon, as well as indicator of whether it is already
339  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
340  * is idling.
341  */
342 static int bd_request;
343 
344 /*
345  * Request for the buf daemon to write more buffers than is indicated by
346  * lodirtybuf.  This may be necessary to push out excess dependencies or
347  * defragment the address space where a simple count of the number of dirty
348  * buffers is insufficient to characterize the demand for flushing them.
349  */
350 static int bd_speedupreq;
351 
352 /*
353  * Synchronization (sleep/wakeup) variable for active buffer space requests.
354  * Set when wait starts, cleared prior to wakeup().
355  * Used in runningbufwakeup() and waitrunningbufspace().
356  */
357 static int runningbufreq;
358 
359 /*
360  * Synchronization for bwillwrite() waiters.
361  */
362 static int bdirtywait;
363 
364 /*
365  * Definitions for the buffer free lists.
366  */
367 #define QUEUE_NONE	0	/* on no queue */
368 #define QUEUE_EMPTY	1	/* empty buffer headers */
369 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
370 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
371 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
372 
373 /* Maximum number of buffer domains. */
374 #define	BUF_DOMAINS	8
375 
376 struct bufdomainset bdlodirty;		/* Domains > lodirty */
377 struct bufdomainset bdhidirty;		/* Domains > hidirty */
378 
379 /* Configured number of clean queues. */
380 static int __read_mostly buf_domains;
381 
382 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
383 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
384 struct bufqueue __exclusive_cache_line bqempty;
385 
386 /*
387  * per-cpu empty buffer cache.
388  */
389 uma_zone_t buf_zone;
390 
391 /*
392  * Single global constant for BUF_WMESG, to avoid getting multiple references.
393  * buf_wmesg is referred from macros.
394  */
395 const char *buf_wmesg = BUF_WMESG;
396 
397 static int
398 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
399 {
400 	long value;
401 	int error;
402 
403 	value = *(long *)arg1;
404 	error = sysctl_handle_long(oidp, &value, 0, req);
405 	if (error != 0 || req->newptr == NULL)
406 		return (error);
407 	mtx_lock(&rbreqlock);
408 	if (arg1 == &hirunningspace) {
409 		if (value < lorunningspace)
410 			error = EINVAL;
411 		else
412 			hirunningspace = value;
413 	} else {
414 		KASSERT(arg1 == &lorunningspace,
415 		    ("%s: unknown arg1", __func__));
416 		if (value > hirunningspace)
417 			error = EINVAL;
418 		else
419 			lorunningspace = value;
420 	}
421 	mtx_unlock(&rbreqlock);
422 	return (error);
423 }
424 
425 static int
426 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
427 {
428 	int error;
429 	int value;
430 	int i;
431 
432 	value = *(int *)arg1;
433 	error = sysctl_handle_int(oidp, &value, 0, req);
434 	if (error != 0 || req->newptr == NULL)
435 		return (error);
436 	*(int *)arg1 = value;
437 	for (i = 0; i < buf_domains; i++)
438 		*(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
439 		    value / buf_domains;
440 
441 	return (error);
442 }
443 
444 static int
445 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
446 {
447 	long value;
448 	int error;
449 	int i;
450 
451 	value = *(long *)arg1;
452 	error = sysctl_handle_long(oidp, &value, 0, req);
453 	if (error != 0 || req->newptr == NULL)
454 		return (error);
455 	*(long *)arg1 = value;
456 	for (i = 0; i < buf_domains; i++)
457 		*(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
458 		    value / buf_domains;
459 
460 	return (error);
461 }
462 
463 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
464     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
465 static int
466 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
467 {
468 	long lvalue;
469 	int ivalue;
470 	int i;
471 
472 	lvalue = 0;
473 	for (i = 0; i < buf_domains; i++)
474 		lvalue += bdomain[i].bd_bufspace;
475 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
476 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
477 	if (lvalue > INT_MAX)
478 		/* On overflow, still write out a long to trigger ENOMEM. */
479 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
480 	ivalue = lvalue;
481 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
482 }
483 #else
484 static int
485 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
486 {
487 	long lvalue;
488 	int i;
489 
490 	lvalue = 0;
491 	for (i = 0; i < buf_domains; i++)
492 		lvalue += bdomain[i].bd_bufspace;
493 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
494 }
495 #endif
496 
497 static int
498 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
499 {
500 	int value;
501 	int i;
502 
503 	value = 0;
504 	for (i = 0; i < buf_domains; i++)
505 		value += bdomain[i].bd_numdirtybuffers;
506 	return (sysctl_handle_int(oidp, &value, 0, req));
507 }
508 
509 /*
510  *	bdirtywakeup:
511  *
512  *	Wakeup any bwillwrite() waiters.
513  */
514 static void
515 bdirtywakeup(void)
516 {
517 	mtx_lock(&bdirtylock);
518 	if (bdirtywait) {
519 		bdirtywait = 0;
520 		wakeup(&bdirtywait);
521 	}
522 	mtx_unlock(&bdirtylock);
523 }
524 
525 /*
526  *	bd_clear:
527  *
528  *	Clear a domain from the appropriate bitsets when dirtybuffers
529  *	is decremented.
530  */
531 static void
532 bd_clear(struct bufdomain *bd)
533 {
534 
535 	mtx_lock(&bdirtylock);
536 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
537 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
538 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
539 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
540 	mtx_unlock(&bdirtylock);
541 }
542 
543 /*
544  *	bd_set:
545  *
546  *	Set a domain in the appropriate bitsets when dirtybuffers
547  *	is incremented.
548  */
549 static void
550 bd_set(struct bufdomain *bd)
551 {
552 
553 	mtx_lock(&bdirtylock);
554 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
555 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
556 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
557 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
558 	mtx_unlock(&bdirtylock);
559 }
560 
561 /*
562  *	bdirtysub:
563  *
564  *	Decrement the numdirtybuffers count by one and wakeup any
565  *	threads blocked in bwillwrite().
566  */
567 static void
568 bdirtysub(struct buf *bp)
569 {
570 	struct bufdomain *bd;
571 	int num;
572 
573 	bd = bufdomain(bp);
574 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
575 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
576 		bdirtywakeup();
577 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
578 		bd_clear(bd);
579 }
580 
581 /*
582  *	bdirtyadd:
583  *
584  *	Increment the numdirtybuffers count by one and wakeup the buf
585  *	daemon if needed.
586  */
587 static void
588 bdirtyadd(struct buf *bp)
589 {
590 	struct bufdomain *bd;
591 	int num;
592 
593 	/*
594 	 * Only do the wakeup once as we cross the boundary.  The
595 	 * buf daemon will keep running until the condition clears.
596 	 */
597 	bd = bufdomain(bp);
598 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
599 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
600 		bd_wakeup();
601 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
602 		bd_set(bd);
603 }
604 
605 /*
606  *	bufspace_daemon_wakeup:
607  *
608  *	Wakeup the daemons responsible for freeing clean bufs.
609  */
610 static void
611 bufspace_daemon_wakeup(struct bufdomain *bd)
612 {
613 
614 	/*
615 	 * avoid the lock if the daemon is running.
616 	 */
617 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
618 		BD_RUN_LOCK(bd);
619 		atomic_store_int(&bd->bd_running, 1);
620 		wakeup(&bd->bd_running);
621 		BD_RUN_UNLOCK(bd);
622 	}
623 }
624 
625 /*
626  *	bufspace_daemon_wait:
627  *
628  *	Sleep until the domain falls below a limit or one second passes.
629  */
630 static void
631 bufspace_daemon_wait(struct bufdomain *bd)
632 {
633 	/*
634 	 * Re-check our limits and sleep.  bd_running must be
635 	 * cleared prior to checking the limits to avoid missed
636 	 * wakeups.  The waker will adjust one of bufspace or
637 	 * freebuffers prior to checking bd_running.
638 	 */
639 	BD_RUN_LOCK(bd);
640 	atomic_store_int(&bd->bd_running, 0);
641 	if (bd->bd_bufspace < bd->bd_bufspacethresh &&
642 	    bd->bd_freebuffers > bd->bd_lofreebuffers) {
643 		msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
644 		    "-", hz);
645 	} else {
646 		/* Avoid spurious wakeups while running. */
647 		atomic_store_int(&bd->bd_running, 1);
648 		BD_RUN_UNLOCK(bd);
649 	}
650 }
651 
652 /*
653  *	bufspace_adjust:
654  *
655  *	Adjust the reported bufspace for a KVA managed buffer, possibly
656  * 	waking any waiters.
657  */
658 static void
659 bufspace_adjust(struct buf *bp, int bufsize)
660 {
661 	struct bufdomain *bd;
662 	long space;
663 	int diff;
664 
665 	KASSERT((bp->b_flags & B_MALLOC) == 0,
666 	    ("bufspace_adjust: malloc buf %p", bp));
667 	bd = bufdomain(bp);
668 	diff = bufsize - bp->b_bufsize;
669 	if (diff < 0) {
670 		atomic_subtract_long(&bd->bd_bufspace, -diff);
671 	} else if (diff > 0) {
672 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
673 		/* Wake up the daemon on the transition. */
674 		if (space < bd->bd_bufspacethresh &&
675 		    space + diff >= bd->bd_bufspacethresh)
676 			bufspace_daemon_wakeup(bd);
677 	}
678 	bp->b_bufsize = bufsize;
679 }
680 
681 /*
682  *	bufspace_reserve:
683  *
684  *	Reserve bufspace before calling allocbuf().  metadata has a
685  *	different space limit than data.
686  */
687 static int
688 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
689 {
690 	long limit, new;
691 	long space;
692 
693 	if (metadata)
694 		limit = bd->bd_maxbufspace;
695 	else
696 		limit = bd->bd_hibufspace;
697 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
698 	new = space + size;
699 	if (new > limit) {
700 		atomic_subtract_long(&bd->bd_bufspace, size);
701 		return (ENOSPC);
702 	}
703 
704 	/* Wake up the daemon on the transition. */
705 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
706 		bufspace_daemon_wakeup(bd);
707 
708 	return (0);
709 }
710 
711 /*
712  *	bufspace_release:
713  *
714  *	Release reserved bufspace after bufspace_adjust() has consumed it.
715  */
716 static void
717 bufspace_release(struct bufdomain *bd, int size)
718 {
719 
720 	atomic_subtract_long(&bd->bd_bufspace, size);
721 }
722 
723 /*
724  *	bufspace_wait:
725  *
726  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
727  *	supplied.  bd_wanted must be set prior to polling for space.  The
728  *	operation must be re-tried on return.
729  */
730 static void
731 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
732     int slpflag, int slptimeo)
733 {
734 	struct thread *td;
735 	int error, fl, norunbuf;
736 
737 	if ((gbflags & GB_NOWAIT_BD) != 0)
738 		return;
739 
740 	td = curthread;
741 	BD_LOCK(bd);
742 	while (bd->bd_wanted) {
743 		if (vp != NULL && vp->v_type != VCHR &&
744 		    (td->td_pflags & TDP_BUFNEED) == 0) {
745 			BD_UNLOCK(bd);
746 			/*
747 			 * getblk() is called with a vnode locked, and
748 			 * some majority of the dirty buffers may as
749 			 * well belong to the vnode.  Flushing the
750 			 * buffers there would make a progress that
751 			 * cannot be achieved by the buf_daemon, that
752 			 * cannot lock the vnode.
753 			 */
754 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
755 			    (td->td_pflags & TDP_NORUNNINGBUF);
756 
757 			/*
758 			 * Play bufdaemon.  The getnewbuf() function
759 			 * may be called while the thread owns lock
760 			 * for another dirty buffer for the same
761 			 * vnode, which makes it impossible to use
762 			 * VOP_FSYNC() there, due to the buffer lock
763 			 * recursion.
764 			 */
765 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
766 			fl = buf_flush(vp, bd, flushbufqtarget);
767 			td->td_pflags &= norunbuf;
768 			BD_LOCK(bd);
769 			if (fl != 0)
770 				continue;
771 			if (bd->bd_wanted == 0)
772 				break;
773 		}
774 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
775 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
776 		if (error != 0)
777 			break;
778 	}
779 	BD_UNLOCK(bd);
780 }
781 
782 /*
783  *	bufspace_daemon:
784  *
785  *	buffer space management daemon.  Tries to maintain some marginal
786  *	amount of free buffer space so that requesting processes neither
787  *	block nor work to reclaim buffers.
788  */
789 static void
790 bufspace_daemon(void *arg)
791 {
792 	struct bufdomain *bd;
793 
794 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
795 	    SHUTDOWN_PRI_LAST + 100);
796 
797 	bd = arg;
798 	for (;;) {
799 		kthread_suspend_check();
800 
801 		/*
802 		 * Free buffers from the clean queue until we meet our
803 		 * targets.
804 		 *
805 		 * Theory of operation:  The buffer cache is most efficient
806 		 * when some free buffer headers and space are always
807 		 * available to getnewbuf().  This daemon attempts to prevent
808 		 * the excessive blocking and synchronization associated
809 		 * with shortfall.  It goes through three phases according
810 		 * demand:
811 		 *
812 		 * 1)	The daemon wakes up voluntarily once per-second
813 		 *	during idle periods when the counters are below
814 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
815 		 *
816 		 * 2)	The daemon wakes up as we cross the thresholds
817 		 *	ahead of any potential blocking.  This may bounce
818 		 *	slightly according to the rate of consumption and
819 		 *	release.
820 		 *
821 		 * 3)	The daemon and consumers are starved for working
822 		 *	clean buffers.  This is the 'bufspace' sleep below
823 		 *	which will inefficiently trade bufs with bqrelse
824 		 *	until we return to condition 2.
825 		 */
826 		while (bd->bd_bufspace > bd->bd_lobufspace ||
827 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
828 			if (buf_recycle(bd, false) != 0) {
829 				if (bd_flushall(bd))
830 					continue;
831 				/*
832 				 * Speedup dirty if we've run out of clean
833 				 * buffers.  This is possible in particular
834 				 * because softdep may held many bufs locked
835 				 * pending writes to other bufs which are
836 				 * marked for delayed write, exhausting
837 				 * clean space until they are written.
838 				 */
839 				bd_speedup();
840 				BD_LOCK(bd);
841 				if (bd->bd_wanted) {
842 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
843 					    PRIBIO|PDROP, "bufspace", hz/10);
844 				} else
845 					BD_UNLOCK(bd);
846 			}
847 			maybe_yield();
848 		}
849 		bufspace_daemon_wait(bd);
850 	}
851 }
852 
853 /*
854  *	bufmallocadjust:
855  *
856  *	Adjust the reported bufspace for a malloc managed buffer, possibly
857  *	waking any waiters.
858  */
859 static void
860 bufmallocadjust(struct buf *bp, int bufsize)
861 {
862 	int diff;
863 
864 	KASSERT((bp->b_flags & B_MALLOC) != 0,
865 	    ("bufmallocadjust: non-malloc buf %p", bp));
866 	diff = bufsize - bp->b_bufsize;
867 	if (diff < 0)
868 		atomic_subtract_long(&bufmallocspace, -diff);
869 	else
870 		atomic_add_long(&bufmallocspace, diff);
871 	bp->b_bufsize = bufsize;
872 }
873 
874 /*
875  *	runningwakeup:
876  *
877  *	Wake up processes that are waiting on asynchronous writes to fall
878  *	below lorunningspace.
879  */
880 static void
881 runningwakeup(void)
882 {
883 
884 	mtx_lock(&rbreqlock);
885 	if (runningbufreq) {
886 		runningbufreq = 0;
887 		wakeup(&runningbufreq);
888 	}
889 	mtx_unlock(&rbreqlock);
890 }
891 
892 /*
893  *	runningbufwakeup:
894  *
895  *	Decrement the outstanding write count according.
896  */
897 void
898 runningbufwakeup(struct buf *bp)
899 {
900 	long space, bspace;
901 
902 	bspace = bp->b_runningbufspace;
903 	if (bspace == 0)
904 		return;
905 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
906 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
907 	    space, bspace));
908 	bp->b_runningbufspace = 0;
909 	/*
910 	 * Only acquire the lock and wakeup on the transition from exceeding
911 	 * the threshold to falling below it.
912 	 */
913 	if (space < lorunningspace)
914 		return;
915 	if (space - bspace > lorunningspace)
916 		return;
917 	runningwakeup();
918 }
919 
920 /*
921  *	waitrunningbufspace()
922  *
923  *	runningbufspace is a measure of the amount of I/O currently
924  *	running.  This routine is used in async-write situations to
925  *	prevent creating huge backups of pending writes to a device.
926  *	Only asynchronous writes are governed by this function.
927  *
928  *	This does NOT turn an async write into a sync write.  It waits
929  *	for earlier writes to complete and generally returns before the
930  *	caller's write has reached the device.
931  */
932 void
933 waitrunningbufspace(void)
934 {
935 
936 	mtx_lock(&rbreqlock);
937 	while (runningbufspace > hirunningspace) {
938 		runningbufreq = 1;
939 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
940 	}
941 	mtx_unlock(&rbreqlock);
942 }
943 
944 /*
945  *	vfs_buf_test_cache:
946  *
947  *	Called when a buffer is extended.  This function clears the B_CACHE
948  *	bit if the newly extended portion of the buffer does not contain
949  *	valid data.
950  */
951 static __inline void
952 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
953     vm_offset_t size, vm_page_t m)
954 {
955 
956 	/*
957 	 * This function and its results are protected by higher level
958 	 * synchronization requiring vnode and buf locks to page in and
959 	 * validate pages.
960 	 */
961 	if (bp->b_flags & B_CACHE) {
962 		int base = (foff + off) & PAGE_MASK;
963 		if (vm_page_is_valid(m, base, size) == 0)
964 			bp->b_flags &= ~B_CACHE;
965 	}
966 }
967 
968 /* Wake up the buffer daemon if necessary */
969 static void
970 bd_wakeup(void)
971 {
972 
973 	mtx_lock(&bdlock);
974 	if (bd_request == 0) {
975 		bd_request = 1;
976 		wakeup(&bd_request);
977 	}
978 	mtx_unlock(&bdlock);
979 }
980 
981 /*
982  * Adjust the maxbcachbuf tunable.
983  */
984 static void
985 maxbcachebuf_adjust(void)
986 {
987 	int i;
988 
989 	/*
990 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
991 	 */
992 	i = 2;
993 	while (i * 2 <= maxbcachebuf)
994 		i *= 2;
995 	maxbcachebuf = i;
996 	if (maxbcachebuf < MAXBSIZE)
997 		maxbcachebuf = MAXBSIZE;
998 	if (maxbcachebuf > MAXPHYS)
999 		maxbcachebuf = MAXPHYS;
1000 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1001 		printf("maxbcachebuf=%d\n", maxbcachebuf);
1002 }
1003 
1004 /*
1005  * bd_speedup - speedup the buffer cache flushing code
1006  */
1007 void
1008 bd_speedup(void)
1009 {
1010 	int needwake;
1011 
1012 	mtx_lock(&bdlock);
1013 	needwake = 0;
1014 	if (bd_speedupreq == 0 || bd_request == 0)
1015 		needwake = 1;
1016 	bd_speedupreq = 1;
1017 	bd_request = 1;
1018 	if (needwake)
1019 		wakeup(&bd_request);
1020 	mtx_unlock(&bdlock);
1021 }
1022 
1023 #ifdef __i386__
1024 #define	TRANSIENT_DENOM	5
1025 #else
1026 #define	TRANSIENT_DENOM 10
1027 #endif
1028 
1029 /*
1030  * Calculating buffer cache scaling values and reserve space for buffer
1031  * headers.  This is called during low level kernel initialization and
1032  * may be called more then once.  We CANNOT write to the memory area
1033  * being reserved at this time.
