xref: /freebsd/sys/kern/vfs_bio.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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/param.h>
48 #include <sys/systm.h>
49 #include <sys/asan.h>
50 #include <sys/bio.h>
51 #include <sys/bitset.h>
52 #include <sys/boottrace.h>
53 #include <sys/buf.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/devicestat.h>
57 #include <sys/eventhandler.h>
58 #include <sys/fail.h>
59 #include <sys/ktr.h>
60 #include <sys/limits.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/memdesc.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/pctrie.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/sched.h>
75 #include <sys/smp.h>
76 #include <sys/sysctl.h>
77 #include <sys/syscallsubr.h>
78 #include <sys/vmem.h>
79 #include <sys/vmmeter.h>
80 #include <sys/vnode.h>
81 #include <sys/watchdog.h>
82 #include <geom/geom.h>
83 #include <vm/vm.h>
84 #include <vm/vm_param.h>
85 #include <vm/vm_kern.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_pager.h>
90 #include <vm/vm_extern.h>
91 #include <vm/vm_map.h>
92 #include <vm/swap_pager.h>
93 
94 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
95 
96 struct	bio_ops bioops;		/* I/O operation notification */
97 
98 struct	buf_ops buf_ops_bio = {
99 	.bop_name	=	"buf_ops_bio",
100 	.bop_write	=	bufwrite,
101 	.bop_strategy	=	bufstrategy,
102 	.bop_sync	=	bufsync,
103 	.bop_bdflush	=	bufbdflush,
104 };
105 
106 struct bufqueue {
107 	struct mtx_padalign	bq_lock;
108 	TAILQ_HEAD(, buf)	bq_queue;
109 	uint8_t			bq_index;
110 	uint16_t		bq_subqueue;
111 	int			bq_len;
112 } __aligned(CACHE_LINE_SIZE);
113 
114 #define	BQ_LOCKPTR(bq)		(&(bq)->bq_lock)
115 #define	BQ_LOCK(bq)		mtx_lock(BQ_LOCKPTR((bq)))
116 #define	BQ_UNLOCK(bq)		mtx_unlock(BQ_LOCKPTR((bq)))
117 #define	BQ_ASSERT_LOCKED(bq)	mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
118 
119 struct bufdomain {
120 	struct bufqueue	*bd_subq;
121 	struct bufqueue bd_dirtyq;
122 	struct bufqueue	*bd_cleanq;
123 	struct mtx_padalign bd_run_lock;
124 	/* Constants */
125 	long		bd_maxbufspace;
126 	long		bd_hibufspace;
127 	long 		bd_lobufspace;
128 	long 		bd_bufspacethresh;
129 	int		bd_hifreebuffers;
130 	int		bd_lofreebuffers;
131 	int		bd_hidirtybuffers;
132 	int		bd_lodirtybuffers;
133 	int		bd_dirtybufthresh;
134 	int		bd_lim;
135 	/* atomics */
136 	int		bd_wanted;
137 	bool		bd_shutdown;
138 	int __aligned(CACHE_LINE_SIZE)	bd_numdirtybuffers;
139 	int __aligned(CACHE_LINE_SIZE)	bd_running;
140 	long __aligned(CACHE_LINE_SIZE) bd_bufspace;
141 	int __aligned(CACHE_LINE_SIZE)	bd_freebuffers;
142 } __aligned(CACHE_LINE_SIZE);
143 
144 #define	BD_LOCKPTR(bd)		(&(bd)->bd_cleanq->bq_lock)
145 #define	BD_LOCK(bd)		mtx_lock(BD_LOCKPTR((bd)))
146 #define	BD_UNLOCK(bd)		mtx_unlock(BD_LOCKPTR((bd)))
147 #define	BD_ASSERT_LOCKED(bd)	mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
148 #define	BD_RUN_LOCKPTR(bd)	(&(bd)->bd_run_lock)
149 #define	BD_RUN_LOCK(bd)		mtx_lock(BD_RUN_LOCKPTR((bd)))
150 #define	BD_RUN_UNLOCK(bd)	mtx_unlock(BD_RUN_LOCKPTR((bd)))
151 #define	BD_DOMAIN(bd)		(bd - bdomain)
152 
153 static char *buf;		/* buffer header pool */
154 static struct buf *
155 nbufp(unsigned i)
156 {
157 	return ((struct buf *)(buf + (sizeof(struct buf) +
158 	    sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
159 }
160 
161 caddr_t __read_mostly unmapped_buf;
162 #ifdef INVARIANTS
163 caddr_t	poisoned_buf = (void *)-1;
164 #endif
165 
166 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
167 struct proc *bufdaemonproc;
168 
169 static void vm_hold_free_pages(struct buf *bp, int newbsize);
170 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
171 		vm_offset_t to);
172 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
173 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
174 		vm_page_t m);
175 static void vfs_clean_pages_dirty_buf(struct buf *bp);
176 static void vfs_setdirty_range(struct buf *bp);
177 static void vfs_vmio_invalidate(struct buf *bp);
178 static void vfs_vmio_truncate(struct buf *bp, int npages);
179 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
180 static int vfs_bio_clcheck(struct vnode *vp, int size,
181 		daddr_t lblkno, daddr_t blkno);
182 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
183 		void (*)(struct buf *));
184 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
185 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
186 static void buf_daemon(void);
187 static __inline void bd_wakeup(void);
188 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
189 static void bufkva_reclaim(vmem_t *, int);
190 static void bufkva_free(struct buf *);
191 static int buf_import(void *, void **, int, int, int);
192 static void buf_release(void *, void **, int);
193 static void maxbcachebuf_adjust(void);
194 static inline struct bufdomain *bufdomain(struct buf *);
195 static void bq_remove(struct bufqueue *bq, struct buf *bp);
196 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
197 static int buf_recycle(struct bufdomain *, bool kva);
198 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
199 	    const char *lockname);
200 static void bd_init(struct bufdomain *bd);
201 static int bd_flushall(struct bufdomain *bd);
202 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
203 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
204 
205 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
206 int vmiodirenable = TRUE;
207 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
208     "Use the VM system for directory writes");
209 long runningbufspace;
210 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
211     "Amount of presently outstanding async buffer io");
212 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
213     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
214 static counter_u64_t bufkvaspace;
215 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
216     "Kernel virtual memory used for buffers");
217 static long maxbufspace;
218 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
219     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
220     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
221     "Maximum allowed value of bufspace (including metadata)");
222 static long bufmallocspace;
223 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
224     "Amount of malloced memory for buffers");
225 static long maxbufmallocspace;
226 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
227     0, "Maximum amount of malloced memory for buffers");
228 static long lobufspace;
229 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
230     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
231     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
232     "Minimum amount of buffers we want to have");
233 long hibufspace;
234 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
235     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
236     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
237     "Maximum allowed value of bufspace (excluding metadata)");
238 long bufspacethresh;
239 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
240     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
241     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
242     "Bufspace consumed before waking the daemon to free some");
243 static counter_u64_t buffreekvacnt;
244 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
245     "Number of times we have freed the KVA space from some buffer");
246 static counter_u64_t bufdefragcnt;
247 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
248     "Number of times we have had to repeat buffer allocation to defragment");
249 static long lorunningspace;
250 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
251     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
252     "Minimum preferred space used for in-progress I/O");
253 static long hirunningspace;
254 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
255     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
256     "Maximum amount of space to use for in-progress I/O");
257 int dirtybufferflushes;
258 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
259     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
260 int bdwriteskip;
261 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
262     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
263 int altbufferflushes;
264 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
265     &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
266 static int recursiveflushes;
267 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
268     &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
269 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
270 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
271     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
272     "Number of buffers that are dirty (has unwritten changes) at the moment");
273 static int lodirtybuffers;
274 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
275     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
276     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
277     "How many buffers we want to have free before bufdaemon can sleep");
278 static int hidirtybuffers;
279 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
280     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
281     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
282     "When the number of dirty buffers is considered severe");
283 int dirtybufthresh;
284 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
285     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
286     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
287     "Number of bdwrite to bawrite conversions to clear dirty buffers");
288 static int numfreebuffers;
289 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
290     "Number of free buffers");
291 static int lofreebuffers;
292 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
293     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
294     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
295    "Target number of free buffers");
296 static int hifreebuffers;
297 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
298     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
299     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
300    "Threshold for clean buffer recycling");
301 static counter_u64_t getnewbufcalls;
302 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
303    &getnewbufcalls, "Number of calls to getnewbuf");
304 static counter_u64_t getnewbufrestarts;
305 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
306     &getnewbufrestarts,
307     "Number of times getnewbuf has had to restart a buffer acquisition");
308 static counter_u64_t mappingrestarts;
309 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
310     &mappingrestarts,
311     "Number of times getblk has had to restart a buffer mapping for "
312     "unmapped buffer");
313 static counter_u64_t numbufallocfails;
314 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
315     &numbufallocfails, "Number of times buffer allocations failed");
316 static int flushbufqtarget = 100;
317 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
318     "Amount of work to do in flushbufqueues when helping bufdaemon");
319 static counter_u64_t notbufdflushes;
320 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
321     "Number of dirty buffer flushes done by the bufdaemon helpers");
322 static long barrierwrites;
323 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
324     &barrierwrites, 0, "Number of barrier writes");
325 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed,
326     CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
327     &unmapped_buf_allowed, 0,
328     "Permit the use of the unmapped i/o");
329 int maxbcachebuf = MAXBCACHEBUF;
330 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
331     "Maximum size of a buffer cache block");
332 
333 /*
334  * This lock synchronizes access to bd_request.
335  */
336 static struct mtx_padalign __exclusive_cache_line bdlock;
337 
338 /*
339  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
340  * waitrunningbufspace().
341  */
342 static struct mtx_padalign __exclusive_cache_line rbreqlock;
343 
344 /*
345  * Lock that protects bdirtywait.
346  */
347 static struct mtx_padalign __exclusive_cache_line bdirtylock;
348 
349 /*
350  * bufdaemon shutdown request and sleep channel.
351  */
352 static bool bd_shutdown;
353 
354 /*
355  * Wakeup point for bufdaemon, as well as indicator of whether it is already
356  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
357  * is idling.
358  */
359 static int bd_request;
360 
361 /*
362  * Request for the buf daemon to write more buffers than is indicated by
363  * lodirtybuf.  This may be necessary to push out excess dependencies or
364  * defragment the address space where a simple count of the number of dirty
365  * buffers is insufficient to characterize the demand for flushing them.
366  */
367 static int bd_speedupreq;
368 
369 /*
370  * Synchronization (sleep/wakeup) variable for active buffer space requests.
371  * Set when wait starts, cleared prior to wakeup().
372  * Used in runningbufwakeup() and waitrunningbufspace().
373  */
374 static int runningbufreq;
375 
376 /*
377  * Synchronization for bwillwrite() waiters.
