xref: /freebsd/sys/kern/vfs_bio.c (revision 66df505066f51e6d8411b966765d828817f88971)
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/asan.h>
53 #include <sys/bio.h>
54 #include <sys/bitset.h>
55 #include <sys/conf.h>
56 #include <sys/counter.h>
57 #include <sys/buf.h>
58 #include <sys/devicestat.h>
59 #include <sys/eventhandler.h>
60 #include <sys/fail.h>
61 #include <sys/ktr.h>
62 #include <sys/limits.h>
63 #include <sys/lock.h>
64 #include <sys/malloc.h>
65 #include <sys/mount.h>
66 #include <sys/mutex.h>
67 #include <sys/kernel.h>
68 #include <sys/kthread.h>
69 #include <sys/proc.h>
70 #include <sys/racct.h>
71 #include <sys/refcount.h>
72 #include <sys/resourcevar.h>
73 #include <sys/rwlock.h>
74 #include <sys/smp.h>
75 #include <sys/sysctl.h>
76 #include <sys/syscallsubr.h>
77 #include <sys/vmem.h>
78 #include <sys/vmmeter.h>
79 #include <sys/vnode.h>
80 #include <sys/watchdog.h>
81 #include <geom/geom.h>
82 #include <vm/vm.h>
83 #include <vm/vm_param.h>
84 #include <vm/vm_kern.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_pager.h>
89 #include <vm/vm_extern.h>
90 #include <vm/vm_map.h>
91 #include <vm/swap_pager.h>
92 
93 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
94 
95 struct	bio_ops bioops;		/* I/O operation notification */
96 
97 struct	buf_ops buf_ops_bio = {
98 	.bop_name	=	"buf_ops_bio",
99 	.bop_write	=	bufwrite,
100 	.bop_strategy	=	bufstrategy,
101 	.bop_sync	=	bufsync,
102 	.bop_bdflush	=	bufbdflush,
103 };
104 
105 struct bufqueue {
106 	struct mtx_padalign	bq_lock;
107 	TAILQ_HEAD(, buf)	bq_queue;
108 	uint8_t			bq_index;
109 	uint16_t		bq_subqueue;
110 	int			bq_len;
111 } __aligned(CACHE_LINE_SIZE);
112 
113 #define	BQ_LOCKPTR(bq)		(&(bq)->bq_lock)
114 #define	BQ_LOCK(bq)		mtx_lock(BQ_LOCKPTR((bq)))
115 #define	BQ_UNLOCK(bq)		mtx_unlock(BQ_LOCKPTR((bq)))
116 #define	BQ_ASSERT_LOCKED(bq)	mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
117 
118 struct bufdomain {
119 	struct bufqueue	bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
120 	struct bufqueue bd_dirtyq;
121 	struct bufqueue	*bd_cleanq;
122 	struct mtx_padalign bd_run_lock;
123 	/* Constants */
124 	long		bd_maxbufspace;
125 	long		bd_hibufspace;
126 	long 		bd_lobufspace;
127 	long 		bd_bufspacethresh;
128 	int		bd_hifreebuffers;
129 	int		bd_lofreebuffers;
130 	int		bd_hidirtybuffers;
131 	int		bd_lodirtybuffers;
132 	int		bd_dirtybufthresh;
133 	int		bd_lim;
134 	/* atomics */
135 	int		bd_wanted;
136 	int __aligned(CACHE_LINE_SIZE)	bd_numdirtybuffers;
137 	int __aligned(CACHE_LINE_SIZE)	bd_running;
138 	long __aligned(CACHE_LINE_SIZE) bd_bufspace;
139 	int __aligned(CACHE_LINE_SIZE)	bd_freebuffers;
140 } __aligned(CACHE_LINE_SIZE);
141 
142 #define	BD_LOCKPTR(bd)		(&(bd)->bd_cleanq->bq_lock)
143 #define	BD_LOCK(bd)		mtx_lock(BD_LOCKPTR((bd)))
144 #define	BD_UNLOCK(bd)		mtx_unlock(BD_LOCKPTR((bd)))
145 #define	BD_ASSERT_LOCKED(bd)	mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
146 #define	BD_RUN_LOCKPTR(bd)	(&(bd)->bd_run_lock)
147 #define	BD_RUN_LOCK(bd)		mtx_lock(BD_RUN_LOCKPTR((bd)))
148 #define	BD_RUN_UNLOCK(bd)	mtx_unlock(BD_RUN_LOCKPTR((bd)))
149 #define	BD_DOMAIN(bd)		(bd - bdomain)
150 
151 static char *buf;		/* buffer header pool */
152 static struct buf *
153 nbufp(unsigned i)
154 {
155 	return ((struct buf *)(buf + (sizeof(struct buf) +
156 	    sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
157 }
158 
159 caddr_t __read_mostly unmapped_buf;
160 
161 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
162 struct proc *bufdaemonproc;
163 
164 static void vm_hold_free_pages(struct buf *bp, int newbsize);
165 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
166 		vm_offset_t to);
167 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
168 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
169 		vm_page_t m);
170 static void vfs_clean_pages_dirty_buf(struct buf *bp);
171 static void vfs_setdirty_range(struct buf *bp);
172 static void vfs_vmio_invalidate(struct buf *bp);
173 static void vfs_vmio_truncate(struct buf *bp, int npages);
174 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
175 static int vfs_bio_clcheck(struct vnode *vp, int size,
176 		daddr_t lblkno, daddr_t blkno);
177 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
178 		void (*)(struct buf *));
179 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
180 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
181 static void buf_daemon(void);
182 static __inline void bd_wakeup(void);
183 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
184 static void bufkva_reclaim(vmem_t *, int);
185 static void bufkva_free(struct buf *);
186 static int buf_import(void *, void **, int, int, int);
187 static void buf_release(void *, void **, int);
188 static void maxbcachebuf_adjust(void);
189 static inline struct bufdomain *bufdomain(struct buf *);
190 static void bq_remove(struct bufqueue *bq, struct buf *bp);
191 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
192 static int buf_recycle(struct bufdomain *, bool kva);
193 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
194 	    const char *lockname);
195 static void bd_init(struct bufdomain *bd);
196 static int bd_flushall(struct bufdomain *bd);
197 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
198 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
199 
200 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
201 int vmiodirenable = TRUE;
202 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
203     "Use the VM system for directory writes");
204 long runningbufspace;
205 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
206     "Amount of presently outstanding async buffer io");
207 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
208     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
209 static counter_u64_t bufkvaspace;
210 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
211     "Kernel virtual memory used for buffers");
212 static long maxbufspace;
213 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
214     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
215     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
216     "Maximum allowed value of bufspace (including metadata)");
217 static long bufmallocspace;
218 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
219     "Amount of malloced memory for buffers");
220 static long maxbufmallocspace;
221 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
222     0, "Maximum amount of malloced memory for buffers");
223 static long lobufspace;
224 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
225     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
226     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
227     "Minimum amount of buffers we want to have");
228 long hibufspace;
229 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
230     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
231     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
232     "Maximum allowed value of bufspace (excluding metadata)");
233 long bufspacethresh;
234 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
235     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
236     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
237     "Bufspace consumed before waking the daemon to free some");
238 static counter_u64_t buffreekvacnt;
239 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
240     "Number of times we have freed the KVA space from some buffer");
241 static counter_u64_t bufdefragcnt;
242 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
243     "Number of times we have had to repeat buffer allocation to defragment");
244 static long lorunningspace;
245 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
246     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
247     "Minimum preferred space used for in-progress I/O");
248 static long hirunningspace;
249 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
250     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
251     "Maximum amount of space to use for in-progress I/O");
252 int dirtybufferflushes;
253 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
254     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
255 int bdwriteskip;
256 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
257     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
258 int altbufferflushes;
259 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
260     &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
261 static int recursiveflushes;
262 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
263     &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
264 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
265 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
266     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
267     "Number of buffers that are dirty (has unwritten changes) at the moment");
268 static int lodirtybuffers;
269 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
270     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
271     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
272     "How many buffers we want to have free before bufdaemon can sleep");
273 static int hidirtybuffers;
274 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
275     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
276     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
277     "When the number of dirty buffers is considered severe");
278 int dirtybufthresh;
279 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
280     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
281     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
282     "Number of bdwrite to bawrite conversions to clear dirty buffers");
283 static int numfreebuffers;
284 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
285     "Number of free buffers");
286 static int lofreebuffers;
287 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
288     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
289     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
290    "Target number of free buffers");
291 static int hifreebuffers;
292 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
293     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
294     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
295    "Threshold for clean buffer recycling");
296 static counter_u64_t getnewbufcalls;
297 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
298    &getnewbufcalls, "Number of calls to getnewbuf");
299 static counter_u64_t getnewbufrestarts;
300 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
301     &getnewbufrestarts,
302     "Number of times getnewbuf has had to restart a buffer acquisition");
303 static counter_u64_t mappingrestarts;
304 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
305     &mappingrestarts,
306     "Number of times getblk has had to restart a buffer mapping for "
307     "unmapped buffer");
308 static counter_u64_t numbufallocfails;
309 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
310     &numbufallocfails, "Number of times buffer allocations failed");
311 static int flushbufqtarget = 100;
312 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
313     "Amount of work to do in flushbufqueues when helping bufdaemon");
314 static counter_u64_t notbufdflushes;
315 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
316     "Number of dirty buffer flushes done by the bufdaemon helpers");
317 static long barrierwrites;
318 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
319     &barrierwrites, 0, "Number of barrier writes");
320 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
321     &unmapped_buf_allowed, 0,
322     "Permit the use of the unmapped i/o");
323 int maxbcachebuf = MAXBCACHEBUF;
324 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
325     "Maximum size of a buffer cache block");
326 
327 /*
328  * This lock synchronizes access to bd_request.
329  */
330 static struct mtx_padalign __exclusive_cache_line bdlock;
331 
332 /*
333  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
334  * waitrunningbufspace().
335  */
336 static struct mtx_padalign __exclusive_cache_line rbreqlock;
337 
338 /*
339  * Lock that protects bdirtywait.
340  */
341 static struct mtx_padalign __exclusive_cache_line bdirtylock;
342 
343 /*
344  * Wakeup point for bufdaemon, as well as indicator of whether it is already
345  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
346  * is idling.
347  */
348 static int bd_request;
349 
350 /*
351  * Request for the buf daemon to write more buffers than is indicated by
352  * lodirtybuf.  This may be necessary to push out excess dependencies or
353  * defragment the address space where a simple count of the number of dirty
354  * buffers is insufficient to characterize the demand for flushing them.
355  */
356 static int bd_speedupreq;
357 
358 /*
359  * Synchronization (sleep/wakeup) variable for active buffer space requests.
360  * Set when wait starts, cleared prior to wakeup().
361  * Used in runningbufwakeup() and waitrunningbufspace().
362  */
363 static int runningbufreq;
364 
365 /*
366  * Synchronization for bwillwrite() waiters.
367  */
368 static int bdirtywait;
369 
370 /*
371  * Definitions for the buffer free lists.
372  */
373 #define QUEUE_NONE	0	/* on no queue */
374 #define QUEUE_EMPTY	1	/* empty buffer headers */
375 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
376 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
377 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
378 
379 /* Maximum number of buffer domains. */
380 #define	BUF_DOMAINS	8
381 
382 struct bufdomainset bdlodirty;		/* Domains > lodirty */
383 struct bufdomainset bdhidirty;		/* Domains > hidirty */
384 
385 /* Configured number of clean queues. */
386 static int __read_mostly buf_domains;
387 
388 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
389 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
390 struct bufqueue __exclusive_cache_line bqempty;
391 
392 /*
393  * per-cpu empty buffer cache.
394  */
395 uma_zone_t buf_zone;
396 
397 /*
398  * Single global constant for BUF_WMESG, to avoid getting multiple references.
399  * buf_wmesg is referred from macros.
400  */
401 const char *buf_wmesg = BUF_WMESG;
402 
403 static int
404 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
405 {
406 	long value;
407 	int error;
408 
409 	value = *(long *)arg1;
410 	error = sysctl_handle_long(oidp, &value, 0, req);
411 	if (error != 0 || req->newptr == NULL)
412 		return (error);
413 	mtx_lock(&rbreqlock);
414 	if (arg1 == &hirunningspace) {
415 		if (value < lorunningspace)
416 			error = EINVAL;
417 		else
418 			hirunningspace = value;
419 	} else {
420 		KASSERT(arg1 == &lorunningspace,
421 		    ("%s: unknown arg1", __func__));
422 		if (value > hirunningspace)
423 			error = EINVAL;
424 		else
425 			lorunningspace = value;
426 	}
427 	mtx_unlock(&rbreqlock);
428 	return (error);
429 }
430 
431 static int
432 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
433 {
434 	int error;
435 	int value;
436 	int i;
437 
438 	value = *(int *)arg1;
439 	error = sysctl_handle_int(oidp, &value, 0, req);
440 	if (error != 0 || req->newptr == NULL)
441 		return (error);
442 	*(int *)arg1 = value;
443 	for (i = 0; i < buf_domains; i++)
444 		*(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
445 		    value / buf_domains;
446 
447 	return (error);
448 }
449 
450 static int
451 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
452 {
453 	long value;
454 	int error;
455 	int i;
456 
457 	value = *(long *)arg1;
458 	error = sysctl_handle_long(oidp, &value, 0, req);
459 	if (error != 0 || req->newptr == NULL)
460 		return (error);
461 	*(long *)arg1 = value;
462 	for (i = 0; i < buf_domains; i++)
463 		*(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
464 		    value / buf_domains;
465 
466 	return (error);
467 }
468 
469 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
470     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
471 static int
472 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
473 {
474 	long lvalue;
475 	int ivalue;
476 	int i;
477 
478 	lvalue = 0;
479 	for (i = 0; i < buf_domains; i++)
480 		lvalue += bdomain[i].bd_bufspace;
481 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
482 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
483 	if (lvalue > INT_MAX)
484 		/* On overflow, still write out a long to trigger ENOMEM. */
485 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
486 	ivalue = lvalue;
487 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
488 }
489 #else
490 static int
491 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
492 {
493 	long lvalue;
494 	int i;
495 
496 	lvalue = 0;
497 	for (i = 0; i < buf_domains; i++)
498 		lvalue += bdomain[i].bd_bufspace;
499 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
500 }
501 #endif
502 
503 static int
504 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
505 {
506 	int value;
507 	int i;
508 
509 	value = 0;
510 	for (i = 0; i < buf_domains; i++)
511 		value += bdomain[i].bd_numdirtybuffers;
512 	return (sysctl_handle_int(oidp, &value, 0, req));
513 }
514 
515 /*
516  *	bdirtywakeup:
517  *
518  *	Wakeup any bwillwrite() waiters.
519  */
520 static void
521 bdirtywakeup(void)
522 {
523 	mtx_lock(&bdirtylock);
524 	if (bdirtywait) {
525 		bdirtywait = 0;
526 		wakeup(&bdirtywait);
527 	}
528 	mtx_unlock(&bdirtylock);
529 }
530 
531 /*
532  *	bd_clear:
533  *
534  *	Clear a domain from the appropriate bitsets when dirtybuffers
535  *	is decremented.
536  */
537 static void
538 bd_clear(struct bufdomain *bd)
539 {
540 
541 	mtx_lock(&bdirtylock);
542 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
543 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
544 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
545 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
546 	mtx_unlock(&bdirtylock);
547 }
548 
549 /*
550  *	bd_set:
551  *
552  *	Set a domain in the appropriate bitsets when dirtybuffers
553  *	is incremented.
554  */
555 static void
556 bd_set(struct bufdomain *bd)
557 {
558 
559 	mtx_lock(&bdirtylock);
560 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
561 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
562 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
563 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
564 	mtx_unlock(&bdirtylock);
565 }
566 
567 /*
568  *	bdirtysub:
569  *
570  *	Decrement the numdirtybuffers count by one and wakeup any
571  *	threads blocked in bwillwrite().
572  */
573 static void
574 bdirtysub(struct buf *bp)
575 {
576 	struct bufdomain *bd;
577 	int num;
578 
579 	bd = bufdomain(bp);
580 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
581 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
582 		bdirtywakeup();
583 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
584 		bd_clear(bd);
585 }
586 
587 /*
588  *	bdirtyadd:
589  *
590  *	Increment the numdirtybuffers count by one and wakeup the buf
591  *	daemon if needed.
592  */
593 static void
594 bdirtyadd(struct buf *bp)
595 {
596 	struct bufdomain *bd;
597 	int num;
598 
599 	/*
600 	 * Only do the wakeup once as we cross the boundary.  The
601 	 * buf daemon will keep running until the condition clears.
602 	 */
603 	bd = bufdomain(bp);
604 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
605 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
606 		bd_wakeup();
607 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
608 		bd_set(bd);
609 }
610 
611 /*
612  *	bufspace_daemon_wakeup:
613  *
614  *	Wakeup the daemons responsible for freeing clean bufs.
615  */
616 static void
617 bufspace_daemon_wakeup(struct bufdomain *bd)
618 {
619 
620 	/*
621 	 * avoid the lock if the daemon is running.
622 	 */
623 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
624 		BD_RUN_LOCK(bd);
625 		atomic_store_int(&bd->bd_running, 1);
626 		wakeup(&bd->bd_running);
627 		BD_RUN_UNLOCK(bd);
628 	}
629 }
630 
631 /*
632  *	bufspace_daemon_wait:
633  *
634  *	Sleep until the domain falls below a limit or one second passes.
635  */
636 static void
637 bufspace_daemon_wait(struct bufdomain *bd)
638 {
639 	/*
640 	 * Re-check our limits and sleep.  bd_running must be
641 	 * cleared prior to checking the limits to avoid missed
642 	 * wakeups.  The waker will adjust one of bufspace or
643 	 * freebuffers prior to checking bd_running.
644 	 */
645 	BD_RUN_LOCK(bd);
646 	atomic_store_int(&bd->bd_running, 0);
647 	if (bd->bd_bufspace < bd->bd_bufspacethresh &&
648 	    bd->bd_freebuffers > bd->bd_lofreebuffers) {
649 		msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
650 		    "-", hz);
651 	} else {
652 		/* Avoid spurious wakeups while running. */
653 		atomic_store_int(&bd->bd_running, 1);
654 		BD_RUN_UNLOCK(bd);
655 	}
656 }
657 
658 /*
659  *	bufspace_adjust:
660  *
661  *	Adjust the reported bufspace for a KVA managed buffer, possibly
662  * 	waking any waiters.
663  */
664 static void
665 bufspace_adjust(struct buf *bp, int bufsize)
666 {
667 	struct bufdomain *bd;
668 	long space;
669 	int diff;
670 
671 	KASSERT((bp->b_flags & B_MALLOC) == 0,
672 	    ("bufspace_adjust: malloc buf %p", bp));
673 	bd = bufdomain(bp);
674 	diff = bufsize - bp->b_bufsize;
675 	if (diff < 0) {
676 		atomic_subtract_long(&bd->bd_bufspace, -diff);
677 	} else if (diff > 0) {
678 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
679 		/* Wake up the daemon on the transition. */
680 		if (space < bd->bd_bufspacethresh &&
681 		    space + diff >= bd->bd_bufspacethresh)
682 			bufspace_daemon_wakeup(bd);
683 	}
684 	bp->b_bufsize = bufsize;
685 }
686 
687 /*
688  *	bufspace_reserve:
689  *
690  *	Reserve bufspace before calling allocbuf().  metadata has a
691  *	different space limit than data.
692  */
693 static int
694 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
695 {
696 	long limit, new;
697 	long space;
698 
699 	if (metadata)
700 		limit = bd->bd_maxbufspace;
701 	else
702 		limit = bd->bd_hibufspace;
703 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
704 	new = space + size;
705 	if (new > limit) {
706 		atomic_subtract_long(&bd->bd_bufspace, size);
707 		return (ENOSPC);
708 	}
709 
710 	/* Wake up the daemon on the transition. */
711 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
712 		bufspace_daemon_wakeup(bd);
713 
714 	return (0);
715 }
716 
717 /*
718  *	bufspace_release:
719  *
720  *	Release reserved bufspace after bufspace_adjust() has consumed it.
721  */
722 static void
723 bufspace_release(struct bufdomain *bd, int size)
724 {
725 
726 	atomic_subtract_long(&bd->bd_bufspace, size);
727 }
728 
729 /*
730  *	bufspace_wait:
731  *
732  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
733  *	supplied.  bd_wanted must be set prior to polling for space.  The
734  *	operation must be re-tried on return.
735  */
736 static void
737 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
738     int slpflag, int slptimeo)
739 {
740 	struct thread *td;
741 	int error, fl, norunbuf;
742 
743 	if ((gbflags & GB_NOWAIT_BD) != 0)
744 		return;
745 
746 	td = curthread;
747 	BD_LOCK(bd);
748 	while (bd->bd_wanted) {
749 		if (vp != NULL && vp->v_type != VCHR &&
750 		    (td->td_pflags & TDP_BUFNEED) == 0) {
751 			BD_UNLOCK(bd);
752 			/*
753 			 * getblk() is called with a vnode locked, and
754 			 * some majority of the dirty buffers may as
755 			 * well belong to the vnode.  Flushing the
756 			 * buffers there would make a progress that
757 			 * cannot be achieved by the buf_daemon, that
758 			 * cannot lock the vnode.
759 			 */
760 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
761 			    (td->td_pflags & TDP_NORUNNINGBUF);
762 
763 			/*
764 			 * Play bufdaemon.  The getnewbuf() function
765 			 * may be called while the thread owns lock
766 			 * for another dirty buffer for the same
767 			 * vnode, which makes it impossible to use
768 			 * VOP_FSYNC() there, due to the buffer lock
769 			 * recursion.