1034  */
1035 caddr_t
1036 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1037 {
1038 	int tuned_nbuf;
1039 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1040 
1041 	/*
1042 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1043 	 * PAGE_SIZE is >= 1K)
1044 	 */
1045 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1046 
1047 	maxbcachebuf_adjust();
1048 	/*
1049 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1050 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1051 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1052 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1053 	 * the buffer cache we limit the eventual kva reservation to
1054 	 * maxbcache bytes.
1055 	 *
1056 	 * factor represents the 1/4 x ram conversion.
1057 	 */
1058 	if (nbuf == 0) {
1059 		int factor = 4 * BKVASIZE / 1024;
1060 
1061 		nbuf = 50;
1062 		if (physmem_est > 4096)
1063 			nbuf += min((physmem_est - 4096) / factor,
1064 			    65536 / factor);
1065 		if (physmem_est > 65536)
1066 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1067 			    32 * 1024 * 1024 / (factor * 5));
1068 
1069 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1070 			nbuf = maxbcache / BKVASIZE;
1071 		tuned_nbuf = 1;
1072 	} else
1073 		tuned_nbuf = 0;
1074 
1075 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1076 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1077 	if (nbuf > maxbuf) {
1078 		if (!tuned_nbuf)
1079 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1080 			    maxbuf);
1081 		nbuf = maxbuf;
1082 	}
1083 
1084 	/*
1085 	 * Ideal allocation size for the transient bio submap is 10%
1086 	 * of the maximal space buffer map.  This roughly corresponds
1087 	 * to the amount of the buffer mapped for typical UFS load.
1088 	 *
1089 	 * Clip the buffer map to reserve space for the transient
1090 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1091 	 * maximum buffer map extent on the platform.
1092 	 *
1093 	 * The fall-back to the maxbuf in case of maxbcache unset,
1094 	 * allows to not trim the buffer KVA for the architectures
1095 	 * with ample KVA space.
1096 	 */
1097 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1098 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1099 		buf_sz = (long)nbuf * BKVASIZE;
1100 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1101 		    (TRANSIENT_DENOM - 1)) {
1102 			/*
1103 			 * There is more KVA than memory.  Do not
1104 			 * adjust buffer map size, and assign the rest
1105 			 * of maxbuf to transient map.
1106 			 */
1107 			biotmap_sz = maxbuf_sz - buf_sz;
1108 		} else {
1109 			/*
1110 			 * Buffer map spans all KVA we could afford on
1111 			 * this platform.  Give 10% (20% on i386) of
1112 			 * the buffer map to the transient bio map.
1113 			 */
1114 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1115 			buf_sz -= biotmap_sz;
1116 		}
1117 		if (biotmap_sz / INT_MAX > MAXPHYS)
1118 			bio_transient_maxcnt = INT_MAX;
1119 		else
1120 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1121 		/*
1122 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1123 		 * using the transient mapping.
1124 		 */
1125 		if (bio_transient_maxcnt > 1024)
1126 			bio_transient_maxcnt = 1024;
1127 		if (tuned_nbuf)
1128 			nbuf = buf_sz / BKVASIZE;
1129 	}
1130 
1131 	if (nswbuf == 0) {
1132 		nswbuf = min(nbuf / 4, 256);
1133 		if (nswbuf < NSWBUF_MIN)
1134 			nswbuf = NSWBUF_MIN;
1135 	}
1136 
1137 	/*
1138 	 * Reserve space for the buffer cache buffers
1139 	 */
1140 	buf = (void *)v;
1141 	v = (caddr_t)(buf + nbuf);
1142 
1143 	return(v);
1144 }
1145 
1146 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1147 void
1148 bufinit(void)
1149 {
1150 	struct buf *bp;
1151 	int i;
1152 
1153 	KASSERT(maxbcachebuf >= MAXBSIZE,
1154 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1155 	    MAXBSIZE));
1156 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1157 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1158 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1159 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1160 
1161 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1162 
1163 	/* finally, initialize each buffer header and stick on empty q */
1164 	for (i = 0; i < nbuf; i++) {
1165 		bp = &buf[i];
1166 		bzero(bp, sizeof *bp);
1167 		bp->b_flags = B_INVAL;
1168 		bp->b_rcred = NOCRED;
1169 		bp->b_wcred = NOCRED;
1170 		bp->b_qindex = QUEUE_NONE;
1171 		bp->b_domain = -1;
1172 		bp->b_subqueue = mp_maxid + 1;
1173 		bp->b_xflags = 0;
1174 		bp->b_data = bp->b_kvabase = unmapped_buf;
1175 		LIST_INIT(&bp->b_dep);
1176 		BUF_LOCKINIT(bp);
1177 		bq_insert(&bqempty, bp, false);
1178 	}
1179 
1180 	/*
1181 	 * maxbufspace is the absolute maximum amount of buffer space we are
1182 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1183 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1184 	 * used by most other requests.  The differential is required to
1185 	 * ensure that metadata deadlocks don't occur.
1186 	 *
1187 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1188 	 * this may result in KVM fragmentation which is not handled optimally
1189 	 * by the system. XXX This is less true with vmem.  We could use
1190 	 * PAGE_SIZE.
1191 	 */
1192 	maxbufspace = (long)nbuf * BKVASIZE;
1193 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1194 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1195 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1196 
1197 	/*
1198 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1199 	 * arbitrarily and may need further tuning. It corresponds to
1200 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1201 	 * which fits with many RAID controllers' tagged queuing limits.
1202 	 * The lower 1 MiB limit is the historical upper limit for
1203 	 * hirunningspace.
1204 	 */
1205 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1206 	    16 * 1024 * 1024), 1024 * 1024);
1207 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1208 
1209 	/*
1210 	 * Limit the amount of malloc memory since it is wired permanently into
1211 	 * the kernel space.  Even though this is accounted for in the buffer
1212 	 * allocation, we don't want the malloced region to grow uncontrolled.
1213 	 * The malloc scheme improves memory utilization significantly on
1214 	 * average (small) directories.
1215 	 */
1216 	maxbufmallocspace = hibufspace / 20;
1217 
1218 	/*
1219 	 * Reduce the chance of a deadlock occurring by limiting the number
1220 	 * of delayed-write dirty buffers we allow to stack up.
1221 	 */
1222 	hidirtybuffers = nbuf / 4 + 20;
1223 	dirtybufthresh = hidirtybuffers * 9 / 10;
1224 	/*
1225 	 * To support extreme low-memory systems, make sure hidirtybuffers
1226 	 * cannot eat up all available buffer space.  This occurs when our
1227 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1228 	 * buffer space assuming BKVASIZE'd buffers.
1229 	 */
1230 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1231 		hidirtybuffers >>= 1;
1232 	}
1233 	lodirtybuffers = hidirtybuffers / 2;
1234 
1235 	/*
1236 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1237 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1238 	 * are counted as free but will be unavailable to threads executing
1239 	 * on other cpus.
1240 	 *
1241 	 * hifreebuffers is the free target for the bufspace daemon.  This
1242 	 * should be set appropriately to limit work per-iteration.
1243 	 */
1244 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1245 	hifreebuffers = (3 * lofreebuffers) / 2;
1246 	numfreebuffers = nbuf;
1247 
1248 	/* Setup the kva and free list allocators. */
1249 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1250 	buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1251 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1252 
1253 	/*
1254 	 * Size the clean queue according to the amount of buffer space.
1255 	 * One queue per-256mb up to the max.  More queues gives better
1256 	 * concurrency but less accurate LRU.
1257 	 */
1258 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1259 	for (i = 0 ; i < buf_domains; i++) {
1260 		struct bufdomain *bd;
1261 
1262 		bd = &bdomain[i];
1263 		bd_init(bd);
1264 		bd->bd_freebuffers = nbuf / buf_domains;
1265 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1266 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1267 		bd->bd_bufspace = 0;
1268 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1269 		bd->bd_hibufspace = hibufspace / buf_domains;
1270 		bd->bd_lobufspace = lobufspace / buf_domains;
1271 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1272 		bd->bd_numdirtybuffers = 0;
1273 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1274 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1275 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1276 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1277 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1278 	}
1279 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1280 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1281 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1282 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1283 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1284 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1285 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1286 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1287 }
1288 
1289 #ifdef INVARIANTS
1290 static inline void
1291 vfs_buf_check_mapped(struct buf *bp)
1292 {
1293 
1294 	KASSERT(bp->b_kvabase != unmapped_buf,
1295 	    ("mapped buf: b_kvabase was not updated %p", bp));
1296 	KASSERT(bp->b_data != unmapped_buf,
1297 	    ("mapped buf: b_data was not updated %p", bp));
1298 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1299 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1300 }
1301 
1302 static inline void
1303 vfs_buf_check_unmapped(struct buf *bp)
1304 {
1305 
1306 	KASSERT(bp->b_data == unmapped_buf,
1307 	    ("unmapped buf: corrupted b_data %p", bp));
1308 }
1309 
1310 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1311 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1312 #else
1313 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1314 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1315 #endif
1316 
1317 static int
1318 isbufbusy(struct buf *bp)
1319 {
1320 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1321 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1322 		return (1);
1323 	return (0);
1324 }
1325 
1326 /*
1327  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1328  */
1329 void
1330 bufshutdown(int show_busybufs)
1331 {
1332 	static int first_buf_printf = 1;
1333 	struct buf *bp;
1334 	int iter, nbusy, pbusy;
1335 #ifndef PREEMPTION
1336 	int subiter;
1337 #endif
1338 
1339 	/*
1340 	 * Sync filesystems for shutdown
1341 	 */
1342 	wdog_kern_pat(WD_LASTVAL);
1343 	kern_sync(curthread);
1344 
1345 	/*
1346 	 * With soft updates, some buffers that are
1347 	 * written will be remarked as dirty until other
1348 	 * buffers are written.
1349 	 */
1350 	for (iter = pbusy = 0; iter < 20; iter++) {
1351 		nbusy = 0;
1352 		for (bp = &buf[nbuf]; --bp >= buf; )
1353 			if (isbufbusy(bp))
1354 				nbusy++;
1355 		if (nbusy == 0) {
1356 			if (first_buf_printf)
1357 				printf("All buffers synced.");
1358 			break;
1359 		}
1360 		if (first_buf_printf) {
1361 			printf("Syncing disks, buffers remaining... ");
1362 			first_buf_printf = 0;
1363 		}
1364 		printf("%d ", nbusy);
1365 		if (nbusy < pbusy)
1366 			iter = 0;
1367 		pbusy = nbusy;
1368 
1369 		wdog_kern_pat(WD_LASTVAL);
1370 		kern_sync(curthread);
1371 
1372 #ifdef PREEMPTION
1373 		/*
1374 		 * Spin for a while to allow interrupt threads to run.
1375 		 */
1376 		DELAY(50000 * iter);
1377 #else
1378 		/*
1379 		 * Context switch several times to allow interrupt
1380 		 * threads to run.
1381 		 */
1382 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1383 			thread_lock(curthread);
1384 			mi_switch(SW_VOL);
1385 			DELAY(1000);
1386 		}
1387 #endif
1388 	}
1389 	printf("\n");
1390 	/*
1391 	 * Count only busy local buffers to prevent forcing
1392 	 * a fsck if we're just a client of a wedged NFS server
1393 	 */
1394 	nbusy = 0;
1395 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1396 		if (isbufbusy(bp)) {
1397 #if 0
1398 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1399 			if (bp->b_dev == NULL) {
1400 				TAILQ_REMOVE(&mountlist,
1401 				    bp->b_vp->v_mount, mnt_list);
1402 				continue;
1403 			}
1404 #endif
1405 			nbusy++;
1406 			if (show_busybufs > 0) {
1407 				printf(
1408 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1409 				    nbusy, bp, bp->b_vp, bp->b_flags,
1410 				    (intmax_t)bp->b_blkno,
1411 				    (intmax_t)bp->b_lblkno);
1412 				BUF_LOCKPRINTINFO(bp);
1413 				if (show_busybufs > 1)
1414 					vn_printf(bp->b_vp,
1415 					    "vnode content: ");
1416 			}
1417 		}
1418 	}
1419 	if (nbusy) {
1420 		/*
1421 		 * Failed to sync all blocks. Indicate this and don't
1422 		 * unmount filesystems (thus forcing an fsck on reboot).
1423 		 */
1424 		printf("Giving up on %d buffers\n", nbusy);
1425 		DELAY(5000000);	/* 5 seconds */
1426 	} else {
1427 		if (!first_buf_printf)
1428 			printf("Final sync complete\n");
1429 		/*
1430 		 * Unmount filesystems
1431 		 */
1432 		if (!KERNEL_PANICKED())
1433 			vfs_unmountall();
1434 	}
1435 	swapoff_all();
1436 	DELAY(100000);		/* wait for console output to finish */
1437 }
1438 
1439 static void
1440 bpmap_qenter(struct buf *bp)
1441 {
1442 
1443 	BUF_CHECK_MAPPED(bp);
1444 
1445 	/*
1446 	 * bp->b_data is relative to bp->b_offset, but
1447 	 * bp->b_offset may be offset into the first page.
1448 	 */
1449 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1450 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1451 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1452 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1453 }
1454 
1455 static inline struct bufdomain *
1456 bufdomain(struct buf *bp)
1457 {
1458 
1459 	return (&bdomain[bp->b_domain]);
1460 }
1461 
1462 static struct bufqueue *
1463 bufqueue(struct buf *bp)
1464 {
1465 
1466 	switch (bp->b_qindex) {
1467 	case QUEUE_NONE:
1468 		/* FALLTHROUGH */
1469 	case QUEUE_SENTINEL:
1470 		return (NULL);
1471 	case QUEUE_EMPTY:
1472 		return (&bqempty);
1473 	case QUEUE_DIRTY:
1474 		return (&bufdomain(bp)->bd_dirtyq);
1475 	case QUEUE_CLEAN:
1476 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1477 	default:
1478 		break;
1479 	}
1480 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1481 }
1482 
1483 /*
1484  * Return the locked bufqueue that bp is a member of.
1485  */
1486 static struct bufqueue *
1487 bufqueue_acquire(struct buf *bp)
1488 {
1489 	struct bufqueue *bq, *nbq;
1490 
1491 	/*
1492 	 * bp can be pushed from a per-cpu queue to the
1493 	 * cleanq while we're waiting on the lock.  Retry
1494 	 * if the queues don't match.
1495 	 */
1496 	bq = bufqueue(bp);
1497 	BQ_LOCK(bq);
1498 	for (;;) {
1499 		nbq = bufqueue(bp);
1500 		if (bq == nbq)
1501 			break;
1502 		BQ_UNLOCK(bq);
1503 		BQ_LOCK(nbq);
1504 		bq = nbq;
1505 	}
1506 	return (bq);
1507 }
1508 
1509 /*
1510  *	binsfree:
1511  *
1512  *	Insert the buffer into the appropriate free list.  Requires a
1513  *	locked buffer on entry and buffer is unlocked before return.
1514  */
1515 static void
1516 binsfree(struct buf *bp, int qindex)
1517 {
1518 	struct bufdomain *bd;
1519 	struct bufqueue *bq;
1520 
1521 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1522 	    ("binsfree: Invalid qindex %d", qindex));
1523 	BUF_ASSERT_XLOCKED(bp);
1524 
1525 	/*
1526 	 * Handle delayed bremfree() processing.
1527 	 */
1528 	if (bp->b_flags & B_REMFREE) {
1529 		if (bp->b_qindex == qindex) {
1530 			bp->b_flags |= B_REUSE;
1531 			bp->b_flags &= ~B_REMFREE;
1532 			BUF_UNLOCK(bp);
1533 			return;
1534 		}
1535 		bq = bufqueue_acquire(bp);
1536 		bq_remove(bq, bp);
1537 		BQ_UNLOCK(bq);
1538 	}
1539 	bd = bufdomain(bp);
1540 	if (qindex == QUEUE_CLEAN) {
1541 		if (bd->bd_lim != 0)
1542 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1543 		else
1544 			bq = bd->bd_cleanq;
1545 	} else
1546 		bq = &bd->bd_dirtyq;
1547 	bq_insert(bq, bp, true);
1548 }
1549 
1550 /*
1551  * buf_free:
1552  *
1553  *	Free a buffer to the buf zone once it no longer has valid contents.
1554  */
1555 static void
1556 buf_free(struct buf *bp)
1557 {
1558 
1559 	if (bp->b_flags & B_REMFREE)
1560 		bremfreef(bp);
1561 	if (bp->b_vflags & BV_BKGRDINPROG)
1562 		panic("losing buffer 1");
1563 	if (bp->b_rcred != NOCRED) {
1564 		crfree(bp->b_rcred);
1565 		bp->b_rcred = NOCRED;
1566 	}
1567 	if (bp->b_wcred != NOCRED) {
1568 		crfree(bp->b_wcred);
1569 		bp->b_wcred = NOCRED;
1570 	}
1571 	if (!LIST_EMPTY(&bp->b_dep))
1572 		buf_deallocate(bp);
1573 	bufkva_free(bp);
1574 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1575 	BUF_UNLOCK(bp);
1576 	uma_zfree(buf_zone, bp);
1577 }
1578 
1579 /*
1580  * buf_import:
1581  *
1582  *	Import bufs into the uma cache from the buf list.  The system still
1583  *	expects a static array of bufs and much of the synchronization
1584  *	around bufs assumes type stable storage.  As a result, UMA is used
1585  *	only as a per-cpu cache of bufs still maintained on a global list.
1586  */
1587 static int
1588 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1589 {
1590 	struct buf *bp;
1591 	int i;
1592 
1593 	BQ_LOCK(&bqempty);
1594 	for (i = 0; i < cnt; i++) {
1595 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1596 		if (bp == NULL)
1597 			break;
1598 		bq_remove(&bqempty, bp);
1599 		store[i] = bp;
1600 	}
1601 	BQ_UNLOCK(&bqempty);
1602 
1603 	return (i);
1604 }
1605 
1606 /*
1607  * buf_release:
1608  *
1609  *	Release bufs from the uma cache back to the buffer queues.
1610  */
1611 static void
1612 buf_release(void *arg, void **store, int cnt)
1613 {
1614 	struct bufqueue *bq;
1615 	struct buf *bp;
1616         int i;
1617 
1618 	bq = &bqempty;
1619 	BQ_LOCK(bq);
1620         for (i = 0; i < cnt; i++) {
1621 		bp = store[i];
1622 		/* Inline bq_insert() to batch locking. */
1623 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1624 		bp->b_flags &= ~(B_AGE | B_REUSE);
1625 		bq->bq_len++;
1626 		bp->b_qindex = bq->bq_index;
1627 	}
1628 	BQ_UNLOCK(bq);
1629 }
1630 
1631 /*
1632  * buf_alloc:
1633  *
1634  *	Allocate an empty buffer header.
1635  */
1636 static struct buf *
1637 buf_alloc(struct bufdomain *bd)
1638 {
1639 	struct buf *bp;
1640 	int freebufs;
1641 
1642 	/*
1643 	 * We can only run out of bufs in the buf zone if the average buf
1644 	 * is less than BKVASIZE.  In this case the actual wait/block will
1645 	 * come from buf_reycle() failing to flush one of these small bufs.
1646 	 */
1647 	bp = NULL;
1648 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1649 	if (freebufs > 0)
1650 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1651 	if (bp == NULL) {
1652 		atomic_add_int(&bd->bd_freebuffers, 1);
1653 		bufspace_daemon_wakeup(bd);
1654 		counter_u64_add(numbufallocfails, 1);
1655 		return (NULL);
1656 	}
1657 	/*
1658 	 * Wake-up the bufspace daemon on transition below threshold.