378  */
379 static int bdirtywait;
380 
381 /*
382  * Definitions for the buffer free lists.
383  */
384 #define QUEUE_NONE	0	/* on no queue */
385 #define QUEUE_EMPTY	1	/* empty buffer headers */
386 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
387 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
388 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
389 
390 /* Maximum number of buffer domains. */
391 #define	BUF_DOMAINS	8
392 
393 struct bufdomainset bdlodirty;		/* Domains > lodirty */
394 struct bufdomainset bdhidirty;		/* Domains > hidirty */
395 
396 /* Configured number of clean queues. */
397 static int __read_mostly buf_domains;
398 
399 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
400 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
401 struct bufqueue __exclusive_cache_line bqempty;
402 
403 /*
404  * per-cpu empty buffer cache.
405  */
406 uma_zone_t buf_zone;
407 
408 static int
409 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
410 {
411 	long value;
412 	int error;
413 
414 	value = *(long *)arg1;
415 	error = sysctl_handle_long(oidp, &value, 0, req);
416 	if (error != 0 || req->newptr == NULL)
417 		return (error);
418 	mtx_lock(&rbreqlock);
419 	if (arg1 == &hirunningspace) {
420 		if (value < lorunningspace)
421 			error = EINVAL;
422 		else
423 			hirunningspace = value;
424 	} else {
425 		KASSERT(arg1 == &lorunningspace,
426 		    ("%s: unknown arg1", __func__));
427 		if (value > hirunningspace)
428 			error = EINVAL;
429 		else
430 			lorunningspace = value;
431 	}
432 	mtx_unlock(&rbreqlock);
433 	return (error);
434 }
435 
436 static int
437 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
438 {
439 	int error;
440 	int value;
441 	int i;
442 
443 	value = *(int *)arg1;
444 	error = sysctl_handle_int(oidp, &value, 0, req);
445 	if (error != 0 || req->newptr == NULL)
446 		return (error);
447 	*(int *)arg1 = value;
448 	for (i = 0; i < buf_domains; i++)
449 		*(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
450 		    value / buf_domains;
451 
452 	return (error);
453 }
454 
455 static int
456 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
457 {
458 	long value;
459 	int error;
460 	int i;
461 
462 	value = *(long *)arg1;
463 	error = sysctl_handle_long(oidp, &value, 0, req);
464 	if (error != 0 || req->newptr == NULL)
465 		return (error);
466 	*(long *)arg1 = value;
467 	for (i = 0; i < buf_domains; i++)
468 		*(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
469 		    value / buf_domains;
470 
471 	return (error);
472 }
473 
474 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
475     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
476 static int
477 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
478 {
479 	long lvalue;
480 	int ivalue;
481 	int i;
482 
483 	lvalue = 0;
484 	for (i = 0; i < buf_domains; i++)
485 		lvalue += bdomain[i].bd_bufspace;
486 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
487 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
488 	if (lvalue > INT_MAX)
489 		/* On overflow, still write out a long to trigger ENOMEM. */
490 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
491 	ivalue = lvalue;
492 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
493 }
494 #else
495 static int
496 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
497 {
498 	long lvalue;
499 	int i;
500 
501 	lvalue = 0;
502 	for (i = 0; i < buf_domains; i++)
503 		lvalue += bdomain[i].bd_bufspace;
504 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
505 }
506 #endif
507 
508 static int
509 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
510 {
511 	int value;
512 	int i;
513 
514 	value = 0;
515 	for (i = 0; i < buf_domains; i++)
516 		value += bdomain[i].bd_numdirtybuffers;
517 	return (sysctl_handle_int(oidp, &value, 0, req));
518 }
519 
520 /*
521  *	bdirtywakeup:
522  *
523  *	Wakeup any bwillwrite() waiters.
524  */
525 static void
526 bdirtywakeup(void)
527 {
528 	mtx_lock(&bdirtylock);
529 	if (bdirtywait) {
530 		bdirtywait = 0;
531 		wakeup(&bdirtywait);
532 	}
533 	mtx_unlock(&bdirtylock);
534 }
535 
536 /*
537  *	bd_clear:
538  *
539  *	Clear a domain from the appropriate bitsets when dirtybuffers
540  *	is decremented.
541  */
542 static void
543 bd_clear(struct bufdomain *bd)
544 {
545 
546 	mtx_lock(&bdirtylock);
547 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
548 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
549 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
550 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
551 	mtx_unlock(&bdirtylock);
552 }
553 
554 /*
555  *	bd_set:
556  *
557  *	Set a domain in the appropriate bitsets when dirtybuffers
558  *	is incremented.
559  */
560 static void
561 bd_set(struct bufdomain *bd)
562 {
563 
564 	mtx_lock(&bdirtylock);
565 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
566 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
567 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
568 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
569 	mtx_unlock(&bdirtylock);
570 }
571 
572 /*
573  *	bdirtysub:
574  *
575  *	Decrement the numdirtybuffers count by one and wakeup any
576  *	threads blocked in bwillwrite().
577  */
578 static void
579 bdirtysub(struct buf *bp)
580 {
581 	struct bufdomain *bd;
582 	int num;
583 
584 	bd = bufdomain(bp);
585 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
586 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
587 		bdirtywakeup();
588 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
589 		bd_clear(bd);
590 }
591 
592 /*
593  *	bdirtyadd:
594  *
595  *	Increment the numdirtybuffers count by one and wakeup the buf
596  *	daemon if needed.
597  */
598 static void
599 bdirtyadd(struct buf *bp)
600 {
601 	struct bufdomain *bd;
602 	int num;
603 
604 	/*
605 	 * Only do the wakeup once as we cross the boundary.  The
606 	 * buf daemon will keep running until the condition clears.
607 	 */
608 	bd = bufdomain(bp);
609 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
610 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
611 		bd_wakeup();
612 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
613 		bd_set(bd);
614 }
615 
616 /*
617  *	bufspace_daemon_wakeup:
618  *
619  *	Wakeup the daemons responsible for freeing clean bufs.
620  */
621 static void
622 bufspace_daemon_wakeup(struct bufdomain *bd)
623 {
624 
625 	/*
626 	 * avoid the lock if the daemon is running.
627 	 */
628 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
629 		BD_RUN_LOCK(bd);
630 		atomic_store_int(&bd->bd_running, 1);
631 		wakeup(&bd->bd_running);
632 		BD_RUN_UNLOCK(bd);
633 	}
634 }
635 
636 /*
637  *	bufspace_adjust:
638  *
639  *	Adjust the reported bufspace for a KVA managed buffer, possibly
640  * 	waking any waiters.
641  */
642 static void
643 bufspace_adjust(struct buf *bp, int bufsize)
644 {
645 	struct bufdomain *bd;
646 	long space;
647 	int diff;
648 
649 	KASSERT((bp->b_flags & B_MALLOC) == 0,
650 	    ("bufspace_adjust: malloc buf %p", bp));
651 	bd = bufdomain(bp);
652 	diff = bufsize - bp->b_bufsize;
653 	if (diff < 0) {
654 		atomic_subtract_long(&bd->bd_bufspace, -diff);
655 	} else if (diff > 0) {
656 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
657 		/* Wake up the daemon on the transition. */
658 		if (space < bd->bd_bufspacethresh &&
659 		    space + diff >= bd->bd_bufspacethresh)
660 			bufspace_daemon_wakeup(bd);
661 	}
662 	bp->b_bufsize = bufsize;
663 }
664 
665 /*
666  *	bufspace_reserve:
667  *
668  *	Reserve bufspace before calling allocbuf().  metadata has a
669  *	different space limit than data.
670  */
671 static int
672 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
673 {
674 	long limit, new;
675 	long space;
676 
677 	if (metadata)
678 		limit = bd->bd_maxbufspace;
679 	else
680 		limit = bd->bd_hibufspace;
681 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
682 	new = space + size;
683 	if (new > limit) {
684 		atomic_subtract_long(&bd->bd_bufspace, size);
685 		return (ENOSPC);
686 	}
687 
688 	/* Wake up the daemon on the transition. */
689 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
690 		bufspace_daemon_wakeup(bd);
691 
692 	return (0);
693 }
694 
695 /*
696  *	bufspace_release:
697  *
698  *	Release reserved bufspace after bufspace_adjust() has consumed it.
699  */
700 static void
701 bufspace_release(struct bufdomain *bd, int size)
702 {
703 
704 	atomic_subtract_long(&bd->bd_bufspace, size);
705 }
706 
707 /*
708  *	bufspace_wait:
709  *
710  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
711  *	supplied.  bd_wanted must be set prior to polling for space.  The
712  *	operation must be re-tried on return.
713  */
714 static void
715 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
716     int slpflag, int slptimeo)
717 {
718 	struct thread *td;
719 	int error, fl, norunbuf;
720 
721 	if ((gbflags & GB_NOWAIT_BD) != 0)
722 		return;
723 
724 	td = curthread;
725 	BD_LOCK(bd);
726 	while (bd->bd_wanted) {
727 		if (vp != NULL && vp->v_type != VCHR &&
728 		    (td->td_pflags & TDP_BUFNEED) == 0) {
729 			BD_UNLOCK(bd);
730 			/*
731 			 * getblk() is called with a vnode locked, and
732 			 * some majority of the dirty buffers may as
733 			 * well belong to the vnode.  Flushing the
734 			 * buffers there would make a progress that
735 			 * cannot be achieved by the buf_daemon, that
736 			 * cannot lock the vnode.
737 			 */
738 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
739 			    (td->td_pflags & TDP_NORUNNINGBUF);
740 
741 			/*
742 			 * Play bufdaemon.  The getnewbuf() function
743 			 * may be called while the thread owns lock
744 			 * for another dirty buffer for the same
745 			 * vnode, which makes it impossible to use
746 			 * VOP_FSYNC() there, due to the buffer lock
747 			 * recursion.
748 			 */
749 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
750 			fl = buf_flush(vp, bd, flushbufqtarget);
751 			td->td_pflags &= norunbuf;
752 			BD_LOCK(bd);
753 			if (fl != 0)
754 				continue;
755 			if (bd->bd_wanted == 0)
756 				break;
757 		}
758 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
759 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
760 		if (error != 0)
761 			break;
762 	}
763 	BD_UNLOCK(bd);
764 }
765 
766 static void
767 bufspace_daemon_shutdown(void *arg, int howto __unused)
768 {
769 	struct bufdomain *bd = arg;
770 	int error;
771 
772 	if (KERNEL_PANICKED())
773 		return;
774 
775 	BD_RUN_LOCK(bd);
776 	bd->bd_shutdown = true;
777 	wakeup(&bd->bd_running);
778 	error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0,
779 	    "bufspace_shutdown", 60 * hz);
780 	BD_RUN_UNLOCK(bd);
781 	if (error != 0)
782 		printf("bufspacedaemon wait error: %d\n", error);
783 }
784 
785 /*
786  *	bufspace_daemon:
787  *
788  *	buffer space management daemon.  Tries to maintain some marginal
789  *	amount of free buffer space so that requesting processes neither
790  *	block nor work to reclaim buffers.