770 			 */
771 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
772 			fl = buf_flush(vp, bd, flushbufqtarget);
773 			td->td_pflags &= norunbuf;
774 			BD_LOCK(bd);
775 			if (fl != 0)
776 				continue;
777 			if (bd->bd_wanted == 0)
778 				break;
779 		}
780 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
781 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
782 		if (error != 0)
783 			break;
784 	}
785 	BD_UNLOCK(bd);
786 }
787 
788 /*
789  *	bufspace_daemon:
790  *
791  *	buffer space management daemon.  Tries to maintain some marginal
792  *	amount of free buffer space so that requesting processes neither
793  *	block nor work to reclaim buffers.
794  */
795 static void
796 bufspace_daemon(void *arg)
797 {
798 	struct bufdomain *bd;
799 
800 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
801 	    SHUTDOWN_PRI_LAST + 100);
802 
803 	bd = arg;
804 	for (;;) {
805 		kthread_suspend_check();
806 
807 		/*
808 		 * Free buffers from the clean queue until we meet our
809 		 * targets.
810 		 *
811 		 * Theory of operation:  The buffer cache is most efficient
812 		 * when some free buffer headers and space are always
813 		 * available to getnewbuf().  This daemon attempts to prevent
814 		 * the excessive blocking and synchronization associated
815 		 * with shortfall.  It goes through three phases according
816 		 * demand:
817 		 *
818 		 * 1)	The daemon wakes up voluntarily once per-second
819 		 *	during idle periods when the counters are below
820 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
821 		 *
822 		 * 2)	The daemon wakes up as we cross the thresholds
823 		 *	ahead of any potential blocking.  This may bounce
824 		 *	slightly according to the rate of consumption and
825 		 *	release.
826 		 *
827 		 * 3)	The daemon and consumers are starved for working
828 		 *	clean buffers.  This is the 'bufspace' sleep below
829 		 *	which will inefficiently trade bufs with bqrelse
830 		 *	until we return to condition 2.
831 		 */
832 		while (bd->bd_bufspace > bd->bd_lobufspace ||
833 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
834 			if (buf_recycle(bd, false) != 0) {
835 				if (bd_flushall(bd))
836 					continue;
837 				/*
838 				 * Speedup dirty if we've run out of clean
839 				 * buffers.  This is possible in particular
840 				 * because softdep may held many bufs locked
841 				 * pending writes to other bufs which are
842 				 * marked for delayed write, exhausting
843 				 * clean space until they are written.
844 				 */
845 				bd_speedup();
846 				BD_LOCK(bd);
847 				if (bd->bd_wanted) {
848 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
849 					    PRIBIO|PDROP, "bufspace", hz/10);
850 				} else
851 					BD_UNLOCK(bd);
852 			}
853 			maybe_yield();
854 		}
855 		bufspace_daemon_wait(bd);
856 	}
857 }
858 
859 /*
860  *	bufmallocadjust:
861  *
862  *	Adjust the reported bufspace for a malloc managed buffer, possibly
863  *	waking any waiters.
864  */
865 static void
866 bufmallocadjust(struct buf *bp, int bufsize)
867 {
868 	int diff;
869 
870 	KASSERT((bp->b_flags & B_MALLOC) != 0,
871 	    ("bufmallocadjust: non-malloc buf %p", bp));
872 	diff = bufsize - bp->b_bufsize;
873 	if (diff < 0)
874 		atomic_subtract_long(&bufmallocspace, -diff);
875 	else
876 		atomic_add_long(&bufmallocspace, diff);
877 	bp->b_bufsize = bufsize;
878 }
879 
880 /*
881  *	runningwakeup:
882  *
883  *	Wake up processes that are waiting on asynchronous writes to fall
884  *	below lorunningspace.
885  */
886 static void
887 runningwakeup(void)
888 {
889 
890 	mtx_lock(&rbreqlock);
891 	if (runningbufreq) {
892 		runningbufreq = 0;
893 		wakeup(&runningbufreq);
894 	}
895 	mtx_unlock(&rbreqlock);
896 }
897 
898 /*
899  *	runningbufwakeup:
900  *
901  *	Decrement the outstanding write count according.
902  */
903 void
904 runningbufwakeup(struct buf *bp)
905 {
906 	long space, bspace;
907 
908 	bspace = bp->b_runningbufspace;
909 	if (bspace == 0)
910 		return;
911 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
912 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
913 	    space, bspace));
914 	bp->b_runningbufspace = 0;
915 	/*
916 	 * Only acquire the lock and wakeup on the transition from exceeding
917 	 * the threshold to falling below it.
918 	 */
919 	if (space < lorunningspace)
920 		return;
921 	if (space - bspace > lorunningspace)
922 		return;
923 	runningwakeup();
924 }
925 
926 /*
927  *	waitrunningbufspace()
928  *
929  *	runningbufspace is a measure of the amount of I/O currently
930  *	running.  This routine is used in async-write situations to
931  *	prevent creating huge backups of pending writes to a device.
932  *	Only asynchronous writes are governed by this function.
933  *
934  *	This does NOT turn an async write into a sync write.  It waits
935  *	for earlier writes to complete and generally returns before the
936  *	caller's write has reached the device.
937  */
938 void
939 waitrunningbufspace(void)
940 {
941 
942 	mtx_lock(&rbreqlock);
943 	while (runningbufspace > hirunningspace) {
944 		runningbufreq = 1;
945 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
946 	}
947 	mtx_unlock(&rbreqlock);
948 }
949 
950 /*
951  *	vfs_buf_test_cache:
952  *
953  *	Called when a buffer is extended.  This function clears the B_CACHE
954  *	bit if the newly extended portion of the buffer does not contain
955  *	valid data.
956  */
957 static __inline void
958 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
959     vm_offset_t size, vm_page_t m)
960 {
961 
962 	/*
963 	 * This function and its results are protected by higher level
964 	 * synchronization requiring vnode and buf locks to page in and
965 	 * validate pages.
966 	 */
967 	if (bp->b_flags & B_CACHE) {
968 		int base = (foff + off) & PAGE_MASK;
969 		if (vm_page_is_valid(m, base, size) == 0)
970 			bp->b_flags &= ~B_CACHE;
971 	}
972 }
973 
974 /* Wake up the buffer daemon if necessary */
975 static void
976 bd_wakeup(void)
977 {
978 
979 	mtx_lock(&bdlock);
980 	if (bd_request == 0) {
981 		bd_request = 1;
982 		wakeup(&bd_request);
983 	}
984 	mtx_unlock(&bdlock);
985 }
986 
987 /*
988  * Adjust the maxbcachbuf tunable.
989  */
990 static void
991 maxbcachebuf_adjust(void)
992 {
993 	int i;
994 
995 	/*
996 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
997 	 */
998 	i = 2;
999 	while (i * 2 <= maxbcachebuf)
1000 		i *= 2;
1001 	maxbcachebuf = i;
1002 	if (maxbcachebuf < MAXBSIZE)
1003 		maxbcachebuf = MAXBSIZE;
1004 	if (maxbcachebuf > maxphys)
1005 		maxbcachebuf = maxphys;
1006 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1007 		printf("maxbcachebuf=%d\n", maxbcachebuf);
1008 }
1009 
1010 /*
1011  * bd_speedup - speedup the buffer cache flushing code
1012  */
1013 void
1014 bd_speedup(void)
1015 {
1016 	int needwake;
1017 
1018 	mtx_lock(&bdlock);
1019 	needwake = 0;
1020 	if (bd_speedupreq == 0 || bd_request == 0)
1021 		needwake = 1;
1022 	bd_speedupreq = 1;
1023 	bd_request = 1;
1024 	if (needwake)
1025 		wakeup(&bd_request);
1026 	mtx_unlock(&bdlock);
1027 }
1028 
1029 #ifdef __i386__
1030 #define	TRANSIENT_DENOM	5
1031 #else
1032 #define	TRANSIENT_DENOM 10
1033 #endif
1034 
1035 /*
1036  * Calculating buffer cache scaling values and reserve space for buffer
1037  * headers.  This is called during low level kernel initialization and
1038  * may be called more then once.  We CANNOT write to the memory area
1039  * being reserved at this time.
1040  */
1041 caddr_t
1042 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1043 {
1044 	int tuned_nbuf;
1045 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1046 
1047 	/*
1048 	 * With KASAN or KMSAN enabled, the kernel map is shadowed.  Account for
1049 	 * this when sizing maps based on the amount of physical memory
1050 	 * available.
1051 	 */
1052 #if defined(KASAN)
1053 	physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1054 	    (KASAN_SHADOW_SCALE + 1);
1055 #elif defined(KMSAN)
1056 	physmem_est /= 3;
1057 
1058 	/*
1059 	 * KMSAN cannot reliably determine whether buffer data is initialized
1060 	 * unless it is updated through a KVA mapping.
1061 	 */
1062 	unmapped_buf_allowed = 0;
1063 #endif
1064 
1065 	/*
1066 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1067 	 * PAGE_SIZE is >= 1K)
1068 	 */
1069 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1070 
1071 	maxbcachebuf_adjust();
1072 	/*
1073 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1074 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1075 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1076 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1077 	 * the buffer cache we limit the eventual kva reservation to
1078 	 * maxbcache bytes.
1079 	 *
1080 	 * factor represents the 1/4 x ram conversion.
1081 	 */
1082 	if (nbuf == 0) {
1083 		int factor = 4 * BKVASIZE / 1024;
1084 
1085 		nbuf = 50;
1086 		if (physmem_est > 4096)
1087 			nbuf += min((physmem_est - 4096) / factor,
1088 			    65536 / factor);
1089 		if (physmem_est > 65536)
1090 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1091 			    32 * 1024 * 1024 / (factor * 5));
1092 
1093 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1094 			nbuf = maxbcache / BKVASIZE;
1095 		tuned_nbuf = 1;
1096 	} else
1097 		tuned_nbuf = 0;
1098 
1099 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1100 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1101 	if (nbuf > maxbuf) {
1102 		if (!tuned_nbuf)
1103 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1104 			    maxbuf);
1105 		nbuf = maxbuf;
1106 	}
1107 
1108 	/*
1109 	 * Ideal allocation size for the transient bio submap is 10%
1110 	 * of the maximal space buffer map.  This roughly corresponds
1111 	 * to the amount of the buffer mapped for typical UFS load.
1112 	 *
1113 	 * Clip the buffer map to reserve space for the transient
1114 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1115 	 * maximum buffer map extent on the platform.
1116 	 *
1117 	 * The fall-back to the maxbuf in case of maxbcache unset,
1118 	 * allows to not trim the buffer KVA for the architectures
1119 	 * with ample KVA space.
1120 	 */
1121 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1122 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1123 		buf_sz = (long)nbuf * BKVASIZE;
1124 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1125 		    (TRANSIENT_DENOM - 1)) {
1126 			/*
1127 			 * There is more KVA than memory.  Do not
1128 			 * adjust buffer map size, and assign the rest
1129 			 * of maxbuf to transient map.
1130 			 */
1131 			biotmap_sz = maxbuf_sz - buf_sz;
1132 		} else {
1133 			/*
1134 			 * Buffer map spans all KVA we could afford on
1135 			 * this platform.  Give 10% (20% on i386) of
1136 			 * the buffer map to the transient bio map.
1137 			 */
1138 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1139 			buf_sz -= biotmap_sz;
1140 		}
1141 		if (biotmap_sz / INT_MAX > maxphys)
1142 			bio_transient_maxcnt = INT_MAX;
1143 		else
1144 			bio_transient_maxcnt = biotmap_sz / maxphys;
1145 		/*
1146 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1147 		 * using the transient mapping.
1148 		 */
1149 		if (bio_transient_maxcnt > 1024)
1150 			bio_transient_maxcnt = 1024;
1151 		if (tuned_nbuf)
1152 			nbuf = buf_sz / BKVASIZE;
1153 	}
1154 
1155 	if (nswbuf == 0) {
1156 		nswbuf = min(nbuf / 4, 256);
1157 		if (nswbuf < NSWBUF_MIN)
1158 			nswbuf = NSWBUF_MIN;
1159 	}
1160 
1161 	/*
1162 	 * Reserve space for the buffer cache buffers
1163 	 */
1164 	buf = (char *)v;
1165 	v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1166 	    atop(maxbcachebuf)) * nbuf;
1167 
1168 	return (v);
1169 }
1170 
1171 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1172 void
1173 bufinit(void)
1174 {
1175 	struct buf *bp;
1176 	int i;
1177 
1178 	KASSERT(maxbcachebuf >= MAXBSIZE,
1179 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1180 	    MAXBSIZE));
1181 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1182 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1183 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1184 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1185 
1186 	unmapped_buf = (caddr_t)kva_alloc(maxphys);
1187 
1188 	/* finally, initialize each buffer header and stick on empty q */
1189 	for (i = 0; i < nbuf; i++) {
1190 		bp = nbufp(i);
1191 		bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1192 		bp->b_flags = B_INVAL;
1193 		bp->b_rcred = NOCRED;
1194 		bp->b_wcred = NOCRED;
1195 		bp->b_qindex = QUEUE_NONE;
1196 		bp->b_domain = -1;
1197 		bp->b_subqueue = mp_maxid + 1;
1198 		bp->b_xflags = 0;
1199 		bp->b_data = bp->b_kvabase = unmapped_buf;
1200 		LIST_INIT(&bp->b_dep);
1201 		BUF_LOCKINIT(bp);
1202 		bq_insert(&bqempty, bp, false);
1203 	}
1204 
1205 	/*
1206 	 * maxbufspace is the absolute maximum amount of buffer space we are
1207 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1208 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1209 	 * used by most other requests.  The differential is required to
1210 	 * ensure that metadata deadlocks don't occur.
1211 	 *
1212 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1213 	 * this may result in KVM fragmentation which is not handled optimally
1214 	 * by the system. XXX This is less true with vmem.  We could use
1215 	 * PAGE_SIZE.
1216 	 */
1217 	maxbufspace = (long)nbuf * BKVASIZE;
1218 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1219 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1220 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1221 
1222 	/*
1223 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1224 	 * arbitrarily and may need further tuning. It corresponds to
1225 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1226 	 * which fits with many RAID controllers' tagged queuing limits.
1227 	 * The lower 1 MiB limit is the historical upper limit for
1228 	 * hirunningspace.
1229 	 */
1230 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1231 	    16 * 1024 * 1024), 1024 * 1024);
1232 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1233 
1234 	/*
1235 	 * Limit the amount of malloc memory since it is wired permanently into
1236 	 * the kernel space.  Even though this is accounted for in the buffer
1237 	 * allocation, we don't want the malloced region to grow uncontrolled.
1238 	 * The malloc scheme improves memory utilization significantly on
1239 	 * average (small) directories.
1240 	 */
1241 	maxbufmallocspace = hibufspace / 20;
1242 
1243 	/*
1244 	 * Reduce the chance of a deadlock occurring by limiting the number
1245 	 * of delayed-write dirty buffers we allow to stack up.
1246 	 */
1247 	hidirtybuffers = nbuf / 4 + 20;
1248 	dirtybufthresh = hidirtybuffers * 9 / 10;
1249 	/*
1250 	 * To support extreme low-memory systems, make sure hidirtybuffers
1251 	 * cannot eat up all available buffer space.  This occurs when our
1252 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1253 	 * buffer space assuming BKVASIZE'd buffers.
1254 	 */
1255 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1256 		hidirtybuffers >>= 1;
1257 	}
1258 	lodirtybuffers = hidirtybuffers / 2;
1259 
1260 	/*
1261 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1262 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1263 	 * are counted as free but will be unavailable to threads executing
1264 	 * on other cpus.
1265 	 *
1266 	 * hifreebuffers is the free target for the bufspace daemon.  This
1267 	 * should be set appropriately to limit work per-iteration.
1268 	 */
1269 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1270 	hifreebuffers = (3 * lofreebuffers) / 2;
1271 	numfreebuffers = nbuf;
1272 
1273 	/* Setup the kva and free list allocators. */
1274 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1275 	buf_zone = uma_zcache_create("buf free cache",
1276 	    sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1277 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1278 
1279 	/*
1280 	 * Size the clean queue according to the amount of buffer space.
1281 	 * One queue per-256mb up to the max.  More queues gives better
1282 	 * concurrency but less accurate LRU.
1283 	 */
1284 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1285 	for (i = 0 ; i < buf_domains; i++) {
1286 		struct bufdomain *bd;
1287 
1288 		bd = &bdomain[i];
1289 		bd_init(bd);
1290 		bd->bd_freebuffers = nbuf / buf_domains;
1291 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1292 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1293 		bd->bd_bufspace = 0;
1294 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1295 		bd->bd_hibufspace = hibufspace / buf_domains;
1296 		bd->bd_lobufspace = lobufspace / buf_domains;
1297 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1298 		bd->bd_numdirtybuffers = 0;
1299 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1300 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1301 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1302 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1303 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1304 	}
1305 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1306 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1307 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1308 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1309 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1310 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1311 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1312 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1313 }
1314 
1315 #ifdef INVARIANTS
1316 static inline void
1317 vfs_buf_check_mapped(struct buf *bp)
1318 {
1319 
1320 	KASSERT(bp->b_kvabase != unmapped_buf,
1321 	    ("mapped buf: b_kvabase was not updated %p", bp));
1322 	KASSERT(bp->b_data != unmapped_buf,
1323 	    ("mapped buf: b_data was not updated %p", bp));
1324 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1325 	    maxphys, ("b_data + b_offset unmapped %p", bp));
1326 }
1327 
1328 static inline void
1329 vfs_buf_check_unmapped(struct buf *bp)
1330 {
1331 
1332 	KASSERT(bp->b_data == unmapped_buf,
1333 	    ("unmapped buf: corrupted b_data %p", bp));
1334 }
1335 
1336 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1337 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1338 #else
1339 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1340 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1341 #endif
1342 
1343 static int
1344 isbufbusy(struct buf *bp)
1345 {
1346 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1347 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1348 		return (1);
1349 	return (0);
1350 }
1351 
1352 /*
1353  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1354  */
1355 void
1356 bufshutdown(int show_busybufs)
1357 {
1358 	static int first_buf_printf = 1;
1359 	struct buf *bp;
1360 	int i, iter, nbusy, pbusy;
1361 #ifndef PREEMPTION
1362 	int subiter;
1363 #endif
1364 
1365 	/*
1366 	 * Sync filesystems for shutdown
1367 	 */
1368 	wdog_kern_pat(WD_LASTVAL);
1369 	kern_sync(curthread);
1370 
1371 	/*
1372 	 * With soft updates, some buffers that are
1373 	 * written will be remarked as dirty until other
1374 	 * buffers are written.
1375 	 */
1376 	for (iter = pbusy = 0; iter < 20; iter++) {
1377 		nbusy = 0;
1378 		for (i = nbuf - 1; i >= 0; i--) {
1379 			bp = nbufp(i);
1380 			if (isbufbusy(bp))
1381 				nbusy++;
1382 		}
1383 		if (nbusy == 0) {
1384 			if (first_buf_printf)
1385 				printf("All buffers synced.");
1386 			break;
1387 		}
1388 		if (first_buf_printf) {
1389 			printf("Syncing disks, buffers remaining... ");
1390 			first_buf_printf = 0;
1391 		}
1392 		printf("%d ", nbusy);
1393 		if (nbusy < pbusy)
1394 			iter = 0;
1395 		pbusy = nbusy;
1396 
1397 		wdog_kern_pat(WD_LASTVAL);
1398 		kern_sync(curthread);
1399 
1400 #ifdef PREEMPTION
1401 		/*
1402 		 * Spin for a while to allow interrupt threads to run.
1403 		 */
1404 		DELAY(50000 * iter);
1405 #else
1406 		/*
1407 		 * Context switch several times to allow interrupt
1408 		 * threads to run.
1409 		 */
1410 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1411 			thread_lock(curthread);
1412 			mi_switch(SW_VOL);
1413 			DELAY(1000);
1414 		}
1415 #endif
1416 	}
1417 	printf("\n");
1418 	/*
1419 	 * Count only busy local buffers to prevent forcing
1420 	 * a fsck if we're just a client of a wedged NFS server
1421 	 */
1422 	nbusy = 0;
1423 	for (i = nbuf - 1; i >= 0; i--) {
1424 		bp = nbufp(i);
1425 		if (isbufbusy(bp)) {
1426 #if 0
1427 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1428 			if (bp->b_dev == NULL) {
1429 				TAILQ_REMOVE(&mountlist,
1430 				    bp->b_vp->v_mount, mnt_list);
1431 				continue;
1432 			}
1433 #endif
1434 			nbusy++;
1435 			if (show_busybufs > 0) {
1436 				printf(
1437 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1438 				    nbusy, bp, bp->b_vp, bp->b_flags,
1439 				    (intmax_t)bp->b_blkno,
1440 				    (intmax_t)bp->b_lblkno);
1441 				BUF_LOCKPRINTINFO(bp);
1442 				if (show_busybufs > 1)
1443 					vn_printf(bp->b_vp,
1444 					    "vnode content: ");
1445 			}
1446 		}
1447 	}
1448 	if (nbusy) {
1449 		/*
1450 		 * Failed to sync all blocks. Indicate this and don't
1451 		 * unmount filesystems (thus forcing an fsck on reboot).
1452 		 */
1453 		printf("Giving up on %d buffers\n", nbusy);
1454 		DELAY(5000000);	/* 5 seconds */
1455 		swapoff_all();
1456 	} else {
1457 		if (!first_buf_printf)
1458 			printf("Final sync complete\n");
1459 
1460 		/*
1461 		 * Unmount filesystems and perform swapoff, to quiesce
1462 		 * the system as much as possible.  In particular, no
1463 		 * I/O should be initiated from top levels since it
1464 		 * might be abruptly terminated by reset, or otherwise
1465 		 * erronously handled because other parts of the
1466 		 * system are disabled.