1659 	 */
1660 	if (freebufs == bd->bd_lofreebuffers)
1661 		bufspace_daemon_wakeup(bd);
1662 
1663 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1664 		panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1665 
1666 	KASSERT(bp->b_vp == NULL,
1667 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1668 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1669 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1670 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1671 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1672 	KASSERT(bp->b_npages == 0,
1673 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1674 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1675 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1676 
1677 	bp->b_domain = BD_DOMAIN(bd);
1678 	bp->b_flags = 0;
1679 	bp->b_ioflags = 0;
1680 	bp->b_xflags = 0;
1681 	bp->b_vflags = 0;
1682 	bp->b_vp = NULL;
1683 	bp->b_blkno = bp->b_lblkno = 0;
1684 	bp->b_offset = NOOFFSET;
1685 	bp->b_iodone = 0;
1686 	bp->b_error = 0;
1687 	bp->b_resid = 0;
1688 	bp->b_bcount = 0;
1689 	bp->b_npages = 0;
1690 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1691 	bp->b_bufobj = NULL;
1692 	bp->b_data = bp->b_kvabase = unmapped_buf;
1693 	bp->b_fsprivate1 = NULL;
1694 	bp->b_fsprivate2 = NULL;
1695 	bp->b_fsprivate3 = NULL;
1696 	LIST_INIT(&bp->b_dep);
1697 
1698 	return (bp);
1699 }
1700 
1701 /*
1702  *	buf_recycle:
1703  *
1704  *	Free a buffer from the given bufqueue.  kva controls whether the
1705  *	freed buf must own some kva resources.  This is used for
1706  *	defragmenting.
1707  */
1708 static int
1709 buf_recycle(struct bufdomain *bd, bool kva)
1710 {
1711 	struct bufqueue *bq;
1712 	struct buf *bp, *nbp;
1713 
1714 	if (kva)
1715 		counter_u64_add(bufdefragcnt, 1);
1716 	nbp = NULL;
1717 	bq = bd->bd_cleanq;
1718 	BQ_LOCK(bq);
1719 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1720 	    ("buf_recycle: Locks don't match"));
1721 	nbp = TAILQ_FIRST(&bq->bq_queue);
1722 
1723 	/*
1724 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1725 	 * depending.
1726 	 */
1727 	while ((bp = nbp) != NULL) {
1728 		/*
1729 		 * Calculate next bp (we can only use it if we do not
1730 		 * release the bqlock).
1731 		 */
1732 		nbp = TAILQ_NEXT(bp, b_freelist);
1733 
1734 		/*
1735 		 * If we are defragging then we need a buffer with
1736 		 * some kva to reclaim.
1737 		 */
1738 		if (kva && bp->b_kvasize == 0)
1739 			continue;
1740 
1741 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1742 			continue;
1743 
1744 		/*
1745 		 * Implement a second chance algorithm for frequently
1746 		 * accessed buffers.
1747 		 */
1748 		if ((bp->b_flags & B_REUSE) != 0) {
1749 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1750 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1751 			bp->b_flags &= ~B_REUSE;
1752 			BUF_UNLOCK(bp);
1753 			continue;
1754 		}
1755 
1756 		/*
1757 		 * Skip buffers with background writes in progress.
1758 		 */
1759 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1760 			BUF_UNLOCK(bp);
1761 			continue;
1762 		}
1763 
1764 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1765 		    ("buf_recycle: inconsistent queue %d bp %p",
1766 		    bp->b_qindex, bp));
1767 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1768 		    ("getnewbuf: queue domain %d doesn't match request %d",
1769 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1770 		/*
1771 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1772 		 * the scan from this point on.
1773 		 */
1774 		bq_remove(bq, bp);
1775 		BQ_UNLOCK(bq);
1776 
1777 		/*
1778 		 * Requeue the background write buffer with error and
1779 		 * restart the scan.
1780 		 */
1781 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1782 			bqrelse(bp);
1783 			BQ_LOCK(bq);
1784 			nbp = TAILQ_FIRST(&bq->bq_queue);
1785 			continue;
1786 		}
1787 		bp->b_flags |= B_INVAL;
1788 		brelse(bp);
1789 		return (0);
1790 	}
1791 	bd->bd_wanted = 1;
1792 	BQ_UNLOCK(bq);
1793 
1794 	return (ENOBUFS);
1795 }
1796 
1797 /*
1798  *	bremfree:
1799  *
1800  *	Mark the buffer for removal from the appropriate free list.
1801  *
1802  */
1803 void
1804 bremfree(struct buf *bp)
1805 {
1806 
1807 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1808 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1809 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1810 	KASSERT(bp->b_qindex != QUEUE_NONE,
1811 	    ("bremfree: buffer %p not on a queue.", bp));
1812 	BUF_ASSERT_XLOCKED(bp);
1813 
1814 	bp->b_flags |= B_REMFREE;
1815 }
1816 
1817 /*
1818  *	bremfreef:
1819  *
1820  *	Force an immediate removal from a free list.  Used only in nfs when
1821  *	it abuses the b_freelist pointer.
1822  */
1823 void
1824 bremfreef(struct buf *bp)
1825 {
1826 	struct bufqueue *bq;
1827 
1828 	bq = bufqueue_acquire(bp);
1829 	bq_remove(bq, bp);
1830 	BQ_UNLOCK(bq);
1831 }
1832 
1833 static void
1834 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1835 {
1836 
1837 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1838 	TAILQ_INIT(&bq->bq_queue);
1839 	bq->bq_len = 0;
1840 	bq->bq_index = qindex;
1841 	bq->bq_subqueue = subqueue;
1842 }
1843 
1844 static void
1845 bd_init(struct bufdomain *bd)
1846 {
1847 	int i;
1848 
1849 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1850 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1851 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1852 	for (i = 0; i <= mp_maxid; i++)
1853 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1854 		    "bufq clean subqueue lock");
1855 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1856 }
1857 
1858 /*
1859  *	bq_remove:
1860  *
1861  *	Removes a buffer from the free list, must be called with the
1862  *	correct qlock held.
1863  */
1864 static void
1865 bq_remove(struct bufqueue *bq, struct buf *bp)
1866 {
1867 
1868 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1869 	    bp, bp->b_vp, bp->b_flags);
1870 	KASSERT(bp->b_qindex != QUEUE_NONE,
1871 	    ("bq_remove: buffer %p not on a queue.", bp));
1872 	KASSERT(bufqueue(bp) == bq,
1873 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1874 
1875 	BQ_ASSERT_LOCKED(bq);
1876 	if (bp->b_qindex != QUEUE_EMPTY) {
1877 		BUF_ASSERT_XLOCKED(bp);
1878 	}
1879 	KASSERT(bq->bq_len >= 1,
1880 	    ("queue %d underflow", bp->b_qindex));
1881 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1882 	bq->bq_len--;
1883 	bp->b_qindex = QUEUE_NONE;
1884 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1885 }
1886 
1887 static void
1888 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1889 {
1890 	struct buf *bp;
1891 
1892 	BQ_ASSERT_LOCKED(bq);
1893 	if (bq != bd->bd_cleanq) {
1894 		BD_LOCK(bd);
1895 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1896 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1897 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1898 			    b_freelist);
1899 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1900 		}
1901 		bd->bd_cleanq->bq_len += bq->bq_len;
1902 		bq->bq_len = 0;
1903 	}
1904 	if (bd->bd_wanted) {
1905 		bd->bd_wanted = 0;
1906 		wakeup(&bd->bd_wanted);
1907 	}
1908 	if (bq != bd->bd_cleanq)
1909 		BD_UNLOCK(bd);
1910 }
1911 
1912 static int
1913 bd_flushall(struct bufdomain *bd)
1914 {
1915 	struct bufqueue *bq;
1916 	int flushed;
1917 	int i;
1918 
1919 	if (bd->bd_lim == 0)
1920 		return (0);
1921 	flushed = 0;
1922 	for (i = 0; i <= mp_maxid; i++) {
1923 		bq = &bd->bd_subq[i];
1924 		if (bq->bq_len == 0)
1925 			continue;
1926 		BQ_LOCK(bq);
1927 		bd_flush(bd, bq);
1928 		BQ_UNLOCK(bq);
1929 		flushed++;
1930 	}
1931 
1932 	return (flushed);
1933 }
1934 
1935 static void
1936 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1937 {
1938 	struct bufdomain *bd;
1939 
1940 	if (bp->b_qindex != QUEUE_NONE)
1941 		panic("bq_insert: free buffer %p onto another queue?", bp);
1942 
1943 	bd = bufdomain(bp);
1944 	if (bp->b_flags & B_AGE) {
1945 		/* Place this buf directly on the real queue. */
1946 		if (bq->bq_index == QUEUE_CLEAN)
1947 			bq = bd->bd_cleanq;
1948 		BQ_LOCK(bq);
1949 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1950 	} else {
1951 		BQ_LOCK(bq);
1952 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1953 	}
1954 	bp->b_flags &= ~(B_AGE | B_REUSE);
1955 	bq->bq_len++;
1956 	bp->b_qindex = bq->bq_index;
1957 	bp->b_subqueue = bq->bq_subqueue;
1958 
1959 	/*
1960 	 * Unlock before we notify so that we don't wakeup a waiter that
1961 	 * fails a trylock on the buf and sleeps again.
1962 	 */
1963 	if (unlock)
1964 		BUF_UNLOCK(bp);
1965 
1966 	if (bp->b_qindex == QUEUE_CLEAN) {
1967 		/*
1968 		 * Flush the per-cpu queue and notify any waiters.
1969 		 */
1970 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1971 		    bq->bq_len >= bd->bd_lim))
1972 			bd_flush(bd, bq);
1973 	}
1974 	BQ_UNLOCK(bq);
1975 }
1976 
1977 /*
1978  *	bufkva_free:
1979  *
1980  *	Free the kva allocation for a buffer.
1981  *
1982  */
1983 static void
1984 bufkva_free(struct buf *bp)
1985 {
1986 
1987 #ifdef INVARIANTS
1988 	if (bp->b_kvasize == 0) {
1989 		KASSERT(bp->b_kvabase == unmapped_buf &&
1990 		    bp->b_data == unmapped_buf,
1991 		    ("Leaked KVA space on %p", bp));
1992 	} else if (buf_mapped(bp))
1993 		BUF_CHECK_MAPPED(bp);
1994 	else
1995 		BUF_CHECK_UNMAPPED(bp);
1996 #endif
1997 	if (bp->b_kvasize == 0)
1998 		return;
1999 
2000 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2001 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2002 	counter_u64_add(buffreekvacnt, 1);
2003 	bp->b_data = bp->b_kvabase = unmapped_buf;
2004 	bp->b_kvasize = 0;
2005 }
2006 
2007 /*
2008  *	bufkva_alloc:
2009  *
2010  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2011  */
2012 static int
2013 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2014 {
2015 	vm_offset_t addr;
2016 	int error;
2017 
2018 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2019 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2020 
2021 	bufkva_free(bp);
2022 
2023 	addr = 0;
2024 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2025 	if (error != 0) {
2026 		/*
2027 		 * Buffer map is too fragmented.  Request the caller
2028 		 * to defragment the map.
2029 		 */
2030 		return (error);
2031 	}
2032 	bp->b_kvabase = (caddr_t)addr;
2033 	bp->b_kvasize = maxsize;
2034 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2035 	if ((gbflags & GB_UNMAPPED) != 0) {
2036 		bp->b_data = unmapped_buf;
2037 		BUF_CHECK_UNMAPPED(bp);
2038 	} else {
2039 		bp->b_data = bp->b_kvabase;
2040 		BUF_CHECK_MAPPED(bp);
2041 	}
2042 	return (0);
2043 }
2044 
2045 /*
2046  *	bufkva_reclaim:
2047  *
2048  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2049  *	callback that fires to avoid returning failure.
2050  */
2051 static void
2052 bufkva_reclaim(vmem_t *vmem, int flags)
2053 {
2054 	bool done;
2055 	int q;
2056 	int i;
2057 
2058 	done = false;
2059 	for (i = 0; i < 5; i++) {
2060 		for (q = 0; q < buf_domains; q++)
2061 			if (buf_recycle(&bdomain[q], true) != 0)
2062 				done = true;
2063 		if (done)
2064 			break;
2065 	}
2066 	return;
2067 }
2068 
2069 /*
2070  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2071  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2072  * the buffer is valid and we do not have to do anything.
2073  */
2074 static void
2075 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2076     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2077 {
2078 	struct buf *rabp;
2079 	struct thread *td;
2080 	int i;
2081 
2082 	td = curthread;
2083 
2084 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2085 		if (inmem(vp, *rablkno))
2086 			continue;
2087 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2088 		if ((rabp->b_flags & B_CACHE) != 0) {
2089 			brelse(rabp);
2090 			continue;
2091 		}
2092 #ifdef RACCT
2093 		if (racct_enable) {
2094 			PROC_LOCK(curproc);
2095 			racct_add_buf(curproc, rabp, 0);
2096 			PROC_UNLOCK(curproc);
2097 		}
2098 #endif /* RACCT */
2099 		td->td_ru.ru_inblock++;
2100 		rabp->b_flags |= B_ASYNC;
2101 		rabp->b_flags &= ~B_INVAL;
2102 		if ((flags & GB_CKHASH) != 0) {
2103 			rabp->b_flags |= B_CKHASH;
2104 			rabp->b_ckhashcalc = ckhashfunc;
2105 		}
2106 		rabp->b_ioflags &= ~BIO_ERROR;
2107 		rabp->b_iocmd = BIO_READ;
2108 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2109 			rabp->b_rcred = crhold(cred);
2110 		vfs_busy_pages(rabp, 0);
2111 		BUF_KERNPROC(rabp);
2112 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2113 		bstrategy(rabp);
2114 	}
2115 }
2116 
2117 /*
2118  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2119  *
2120  * Get a buffer with the specified data.  Look in the cache first.  We
2121  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2122  * is set, the buffer is valid and we do not have to do anything, see
2123  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2124  *
2125  * Always return a NULL buffer pointer (in bpp) when returning an error.
2126  *
2127  * The blkno parameter is the logical block being requested. Normally
2128  * the mapping of logical block number to disk block address is done
2129  * by calling VOP_BMAP(). However, if the mapping is already known, the
2130  * disk block address can be passed using the dblkno parameter. If the
2131  * disk block address is not known, then the same value should be passed
2132  * for blkno and dblkno.
2133  */
2134 int
2135 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2136     daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2137     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2138 {
2139 	struct buf *bp;
2140 	struct thread *td;
2141 	int error, readwait, rv;
2142 
2143 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2144 	td = curthread;
2145 	/*
2146 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2147 	 * are specified.
2148 	 */
2149 	error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2150 	if (error != 0) {
2151 		*bpp = NULL;
2152 		return (error);
2153 	}
2154 	KASSERT(blkno == bp->b_lblkno,
2155 	    ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2156 	    (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2157 	flags &= ~GB_NOSPARSE;
2158 	*bpp = bp;
2159 
2160 	/*
2161 	 * If not found in cache, do some I/O
2162 	 */
2163 	readwait = 0;
2164 	if ((bp->b_flags & B_CACHE) == 0) {
2165 #ifdef RACCT
2166 		if (racct_enable) {
2167 			PROC_LOCK(td->td_proc);
2168 			racct_add_buf(td->td_proc, bp, 0);
2169 			PROC_UNLOCK(td->td_proc);
2170 		}
2171 #endif /* RACCT */
2172 		td->td_ru.ru_inblock++;
2173 		bp->b_iocmd = BIO_READ;
2174 		bp->b_flags &= ~B_INVAL;
2175 		if ((flags & GB_CKHASH) != 0) {
2176 			bp->b_flags |= B_CKHASH;
2177 			bp->b_ckhashcalc = ckhashfunc;
2178 		}
2179 		bp->b_ioflags &= ~BIO_ERROR;
2180 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2181 			bp->b_rcred = crhold(cred);
2182 		vfs_busy_pages(bp, 0);
2183 		bp->b_iooffset = dbtob(bp->b_blkno);
2184 		bstrategy(bp);
2185 		++readwait;
2186 	}
2187 
2188 	/*
2189 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2190 	 */
2191 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2192 
2193 	rv = 0;
2194 	if (readwait) {
2195 		rv = bufwait(bp);
2196 		if (rv != 0) {
2197 			brelse(bp);
2198 			*bpp = NULL;
2199 		}
2200 	}
2201 	return (rv);
2202 }
2203 
2204 /*
2205  * Write, release buffer on completion.  (Done by iodone
2206  * if async).  Do not bother writing anything if the buffer
2207  * is invalid.
2208  *
2209  * Note that we set B_CACHE here, indicating that buffer is
2210  * fully valid and thus cacheable.  This is true even of NFS
2211  * now so we set it generally.  This could be set either here
2212  * or in biodone() since the I/O is synchronous.  We put it
2213  * here.
2214  */
2215 int
2216 bufwrite(struct buf *bp)
2217 {
2218 	int oldflags;
2219 	struct vnode *vp;
2220 	long space;
2221 	int vp_md;
2222 
2223 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2224 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2225 		bp->b_flags |= B_INVAL | B_RELBUF;
2226 		bp->b_flags &= ~B_CACHE;
2227 		brelse(bp);
2228 		return (ENXIO);
2229 	}
2230 	if (bp->b_flags & B_INVAL) {
2231 		brelse(bp);
2232 		return (0);
2233 	}
2234 
2235 	if (bp->b_flags & B_BARRIER)
2236 		atomic_add_long(&barrierwrites, 1);
2237 
2238 	oldflags = bp->b_flags;
2239 
2240 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2241 	    ("FFS background buffer should not get here %p", bp));
2242 
2243 	vp = bp->b_vp;
2244 	if (vp)
2245 		vp_md = vp->v_vflag & VV_MD;
2246 	else
2247 		vp_md = 0;
2248 
2249 	/*
2250 	 * Mark the buffer clean.  Increment the bufobj write count
2251 	 * before bundirty() call, to prevent other thread from seeing
2252 	 * empty dirty list and zero counter for writes in progress,
2253 	 * falsely indicating that the bufobj is clean.
2254 	 */
2255 	bufobj_wref(bp->b_bufobj);
2256 	bundirty(bp);
2257 
2258 	bp->b_flags &= ~B_DONE;
2259 	bp->b_ioflags &= ~BIO_ERROR;
2260 	bp->b_flags |= B_CACHE;
2261 	bp->b_iocmd = BIO_WRITE;
2262 
2263 	vfs_busy_pages(bp, 1);
2264 
2265 	/*
2266 	 * Normal bwrites pipeline writes
2267 	 */
2268 	bp->b_runningbufspace = bp->b_bufsize;
2269 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2270 
2271 #ifdef RACCT
2272 	if (racct_enable) {
2273 		PROC_LOCK(curproc);
2274 		racct_add_buf(curproc, bp, 1);
2275 		PROC_UNLOCK(curproc);
2276 	}
2277 #endif /* RACCT */
2278 	curthread->td_ru.ru_oublock++;
2279 	if (oldflags & B_ASYNC)
2280 		BUF_KERNPROC(bp);
2281 	bp->b_iooffset = dbtob(bp->b_blkno);
2282 	buf_track(bp, __func__);
2283 	bstrategy(bp);
2284 
2285 	if ((oldflags & B_ASYNC) == 0) {
2286 		int rtval = bufwait(bp);
2287 		brelse(bp);
2288 		return (rtval);
2289 	} else if (space > hirunningspace) {
2290 		/*
2291 		 * don't allow the async write to saturate the I/O
2292 		 * system.  We will not deadlock here because
2293 		 * we are blocking waiting for I/O that is already in-progress
2294 		 * to complete. We do not block here if it is the update
2295 		 * or syncer daemon trying to clean up as that can lead
2296 		 * to deadlock.
2297 		 */
2298 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2299 			waitrunningbufspace();
2300 	}
2301 
2302 	return (0);
2303 }
2304 
2305 void
2306 bufbdflush(struct bufobj *bo, struct buf *bp)
2307 {
2308 	struct buf *nbp;
2309 
2310 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2311 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2312 		altbufferflushes++;
2313 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2314 		BO_LOCK(bo);
2315 		/*
2316 		 * Try to find a buffer to flush.