791  */
792 static void
793 bufspace_daemon(void *arg)
794 {
795 	struct bufdomain *bd = arg;
796 
797 	EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd,
798 	    SHUTDOWN_PRI_LAST + 100);
799 
800 	BD_RUN_LOCK(bd);
801 	while (!bd->bd_shutdown) {
802 		BD_RUN_UNLOCK(bd);
803 
804 		/*
805 		 * Free buffers from the clean queue until we meet our
806 		 * targets.
807 		 *
808 		 * Theory of operation:  The buffer cache is most efficient
809 		 * when some free buffer headers and space are always
810 		 * available to getnewbuf().  This daemon attempts to prevent
811 		 * the excessive blocking and synchronization associated
812 		 * with shortfall.  It goes through three phases according
813 		 * demand:
814 		 *
815 		 * 1)	The daemon wakes up voluntarily once per-second
816 		 *	during idle periods when the counters are below
817 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
818 		 *
819 		 * 2)	The daemon wakes up as we cross the thresholds
820 		 *	ahead of any potential blocking.  This may bounce
821 		 *	slightly according to the rate of consumption and
822 		 *	release.
823 		 *
824 		 * 3)	The daemon and consumers are starved for working
825 		 *	clean buffers.  This is the 'bufspace' sleep below
826 		 *	which will inefficiently trade bufs with bqrelse
827 		 *	until we return to condition 2.
828 		 */
829 		while (bd->bd_bufspace > bd->bd_lobufspace ||
830 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
831 			if (buf_recycle(bd, false) != 0) {
832 				if (bd_flushall(bd))
833 					continue;
834 				/*
835 				 * Speedup dirty if we've run out of clean
836 				 * buffers.  This is possible in particular
837 				 * because softdep may held many bufs locked
838 				 * pending writes to other bufs which are
839 				 * marked for delayed write, exhausting
840 				 * clean space until they are written.
841 				 */
842 				bd_speedup();
843 				BD_LOCK(bd);
844 				if (bd->bd_wanted) {
845 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
846 					    PRIBIO|PDROP, "bufspace", hz/10);
847 				} else
848 					BD_UNLOCK(bd);
849 			}
850 			maybe_yield();
851 		}
852 
853 		/*
854 		 * Re-check our limits and sleep.  bd_running must be
855 		 * cleared prior to checking the limits to avoid missed
856 		 * wakeups.  The waker will adjust one of bufspace or
857 		 * freebuffers prior to checking bd_running.
858 		 */
859 		BD_RUN_LOCK(bd);
860 		if (bd->bd_shutdown)
861 			break;
862 		atomic_store_int(&bd->bd_running, 0);
863 		if (bd->bd_bufspace < bd->bd_bufspacethresh &&
864 		    bd->bd_freebuffers > bd->bd_lofreebuffers) {
865 			msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd),
866 			    PRIBIO, "-", hz);
867 		} else {
868 			/* Avoid spurious wakeups while running. */
869 			atomic_store_int(&bd->bd_running, 1);
870 		}
871 	}
872 	wakeup(&bd->bd_shutdown);
873 	BD_RUN_UNLOCK(bd);
874 	kthread_exit();
875 }
876 
877 /*
878  *	bufmallocadjust:
879  *
880  *	Adjust the reported bufspace for a malloc managed buffer, possibly
881  *	waking any waiters.
882  */
883 static void
884 bufmallocadjust(struct buf *bp, int bufsize)
885 {
886 	int diff;
887 
888 	KASSERT((bp->b_flags & B_MALLOC) != 0,
889 	    ("bufmallocadjust: non-malloc buf %p", bp));
890 	diff = bufsize - bp->b_bufsize;
891 	if (diff < 0)
892 		atomic_subtract_long(&bufmallocspace, -diff);
893 	else
894 		atomic_add_long(&bufmallocspace, diff);
895 	bp->b_bufsize = bufsize;
896 }
897 
898 /*
899  *	runningwakeup:
900  *
901  *	Wake up processes that are waiting on asynchronous writes to fall
902  *	below lorunningspace.
903  */
904 static void
905 runningwakeup(void)
906 {
907 
908 	mtx_lock(&rbreqlock);
909 	if (runningbufreq) {
910 		runningbufreq = 0;
911 		wakeup(&runningbufreq);
912 	}
913 	mtx_unlock(&rbreqlock);
914 }
915 
916 /*
917  *	runningbufwakeup:
918  *
919  *	Decrement the outstanding write count according.
920  */
921 void
922 runningbufwakeup(struct buf *bp)
923 {
924 	long space, bspace;
925 
926 	bspace = bp->b_runningbufspace;
927 	if (bspace == 0)
928 		return;
929 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
930 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
931 	    space, bspace));
932 	bp->b_runningbufspace = 0;
933 	/*
934 	 * Only acquire the lock and wakeup on the transition from exceeding
935 	 * the threshold to falling below it.
936 	 */
937 	if (space < lorunningspace)
938 		return;
939 	if (space - bspace > lorunningspace)
940 		return;
941 	runningwakeup();
942 }
943 
944 /*
945  *	waitrunningbufspace()
946  *
947  *	runningbufspace is a measure of the amount of I/O currently
948  *	running.  This routine is used in async-write situations to
949  *	prevent creating huge backups of pending writes to a device.
950  *	Only asynchronous writes are governed by this function.
951  *
952  *	This does NOT turn an async write into a sync write.  It waits
953  *	for earlier writes to complete and generally returns before the
954  *	caller's write has reached the device.
955  */
956 void
957 waitrunningbufspace(void)
958 {
959 
960 	mtx_lock(&rbreqlock);
961 	while (runningbufspace > hirunningspace) {
962 		runningbufreq = 1;
963 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
964 	}
965 	mtx_unlock(&rbreqlock);
966 }
967 
968 /*
969  *	vfs_buf_test_cache:
970  *
971  *	Called when a buffer is extended.  This function clears the B_CACHE
972  *	bit if the newly extended portion of the buffer does not contain
973  *	valid data.
974  */
975 static __inline void
976 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
977     vm_offset_t size, vm_page_t m)
978 {
979 
980 	/*
981 	 * This function and its results are protected by higher level
982 	 * synchronization requiring vnode and buf locks to page in and
983 	 * validate pages.
984 	 */
985 	if (bp->b_flags & B_CACHE) {
986 		int base = (foff + off) & PAGE_MASK;
987 		if (vm_page_is_valid(m, base, size) == 0)
988 			bp->b_flags &= ~B_CACHE;
989 	}
990 }
991 
992 /* Wake up the buffer daemon if necessary */
993 static void
994 bd_wakeup(void)
995 {
996 
997 	mtx_lock(&bdlock);
998 	if (bd_request == 0) {
999 		bd_request = 1;
1000 		wakeup(&bd_request);
1001 	}
1002 	mtx_unlock(&bdlock);
1003 }
1004 
1005 /*
1006  * Adjust the maxbcachbuf tunable.
1007  */
1008 static void
1009 maxbcachebuf_adjust(void)
1010 {
1011 	int i;
1012 
1013 	/*
1014 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
1015 	 */
1016 	i = 2;
1017 	while (i * 2 <= maxbcachebuf)
1018 		i *= 2;
1019 	maxbcachebuf = i;
1020 	if (maxbcachebuf < MAXBSIZE)
1021 		maxbcachebuf = MAXBSIZE;
1022 	if (maxbcachebuf > maxphys)
1023 		maxbcachebuf = maxphys;
1024 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1025 		printf("maxbcachebuf=%d\n", maxbcachebuf);
1026 }
1027 
1028 /*
1029  * bd_speedup - speedup the buffer cache flushing code
1030  */
1031 void
1032 bd_speedup(void)
1033 {
1034 	int needwake;
1035 
1036 	mtx_lock(&bdlock);
1037 	needwake = 0;
1038 	if (bd_speedupreq == 0 || bd_request == 0)
1039 		needwake = 1;
1040 	bd_speedupreq = 1;
1041 	bd_request = 1;
1042 	if (needwake)
1043 		wakeup(&bd_request);
1044 	mtx_unlock(&bdlock);
1045 }
1046 
1047 #ifdef __i386__
1048 #define	TRANSIENT_DENOM	5
1049 #else
1050 #define	TRANSIENT_DENOM 10
1051 #endif
1052 
1053 /*
1054  * Calculating buffer cache scaling values and reserve space for buffer
1055  * headers.  This is called during low level kernel initialization and
1056  * may be called more then once.  We CANNOT write to the memory area
1057  * being reserved at this time.
1058  */
1059 caddr_t
1060 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1061 {
1062 	int tuned_nbuf;
1063 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1064 
1065 	/*
1066 	 * With KASAN or KMSAN enabled, the kernel map is shadowed.  Account for
1067 	 * this when sizing maps based on the amount of physical memory
1068 	 * available.
1069 	 */
1070 #if defined(KASAN)
1071 	physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1072 	    (KASAN_SHADOW_SCALE + 1);
1073 #elif defined(KMSAN)
1074 	physmem_est /= 3;
1075 
1076 	/*
1077 	 * KMSAN cannot reliably determine whether buffer data is initialized
1078 	 * unless it is updated through a KVA mapping.
1079 	 */
1080 	unmapped_buf_allowed = 0;
1081 #endif
1082 
1083 	/*
1084 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1085 	 * PAGE_SIZE is >= 1K)
1086 	 */
1087 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1088 
1089 	maxbcachebuf_adjust();
1090 	/*
1091 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1092 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1093 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1094 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1095 	 * the buffer cache we limit the eventual kva reservation to
1096 	 * maxbcache bytes.
1097 	 *
1098 	 * factor represents the 1/4 x ram conversion.
1099 	 */
1100 	if (nbuf == 0) {
1101 		int factor = 4 * BKVASIZE / 1024;
1102 
1103 		nbuf = 50;
1104 		if (physmem_est > 4096)
1105 			nbuf += min((physmem_est - 4096) / factor,
1106 			    65536 / factor);
1107 		if (physmem_est > 65536)
1108 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1109 			    32 * 1024 * 1024 / (factor * 5));
1110 
1111 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1112 			nbuf = maxbcache / BKVASIZE;
1113 		tuned_nbuf = 1;
1114 	} else
1115 		tuned_nbuf = 0;
1116 
1117 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1118 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1119 	if (nbuf > maxbuf) {
1120 		if (!tuned_nbuf)
1121 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1122 			    maxbuf);
1123 		nbuf = maxbuf;
1124 	}
1125 
1126 	/*
1127 	 * Ideal allocation size for the transient bio submap is 10%
1128 	 * of the maximal space buffer map.  This roughly corresponds
1129 	 * to the amount of the buffer mapped for typical UFS load.