1467 		 *
1468 		 * Swapoff before unmount, because file-backed swap is
1469 		 * non-operational after unmount of the underlying
1470 		 * filesystem.
1471 		 */
1472 		if (!KERNEL_PANICKED()) {
1473 			swapoff_all();
1474 			vfs_unmountall();
1475 		}
1476 	}
1477 	DELAY(100000);		/* wait for console output to finish */
1478 }
1479 
1480 static void
1481 bpmap_qenter(struct buf *bp)
1482 {
1483 
1484 	BUF_CHECK_MAPPED(bp);
1485 
1486 	/*
1487 	 * bp->b_data is relative to bp->b_offset, but
1488 	 * bp->b_offset may be offset into the first page.
1489 	 */
1490 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1491 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1492 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1493 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1494 }
1495 
1496 static inline struct bufdomain *
1497 bufdomain(struct buf *bp)
1498 {
1499 
1500 	return (&bdomain[bp->b_domain]);
1501 }
1502 
1503 static struct bufqueue *
1504 bufqueue(struct buf *bp)
1505 {
1506 
1507 	switch (bp->b_qindex) {
1508 	case QUEUE_NONE:
1509 		/* FALLTHROUGH */
1510 	case QUEUE_SENTINEL:
1511 		return (NULL);
1512 	case QUEUE_EMPTY:
1513 		return (&bqempty);
1514 	case QUEUE_DIRTY:
1515 		return (&bufdomain(bp)->bd_dirtyq);
1516 	case QUEUE_CLEAN:
1517 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1518 	default:
1519 		break;
1520 	}
1521 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1522 }
1523 
1524 /*
1525  * Return the locked bufqueue that bp is a member of.
1526  */
1527 static struct bufqueue *
1528 bufqueue_acquire(struct buf *bp)
1529 {
1530 	struct bufqueue *bq, *nbq;
1531 
1532 	/*
1533 	 * bp can be pushed from a per-cpu queue to the
1534 	 * cleanq while we're waiting on the lock.  Retry
1535 	 * if the queues don't match.
1536 	 */
1537 	bq = bufqueue(bp);
1538 	BQ_LOCK(bq);
1539 	for (;;) {
1540 		nbq = bufqueue(bp);
1541 		if (bq == nbq)
1542 			break;
1543 		BQ_UNLOCK(bq);
1544 		BQ_LOCK(nbq);
1545 		bq = nbq;
1546 	}
1547 	return (bq);
1548 }
1549 
1550 /*
1551  *	binsfree:
1552  *
1553  *	Insert the buffer into the appropriate free list.  Requires a
1554  *	locked buffer on entry and buffer is unlocked before return.
1555  */
1556 static void
1557 binsfree(struct buf *bp, int qindex)
1558 {
1559 	struct bufdomain *bd;
1560 	struct bufqueue *bq;
1561 
1562 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1563 	    ("binsfree: Invalid qindex %d", qindex));
1564 	BUF_ASSERT_XLOCKED(bp);
1565 
1566 	/*
1567 	 * Handle delayed bremfree() processing.
1568 	 */
1569 	if (bp->b_flags & B_REMFREE) {
1570 		if (bp->b_qindex == qindex) {
1571 			bp->b_flags |= B_REUSE;
1572 			bp->b_flags &= ~B_REMFREE;
1573 			BUF_UNLOCK(bp);
1574 			return;
1575 		}
1576 		bq = bufqueue_acquire(bp);
1577 		bq_remove(bq, bp);
1578 		BQ_UNLOCK(bq);
1579 	}
1580 	bd = bufdomain(bp);
1581 	if (qindex == QUEUE_CLEAN) {
1582 		if (bd->bd_lim != 0)
1583 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1584 		else
1585 			bq = bd->bd_cleanq;
1586 	} else
1587 		bq = &bd->bd_dirtyq;
1588 	bq_insert(bq, bp, true);
1589 }
1590 
1591 /*
1592  * buf_free:
1593  *
1594  *	Free a buffer to the buf zone once it no longer has valid contents.
1595  */
1596 static void
1597 buf_free(struct buf *bp)
1598 {
1599 
1600 	if (bp->b_flags & B_REMFREE)
1601 		bremfreef(bp);
1602 	if (bp->b_vflags & BV_BKGRDINPROG)
1603 		panic("losing buffer 1");
1604 	if (bp->b_rcred != NOCRED) {
1605 		crfree(bp->b_rcred);
1606 		bp->b_rcred = NOCRED;
1607 	}
1608 	if (bp->b_wcred != NOCRED) {
1609 		crfree(bp->b_wcred);
1610 		bp->b_wcred = NOCRED;
1611 	}
1612 	if (!LIST_EMPTY(&bp->b_dep))
1613 		buf_deallocate(bp);
1614 	bufkva_free(bp);
1615 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1616 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
1617 	BUF_UNLOCK(bp);
1618 	uma_zfree(buf_zone, bp);
1619 }
1620 
1621 /*
1622  * buf_import:
1623  *
1624  *	Import bufs into the uma cache from the buf list.  The system still
1625  *	expects a static array of bufs and much of the synchronization
1626  *	around bufs assumes type stable storage.  As a result, UMA is used
1627  *	only as a per-cpu cache of bufs still maintained on a global list.
1628  */
1629 static int
1630 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1631 {
1632 	struct buf *bp;
1633 	int i;
1634 
1635 	BQ_LOCK(&bqempty);
1636 	for (i = 0; i < cnt; i++) {
1637 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1638 		if (bp == NULL)
1639 			break;
1640 		bq_remove(&bqempty, bp);
1641 		store[i] = bp;
1642 	}
1643 	BQ_UNLOCK(&bqempty);
1644 
1645 	return (i);
1646 }
1647 
1648 /*
1649  * buf_release:
1650  *
1651  *	Release bufs from the uma cache back to the buffer queues.
1652  */
1653 static void
1654 buf_release(void *arg, void **store, int cnt)
1655 {
1656 	struct bufqueue *bq;
1657 	struct buf *bp;
1658         int i;
1659 
1660 	bq = &bqempty;
1661 	BQ_LOCK(bq);
1662         for (i = 0; i < cnt; i++) {
1663 		bp = store[i];
1664 		/* Inline bq_insert() to batch locking. */
1665 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1666 		bp->b_flags &= ~(B_AGE | B_REUSE);
1667 		bq->bq_len++;
1668 		bp->b_qindex = bq->bq_index;
1669 	}
1670 	BQ_UNLOCK(bq);
1671 }
1672 
1673 /*
1674  * buf_alloc:
1675  *
1676  *	Allocate an empty buffer header.
1677  */
1678 static struct buf *
1679 buf_alloc(struct bufdomain *bd)
1680 {
1681 	struct buf *bp;
1682 	int freebufs, error;
1683 
1684 	/*
1685 	 * We can only run out of bufs in the buf zone if the average buf
1686 	 * is less than BKVASIZE.  In this case the actual wait/block will
1687 	 * come from buf_reycle() failing to flush one of these small bufs.
1688 	 */
1689 	bp = NULL;
1690 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1691 	if (freebufs > 0)
1692 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1693 	if (bp == NULL) {
1694 		atomic_add_int(&bd->bd_freebuffers, 1);
1695 		bufspace_daemon_wakeup(bd);
1696 		counter_u64_add(numbufallocfails, 1);
1697 		return (NULL);
1698 	}
1699 	/*
1700 	 * Wake-up the bufspace daemon on transition below threshold.
1701 	 */
1702 	if (freebufs == bd->bd_lofreebuffers)
1703 		bufspace_daemon_wakeup(bd);
1704 
1705 	error = BUF_LOCK(bp, LK_EXCLUSIVE, NULL);
1706 	KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1707 	    error));
1708 	(void)error;
1709 
1710 	KASSERT(bp->b_vp == NULL,
1711 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1712 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1713 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1714 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1715 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1716 	KASSERT(bp->b_npages == 0,
1717 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1718 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1719 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1720 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
1721 
1722 	bp->b_domain = BD_DOMAIN(bd);
1723 	bp->b_flags = 0;
1724 	bp->b_ioflags = 0;
1725 	bp->b_xflags = 0;
1726 	bp->b_vflags = 0;
1727 	bp->b_vp = NULL;
1728 	bp->b_blkno = bp->b_lblkno = 0;
1729 	bp->b_offset = NOOFFSET;
1730 	bp->b_iodone = 0;
1731 	bp->b_error = 0;
1732 	bp->b_resid = 0;
1733 	bp->b_bcount = 0;
1734 	bp->b_npages = 0;
1735 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1736 	bp->b_bufobj = NULL;
1737 	bp->b_data = bp->b_kvabase = unmapped_buf;
1738 	bp->b_fsprivate1 = NULL;
1739 	bp->b_fsprivate2 = NULL;
1740 	bp->b_fsprivate3 = NULL;
1741 	LIST_INIT(&bp->b_dep);
1742 
1743 	return (bp);
1744 }
1745 
1746 /*
1747  *	buf_recycle:
1748  *
1749  *	Free a buffer from the given bufqueue.  kva controls whether the
1750  *	freed buf must own some kva resources.  This is used for
1751  *	defragmenting.
1752  */
1753 static int
1754 buf_recycle(struct bufdomain *bd, bool kva)
1755 {
1756 	struct bufqueue *bq;
1757 	struct buf *bp, *nbp;
1758 
1759 	if (kva)
1760 		counter_u64_add(bufdefragcnt, 1);
1761 	nbp = NULL;
1762 	bq = bd->bd_cleanq;
1763 	BQ_LOCK(bq);
1764 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1765 	    ("buf_recycle: Locks don't match"));
1766 	nbp = TAILQ_FIRST(&bq->bq_queue);
1767 
1768 	/*
1769 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1770 	 * depending.
1771 	 */
1772 	while ((bp = nbp) != NULL) {
1773 		/*
1774 		 * Calculate next bp (we can only use it if we do not
1775 		 * release the bqlock).
1776 		 */
1777 		nbp = TAILQ_NEXT(bp, b_freelist);
1778 
1779 		/*
1780 		 * If we are defragging then we need a buffer with
1781 		 * some kva to reclaim.
1782 		 */
1783 		if (kva && bp->b_kvasize == 0)
1784 			continue;
1785 
1786 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1787 			continue;
1788 
1789 		/*
1790 		 * Implement a second chance algorithm for frequently
1791 		 * accessed buffers.
1792 		 */
1793 		if ((bp->b_flags & B_REUSE) != 0) {
1794 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1795 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1796 			bp->b_flags &= ~B_REUSE;
1797 			BUF_UNLOCK(bp);
1798 			continue;
1799 		}
1800 
1801 		/*
1802 		 * Skip buffers with background writes in progress.
1803 		 */
1804 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1805 			BUF_UNLOCK(bp);
1806 			continue;
1807 		}
1808 
1809 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1810 		    ("buf_recycle: inconsistent queue %d bp %p",
1811 		    bp->b_qindex, bp));
1812 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1813 		    ("getnewbuf: queue domain %d doesn't match request %d",
1814 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1815 		/*
1816 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1817 		 * the scan from this point on.
1818 		 */
1819 		bq_remove(bq, bp);
1820 		BQ_UNLOCK(bq);
1821 
1822 		/*
1823 		 * Requeue the background write buffer with error and
1824 		 * restart the scan.
1825 		 */
1826 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1827 			bqrelse(bp);
1828 			BQ_LOCK(bq);
1829 			nbp = TAILQ_FIRST(&bq->bq_queue);
1830 			continue;
1831 		}
1832 		bp->b_flags |= B_INVAL;
1833 		brelse(bp);
1834 		return (0);
1835 	}
1836 	bd->bd_wanted = 1;
1837 	BQ_UNLOCK(bq);
1838 
1839 	return (ENOBUFS);
1840 }
1841 
1842 /*
1843  *	bremfree:
1844  *
1845  *	Mark the buffer for removal from the appropriate free list.
1846  *
1847  */
1848 void
1849 bremfree(struct buf *bp)
1850 {
1851 
1852 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1853 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1854 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1855 	KASSERT(bp->b_qindex != QUEUE_NONE,
1856 	    ("bremfree: buffer %p not on a queue.", bp));
1857 	BUF_ASSERT_XLOCKED(bp);
1858 
1859 	bp->b_flags |= B_REMFREE;
1860 }
1861 
1862 /*
1863  *	bremfreef:
1864  *
1865  *	Force an immediate removal from a free list.  Used only in nfs when
1866  *	it abuses the b_freelist pointer.
1867  */
1868 void
1869 bremfreef(struct buf *bp)
1870 {
1871 	struct bufqueue *bq;
1872 
1873 	bq = bufqueue_acquire(bp);
1874 	bq_remove(bq, bp);
1875 	BQ_UNLOCK(bq);
1876 }
1877 
1878 static void
1879 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1880 {
1881 
1882 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1883 	TAILQ_INIT(&bq->bq_queue);
1884 	bq->bq_len = 0;
1885 	bq->bq_index = qindex;
1886 	bq->bq_subqueue = subqueue;
1887 }
1888 
1889 static void
1890 bd_init(struct bufdomain *bd)
1891 {
1892 	int i;
1893 
1894 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1895 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1896 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1897 	for (i = 0; i <= mp_maxid; i++)
1898 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1899 		    "bufq clean subqueue lock");
1900 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1901 }
1902 
1903 /*
1904  *	bq_remove:
1905  *
1906  *	Removes a buffer from the free list, must be called with the
1907  *	correct qlock held.
1908  */
1909 static void
1910 bq_remove(struct bufqueue *bq, struct buf *bp)
1911 {
1912 
1913 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1914 	    bp, bp->b_vp, bp->b_flags);
1915 	KASSERT(bp->b_qindex != QUEUE_NONE,
1916 	    ("bq_remove: buffer %p not on a queue.", bp));
1917 	KASSERT(bufqueue(bp) == bq,
1918 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1919 
1920 	BQ_ASSERT_LOCKED(bq);
1921 	if (bp->b_qindex != QUEUE_EMPTY) {
1922 		BUF_ASSERT_XLOCKED(bp);
1923 	}
1924 	KASSERT(bq->bq_len >= 1,
1925 	    ("queue %d underflow", bp->b_qindex));
1926 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1927 	bq->bq_len--;
1928 	bp->b_qindex = QUEUE_NONE;
1929 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1930 }
1931 
1932 static void
1933 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1934 {
1935 	struct buf *bp;
1936 
1937 	BQ_ASSERT_LOCKED(bq);
1938 	if (bq != bd->bd_cleanq) {
1939 		BD_LOCK(bd);
1940 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1941 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1942 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1943 			    b_freelist);
1944 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1945 		}
1946 		bd->bd_cleanq->bq_len += bq->bq_len;
1947 		bq->bq_len = 0;
1948 	}
1949 	if (bd->bd_wanted) {
1950 		bd->bd_wanted = 0;
1951 		wakeup(&bd->bd_wanted);
1952 	}
1953 	if (bq != bd->bd_cleanq)
1954 		BD_UNLOCK(bd);
1955 }
1956 
1957 static int
1958 bd_flushall(struct bufdomain *bd)
1959 {
1960 	struct bufqueue *bq;
1961 	int flushed;
1962 	int i;
1963 
1964 	if (bd->bd_lim == 0)
1965 		return (0);
1966 	flushed = 0;
1967 	for (i = 0; i <= mp_maxid; i++) {
1968 		bq = &bd->bd_subq[i];
1969 		if (bq->bq_len == 0)
1970 			continue;
1971 		BQ_LOCK(bq);
1972 		bd_flush(bd, bq);
1973 		BQ_UNLOCK(bq);
1974 		flushed++;
1975 	}
1976 
1977 	return (flushed);
1978 }
1979 
1980 static void
1981 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1982 {
1983 	struct bufdomain *bd;
1984 
1985 	if (bp->b_qindex != QUEUE_NONE)
1986 		panic("bq_insert: free buffer %p onto another queue?", bp);
1987 
1988 	bd = bufdomain(bp);
1989 	if (bp->b_flags & B_AGE) {
1990 		/* Place this buf directly on the real queue. */
1991 		if (bq->bq_index == QUEUE_CLEAN)
1992 			bq = bd->bd_cleanq;
1993 		BQ_LOCK(bq);
1994 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1995 	} else {
1996 		BQ_LOCK(bq);
1997 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1998 	}
1999 	bp->b_flags &= ~(B_AGE | B_REUSE);
2000 	bq->bq_len++;
2001 	bp->b_qindex = bq->bq_index;
2002 	bp->b_subqueue = bq->bq_subqueue;
2003 
2004 	/*
2005 	 * Unlock before we notify so that we don't wakeup a waiter that
2006 	 * fails a trylock on the buf and sleeps again.
2007 	 */
2008 	if (unlock)
2009 		BUF_UNLOCK(bp);
2010 
2011 	if (bp->b_qindex == QUEUE_CLEAN) {
2012 		/*
2013 		 * Flush the per-cpu queue and notify any waiters.
2014 		 */
2015 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2016 		    bq->bq_len >= bd->bd_lim))
2017 			bd_flush(bd, bq);
2018 	}
2019 	BQ_UNLOCK(bq);
2020 }
2021 
2022 /*
2023  *	bufkva_free:
2024  *
2025  *	Free the kva allocation for a buffer.
2026  *
2027  */
2028 static void
2029 bufkva_free(struct buf *bp)
2030 {
2031 
2032 #ifdef INVARIANTS
2033 	if (bp->b_kvasize == 0) {
2034 		KASSERT(bp->b_kvabase == unmapped_buf &&
2035 		    bp->b_data == unmapped_buf,
2036 		    ("Leaked KVA space on %p", bp));
2037 	} else if (buf_mapped(bp))
2038 		BUF_CHECK_MAPPED(bp);
2039 	else
2040 		BUF_CHECK_UNMAPPED(bp);
2041 #endif
2042 	if (bp->b_kvasize == 0)
2043 		return;
2044 
2045 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2046 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2047 	counter_u64_add(buffreekvacnt, 1);
2048 	bp->b_data = bp->b_kvabase = unmapped_buf;
2049 	bp->b_kvasize = 0;
2050 }
2051 
2052 /*
2053  *	bufkva_alloc:
2054  *
2055  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2056  */
2057 static int
2058 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2059 {
2060 	vm_offset_t addr;
2061 	int error;
2062 
2063 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2064 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2065 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
2066 	KASSERT(maxsize <= maxbcachebuf,
2067 	    ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2068 
2069 	bufkva_free(bp);
2070 
2071 	addr = 0;
2072 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2073 	if (error != 0) {
2074 		/*
2075 		 * Buffer map is too fragmented.  Request the caller
2076 		 * to defragment the map.
2077 		 */
2078 		return (error);
2079 	}
2080 	bp->b_kvabase = (caddr_t)addr;
2081 	bp->b_kvasize = maxsize;
2082 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2083 	if ((gbflags & GB_UNMAPPED) != 0) {
2084 		bp->b_data = unmapped_buf;
2085 		BUF_CHECK_UNMAPPED(bp);
2086 	} else {
2087 		bp->b_data = bp->b_kvabase;
2088 		BUF_CHECK_MAPPED(bp);
2089 	}
2090 	return (0);
2091 }
2092 
2093 /*
2094  *	bufkva_reclaim:
2095  *
2096  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2097  *	callback that fires to avoid returning failure.
2098  */
2099 static void
2100 bufkva_reclaim(vmem_t *vmem, int flags)
2101 {
2102 	bool done;
2103 	int q;
2104 	int i;
2105 
2106 	done = false;
2107 	for (i = 0; i < 5; i++) {
2108 		for (q = 0; q < buf_domains; q++)
2109 			if (buf_recycle(&bdomain[q], true) != 0)
2110 				done = true;
2111 		if (done)
2112 			break;
2113 	}
2114 	return;
2115 }
2116 
2117 /*
2118  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2119  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2120  * the buffer is valid and we do not have to do anything.
2121  */
2122 static void
2123 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2124     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2125 {
2126 	struct buf *rabp;
2127 	struct thread *td;
2128 	int i;
2129 
2130 	td = curthread;
2131 
2132 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2133 		if (inmem(vp, *rablkno))
2134 			continue;
2135 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2136 		if ((rabp->b_flags & B_CACHE) != 0) {
2137 			brelse(rabp);
2138 			continue;
2139 		}
2140 #ifdef RACCT
2141 		if (racct_enable) {
2142 			PROC_LOCK(curproc);
2143 			racct_add_buf(curproc, rabp, 0);
2144 			PROC_UNLOCK(curproc);
2145 		}
2146 #endif /* RACCT */
2147 		td->td_ru.ru_inblock++;
2148 		rabp->b_flags |= B_ASYNC;
2149 		rabp->b_flags &= ~B_INVAL;
2150 		if ((flags & GB_CKHASH) != 0) {
2151 			rabp->b_flags |= B_CKHASH;
2152 			rabp->b_ckhashcalc = ckhashfunc;
2153 		}
2154 		rabp->b_ioflags &= ~BIO_ERROR;
2155 		rabp->b_iocmd = BIO_READ;
2156 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2157 			rabp->b_rcred = crhold(cred);
2158 		vfs_busy_pages(rabp, 0);
2159 		BUF_KERNPROC(rabp);
2160 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2161 		bstrategy(rabp);
2162 	}
2163 }
2164 
2165 /*
2166  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2167  *
2168  * Get a buffer with the specified data.  Look in the cache first.  We
2169  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2170  * is set, the buffer is valid and we do not have to do anything, see
2171  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2172  *
2173  * Always return a NULL buffer pointer (in bpp) when returning an error.