2317 		 */
2318 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2319 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2320 			    BUF_LOCK(nbp,
2321 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2322 				continue;
2323 			if (bp == nbp)
2324 				panic("bdwrite: found ourselves");
2325 			BO_UNLOCK(bo);
2326 			/* Don't countdeps with the bo lock held. */
2327 			if (buf_countdeps(nbp, 0)) {
2328 				BO_LOCK(bo);
2329 				BUF_UNLOCK(nbp);
2330 				continue;
2331 			}
2332 			if (nbp->b_flags & B_CLUSTEROK) {
2333 				vfs_bio_awrite(nbp);
2334 			} else {
2335 				bremfree(nbp);
2336 				bawrite(nbp);
2337 			}
2338 			dirtybufferflushes++;
2339 			break;
2340 		}
2341 		if (nbp == NULL)
2342 			BO_UNLOCK(bo);
2343 	}
2344 }
2345 
2346 /*
2347  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2348  * anything if the buffer is marked invalid.
2349  *
2350  * Note that since the buffer must be completely valid, we can safely
2351  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2352  * biodone() in order to prevent getblk from writing the buffer
2353  * out synchronously.
2354  */
2355 void
2356 bdwrite(struct buf *bp)
2357 {
2358 	struct thread *td = curthread;
2359 	struct vnode *vp;
2360 	struct bufobj *bo;
2361 
2362 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2363 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2364 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2365 	    ("Barrier request in delayed write %p", bp));
2366 
2367 	if (bp->b_flags & B_INVAL) {
2368 		brelse(bp);
2369 		return;
2370 	}
2371 
2372 	/*
2373 	 * If we have too many dirty buffers, don't create any more.
2374 	 * If we are wildly over our limit, then force a complete
2375 	 * cleanup. Otherwise, just keep the situation from getting
2376 	 * out of control. Note that we have to avoid a recursive
2377 	 * disaster and not try to clean up after our own cleanup!
2378 	 */
2379 	vp = bp->b_vp;
2380 	bo = bp->b_bufobj;
2381 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2382 		td->td_pflags |= TDP_INBDFLUSH;
2383 		BO_BDFLUSH(bo, bp);
2384 		td->td_pflags &= ~TDP_INBDFLUSH;
2385 	} else
2386 		recursiveflushes++;
2387 
2388 	bdirty(bp);
2389 	/*
2390 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2391 	 * true even of NFS now.
2392 	 */
2393 	bp->b_flags |= B_CACHE;
2394 
2395 	/*
2396 	 * This bmap keeps the system from needing to do the bmap later,
2397 	 * perhaps when the system is attempting to do a sync.  Since it
2398 	 * is likely that the indirect block -- or whatever other datastructure
2399 	 * that the filesystem needs is still in memory now, it is a good
2400 	 * thing to do this.  Note also, that if the pageout daemon is
2401 	 * requesting a sync -- there might not be enough memory to do
2402 	 * the bmap then...  So, this is important to do.
2403 	 */
2404 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2405 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2406 	}
2407 
2408 	buf_track(bp, __func__);
2409 
2410 	/*
2411 	 * Set the *dirty* buffer range based upon the VM system dirty
2412 	 * pages.
2413 	 *
2414 	 * Mark the buffer pages as clean.  We need to do this here to
2415 	 * satisfy the vnode_pager and the pageout daemon, so that it
2416 	 * thinks that the pages have been "cleaned".  Note that since
2417 	 * the pages are in a delayed write buffer -- the VFS layer
2418 	 * "will" see that the pages get written out on the next sync,
2419 	 * or perhaps the cluster will be completed.
2420 	 */
2421 	vfs_clean_pages_dirty_buf(bp);
2422 	bqrelse(bp);
2423 
2424 	/*
2425 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2426 	 * due to the softdep code.
2427 	 */
2428 }
2429 
2430 /*
2431  *	bdirty:
2432  *
2433  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2434  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2435  *	itself to properly update it in the dirty/clean lists.  We mark it
2436  *	B_DONE to ensure that any asynchronization of the buffer properly
2437  *	clears B_DONE ( else a panic will occur later ).
2438  *
2439  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2440  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2441  *	should only be called if the buffer is known-good.
2442  *
2443  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2444  *	count.
2445  *
2446  *	The buffer must be on QUEUE_NONE.
2447  */
2448 void
2449 bdirty(struct buf *bp)
2450 {
2451 
2452 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2453 	    bp, bp->b_vp, bp->b_flags);
2454 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2455 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2456 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2457 	bp->b_flags &= ~(B_RELBUF);
2458 	bp->b_iocmd = BIO_WRITE;
2459 
2460 	if ((bp->b_flags & B_DELWRI) == 0) {
2461 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2462 		reassignbuf(bp);
2463 		bdirtyadd(bp);
2464 	}
2465 }
2466 
2467 /*
2468  *	bundirty:
2469  *
2470  *	Clear B_DELWRI for buffer.
2471  *
2472  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2473  *	count.
2474  *
2475  *	The buffer must be on QUEUE_NONE.
2476  */
2477 
2478 void
2479 bundirty(struct buf *bp)
2480 {
2481 
2482 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2483 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2484 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2485 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2486 
2487 	if (bp->b_flags & B_DELWRI) {
2488 		bp->b_flags &= ~B_DELWRI;
2489 		reassignbuf(bp);
2490 		bdirtysub(bp);
2491 	}
2492 	/*
2493 	 * Since it is now being written, we can clear its deferred write flag.
2494 	 */
2495 	bp->b_flags &= ~B_DEFERRED;
2496 }
2497 
2498 /*
2499  *	bawrite:
2500  *
2501  *	Asynchronous write.  Start output on a buffer, but do not wait for
2502  *	it to complete.  The buffer is released when the output completes.
2503  *
2504  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2505  *	B_INVAL buffers.  Not us.
2506  */
2507 void
2508 bawrite(struct buf *bp)
2509 {
2510 
2511 	bp->b_flags |= B_ASYNC;
2512 	(void) bwrite(bp);
2513 }
2514 
2515 /*
2516  *	babarrierwrite:
2517  *
2518  *	Asynchronous barrier write.  Start output on a buffer, but do not
2519  *	wait for it to complete.  Place a write barrier after this write so
2520  *	that this buffer and all buffers written before it are committed to
2521  *	the disk before any buffers written after this write are committed
2522  *	to the disk.  The buffer is released when the output completes.
2523  */
2524 void
2525 babarrierwrite(struct buf *bp)
2526 {
2527 
2528 	bp->b_flags |= B_ASYNC | B_BARRIER;
2529 	(void) bwrite(bp);
2530 }
2531 
2532 /*
2533  *	bbarrierwrite:
2534  *
2535  *	Synchronous barrier write.  Start output on a buffer and wait for
2536  *	it to complete.  Place a write barrier after this write so that
2537  *	this buffer and all buffers written before it are committed to
2538  *	the disk before any buffers written after this write are committed
2539  *	to the disk.  The buffer is released when the output completes.
2540  */
2541 int
2542 bbarrierwrite(struct buf *bp)
2543 {
2544 
2545 	bp->b_flags |= B_BARRIER;
2546 	return (bwrite(bp));
2547 }
2548 
2549 /*
2550  *	bwillwrite:
2551  *
2552  *	Called prior to the locking of any vnodes when we are expecting to
2553  *	write.  We do not want to starve the buffer cache with too many
2554  *	dirty buffers so we block here.  By blocking prior to the locking
2555  *	of any vnodes we attempt to avoid the situation where a locked vnode
2556  *	prevents the various system daemons from flushing related buffers.
2557  */
2558 void
2559 bwillwrite(void)
2560 {
2561 
2562 	if (buf_dirty_count_severe()) {
2563 		mtx_lock(&bdirtylock);
2564 		while (buf_dirty_count_severe()) {
2565 			bdirtywait = 1;
2566 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2567 			    "flswai", 0);
2568 		}
2569 		mtx_unlock(&bdirtylock);
2570 	}
2571 }
2572 
2573 /*
2574  * Return true if we have too many dirty buffers.
2575  */
2576 int
2577 buf_dirty_count_severe(void)
2578 {
2579 
2580 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2581 }
2582 
2583 /*
2584  *	brelse:
2585  *
2586  *	Release a busy buffer and, if requested, free its resources.  The
2587  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2588  *	to be accessed later as a cache entity or reused for other purposes.
2589  */
2590 void
2591 brelse(struct buf *bp)
2592 {
2593 	struct mount *v_mnt;
2594 	int qindex;
2595 
2596 	/*
2597 	 * Many functions erroneously call brelse with a NULL bp under rare
2598 	 * error conditions. Simply return when called with a NULL bp.
2599 	 */
2600 	if (bp == NULL)
2601 		return;
2602 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2603 	    bp, bp->b_vp, bp->b_flags);
2604 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2605 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2606 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2607 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2608 
2609 	if (BUF_LOCKRECURSED(bp)) {
2610 		/*
2611 		 * Do not process, in particular, do not handle the
2612 		 * B_INVAL/B_RELBUF and do not release to free list.
2613 		 */
2614 		BUF_UNLOCK(bp);
2615 		return;
2616 	}
2617 
2618 	if (bp->b_flags & B_MANAGED) {
2619 		bqrelse(bp);
2620 		return;
2621 	}
2622 
2623 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2624 		BO_LOCK(bp->b_bufobj);
2625 		bp->b_vflags &= ~BV_BKGRDERR;
2626 		BO_UNLOCK(bp->b_bufobj);
2627 		bdirty(bp);
2628 	}
2629 
2630 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2631 	    (bp->b_flags & B_INVALONERR)) {
2632 		/*
2633 		 * Forced invalidation of dirty buffer contents, to be used
2634 		 * after a failed write in the rare case that the loss of the
2635 		 * contents is acceptable.  The buffer is invalidated and
2636 		 * freed.
2637 		 */
2638 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2639 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2640 	}
2641 
2642 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2643 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2644 	    !(bp->b_flags & B_INVAL)) {
2645 		/*
2646 		 * Failed write, redirty.  All errors except ENXIO (which
2647 		 * means the device is gone) are treated as being
2648 		 * transient.
2649 		 *
2650 		 * XXX Treating EIO as transient is not correct; the
2651 		 * contract with the local storage device drivers is that
2652 		 * they will only return EIO once the I/O is no longer
2653 		 * retriable.  Network I/O also respects this through the
2654 		 * guarantees of TCP and/or the internal retries of NFS.
2655 		 * ENOMEM might be transient, but we also have no way of
2656 		 * knowing when its ok to retry/reschedule.  In general,
2657 		 * this entire case should be made obsolete through better
2658 		 * error handling/recovery and resource scheduling.
2659 		 *
2660 		 * Do this also for buffers that failed with ENXIO, but have
2661 		 * non-empty dependencies - the soft updates code might need
2662 		 * to access the buffer to untangle them.
2663 		 *
2664 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2665 		 */
2666 		bp->b_ioflags &= ~BIO_ERROR;
2667 		bdirty(bp);
2668 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2669 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2670 		/*
2671 		 * Either a failed read I/O, or we were asked to free or not
2672 		 * cache the buffer, or we failed to write to a device that's
2673 		 * no longer present.
2674 		 */
2675 		bp->b_flags |= B_INVAL;
2676 		if (!LIST_EMPTY(&bp->b_dep))
2677 			buf_deallocate(bp);
2678 		if (bp->b_flags & B_DELWRI)
2679 			bdirtysub(bp);
2680 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2681 		if ((bp->b_flags & B_VMIO) == 0) {
2682 			allocbuf(bp, 0);
2683 			if (bp->b_vp)
2684 				brelvp(bp);
2685 		}
2686 	}
2687 
2688 	/*
2689 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2690 	 * is called with B_DELWRI set, the underlying pages may wind up
2691 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2692 	 * because pages associated with a B_DELWRI bp are marked clean.
2693 	 *
2694 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2695 	 * if B_DELWRI is set.
2696 	 */
2697 	if (bp->b_flags & B_DELWRI)
2698 		bp->b_flags &= ~B_RELBUF;
2699 
2700 	/*
2701 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2702 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2703 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2704 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2705 	 *
2706 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2707 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2708 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2709 	 *
2710 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2711 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2712 	 * the commit state and we cannot afford to lose the buffer. If the
2713 	 * buffer has a background write in progress, we need to keep it
2714 	 * around to prevent it from being reconstituted and starting a second
2715 	 * background write.
2716 	 */
2717 
2718 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2719 
2720 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2721 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2722 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2723 	    vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2724 		vfs_vmio_invalidate(bp);
2725 		allocbuf(bp, 0);
2726 	}
2727 
2728 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2729 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2730 		allocbuf(bp, 0);
2731 		bp->b_flags &= ~B_NOREUSE;
2732 		if (bp->b_vp != NULL)
2733 			brelvp(bp);
2734 	}
2735 
2736 	/*
2737 	 * If the buffer has junk contents signal it and eventually
2738 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2739 	 * doesn't find it.
2740 	 */
2741 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2742 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2743 		bp->b_flags |= B_INVAL;
2744 	if (bp->b_flags & B_INVAL) {
2745 		if (bp->b_flags & B_DELWRI)
2746 			bundirty(bp);
2747 		if (bp->b_vp)
2748 			brelvp(bp);
2749 	}
2750 
2751 	buf_track(bp, __func__);
2752 
2753 	/* buffers with no memory */
2754 	if (bp->b_bufsize == 0) {
2755 		buf_free(bp);
2756 		return;
2757 	}
2758 	/* buffers with junk contents */
2759 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2760 	    (bp->b_ioflags & BIO_ERROR)) {
2761 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2762 		if (bp->b_vflags & BV_BKGRDINPROG)
2763 			panic("losing buffer 2");
2764 		qindex = QUEUE_CLEAN;
2765 		bp->b_flags |= B_AGE;
2766 	/* remaining buffers */
2767 	} else if (bp->b_flags & B_DELWRI)
2768 		qindex = QUEUE_DIRTY;
2769 	else
2770 		qindex = QUEUE_CLEAN;
2771 
2772 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2773 		panic("brelse: not dirty");
2774 
2775 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2776 	/* binsfree unlocks bp. */
2777 	binsfree(bp, qindex);
2778 }
2779 
2780 /*
2781  * Release a buffer back to the appropriate queue but do not try to free
2782  * it.  The buffer is expected to be used again soon.
2783  *
2784  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2785  * biodone() to requeue an async I/O on completion.  It is also used when
2786  * known good buffers need to be requeued but we think we may need the data
2787  * again soon.
2788  *
2789  * XXX we should be able to leave the B_RELBUF hint set on completion.
2790  */
2791 void
2792 bqrelse(struct buf *bp)
2793 {
2794 	int qindex;
2795 
2796 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2797 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2798 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2799 
2800 	qindex = QUEUE_NONE;
2801 	if (BUF_LOCKRECURSED(bp)) {
2802 		/* do not release to free list */
2803 		BUF_UNLOCK(bp);
2804 		return;
2805 	}
2806 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2807 
2808 	if (bp->b_flags & B_MANAGED) {
2809 		if (bp->b_flags & B_REMFREE)
2810 			bremfreef(bp);
2811 		goto out;
2812 	}
2813 
2814 	/* buffers with stale but valid contents */
2815 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2816 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2817 		BO_LOCK(bp->b_bufobj);
2818 		bp->b_vflags &= ~BV_BKGRDERR;
2819 		BO_UNLOCK(bp->b_bufobj);
2820 		qindex = QUEUE_DIRTY;
2821 	} else {
2822 		if ((bp->b_flags & B_DELWRI) == 0 &&
2823 		    (bp->b_xflags & BX_VNDIRTY))
2824 			panic("bqrelse: not dirty");
2825 		if ((bp->b_flags & B_NOREUSE) != 0) {
2826 			brelse(bp);
2827 			return;
2828 		}
2829 		qindex = QUEUE_CLEAN;
2830 	}
2831 	buf_track(bp, __func__);
2832 	/* binsfree unlocks bp. */
2833 	binsfree(bp, qindex);
2834 	return;
2835 
2836 out:
2837 	buf_track(bp, __func__);
2838 	/* unlock */
2839 	BUF_UNLOCK(bp);
2840 }
2841 
2842 /*
2843  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2844  * restore bogus pages.
2845  */
2846 static void
2847 vfs_vmio_iodone(struct buf *bp)
2848 {
2849 	vm_ooffset_t foff;
2850 	vm_page_t m;
2851 	vm_object_t obj;
2852 	struct vnode *vp __unused;
2853 	int i, iosize, resid;
2854 	bool bogus;
2855 
2856 	obj = bp->b_bufobj->bo_object;
2857 	KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2858 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2859 	    blockcount_read(&obj->paging_in_progress), bp->b_npages));
2860 
2861 	vp = bp->b_vp;
2862 	VNPASS(vp->v_holdcnt > 0, vp);
2863 	VNPASS(vp->v_object != NULL, vp);
2864 
2865 	foff = bp->b_offset;
2866 	KASSERT(bp->b_offset != NOOFFSET,
2867 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2868 
2869 	bogus = false;
2870 	iosize = bp->b_bcount - bp->b_resid;
2871 	for (i = 0; i < bp->b_npages; i++) {
2872 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2873 		if (resid > iosize)
2874 			resid = iosize;
2875 
2876 		/*
2877 		 * cleanup bogus pages, restoring the originals
2878 		 */
2879 		m = bp->b_pages[i];
2880 		if (m == bogus_page) {
2881 			bogus = true;
2882 			m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2883 			if (m == NULL)
2884 				panic("biodone: page disappeared!");
2885 			bp->b_pages[i] = m;
2886 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2887 			/*
2888 			 * In the write case, the valid and clean bits are
2889 			 * already changed correctly ( see bdwrite() ), so we
2890 			 * only need to do this here in the read case.
2891 			 */
2892 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2893 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2894 			    "has unexpected dirty bits", m));
2895 			vfs_page_set_valid(bp, foff, m);
2896 		}
2897 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2898 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2899 		    (intmax_t)foff, (uintmax_t)m->pindex));
2900 
2901 		vm_page_sunbusy(m);
2902 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2903 		iosize -= resid;
2904 	}
2905 	vm_object_pip_wakeupn(obj, bp->b_npages);
2906 	if (bogus && buf_mapped(bp)) {
2907 		BUF_CHECK_MAPPED(bp);
2908 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2909 		    bp->b_pages, bp->b_npages);
2910 	}
2911 }
2912 
2913 /*
2914  * Perform page invalidation when a buffer is released.  The fully invalid
2915  * pages will be reclaimed later in vfs_vmio_truncate().
2916  */
2917 static void
2918 vfs_vmio_invalidate(struct buf *bp)
2919 {
2920 	vm_object_t obj;
2921 	vm_page_t m;
2922 	int flags, i, resid, poffset, presid;
2923 
2924 	if (buf_mapped(bp)) {
2925 		BUF_CHECK_MAPPED(bp);
2926 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2927 	} else
2928 		BUF_CHECK_UNMAPPED(bp);
2929 	/*
2930 	 * Get the base offset and length of the buffer.  Note that
2931 	 * in the VMIO case if the buffer block size is not
2932 	 * page-aligned then b_data pointer may not be page-aligned.
2933 	 * But our b_pages[] array *IS* page aligned.
2934 	 *
2935 	 * block sizes less then DEV_BSIZE (usually 512) are not
2936 	 * supported due to the page granularity bits (m->valid,
2937 	 * m->dirty, etc...).
2938 	 *
2939 	 * See man buf(9) for more information
2940 	 */
2941 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2942 	obj = bp->b_bufobj->bo_object;
2943 	resid = bp->b_bufsize;
2944 	poffset = bp->b_offset & PAGE_MASK;
2945 	VM_OBJECT_WLOCK(obj);
2946 	for (i = 0; i < bp->b_npages; i++) {
2947 		m = bp->b_pages[i];
2948 		if (m == bogus_page)
2949 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2950 		bp->b_pages[i] = NULL;
2951 
2952 		presid = resid > (PAGE_SIZE - poffset) ?