1130 	 *
1131 	 * Clip the buffer map to reserve space for the transient
1132 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1133 	 * maximum buffer map extent on the platform.
1134 	 *
1135 	 * The fall-back to the maxbuf in case of maxbcache unset,
1136 	 * allows to not trim the buffer KVA for the architectures
1137 	 * with ample KVA space.
1138 	 */
1139 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1140 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1141 		buf_sz = (long)nbuf * BKVASIZE;
1142 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1143 		    (TRANSIENT_DENOM - 1)) {
1144 			/*
1145 			 * There is more KVA than memory.  Do not
1146 			 * adjust buffer map size, and assign the rest
1147 			 * of maxbuf to transient map.
1148 			 */
1149 			biotmap_sz = maxbuf_sz - buf_sz;
1150 		} else {
1151 			/*
1152 			 * Buffer map spans all KVA we could afford on
1153 			 * this platform.  Give 10% (20% on i386) of
1154 			 * the buffer map to the transient bio map.
1155 			 */
1156 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1157 			buf_sz -= biotmap_sz;
1158 		}
1159 		if (biotmap_sz / INT_MAX > maxphys)
1160 			bio_transient_maxcnt = INT_MAX;
1161 		else
1162 			bio_transient_maxcnt = biotmap_sz / maxphys;
1163 		/*
1164 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1165 		 * using the transient mapping.
1166 		 */
1167 		if (bio_transient_maxcnt > 1024)
1168 			bio_transient_maxcnt = 1024;
1169 		if (tuned_nbuf)
1170 			nbuf = buf_sz / BKVASIZE;
1171 	}
1172 
1173 	if (nswbuf == 0) {
1174 		/*
1175 		 * Pager buffers are allocated for short periods, so scale the
1176 		 * number of reserved buffers based on the number of CPUs rather
1177 		 * than amount of memory.
1178 		 */
1179 		nswbuf = min(nbuf / 4, 32 * mp_ncpus);
1180 		if (nswbuf < NSWBUF_MIN)
1181 			nswbuf = NSWBUF_MIN;
1182 	}
1183 
1184 	/*
1185 	 * Reserve space for the buffer cache buffers
1186 	 */
1187 	buf = (char *)v;
1188 	v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1189 	    atop(maxbcachebuf)) * nbuf;
1190 
1191 	return (v);
1192 }
1193 
1194 /*
1195  * Single global constant for BUF_WMESG, to avoid getting multiple
1196  * references.
1197  */
1198 static const char buf_wmesg[] = "bufwait";
1199 
1200 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1201 void
1202 bufinit(void)
1203 {
1204 	struct buf *bp;
1205 	int i;
1206 
1207 	TSENTER();
1208 	KASSERT(maxbcachebuf >= MAXBSIZE,
1209 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1210 	    MAXBSIZE));
1211 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1212 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1213 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1214 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1215 
1216 	unmapped_buf = (caddr_t)kva_alloc(maxphys);
1217 #ifdef INVARIANTS
1218 	poisoned_buf = unmapped_buf;
1219 #endif
1220 
1221 	/* finally, initialize each buffer header and stick on empty q */
1222 	for (i = 0; i < nbuf; i++) {
1223 		bp = nbufp(i);
1224 		bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1225 		bp->b_flags = B_INVAL;
1226 		bp->b_rcred = NOCRED;
1227 		bp->b_wcred = NOCRED;
1228 		bp->b_qindex = QUEUE_NONE;
1229 		bp->b_domain = -1;
1230 		bp->b_subqueue = mp_maxid + 1;
1231 		bp->b_xflags = 0;
1232 		bp->b_data = bp->b_kvabase = unmapped_buf;
1233 		LIST_INIT(&bp->b_dep);
1234 		BUF_LOCKINIT(bp, buf_wmesg);
1235 		bq_insert(&bqempty, bp, false);
1236 	}
1237 
1238 	/*
1239 	 * maxbufspace is the absolute maximum amount of buffer space we are
1240 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1241 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1242 	 * used by most other requests.  The differential is required to
1243 	 * ensure that metadata deadlocks don't occur.
1244 	 *
1245 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1246 	 * this may result in KVM fragmentation which is not handled optimally
1247 	 * by the system. XXX This is less true with vmem.  We could use
1248 	 * PAGE_SIZE.
1249 	 */
1250 	maxbufspace = (long)nbuf * BKVASIZE;
1251 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1252 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1253 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1254 
1255 	/*
1256 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1257 	 * arbitrarily and may need further tuning. It corresponds to
1258 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1259 	 * which fits with many RAID controllers' tagged queuing limits.
1260 	 * The lower 1 MiB limit is the historical upper limit for
1261 	 * hirunningspace.
1262 	 */
1263 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1264 	    16 * 1024 * 1024), 1024 * 1024);
1265 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1266 
1267 	/*
1268 	 * Limit the amount of malloc memory since it is wired permanently into
1269 	 * the kernel space.  Even though this is accounted for in the buffer
1270 	 * allocation, we don't want the malloced region to grow uncontrolled.
1271 	 * The malloc scheme improves memory utilization significantly on
1272 	 * average (small) directories.
1273 	 */
1274 	maxbufmallocspace = hibufspace / 20;
1275 
1276 	/*
1277 	 * Reduce the chance of a deadlock occurring by limiting the number
1278 	 * of delayed-write dirty buffers we allow to stack up.
1279 	 */
1280 	hidirtybuffers = nbuf / 4 + 20;
1281 	dirtybufthresh = hidirtybuffers * 9 / 10;
1282 	/*
1283 	 * To support extreme low-memory systems, make sure hidirtybuffers
1284 	 * cannot eat up all available buffer space.  This occurs when our
1285 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1286 	 * buffer space assuming BKVASIZE'd buffers.
1287 	 */
1288 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1289 		hidirtybuffers >>= 1;
1290 	}
1291 	lodirtybuffers = hidirtybuffers / 2;
1292 
1293 	/*
1294 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1295 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1296 	 * are counted as free but will be unavailable to threads executing
1297 	 * on other cpus.
1298 	 *
1299 	 * hifreebuffers is the free target for the bufspace daemon.  This
1300 	 * should be set appropriately to limit work per-iteration.
1301 	 */
1302 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1303 	hifreebuffers = (3 * lofreebuffers) / 2;
1304 	numfreebuffers = nbuf;
1305 
1306 	/* Setup the kva and free list allocators. */
1307 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1308 	buf_zone = uma_zcache_create("buf free cache",
1309 	    sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1310 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1311 
1312 	/*
1313 	 * Size the clean queue according to the amount of buffer space.
1314 	 * One queue per-256mb up to the max.  More queues gives better
1315 	 * concurrency but less accurate LRU.
1316 	 */
1317 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1318 	for (i = 0 ; i < buf_domains; i++) {
1319 		struct bufdomain *bd;
1320 
1321 		bd = &bdomain[i];
1322 		bd_init(bd);
1323 		bd->bd_freebuffers = nbuf / buf_domains;
1324 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1325 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1326 		bd->bd_bufspace = 0;
1327 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1328 		bd->bd_hibufspace = hibufspace / buf_domains;
1329 		bd->bd_lobufspace = lobufspace / buf_domains;
1330 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1331 		bd->bd_numdirtybuffers = 0;
1332 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1333 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1334 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1335 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1336 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1337 	}
1338 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1339 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1340 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1341 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1342 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1343 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1344 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1345 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1346 	TSEXIT();
1347 }
1348 
1349 #ifdef INVARIANTS
1350 static inline void
1351 vfs_buf_check_mapped(struct buf *bp)
1352 {
1353 
1354 	KASSERT(bp->b_kvabase != unmapped_buf,
1355 	    ("mapped buf: b_kvabase was not updated %p", bp));
1356 	KASSERT(bp->b_data != unmapped_buf,
1357 	    ("mapped buf: b_data was not updated %p", bp));
1358 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1359 	    maxphys, ("b_data + b_offset unmapped %p", bp));
1360 }
1361 
1362 static inline void
1363 vfs_buf_check_unmapped(struct buf *bp)
1364 {
1365 
1366 	KASSERT(bp->b_data == unmapped_buf,
1367 	    ("unmapped buf: corrupted b_data %p", bp));
1368 }
1369 
1370 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1371 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1372 #else
1373 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1374 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1375 #endif
1376 
1377 static int
1378 isbufbusy(struct buf *bp)
1379 {
1380 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1381 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1382 		return (1);
1383 	return (0);
1384 }
1385 
1386 /*
1387  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1388  */
1389 void
1390 bufshutdown(int show_busybufs)
1391 {
1392 	static int first_buf_printf = 1;
1393 	struct buf *bp;
1394 	int i, iter, nbusy, pbusy;
1395 #ifndef PREEMPTION
1396 	int subiter;
1397 #endif
1398 
1399 	/*
1400 	 * Sync filesystems for shutdown
1401 	 */
1402 	wdog_kern_pat(WD_LASTVAL);
1403 	kern_sync(curthread);
1404 
1405 	/*
1406 	 * With soft updates, some buffers that are
1407 	 * written will be remarked as dirty until other
1408 	 * buffers are written.
1409 	 */
1410 	for (iter = pbusy = 0; iter < 20; iter++) {
1411 		nbusy = 0;
1412 		for (i = nbuf - 1; i >= 0; i--) {
1413 			bp = nbufp(i);
1414 			if (isbufbusy(bp))
1415 				nbusy++;
1416 		}
1417 		if (nbusy == 0) {
1418 			if (first_buf_printf)
1419 				printf("All buffers synced.");
1420 			break;
1421 		}
1422 		if (first_buf_printf) {
1423 			printf("Syncing disks, buffers remaining... ");
1424 			first_buf_printf = 0;
1425 		}
1426 		printf("%d ", nbusy);
1427 		if (nbusy < pbusy)
1428 			iter = 0;
1429 		pbusy = nbusy;
1430 
1431 		wdog_kern_pat(WD_LASTVAL);
1432 		kern_sync(curthread);
1433 
1434 #ifdef PREEMPTION
1435 		/*
1436 		 * Spin for a while to allow interrupt threads to run.
1437 		 */
1438 		DELAY(50000 * iter);
1439 #else
1440 		/*
1441 		 * Context switch several times to allow interrupt
1442 		 * threads to run.