2174  *
2175  * The blkno parameter is the logical block being requested. Normally
2176  * the mapping of logical block number to disk block address is done
2177  * by calling VOP_BMAP(). However, if the mapping is already known, the
2178  * disk block address can be passed using the dblkno parameter. If the
2179  * disk block address is not known, then the same value should be passed
2180  * for blkno and dblkno.
2181  */
2182 int
2183 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2184     daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2185     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2186 {
2187 	struct buf *bp;
2188 	struct thread *td;
2189 	int error, readwait, rv;
2190 
2191 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2192 	td = curthread;
2193 	/*
2194 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2195 	 * are specified.
2196 	 */
2197 	error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2198 	if (error != 0) {
2199 		*bpp = NULL;
2200 		return (error);
2201 	}
2202 	KASSERT(blkno == bp->b_lblkno,
2203 	    ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2204 	    (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2205 	flags &= ~GB_NOSPARSE;
2206 	*bpp = bp;
2207 
2208 	/*
2209 	 * If not found in cache, do some I/O
2210 	 */
2211 	readwait = 0;
2212 	if ((bp->b_flags & B_CACHE) == 0) {
2213 #ifdef RACCT
2214 		if (racct_enable) {
2215 			PROC_LOCK(td->td_proc);
2216 			racct_add_buf(td->td_proc, bp, 0);
2217 			PROC_UNLOCK(td->td_proc);
2218 		}
2219 #endif /* RACCT */
2220 		td->td_ru.ru_inblock++;
2221 		bp->b_iocmd = BIO_READ;
2222 		bp->b_flags &= ~B_INVAL;
2223 		if ((flags & GB_CKHASH) != 0) {
2224 			bp->b_flags |= B_CKHASH;
2225 			bp->b_ckhashcalc = ckhashfunc;
2226 		}
2227 		if ((flags & GB_CVTENXIO) != 0)
2228 			bp->b_xflags |= BX_CVTENXIO;
2229 		bp->b_ioflags &= ~BIO_ERROR;
2230 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2231 			bp->b_rcred = crhold(cred);
2232 		vfs_busy_pages(bp, 0);
2233 		bp->b_iooffset = dbtob(bp->b_blkno);
2234 		bstrategy(bp);
2235 		++readwait;
2236 	}
2237 
2238 	/*
2239 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2240 	 */
2241 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2242 
2243 	rv = 0;
2244 	if (readwait) {
2245 		rv = bufwait(bp);
2246 		if (rv != 0) {
2247 			brelse(bp);
2248 			*bpp = NULL;
2249 		}
2250 	}
2251 	return (rv);
2252 }
2253 
2254 /*
2255  * Write, release buffer on completion.  (Done by iodone
2256  * if async).  Do not bother writing anything if the buffer
2257  * is invalid.
2258  *
2259  * Note that we set B_CACHE here, indicating that buffer is
2260  * fully valid and thus cacheable.  This is true even of NFS
2261  * now so we set it generally.  This could be set either here
2262  * or in biodone() since the I/O is synchronous.  We put it
2263  * here.
2264  */
2265 int
2266 bufwrite(struct buf *bp)
2267 {
2268 	int oldflags;
2269 	struct vnode *vp;
2270 	long space;
2271 	int vp_md;
2272 
2273 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2274 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2275 		bp->b_flags |= B_INVAL | B_RELBUF;
2276 		bp->b_flags &= ~B_CACHE;
2277 		brelse(bp);
2278 		return (ENXIO);
2279 	}
2280 	if (bp->b_flags & B_INVAL) {
2281 		brelse(bp);
2282 		return (0);
2283 	}
2284 
2285 	if (bp->b_flags & B_BARRIER)
2286 		atomic_add_long(&barrierwrites, 1);
2287 
2288 	oldflags = bp->b_flags;
2289 
2290 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2291 	    ("FFS background buffer should not get here %p", bp));
2292 
2293 	vp = bp->b_vp;
2294 	if (vp)
2295 		vp_md = vp->v_vflag & VV_MD;
2296 	else
2297 		vp_md = 0;
2298 
2299 	/*
2300 	 * Mark the buffer clean.  Increment the bufobj write count
2301 	 * before bundirty() call, to prevent other thread from seeing
2302 	 * empty dirty list and zero counter for writes in progress,
2303 	 * falsely indicating that the bufobj is clean.
2304 	 */
2305 	bufobj_wref(bp->b_bufobj);
2306 	bundirty(bp);
2307 
2308 	bp->b_flags &= ~B_DONE;
2309 	bp->b_ioflags &= ~BIO_ERROR;
2310 	bp->b_flags |= B_CACHE;
2311 	bp->b_iocmd = BIO_WRITE;
2312 
2313 	vfs_busy_pages(bp, 1);
2314 
2315 	/*
2316 	 * Normal bwrites pipeline writes
2317 	 */
2318 	bp->b_runningbufspace = bp->b_bufsize;
2319 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2320 
2321 #ifdef RACCT
2322 	if (racct_enable) {
2323 		PROC_LOCK(curproc);
2324 		racct_add_buf(curproc, bp, 1);
2325 		PROC_UNLOCK(curproc);
2326 	}
2327 #endif /* RACCT */
2328 	curthread->td_ru.ru_oublock++;
2329 	if (oldflags & B_ASYNC)
2330 		BUF_KERNPROC(bp);
2331 	bp->b_iooffset = dbtob(bp->b_blkno);
2332 	buf_track(bp, __func__);
2333 	bstrategy(bp);
2334 
2335 	if ((oldflags & B_ASYNC) == 0) {
2336 		int rtval = bufwait(bp);
2337 		brelse(bp);
2338 		return (rtval);
2339 	} else if (space > hirunningspace) {
2340 		/*
2341 		 * don't allow the async write to saturate the I/O
2342 		 * system.  We will not deadlock here because
2343 		 * we are blocking waiting for I/O that is already in-progress
2344 		 * to complete. We do not block here if it is the update
2345 		 * or syncer daemon trying to clean up as that can lead
2346 		 * to deadlock.
2347 		 */
2348 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2349 			waitrunningbufspace();
2350 	}
2351 
2352 	return (0);
2353 }
2354 
2355 void
2356 bufbdflush(struct bufobj *bo, struct buf *bp)
2357 {
2358 	struct buf *nbp;
2359 	struct bufdomain *bd;
2360 
2361 	bd = &bdomain[bo->bo_domain];
2362 	if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2363 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2364 		altbufferflushes++;
2365 	} else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2366 		BO_LOCK(bo);
2367 		/*
2368 		 * Try to find a buffer to flush.
2369 		 */
2370 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2371 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2372 			    BUF_LOCK(nbp,
2373 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2374 				continue;
2375 			if (bp == nbp)
2376 				panic("bdwrite: found ourselves");
2377 			BO_UNLOCK(bo);
2378 			/* Don't countdeps with the bo lock held. */
2379 			if (buf_countdeps(nbp, 0)) {
2380 				BO_LOCK(bo);
2381 				BUF_UNLOCK(nbp);
2382 				continue;
2383 			}
2384 			if (nbp->b_flags & B_CLUSTEROK) {
2385 				vfs_bio_awrite(nbp);
2386 			} else {
2387 				bremfree(nbp);
2388 				bawrite(nbp);
2389 			}
2390 			dirtybufferflushes++;
2391 			break;
2392 		}
2393 		if (nbp == NULL)
2394 			BO_UNLOCK(bo);
2395 	}
2396 }
2397 
2398 /*
2399  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2400  * anything if the buffer is marked invalid.
2401  *
2402  * Note that since the buffer must be completely valid, we can safely
2403  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2404  * biodone() in order to prevent getblk from writing the buffer
2405  * out synchronously.
2406  */
2407 void
2408 bdwrite(struct buf *bp)
2409 {
2410 	struct thread *td = curthread;
2411 	struct vnode *vp;
2412 	struct bufobj *bo;
2413 
2414 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2415 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2416 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2417 	    ("Barrier request in delayed write %p", bp));
2418 
2419 	if (bp->b_flags & B_INVAL) {
2420 		brelse(bp);
2421 		return;
2422 	}
2423 
2424 	/*
2425 	 * If we have too many dirty buffers, don't create any more.
2426 	 * If we are wildly over our limit, then force a complete
2427 	 * cleanup. Otherwise, just keep the situation from getting
2428 	 * out of control. Note that we have to avoid a recursive
2429 	 * disaster and not try to clean up after our own cleanup!
2430 	 */
2431 	vp = bp->b_vp;
2432 	bo = bp->b_bufobj;
2433 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2434 		td->td_pflags |= TDP_INBDFLUSH;
2435 		BO_BDFLUSH(bo, bp);
2436 		td->td_pflags &= ~TDP_INBDFLUSH;
2437 	} else
2438 		recursiveflushes++;
2439 
2440 	bdirty(bp);
2441 	/*
2442 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2443 	 * true even of NFS now.
2444 	 */
2445 	bp->b_flags |= B_CACHE;
2446 
2447 	/*
2448 	 * This bmap keeps the system from needing to do the bmap later,
2449 	 * perhaps when the system is attempting to do a sync.  Since it
2450 	 * is likely that the indirect block -- or whatever other datastructure
2451 	 * that the filesystem needs is still in memory now, it is a good
2452 	 * thing to do this.  Note also, that if the pageout daemon is
2453 	 * requesting a sync -- there might not be enough memory to do
2454 	 * the bmap then...  So, this is important to do.
2455 	 */
2456 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2457 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2458 	}
2459 
2460 	buf_track(bp, __func__);
2461 
2462 	/*
2463 	 * Set the *dirty* buffer range based upon the VM system dirty
2464 	 * pages.
2465 	 *
2466 	 * Mark the buffer pages as clean.  We need to do this here to
2467 	 * satisfy the vnode_pager and the pageout daemon, so that it
2468 	 * thinks that the pages have been "cleaned".  Note that since
2469 	 * the pages are in a delayed write buffer -- the VFS layer
2470 	 * "will" see that the pages get written out on the next sync,
2471 	 * or perhaps the cluster will be completed.
2472 	 */
2473 	vfs_clean_pages_dirty_buf(bp);
2474 	bqrelse(bp);
2475 
2476 	/*
2477 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2478 	 * due to the softdep code.
2479 	 */
2480 }
2481 
2482 /*
2483  *	bdirty:
2484  *
2485  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2486  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2487  *	itself to properly update it in the dirty/clean lists.  We mark it
2488  *	B_DONE to ensure that any asynchronization of the buffer properly
2489  *	clears B_DONE ( else a panic will occur later ).
2490  *
2491  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2492  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2493  *	should only be called if the buffer is known-good.
2494  *
2495  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2496  *	count.
2497  *
2498  *	The buffer must be on QUEUE_NONE.
2499  */
2500 void
2501 bdirty(struct buf *bp)
2502 {
2503 
2504 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2505 	    bp, bp->b_vp, bp->b_flags);
2506 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2507 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2508 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2509 	bp->b_flags &= ~(B_RELBUF);
2510 	bp->b_iocmd = BIO_WRITE;
2511 
2512 	if ((bp->b_flags & B_DELWRI) == 0) {
2513 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2514 		reassignbuf(bp);
2515 		bdirtyadd(bp);
2516 	}
2517 }
2518 
2519 /*
2520  *	bundirty:
2521  *
2522  *	Clear B_DELWRI for buffer.
2523  *
2524  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2525  *	count.
2526  *
2527  *	The buffer must be on QUEUE_NONE.
2528  */
2529 
2530 void
2531 bundirty(struct buf *bp)
2532 {
2533 
2534 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2535 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2536 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2537 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2538 
2539 	if (bp->b_flags & B_DELWRI) {
2540 		bp->b_flags &= ~B_DELWRI;
2541 		reassignbuf(bp);
2542 		bdirtysub(bp);
2543 	}
2544 	/*
2545 	 * Since it is now being written, we can clear its deferred write flag.
2546 	 */
2547 	bp->b_flags &= ~B_DEFERRED;
2548 }
2549 
2550 /*
2551  *	bawrite:
2552  *
2553  *	Asynchronous write.  Start output on a buffer, but do not wait for
2554  *	it to complete.  The buffer is released when the output completes.
2555  *
2556  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2557  *	B_INVAL buffers.  Not us.
2558  */
2559 void
2560 bawrite(struct buf *bp)
2561 {
2562 
2563 	bp->b_flags |= B_ASYNC;
2564 	(void) bwrite(bp);
2565 }
2566 
2567 /*
2568  *	babarrierwrite:
2569  *
2570  *	Asynchronous barrier write.  Start output on a buffer, but do not
2571  *	wait for it to complete.  Place a write barrier after this write so
2572  *	that this buffer and all buffers written before it are committed to
2573  *	the disk before any buffers written after this write are committed
2574  *	to the disk.  The buffer is released when the output completes.
2575  */
2576 void
2577 babarrierwrite(struct buf *bp)
2578 {
2579 
2580 	bp->b_flags |= B_ASYNC | B_BARRIER;
2581 	(void) bwrite(bp);
2582 }
2583 
2584 /*
2585  *	bbarrierwrite:
2586  *
2587  *	Synchronous barrier write.  Start output on a buffer and wait for
2588  *	it to complete.  Place a write barrier after this write so that
2589  *	this buffer and all buffers written before it are committed to
2590  *	the disk before any buffers written after this write are committed
2591  *	to the disk.  The buffer is released when the output completes.
2592  */
2593 int
2594 bbarrierwrite(struct buf *bp)
2595 {
2596 
2597 	bp->b_flags |= B_BARRIER;
2598 	return (bwrite(bp));
2599 }
2600 
2601 /*
2602  *	bwillwrite:
2603  *
2604  *	Called prior to the locking of any vnodes when we are expecting to
2605  *	write.  We do not want to starve the buffer cache with too many
2606  *	dirty buffers so we block here.  By blocking prior to the locking
2607  *	of any vnodes we attempt to avoid the situation where a locked vnode
2608  *	prevents the various system daemons from flushing related buffers.
2609  */
2610 void
2611 bwillwrite(void)
2612 {
2613 
2614 	if (buf_dirty_count_severe()) {
2615 		mtx_lock(&bdirtylock);
2616 		while (buf_dirty_count_severe()) {
2617 			bdirtywait = 1;
2618 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2619 			    "flswai", 0);
2620 		}
2621 		mtx_unlock(&bdirtylock);
2622 	}
2623 }
2624 
2625 /*
2626  * Return true if we have too many dirty buffers.
2627  */
2628 int
2629 buf_dirty_count_severe(void)
2630 {
2631 
2632 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2633 }
2634 
2635 /*
2636  *	brelse:
2637  *
2638  *	Release a busy buffer and, if requested, free its resources.  The
2639  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2640  *	to be accessed later as a cache entity or reused for other purposes.
2641  */
2642 void
2643 brelse(struct buf *bp)
2644 {
2645 	struct mount *v_mnt;
2646 	int qindex;
2647 
2648 	/*
2649 	 * Many functions erroneously call brelse with a NULL bp under rare
2650 	 * error conditions. Simply return when called with a NULL bp.
2651 	 */
2652 	if (bp == NULL)
2653 		return;
2654 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2655 	    bp, bp->b_vp, bp->b_flags);
2656 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2657 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2658 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2659 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2660 
2661 	if (BUF_LOCKRECURSED(bp)) {
2662 		/*
2663 		 * Do not process, in particular, do not handle the
2664 		 * B_INVAL/B_RELBUF and do not release to free list.
2665 		 */
2666 		BUF_UNLOCK(bp);
2667 		return;
2668 	}
2669 
2670 	if (bp->b_flags & B_MANAGED) {
2671 		bqrelse(bp);
2672 		return;
2673 	}
2674 
2675 	if (LIST_EMPTY(&bp->b_dep)) {
2676 		bp->b_flags &= ~B_IOSTARTED;
2677 	} else {
2678 		KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2679 		    ("brelse: SU io not finished bp %p", bp));
2680 	}
2681 
2682 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2683 		BO_LOCK(bp->b_bufobj);
2684 		bp->b_vflags &= ~BV_BKGRDERR;
2685 		BO_UNLOCK(bp->b_bufobj);
2686 		bdirty(bp);
2687 	}
2688 
2689 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2690 	    (bp->b_flags & B_INVALONERR)) {
2691 		/*
2692 		 * Forced invalidation of dirty buffer contents, to be used
2693 		 * after a failed write in the rare case that the loss of the
2694 		 * contents is acceptable.  The buffer is invalidated and
2695 		 * freed.
2696 		 */
2697 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2698 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2699 	}
2700 
2701 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2702 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2703 	    !(bp->b_flags & B_INVAL)) {
2704 		/*
2705 		 * Failed write, redirty.  All errors except ENXIO (which
2706 		 * means the device is gone) are treated as being
2707 		 * transient.
2708 		 *
2709 		 * XXX Treating EIO as transient is not correct; the
2710 		 * contract with the local storage device drivers is that
2711 		 * they will only return EIO once the I/O is no longer
2712 		 * retriable.  Network I/O also respects this through the
2713 		 * guarantees of TCP and/or the internal retries of NFS.
2714 		 * ENOMEM might be transient, but we also have no way of
2715 		 * knowing when its ok to retry/reschedule.  In general,
2716 		 * this entire case should be made obsolete through better
2717 		 * error handling/recovery and resource scheduling.
2718 		 *
2719 		 * Do this also for buffers that failed with ENXIO, but have
2720 		 * non-empty dependencies - the soft updates code might need
2721 		 * to access the buffer to untangle them.
2722 		 *
2723 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2724 		 */
2725 		bp->b_ioflags &= ~BIO_ERROR;
2726 		bdirty(bp);
2727 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2728 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2729 		/*
2730 		 * Either a failed read I/O, or we were asked to free or not
2731 		 * cache the buffer, or we failed to write to a device that's
2732 		 * no longer present.
2733 		 */
2734 		bp->b_flags |= B_INVAL;
2735 		if (!LIST_EMPTY(&bp->b_dep))
2736 			buf_deallocate(bp);
2737 		if (bp->b_flags & B_DELWRI)
2738 			bdirtysub(bp);
2739 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2740 		if ((bp->b_flags & B_VMIO) == 0) {
2741 			allocbuf(bp, 0);
2742 			if (bp->b_vp)
2743 				brelvp(bp);
2744 		}
2745 	}
2746 
2747 	/*
2748 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2749 	 * is called with B_DELWRI set, the underlying pages may wind up
2750 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2751 	 * because pages associated with a B_DELWRI bp are marked clean.
2752 	 *
2753 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2754 	 * if B_DELWRI is set.
2755 	 */
2756 	if (bp->b_flags & B_DELWRI)
2757 		bp->b_flags &= ~B_RELBUF;
2758 
2759 	/*
2760 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2761 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2762 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2763 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2764 	 *
2765 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2766 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2767 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2768 	 *
2769 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2770 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2771 	 * the commit state and we cannot afford to lose the buffer. If the
2772 	 * buffer has a background write in progress, we need to keep it
2773 	 * around to prevent it from being reconstituted and starting a second
2774 	 * background write.
2775 	 */
2776 
2777 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2778 
2779 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2780 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2781 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2782 	    vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2783 		vfs_vmio_invalidate(bp);
2784 		allocbuf(bp, 0);
2785 	}
2786 
2787 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2788 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2789 		allocbuf(bp, 0);
2790 		bp->b_flags &= ~B_NOREUSE;
2791 		if (bp->b_vp != NULL)
2792 			brelvp(bp);
2793 	}
2794 
2795 	/*
2796 	 * If the buffer has junk contents signal it and eventually
2797 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2798 	 * doesn't find it.
2799 	 */
2800 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2801 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2802 		bp->b_flags |= B_INVAL;
2803 	if (bp->b_flags & B_INVAL) {
2804 		if (bp->b_flags & B_DELWRI)
2805 			bundirty(bp);
2806 		if (bp->b_vp)
2807 			brelvp(bp);
2808 	}
2809 
2810 	buf_track(bp, __func__);
2811 
2812 	/* buffers with no memory */
2813 	if (bp->b_bufsize == 0) {
2814 		buf_free(bp);
2815 		return;
2816 	}
2817 	/* buffers with junk contents */
2818 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2819 	    (bp->b_ioflags & BIO_ERROR)) {
2820 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2821 		if (bp->b_vflags & BV_BKGRDINPROG)
2822 			panic("losing buffer 2");
2823 		qindex = QUEUE_CLEAN;
2824 		bp->b_flags |= B_AGE;
2825 	/* remaining buffers */
2826 	} else if (bp->b_flags & B_DELWRI)
2827 		qindex = QUEUE_DIRTY;
2828 	else
2829 		qindex = QUEUE_CLEAN;
2830 
2831 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2832 		panic("brelse: not dirty");
2833 
2834 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2835 	bp->b_xflags &= ~(BX_CVTENXIO);
2836 	/* binsfree unlocks bp. */
2837 	binsfree(bp, qindex);
2838 }
2839 
2840 /*
2841  * Release a buffer back to the appropriate queue but do not try to free
2842  * it.  The buffer is expected to be used again soon.
2843  *
2844  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2845  * biodone() to requeue an async I/O on completion.  It is also used when
2846  * known good buffers need to be requeued but we think we may need the data
2847  * again soon.
2848  *
2849  * XXX we should be able to leave the B_RELBUF hint set on completion.