2953 		    (PAGE_SIZE - poffset) : resid;
2954 		KASSERT(presid >= 0, ("brelse: extra page"));
2955 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
2956 		if (pmap_page_wired_mappings(m) == 0)
2957 			vm_page_set_invalid(m, poffset, presid);
2958 		vm_page_sunbusy(m);
2959 		vm_page_release_locked(m, flags);
2960 		resid -= presid;
2961 		poffset = 0;
2962 	}
2963 	VM_OBJECT_WUNLOCK(obj);
2964 	bp->b_npages = 0;
2965 }
2966 
2967 /*
2968  * Page-granular truncation of an existing VMIO buffer.
2969  */
2970 static void
2971 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2972 {
2973 	vm_object_t obj;
2974 	vm_page_t m;
2975 	int flags, i;
2976 
2977 	if (bp->b_npages == desiredpages)
2978 		return;
2979 
2980 	if (buf_mapped(bp)) {
2981 		BUF_CHECK_MAPPED(bp);
2982 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2983 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2984 	} else
2985 		BUF_CHECK_UNMAPPED(bp);
2986 
2987 	/*
2988 	 * The object lock is needed only if we will attempt to free pages.
2989 	 */
2990 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2991 	if ((bp->b_flags & B_DIRECT) != 0) {
2992 		flags |= VPR_TRYFREE;
2993 		obj = bp->b_bufobj->bo_object;
2994 		VM_OBJECT_WLOCK(obj);
2995 	} else {
2996 		obj = NULL;
2997 	}
2998 	for (i = desiredpages; i < bp->b_npages; i++) {
2999 		m = bp->b_pages[i];
3000 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3001 		bp->b_pages[i] = NULL;
3002 		if (obj != NULL)
3003 			vm_page_release_locked(m, flags);
3004 		else
3005 			vm_page_release(m, flags);
3006 	}
3007 	if (obj != NULL)
3008 		VM_OBJECT_WUNLOCK(obj);
3009 	bp->b_npages = desiredpages;
3010 }
3011 
3012 /*
3013  * Byte granular extension of VMIO buffers.
3014  */
3015 static void
3016 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3017 {
3018 	/*
3019 	 * We are growing the buffer, possibly in a
3020 	 * byte-granular fashion.
3021 	 */
3022 	vm_object_t obj;
3023 	vm_offset_t toff;
3024 	vm_offset_t tinc;
3025 	vm_page_t m;
3026 
3027 	/*
3028 	 * Step 1, bring in the VM pages from the object, allocating
3029 	 * them if necessary.  We must clear B_CACHE if these pages
3030 	 * are not valid for the range covered by the buffer.
3031 	 */
3032 	obj = bp->b_bufobj->bo_object;
3033 	if (bp->b_npages < desiredpages) {
3034 		/*
3035 		 * We must allocate system pages since blocking
3036 		 * here could interfere with paging I/O, no
3037 		 * matter which process we are.
3038 		 *
3039 		 * Only exclusive busy can be tested here.
3040 		 * Blocking on shared busy might lead to
3041 		 * deadlocks once allocbuf() is called after
3042 		 * pages are vfs_busy_pages().
3043 		 */
3044 		(void)vm_page_grab_pages_unlocked(obj,
3045 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3046 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3047 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3048 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3049 		bp->b_npages = desiredpages;
3050 	}
3051 
3052 	/*
3053 	 * Step 2.  We've loaded the pages into the buffer,
3054 	 * we have to figure out if we can still have B_CACHE
3055 	 * set.  Note that B_CACHE is set according to the
3056 	 * byte-granular range ( bcount and size ), not the
3057 	 * aligned range ( newbsize ).
3058 	 *
3059 	 * The VM test is against m->valid, which is DEV_BSIZE
3060 	 * aligned.  Needless to say, the validity of the data
3061 	 * needs to also be DEV_BSIZE aligned.  Note that this
3062 	 * fails with NFS if the server or some other client
3063 	 * extends the file's EOF.  If our buffer is resized,
3064 	 * B_CACHE may remain set! XXX
3065 	 */
3066 	toff = bp->b_bcount;
3067 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3068 	while ((bp->b_flags & B_CACHE) && toff < size) {
3069 		vm_pindex_t pi;
3070 
3071 		if (tinc > (size - toff))
3072 			tinc = size - toff;
3073 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3074 		m = bp->b_pages[pi];
3075 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3076 		toff += tinc;
3077 		tinc = PAGE_SIZE;
3078 	}
3079 
3080 	/*
3081 	 * Step 3, fixup the KVA pmap.
3082 	 */
3083 	if (buf_mapped(bp))
3084 		bpmap_qenter(bp);
3085 	else
3086 		BUF_CHECK_UNMAPPED(bp);
3087 }
3088 
3089 /*
3090  * Check to see if a block at a particular lbn is available for a clustered
3091  * write.
3092  */
3093 static int
3094 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3095 {
3096 	struct buf *bpa;
3097 	int match;
3098 
3099 	match = 0;
3100 
3101 	/* If the buf isn't in core skip it */
3102 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3103 		return (0);
3104 
3105 	/* If the buf is busy we don't want to wait for it */
3106 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3107 		return (0);
3108 
3109 	/* Only cluster with valid clusterable delayed write buffers */
3110 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3111 	    (B_DELWRI | B_CLUSTEROK))
3112 		goto done;
3113 
3114 	if (bpa->b_bufsize != size)
3115 		goto done;
3116 
3117 	/*
3118 	 * Check to see if it is in the expected place on disk and that the
3119 	 * block has been mapped.
3120 	 */
3121 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3122 		match = 1;
3123 done:
3124 	BUF_UNLOCK(bpa);
3125 	return (match);
3126 }
3127 
3128 /*
3129  *	vfs_bio_awrite:
3130  *
3131  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3132  *	This is much better then the old way of writing only one buffer at
3133  *	a time.  Note that we may not be presented with the buffers in the
3134  *	correct order, so we search for the cluster in both directions.
3135  */
3136 int
3137 vfs_bio_awrite(struct buf *bp)
3138 {
3139 	struct bufobj *bo;
3140 	int i;
3141 	int j;
3142 	daddr_t lblkno = bp->b_lblkno;
3143 	struct vnode *vp = bp->b_vp;
3144 	int ncl;
3145 	int nwritten;
3146 	int size;
3147 	int maxcl;
3148 	int gbflags;
3149 
3150 	bo = &vp->v_bufobj;
3151 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3152 	/*
3153 	 * right now we support clustered writing only to regular files.  If
3154 	 * we find a clusterable block we could be in the middle of a cluster
3155 	 * rather then at the beginning.
3156 	 */
3157 	if ((vp->v_type == VREG) &&
3158 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3159 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3160 
3161 		size = vp->v_mount->mnt_stat.f_iosize;
3162 		maxcl = MAXPHYS / size;
3163 
3164 		BO_RLOCK(bo);
3165 		for (i = 1; i < maxcl; i++)
3166 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3167 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3168 				break;
3169 
3170 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3171 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3172 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3173 				break;
3174 		BO_RUNLOCK(bo);
3175 		--j;
3176 		ncl = i + j;
3177 		/*
3178 		 * this is a possible cluster write
3179 		 */
3180 		if (ncl != 1) {
3181 			BUF_UNLOCK(bp);
3182 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3183 			    gbflags);
3184 			return (nwritten);
3185 		}
3186 	}
3187 	bremfree(bp);
3188 	bp->b_flags |= B_ASYNC;
3189 	/*
3190 	 * default (old) behavior, writing out only one block
3191 	 *
3192 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3193 	 */
3194 	nwritten = bp->b_bufsize;
3195 	(void) bwrite(bp);
3196 
3197 	return (nwritten);
3198 }
3199 
3200 /*
3201  *	getnewbuf_kva:
3202  *
3203  *	Allocate KVA for an empty buf header according to gbflags.
3204  */
3205 static int
3206 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3207 {
3208 
3209 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3210 		/*
3211 		 * In order to keep fragmentation sane we only allocate kva
3212 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3213 		 */
3214 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3215 
3216 		if (maxsize != bp->b_kvasize &&
3217 		    bufkva_alloc(bp, maxsize, gbflags))
3218 			return (ENOSPC);
3219 	}
3220 	return (0);
3221 }
3222 
3223 /*
3224  *	getnewbuf:
3225  *
3226  *	Find and initialize a new buffer header, freeing up existing buffers
3227  *	in the bufqueues as necessary.  The new buffer is returned locked.
3228  *
3229  *	We block if:
3230  *		We have insufficient buffer headers
3231  *		We have insufficient buffer space
3232  *		buffer_arena is too fragmented ( space reservation fails )
3233  *		If we have to flush dirty buffers ( but we try to avoid this )
3234  *
3235  *	The caller is responsible for releasing the reserved bufspace after
3236  *	allocbuf() is called.
3237  */
3238 static struct buf *
3239 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3240 {
3241 	struct bufdomain *bd;
3242 	struct buf *bp;
3243 	bool metadata, reserved;
3244 
3245 	bp = NULL;
3246 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3247 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3248 	if (!unmapped_buf_allowed)
3249 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3250 
3251 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3252 	    vp->v_type == VCHR)
3253 		metadata = true;
3254 	else
3255 		metadata = false;
3256 	if (vp == NULL)
3257 		bd = &bdomain[0];
3258 	else
3259 		bd = &bdomain[vp->v_bufobj.bo_domain];
3260 
3261 	counter_u64_add(getnewbufcalls, 1);
3262 	reserved = false;
3263 	do {
3264 		if (reserved == false &&
3265 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3266 			counter_u64_add(getnewbufrestarts, 1);
3267 			continue;
3268 		}
3269 		reserved = true;
3270 		if ((bp = buf_alloc(bd)) == NULL) {
3271 			counter_u64_add(getnewbufrestarts, 1);
3272 			continue;
3273 		}
3274 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3275 			return (bp);
3276 		break;
3277 	} while (buf_recycle(bd, false) == 0);
3278 
3279 	if (reserved)
3280 		bufspace_release(bd, maxsize);
3281 	if (bp != NULL) {
3282 		bp->b_flags |= B_INVAL;
3283 		brelse(bp);
3284 	}
3285 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3286 
3287 	return (NULL);
3288 }
3289 
3290 /*
3291  *	buf_daemon:
3292  *
3293  *	buffer flushing daemon.  Buffers are normally flushed by the
3294  *	update daemon but if it cannot keep up this process starts to
3295  *	take the load in an attempt to prevent getnewbuf() from blocking.
3296  */
3297 static struct kproc_desc buf_kp = {
3298 	"bufdaemon",
3299 	buf_daemon,
3300 	&bufdaemonproc
3301 };
3302 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3303 
3304 static int
3305 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3306 {
3307 	int flushed;
3308 
3309 	flushed = flushbufqueues(vp, bd, target, 0);
3310 	if (flushed == 0) {
3311 		/*
3312 		 * Could not find any buffers without rollback
3313 		 * dependencies, so just write the first one
3314 		 * in the hopes of eventually making progress.
3315 		 */
3316 		if (vp != NULL && target > 2)
3317 			target /= 2;
3318 		flushbufqueues(vp, bd, target, 1);
3319 	}
3320 	return (flushed);
3321 }
3322 
3323 static void
3324 buf_daemon()
3325 {
3326 	struct bufdomain *bd;
3327 	int speedupreq;
3328 	int lodirty;
3329 	int i;
3330 
3331 	/*
3332 	 * This process needs to be suspended prior to shutdown sync.
3333 	 */
3334 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3335 	    SHUTDOWN_PRI_LAST + 100);
3336 
3337 	/*
3338 	 * Start the buf clean daemons as children threads.
3339 	 */
3340 	for (i = 0 ; i < buf_domains; i++) {
3341 		int error;
3342 
3343 		error = kthread_add((void (*)(void *))bufspace_daemon,
3344 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3345 		if (error)
3346 			panic("error %d spawning bufspace daemon", error);
3347 	}
3348 
3349 	/*
3350 	 * This process is allowed to take the buffer cache to the limit
3351 	 */
3352 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3353 	mtx_lock(&bdlock);
3354 	for (;;) {
3355 		bd_request = 0;
3356 		mtx_unlock(&bdlock);
3357 
3358 		kthread_suspend_check();
3359 
3360 		/*
3361 		 * Save speedupreq for this pass and reset to capture new
3362 		 * requests.
3363 		 */
3364 		speedupreq = bd_speedupreq;
3365 		bd_speedupreq = 0;
3366 
3367 		/*
3368 		 * Flush each domain sequentially according to its level and
3369 		 * the speedup request.
3370 		 */
3371 		for (i = 0; i < buf_domains; i++) {
3372 			bd = &bdomain[i];
3373 			if (speedupreq)
3374 				lodirty = bd->bd_numdirtybuffers / 2;
3375 			else
3376 				lodirty = bd->bd_lodirtybuffers;
3377 			while (bd->bd_numdirtybuffers > lodirty) {
3378 				if (buf_flush(NULL, bd,
3379 				    bd->bd_numdirtybuffers - lodirty) == 0)
3380 					break;
3381 				kern_yield(PRI_USER);
3382 			}
3383 		}
3384 
3385 		/*
3386 		 * Only clear bd_request if we have reached our low water
3387 		 * mark.  The buf_daemon normally waits 1 second and
3388 		 * then incrementally flushes any dirty buffers that have
3389 		 * built up, within reason.
3390 		 *
3391 		 * If we were unable to hit our low water mark and couldn't
3392 		 * find any flushable buffers, we sleep for a short period
3393 		 * to avoid endless loops on unlockable buffers.
3394 		 */
3395 		mtx_lock(&bdlock);
3396 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3397 			/*
3398 			 * We reached our low water mark, reset the
3399 			 * request and sleep until we are needed again.
3400 			 * The sleep is just so the suspend code works.
3401 			 */
3402 			bd_request = 0;
3403 			/*
3404 			 * Do an extra wakeup in case dirty threshold
3405 			 * changed via sysctl and the explicit transition
3406 			 * out of shortfall was missed.
3407 			 */
3408 			bdirtywakeup();
3409 			if (runningbufspace <= lorunningspace)
3410 				runningwakeup();
3411 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3412 		} else {
3413 			/*
3414 			 * We couldn't find any flushable dirty buffers but
3415 			 * still have too many dirty buffers, we
3416 			 * have to sleep and try again.  (rare)
3417 			 */
3418 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3419 		}
3420 	}
3421 }
3422 
3423 /*
3424  *	flushbufqueues:
3425  *
3426  *	Try to flush a buffer in the dirty queue.  We must be careful to
3427  *	free up B_INVAL buffers instead of write them, which NFS is
3428  *	particularly sensitive to.
3429  */
3430 static int flushwithdeps = 0;
3431 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3432     &flushwithdeps, 0,
3433     "Number of buffers flushed with dependecies that require rollbacks");
3434 
3435 static int
3436 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3437     int flushdeps)
3438 {
3439 	struct bufqueue *bq;
3440 	struct buf *sentinel;
3441 	struct vnode *vp;
3442 	struct mount *mp;
3443 	struct buf *bp;
3444 	int hasdeps;
3445 	int flushed;
3446 	int error;
3447 	bool unlock;
3448 
3449 	flushed = 0;
3450 	bq = &bd->bd_dirtyq;
3451 	bp = NULL;
3452 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3453 	sentinel->b_qindex = QUEUE_SENTINEL;
3454 	BQ_LOCK(bq);
3455 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3456 	BQ_UNLOCK(bq);
3457 	while (flushed != target) {
3458 		maybe_yield();
3459 		BQ_LOCK(bq);
3460 		bp = TAILQ_NEXT(sentinel, b_freelist);
3461 		if (bp != NULL) {
3462 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3463 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3464 			    b_freelist);
3465 		} else {
3466 			BQ_UNLOCK(bq);
3467 			break;
3468 		}
3469 		/*
3470 		 * Skip sentinels inserted by other invocations of the
3471 		 * flushbufqueues(), taking care to not reorder them.
3472 		 *
3473 		 * Only flush the buffers that belong to the
3474 		 * vnode locked by the curthread.
3475 		 */
3476 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3477 		    bp->b_vp != lvp)) {
3478 			BQ_UNLOCK(bq);
3479 			continue;
3480 		}
3481 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3482 		BQ_UNLOCK(bq);
3483 		if (error != 0)
3484 			continue;
3485 
3486 		/*
3487 		 * BKGRDINPROG can only be set with the buf and bufobj
3488 		 * locks both held.  We tolerate a race to clear it here.
3489 		 */
3490 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3491 		    (bp->b_flags & B_DELWRI) == 0) {
3492 			BUF_UNLOCK(bp);
3493 			continue;
3494 		}
3495 		if (bp->b_flags & B_INVAL) {
3496 			bremfreef(bp);
3497 			brelse(bp);
3498 			flushed++;
3499 			continue;
3500 		}
3501 
3502 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3503 			if (flushdeps == 0) {
3504 				BUF_UNLOCK(bp);
3505 				continue;
3506 			}
3507 			hasdeps = 1;
3508 		} else
3509 			hasdeps = 0;
3510 		/*
3511 		 * We must hold the lock on a vnode before writing
3512 		 * one of its buffers. Otherwise we may confuse, or
3513 		 * in the case of a snapshot vnode, deadlock the
3514 		 * system.
3515 		 *
3516 		 * The lock order here is the reverse of the normal
3517 		 * of vnode followed by buf lock.  This is ok because
3518 		 * the NOWAIT will prevent deadlock.
3519 		 */
3520 		vp = bp->b_vp;
3521 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3522 			BUF_UNLOCK(bp);
3523 			continue;
3524 		}
3525 		if (lvp == NULL) {
3526 			unlock = true;
3527 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3528 		} else {
3529 			ASSERT_VOP_LOCKED(vp, "getbuf");
3530 			unlock = false;
3531 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3532 			    vn_lock(vp, LK_TRYUPGRADE);
3533 		}
3534 		if (error == 0) {
3535 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3536 			    bp, bp->b_vp, bp->b_flags);
3537 			if (curproc == bufdaemonproc) {
3538 				vfs_bio_awrite(bp);
3539 			} else {
3540 				bremfree(bp);
3541 				bwrite(bp);
3542 				counter_u64_add(notbufdflushes, 1);
3543 			}
3544 			vn_finished_write(mp);
3545 			if (unlock)
3546 				VOP_UNLOCK(vp);
3547 			flushwithdeps += hasdeps;
3548 			flushed++;
3549 
3550 			/*
3551 			 * Sleeping on runningbufspace while holding
3552 			 * vnode lock leads to deadlock.
3553 			 */
3554 			if (curproc == bufdaemonproc &&
3555 			    runningbufspace > hirunningspace)
3556 				waitrunningbufspace();
3557 			continue;
3558 		}
3559 		vn_finished_write(mp);
3560 		BUF_UNLOCK(bp);
3561 	}
3562 	BQ_LOCK(bq);
3563 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3564 	BQ_UNLOCK(bq);
3565 	free(sentinel, M_TEMP);
3566 	return (flushed);
3567 }
3568 
3569 /*
3570  * Check to see if a block is currently memory resident.
3571  */
3572 struct buf *
3573 incore(struct bufobj *bo, daddr_t blkno)
3574 {
3575 	struct buf *bp;
3576 
3577 	BO_RLOCK(bo);
3578 	bp = gbincore(bo, blkno);
3579 	BO_RUNLOCK(bo);
3580 	return (bp);
3581 }
3582 
3583 /*
3584  * Returns true if no I/O is needed to access the
3585  * associated VM object.  This is like incore except
3586  * it also hunts around in the VM system for the data.