1443 		 */
1444 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1445 			sched_relinquish(curthread);
1446 			DELAY(1000);
1447 		}
1448 #endif
1449 	}
1450 	printf("\n");
1451 	/*
1452 	 * Count only busy local buffers to prevent forcing
1453 	 * a fsck if we're just a client of a wedged NFS server
1454 	 */
1455 	nbusy = 0;
1456 	for (i = nbuf - 1; i >= 0; i--) {
1457 		bp = nbufp(i);
1458 		if (isbufbusy(bp)) {
1459 #if 0
1460 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1461 			if (bp->b_dev == NULL) {
1462 				TAILQ_REMOVE(&mountlist,
1463 				    bp->b_vp->v_mount, mnt_list);
1464 				continue;
1465 			}
1466 #endif
1467 			nbusy++;
1468 			if (show_busybufs > 0) {
1469 				printf(
1470 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1471 				    nbusy, bp, bp->b_vp, bp->b_flags,
1472 				    (intmax_t)bp->b_blkno,
1473 				    (intmax_t)bp->b_lblkno);
1474 				BUF_LOCKPRINTINFO(bp);
1475 				if (show_busybufs > 1)
1476 					vn_printf(bp->b_vp,
1477 					    "vnode content: ");
1478 			}
1479 		}
1480 	}
1481 	if (nbusy) {
1482 		/*
1483 		 * Failed to sync all blocks. Indicate this and don't
1484 		 * unmount filesystems (thus forcing an fsck on reboot).
1485 		 */
1486 		BOOTTRACE("shutdown failed to sync buffers");
1487 		printf("Giving up on %d buffers\n", nbusy);
1488 		DELAY(5000000);	/* 5 seconds */
1489 		swapoff_all();
1490 	} else {
1491 		BOOTTRACE("shutdown sync complete");
1492 		if (!first_buf_printf)
1493 			printf("Final sync complete\n");
1494 
1495 		/*
1496 		 * Unmount filesystems and perform swapoff, to quiesce
1497 		 * the system as much as possible.  In particular, no
1498 		 * I/O should be initiated from top levels since it
1499 		 * might be abruptly terminated by reset, or otherwise
1500 		 * erronously handled because other parts of the
1501 		 * system are disabled.
1502 		 *
1503 		 * Swapoff before unmount, because file-backed swap is
1504 		 * non-operational after unmount of the underlying
1505 		 * filesystem.
1506 		 */
1507 		if (!KERNEL_PANICKED()) {
1508 			swapoff_all();
1509 			vfs_unmountall();
1510 		}
1511 		BOOTTRACE("shutdown unmounted all filesystems");
1512 	}
1513 	DELAY(100000);		/* wait for console output to finish */
1514 }
1515 
1516 static void
1517 bpmap_qenter(struct buf *bp)
1518 {
1519 
1520 	BUF_CHECK_MAPPED(bp);
1521 
1522 	/*
1523 	 * bp->b_data is relative to bp->b_offset, but
1524 	 * bp->b_offset may be offset into the first page.
1525 	 */
1526 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1527 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1528 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1529 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1530 }
1531 
1532 static inline struct bufdomain *
1533 bufdomain(struct buf *bp)
1534 {
1535 
1536 	return (&bdomain[bp->b_domain]);
1537 }
1538 
1539 static struct bufqueue *
1540 bufqueue(struct buf *bp)
1541 {
1542 
1543 	switch (bp->b_qindex) {
1544 	case QUEUE_NONE:
1545 		/* FALLTHROUGH */
1546 	case QUEUE_SENTINEL:
1547 		return (NULL);
1548 	case QUEUE_EMPTY:
1549 		return (&bqempty);
1550 	case QUEUE_DIRTY:
1551 		return (&bufdomain(bp)->bd_dirtyq);
1552 	case QUEUE_CLEAN:
1553 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1554 	default:
1555 		break;
1556 	}
1557 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1558 }
1559 
1560 /*
1561  * Return the locked bufqueue that bp is a member of.
1562  */
1563 static struct bufqueue *
1564 bufqueue_acquire(struct buf *bp)
1565 {
1566 	struct bufqueue *bq, *nbq;
1567 
1568 	/*
1569 	 * bp can be pushed from a per-cpu queue to the
1570 	 * cleanq while we're waiting on the lock.  Retry
1571 	 * if the queues don't match.
1572 	 */
1573 	bq = bufqueue(bp);
1574 	BQ_LOCK(bq);
1575 	for (;;) {
1576 		nbq = bufqueue(bp);
1577 		if (bq == nbq)
1578 			break;
1579 		BQ_UNLOCK(bq);
1580 		BQ_LOCK(nbq);
1581 		bq = nbq;
1582 	}
1583 	return (bq);
1584 }
1585 
1586 /*
1587  *	binsfree:
1588  *
1589  *	Insert the buffer into the appropriate free list.  Requires a
1590  *	locked buffer on entry and buffer is unlocked before return.
1591  */
1592 static void
1593 binsfree(struct buf *bp, int qindex)
1594 {
1595 	struct bufdomain *bd;
1596 	struct bufqueue *bq;
1597 
1598 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1599 	    ("binsfree: Invalid qindex %d", qindex));
1600 	BUF_ASSERT_XLOCKED(bp);
1601 
1602 	/*
1603 	 * Handle delayed bremfree() processing.
1604 	 */
1605 	if (bp->b_flags & B_REMFREE) {
1606 		if (bp->b_qindex == qindex) {
1607 			bp->b_flags |= B_REUSE;
1608 			bp->b_flags &= ~B_REMFREE;
1609 			BUF_UNLOCK(bp);
1610 			return;
1611 		}
1612 		bq = bufqueue_acquire(bp);
1613 		bq_remove(bq, bp);
1614 		BQ_UNLOCK(bq);
1615 	}
1616 	bd = bufdomain(bp);
1617 	if (qindex == QUEUE_CLEAN) {
1618 		if (bd->bd_lim != 0)
1619 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1620 		else
1621 			bq = bd->bd_cleanq;
1622 	} else
1623 		bq = &bd->bd_dirtyq;
1624 	bq_insert(bq, bp, true);
1625 }
1626 
1627 /*
1628  * buf_free:
1629  *
1630  *	Free a buffer to the buf zone once it no longer has valid contents.
1631  */
1632 static void
1633 buf_free(struct buf *bp)
1634 {
1635 
1636 	if (bp->b_flags & B_REMFREE)
1637 		bremfreef(bp);
1638 	if (bp->b_vflags & BV_BKGRDINPROG)
1639 		panic("losing buffer 1");
1640 	if (bp->b_rcred != NOCRED) {
1641 		crfree(bp->b_rcred);
1642 		bp->b_rcred = NOCRED;
1643 	}
1644 	if (bp->b_wcred != NOCRED) {
1645 		crfree(bp->b_wcred);
1646 		bp->b_wcred = NOCRED;
1647 	}
1648 	if (!LIST_EMPTY(&bp->b_dep))
1649 		buf_deallocate(bp);
1650 	bufkva_free(bp);
1651 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1652 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
1653 	BUF_UNLOCK(bp);
1654 	uma_zfree(buf_zone, bp);
1655 }
1656 
1657 /*
1658  * buf_import:
1659  *
1660  *	Import bufs into the uma cache from the buf list.  The system still
1661  *	expects a static array of bufs and much of the synchronization
1662  *	around bufs assumes type stable storage.  As a result, UMA is used
1663  *	only as a per-cpu cache of bufs still maintained on a global list.
1664  */
1665 static int
1666 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1667 {
1668 	struct buf *bp;
1669 	int i;
1670 
1671 	BQ_LOCK(&bqempty);
1672 	for (i = 0; i < cnt; i++) {
1673 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1674 		if (bp == NULL)
1675 			break;
1676 		bq_remove(&bqempty, bp);
1677 		store[i] = bp;
1678 	}
1679 	BQ_UNLOCK(&bqempty);
1680 
1681 	return (i);
1682 }
1683 
1684 /*
1685  * buf_release:
1686  *
1687  *	Release bufs from the uma cache back to the buffer queues.
1688  */
1689 static void
1690 buf_release(void *arg, void **store, int cnt)
1691 {
1692 	struct bufqueue *bq;
1693 	struct buf *bp;
1694         int i;
1695 
1696 	bq = &bqempty;
1697 	BQ_LOCK(bq);
1698         for (i = 0; i < cnt; i++) {
1699 		bp = store[i];
1700 		/* Inline bq_insert() to batch locking. */
1701 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1702 		bp->b_flags &= ~(B_AGE | B_REUSE);
1703 		bq->bq_len++;
1704 		bp->b_qindex = bq->bq_index;
1705 	}
1706 	BQ_UNLOCK(bq);
1707 }
1708 
1709 /*
1710  * buf_alloc:
1711  *
1712  *	Allocate an empty buffer header.
1713  */
1714 static struct buf *
1715 buf_alloc(struct bufdomain *bd)
1716 {
1717 	struct buf *bp;
1718 	int freebufs, error;
1719 
1720 	/*
1721 	 * We can only run out of bufs in the buf zone if the average buf
1722 	 * is less than BKVASIZE.  In this case the actual wait/block will
1723 	 * come from buf_reycle() failing to flush one of these small bufs.
1724 	 */
1725 	bp = NULL;
1726 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1727 	if (freebufs > 0)
1728 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1729 	if (bp == NULL) {
1730 		atomic_add_int(&bd->bd_freebuffers, 1);
1731 		bufspace_daemon_wakeup(bd);
1732 		counter_u64_add(numbufallocfails, 1);
1733 		return (NULL);
1734 	}
1735 	/*
1736 	 * Wake-up the bufspace daemon on transition below threshold.
1737 	 */
1738 	if (freebufs == bd->bd_lofreebuffers)
1739 		bufspace_daemon_wakeup(bd);
1740 
1741 	error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWITNESS, NULL);
1742 	KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1743 	    error));
1744 	(void)error;
1745 
1746 	KASSERT(bp->b_vp == NULL,
1747 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1748 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1749 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1750 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1751 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1752 	KASSERT(bp->b_npages == 0,
1753 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1754 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1755 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1756 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
1757 
1758 	bp->b_domain = BD_DOMAIN(bd);
1759 	bp->b_flags = 0;
1760 	bp->b_ioflags = 0;
1761 	bp->b_xflags = 0;
1762 	bp->b_vflags = 0;
1763 	bp->b_vp = NULL;
1764 	bp->b_blkno = bp->b_lblkno = 0;
1765 	bp->b_offset = NOOFFSET;
1766 	bp->b_iodone = 0;
1767 	bp->b_error = 0;
1768 	bp->b_resid = 0;
1769 	bp->b_bcount = 0;
1770 	bp->b_npages = 0;
1771 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1772 	bp->b_bufobj = NULL;
1773 	bp->b_data = bp->b_kvabase = unmapped_buf;
1774 	bp->b_fsprivate1 = NULL;
1775 	bp->b_fsprivate2 = NULL;
1776 	bp->b_fsprivate3 = NULL;
1777 	LIST_INIT(&bp->b_dep);
1778 
1779 	return (bp);
1780 }
1781 
1782 /*
1783  *	buf_recycle:
1784  *
1785  *	Free a buffer from the given bufqueue.  kva controls whether the
1786  *	freed buf must own some kva resources.  This is used for
1787  *	defragmenting.