2850  */
2851 void
2852 bqrelse(struct buf *bp)
2853 {
2854 	int qindex;
2855 
2856 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2857 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2858 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2859 
2860 	qindex = QUEUE_NONE;
2861 	if (BUF_LOCKRECURSED(bp)) {
2862 		/* do not release to free list */
2863 		BUF_UNLOCK(bp);
2864 		return;
2865 	}
2866 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2867 	bp->b_xflags &= ~(BX_CVTENXIO);
2868 
2869 	if (LIST_EMPTY(&bp->b_dep)) {
2870 		bp->b_flags &= ~B_IOSTARTED;
2871 	} else {
2872 		KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2873 		    ("bqrelse: SU io not finished bp %p", bp));
2874 	}
2875 
2876 	if (bp->b_flags & B_MANAGED) {
2877 		if (bp->b_flags & B_REMFREE)
2878 			bremfreef(bp);
2879 		goto out;
2880 	}
2881 
2882 	/* buffers with stale but valid contents */
2883 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2884 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2885 		BO_LOCK(bp->b_bufobj);
2886 		bp->b_vflags &= ~BV_BKGRDERR;
2887 		BO_UNLOCK(bp->b_bufobj);
2888 		qindex = QUEUE_DIRTY;
2889 	} else {
2890 		if ((bp->b_flags & B_DELWRI) == 0 &&
2891 		    (bp->b_xflags & BX_VNDIRTY))
2892 			panic("bqrelse: not dirty");
2893 		if ((bp->b_flags & B_NOREUSE) != 0) {
2894 			brelse(bp);
2895 			return;
2896 		}
2897 		qindex = QUEUE_CLEAN;
2898 	}
2899 	buf_track(bp, __func__);
2900 	/* binsfree unlocks bp. */
2901 	binsfree(bp, qindex);
2902 	return;
2903 
2904 out:
2905 	buf_track(bp, __func__);
2906 	/* unlock */
2907 	BUF_UNLOCK(bp);
2908 }
2909 
2910 /*
2911  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2912  * restore bogus pages.
2913  */
2914 static void
2915 vfs_vmio_iodone(struct buf *bp)
2916 {
2917 	vm_ooffset_t foff;
2918 	vm_page_t m;
2919 	vm_object_t obj;
2920 	struct vnode *vp __unused;
2921 	int i, iosize, resid;
2922 	bool bogus;
2923 
2924 	obj = bp->b_bufobj->bo_object;
2925 	KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2926 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2927 	    blockcount_read(&obj->paging_in_progress), bp->b_npages));
2928 
2929 	vp = bp->b_vp;
2930 	VNPASS(vp->v_holdcnt > 0, vp);
2931 	VNPASS(vp->v_object != NULL, vp);
2932 
2933 	foff = bp->b_offset;
2934 	KASSERT(bp->b_offset != NOOFFSET,
2935 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2936 
2937 	bogus = false;
2938 	iosize = bp->b_bcount - bp->b_resid;
2939 	for (i = 0; i < bp->b_npages; i++) {
2940 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2941 		if (resid > iosize)
2942 			resid = iosize;
2943 
2944 		/*
2945 		 * cleanup bogus pages, restoring the originals
2946 		 */
2947 		m = bp->b_pages[i];
2948 		if (m == bogus_page) {
2949 			bogus = true;
2950 			m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2951 			if (m == NULL)
2952 				panic("biodone: page disappeared!");
2953 			bp->b_pages[i] = m;
2954 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2955 			/*
2956 			 * In the write case, the valid and clean bits are
2957 			 * already changed correctly ( see bdwrite() ), so we
2958 			 * only need to do this here in the read case.
2959 			 */
2960 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2961 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2962 			    "has unexpected dirty bits", m));
2963 			vfs_page_set_valid(bp, foff, m);
2964 		}
2965 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2966 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2967 		    (intmax_t)foff, (uintmax_t)m->pindex));
2968 
2969 		vm_page_sunbusy(m);
2970 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2971 		iosize -= resid;
2972 	}
2973 	vm_object_pip_wakeupn(obj, bp->b_npages);
2974 	if (bogus && buf_mapped(bp)) {
2975 		BUF_CHECK_MAPPED(bp);
2976 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2977 		    bp->b_pages, bp->b_npages);
2978 	}
2979 }
2980 
2981 /*
2982  * Perform page invalidation when a buffer is released.  The fully invalid
2983  * pages will be reclaimed later in vfs_vmio_truncate().
2984  */
2985 static void
2986 vfs_vmio_invalidate(struct buf *bp)
2987 {
2988 	vm_object_t obj;
2989 	vm_page_t m;
2990 	int flags, i, resid, poffset, presid;
2991 
2992 	if (buf_mapped(bp)) {
2993 		BUF_CHECK_MAPPED(bp);
2994 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2995 	} else
2996 		BUF_CHECK_UNMAPPED(bp);
2997 	/*
2998 	 * Get the base offset and length of the buffer.  Note that
2999 	 * in the VMIO case if the buffer block size is not
3000 	 * page-aligned then b_data pointer may not be page-aligned.
3001 	 * But our b_pages[] array *IS* page aligned.
3002 	 *
3003 	 * block sizes less then DEV_BSIZE (usually 512) are not
3004 	 * supported due to the page granularity bits (m->valid,
3005 	 * m->dirty, etc...).
3006 	 *
3007 	 * See man buf(9) for more information
3008 	 */
3009 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3010 	obj = bp->b_bufobj->bo_object;
3011 	resid = bp->b_bufsize;
3012 	poffset = bp->b_offset & PAGE_MASK;
3013 	VM_OBJECT_WLOCK(obj);
3014 	for (i = 0; i < bp->b_npages; i++) {
3015 		m = bp->b_pages[i];
3016 		if (m == bogus_page)
3017 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
3018 		bp->b_pages[i] = NULL;
3019 
3020 		presid = resid > (PAGE_SIZE - poffset) ?
3021 		    (PAGE_SIZE - poffset) : resid;
3022 		KASSERT(presid >= 0, ("brelse: extra page"));
3023 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3024 		if (pmap_page_wired_mappings(m) == 0)
3025 			vm_page_set_invalid(m, poffset, presid);
3026 		vm_page_sunbusy(m);
3027 		vm_page_release_locked(m, flags);
3028 		resid -= presid;
3029 		poffset = 0;
3030 	}
3031 	VM_OBJECT_WUNLOCK(obj);
3032 	bp->b_npages = 0;
3033 }
3034 
3035 /*
3036  * Page-granular truncation of an existing VMIO buffer.
3037  */
3038 static void
3039 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3040 {
3041 	vm_object_t obj;
3042 	vm_page_t m;
3043 	int flags, i;
3044 
3045 	if (bp->b_npages == desiredpages)
3046 		return;
3047 
3048 	if (buf_mapped(bp)) {
3049 		BUF_CHECK_MAPPED(bp);
3050 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3051 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3052 	} else
3053 		BUF_CHECK_UNMAPPED(bp);
3054 
3055 	/*
3056 	 * The object lock is needed only if we will attempt to free pages.
3057 	 */
3058 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3059 	if ((bp->b_flags & B_DIRECT) != 0) {
3060 		flags |= VPR_TRYFREE;
3061 		obj = bp->b_bufobj->bo_object;
3062 		VM_OBJECT_WLOCK(obj);
3063 	} else {
3064 		obj = NULL;
3065 	}
3066 	for (i = desiredpages; i < bp->b_npages; i++) {
3067 		m = bp->b_pages[i];
3068 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3069 		bp->b_pages[i] = NULL;
3070 		if (obj != NULL)
3071 			vm_page_release_locked(m, flags);
3072 		else
3073 			vm_page_release(m, flags);
3074 	}
3075 	if (obj != NULL)
3076 		VM_OBJECT_WUNLOCK(obj);
3077 	bp->b_npages = desiredpages;
3078 }
3079 
3080 /*
3081  * Byte granular extension of VMIO buffers.
3082  */
3083 static void
3084 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3085 {
3086 	/*
3087 	 * We are growing the buffer, possibly in a
3088 	 * byte-granular fashion.
3089 	 */
3090 	vm_object_t obj;
3091 	vm_offset_t toff;
3092 	vm_offset_t tinc;
3093 	vm_page_t m;
3094 
3095 	/*
3096 	 * Step 1, bring in the VM pages from the object, allocating
3097 	 * them if necessary.  We must clear B_CACHE if these pages
3098 	 * are not valid for the range covered by the buffer.
3099 	 */
3100 	obj = bp->b_bufobj->bo_object;
3101 	if (bp->b_npages < desiredpages) {
3102 		KASSERT(desiredpages <= atop(maxbcachebuf),
3103 		    ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3104 		    bp, desiredpages, maxbcachebuf));
3105 
3106 		/*
3107 		 * We must allocate system pages since blocking
3108 		 * here could interfere with paging I/O, no
3109 		 * matter which process we are.
3110 		 *
3111 		 * Only exclusive busy can be tested here.
3112 		 * Blocking on shared busy might lead to
3113 		 * deadlocks once allocbuf() is called after
3114 		 * pages are vfs_busy_pages().
3115 		 */
3116 		(void)vm_page_grab_pages_unlocked(obj,
3117 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3118 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3119 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3120 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3121 		bp->b_npages = desiredpages;
3122 	}
3123 
3124 	/*
3125 	 * Step 2.  We've loaded the pages into the buffer,
3126 	 * we have to figure out if we can still have B_CACHE
3127 	 * set.  Note that B_CACHE is set according to the
3128 	 * byte-granular range ( bcount and size ), not the
3129 	 * aligned range ( newbsize ).
3130 	 *
3131 	 * The VM test is against m->valid, which is DEV_BSIZE
3132 	 * aligned.  Needless to say, the validity of the data
3133 	 * needs to also be DEV_BSIZE aligned.  Note that this
3134 	 * fails with NFS if the server or some other client
3135 	 * extends the file's EOF.  If our buffer is resized,
3136 	 * B_CACHE may remain set! XXX
3137 	 */
3138 	toff = bp->b_bcount;
3139 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3140 	while ((bp->b_flags & B_CACHE) && toff < size) {
3141 		vm_pindex_t pi;
3142 
3143 		if (tinc > (size - toff))
3144 			tinc = size - toff;
3145 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3146 		m = bp->b_pages[pi];
3147 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3148 		toff += tinc;
3149 		tinc = PAGE_SIZE;
3150 	}
3151 
3152 	/*
3153 	 * Step 3, fixup the KVA pmap.
3154 	 */
3155 	if (buf_mapped(bp))
3156 		bpmap_qenter(bp);
3157 	else
3158 		BUF_CHECK_UNMAPPED(bp);
3159 }
3160 
3161 /*
3162  * Check to see if a block at a particular lbn is available for a clustered
3163  * write.
3164  */
3165 static int
3166 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3167 {
3168 	struct buf *bpa;
3169 	int match;
3170 
3171 	match = 0;
3172 
3173 	/* If the buf isn't in core skip it */
3174 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3175 		return (0);
3176 
3177 	/* If the buf is busy we don't want to wait for it */
3178 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3179 		return (0);
3180 
3181 	/* Only cluster with valid clusterable delayed write buffers */
3182 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3183 	    (B_DELWRI | B_CLUSTEROK))
3184 		goto done;
3185 
3186 	if (bpa->b_bufsize != size)
3187 		goto done;
3188 
3189 	/*
3190 	 * Check to see if it is in the expected place on disk and that the
3191 	 * block has been mapped.
3192 	 */
3193 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3194 		match = 1;
3195 done:
3196 	BUF_UNLOCK(bpa);
3197 	return (match);
3198 }
3199 
3200 /*
3201  *	vfs_bio_awrite:
3202  *
3203  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3204  *	This is much better then the old way of writing only one buffer at
3205  *	a time.  Note that we may not be presented with the buffers in the
3206  *	correct order, so we search for the cluster in both directions.
3207  */
3208 int
3209 vfs_bio_awrite(struct buf *bp)
3210 {
3211 	struct bufobj *bo;
3212 	int i;
3213 	int j;
3214 	daddr_t lblkno = bp->b_lblkno;
3215 	struct vnode *vp = bp->b_vp;
3216 	int ncl;
3217 	int nwritten;
3218 	int size;
3219 	int maxcl;
3220 	int gbflags;
3221 
3222 	bo = &vp->v_bufobj;
3223 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3224 	/*
3225 	 * right now we support clustered writing only to regular files.  If
3226 	 * we find a clusterable block we could be in the middle of a cluster
3227 	 * rather then at the beginning.
3228 	 */
3229 	if ((vp->v_type == VREG) &&
3230 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3231 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3232 		size = vp->v_mount->mnt_stat.f_iosize;
3233 		maxcl = maxphys / size;
3234 
3235 		BO_RLOCK(bo);
3236 		for (i = 1; i < maxcl; i++)
3237 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3238 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3239 				break;
3240 
3241 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3242 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3243 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3244 				break;
3245 		BO_RUNLOCK(bo);
3246 		--j;
3247 		ncl = i + j;
3248 		/*
3249 		 * this is a possible cluster write
3250 		 */
3251 		if (ncl != 1) {
3252 			BUF_UNLOCK(bp);
3253 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3254 			    gbflags);
3255 			return (nwritten);
3256 		}
3257 	}
3258 	bremfree(bp);
3259 	bp->b_flags |= B_ASYNC;
3260 	/*
3261 	 * default (old) behavior, writing out only one block
3262 	 *
3263 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3264 	 */
3265 	nwritten = bp->b_bufsize;
3266 	(void) bwrite(bp);
3267 
3268 	return (nwritten);
3269 }
3270 
3271 /*
3272  *	getnewbuf_kva:
3273  *
3274  *	Allocate KVA for an empty buf header according to gbflags.
3275  */
3276 static int
3277 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3278 {
3279 
3280 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3281 		/*
3282 		 * In order to keep fragmentation sane we only allocate kva
3283 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3284 		 */
3285 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3286 
3287 		if (maxsize != bp->b_kvasize &&
3288 		    bufkva_alloc(bp, maxsize, gbflags))
3289 			return (ENOSPC);
3290 	}
3291 	return (0);
3292 }
3293 
3294 /*
3295  *	getnewbuf:
3296  *
3297  *	Find and initialize a new buffer header, freeing up existing buffers
3298  *	in the bufqueues as necessary.  The new buffer is returned locked.
3299  *
3300  *	We block if:
3301  *		We have insufficient buffer headers
3302  *		We have insufficient buffer space
3303  *		buffer_arena is too fragmented ( space reservation fails )
3304  *		If we have to flush dirty buffers ( but we try to avoid this )
3305  *
3306  *	The caller is responsible for releasing the reserved bufspace after
3307  *	allocbuf() is called.
3308  */
3309 static struct buf *
3310 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3311 {
3312 	struct bufdomain *bd;
3313 	struct buf *bp;
3314 	bool metadata, reserved;
3315 
3316 	bp = NULL;
3317 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3318 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3319 	if (!unmapped_buf_allowed)
3320 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3321 
3322 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3323 	    vp->v_type == VCHR)
3324 		metadata = true;
3325 	else
3326 		metadata = false;
3327 	if (vp == NULL)
3328 		bd = &bdomain[0];
3329 	else
3330 		bd = &bdomain[vp->v_bufobj.bo_domain];
3331 
3332 	counter_u64_add(getnewbufcalls, 1);
3333 	reserved = false;
3334 	do {
3335 		if (reserved == false &&
3336 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3337 			counter_u64_add(getnewbufrestarts, 1);
3338 			continue;
3339 		}
3340 		reserved = true;
3341 		if ((bp = buf_alloc(bd)) == NULL) {
3342 			counter_u64_add(getnewbufrestarts, 1);
3343 			continue;
3344 		}
3345 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3346 			return (bp);
3347 		break;
3348 	} while (buf_recycle(bd, false) == 0);
3349 
3350 	if (reserved)
3351 		bufspace_release(bd, maxsize);
3352 	if (bp != NULL) {
3353 		bp->b_flags |= B_INVAL;
3354 		brelse(bp);
3355 	}
3356 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3357 
3358 	return (NULL);
3359 }
3360 
3361 /*
3362  *	buf_daemon:
3363  *
3364  *	buffer flushing daemon.  Buffers are normally flushed by the
3365  *	update daemon but if it cannot keep up this process starts to
3366  *	take the load in an attempt to prevent getnewbuf() from blocking.
3367  */
3368 static struct kproc_desc buf_kp = {
3369 	"bufdaemon",
3370 	buf_daemon,
3371 	&bufdaemonproc
3372 };
3373 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3374 
3375 static int
3376 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3377 {
3378 	int flushed;
3379 
3380 	flushed = flushbufqueues(vp, bd, target, 0);
3381 	if (flushed == 0) {
3382 		/*
3383 		 * Could not find any buffers without rollback
3384 		 * dependencies, so just write the first one
3385 		 * in the hopes of eventually making progress.
3386 		 */
3387 		if (vp != NULL && target > 2)
3388 			target /= 2;
3389 		flushbufqueues(vp, bd, target, 1);
3390 	}
3391 	return (flushed);
3392 }
3393 
3394 static void
3395 buf_daemon()
3396 {
3397 	struct bufdomain *bd;
3398 	int speedupreq;
3399 	int lodirty;
3400 	int i;
3401 
3402 	/*
3403 	 * This process needs to be suspended prior to shutdown sync.
3404 	 */
3405 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3406 	    SHUTDOWN_PRI_LAST + 100);
3407 
3408 	/*
3409 	 * Start the buf clean daemons as children threads.
3410 	 */
3411 	for (i = 0 ; i < buf_domains; i++) {
3412 		int error;
3413 
3414 		error = kthread_add((void (*)(void *))bufspace_daemon,
3415 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3416 		if (error)
3417 			panic("error %d spawning bufspace daemon", error);
3418 	}
3419 
3420 	/*
3421 	 * This process is allowed to take the buffer cache to the limit
3422 	 */
3423 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3424 	mtx_lock(&bdlock);
3425 	for (;;) {
3426 		bd_request = 0;
3427 		mtx_unlock(&bdlock);
3428 
3429 		kthread_suspend_check();
3430 
3431 		/*
3432 		 * Save speedupreq for this pass and reset to capture new
3433 		 * requests.
3434 		 */
3435 		speedupreq = bd_speedupreq;
3436 		bd_speedupreq = 0;
3437 
3438 		/*
3439 		 * Flush each domain sequentially according to its level and
3440 		 * the speedup request.
3441 		 */
3442 		for (i = 0; i < buf_domains; i++) {
3443 			bd = &bdomain[i];
3444 			if (speedupreq)
3445 				lodirty = bd->bd_numdirtybuffers / 2;
3446 			else
3447 				lodirty = bd->bd_lodirtybuffers;
3448 			while (bd->bd_numdirtybuffers > lodirty) {
3449 				if (buf_flush(NULL, bd,
3450 				    bd->bd_numdirtybuffers - lodirty) == 0)
3451 					break;
3452 				kern_yield(PRI_USER);
3453 			}
3454 		}
3455 
3456 		/*
3457 		 * Only clear bd_request if we have reached our low water
3458 		 * mark.  The buf_daemon normally waits 1 second and
3459 		 * then incrementally flushes any dirty buffers that have
3460 		 * built up, within reason.
3461 		 *
3462 		 * If we were unable to hit our low water mark and couldn't
3463 		 * find any flushable buffers, we sleep for a short period
3464 		 * to avoid endless loops on unlockable buffers.
3465 		 */
3466 		mtx_lock(&bdlock);
3467 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3468 			/*
3469 			 * We reached our low water mark, reset the
3470 			 * request and sleep until we are needed again.
3471 			 * The sleep is just so the suspend code works.
3472 			 */
3473 			bd_request = 0;
3474 			/*
3475 			 * Do an extra wakeup in case dirty threshold
3476 			 * changed via sysctl and the explicit transition
3477 			 * out of shortfall was missed.
3478 			 */
3479 			bdirtywakeup();
3480 			if (runningbufspace <= lorunningspace)
3481 				runningwakeup();
3482 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3483 		} else {
3484 			/*
3485 			 * We couldn't find any flushable dirty buffers but
3486 			 * still have too many dirty buffers, we
3487 			 * have to sleep and try again.  (rare)
3488 			 */
3489 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3490 		}
3491 	}
3492 }
3493 
3494 /*
3495  *	flushbufqueues:
3496  *
3497  *	Try to flush a buffer in the dirty queue.  We must be careful to
3498  *	free up B_INVAL buffers instead of write them, which NFS is
3499  *	particularly sensitive to.
3500  */
3501 static int flushwithdeps = 0;
3502 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3503     &flushwithdeps, 0,
3504     "Number of buffers flushed with dependencies that require rollbacks");
3505 
3506 static int
3507 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3508     int flushdeps)
3509 {
3510 	struct bufqueue *bq;
3511 	struct buf *sentinel;
3512 	struct vnode *vp;
3513 	struct mount *mp;
3514 	struct buf *bp;
3515 	int hasdeps;
3516 	int flushed;
3517 	int error;
3518 	bool unlock;
3519 
3520 	flushed = 0;
3521 	bq = &bd->bd_dirtyq;
3522 	bp = NULL;
3523 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3524 	sentinel->b_qindex = QUEUE_SENTINEL;
3525 	BQ_LOCK(bq);
3526 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3527 	BQ_UNLOCK(bq);
3528 	while (flushed != target) {
3529 		maybe_yield();
3530 		BQ_LOCK(bq);
3531 		bp = TAILQ_NEXT(sentinel, b_freelist);
3532 		if (bp != NULL) {
3533 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3534 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3535 			    b_freelist);
3536 		} else {
3537 			BQ_UNLOCK(bq);
3538 			break;
3539 		}
3540 		/*
3541 		 * Skip sentinels inserted by other invocations of the
3542 		 * flushbufqueues(), taking care to not reorder them.
3543 		 *
3544 		 * Only flush the buffers that belong to the
3545 		 * vnode locked by the curthread.