3587  */
3588 
3589 static int
3590 inmem(struct vnode * vp, daddr_t blkno)
3591 {
3592 	vm_object_t obj;
3593 	vm_offset_t toff, tinc, size;
3594 	vm_page_t m;
3595 	vm_ooffset_t off;
3596 
3597 	ASSERT_VOP_LOCKED(vp, "inmem");
3598 
3599 	if (incore(&vp->v_bufobj, blkno))
3600 		return 1;
3601 	if (vp->v_mount == NULL)
3602 		return 0;
3603 	obj = vp->v_object;
3604 	if (obj == NULL)
3605 		return (0);
3606 
3607 	size = PAGE_SIZE;
3608 	if (size > vp->v_mount->mnt_stat.f_iosize)
3609 		size = vp->v_mount->mnt_stat.f_iosize;
3610 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3611 
3612 	VM_OBJECT_RLOCK(obj);
3613 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3614 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3615 		if (!m)
3616 			goto notinmem;
3617 		tinc = size;
3618 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3619 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3620 		if (vm_page_is_valid(m,
3621 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3622 			goto notinmem;
3623 	}
3624 	VM_OBJECT_RUNLOCK(obj);
3625 	return 1;
3626 
3627 notinmem:
3628 	VM_OBJECT_RUNLOCK(obj);
3629 	return (0);
3630 }
3631 
3632 /*
3633  * Set the dirty range for a buffer based on the status of the dirty
3634  * bits in the pages comprising the buffer.  The range is limited
3635  * to the size of the buffer.
3636  *
3637  * Tell the VM system that the pages associated with this buffer
3638  * are clean.  This is used for delayed writes where the data is
3639  * going to go to disk eventually without additional VM intevention.
3640  *
3641  * Note that while we only really need to clean through to b_bcount, we
3642  * just go ahead and clean through to b_bufsize.
3643  */
3644 static void
3645 vfs_clean_pages_dirty_buf(struct buf *bp)
3646 {
3647 	vm_ooffset_t foff, noff, eoff;
3648 	vm_page_t m;
3649 	int i;
3650 
3651 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3652 		return;
3653 
3654 	foff = bp->b_offset;
3655 	KASSERT(bp->b_offset != NOOFFSET,
3656 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3657 
3658 	vfs_busy_pages_acquire(bp);
3659 	vfs_setdirty_range(bp);
3660 	for (i = 0; i < bp->b_npages; i++) {
3661 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3662 		eoff = noff;
3663 		if (eoff > bp->b_offset + bp->b_bufsize)
3664 			eoff = bp->b_offset + bp->b_bufsize;
3665 		m = bp->b_pages[i];
3666 		vfs_page_set_validclean(bp, foff, m);
3667 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3668 		foff = noff;
3669 	}
3670 	vfs_busy_pages_release(bp);
3671 }
3672 
3673 static void
3674 vfs_setdirty_range(struct buf *bp)
3675 {
3676 	vm_offset_t boffset;
3677 	vm_offset_t eoffset;
3678 	int i;
3679 
3680 	/*
3681 	 * test the pages to see if they have been modified directly
3682 	 * by users through the VM system.
3683 	 */
3684 	for (i = 0; i < bp->b_npages; i++)
3685 		vm_page_test_dirty(bp->b_pages[i]);
3686 
3687 	/*
3688 	 * Calculate the encompassing dirty range, boffset and eoffset,
3689 	 * (eoffset - boffset) bytes.
3690 	 */
3691 
3692 	for (i = 0; i < bp->b_npages; i++) {
3693 		if (bp->b_pages[i]->dirty)
3694 			break;
3695 	}
3696 	boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3697 
3698 	for (i = bp->b_npages - 1; i >= 0; --i) {
3699 		if (bp->b_pages[i]->dirty) {
3700 			break;
3701 		}
3702 	}
3703 	eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3704 
3705 	/*
3706 	 * Fit it to the buffer.
3707 	 */
3708 
3709 	if (eoffset > bp->b_bcount)
3710 		eoffset = bp->b_bcount;
3711 
3712 	/*
3713 	 * If we have a good dirty range, merge with the existing
3714 	 * dirty range.
3715 	 */
3716 
3717 	if (boffset < eoffset) {
3718 		if (bp->b_dirtyoff > boffset)
3719 			bp->b_dirtyoff = boffset;
3720 		if (bp->b_dirtyend < eoffset)
3721 			bp->b_dirtyend = eoffset;
3722 	}
3723 }
3724 
3725 /*
3726  * Allocate the KVA mapping for an existing buffer.
3727  * If an unmapped buffer is provided but a mapped buffer is requested, take
3728  * also care to properly setup mappings between pages and KVA.
3729  */
3730 static void
3731 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3732 {
3733 	int bsize, maxsize, need_mapping, need_kva;
3734 	off_t offset;
3735 
3736 	need_mapping = bp->b_data == unmapped_buf &&
3737 	    (gbflags & GB_UNMAPPED) == 0;
3738 	need_kva = bp->b_kvabase == unmapped_buf &&
3739 	    bp->b_data == unmapped_buf &&
3740 	    (gbflags & GB_KVAALLOC) != 0;
3741 	if (!need_mapping && !need_kva)
3742 		return;
3743 
3744 	BUF_CHECK_UNMAPPED(bp);
3745 
3746 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3747 		/*
3748 		 * Buffer is not mapped, but the KVA was already
3749 		 * reserved at the time of the instantiation.  Use the
3750 		 * allocated space.
3751 		 */
3752 		goto has_addr;
3753 	}
3754 
3755 	/*
3756 	 * Calculate the amount of the address space we would reserve
3757 	 * if the buffer was mapped.
3758 	 */
3759 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3760 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3761 	offset = blkno * bsize;
3762 	maxsize = size + (offset & PAGE_MASK);
3763 	maxsize = imax(maxsize, bsize);
3764 
3765 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3766 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3767 			/*
3768 			 * XXXKIB: defragmentation cannot
3769 			 * succeed, not sure what else to do.
3770 			 */
3771 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3772 		}
3773 		counter_u64_add(mappingrestarts, 1);
3774 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3775 	}
3776 has_addr:
3777 	if (need_mapping) {
3778 		/* b_offset is handled by bpmap_qenter. */
3779 		bp->b_data = bp->b_kvabase;
3780 		BUF_CHECK_MAPPED(bp);
3781 		bpmap_qenter(bp);
3782 	}
3783 }
3784 
3785 struct buf *
3786 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3787     int flags)
3788 {
3789 	struct buf *bp;
3790 	int error;
3791 
3792 	error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3793 	if (error != 0)
3794 		return (NULL);
3795 	return (bp);
3796 }
3797 
3798 /*
3799  *	getblkx:
3800  *
3801  *	Get a block given a specified block and offset into a file/device.
3802  *	The buffers B_DONE bit will be cleared on return, making it almost
3803  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3804  *	return.  The caller should clear B_INVAL prior to initiating a
3805  *	READ.
3806  *
3807  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3808  *	an existing buffer.
3809  *
3810  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3811  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3812  *	and then cleared based on the backing VM.  If the previous buffer is
3813  *	non-0-sized but invalid, B_CACHE will be cleared.
3814  *
3815  *	If getblk() must create a new buffer, the new buffer is returned with
3816  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3817  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3818  *	backing VM.
3819  *
3820  *	getblk() also forces a bwrite() for any B_DELWRI buffer whose
3821  *	B_CACHE bit is clear.
3822  *
3823  *	What this means, basically, is that the caller should use B_CACHE to
3824  *	determine whether the buffer is fully valid or not and should clear
3825  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3826  *	the buffer by loading its data area with something, the caller needs
3827  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3828  *	the caller should set B_CACHE ( as an optimization ), else the caller
3829  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3830  *	a write attempt or if it was a successful read.  If the caller
3831  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3832  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3833  *
3834  *	The blkno parameter is the logical block being requested. Normally
3835  *	the mapping of logical block number to disk block address is done
3836  *	by calling VOP_BMAP(). However, if the mapping is already known, the
3837  *	disk block address can be passed using the dblkno parameter. If the
3838  *	disk block address is not known, then the same value should be passed
3839  *	for blkno and dblkno.
3840  */
3841 int
3842 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3843     int slptimeo, int flags, struct buf **bpp)
3844 {
3845 	struct buf *bp;
3846 	struct bufobj *bo;
3847 	daddr_t d_blkno;
3848 	int bsize, error, maxsize, vmio;
3849 	off_t offset;
3850 
3851 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3852 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3853 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3854 	ASSERT_VOP_LOCKED(vp, "getblk");
3855 	if (size > maxbcachebuf)
3856 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3857 		    maxbcachebuf);
3858 	if (!unmapped_buf_allowed)
3859 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3860 
3861 	bo = &vp->v_bufobj;
3862 	d_blkno = dblkno;
3863 loop:
3864 	BO_RLOCK(bo);
3865 	bp = gbincore(bo, blkno);
3866 	if (bp != NULL) {
3867 		int lockflags;
3868 		/*
3869 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3870 		 * it must be on a queue.
3871 		 */
3872 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3873 
3874 		if ((flags & GB_LOCK_NOWAIT) != 0)
3875 			lockflags |= LK_NOWAIT;
3876 
3877 		error = BUF_TIMELOCK(bp, lockflags,
3878 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3879 
3880 		/*
3881 		 * If we slept and got the lock we have to restart in case
3882 		 * the buffer changed identities.
3883 		 */
3884 		if (error == ENOLCK)
3885 			goto loop;
3886 		/* We timed out or were interrupted. */
3887 		else if (error != 0)
3888 			return (error);
3889 		/* If recursed, assume caller knows the rules. */
3890 		else if (BUF_LOCKRECURSED(bp))
3891 			goto end;
3892 
3893 		/*
3894 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3895 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3896 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3897 		 * backing VM cache.
3898 		 */
3899 		if (bp->b_flags & B_INVAL)
3900 			bp->b_flags &= ~B_CACHE;
3901 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3902 			bp->b_flags |= B_CACHE;
3903 		if (bp->b_flags & B_MANAGED)
3904 			MPASS(bp->b_qindex == QUEUE_NONE);
3905 		else
3906 			bremfree(bp);
3907 
3908 		/*
3909 		 * check for size inconsistencies for non-VMIO case.
3910 		 */
3911 		if (bp->b_bcount != size) {
3912 			if ((bp->b_flags & B_VMIO) == 0 ||
3913 			    (size > bp->b_kvasize)) {
3914 				if (bp->b_flags & B_DELWRI) {
3915 					bp->b_flags |= B_NOCACHE;
3916 					bwrite(bp);
3917 				} else {
3918 					if (LIST_EMPTY(&bp->b_dep)) {
3919 						bp->b_flags |= B_RELBUF;
3920 						brelse(bp);
3921 					} else {
3922 						bp->b_flags |= B_NOCACHE;
3923 						bwrite(bp);
3924 					}
3925 				}
3926 				goto loop;
3927 			}
3928 		}
3929 
3930 		/*
3931 		 * Handle the case of unmapped buffer which should
3932 		 * become mapped, or the buffer for which KVA
3933 		 * reservation is requested.
3934 		 */
3935 		bp_unmapped_get_kva(bp, blkno, size, flags);
3936 
3937 		/*
3938 		 * If the size is inconsistent in the VMIO case, we can resize
3939 		 * the buffer.  This might lead to B_CACHE getting set or
3940 		 * cleared.  If the size has not changed, B_CACHE remains
3941 		 * unchanged from its previous state.
3942 		 */
3943 		allocbuf(bp, size);
3944 
3945 		KASSERT(bp->b_offset != NOOFFSET,
3946 		    ("getblk: no buffer offset"));
3947 
3948 		/*
3949 		 * A buffer with B_DELWRI set and B_CACHE clear must
3950 		 * be committed before we can return the buffer in
3951 		 * order to prevent the caller from issuing a read
3952 		 * ( due to B_CACHE not being set ) and overwriting
3953 		 * it.
3954 		 *
3955 		 * Most callers, including NFS and FFS, need this to
3956 		 * operate properly either because they assume they
3957 		 * can issue a read if B_CACHE is not set, or because
3958 		 * ( for example ) an uncached B_DELWRI might loop due
3959 		 * to softupdates re-dirtying the buffer.  In the latter
3960 		 * case, B_CACHE is set after the first write completes,
3961 		 * preventing further loops.
3962 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3963 		 * above while extending the buffer, we cannot allow the
3964 		 * buffer to remain with B_CACHE set after the write
3965 		 * completes or it will represent a corrupt state.  To
3966 		 * deal with this we set B_NOCACHE to scrap the buffer
3967 		 * after the write.
3968 		 *
3969 		 * We might be able to do something fancy, like setting
3970 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3971 		 * so the below call doesn't set B_CACHE, but that gets real
3972 		 * confusing.  This is much easier.
3973 		 */
3974 
3975 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3976 			bp->b_flags |= B_NOCACHE;
3977 			bwrite(bp);
3978 			goto loop;
3979 		}
3980 		bp->b_flags &= ~B_DONE;
3981 	} else {
3982 		/*
3983 		 * Buffer is not in-core, create new buffer.  The buffer
3984 		 * returned by getnewbuf() is locked.  Note that the returned
3985 		 * buffer is also considered valid (not marked B_INVAL).
3986 		 */
3987 		BO_RUNLOCK(bo);
3988 		/*
3989 		 * If the user does not want us to create the buffer, bail out
3990 		 * here.
3991 		 */
3992 		if (flags & GB_NOCREAT)
3993 			return (EEXIST);
3994 
3995 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3996 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3997 		offset = blkno * bsize;
3998 		vmio = vp->v_object != NULL;
3999 		if (vmio) {
4000 			maxsize = size + (offset & PAGE_MASK);
4001 		} else {
4002 			maxsize = size;
4003 			/* Do not allow non-VMIO notmapped buffers. */
4004 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4005 		}
4006 		maxsize = imax(maxsize, bsize);
4007 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4008 		    !vn_isdisk(vp, NULL)) {
4009 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4010 			KASSERT(error != EOPNOTSUPP,
4011 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4012 			    vp));
4013 			if (error != 0)
4014 				return (error);
4015 			if (d_blkno == -1)
4016 				return (EJUSTRETURN);
4017 		}
4018 
4019 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4020 		if (bp == NULL) {
4021 			if (slpflag || slptimeo)
4022 				return (ETIMEDOUT);
4023 			/*
4024 			 * XXX This is here until the sleep path is diagnosed
4025 			 * enough to work under very low memory conditions.
4026 			 *
4027 			 * There's an issue on low memory, 4BSD+non-preempt
4028 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4029 			 * exhaustion occurs without sleeping for buffer
4030 			 * reclaimation.  This just sticks in a loop and
4031 			 * constantly attempts to allocate a buffer, which
4032 			 * hits exhaustion and tries to wakeup bufdaemon.
4033 			 * This never happens because we never yield.
4034 			 *
4035 			 * The real solution is to identify and fix these cases
4036 			 * so we aren't effectively busy-waiting in a loop
4037 			 * until the reclaimation path has cycles to run.
4038 			 */
4039 			kern_yield(PRI_USER);
4040 			goto loop;
4041 		}
4042 
4043 		/*
4044 		 * This code is used to make sure that a buffer is not
4045 		 * created while the getnewbuf routine is blocked.
4046 		 * This can be a problem whether the vnode is locked or not.
4047 		 * If the buffer is created out from under us, we have to
4048 		 * throw away the one we just created.
4049 		 *
4050 		 * Note: this must occur before we associate the buffer
4051 		 * with the vp especially considering limitations in
4052 		 * the splay tree implementation when dealing with duplicate
4053 		 * lblkno's.
4054 		 */
4055 		BO_LOCK(bo);
4056 		if (gbincore(bo, blkno)) {
4057 			BO_UNLOCK(bo);
4058 			bp->b_flags |= B_INVAL;
4059 			bufspace_release(bufdomain(bp), maxsize);
4060 			brelse(bp);
4061 			goto loop;
4062 		}
4063 
4064 		/*
4065 		 * Insert the buffer into the hash, so that it can
4066 		 * be found by incore.
4067 		 */
4068 		bp->b_lblkno = blkno;
4069 		bp->b_blkno = d_blkno;
4070 		bp->b_offset = offset;
4071 		bgetvp(vp, bp);
4072 		BO_UNLOCK(bo);
4073 
4074 		/*
4075 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4076 		 * buffer size starts out as 0, B_CACHE will be set by
4077 		 * allocbuf() for the VMIO case prior to it testing the
4078 		 * backing store for validity.
4079 		 */
4080 
4081 		if (vmio) {
4082 			bp->b_flags |= B_VMIO;
4083 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4084 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4085 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4086 		} else {
4087 			bp->b_flags &= ~B_VMIO;
4088 			KASSERT(bp->b_bufobj->bo_object == NULL,
4089 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4090 			    bp, bp->b_bufobj->bo_object));
4091 			BUF_CHECK_MAPPED(bp);
4092 		}
4093 
4094 		allocbuf(bp, size);
4095 		bufspace_release(bufdomain(bp), maxsize);
4096 		bp->b_flags &= ~B_DONE;
4097 	}
4098 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4099 end:
4100 	buf_track(bp, __func__);
4101 	KASSERT(bp->b_bufobj == bo,
4102 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4103 	*bpp = bp;
4104 	return (0);
4105 }
4106 
4107 /*
4108  * Get an empty, disassociated buffer of given size.  The buffer is initially
4109  * set to B_INVAL.
4110  */
4111 struct buf *
4112 geteblk(int size, int flags)
4113 {
4114 	struct buf *bp;
4115 	int maxsize;
4116 
4117 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4118 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4119 		if ((flags & GB_NOWAIT_BD) &&
4120 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4121 			return (NULL);
4122 	}
4123 	allocbuf(bp, size);
4124 	bufspace_release(bufdomain(bp), maxsize);
4125 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4126 	return (bp);
4127 }
4128 
4129 /*
4130  * Truncate the backing store for a non-vmio buffer.
4131  */
4132 static void
4133 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4134 {
4135 
4136 	if (bp->b_flags & B_MALLOC) {
4137 		/*
4138 		 * malloced buffers are not shrunk
4139 		 */
4140 		if (newbsize == 0) {
4141 			bufmallocadjust(bp, 0);
4142 			free(bp->b_data, M_BIOBUF);
4143 			bp->b_data = bp->b_kvabase;
4144 			bp->b_flags &= ~B_MALLOC;
4145 		}
4146 		return;
4147 	}
4148 	vm_hold_free_pages(bp, newbsize);
4149 	bufspace_adjust(bp, newbsize);
4150 }
4151 
4152 /*
4153  * Extend the backing for a non-VMIO buffer.
4154  */
4155 static void
4156 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4157 {
4158 	caddr_t origbuf;
4159 	int origbufsize;
4160 
4161 	/*
4162 	 * We only use malloced memory on the first allocation.
4163 	 * and revert to page-allocated memory when the buffer
4164 	 * grows.
4165 	 *
4166 	 * There is a potential smp race here that could lead
4167 	 * to bufmallocspace slightly passing the max.  It
4168 	 * is probably extremely rare and not worth worrying
4169 	 * over.
4170 	 */
4171 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4172 	    bufmallocspace < maxbufmallocspace) {
4173 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4174 		bp->b_flags |= B_MALLOC;
4175 		bufmallocadjust(bp, newbsize);
4176 		return;
4177 	}
4178 
4179 	/*
4180 	 * If the buffer is growing on its other-than-first
4181 	 * allocation then we revert to the page-allocation
4182 	 * scheme.
4183 	 */
4184 	origbuf = NULL;
4185 	origbufsize = 0;
4186 	if (bp->b_flags & B_MALLOC) {
4187 		origbuf = bp->b_data;
4188 		origbufsize = bp->b_bufsize;
4189 		bp->b_data = bp->b_kvabase;
4190 		bufmallocadjust(bp, 0);
4191 		bp->b_flags &= ~B_MALLOC;
4192 		newbsize = round_page(newbsize);
4193 	}
4194 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4195 	    (vm_offset_t) bp->b_data + newbsize);
4196 	if (origbuf != NULL) {
4197 		bcopy(origbuf, bp->b_data, origbufsize);
4198 		free(origbuf, M_BIOBUF);
4199 	}
4200 	bufspace_adjust(bp, newbsize);
4201 }
4202 
4203 /*
4204  * This code constitutes the buffer memory from either anonymous system
4205  * memory (in the case of non-VMIO operations) or from an associated
4206  * VM object (in the case of VMIO operations).  This code is able to
4207  * resize a buffer up or down.