1788  */
1789 static int
1790 buf_recycle(struct bufdomain *bd, bool kva)
1791 {
1792 	struct bufqueue *bq;
1793 	struct buf *bp, *nbp;
1794 
1795 	if (kva)
1796 		counter_u64_add(bufdefragcnt, 1);
1797 	nbp = NULL;
1798 	bq = bd->bd_cleanq;
1799 	BQ_LOCK(bq);
1800 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1801 	    ("buf_recycle: Locks don't match"));
1802 	nbp = TAILQ_FIRST(&bq->bq_queue);
1803 
1804 	/*
1805 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1806 	 * depending.
1807 	 */
1808 	while ((bp = nbp) != NULL) {
1809 		/*
1810 		 * Calculate next bp (we can only use it if we do not
1811 		 * release the bqlock).
1812 		 */
1813 		nbp = TAILQ_NEXT(bp, b_freelist);
1814 
1815 		/*
1816 		 * If we are defragging then we need a buffer with
1817 		 * some kva to reclaim.
1818 		 */
1819 		if (kva && bp->b_kvasize == 0)
1820 			continue;
1821 
1822 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1823 			continue;
1824 
1825 		/*
1826 		 * Implement a second chance algorithm for frequently
1827 		 * accessed buffers.
1828 		 */
1829 		if ((bp->b_flags & B_REUSE) != 0) {
1830 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1831 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1832 			bp->b_flags &= ~B_REUSE;
1833 			BUF_UNLOCK(bp);
1834 			continue;
1835 		}
1836 
1837 		/*
1838 		 * Skip buffers with background writes in progress.
1839 		 */
1840 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1841 			BUF_UNLOCK(bp);
1842 			continue;
1843 		}
1844 
1845 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1846 		    ("buf_recycle: inconsistent queue %d bp %p",
1847 		    bp->b_qindex, bp));
1848 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1849 		    ("getnewbuf: queue domain %d doesn't match request %d",
1850 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1851 		/*
1852 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1853 		 * the scan from this point on.
1854 		 */
1855 		bq_remove(bq, bp);
1856 		BQ_UNLOCK(bq);
1857 
1858 		/*
1859 		 * Requeue the background write buffer with error and
1860 		 * restart the scan.
1861 		 */
1862 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1863 			bqrelse(bp);
1864 			BQ_LOCK(bq);
1865 			nbp = TAILQ_FIRST(&bq->bq_queue);
1866 			continue;
1867 		}
1868 		bp->b_flags |= B_INVAL;
1869 		brelse(bp);
1870 		return (0);
1871 	}
1872 	bd->bd_wanted = 1;
1873 	BQ_UNLOCK(bq);
1874 
1875 	return (ENOBUFS);
1876 }
1877 
1878 /*
1879  *	bremfree:
1880  *
1881  *	Mark the buffer for removal from the appropriate free list.
1882  *
1883  */
1884 void
1885 bremfree(struct buf *bp)
1886 {
1887 
1888 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1889 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1890 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1891 	KASSERT(bp->b_qindex != QUEUE_NONE,
1892 	    ("bremfree: buffer %p not on a queue.", bp));
1893 	BUF_ASSERT_XLOCKED(bp);
1894 
1895 	bp->b_flags |= B_REMFREE;
1896 }
1897 
1898 /*
1899  *	bremfreef:
1900  *
1901  *	Force an immediate removal from a free list.  Used only in nfs when
1902  *	it abuses the b_freelist pointer.
1903  */
1904 void
1905 bremfreef(struct buf *bp)
1906 {
1907 	struct bufqueue *bq;
1908 
1909 	bq = bufqueue_acquire(bp);
1910 	bq_remove(bq, bp);
1911 	BQ_UNLOCK(bq);
1912 }
1913 
1914 static void
1915 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1916 {
1917 
1918 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1919 	TAILQ_INIT(&bq->bq_queue);
1920 	bq->bq_len = 0;
1921 	bq->bq_index = qindex;
1922 	bq->bq_subqueue = subqueue;
1923 }
1924 
1925 static void
1926 bd_init(struct bufdomain *bd)
1927 {
1928 	int i;
1929 
1930 	/* Per-CPU clean buf queues, plus one global queue. */
1931 	bd->bd_subq = mallocarray(mp_maxid + 2, sizeof(struct bufqueue),
1932 	    M_BIOBUF, M_WAITOK | M_ZERO);
1933 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1934 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1935 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1936 	for (i = 0; i <= mp_maxid; i++)
1937 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1938 		    "bufq clean subqueue lock");
1939 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1940 }
1941 
1942 /*
1943  *	bq_remove:
1944  *
1945  *	Removes a buffer from the free list, must be called with the
1946  *	correct qlock held.
1947  */
1948 static void
1949 bq_remove(struct bufqueue *bq, struct buf *bp)
1950 {
1951 
1952 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1953 	    bp, bp->b_vp, bp->b_flags);
1954 	KASSERT(bp->b_qindex != QUEUE_NONE,
1955 	    ("bq_remove: buffer %p not on a queue.", bp));
1956 	KASSERT(bufqueue(bp) == bq,
1957 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1958 
1959 	BQ_ASSERT_LOCKED(bq);
1960 	if (bp->b_qindex != QUEUE_EMPTY) {
1961 		BUF_ASSERT_XLOCKED(bp);
1962 	}
1963 	KASSERT(bq->bq_len >= 1,
1964 	    ("queue %d underflow", bp->b_qindex));
1965 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1966 	bq->bq_len--;
1967 	bp->b_qindex = QUEUE_NONE;
1968 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1969 }
1970 
1971 static void
1972 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1973 {
1974 	struct buf *bp;
1975 
1976 	BQ_ASSERT_LOCKED(bq);
1977 	if (bq != bd->bd_cleanq) {
1978 		BD_LOCK(bd);
1979 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1980 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1981 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1982 			    b_freelist);
1983 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1984 		}
1985 		bd->bd_cleanq->bq_len += bq->bq_len;
1986 		bq->bq_len = 0;
1987 	}
1988 	if (bd->bd_wanted) {
1989 		bd->bd_wanted = 0;
1990 		wakeup(&bd->bd_wanted);
1991 	}
1992 	if (bq != bd->bd_cleanq)
1993 		BD_UNLOCK(bd);
1994 }
1995 
1996 static int
1997 bd_flushall(struct bufdomain *bd)
1998 {
1999 	struct bufqueue *bq;
2000 	int flushed;
2001 	int i;
2002 
2003 	if (bd->bd_lim == 0)
2004 		return (0);
2005 	flushed = 0;
2006 	for (i = 0; i <= mp_maxid; i++) {
2007 		bq = &bd->bd_subq[i];
2008 		if (bq->bq_len == 0)
2009 			continue;
2010 		BQ_LOCK(bq);
2011 		bd_flush(bd, bq);
2012 		BQ_UNLOCK(bq);
2013 		flushed++;
2014 	}
2015 
2016 	return (flushed);
2017 }
2018 
2019 static void
2020 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
2021 {
2022 	struct bufdomain *bd;
2023 
2024 	if (bp->b_qindex != QUEUE_NONE)
2025 		panic("bq_insert: free buffer %p onto another queue?", bp);
2026 
2027 	bd = bufdomain(bp);
2028 	if (bp->b_flags & B_AGE) {
2029 		/* Place this buf directly on the real queue. */
2030 		if (bq->bq_index == QUEUE_CLEAN)
2031 			bq = bd->bd_cleanq;
2032 		BQ_LOCK(bq);
2033 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
2034 	} else {
2035 		BQ_LOCK(bq);
2036 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
2037 	}
2038 	bp->b_flags &= ~(B_AGE | B_REUSE);
2039 	bq->bq_len++;
2040 	bp->b_qindex = bq->bq_index;
2041 	bp->b_subqueue = bq->bq_subqueue;
2042 
2043 	/*
2044 	 * Unlock before we notify so that we don't wakeup a waiter that
2045 	 * fails a trylock on the buf and sleeps again.
2046 	 */
2047 	if (unlock)
2048 		BUF_UNLOCK(bp);
2049 
2050 	if (bp->b_qindex == QUEUE_CLEAN) {
2051 		/*
2052 		 * Flush the per-cpu queue and notify any waiters.
2053 		 */
2054 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2055 		    bq->bq_len >= bd->bd_lim))
2056 			bd_flush(bd, bq);
2057 	}
2058 	BQ_UNLOCK(bq);
2059 }
2060 
2061 /*
2062  *	bufkva_free:
2063  *
2064  *	Free the kva allocation for a buffer.
2065  *
2066  */
2067 static void
2068 bufkva_free(struct buf *bp)
2069 {
2070 
2071 #ifdef INVARIANTS
2072 	if (bp->b_kvasize == 0) {
2073 		KASSERT(bp->b_kvabase == unmapped_buf &&
2074 		    bp->b_data == unmapped_buf,
2075 		    ("Leaked KVA space on %p", bp));
2076 	} else if (buf_mapped(bp))
2077 		BUF_CHECK_MAPPED(bp);
2078 	else
2079 		BUF_CHECK_UNMAPPED(bp);
2080 #endif
2081 	if (bp->b_kvasize == 0)
2082 		return;
2083 
2084 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2085 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
2086 	counter_u64_add(buffreekvacnt, 1);
2087 	bp->b_data = bp->b_kvabase = unmapped_buf;
2088 	bp->b_kvasize = 0;
2089 }
2090 
2091 /*
2092  *	bufkva_alloc:
2093  *
2094  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2095  */
2096 static int
2097 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2098 {
2099 	vm_offset_t addr;
2100 	int error;
2101 
2102 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2103 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2104 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
2105 	KASSERT(maxsize <= maxbcachebuf,
2106 	    ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2107 
2108 	bufkva_free(bp);
2109 
2110 	addr = 0;
2111 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2112 	if (error != 0) {
2113 		/*
2114 		 * Buffer map is too fragmented.  Request the caller
2115 		 * to defragment the map.
2116 		 */
2117 		return (error);
2118 	}
2119 	bp->b_kvabase = (caddr_t)addr;
2120 	bp->b_kvasize = maxsize;
2121 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2122 	if ((gbflags & GB_UNMAPPED) != 0) {
2123 		bp->b_data = unmapped_buf;
2124 		BUF_CHECK_UNMAPPED(bp);
2125 	} else {
2126 		bp->b_data = bp->b_kvabase;
2127 		BUF_CHECK_MAPPED(bp);
2128 	}
2129 	return (0);
2130 }
2131 
2132 /*
2133  *	bufkva_reclaim:
2134  *
2135  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2136  *	callback that fires to avoid returning failure.
2137  */
2138 static void
2139 bufkva_reclaim(vmem_t *vmem, int flags)
2140 {
2141 	bool done;
2142 	int q;
2143 	int i;
2144 
2145 	done = false;
2146 	for (i = 0; i < 5; i++) {
2147 		for (q = 0; q < buf_domains; q++)
2148 			if (buf_recycle(&bdomain[q], true) != 0)
2149 				done = true;
2150 		if (done)
2151 			break;
2152 	}
2153 	return;
2154 }
2155 
2156 /*
2157  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2158  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2159  * the buffer is valid and we do not have to do anything.