3546 		 */
3547 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3548 		    bp->b_vp != lvp)) {
3549 			BQ_UNLOCK(bq);
3550 			continue;
3551 		}
3552 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3553 		BQ_UNLOCK(bq);
3554 		if (error != 0)
3555 			continue;
3556 
3557 		/*
3558 		 * BKGRDINPROG can only be set with the buf and bufobj
3559 		 * locks both held.  We tolerate a race to clear it here.
3560 		 */
3561 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3562 		    (bp->b_flags & B_DELWRI) == 0) {
3563 			BUF_UNLOCK(bp);
3564 			continue;
3565 		}
3566 		if (bp->b_flags & B_INVAL) {
3567 			bremfreef(bp);
3568 			brelse(bp);
3569 			flushed++;
3570 			continue;
3571 		}
3572 
3573 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3574 			if (flushdeps == 0) {
3575 				BUF_UNLOCK(bp);
3576 				continue;
3577 			}
3578 			hasdeps = 1;
3579 		} else
3580 			hasdeps = 0;
3581 		/*
3582 		 * We must hold the lock on a vnode before writing
3583 		 * one of its buffers. Otherwise we may confuse, or
3584 		 * in the case of a snapshot vnode, deadlock the
3585 		 * system.
3586 		 *
3587 		 * The lock order here is the reverse of the normal
3588 		 * of vnode followed by buf lock.  This is ok because
3589 		 * the NOWAIT will prevent deadlock.
3590 		 */
3591 		vp = bp->b_vp;
3592 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3593 			BUF_UNLOCK(bp);
3594 			continue;
3595 		}
3596 		if (lvp == NULL) {
3597 			unlock = true;
3598 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3599 		} else {
3600 			ASSERT_VOP_LOCKED(vp, "getbuf");
3601 			unlock = false;
3602 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3603 			    vn_lock(vp, LK_TRYUPGRADE);
3604 		}
3605 		if (error == 0) {
3606 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3607 			    bp, bp->b_vp, bp->b_flags);
3608 			if (curproc == bufdaemonproc) {
3609 				vfs_bio_awrite(bp);
3610 			} else {
3611 				bremfree(bp);
3612 				bwrite(bp);
3613 				counter_u64_add(notbufdflushes, 1);
3614 			}
3615 			vn_finished_write(mp);
3616 			if (unlock)
3617 				VOP_UNLOCK(vp);
3618 			flushwithdeps += hasdeps;
3619 			flushed++;
3620 
3621 			/*
3622 			 * Sleeping on runningbufspace while holding
3623 			 * vnode lock leads to deadlock.
3624 			 */
3625 			if (curproc == bufdaemonproc &&
3626 			    runningbufspace > hirunningspace)
3627 				waitrunningbufspace();
3628 			continue;
3629 		}
3630 		vn_finished_write(mp);
3631 		BUF_UNLOCK(bp);
3632 	}
3633 	BQ_LOCK(bq);
3634 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3635 	BQ_UNLOCK(bq);
3636 	free(sentinel, M_TEMP);
3637 	return (flushed);
3638 }
3639 
3640 /*
3641  * Check to see if a block is currently memory resident.
3642  */
3643 struct buf *
3644 incore(struct bufobj *bo, daddr_t blkno)
3645 {
3646 	return (gbincore_unlocked(bo, blkno));
3647 }
3648 
3649 /*
3650  * Returns true if no I/O is needed to access the
3651  * associated VM object.  This is like incore except
3652  * it also hunts around in the VM system for the data.
3653  */
3654 bool
3655 inmem(struct vnode * vp, daddr_t blkno)
3656 {
3657 	vm_object_t obj;
3658 	vm_offset_t toff, tinc, size;
3659 	vm_page_t m, n;
3660 	vm_ooffset_t off;
3661 	int valid;
3662 
3663 	ASSERT_VOP_LOCKED(vp, "inmem");
3664 
3665 	if (incore(&vp->v_bufobj, blkno))
3666 		return (true);
3667 	if (vp->v_mount == NULL)
3668 		return (false);
3669 	obj = vp->v_object;
3670 	if (obj == NULL)
3671 		return (false);
3672 
3673 	size = PAGE_SIZE;
3674 	if (size > vp->v_mount->mnt_stat.f_iosize)
3675 		size = vp->v_mount->mnt_stat.f_iosize;
3676 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3677 
3678 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3679 		m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3680 recheck:
3681 		if (m == NULL)
3682 			return (false);
3683 
3684 		tinc = size;
3685 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3686 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3687 		/*
3688 		 * Consider page validity only if page mapping didn't change
3689 		 * during the check.
3690 		 */
3691 		valid = vm_page_is_valid(m,
3692 		    (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3693 		n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3694 		if (m != n) {
3695 			m = n;
3696 			goto recheck;
3697 		}
3698 		if (!valid)
3699 			return (false);
3700 	}
3701 	return (true);
3702 }
3703 
3704 /*
3705  * Set the dirty range for a buffer based on the status of the dirty
3706  * bits in the pages comprising the buffer.  The range is limited
3707  * to the size of the buffer.
3708  *
3709  * Tell the VM system that the pages associated with this buffer
3710  * are clean.  This is used for delayed writes where the data is
3711  * going to go to disk eventually without additional VM intevention.
3712  *
3713  * Note that while we only really need to clean through to b_bcount, we
3714  * just go ahead and clean through to b_bufsize.
3715  */
3716 static void
3717 vfs_clean_pages_dirty_buf(struct buf *bp)
3718 {
3719 	vm_ooffset_t foff, noff, eoff;
3720 	vm_page_t m;
3721 	int i;
3722 
3723 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3724 		return;
3725 
3726 	foff = bp->b_offset;
3727 	KASSERT(bp->b_offset != NOOFFSET,
3728 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3729 
3730 	vfs_busy_pages_acquire(bp);
3731 	vfs_setdirty_range(bp);
3732 	for (i = 0; i < bp->b_npages; i++) {
3733 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3734 		eoff = noff;
3735 		if (eoff > bp->b_offset + bp->b_bufsize)
3736 			eoff = bp->b_offset + bp->b_bufsize;
3737 		m = bp->b_pages[i];
3738 		vfs_page_set_validclean(bp, foff, m);
3739 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3740 		foff = noff;
3741 	}
3742 	vfs_busy_pages_release(bp);
3743 }
3744 
3745 static void
3746 vfs_setdirty_range(struct buf *bp)
3747 {
3748 	vm_offset_t boffset;
3749 	vm_offset_t eoffset;
3750 	int i;
3751 
3752 	/*
3753 	 * test the pages to see if they have been modified directly
3754 	 * by users through the VM system.
3755 	 */
3756 	for (i = 0; i < bp->b_npages; i++)
3757 		vm_page_test_dirty(bp->b_pages[i]);
3758 
3759 	/*
3760 	 * Calculate the encompassing dirty range, boffset and eoffset,
3761 	 * (eoffset - boffset) bytes.
3762 	 */
3763 
3764 	for (i = 0; i < bp->b_npages; i++) {
3765 		if (bp->b_pages[i]->dirty)
3766 			break;
3767 	}
3768 	boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3769 
3770 	for (i = bp->b_npages - 1; i >= 0; --i) {
3771 		if (bp->b_pages[i]->dirty) {
3772 			break;
3773 		}
3774 	}
3775 	eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3776 
3777 	/*
3778 	 * Fit it to the buffer.
3779 	 */
3780 
3781 	if (eoffset > bp->b_bcount)
3782 		eoffset = bp->b_bcount;
3783 
3784 	/*
3785 	 * If we have a good dirty range, merge with the existing
3786 	 * dirty range.
3787 	 */
3788 
3789 	if (boffset < eoffset) {
3790 		if (bp->b_dirtyoff > boffset)
3791 			bp->b_dirtyoff = boffset;
3792 		if (bp->b_dirtyend < eoffset)
3793 			bp->b_dirtyend = eoffset;
3794 	}
3795 }
3796 
3797 /*
3798  * Allocate the KVA mapping for an existing buffer.
3799  * If an unmapped buffer is provided but a mapped buffer is requested, take
3800  * also care to properly setup mappings between pages and KVA.
3801  */
3802 static void
3803 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3804 {
3805 	int bsize, maxsize, need_mapping, need_kva;
3806 	off_t offset;
3807 
3808 	need_mapping = bp->b_data == unmapped_buf &&
3809 	    (gbflags & GB_UNMAPPED) == 0;
3810 	need_kva = bp->b_kvabase == unmapped_buf &&
3811 	    bp->b_data == unmapped_buf &&
3812 	    (gbflags & GB_KVAALLOC) != 0;
3813 	if (!need_mapping && !need_kva)
3814 		return;
3815 
3816 	BUF_CHECK_UNMAPPED(bp);
3817 
3818 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3819 		/*
3820 		 * Buffer is not mapped, but the KVA was already
3821 		 * reserved at the time of the instantiation.  Use the
3822 		 * allocated space.
3823 		 */
3824 		goto has_addr;
3825 	}
3826 
3827 	/*
3828 	 * Calculate the amount of the address space we would reserve
3829 	 * if the buffer was mapped.
3830 	 */
3831 	bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3832 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3833 	offset = blkno * bsize;
3834 	maxsize = size + (offset & PAGE_MASK);
3835 	maxsize = imax(maxsize, bsize);
3836 
3837 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3838 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3839 			/*
3840 			 * XXXKIB: defragmentation cannot
3841 			 * succeed, not sure what else to do.
3842 			 */
3843 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3844 		}
3845 		counter_u64_add(mappingrestarts, 1);
3846 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3847 	}
3848 has_addr:
3849 	if (need_mapping) {
3850 		/* b_offset is handled by bpmap_qenter. */
3851 		bp->b_data = bp->b_kvabase;
3852 		BUF_CHECK_MAPPED(bp);
3853 		bpmap_qenter(bp);
3854 	}
3855 }
3856 
3857 struct buf *
3858 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3859     int flags)
3860 {
3861 	struct buf *bp;
3862 	int error;
3863 
3864 	error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3865 	if (error != 0)
3866 		return (NULL);
3867 	return (bp);
3868 }
3869 
3870 /*
3871  *	getblkx:
3872  *
3873  *	Get a block given a specified block and offset into a file/device.
3874  *	The buffers B_DONE bit will be cleared on return, making it almost
3875  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3876  *	return.  The caller should clear B_INVAL prior to initiating a
3877  *	READ.
3878  *
3879  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3880  *	an existing buffer.
3881  *
3882  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3883  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3884  *	and then cleared based on the backing VM.  If the previous buffer is
3885  *	non-0-sized but invalid, B_CACHE will be cleared.
3886  *
3887  *	If getblk() must create a new buffer, the new buffer is returned with
3888  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3889  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3890  *	backing VM.
3891  *
3892  *	getblk() also forces a bwrite() for any B_DELWRI buffer whose
3893  *	B_CACHE bit is clear.
3894  *
3895  *	What this means, basically, is that the caller should use B_CACHE to
3896  *	determine whether the buffer is fully valid or not and should clear
3897  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3898  *	the buffer by loading its data area with something, the caller needs
3899  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3900  *	the caller should set B_CACHE ( as an optimization ), else the caller
3901  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3902  *	a write attempt or if it was a successful read.  If the caller
3903  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3904  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3905  *
3906  *	The blkno parameter is the logical block being requested. Normally
3907  *	the mapping of logical block number to disk block address is done
3908  *	by calling VOP_BMAP(). However, if the mapping is already known, the
3909  *	disk block address can be passed using the dblkno parameter. If the
3910  *	disk block address is not known, then the same value should be passed
3911  *	for blkno and dblkno.
3912  */
3913 int
3914 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3915     int slptimeo, int flags, struct buf **bpp)
3916 {
3917 	struct buf *bp;
3918 	struct bufobj *bo;
3919 	daddr_t d_blkno;
3920 	int bsize, error, maxsize, vmio;
3921 	off_t offset;
3922 
3923 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3924 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3925 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3926 	if (vp->v_type != VCHR)
3927 		ASSERT_VOP_LOCKED(vp, "getblk");
3928 	if (size > maxbcachebuf)
3929 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3930 		    maxbcachebuf);
3931 	if (!unmapped_buf_allowed)
3932 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3933 
3934 	bo = &vp->v_bufobj;
3935 	d_blkno = dblkno;
3936 
3937 	/* Attempt lockless lookup first. */
3938 	bp = gbincore_unlocked(bo, blkno);
3939 	if (bp == NULL) {
3940 		/*
3941 		 * With GB_NOCREAT we must be sure about not finding the buffer
3942 		 * as it may have been reassigned during unlocked lookup.
3943 		 */
3944 		if ((flags & GB_NOCREAT) != 0)
3945 			goto loop;
3946 		goto newbuf_unlocked;
3947 	}
3948 
3949 	error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3950 	    0);
3951 	if (error != 0)
3952 		goto loop;
3953 
3954 	/* Verify buf identify has not changed since lookup. */
3955 	if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3956 		goto foundbuf_fastpath;
3957 
3958 	/* It changed, fallback to locked lookup. */
3959 	BUF_UNLOCK_RAW(bp);
3960 
3961 loop:
3962 	BO_RLOCK(bo);
3963 	bp = gbincore(bo, blkno);
3964 	if (bp != NULL) {
3965 		int lockflags;
3966 
3967 		/*
3968 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3969 		 * it must be on a queue.
3970 		 */
3971 		lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3972 		    ((flags & GB_LOCK_NOWAIT) ? LK_NOWAIT : LK_SLEEPFAIL);
3973 
3974 		error = BUF_TIMELOCK(bp, lockflags,
3975 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3976 
3977 		/*
3978 		 * If we slept and got the lock we have to restart in case
3979 		 * the buffer changed identities.
3980 		 */
3981 		if (error == ENOLCK)
3982 			goto loop;
3983 		/* We timed out or were interrupted. */
3984 		else if (error != 0)
3985 			return (error);
3986 
3987 foundbuf_fastpath:
3988 		/* If recursed, assume caller knows the rules. */
3989 		if (BUF_LOCKRECURSED(bp))
3990 			goto end;
3991 
3992 		/*
3993 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3994 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3995 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3996 		 * backing VM cache.
3997 		 */
3998 		if (bp->b_flags & B_INVAL)
3999 			bp->b_flags &= ~B_CACHE;
4000 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4001 			bp->b_flags |= B_CACHE;
4002 		if (bp->b_flags & B_MANAGED)
4003 			MPASS(bp->b_qindex == QUEUE_NONE);
4004 		else
4005 			bremfree(bp);
4006 
4007 		/*
4008 		 * check for size inconsistencies for non-VMIO case.
4009 		 */
4010 		if (bp->b_bcount != size) {
4011 			if ((bp->b_flags & B_VMIO) == 0 ||
4012 			    (size > bp->b_kvasize)) {
4013 				if (bp->b_flags & B_DELWRI) {
4014 					bp->b_flags |= B_NOCACHE;
4015 					bwrite(bp);
4016 				} else {
4017 					if (LIST_EMPTY(&bp->b_dep)) {
4018 						bp->b_flags |= B_RELBUF;
4019 						brelse(bp);
4020 					} else {
4021 						bp->b_flags |= B_NOCACHE;
4022 						bwrite(bp);
4023 					}
4024 				}
4025 				goto loop;
4026 			}
4027 		}
4028 
4029 		/*
4030 		 * Handle the case of unmapped buffer which should
4031 		 * become mapped, or the buffer for which KVA
4032 		 * reservation is requested.
4033 		 */
4034 		bp_unmapped_get_kva(bp, blkno, size, flags);
4035 
4036 		/*
4037 		 * If the size is inconsistent in the VMIO case, we can resize
4038 		 * the buffer.  This might lead to B_CACHE getting set or
4039 		 * cleared.  If the size has not changed, B_CACHE remains
4040 		 * unchanged from its previous state.
4041 		 */
4042 		allocbuf(bp, size);
4043 
4044 		KASSERT(bp->b_offset != NOOFFSET,
4045 		    ("getblk: no buffer offset"));
4046 
4047 		/*
4048 		 * A buffer with B_DELWRI set and B_CACHE clear must
4049 		 * be committed before we can return the buffer in
4050 		 * order to prevent the caller from issuing a read
4051 		 * ( due to B_CACHE not being set ) and overwriting
4052 		 * it.
4053 		 *
4054 		 * Most callers, including NFS and FFS, need this to
4055 		 * operate properly either because they assume they
4056 		 * can issue a read if B_CACHE is not set, or because
4057 		 * ( for example ) an uncached B_DELWRI might loop due
4058 		 * to softupdates re-dirtying the buffer.  In the latter
4059 		 * case, B_CACHE is set after the first write completes,
4060 		 * preventing further loops.
4061 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
4062 		 * above while extending the buffer, we cannot allow the
4063 		 * buffer to remain with B_CACHE set after the write
4064 		 * completes or it will represent a corrupt state.  To
4065 		 * deal with this we set B_NOCACHE to scrap the buffer
4066 		 * after the write.
4067 		 *
4068 		 * We might be able to do something fancy, like setting
4069 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
4070 		 * so the below call doesn't set B_CACHE, but that gets real
4071 		 * confusing.  This is much easier.
4072 		 */
4073 
4074 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4075 			bp->b_flags |= B_NOCACHE;
4076 			bwrite(bp);
4077 			goto loop;
4078 		}
4079 		bp->b_flags &= ~B_DONE;
4080 	} else {
4081 		/*
4082 		 * Buffer is not in-core, create new buffer.  The buffer
4083 		 * returned by getnewbuf() is locked.  Note that the returned
4084 		 * buffer is also considered valid (not marked B_INVAL).
4085 		 */
4086 		BO_RUNLOCK(bo);
4087 newbuf_unlocked:
4088 		/*
4089 		 * If the user does not want us to create the buffer, bail out
4090 		 * here.
4091 		 */
4092 		if (flags & GB_NOCREAT)
4093 			return (EEXIST);
4094 
4095 		bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4096 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4097 		offset = blkno * bsize;
4098 		vmio = vp->v_object != NULL;
4099 		if (vmio) {
4100 			maxsize = size + (offset & PAGE_MASK);
4101 		} else {
4102 			maxsize = size;
4103 			/* Do not allow non-VMIO notmapped buffers. */
4104 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4105 		}
4106 		maxsize = imax(maxsize, bsize);
4107 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4108 		    !vn_isdisk(vp)) {
4109 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4110 			KASSERT(error != EOPNOTSUPP,
4111 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4112 			    vp));
4113 			if (error != 0)
4114 				return (error);
4115 			if (d_blkno == -1)
4116 				return (EJUSTRETURN);
4117 		}
4118 
4119 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4120 		if (bp == NULL) {
4121 			if (slpflag || slptimeo)
4122 				return (ETIMEDOUT);
4123 			/*
4124 			 * XXX This is here until the sleep path is diagnosed
4125 			 * enough to work under very low memory conditions.
4126 			 *
4127 			 * There's an issue on low memory, 4BSD+non-preempt
4128 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4129 			 * exhaustion occurs without sleeping for buffer
4130 			 * reclaimation.  This just sticks in a loop and
4131 			 * constantly attempts to allocate a buffer, which
4132 			 * hits exhaustion and tries to wakeup bufdaemon.
4133 			 * This never happens because we never yield.
4134 			 *
4135 			 * The real solution is to identify and fix these cases
4136 			 * so we aren't effectively busy-waiting in a loop
4137 			 * until the reclaimation path has cycles to run.
4138 			 */
4139 			kern_yield(PRI_USER);
4140 			goto loop;
4141 		}
4142 
4143 		/*
4144 		 * This code is used to make sure that a buffer is not
4145 		 * created while the getnewbuf routine is blocked.
4146 		 * This can be a problem whether the vnode is locked or not.
4147 		 * If the buffer is created out from under us, we have to
4148 		 * throw away the one we just created.
4149 		 *
4150 		 * Note: this must occur before we associate the buffer
4151 		 * with the vp especially considering limitations in
4152 		 * the splay tree implementation when dealing with duplicate
4153 		 * lblkno's.
4154 		 */
4155 		BO_LOCK(bo);
4156 		if (gbincore(bo, blkno)) {
4157 			BO_UNLOCK(bo);
4158 			bp->b_flags |= B_INVAL;
4159 			bufspace_release(bufdomain(bp), maxsize);
4160 			brelse(bp);
4161 			goto loop;
4162 		}
4163 
4164 		/*
4165 		 * Insert the buffer into the hash, so that it can
4166 		 * be found by incore.
4167 		 */
4168 		bp->b_lblkno = blkno;
4169 		bp->b_blkno = d_blkno;
4170 		bp->b_offset = offset;
4171 		bgetvp(vp, bp);
4172 		BO_UNLOCK(bo);
4173 
4174 		/*
4175 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4176 		 * buffer size starts out as 0, B_CACHE will be set by
4177 		 * allocbuf() for the VMIO case prior to it testing the
4178 		 * backing store for validity.
4179 		 */
4180 
4181 		if (vmio) {
4182 			bp->b_flags |= B_VMIO;
4183 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4184 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4185 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4186 		} else {
4187 			bp->b_flags &= ~B_VMIO;
4188 			KASSERT(bp->b_bufobj->bo_object == NULL,
4189 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4190 			    bp, bp->b_bufobj->bo_object));
4191 			BUF_CHECK_MAPPED(bp);
4192 		}
4193 
4194 		allocbuf(bp, size);
4195 		bufspace_release(bufdomain(bp), maxsize);
4196 		bp->b_flags &= ~B_DONE;
4197 	}
4198 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4199 end:
4200 	buf_track(bp, __func__);
4201 	KASSERT(bp->b_bufobj == bo,
4202 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4203 	*bpp = bp;
4204 	return (0);
4205 }
4206 
4207 /*
4208  * Get an empty, disassociated buffer of given size.  The buffer is initially
4209  * set to B_INVAL.