4208  *
4209  * Note that this code is tricky, and has many complications to resolve
4210  * deadlock or inconsistent data situations.  Tread lightly!!!
4211  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4212  * the caller.  Calling this code willy nilly can result in the loss of data.
4213  *
4214  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4215  * B_CACHE for the non-VMIO case.
4216  */
4217 int
4218 allocbuf(struct buf *bp, int size)
4219 {
4220 	int newbsize;
4221 
4222 	if (bp->b_bcount == size)
4223 		return (1);
4224 
4225 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4226 		panic("allocbuf: buffer too small");
4227 
4228 	newbsize = roundup2(size, DEV_BSIZE);
4229 	if ((bp->b_flags & B_VMIO) == 0) {
4230 		if ((bp->b_flags & B_MALLOC) == 0)
4231 			newbsize = round_page(newbsize);
4232 		/*
4233 		 * Just get anonymous memory from the kernel.  Don't
4234 		 * mess with B_CACHE.
4235 		 */
4236 		if (newbsize < bp->b_bufsize)
4237 			vfs_nonvmio_truncate(bp, newbsize);
4238 		else if (newbsize > bp->b_bufsize)
4239 			vfs_nonvmio_extend(bp, newbsize);
4240 	} else {
4241 		int desiredpages;
4242 
4243 		desiredpages = (size == 0) ? 0 :
4244 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4245 
4246 		if (bp->b_flags & B_MALLOC)
4247 			panic("allocbuf: VMIO buffer can't be malloced");
4248 		/*
4249 		 * Set B_CACHE initially if buffer is 0 length or will become
4250 		 * 0-length.
4251 		 */
4252 		if (size == 0 || bp->b_bufsize == 0)
4253 			bp->b_flags |= B_CACHE;
4254 
4255 		if (newbsize < bp->b_bufsize)
4256 			vfs_vmio_truncate(bp, desiredpages);
4257 		/* XXX This looks as if it should be newbsize > b_bufsize */
4258 		else if (size > bp->b_bcount)
4259 			vfs_vmio_extend(bp, desiredpages, size);
4260 		bufspace_adjust(bp, newbsize);
4261 	}
4262 	bp->b_bcount = size;		/* requested buffer size. */
4263 	return (1);
4264 }
4265 
4266 extern int inflight_transient_maps;
4267 
4268 static struct bio_queue nondump_bios;
4269 
4270 void
4271 biodone(struct bio *bp)
4272 {
4273 	struct mtx *mtxp;
4274 	void (*done)(struct bio *);
4275 	vm_offset_t start, end;
4276 
4277 	biotrack(bp, __func__);
4278 
4279 	/*
4280 	 * Avoid completing I/O when dumping after a panic since that may
4281 	 * result in a deadlock in the filesystem or pager code.  Note that
4282 	 * this doesn't affect dumps that were started manually since we aim
4283 	 * to keep the system usable after it has been resumed.
4284 	 */
4285 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4286 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4287 		return;
4288 	}
4289 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4290 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4291 		bp->bio_flags |= BIO_UNMAPPED;
4292 		start = trunc_page((vm_offset_t)bp->bio_data);
4293 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4294 		bp->bio_data = unmapped_buf;
4295 		pmap_qremove(start, atop(end - start));
4296 		vmem_free(transient_arena, start, end - start);
4297 		atomic_add_int(&inflight_transient_maps, -1);
4298 	}
4299 	done = bp->bio_done;
4300 	if (done == NULL) {
4301 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4302 		mtx_lock(mtxp);
4303 		bp->bio_flags |= BIO_DONE;
4304 		wakeup(bp);
4305 		mtx_unlock(mtxp);
4306 	} else
4307 		done(bp);
4308 }
4309 
4310 /*
4311  * Wait for a BIO to finish.
4312  */
4313 int
4314 biowait(struct bio *bp, const char *wchan)
4315 {
4316 	struct mtx *mtxp;
4317 
4318 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4319 	mtx_lock(mtxp);
4320 	while ((bp->bio_flags & BIO_DONE) == 0)
4321 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4322 	mtx_unlock(mtxp);
4323 	if (bp->bio_error != 0)
4324 		return (bp->bio_error);
4325 	if (!(bp->bio_flags & BIO_ERROR))
4326 		return (0);
4327 	return (EIO);
4328 }
4329 
4330 void
4331 biofinish(struct bio *bp, struct devstat *stat, int error)
4332 {
4333 
4334 	if (error) {
4335 		bp->bio_error = error;
4336 		bp->bio_flags |= BIO_ERROR;
4337 	}
4338 	if (stat != NULL)
4339 		devstat_end_transaction_bio(stat, bp);
4340 	biodone(bp);
4341 }
4342 
4343 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4344 void
4345 biotrack_buf(struct bio *bp, const char *location)
4346 {
4347 
4348 	buf_track(bp->bio_track_bp, location);
4349 }
4350 #endif
4351 
4352 /*
4353  *	bufwait:
4354  *
4355  *	Wait for buffer I/O completion, returning error status.  The buffer
4356  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4357  *	error and cleared.
4358  */
4359 int
4360 bufwait(struct buf *bp)
4361 {
4362 	if (bp->b_iocmd == BIO_READ)
4363 		bwait(bp, PRIBIO, "biord");
4364 	else
4365 		bwait(bp, PRIBIO, "biowr");
4366 	if (bp->b_flags & B_EINTR) {
4367 		bp->b_flags &= ~B_EINTR;
4368 		return (EINTR);
4369 	}
4370 	if (bp->b_ioflags & BIO_ERROR) {
4371 		return (bp->b_error ? bp->b_error : EIO);
4372 	} else {
4373 		return (0);
4374 	}
4375 }
4376 
4377 /*
4378  *	bufdone:
4379  *
4380  *	Finish I/O on a buffer, optionally calling a completion function.
4381  *	This is usually called from an interrupt so process blocking is
4382  *	not allowed.
4383  *
4384  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4385  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4386  *	assuming B_INVAL is clear.
4387  *
4388  *	For the VMIO case, we set B_CACHE if the op was a read and no
4389  *	read error occurred, or if the op was a write.  B_CACHE is never
4390  *	set if the buffer is invalid or otherwise uncacheable.
4391  *
4392  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4393  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4394  *	in the biodone routine.
4395  */
4396 void
4397 bufdone(struct buf *bp)
4398 {
4399 	struct bufobj *dropobj;
4400 	void    (*biodone)(struct buf *);
4401 
4402 	buf_track(bp, __func__);
4403 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4404 	dropobj = NULL;
4405 
4406 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4407 
4408 	runningbufwakeup(bp);
4409 	if (bp->b_iocmd == BIO_WRITE)
4410 		dropobj = bp->b_bufobj;
4411 	/* call optional completion function if requested */
4412 	if (bp->b_iodone != NULL) {
4413 		biodone = bp->b_iodone;
4414 		bp->b_iodone = NULL;
4415 		(*biodone) (bp);
4416 		if (dropobj)
4417 			bufobj_wdrop(dropobj);
4418 		return;
4419 	}
4420 	if (bp->b_flags & B_VMIO) {
4421 		/*
4422 		 * Set B_CACHE if the op was a normal read and no error
4423 		 * occurred.  B_CACHE is set for writes in the b*write()
4424 		 * routines.
4425 		 */
4426 		if (bp->b_iocmd == BIO_READ &&
4427 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4428 		    !(bp->b_ioflags & BIO_ERROR))
4429 			bp->b_flags |= B_CACHE;
4430 		vfs_vmio_iodone(bp);
4431 	}
4432 	if (!LIST_EMPTY(&bp->b_dep))
4433 		buf_complete(bp);
4434 	if ((bp->b_flags & B_CKHASH) != 0) {
4435 		KASSERT(bp->b_iocmd == BIO_READ,
4436 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4437 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4438 		(*bp->b_ckhashcalc)(bp);
4439 	}
4440 	/*
4441 	 * For asynchronous completions, release the buffer now. The brelse
4442 	 * will do a wakeup there if necessary - so no need to do a wakeup
4443 	 * here in the async case. The sync case always needs to do a wakeup.
4444 	 */
4445 	if (bp->b_flags & B_ASYNC) {
4446 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4447 		    (bp->b_ioflags & BIO_ERROR))
4448 			brelse(bp);
4449 		else
4450 			bqrelse(bp);
4451 	} else
4452 		bdone(bp);
4453 	if (dropobj)
4454 		bufobj_wdrop(dropobj);
4455 }
4456 
4457 /*
4458  * This routine is called in lieu of iodone in the case of
4459  * incomplete I/O.  This keeps the busy status for pages
4460  * consistent.
4461  */
4462 void
4463 vfs_unbusy_pages(struct buf *bp)
4464 {
4465 	int i;
4466 	vm_object_t obj;
4467 	vm_page_t m;
4468 
4469 	runningbufwakeup(bp);
4470 	if (!(bp->b_flags & B_VMIO))
4471 		return;
4472 
4473 	obj = bp->b_bufobj->bo_object;
4474 	for (i = 0; i < bp->b_npages; i++) {
4475 		m = bp->b_pages[i];
4476 		if (m == bogus_page) {
4477 			m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4478 			if (!m)
4479 				panic("vfs_unbusy_pages: page missing\n");
4480 			bp->b_pages[i] = m;
4481 			if (buf_mapped(bp)) {
4482 				BUF_CHECK_MAPPED(bp);
4483 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4484 				    bp->b_pages, bp->b_npages);
4485 			} else
4486 				BUF_CHECK_UNMAPPED(bp);
4487 		}
4488 		vm_page_sunbusy(m);
4489 	}
4490 	vm_object_pip_wakeupn(obj, bp->b_npages);
4491 }
4492 
4493 /*
4494  * vfs_page_set_valid:
4495  *
4496  *	Set the valid bits in a page based on the supplied offset.   The
4497  *	range is restricted to the buffer's size.
4498  *
4499  *	This routine is typically called after a read completes.
4500  */
4501 static void
4502 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4503 {
4504 	vm_ooffset_t eoff;
4505 
4506 	/*
4507 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4508 	 * page boundary and eoff is not greater than the end of the buffer.
4509 	 * The end of the buffer, in this case, is our file EOF, not the
4510 	 * allocation size of the buffer.
4511 	 */
4512 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4513 	if (eoff > bp->b_offset + bp->b_bcount)
4514 		eoff = bp->b_offset + bp->b_bcount;
4515 
4516 	/*
4517 	 * Set valid range.  This is typically the entire buffer and thus the
4518 	 * entire page.
4519 	 */
4520 	if (eoff > off)
4521 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4522 }
4523 
4524 /*
4525  * vfs_page_set_validclean:
4526  *
4527  *	Set the valid bits and clear the dirty bits in a page based on the
4528  *	supplied offset.   The range is restricted to the buffer's size.
4529  */
4530 static void
4531 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4532 {
4533 	vm_ooffset_t soff, eoff;
4534 
4535 	/*
4536 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4537 	 * page boundary or cross the end of the buffer.  The end of the
4538 	 * buffer, in this case, is our file EOF, not the allocation size
4539 	 * of the buffer.
4540 	 */
4541 	soff = off;
4542 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4543 	if (eoff > bp->b_offset + bp->b_bcount)
4544 		eoff = bp->b_offset + bp->b_bcount;
4545 
4546 	/*
4547 	 * Set valid range.  This is typically the entire buffer and thus the
4548 	 * entire page.
4549 	 */
4550 	if (eoff > soff) {
4551 		vm_page_set_validclean(
4552 		    m,
4553 		   (vm_offset_t) (soff & PAGE_MASK),
4554 		   (vm_offset_t) (eoff - soff)
4555 		);
4556 	}
4557 }
4558 
4559 /*
4560  * Acquire a shared busy on all pages in the buf.
4561  */
4562 void
4563 vfs_busy_pages_acquire(struct buf *bp)
4564 {
4565 	int i;
4566 
4567 	for (i = 0; i < bp->b_npages; i++)
4568 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4569 }
4570 
4571 void
4572 vfs_busy_pages_release(struct buf *bp)
4573 {
4574 	int i;
4575 
4576 	for (i = 0; i < bp->b_npages; i++)
4577 		vm_page_sunbusy(bp->b_pages[i]);
4578 }
4579 
4580 /*
4581  * This routine is called before a device strategy routine.
4582  * It is used to tell the VM system that paging I/O is in
4583  * progress, and treat the pages associated with the buffer
4584  * almost as being exclusive busy.  Also the object paging_in_progress
4585  * flag is handled to make sure that the object doesn't become
4586  * inconsistent.
4587  *
4588  * Since I/O has not been initiated yet, certain buffer flags
4589  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4590  * and should be ignored.
4591  */
4592 void
4593 vfs_busy_pages(struct buf *bp, int clear_modify)
4594 {
4595 	vm_object_t obj;
4596 	vm_ooffset_t foff;
4597 	vm_page_t m;
4598 	int i;
4599 	bool bogus;
4600 
4601 	if (!(bp->b_flags & B_VMIO))
4602 		return;
4603 
4604 	obj = bp->b_bufobj->bo_object;
4605 	foff = bp->b_offset;
4606 	KASSERT(bp->b_offset != NOOFFSET,
4607 	    ("vfs_busy_pages: no buffer offset"));
4608 	if ((bp->b_flags & B_CLUSTER) == 0) {
4609 		vm_object_pip_add(obj, bp->b_npages);
4610 		vfs_busy_pages_acquire(bp);
4611 	}
4612 	if (bp->b_bufsize != 0)
4613 		vfs_setdirty_range(bp);
4614 	bogus = false;
4615 	for (i = 0; i < bp->b_npages; i++) {
4616 		m = bp->b_pages[i];
4617 		vm_page_assert_sbusied(m);
4618 
4619 		/*
4620 		 * When readying a buffer for a read ( i.e
4621 		 * clear_modify == 0 ), it is important to do
4622 		 * bogus_page replacement for valid pages in
4623 		 * partially instantiated buffers.  Partially
4624 		 * instantiated buffers can, in turn, occur when
4625 		 * reconstituting a buffer from its VM backing store
4626 		 * base.  We only have to do this if B_CACHE is
4627 		 * clear ( which causes the I/O to occur in the
4628 		 * first place ).  The replacement prevents the read
4629 		 * I/O from overwriting potentially dirty VM-backed
4630 		 * pages.  XXX bogus page replacement is, uh, bogus.
4631 		 * It may not work properly with small-block devices.
4632 		 * We need to find a better way.
4633 		 */
4634 		if (clear_modify) {
4635 			pmap_remove_write(m);
4636 			vfs_page_set_validclean(bp, foff, m);
4637 		} else if (vm_page_all_valid(m) &&
4638 		    (bp->b_flags & B_CACHE) == 0) {
4639 			bp->b_pages[i] = bogus_page;
4640 			bogus = true;
4641 		}
4642 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4643 	}
4644 	if (bogus && buf_mapped(bp)) {
4645 		BUF_CHECK_MAPPED(bp);
4646 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4647 		    bp->b_pages, bp->b_npages);
4648 	}
4649 }
4650 
4651 /*
4652  *	vfs_bio_set_valid:
4653  *
4654  *	Set the range within the buffer to valid.  The range is
4655  *	relative to the beginning of the buffer, b_offset.  Note that
4656  *	b_offset itself may be offset from the beginning of the first
4657  *	page.
4658  */
4659 void
4660 vfs_bio_set_valid(struct buf *bp, int base, int size)
4661 {
4662 	int i, n;
4663 	vm_page_t m;
4664 
4665 	if (!(bp->b_flags & B_VMIO))
4666 		return;
4667 
4668 	/*
4669 	 * Fixup base to be relative to beginning of first page.
4670 	 * Set initial n to be the maximum number of bytes in the
4671 	 * first page that can be validated.
4672 	 */
4673 	base += (bp->b_offset & PAGE_MASK);
4674 	n = PAGE_SIZE - (base & PAGE_MASK);
4675 
4676 	/*
4677 	 * Busy may not be strictly necessary here because the pages are
4678 	 * unlikely to be fully valid and the vnode lock will synchronize
4679 	 * their access via getpages.  It is grabbed for consistency with
4680 	 * other page validation.
4681 	 */
4682 	vfs_busy_pages_acquire(bp);
4683 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4684 		m = bp->b_pages[i];
4685 		if (n > size)
4686 			n = size;
4687 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4688 		base += n;
4689 		size -= n;
4690 		n = PAGE_SIZE;
4691 	}
4692 	vfs_busy_pages_release(bp);
4693 }
4694 
4695 /*
4696  *	vfs_bio_clrbuf:
4697  *
4698  *	If the specified buffer is a non-VMIO buffer, clear the entire
4699  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4700  *	validate only the previously invalid portions of the buffer.
4701  *	This routine essentially fakes an I/O, so we need to clear
4702  *	BIO_ERROR and B_INVAL.
4703  *
4704  *	Note that while we only theoretically need to clear through b_bcount,
4705  *	we go ahead and clear through b_bufsize.
4706  */
4707 void
4708 vfs_bio_clrbuf(struct buf *bp)
4709 {
4710 	int i, j, mask, sa, ea, slide;
4711 
4712 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4713 		clrbuf(bp);
4714 		return;
4715 	}
4716 	bp->b_flags &= ~B_INVAL;
4717 	bp->b_ioflags &= ~BIO_ERROR;
4718 	vfs_busy_pages_acquire(bp);
4719 	sa = bp->b_offset & PAGE_MASK;
4720 	slide = 0;
4721 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4722 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4723 		ea = slide & PAGE_MASK;
4724 		if (ea == 0)
4725 			ea = PAGE_SIZE;
4726 		if (bp->b_pages[i] == bogus_page)
4727 			continue;
4728 		j = sa / DEV_BSIZE;
4729 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4730 		if ((bp->b_pages[i]->valid & mask) == mask)
4731 			continue;
4732 		if ((bp->b_pages[i]->valid & mask) == 0)
4733 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4734 		else {
4735 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4736 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4737 					pmap_zero_page_area(bp->b_pages[i],
4738 					    sa, DEV_BSIZE);
4739 				}
4740 			}
4741 		}
4742 		vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4743 		    roundup2(ea - sa, DEV_BSIZE));
4744 	}
4745 	vfs_busy_pages_release(bp);
4746 	bp->b_resid = 0;
4747 }
4748 
4749 void
4750 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4751 {
4752 	vm_page_t m;
4753 	int i, n;
4754 
4755 	if (buf_mapped(bp)) {
4756 		BUF_CHECK_MAPPED(bp);
4757 		bzero(bp->b_data + base, size);
4758 	} else {
4759 		BUF_CHECK_UNMAPPED(bp);
4760 		n = PAGE_SIZE - (base & PAGE_MASK);
4761 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4762 			m = bp->b_pages[i];
4763 			if (n > size)
4764 				n = size;
4765 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4766 			base += n;
4767 			size -= n;
4768 			n = PAGE_SIZE;
4769 		}
4770 	}
4771 }
4772 
4773 /*
4774  * Update buffer flags based on I/O request parameters, optionally releasing the
4775  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4776  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4777  * I/O).  Otherwise the buffer is released to the cache.