2160  */
2161 static void
2162 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2163     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2164 {
2165 	struct buf *rabp;
2166 	struct thread *td;
2167 	int i;
2168 
2169 	td = curthread;
2170 
2171 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2172 		if (inmem(vp, *rablkno))
2173 			continue;
2174 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2175 		if ((rabp->b_flags & B_CACHE) != 0) {
2176 			brelse(rabp);
2177 			continue;
2178 		}
2179 #ifdef RACCT
2180 		if (racct_enable) {
2181 			PROC_LOCK(curproc);
2182 			racct_add_buf(curproc, rabp, 0);
2183 			PROC_UNLOCK(curproc);
2184 		}
2185 #endif /* RACCT */
2186 		td->td_ru.ru_inblock++;
2187 		rabp->b_flags |= B_ASYNC;
2188 		rabp->b_flags &= ~B_INVAL;
2189 		if ((flags & GB_CKHASH) != 0) {
2190 			rabp->b_flags |= B_CKHASH;
2191 			rabp->b_ckhashcalc = ckhashfunc;
2192 		}
2193 		rabp->b_ioflags &= ~BIO_ERROR;
2194 		rabp->b_iocmd = BIO_READ;
2195 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2196 			rabp->b_rcred = crhold(cred);
2197 		vfs_busy_pages(rabp, 0);
2198 		BUF_KERNPROC(rabp);
2199 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2200 		bstrategy(rabp);
2201 	}
2202 }
2203 
2204 /*
2205  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2206  *
2207  * Get a buffer with the specified data.  Look in the cache first.  We
2208  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2209  * is set, the buffer is valid and we do not have to do anything, see
2210  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2211  *
2212  * Always return a NULL buffer pointer (in bpp) when returning an error.
2213  *
2214  * The blkno parameter is the logical block being requested. Normally
2215  * the mapping of logical block number to disk block address is done
2216  * by calling VOP_BMAP(). However, if the mapping is already known, the
2217  * disk block address can be passed using the dblkno parameter. If the
2218  * disk block address is not known, then the same value should be passed
2219  * for blkno and dblkno.
2220  */
2221 int
2222 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2223     daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2224     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2225 {
2226 	struct buf *bp;
2227 	struct thread *td;
2228 	int error, readwait, rv;
2229 
2230 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2231 	td = curthread;
2232 	/*
2233 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2234 	 * are specified.
2235 	 */
2236 	error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2237 	if (error != 0) {
2238 		*bpp = NULL;
2239 		return (error);
2240 	}
2241 	KASSERT(blkno == bp->b_lblkno,
2242 	    ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2243 	    (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2244 	flags &= ~GB_NOSPARSE;
2245 	*bpp = bp;
2246 
2247 	/*
2248 	 * If not found in cache, do some I/O
2249 	 */
2250 	readwait = 0;
2251 	if ((bp->b_flags & B_CACHE) == 0) {
2252 #ifdef RACCT
2253 		if (racct_enable) {
2254 			PROC_LOCK(td->td_proc);
2255 			racct_add_buf(td->td_proc, bp, 0);
2256 			PROC_UNLOCK(td->td_proc);
2257 		}
2258 #endif /* RACCT */
2259 		td->td_ru.ru_inblock++;
2260 		bp->b_iocmd = BIO_READ;
2261 		bp->b_flags &= ~B_INVAL;
2262 		if ((flags & GB_CKHASH) != 0) {
2263 			bp->b_flags |= B_CKHASH;
2264 			bp->b_ckhashcalc = ckhashfunc;
2265 		}
2266 		if ((flags & GB_CVTENXIO) != 0)
2267 			bp->b_xflags |= BX_CVTENXIO;
2268 		bp->b_ioflags &= ~BIO_ERROR;
2269 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2270 			bp->b_rcred = crhold(cred);
2271 		vfs_busy_pages(bp, 0);
2272 		bp->b_iooffset = dbtob(bp->b_blkno);
2273 		bstrategy(bp);
2274 		++readwait;
2275 	}
2276 
2277 	/*
2278 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2279 	 */
2280 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2281 
2282 	rv = 0;
2283 	if (readwait) {
2284 		rv = bufwait(bp);
2285 		if (rv != 0) {
2286 			brelse(bp);
2287 			*bpp = NULL;
2288 		}
2289 	}
2290 	return (rv);
2291 }
2292 
2293 /*
2294  * Write, release buffer on completion.  (Done by iodone
2295  * if async).  Do not bother writing anything if the buffer
2296  * is invalid.
2297  *
2298  * Note that we set B_CACHE here, indicating that buffer is
2299  * fully valid and thus cacheable.  This is true even of NFS
2300  * now so we set it generally.  This could be set either here
2301  * or in biodone() since the I/O is synchronous.  We put it
2302  * here.
2303  */
2304 int
2305 bufwrite(struct buf *bp)
2306 {
2307 	int oldflags;
2308 	struct vnode *vp;
2309 	long space;
2310 	int vp_md;
2311 
2312 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2313 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2314 		bp->b_flags |= B_INVAL | B_RELBUF;
2315 		bp->b_flags &= ~B_CACHE;
2316 		brelse(bp);
2317 		return (ENXIO);
2318 	}
2319 	if (bp->b_flags & B_INVAL) {
2320 		brelse(bp);
2321 		return (0);
2322 	}
2323 
2324 	if (bp->b_flags & B_BARRIER)
2325 		atomic_add_long(&barrierwrites, 1);
2326 
2327 	oldflags = bp->b_flags;
2328 
2329 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2330 	    ("FFS background buffer should not get here %p", bp));
2331 
2332 	vp = bp->b_vp;
2333 	if (vp)
2334 		vp_md = vp->v_vflag & VV_MD;
2335 	else
2336 		vp_md = 0;
2337 
2338 	/*
2339 	 * Mark the buffer clean.  Increment the bufobj write count
2340 	 * before bundirty() call, to prevent other thread from seeing
2341 	 * empty dirty list and zero counter for writes in progress,
2342 	 * falsely indicating that the bufobj is clean.
2343 	 */
2344 	bufobj_wref(bp->b_bufobj);
2345 	bundirty(bp);
2346 
2347 	bp->b_flags &= ~B_DONE;
2348 	bp->b_ioflags &= ~BIO_ERROR;
2349 	bp->b_flags |= B_CACHE;
2350 	bp->b_iocmd = BIO_WRITE;
2351 
2352 	vfs_busy_pages(bp, 1);
2353 
2354 	/*
2355 	 * Normal bwrites pipeline writes
2356 	 */
2357 	bp->b_runningbufspace = bp->b_bufsize;
2358 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2359 
2360 #ifdef RACCT
2361 	if (racct_enable) {
2362 		PROC_LOCK(curproc);
2363 		racct_add_buf(curproc, bp, 1);
2364 		PROC_UNLOCK(curproc);
2365 	}
2366 #endif /* RACCT */
2367 	curthread->td_ru.ru_oublock++;
2368 	if (oldflags & B_ASYNC)
2369 		BUF_KERNPROC(bp);
2370 	bp->b_iooffset = dbtob(bp->b_blkno);
2371 	buf_track(bp, __func__);
2372 	bstrategy(bp);
2373 
2374 	if ((oldflags & B_ASYNC) == 0) {
2375 		int rtval = bufwait(bp);
2376 		brelse(bp);
2377 		return (rtval);
2378 	} else if (space > hirunningspace) {
2379 		/*
2380 		 * don't allow the async write to saturate the I/O
2381 		 * system.  We will not deadlock here because
2382 		 * we are blocking waiting for I/O that is already in-progress
2383 		 * to complete. We do not block here if it is the update
2384 		 * or syncer daemon trying to clean up as that can lead
2385 		 * to deadlock.
2386 		 */
2387 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2388 			waitrunningbufspace();
2389 	}
2390 
2391 	return (0);
2392 }
2393 
2394 void
2395 bufbdflush(struct bufobj *bo, struct buf *bp)
2396 {
2397 	struct buf *nbp;
2398 	struct bufdomain *bd;
2399 
2400 	bd = &bdomain[bo->bo_domain];
2401 	if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2402 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2403 		altbufferflushes++;
2404 	} else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2405 		BO_LOCK(bo);
2406 		/*
2407 		 * Try to find a buffer to flush.
2408 		 */
2409 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2410 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2411 			    BUF_LOCK(nbp,
2412 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2413 				continue;
2414 			if (bp == nbp)
2415 				panic("bdwrite: found ourselves");
2416 			BO_UNLOCK(bo);
2417 			/* Don't countdeps with the bo lock held. */
2418 			if (buf_countdeps(nbp, 0)) {
2419 				BO_LOCK(bo);
2420 				BUF_UNLOCK(nbp);
2421 				continue;
2422 			}
2423 			if (nbp->b_flags & B_CLUSTEROK) {
2424 				vfs_bio_awrite(nbp);
2425 			} else {
2426 				bremfree(nbp);
2427 				bawrite(nbp);
2428 			}
2429 			dirtybufferflushes++;
2430 			break;
2431 		}
2432 		if (nbp == NULL)
2433 			BO_UNLOCK(bo);
2434 	}
2435 }
2436 
2437 /*
2438  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2439  * anything if the buffer is marked invalid.
2440  *
2441  * Note that since the buffer must be completely valid, we can safely
2442  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2443  * biodone() in order to prevent getblk from writing the buffer
2444  * out synchronously.
2445  */
2446 void
2447 bdwrite(struct buf *bp)
2448 {
2449 	struct thread *td = curthread;
2450 	struct vnode *vp;
2451 	struct bufobj *bo;
2452 
2453 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2454 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2455 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2456 	    ("Barrier request in delayed write %p", bp));
2457 
2458 	if (bp->b_flags & B_INVAL) {
2459 		brelse(bp);
2460 		return;
2461 	}
2462 
2463 	/*
2464 	 * If we have too many dirty buffers, don't create any more.
2465 	 * If we are wildly over our limit, then force a complete
2466 	 * cleanup. Otherwise, just keep the situation from getting
2467 	 * out of control. Note that we have to avoid a recursive
2468 	 * disaster and not try to clean up after our own cleanup!
2469 	 */
2470 	vp = bp->b_vp;
2471 	bo = bp->b_bufobj;
2472 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2473 		td->td_pflags |= TDP_INBDFLUSH;
2474 		BO_BDFLUSH(bo, bp);
2475 		td->td_pflags &= ~TDP_INBDFLUSH;
2476 	} else
2477 		recursiveflushes++;
2478 
2479 	bdirty(bp);
2480 	/*
2481 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2482 	 * true even of NFS now.