4210  */
4211 struct buf *
4212 geteblk(int size, int flags)
4213 {
4214 	struct buf *bp;
4215 	int maxsize;
4216 
4217 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4218 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4219 		if ((flags & GB_NOWAIT_BD) &&
4220 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4221 			return (NULL);
4222 	}
4223 	allocbuf(bp, size);
4224 	bufspace_release(bufdomain(bp), maxsize);
4225 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4226 	return (bp);
4227 }
4228 
4229 /*
4230  * Truncate the backing store for a non-vmio buffer.
4231  */
4232 static void
4233 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4234 {
4235 
4236 	if (bp->b_flags & B_MALLOC) {
4237 		/*
4238 		 * malloced buffers are not shrunk
4239 		 */
4240 		if (newbsize == 0) {
4241 			bufmallocadjust(bp, 0);
4242 			free(bp->b_data, M_BIOBUF);
4243 			bp->b_data = bp->b_kvabase;
4244 			bp->b_flags &= ~B_MALLOC;
4245 		}
4246 		return;
4247 	}
4248 	vm_hold_free_pages(bp, newbsize);
4249 	bufspace_adjust(bp, newbsize);
4250 }
4251 
4252 /*
4253  * Extend the backing for a non-VMIO buffer.
4254  */
4255 static void
4256 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4257 {
4258 	caddr_t origbuf;
4259 	int origbufsize;
4260 
4261 	/*
4262 	 * We only use malloced memory on the first allocation.
4263 	 * and revert to page-allocated memory when the buffer
4264 	 * grows.
4265 	 *
4266 	 * There is a potential smp race here that could lead
4267 	 * to bufmallocspace slightly passing the max.  It
4268 	 * is probably extremely rare and not worth worrying
4269 	 * over.
4270 	 */
4271 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4272 	    bufmallocspace < maxbufmallocspace) {
4273 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4274 		bp->b_flags |= B_MALLOC;
4275 		bufmallocadjust(bp, newbsize);
4276 		return;
4277 	}
4278 
4279 	/*
4280 	 * If the buffer is growing on its other-than-first
4281 	 * allocation then we revert to the page-allocation
4282 	 * scheme.
4283 	 */
4284 	origbuf = NULL;
4285 	origbufsize = 0;
4286 	if (bp->b_flags & B_MALLOC) {
4287 		origbuf = bp->b_data;
4288 		origbufsize = bp->b_bufsize;
4289 		bp->b_data = bp->b_kvabase;
4290 		bufmallocadjust(bp, 0);
4291 		bp->b_flags &= ~B_MALLOC;
4292 		newbsize = round_page(newbsize);
4293 	}
4294 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4295 	    (vm_offset_t) bp->b_data + newbsize);
4296 	if (origbuf != NULL) {
4297 		bcopy(origbuf, bp->b_data, origbufsize);
4298 		free(origbuf, M_BIOBUF);
4299 	}
4300 	bufspace_adjust(bp, newbsize);
4301 }
4302 
4303 /*
4304  * This code constitutes the buffer memory from either anonymous system
4305  * memory (in the case of non-VMIO operations) or from an associated
4306  * VM object (in the case of VMIO operations).  This code is able to
4307  * resize a buffer up or down.
4308  *
4309  * Note that this code is tricky, and has many complications to resolve
4310  * deadlock or inconsistent data situations.  Tread lightly!!!
4311  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4312  * the caller.  Calling this code willy nilly can result in the loss of data.
4313  *
4314  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4315  * B_CACHE for the non-VMIO case.
4316  */
4317 int
4318 allocbuf(struct buf *bp, int size)
4319 {
4320 	int newbsize;
4321 
4322 	if (bp->b_bcount == size)
4323 		return (1);
4324 
4325 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4326 		panic("allocbuf: buffer too small");
4327 
4328 	newbsize = roundup2(size, DEV_BSIZE);
4329 	if ((bp->b_flags & B_VMIO) == 0) {
4330 		if ((bp->b_flags & B_MALLOC) == 0)
4331 			newbsize = round_page(newbsize);
4332 		/*
4333 		 * Just get anonymous memory from the kernel.  Don't
4334 		 * mess with B_CACHE.
4335 		 */
4336 		if (newbsize < bp->b_bufsize)
4337 			vfs_nonvmio_truncate(bp, newbsize);
4338 		else if (newbsize > bp->b_bufsize)
4339 			vfs_nonvmio_extend(bp, newbsize);
4340 	} else {
4341 		int desiredpages;
4342 
4343 		desiredpages = (size == 0) ? 0 :
4344 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4345 
4346 		if (bp->b_flags & B_MALLOC)
4347 			panic("allocbuf: VMIO buffer can't be malloced");
4348 		/*
4349 		 * Set B_CACHE initially if buffer is 0 length or will become
4350 		 * 0-length.
4351 		 */
4352 		if (size == 0 || bp->b_bufsize == 0)
4353 			bp->b_flags |= B_CACHE;
4354 
4355 		if (newbsize < bp->b_bufsize)
4356 			vfs_vmio_truncate(bp, desiredpages);
4357 		/* XXX This looks as if it should be newbsize > b_bufsize */
4358 		else if (size > bp->b_bcount)
4359 			vfs_vmio_extend(bp, desiredpages, size);
4360 		bufspace_adjust(bp, newbsize);
4361 	}
4362 	bp->b_bcount = size;		/* requested buffer size. */
4363 	return (1);
4364 }
4365 
4366 extern int inflight_transient_maps;
4367 
4368 static struct bio_queue nondump_bios;
4369 
4370 void
4371 biodone(struct bio *bp)
4372 {
4373 	struct mtx *mtxp;
4374 	void (*done)(struct bio *);
4375 	vm_offset_t start, end;
4376 
4377 	biotrack(bp, __func__);
4378 
4379 	/*
4380 	 * Avoid completing I/O when dumping after a panic since that may
4381 	 * result in a deadlock in the filesystem or pager code.  Note that
4382 	 * this doesn't affect dumps that were started manually since we aim
4383 	 * to keep the system usable after it has been resumed.
4384 	 */
4385 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4386 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4387 		return;
4388 	}
4389 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4390 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4391 		bp->bio_flags |= BIO_UNMAPPED;
4392 		start = trunc_page((vm_offset_t)bp->bio_data);
4393 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4394 		bp->bio_data = unmapped_buf;
4395 		pmap_qremove(start, atop(end - start));
4396 		vmem_free(transient_arena, start, end - start);
4397 		atomic_add_int(&inflight_transient_maps, -1);
4398 	}
4399 	done = bp->bio_done;
4400 	/*
4401 	 * The check for done == biodone is to allow biodone to be
4402 	 * used as a bio_done routine.
4403 	 */
4404 	if (done == NULL || done == biodone) {
4405 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4406 		mtx_lock(mtxp);
4407 		bp->bio_flags |= BIO_DONE;
4408 		wakeup(bp);
4409 		mtx_unlock(mtxp);
4410 	} else
4411 		done(bp);
4412 }
4413 
4414 /*
4415  * Wait for a BIO to finish.
4416  */
4417 int
4418 biowait(struct bio *bp, const char *wmesg)
4419 {
4420 	struct mtx *mtxp;
4421 
4422 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4423 	mtx_lock(mtxp);
4424 	while ((bp->bio_flags & BIO_DONE) == 0)
4425 		msleep(bp, mtxp, PRIBIO, wmesg, 0);
4426 	mtx_unlock(mtxp);
4427 	if (bp->bio_error != 0)
4428 		return (bp->bio_error);
4429 	if (!(bp->bio_flags & BIO_ERROR))
4430 		return (0);
4431 	return (EIO);
4432 }
4433 
4434 void
4435 biofinish(struct bio *bp, struct devstat *stat, int error)
4436 {
4437 
4438 	if (error) {
4439 		bp->bio_error = error;
4440 		bp->bio_flags |= BIO_ERROR;
4441 	}
4442 	if (stat != NULL)
4443 		devstat_end_transaction_bio(stat, bp);
4444 	biodone(bp);
4445 }
4446 
4447 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4448 void
4449 biotrack_buf(struct bio *bp, const char *location)
4450 {
4451 
4452 	buf_track(bp->bio_track_bp, location);
4453 }
4454 #endif
4455 
4456 /*
4457  *	bufwait:
4458  *
4459  *	Wait for buffer I/O completion, returning error status.  The buffer
4460  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4461  *	error and cleared.
4462  */
4463 int
4464 bufwait(struct buf *bp)
4465 {
4466 	if (bp->b_iocmd == BIO_READ)
4467 		bwait(bp, PRIBIO, "biord");
4468 	else
4469 		bwait(bp, PRIBIO, "biowr");
4470 	if (bp->b_flags & B_EINTR) {
4471 		bp->b_flags &= ~B_EINTR;
4472 		return (EINTR);
4473 	}
4474 	if (bp->b_ioflags & BIO_ERROR) {
4475 		return (bp->b_error ? bp->b_error : EIO);
4476 	} else {
4477 		return (0);
4478 	}
4479 }
4480 
4481 /*
4482  *	bufdone:
4483  *
4484  *	Finish I/O on a buffer, optionally calling a completion function.
4485  *	This is usually called from an interrupt so process blocking is
4486  *	not allowed.
4487  *
4488  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4489  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4490  *	assuming B_INVAL is clear.
4491  *
4492  *	For the VMIO case, we set B_CACHE if the op was a read and no
4493  *	read error occurred, or if the op was a write.  B_CACHE is never
4494  *	set if the buffer is invalid or otherwise uncacheable.
4495  *
4496  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4497  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4498  *	in the biodone routine.
4499  */
4500 void
4501 bufdone(struct buf *bp)
4502 {
4503 	struct bufobj *dropobj;
4504 	void    (*biodone)(struct buf *);
4505 
4506 	buf_track(bp, __func__);
4507 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4508 	dropobj = NULL;
4509 
4510 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4511 
4512 	runningbufwakeup(bp);
4513 	if (bp->b_iocmd == BIO_WRITE)
4514 		dropobj = bp->b_bufobj;
4515 	/* call optional completion function if requested */
4516 	if (bp->b_iodone != NULL) {
4517 		biodone = bp->b_iodone;
4518 		bp->b_iodone = NULL;
4519 		(*biodone) (bp);
4520 		if (dropobj)
4521 			bufobj_wdrop(dropobj);
4522 		return;
4523 	}
4524 	if (bp->b_flags & B_VMIO) {
4525 		/*
4526 		 * Set B_CACHE if the op was a normal read and no error
4527 		 * occurred.  B_CACHE is set for writes in the b*write()
4528 		 * routines.
4529 		 */
4530 		if (bp->b_iocmd == BIO_READ &&
4531 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4532 		    !(bp->b_ioflags & BIO_ERROR))
4533 			bp->b_flags |= B_CACHE;
4534 		vfs_vmio_iodone(bp);
4535 	}
4536 	if (!LIST_EMPTY(&bp->b_dep))
4537 		buf_complete(bp);
4538 	if ((bp->b_flags & B_CKHASH) != 0) {
4539 		KASSERT(bp->b_iocmd == BIO_READ,
4540 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4541 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4542 		(*bp->b_ckhashcalc)(bp);
4543 	}
4544 	/*
4545 	 * For asynchronous completions, release the buffer now. The brelse
4546 	 * will do a wakeup there if necessary - so no need to do a wakeup
4547 	 * here in the async case. The sync case always needs to do a wakeup.
4548 	 */
4549 	if (bp->b_flags & B_ASYNC) {
4550 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4551 		    (bp->b_ioflags & BIO_ERROR))
4552 			brelse(bp);
4553 		else
4554 			bqrelse(bp);
4555 	} else
4556 		bdone(bp);
4557 	if (dropobj)
4558 		bufobj_wdrop(dropobj);
4559 }
4560 
4561 /*
4562  * This routine is called in lieu of iodone in the case of
4563  * incomplete I/O.  This keeps the busy status for pages
4564  * consistent.
4565  */
4566 void
4567 vfs_unbusy_pages(struct buf *bp)
4568 {
4569 	int i;
4570 	vm_object_t obj;
4571 	vm_page_t m;
4572 
4573 	runningbufwakeup(bp);
4574 	if (!(bp->b_flags & B_VMIO))
4575 		return;
4576 
4577 	obj = bp->b_bufobj->bo_object;
4578 	for (i = 0; i < bp->b_npages; i++) {
4579 		m = bp->b_pages[i];
4580 		if (m == bogus_page) {
4581 			m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4582 			if (!m)
4583 				panic("vfs_unbusy_pages: page missing\n");
4584 			bp->b_pages[i] = m;
4585 			if (buf_mapped(bp)) {
4586 				BUF_CHECK_MAPPED(bp);
4587 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4588 				    bp->b_pages, bp->b_npages);
4589 			} else
4590 				BUF_CHECK_UNMAPPED(bp);
4591 		}
4592 		vm_page_sunbusy(m);
4593 	}
4594 	vm_object_pip_wakeupn(obj, bp->b_npages);
4595 }
4596 
4597 /*
4598  * vfs_page_set_valid:
4599  *
4600  *	Set the valid bits in a page based on the supplied offset.   The
4601  *	range is restricted to the buffer's size.
4602  *
4603  *	This routine is typically called after a read completes.
4604  */
4605 static void
4606 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4607 {
4608 	vm_ooffset_t eoff;
4609 
4610 	/*
4611 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4612 	 * page boundary and eoff is not greater than the end of the buffer.
4613 	 * The end of the buffer, in this case, is our file EOF, not the
4614 	 * allocation size of the buffer.
4615 	 */
4616 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4617 	if (eoff > bp->b_offset + bp->b_bcount)
4618 		eoff = bp->b_offset + bp->b_bcount;
4619 
4620 	/*
4621 	 * Set valid range.  This is typically the entire buffer and thus the
4622 	 * entire page.
4623 	 */
4624 	if (eoff > off)
4625 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4626 }
4627 
4628 /*
4629  * vfs_page_set_validclean:
4630  *
4631  *	Set the valid bits and clear the dirty bits in a page based on the
4632  *	supplied offset.   The range is restricted to the buffer's size.
4633  */
4634 static void
4635 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4636 {
4637 	vm_ooffset_t soff, eoff;
4638 
4639 	/*
4640 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4641 	 * page boundary or cross the end of the buffer.  The end of the
4642 	 * buffer, in this case, is our file EOF, not the allocation size
4643 	 * of the buffer.
4644 	 */
4645 	soff = off;
4646 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4647 	if (eoff > bp->b_offset + bp->b_bcount)
4648 		eoff = bp->b_offset + bp->b_bcount;
4649 
4650 	/*
4651 	 * Set valid range.  This is typically the entire buffer and thus the
4652 	 * entire page.
4653 	 */
4654 	if (eoff > soff) {
4655 		vm_page_set_validclean(
4656 		    m,
4657 		   (vm_offset_t) (soff & PAGE_MASK),
4658 		   (vm_offset_t) (eoff - soff)
4659 		);
4660 	}
4661 }
4662 
4663 /*
4664  * Acquire a shared busy on all pages in the buf.
4665  */
4666 void
4667 vfs_busy_pages_acquire(struct buf *bp)
4668 {
4669 	int i;
4670 
4671 	for (i = 0; i < bp->b_npages; i++)
4672 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4673 }
4674 
4675 void
4676 vfs_busy_pages_release(struct buf *bp)
4677 {
4678 	int i;
4679 
4680 	for (i = 0; i < bp->b_npages; i++)
4681 		vm_page_sunbusy(bp->b_pages[i]);
4682 }
4683 
4684 /*
4685  * This routine is called before a device strategy routine.
4686  * It is used to tell the VM system that paging I/O is in
4687  * progress, and treat the pages associated with the buffer
4688  * almost as being exclusive busy.  Also the object paging_in_progress
4689  * flag is handled to make sure that the object doesn't become
4690  * inconsistent.
4691  *
4692  * Since I/O has not been initiated yet, certain buffer flags
4693  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4694  * and should be ignored.
4695  */
4696 void
4697 vfs_busy_pages(struct buf *bp, int clear_modify)
4698 {
4699 	vm_object_t obj;
4700 	vm_ooffset_t foff;
4701 	vm_page_t m;
4702 	int i;
4703 	bool bogus;
4704 
4705 	if (!(bp->b_flags & B_VMIO))
4706 		return;
4707 
4708 	obj = bp->b_bufobj->bo_object;
4709 	foff = bp->b_offset;
4710 	KASSERT(bp->b_offset != NOOFFSET,
4711 	    ("vfs_busy_pages: no buffer offset"));
4712 	if ((bp->b_flags & B_CLUSTER) == 0) {
4713 		vm_object_pip_add(obj, bp->b_npages);
4714 		vfs_busy_pages_acquire(bp);
4715 	}
4716 	if (bp->b_bufsize != 0)
4717 		vfs_setdirty_range(bp);
4718 	bogus = false;
4719 	for (i = 0; i < bp->b_npages; i++) {
4720 		m = bp->b_pages[i];
4721 		vm_page_assert_sbusied(m);
4722 
4723 		/*
4724 		 * When readying a buffer for a read ( i.e
4725 		 * clear_modify == 0 ), it is important to do
4726 		 * bogus_page replacement for valid pages in
4727 		 * partially instantiated buffers.  Partially
4728 		 * instantiated buffers can, in turn, occur when
4729 		 * reconstituting a buffer from its VM backing store
4730 		 * base.  We only have to do this if B_CACHE is
4731 		 * clear ( which causes the I/O to occur in the
4732 		 * first place ).  The replacement prevents the read
4733 		 * I/O from overwriting potentially dirty VM-backed
4734 		 * pages.  XXX bogus page replacement is, uh, bogus.
4735 		 * It may not work properly with small-block devices.
4736 		 * We need to find a better way.
4737 		 */
4738 		if (clear_modify) {
4739 			pmap_remove_write(m);
4740 			vfs_page_set_validclean(bp, foff, m);
4741 		} else if (vm_page_all_valid(m) &&
4742 		    (bp->b_flags & B_CACHE) == 0) {
4743 			bp->b_pages[i] = bogus_page;
4744 			bogus = true;
4745 		}
4746 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4747 	}
4748 	if (bogus && buf_mapped(bp)) {
4749 		BUF_CHECK_MAPPED(bp);
4750 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4751 		    bp->b_pages, bp->b_npages);
4752 	}
4753 }
4754 
4755 /*
4756  *	vfs_bio_set_valid:
4757  *
4758  *	Set the range within the buffer to valid.  The range is
4759  *	relative to the beginning of the buffer, b_offset.  Note that
4760  *	b_offset itself may be offset from the beginning of the first
4761  *	page.
4762  */
4763 void
4764 vfs_bio_set_valid(struct buf *bp, int base, int size)
4765 {
4766 	int i, n;
4767 	vm_page_t m;
4768 
4769 	if (!(bp->b_flags & B_VMIO))
4770 		return;
4771 
4772 	/*
4773 	 * Fixup base to be relative to beginning of first page.
4774 	 * Set initial n to be the maximum number of bytes in the
4775 	 * first page that can be validated.
4776 	 */
4777 	base += (bp->b_offset & PAGE_MASK);
4778 	n = PAGE_SIZE - (base & PAGE_MASK);
4779 
4780 	/*
4781 	 * Busy may not be strictly necessary here because the pages are
4782 	 * unlikely to be fully valid and the vnode lock will synchronize
4783 	 * their access via getpages.  It is grabbed for consistency with
4784 	 * other page validation.
4785 	 */
4786 	vfs_busy_pages_acquire(bp);
4787 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4788 		m = bp->b_pages[i];
4789 		if (n > size)
4790 			n = size;
4791 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4792 		base += n;
4793 		size -= n;
4794 		n = PAGE_SIZE;
4795 	}
4796 	vfs_busy_pages_release(bp);
4797 }
4798 
4799 /*
4800  *	vfs_bio_clrbuf:
4801  *
4802  *	If the specified buffer is a non-VMIO buffer, clear the entire
4803  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4804  *	validate only the previously invalid portions of the buffer.
4805  *	This routine essentially fakes an I/O, so we need to clear
4806  *	BIO_ERROR and B_INVAL.
4807  *
4808  *	Note that while we only theoretically need to clear through b_bcount,
4809  *	we go ahead and clear through b_bufsize.
4810  */
4811 void
4812 vfs_bio_clrbuf(struct buf *bp)
4813 {
4814 	int i, j, mask, sa, ea, slide;
4815 
4816 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4817 		clrbuf(bp);
4818 		return;
4819 	}
4820 	bp->b_flags &= ~B_INVAL;
4821 	bp->b_ioflags &= ~BIO_ERROR;
4822 	vfs_busy_pages_acquire(bp);
4823 	sa = bp->b_offset & PAGE_MASK;
4824 	slide = 0;
4825 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4826 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4827 		ea = slide & PAGE_MASK;
4828 		if (ea == 0)
4829 			ea = PAGE_SIZE;
4830 		if (bp->b_pages[i] == bogus_page)
4831 			continue;
4832 		j = sa / DEV_BSIZE;
4833 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4834 		if ((bp->b_pages[i]->valid & mask) == mask)
4835 			continue;
4836 		if ((bp->b_pages[i]->valid & mask) == 0)
4837 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4838 		else {
4839 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4840 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4841 					pmap_zero_page_area(bp->b_pages[i],
4842 					    sa, DEV_BSIZE);
4843 				}
4844 			}
4845 		}
4846 		vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4847 		    roundup2(ea - sa, DEV_BSIZE));
4848 	}
4849 	vfs_busy_pages_release(bp);
4850 	bp->b_resid = 0;
4851 }
4852 
4853 void
4854 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4855 {
4856 	vm_page_t m;
4857 	int i, n;
4858 
4859 	if (buf_mapped(bp)) {
4860 		BUF_CHECK_MAPPED(bp);
4861 		bzero(bp->b_data + base, size);
4862 	} else {
4863 		BUF_CHECK_UNMAPPED(bp);
4864 		n = PAGE_SIZE - (base & PAGE_MASK);
4865 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4866 			m = bp->b_pages[i];
4867 			if (n > size)
4868 				n = size;
4869 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4870 			base += n;
4871 			size -= n;
4872 			n = PAGE_SIZE;
4873 		}
4874 	}
4875 }
4876 
4877 /*
4878  * Update buffer flags based on I/O request parameters, optionally releasing the
4879  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4880  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4881  * I/O).  Otherwise the buffer is released to the cache.