4778  */
4779 static void
4780 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4781 {
4782 
4783 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4784 	    ("buf %p non-VMIO noreuse", bp));
4785 
4786 	if ((ioflag & IO_DIRECT) != 0)
4787 		bp->b_flags |= B_DIRECT;
4788 	if ((ioflag & IO_EXT) != 0)
4789 		bp->b_xflags |= BX_ALTDATA;
4790 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4791 		bp->b_flags |= B_RELBUF;
4792 		if ((ioflag & IO_NOREUSE) != 0)
4793 			bp->b_flags |= B_NOREUSE;
4794 		if (release)
4795 			brelse(bp);
4796 	} else if (release)
4797 		bqrelse(bp);
4798 }
4799 
4800 void
4801 vfs_bio_brelse(struct buf *bp, int ioflag)
4802 {
4803 
4804 	b_io_dismiss(bp, ioflag, true);
4805 }
4806 
4807 void
4808 vfs_bio_set_flags(struct buf *bp, int ioflag)
4809 {
4810 
4811 	b_io_dismiss(bp, ioflag, false);
4812 }
4813 
4814 /*
4815  * vm_hold_load_pages and vm_hold_free_pages get pages into
4816  * a buffers address space.  The pages are anonymous and are
4817  * not associated with a file object.
4818  */
4819 static void
4820 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4821 {
4822 	vm_offset_t pg;
4823 	vm_page_t p;
4824 	int index;
4825 
4826 	BUF_CHECK_MAPPED(bp);
4827 
4828 	to = round_page(to);
4829 	from = round_page(from);
4830 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4831 
4832 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4833 		/*
4834 		 * note: must allocate system pages since blocking here
4835 		 * could interfere with paging I/O, no matter which
4836 		 * process we are.
4837 		 */
4838 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4839 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4840 		    VM_ALLOC_WAITOK);
4841 		pmap_qenter(pg, &p, 1);
4842 		bp->b_pages[index] = p;
4843 	}
4844 	bp->b_npages = index;
4845 }
4846 
4847 /* Return pages associated with this buf to the vm system */
4848 static void
4849 vm_hold_free_pages(struct buf *bp, int newbsize)
4850 {
4851 	vm_offset_t from;
4852 	vm_page_t p;
4853 	int index, newnpages;
4854 
4855 	BUF_CHECK_MAPPED(bp);
4856 
4857 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4858 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4859 	if (bp->b_npages > newnpages)
4860 		pmap_qremove(from, bp->b_npages - newnpages);
4861 	for (index = newnpages; index < bp->b_npages; index++) {
4862 		p = bp->b_pages[index];
4863 		bp->b_pages[index] = NULL;
4864 		vm_page_unwire_noq(p);
4865 		vm_page_free(p);
4866 	}
4867 	bp->b_npages = newnpages;
4868 }
4869 
4870 /*
4871  * Map an IO request into kernel virtual address space.
4872  *
4873  * All requests are (re)mapped into kernel VA space.
4874  * Notice that we use b_bufsize for the size of the buffer
4875  * to be mapped.  b_bcount might be modified by the driver.
4876  *
4877  * Note that even if the caller determines that the address space should
4878  * be valid, a race or a smaller-file mapped into a larger space may
4879  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4880  * check the return value.
4881  *
4882  * This function only works with pager buffers.
4883  */
4884 int
4885 vmapbuf(struct buf *bp, int mapbuf)
4886 {
4887 	vm_prot_t prot;
4888 	int pidx;
4889 
4890 	if (bp->b_bufsize < 0)
4891 		return (-1);
4892 	prot = VM_PROT_READ;
4893 	if (bp->b_iocmd == BIO_READ)
4894 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4895 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4896 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4897 	    btoc(MAXPHYS))) < 0)
4898 		return (-1);
4899 	bp->b_npages = pidx;
4900 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4901 	if (mapbuf || !unmapped_buf_allowed) {
4902 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4903 		bp->b_data = bp->b_kvabase + bp->b_offset;
4904 	} else
4905 		bp->b_data = unmapped_buf;
4906 	return(0);
4907 }
4908 
4909 /*
4910  * Free the io map PTEs associated with this IO operation.
4911  * We also invalidate the TLB entries and restore the original b_addr.
4912  *
4913  * This function only works with pager buffers.
4914  */
4915 void
4916 vunmapbuf(struct buf *bp)
4917 {
4918 	int npages;
4919 
4920 	npages = bp->b_npages;
4921 	if (buf_mapped(bp))
4922 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4923 	vm_page_unhold_pages(bp->b_pages, npages);
4924 
4925 	bp->b_data = unmapped_buf;
4926 }
4927 
4928 void
4929 bdone(struct buf *bp)
4930 {
4931 	struct mtx *mtxp;
4932 
4933 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4934 	mtx_lock(mtxp);
4935 	bp->b_flags |= B_DONE;
4936 	wakeup(bp);
4937 	mtx_unlock(mtxp);
4938 }
4939 
4940 void
4941 bwait(struct buf *bp, u_char pri, const char *wchan)
4942 {
4943 	struct mtx *mtxp;
4944 
4945 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4946 	mtx_lock(mtxp);
4947 	while ((bp->b_flags & B_DONE) == 0)
4948 		msleep(bp, mtxp, pri, wchan, 0);
4949 	mtx_unlock(mtxp);
4950 }
4951 
4952 int
4953 bufsync(struct bufobj *bo, int waitfor)
4954 {
4955 
4956 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
4957 }
4958 
4959 void
4960 bufstrategy(struct bufobj *bo, struct buf *bp)
4961 {
4962 	int i __unused;
4963 	struct vnode *vp;
4964 
4965 	vp = bp->b_vp;
4966 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4967 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4968 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4969 	i = VOP_STRATEGY(vp, bp);
4970 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4971 }
4972 
4973 /*
4974  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
4975  */
4976 void
4977 bufobj_init(struct bufobj *bo, void *private)
4978 {
4979 	static volatile int bufobj_cleanq;
4980 
4981         bo->bo_domain =
4982             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
4983         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
4984         bo->bo_private = private;
4985         TAILQ_INIT(&bo->bo_clean.bv_hd);
4986         TAILQ_INIT(&bo->bo_dirty.bv_hd);
4987 }
4988 
4989 void
4990 bufobj_wrefl(struct bufobj *bo)
4991 {
4992 
4993 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4994 	ASSERT_BO_WLOCKED(bo);
4995 	bo->bo_numoutput++;
4996 }
4997 
4998 void
4999 bufobj_wref(struct bufobj *bo)
5000 {
5001 
5002 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5003 	BO_LOCK(bo);
5004 	bo->bo_numoutput++;
5005 	BO_UNLOCK(bo);
5006 }
5007 
5008 void
5009 bufobj_wdrop(struct bufobj *bo)
5010 {
5011 
5012 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5013 	BO_LOCK(bo);
5014 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5015 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5016 		bo->bo_flag &= ~BO_WWAIT;
5017 		wakeup(&bo->bo_numoutput);
5018 	}
5019 	BO_UNLOCK(bo);
5020 }
5021 
5022 int
5023 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5024 {
5025 	int error;
5026 
5027 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5028 	ASSERT_BO_WLOCKED(bo);
5029 	error = 0;
5030 	while (bo->bo_numoutput) {
5031 		bo->bo_flag |= BO_WWAIT;
5032 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5033 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5034 		if (error)
5035 			break;
5036 	}
5037 	return (error);
5038 }
5039 
5040 /*
5041  * Set bio_data or bio_ma for struct bio from the struct buf.
5042  */
5043 void
5044 bdata2bio(struct buf *bp, struct bio *bip)
5045 {
5046 
5047 	if (!buf_mapped(bp)) {
5048 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5049 		bip->bio_ma = bp->b_pages;
5050 		bip->bio_ma_n = bp->b_npages;
5051 		bip->bio_data = unmapped_buf;
5052 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5053 		bip->bio_flags |= BIO_UNMAPPED;
5054 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5055 		    PAGE_SIZE == bp->b_npages,
5056 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5057 		    (long long)bip->bio_length, bip->bio_ma_n));
5058 	} else {
5059 		bip->bio_data = bp->b_data;
5060 		bip->bio_ma = NULL;
5061 	}
5062 }
5063 
5064 /*
5065  * The MIPS pmap code currently doesn't handle aliased pages.
5066  * The VIPT caches may not handle page aliasing themselves, leading
5067  * to data corruption.
5068  *
5069  * As such, this code makes a system extremely unhappy if said
5070  * system doesn't support unaliasing the above situation in hardware.
5071  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5072  * this feature at build time, so it has to be handled in software.
5073  *
5074  * Once the MIPS pmap/cache code grows to support this function on
5075  * earlier chips, it should be flipped back off.
5076  */
5077 #ifdef	__mips__
5078 static int buf_pager_relbuf = 1;
5079 #else
5080 static int buf_pager_relbuf = 0;
5081 #endif
5082 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5083     &buf_pager_relbuf, 0,
5084     "Make buffer pager release buffers after reading");
5085 
5086 /*
5087  * The buffer pager.  It uses buffer reads to validate pages.
5088  *
5089  * In contrast to the generic local pager from vm/vnode_pager.c, this
5090  * pager correctly and easily handles volumes where the underlying
5091  * device block size is greater than the machine page size.  The
5092  * buffer cache transparently extends the requested page run to be
5093  * aligned at the block boundary, and does the necessary bogus page
5094  * replacements in the addends to avoid obliterating already valid
5095  * pages.
5096  *
5097  * The only non-trivial issue is that the exclusive busy state for
5098  * pages, which is assumed by the vm_pager_getpages() interface, is
5099  * incompatible with the VMIO buffer cache's desire to share-busy the
5100  * pages.  This function performs a trivial downgrade of the pages'
5101  * state before reading buffers, and a less trivial upgrade from the
5102  * shared-busy to excl-busy state after the read.
5103  */
5104 int
5105 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5106     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5107     vbg_get_blksize_t get_blksize)
5108 {
5109 	vm_page_t m;
5110 	vm_object_t object;
5111 	struct buf *bp;
5112 	struct mount *mp;
5113 	daddr_t lbn, lbnp;
5114 	vm_ooffset_t la, lb, poff, poffe;
5115 	long bsize;
5116 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5117 	bool redo, lpart;
5118 
5119 	object = vp->v_object;
5120 	mp = vp->v_mount;
5121 	error = 0;
5122 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5123 	if (la >= object->un_pager.vnp.vnp_size)
5124 		return (VM_PAGER_BAD);
5125 
5126 	/*
5127 	 * Change the meaning of la from where the last requested page starts
5128 	 * to where it ends, because that's the end of the requested region
5129 	 * and the start of the potential read-ahead region.
5130 	 */
5131 	la += PAGE_SIZE;
5132 	lpart = la > object->un_pager.vnp.vnp_size;
5133 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5134 
5135 	/*
5136 	 * Calculate read-ahead, behind and total pages.
5137 	 */
5138 	pgsin = count;
5139 	lb = IDX_TO_OFF(ma[0]->pindex);
5140 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5141 	pgsin += pgsin_b;
5142 	if (rbehind != NULL)
5143 		*rbehind = pgsin_b;
5144 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5145 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5146 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5147 		    PAGE_SIZE) - la);
5148 	pgsin += pgsin_a;
5149 	if (rahead != NULL)
5150 		*rahead = pgsin_a;
5151 	VM_CNT_INC(v_vnodein);
5152 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5153 
5154 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5155 	    != 0) ? GB_UNMAPPED : 0;
5156 again:
5157 	for (i = 0; i < count; i++) {
5158 		if (ma[i] != bogus_page)
5159 			vm_page_busy_downgrade(ma[i]);
5160 	}
5161 
5162 	lbnp = -1;
5163 	for (i = 0; i < count; i++) {
5164 		m = ma[i];
5165 		if (m == bogus_page)
5166 			continue;
5167 
5168 		/*
5169 		 * Pages are shared busy and the object lock is not
5170 		 * owned, which together allow for the pages'
5171 		 * invalidation.  The racy test for validity avoids
5172 		 * useless creation of the buffer for the most typical
5173 		 * case when invalidation is not used in redo or for
5174 		 * parallel read.  The shared->excl upgrade loop at
5175 		 * the end of the function catches the race in a
5176 		 * reliable way (protected by the object lock).
5177 		 */
5178 		if (vm_page_all_valid(m))
5179 			continue;
5180 
5181 		poff = IDX_TO_OFF(m->pindex);
5182 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5183 		for (; poff < poffe; poff += bsize) {
5184 			lbn = get_lblkno(vp, poff);
5185 			if (lbn == lbnp)
5186 				goto next_page;
5187 			lbnp = lbn;
5188 
5189 			bsize = get_blksize(vp, lbn);
5190 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5191 			    br_flags, &bp);
5192 			if (error != 0)
5193 				goto end_pages;
5194 			if (bp->b_rcred == curthread->td_ucred) {
5195 				crfree(bp->b_rcred);
5196 				bp->b_rcred = NOCRED;
5197 			}
5198 			if (LIST_EMPTY(&bp->b_dep)) {
5199 				/*
5200 				 * Invalidation clears m->valid, but
5201 				 * may leave B_CACHE flag if the
5202 				 * buffer existed at the invalidation
5203 				 * time.  In this case, recycle the
5204 				 * buffer to do real read on next
5205 				 * bread() after redo.
5206 				 *
5207 				 * Otherwise B_RELBUF is not strictly
5208 				 * necessary, enable to reduce buf
5209 				 * cache pressure.
5210 				 */
5211 				if (buf_pager_relbuf ||
5212 				    !vm_page_all_valid(m))
5213 					bp->b_flags |= B_RELBUF;
5214 
5215 				bp->b_flags &= ~B_NOCACHE;
5216 				brelse(bp);
5217 			} else {
5218 				bqrelse(bp);
5219 			}
5220 		}
5221 		KASSERT(1 /* racy, enable for debugging */ ||
5222 		    vm_page_all_valid(m) || i == count - 1,
5223 		    ("buf %d %p invalid", i, m));
5224 		if (i == count - 1 && lpart) {
5225 			if (!vm_page_none_valid(m) &&
5226 			    !vm_page_all_valid(m))
5227 				vm_page_zero_invalid(m, TRUE);
5228 		}
5229 next_page:;
5230 	}
5231 end_pages:
5232 
5233 	redo = false;
5234 	for (i = 0; i < count; i++) {
5235 		if (ma[i] == bogus_page)
5236 			continue;
5237 		if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5238 			vm_page_sunbusy(ma[i]);
5239 			ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5240 			    VM_ALLOC_NORMAL);
5241 		}
5242 
5243 		/*
5244 		 * Since the pages were only sbusy while neither the
5245 		 * buffer nor the object lock was held by us, or
5246 		 * reallocated while vm_page_grab() slept for busy
5247 		 * relinguish, they could have been invalidated.
5248 		 * Recheck the valid bits and re-read as needed.
5249 		 *
5250 		 * Note that the last page is made fully valid in the
5251 		 * read loop, and partial validity for the page at
5252 		 * index count - 1 could mean that the page was
5253 		 * invalidated or removed, so we must restart for
5254 		 * safety as well.
5255 		 */
5256 		if (!vm_page_all_valid(ma[i]))
5257 			redo = true;
5258 	}
5259 	if (redo && error == 0)
5260 		goto again;
5261 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5262 }
5263 
5264 #include "opt_ddb.h"
5265 #ifdef DDB
5266 #include <ddb/ddb.h>
5267 
5268 /* DDB command to show buffer data */
5269 DB_SHOW_COMMAND(buffer, db_show_buffer)
5270 {
5271 	/* get args */
5272 	struct buf *bp = (struct buf *)addr;
5273 #ifdef FULL_BUF_TRACKING
5274 	uint32_t i, j;
5275 #endif
5276 
5277 	if (!have_addr) {
5278 		db_printf("usage: show buffer <addr>\n");
5279 		return;
5280 	}
5281 
5282 	db_printf("buf at %p\n", bp);
5283 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5284 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5285 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5286 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5287 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5288 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5289 	db_printf(
5290 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5291 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5292 	    "b_vp = %p, b_dep = %p\n",
5293 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5294 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5295 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5296 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5297 	    bp->b_kvabase, bp->b_kvasize);
5298 	if (bp->b_npages) {
5299 		int i;
5300 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5301 		for (i = 0; i < bp->b_npages; i++) {
5302 			vm_page_t m;
5303 			m = bp->b_pages[i];
5304 			if (m != NULL)
5305 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5306 				    (u_long)m->pindex,
5307 				    (u_long)VM_PAGE_TO_PHYS(m));
5308 			else
5309 				db_printf("( ??? )");
5310 			if ((i + 1) < bp->b_npages)
5311 				db_printf(",");
5312 		}
5313 		db_printf("\n");
5314 	}
5315 	BUF_LOCKPRINTINFO(bp);
5316 #if defined(FULL_BUF_TRACKING)
5317 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5318 
5319 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5320 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5321 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5322 			continue;
5323 		db_printf(" %2u: %s\n", j,
5324 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5325 	}
5326 #elif defined(BUF_TRACKING)
5327 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5328 #endif
5329 	db_printf(" ");
5330 }
5331 
5332 DB_SHOW_COMMAND(bufqueues, bufqueues)
5333 {
5334 	struct bufdomain *bd;
5335 	struct buf *bp;
5336 	long total;
5337 	int i, j, cnt;
5338 
5339 	db_printf("bqempty: %d\n", bqempty.bq_len);
5340 
5341 	for (i = 0; i < buf_domains; i++) {
5342 		bd = &bdomain[i];
5343 		db_printf("Buf domain %d\n", i);
5344 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5345 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5346 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5347 		db_printf("\n");
5348 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5349 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5350 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5351 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5352 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5353 		db_printf("\n");
5354 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5355 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5356 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5357 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5358 		db_printf("\n");
5359 		total = 0;
5360 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5361 			total += bp->b_bufsize;
5362 		db_printf("\tcleanq count\t%d (%ld)\n",
5363 		    bd->bd_cleanq->bq_len, total);
5364 		total = 0;
5365 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5366 			total += bp->b_bufsize;
5367 		db_printf("\tdirtyq count\t%d (%ld)\n",
5368 		    bd->bd_dirtyq.bq_len, total);
5369 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5370 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5371 		db_printf("\tCPU ");
5372 		for (j = 0; j <= mp_maxid; j++)
5373 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5374 		db_printf("\n");
5375 		cnt = 0;
5376 		total = 0;
5377 		for (j = 0; j < nbuf; j++)
5378 			if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5379 				cnt++;
5380 				total += buf[j].b_bufsize;
5381 			}
5382 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5383 		cnt = 0;
5384 		total = 0;
5385 		for (j = 0; j < nbuf; j++)
5386 			if (buf[j].b_domain == i) {
5387 				cnt++;
5388 				total += buf[j].b_bufsize;
5389 			}
5390 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5391 	}
5392 }
5393 
5394 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5395 {
5396 	struct buf *bp;
5397 	int i;
5398 
5399 	for (i = 0; i < nbuf; i++) {
5400 		bp = &buf[i];
5401 		if (BUF_ISLOCKED(bp)) {
5402 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5403 			db_printf("\n");
5404 			if (db_pager_quit)
5405 				break;
5406 		}
5407 	}
5408 }
5409 
5410 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5411 {
5412 	struct vnode *vp;
5413 	struct buf *bp;
5414 
5415 	if (!have_addr) {
5416 		db_printf("usage: show vnodebufs <addr>\n");
5417 		return;
5418 	}
5419 	vp = (struct vnode *)addr;
5420 	db_printf("Clean buffers:\n");
5421 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5422 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5423 		db_printf("\n");
5424 	}
5425 	db_printf("Dirty buffers:\n");
5426 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5427 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5428 		db_printf("\n");
5429 	}
5430 }
5431 
5432 DB_COMMAND(countfreebufs, db_coundfreebufs)
5433 {
5434 	struct buf *bp;
5435 	int i, used = 0, nfree = 0;
5436 
5437 	if (have_addr) {
5438 		db_printf("usage: countfreebufs\n");
5439 		return;
5440 	}
5441 
5442 	for (i = 0; i < nbuf; i++) {
5443 		bp = &buf[i];
5444 		if (bp->b_qindex == QUEUE_EMPTY)
5445 			nfree++;
5446 		else
5447 			used++;
5448 	}
5449 
5450 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5451 	    nfree + used);
5452 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5453 }
5454 #endif /* DDB */
5455