4882  */
4883 static void
4884 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4885 {
4886 
4887 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4888 	    ("buf %p non-VMIO noreuse", bp));
4889 
4890 	if ((ioflag & IO_DIRECT) != 0)
4891 		bp->b_flags |= B_DIRECT;
4892 	if ((ioflag & IO_EXT) != 0)
4893 		bp->b_xflags |= BX_ALTDATA;
4894 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4895 		bp->b_flags |= B_RELBUF;
4896 		if ((ioflag & IO_NOREUSE) != 0)
4897 			bp->b_flags |= B_NOREUSE;
4898 		if (release)
4899 			brelse(bp);
4900 	} else if (release)
4901 		bqrelse(bp);
4902 }
4903 
4904 void
4905 vfs_bio_brelse(struct buf *bp, int ioflag)
4906 {
4907 
4908 	b_io_dismiss(bp, ioflag, true);
4909 }
4910 
4911 void
4912 vfs_bio_set_flags(struct buf *bp, int ioflag)
4913 {
4914 
4915 	b_io_dismiss(bp, ioflag, false);
4916 }
4917 
4918 /*
4919  * vm_hold_load_pages and vm_hold_free_pages get pages into
4920  * a buffers address space.  The pages are anonymous and are
4921  * not associated with a file object.
4922  */
4923 static void
4924 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4925 {
4926 	vm_offset_t pg;
4927 	vm_page_t p;
4928 	int index;
4929 
4930 	BUF_CHECK_MAPPED(bp);
4931 
4932 	to = round_page(to);
4933 	from = round_page(from);
4934 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4935 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
4936 	KASSERT(to - from <= maxbcachebuf,
4937 	    ("vm_hold_load_pages too large %p %#jx %#jx %u",
4938 	    bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4939 
4940 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4941 		/*
4942 		 * note: must allocate system pages since blocking here
4943 		 * could interfere with paging I/O, no matter which
4944 		 * process we are.
4945 		 */
4946 		p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
4947 		    VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
4948 		pmap_qenter(pg, &p, 1);
4949 		bp->b_pages[index] = p;
4950 	}
4951 	bp->b_npages = index;
4952 }
4953 
4954 /* Return pages associated with this buf to the vm system */
4955 static void
4956 vm_hold_free_pages(struct buf *bp, int newbsize)
4957 {
4958 	vm_offset_t from;
4959 	vm_page_t p;
4960 	int index, newnpages;
4961 
4962 	BUF_CHECK_MAPPED(bp);
4963 
4964 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4965 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4966 	if (bp->b_npages > newnpages)
4967 		pmap_qremove(from, bp->b_npages - newnpages);
4968 	for (index = newnpages; index < bp->b_npages; index++) {
4969 		p = bp->b_pages[index];
4970 		bp->b_pages[index] = NULL;
4971 		vm_page_unwire_noq(p);
4972 		vm_page_free(p);
4973 	}
4974 	bp->b_npages = newnpages;
4975 }
4976 
4977 /*
4978  * Map an IO request into kernel virtual address space.
4979  *
4980  * All requests are (re)mapped into kernel VA space.
4981  * Notice that we use b_bufsize for the size of the buffer
4982  * to be mapped.  b_bcount might be modified by the driver.
4983  *
4984  * Note that even if the caller determines that the address space should
4985  * be valid, a race or a smaller-file mapped into a larger space may
4986  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4987  * check the return value.
4988  *
4989  * This function only works with pager buffers.
4990  */
4991 int
4992 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
4993 {
4994 	vm_prot_t prot;
4995 	int pidx;
4996 
4997 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
4998 	prot = VM_PROT_READ;
4999 	if (bp->b_iocmd == BIO_READ)
5000 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
5001 	pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5002 	    (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5003 	if (pidx < 0)
5004 		return (-1);
5005 	bp->b_bufsize = len;
5006 	bp->b_npages = pidx;
5007 	bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5008 	if (mapbuf || !unmapped_buf_allowed) {
5009 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5010 		bp->b_data = bp->b_kvabase + bp->b_offset;
5011 	} else
5012 		bp->b_data = unmapped_buf;
5013 	return (0);
5014 }
5015 
5016 /*
5017  * Free the io map PTEs associated with this IO operation.
5018  * We also invalidate the TLB entries and restore the original b_addr.
5019  *
5020  * This function only works with pager buffers.
5021  */
5022 void
5023 vunmapbuf(struct buf *bp)
5024 {
5025 	int npages;
5026 
5027 	npages = bp->b_npages;
5028 	if (buf_mapped(bp))
5029 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5030 	vm_page_unhold_pages(bp->b_pages, npages);
5031 
5032 	bp->b_data = unmapped_buf;
5033 }
5034 
5035 void
5036 bdone(struct buf *bp)
5037 {
5038 	struct mtx *mtxp;
5039 
5040 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5041 	mtx_lock(mtxp);
5042 	bp->b_flags |= B_DONE;
5043 	wakeup(bp);
5044 	mtx_unlock(mtxp);
5045 }
5046 
5047 void
5048 bwait(struct buf *bp, u_char pri, const char *wchan)
5049 {
5050 	struct mtx *mtxp;
5051 
5052 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5053 	mtx_lock(mtxp);
5054 	while ((bp->b_flags & B_DONE) == 0)
5055 		msleep(bp, mtxp, pri, wchan, 0);
5056 	mtx_unlock(mtxp);
5057 }
5058 
5059 int
5060 bufsync(struct bufobj *bo, int waitfor)
5061 {
5062 
5063 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5064 }
5065 
5066 void
5067 bufstrategy(struct bufobj *bo, struct buf *bp)
5068 {
5069 	int i __unused;
5070 	struct vnode *vp;
5071 
5072 	vp = bp->b_vp;
5073 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5074 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5075 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5076 	i = VOP_STRATEGY(vp, bp);
5077 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5078 }
5079 
5080 /*
5081  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
5082  */
5083 void
5084 bufobj_init(struct bufobj *bo, void *private)
5085 {
5086 	static volatile int bufobj_cleanq;
5087 
5088         bo->bo_domain =
5089             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5090         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5091         bo->bo_private = private;
5092         TAILQ_INIT(&bo->bo_clean.bv_hd);
5093         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5094 }
5095 
5096 void
5097 bufobj_wrefl(struct bufobj *bo)
5098 {
5099 
5100 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5101 	ASSERT_BO_WLOCKED(bo);
5102 	bo->bo_numoutput++;
5103 }
5104 
5105 void
5106 bufobj_wref(struct bufobj *bo)
5107 {
5108 
5109 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5110 	BO_LOCK(bo);
5111 	bo->bo_numoutput++;
5112 	BO_UNLOCK(bo);
5113 }
5114 
5115 void
5116 bufobj_wdrop(struct bufobj *bo)
5117 {
5118 
5119 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5120 	BO_LOCK(bo);
5121 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5122 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5123 		bo->bo_flag &= ~BO_WWAIT;
5124 		wakeup(&bo->bo_numoutput);
5125 	}
5126 	BO_UNLOCK(bo);
5127 }
5128 
5129 int
5130 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5131 {
5132 	int error;
5133 
5134 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5135 	ASSERT_BO_WLOCKED(bo);
5136 	error = 0;
5137 	while (bo->bo_numoutput) {
5138 		bo->bo_flag |= BO_WWAIT;
5139 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5140 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5141 		if (error)
5142 			break;
5143 	}
5144 	return (error);
5145 }
5146 
5147 /*
5148  * Set bio_data or bio_ma for struct bio from the struct buf.
5149  */
5150 void
5151 bdata2bio(struct buf *bp, struct bio *bip)
5152 {
5153 
5154 	if (!buf_mapped(bp)) {
5155 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5156 		bip->bio_ma = bp->b_pages;
5157 		bip->bio_ma_n = bp->b_npages;
5158 		bip->bio_data = unmapped_buf;
5159 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5160 		bip->bio_flags |= BIO_UNMAPPED;
5161 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5162 		    PAGE_SIZE == bp->b_npages,
5163 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5164 		    (long long)bip->bio_length, bip->bio_ma_n));
5165 	} else {
5166 		bip->bio_data = bp->b_data;
5167 		bip->bio_ma = NULL;
5168 	}
5169 }
5170 
5171 /*
5172  * The MIPS pmap code currently doesn't handle aliased pages.
5173  * The VIPT caches may not handle page aliasing themselves, leading
5174  * to data corruption.
5175  *
5176  * As such, this code makes a system extremely unhappy if said
5177  * system doesn't support unaliasing the above situation in hardware.
5178  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5179  * this feature at build time, so it has to be handled in software.
5180  *
5181  * Once the MIPS pmap/cache code grows to support this function on
5182  * earlier chips, it should be flipped back off.
5183  */
5184 #ifdef	__mips__
5185 static int buf_pager_relbuf = 1;
5186 #else
5187 static int buf_pager_relbuf = 0;
5188 #endif
5189 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5190     &buf_pager_relbuf, 0,
5191     "Make buffer pager release buffers after reading");
5192 
5193 /*
5194  * The buffer pager.  It uses buffer reads to validate pages.
5195  *
5196  * In contrast to the generic local pager from vm/vnode_pager.c, this
5197  * pager correctly and easily handles volumes where the underlying
5198  * device block size is greater than the machine page size.  The
5199  * buffer cache transparently extends the requested page run to be
5200  * aligned at the block boundary, and does the necessary bogus page
5201  * replacements in the addends to avoid obliterating already valid
5202  * pages.
5203  *
5204  * The only non-trivial issue is that the exclusive busy state for
5205  * pages, which is assumed by the vm_pager_getpages() interface, is
5206  * incompatible with the VMIO buffer cache's desire to share-busy the
5207  * pages.  This function performs a trivial downgrade of the pages'
5208  * state before reading buffers, and a less trivial upgrade from the
5209  * shared-busy to excl-busy state after the read.
5210  */
5211 int
5212 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5213     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5214     vbg_get_blksize_t get_blksize)
5215 {
5216 	vm_page_t m;
5217 	vm_object_t object;
5218 	struct buf *bp;
5219 	struct mount *mp;
5220 	daddr_t lbn, lbnp;
5221 	vm_ooffset_t la, lb, poff, poffe;
5222 	long bo_bs, bsize;
5223 	int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5224 	bool redo, lpart;
5225 
5226 	object = vp->v_object;
5227 	mp = vp->v_mount;
5228 	error = 0;
5229 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5230 	if (la >= object->un_pager.vnp.vnp_size)
5231 		return (VM_PAGER_BAD);
5232 
5233 	/*
5234 	 * Change the meaning of la from where the last requested page starts
5235 	 * to where it ends, because that's the end of the requested region
5236 	 * and the start of the potential read-ahead region.
5237 	 */
5238 	la += PAGE_SIZE;
5239 	lpart = la > object->un_pager.vnp.vnp_size;
5240 	error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5241 	    &bo_bs);
5242 	if (error != 0)
5243 		return (VM_PAGER_ERROR);
5244 
5245 	/*
5246 	 * Calculate read-ahead, behind and total pages.
5247 	 */
5248 	pgsin = count;
5249 	lb = IDX_TO_OFF(ma[0]->pindex);
5250 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5251 	pgsin += pgsin_b;
5252 	if (rbehind != NULL)
5253 		*rbehind = pgsin_b;
5254 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5255 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5256 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5257 		    PAGE_SIZE) - la);
5258 	pgsin += pgsin_a;
5259 	if (rahead != NULL)
5260 		*rahead = pgsin_a;
5261 	VM_CNT_INC(v_vnodein);
5262 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5263 
5264 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5265 	    != 0) ? GB_UNMAPPED : 0;
5266 again:
5267 	for (i = 0; i < count; i++) {
5268 		if (ma[i] != bogus_page)
5269 			vm_page_busy_downgrade(ma[i]);
5270 	}
5271 
5272 	lbnp = -1;
5273 	for (i = 0; i < count; i++) {
5274 		m = ma[i];
5275 		if (m == bogus_page)
5276 			continue;
5277 
5278 		/*
5279 		 * Pages are shared busy and the object lock is not
5280 		 * owned, which together allow for the pages'
5281 		 * invalidation.  The racy test for validity avoids
5282 		 * useless creation of the buffer for the most typical
5283 		 * case when invalidation is not used in redo or for
5284 		 * parallel read.  The shared->excl upgrade loop at
5285 		 * the end of the function catches the race in a
5286 		 * reliable way (protected by the object lock).
5287 		 */
5288 		if (vm_page_all_valid(m))
5289 			continue;
5290 
5291 		poff = IDX_TO_OFF(m->pindex);
5292 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5293 		for (; poff < poffe; poff += bsize) {
5294 			lbn = get_lblkno(vp, poff);
5295 			if (lbn == lbnp)
5296 				goto next_page;
5297 			lbnp = lbn;
5298 
5299 			error = get_blksize(vp, lbn, &bsize);
5300 			if (error == 0)
5301 				error = bread_gb(vp, lbn, bsize,
5302 				    curthread->td_ucred, br_flags, &bp);
5303 			if (error != 0)
5304 				goto end_pages;
5305 			if (bp->b_rcred == curthread->td_ucred) {
5306 				crfree(bp->b_rcred);
5307 				bp->b_rcred = NOCRED;
5308 			}
5309 			if (LIST_EMPTY(&bp->b_dep)) {
5310 				/*
5311 				 * Invalidation clears m->valid, but
5312 				 * may leave B_CACHE flag if the
5313 				 * buffer existed at the invalidation
5314 				 * time.  In this case, recycle the
5315 				 * buffer to do real read on next
5316 				 * bread() after redo.
5317 				 *
5318 				 * Otherwise B_RELBUF is not strictly
5319 				 * necessary, enable to reduce buf
5320 				 * cache pressure.
5321 				 */
5322 				if (buf_pager_relbuf ||
5323 				    !vm_page_all_valid(m))
5324 					bp->b_flags |= B_RELBUF;
5325 
5326 				bp->b_flags &= ~B_NOCACHE;
5327 				brelse(bp);
5328 			} else {
5329 				bqrelse(bp);
5330 			}
5331 		}
5332 		KASSERT(1 /* racy, enable for debugging */ ||
5333 		    vm_page_all_valid(m) || i == count - 1,
5334 		    ("buf %d %p invalid", i, m));
5335 		if (i == count - 1 && lpart) {
5336 			if (!vm_page_none_valid(m) &&
5337 			    !vm_page_all_valid(m))
5338 				vm_page_zero_invalid(m, TRUE);
5339 		}
5340 next_page:;
5341 	}
5342 end_pages:
5343 
5344 	redo = false;
5345 	for (i = 0; i < count; i++) {
5346 		if (ma[i] == bogus_page)
5347 			continue;
5348 		if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5349 			vm_page_sunbusy(ma[i]);
5350 			ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5351 			    VM_ALLOC_NORMAL);
5352 		}
5353 
5354 		/*
5355 		 * Since the pages were only sbusy while neither the
5356 		 * buffer nor the object lock was held by us, or
5357 		 * reallocated while vm_page_grab() slept for busy
5358 		 * relinguish, they could have been invalidated.
5359 		 * Recheck the valid bits and re-read as needed.
5360 		 *
5361 		 * Note that the last page is made fully valid in the
5362 		 * read loop, and partial validity for the page at
5363 		 * index count - 1 could mean that the page was
5364 		 * invalidated or removed, so we must restart for
5365 		 * safety as well.
5366 		 */
5367 		if (!vm_page_all_valid(ma[i]))
5368 			redo = true;
5369 	}
5370 	if (redo && error == 0)
5371 		goto again;
5372 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5373 }
5374 
5375 #include "opt_ddb.h"
5376 #ifdef DDB
5377 #include <ddb/ddb.h>
5378 
5379 /* DDB command to show buffer data */
5380 DB_SHOW_COMMAND(buffer, db_show_buffer)
5381 {
5382 	/* get args */
5383 	struct buf *bp = (struct buf *)addr;
5384 #ifdef FULL_BUF_TRACKING
5385 	uint32_t i, j;
5386 #endif
5387 
5388 	if (!have_addr) {
5389 		db_printf("usage: show buffer <addr>\n");
5390 		return;
5391 	}
5392 
5393 	db_printf("buf at %p\n", bp);
5394 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5395 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5396 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5397 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5398 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5399 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5400 	db_printf(
5401 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5402 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5403 	    "b_vp = %p, b_dep = %p\n",
5404 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5405 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5406 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5407 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5408 	    bp->b_kvabase, bp->b_kvasize);
5409 	if (bp->b_npages) {
5410 		int i;
5411 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5412 		for (i = 0; i < bp->b_npages; i++) {
5413 			vm_page_t m;
5414 			m = bp->b_pages[i];
5415 			if (m != NULL)
5416 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5417 				    (u_long)m->pindex,
5418 				    (u_long)VM_PAGE_TO_PHYS(m));
5419 			else
5420 				db_printf("( ??? )");
5421 			if ((i + 1) < bp->b_npages)
5422 				db_printf(",");
5423 		}
5424 		db_printf("\n");
5425 	}
5426 	BUF_LOCKPRINTINFO(bp);
5427 #if defined(FULL_BUF_TRACKING)
5428 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5429 
5430 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5431 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5432 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5433 			continue;
5434 		db_printf(" %2u: %s\n", j,
5435 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5436 	}
5437 #elif defined(BUF_TRACKING)
5438 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5439 #endif
5440 	db_printf(" ");
5441 }
5442 
5443 DB_SHOW_COMMAND(bufqueues, bufqueues)
5444 {
5445 	struct bufdomain *bd;
5446 	struct buf *bp;
5447 	long total;
5448 	int i, j, cnt;
5449 
5450 	db_printf("bqempty: %d\n", bqempty.bq_len);
5451 
5452 	for (i = 0; i < buf_domains; i++) {
5453 		bd = &bdomain[i];
5454 		db_printf("Buf domain %d\n", i);
5455 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5456 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5457 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5458 		db_printf("\n");
5459 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5460 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5461 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5462 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5463 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5464 		db_printf("\n");
5465 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5466 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5467 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5468 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5469 		db_printf("\n");
5470 		total = 0;
5471 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5472 			total += bp->b_bufsize;
5473 		db_printf("\tcleanq count\t%d (%ld)\n",
5474 		    bd->bd_cleanq->bq_len, total);
5475 		total = 0;
5476 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5477 			total += bp->b_bufsize;
5478 		db_printf("\tdirtyq count\t%d (%ld)\n",
5479 		    bd->bd_dirtyq.bq_len, total);
5480 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5481 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5482 		db_printf("\tCPU ");
5483 		for (j = 0; j <= mp_maxid; j++)
5484 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5485 		db_printf("\n");
5486 		cnt = 0;
5487 		total = 0;
5488 		for (j = 0; j < nbuf; j++) {
5489 			bp = nbufp(j);
5490 			if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5491 				cnt++;
5492 				total += bp->b_bufsize;
5493 			}
5494 		}
5495 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5496 		cnt = 0;
5497 		total = 0;
5498 		for (j = 0; j < nbuf; j++) {
5499 			bp = nbufp(j);
5500 			if (bp->b_domain == i) {
5501 				cnt++;
5502 				total += bp->b_bufsize;
5503 			}
5504 		}
5505 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5506 	}
5507 }
5508 
5509 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5510 {
5511 	struct buf *bp;
5512 	int i;
5513 
5514 	for (i = 0; i < nbuf; i++) {
5515 		bp = nbufp(i);
5516 		if (BUF_ISLOCKED(bp)) {
5517 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5518 			db_printf("\n");
5519 			if (db_pager_quit)
5520 				break;
5521 		}
5522 	}
5523 }
5524 
5525 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5526 {
5527 	struct vnode *vp;
5528 	struct buf *bp;
5529 
5530 	if (!have_addr) {
5531 		db_printf("usage: show vnodebufs <addr>\n");
5532 		return;
5533 	}
5534 	vp = (struct vnode *)addr;
5535 	db_printf("Clean buffers:\n");
5536 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5537 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5538 		db_printf("\n");
5539 	}
5540 	db_printf("Dirty buffers:\n");
5541 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5542 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5543 		db_printf("\n");
5544 	}
5545 }
5546 
5547 DB_COMMAND(countfreebufs, db_coundfreebufs)
5548 {
5549 	struct buf *bp;
5550 	int i, used = 0, nfree = 0;
5551 
5552 	if (have_addr) {
5553 		db_printf("usage: countfreebufs\n");
5554 		return;
5555 	}
5556 
5557 	for (i = 0; i < nbuf; i++) {
5558 		bp = nbufp(i);
5559 		if (bp->b_qindex == QUEUE_EMPTY)
5560 			nfree++;
5561 		else
5562 			used++;
5563 	}
5564 
5565 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5566 	    nfree + used);
5567 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5568 }
5569 #endif /* DDB */
5570