2483 	 */
2484 	bp->b_flags |= B_CACHE;
2485 
2486 	/*
2487 	 * This bmap keeps the system from needing to do the bmap later,
2488 	 * perhaps when the system is attempting to do a sync.  Since it
2489 	 * is likely that the indirect block -- or whatever other datastructure
2490 	 * that the filesystem needs is still in memory now, it is a good
2491 	 * thing to do this.  Note also, that if the pageout daemon is
2492 	 * requesting a sync -- there might not be enough memory to do
2493 	 * the bmap then...  So, this is important to do.
2494 	 */
2495 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2496 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2497 	}
2498 
2499 	buf_track(bp, __func__);
2500 
2501 	/*
2502 	 * Set the *dirty* buffer range based upon the VM system dirty
2503 	 * pages.
2504 	 *
2505 	 * Mark the buffer pages as clean.  We need to do this here to
2506 	 * satisfy the vnode_pager and the pageout daemon, so that it
2507 	 * thinks that the pages have been "cleaned".  Note that since
2508 	 * the pages are in a delayed write buffer -- the VFS layer
2509 	 * "will" see that the pages get written out on the next sync,
2510 	 * or perhaps the cluster will be completed.
2511 	 */
2512 	vfs_clean_pages_dirty_buf(bp);
2513 	bqrelse(bp);
2514 
2515 	/*
2516 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2517 	 * due to the softdep code.
2518 	 */
2519 }
2520 
2521 /*
2522  *	bdirty:
2523  *
2524  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2525  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2526  *	itself to properly update it in the dirty/clean lists.  We mark it
2527  *	B_DONE to ensure that any asynchronization of the buffer properly
2528  *	clears B_DONE ( else a panic will occur later ).
2529  *
2530  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2531  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2532  *	should only be called if the buffer is known-good.
2533  *
2534  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2535  *	count.
2536  *
2537  *	The buffer must be on QUEUE_NONE.
2538  */
2539 void
2540 bdirty(struct buf *bp)
2541 {
2542 
2543 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2544 	    bp, bp->b_vp, bp->b_flags);
2545 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2546 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2547 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2548 	bp->b_flags &= ~(B_RELBUF);
2549 	bp->b_iocmd = BIO_WRITE;
2550 
2551 	if ((bp->b_flags & B_DELWRI) == 0) {
2552 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2553 		reassignbuf(bp);
2554 		bdirtyadd(bp);
2555 	}
2556 }
2557 
2558 /*
2559  *	bundirty:
2560  *
2561  *	Clear B_DELWRI for buffer.
2562  *
2563  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2564  *	count.
2565  *
2566  *	The buffer must be on QUEUE_NONE.
2567  */
2568 
2569 void
2570 bundirty(struct buf *bp)
2571 {
2572 
2573 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2574 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2575 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2576 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2577 
2578 	if (bp->b_flags & B_DELWRI) {
2579 		bp->b_flags &= ~B_DELWRI;
2580 		reassignbuf(bp);
2581 		bdirtysub(bp);
2582 	}
2583 	/*
2584 	 * Since it is now being written, we can clear its deferred write flag.
2585 	 */
2586 	bp->b_flags &= ~B_DEFERRED;
2587 }
2588 
2589 /*
2590  *	bawrite:
2591  *
2592  *	Asynchronous write.  Start output on a buffer, but do not wait for
2593  *	it to complete.  The buffer is released when the output completes.
2594  *
2595  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2596  *	B_INVAL buffers.  Not us.
2597  */
2598 void
2599 bawrite(struct buf *bp)
2600 {
2601 
2602 	bp->b_flags |= B_ASYNC;
2603 	(void) bwrite(bp);
2604 }
2605 
2606 /*
2607  *	babarrierwrite:
2608  *
2609  *	Asynchronous barrier write.  Start output on a buffer, but do not
2610  *	wait for it to complete.  Place a write barrier after this write so
2611  *	that this buffer and all buffers written before it are committed to
2612  *	the disk before any buffers written after this write are committed
2613  *	to the disk.  The buffer is released when the output completes.
2614  */
2615 void
2616 babarrierwrite(struct buf *bp)
2617 {
2618 
2619 	bp->b_flags |= B_ASYNC | B_BARRIER;
2620 	(void) bwrite(bp);
2621 }
2622 
2623 /*
2624  *	bbarrierwrite:
2625  *
2626  *	Synchronous barrier write.  Start output on a buffer and wait for
2627  *	it to complete.  Place a write barrier after this write so that
2628  *	this buffer and all buffers written before it are committed to
2629  *	the disk before any buffers written after this write are committed
2630  *	to the disk.  The buffer is released when the output completes.
2631  */
2632 int
2633 bbarrierwrite(struct buf *bp)
2634 {
2635 
2636 	bp->b_flags |= B_BARRIER;
2637 	return (bwrite(bp));
2638 }
2639 
2640 /*
2641  *	bwillwrite:
2642  *
2643  *	Called prior to the locking of any vnodes when we are expecting to
2644  *	write.  We do not want to starve the buffer cache with too many
2645  *	dirty buffers so we block here.  By blocking prior to the locking
2646  *	of any vnodes we attempt to avoid the situation where a locked vnode
2647  *	prevents the various system daemons from flushing related buffers.
2648  */
2649 void
2650 bwillwrite(void)
2651 {
2652 
2653 	if (buf_dirty_count_severe()) {
2654 		mtx_lock(&bdirtylock);
2655 		while (buf_dirty_count_severe()) {
2656 			bdirtywait = 1;
2657 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2658 			    "flswai", 0);
2659 		}
2660 		mtx_unlock(&bdirtylock);
2661 	}
2662 }
2663 
2664 /*
2665  * Return true if we have too many dirty buffers.
2666  */
2667 int
2668 buf_dirty_count_severe(void)
2669 {
2670 
2671 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2672 }
2673 
2674 /*
2675  *	brelse:
2676  *
2677  *	Release a busy buffer and, if requested, free its resources.  The
2678  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2679  *	to be accessed later as a cache entity or reused for other purposes.
2680  */
2681 void
2682 brelse(struct buf *bp)
2683 {
2684 	struct mount *v_mnt;
2685 	int qindex;
2686 
2687 	/*
2688 	 * Many functions erroneously call brelse with a NULL bp under rare
2689 	 * error conditions. Simply return when called with a NULL bp.
2690 	 */
2691 	if (bp == NULL)
2692 		return;
2693 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2694 	    bp, bp->b_vp, bp->b_flags);
2695 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2696 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2697 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2698 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2699 
2700 	if (BUF_LOCKRECURSED(bp)) {
2701 		/*
2702 		 * Do not process, in particular, do not handle the
2703 		 * B_INVAL/B_RELBUF and do not release to free list.
2704 		 */
2705 		BUF_UNLOCK(bp);
2706 		return;
2707 	}
2708 
2709 	if (bp->b_flags & B_MANAGED) {
2710 		bqrelse(bp);
2711 		return;
2712 	}
2713 
2714 	if (LIST_EMPTY(&bp->b_dep)) {
2715 		bp->b_flags &= ~B_IOSTARTED;
2716 	} else {
2717 		KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2718 		    ("brelse: SU io not finished bp %p", bp));
2719 	}
2720 
2721 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2722 		BO_LOCK(bp->b_bufobj);
2723 		bp->b_vflags &= ~BV_BKGRDERR;
2724 		BO_UNLOCK(bp->b_bufobj);
2725 		bdirty(bp);
2726 	}
2727 
2728 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2729 	    (bp->b_flags & B_INVALONERR)) {
2730 		/*
2731 		 * Forced invalidation of dirty buffer contents, to be used
2732 		 * after a failed write in the rare case that the loss of the
2733 		 * contents is acceptable.  The buffer is invalidated and
2734 		 * freed.
2735 		 */
2736 		bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2737 		bp->b_flags &= ~(B_ASYNC | B_CACHE);
2738 	}
2739 
2740 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2741 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2742 	    !(bp->b_flags & B_INVAL)) {
2743 		/*
2744 		 * Failed write, redirty.  All errors except ENXIO (which
2745 		 * means the device is gone) are treated as being
2746 		 * transient.
2747 		 *
2748 		 * XXX Treating EIO as transient is not correct; the
2749 		 * contract with the local storage device drivers is that
2750 		 * they will only return EIO once the I/O is no longer
2751 		 * retriable.  Network I/O also respects this through the
2752 		 * guarantees of TCP and/or the internal retries of NFS.
2753 		 * ENOMEM might be transient, but we also have no way of
2754 		 * knowing when its ok to retry/reschedule.  In general,
2755 		 * this entire case should be made obsolete through better
2756 		 * error handling/recovery and resource scheduling.
2757 		 *
2758 		 * Do this also for buffers that failed with ENXIO, but have
2759 		 * non-empty dependencies - the soft updates code might need
2760 		 * to access the buffer to untangle them.
2761 		 *
2762 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2763 		 */
2764 		bp->b_ioflags &= ~BIO_ERROR;
2765 		bdirty(bp);
2766 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2767 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2768 		/*
2769 		 * Either a failed read I/O, or we were asked to free or not
2770 		 * cache the buffer, or we failed to write to a device that's
2771 		 * no longer present.
2772 		 */
2773 		bp->b_flags |= B_INVAL;
2774 		if (!LIST_EMPTY(&bp->b_dep))
2775 			buf_deallocate(bp);
2776 		if (bp->b_flags & B_DELWRI)
2777 			bdirtysub(bp);
2778 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2779 		if ((bp->b_flags & B_VMIO) == 0) {
2780 			allocbuf(bp, 0);
2781 			if (bp->b_vp)
2782 				brelvp(bp);
2783 		}
2784 	}
2785 
2786 	/*
2787 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2788 	 * is called with B_DELWRI set, the underlying pages may wind up
2789 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2790 	 * because pages associated with a B_DELWRI bp are marked clean.
2791 	 *
2792 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2793 	 * if B_DELWRI is set.
2794 	 */
2795 	if (bp->b_flags & B_DELWRI)
2796 		bp->b_flags &= ~B_RELBUF;
2797 
2798 	/*
2799 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2800 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2801 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2802 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2803 	 *
2804 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2805 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2806 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2807 	 *
2808 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2809 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2810 	 * the commit state and we cannot afford to lose the buffer. If the
2811 	 * buffer has a background write in progress, we need to keep it
2812 	 * around to prevent it from being reconstituted and starting a second
2813 	 * background write.
2814 	 */
2815 
2816 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2817 
2818 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2819 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2820 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2821 	    vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2822 		vfs_vmio_invalidate(bp);
2823 		allocbuf(bp, 0);
2824 	}
2825 
2826 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2827 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2828 		allocbuf(bp, 0);
2829 		bp->b_flags &= ~B_NOREUSE;
2830 		if (bp->b_vp != NULL)
2831 			brelvp(bp);
2832 	}
2833 
2834 	/*
2835 	 * If the buffer has junk contents signal it and eventually
2836 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2837 	 * doesn't find it.
2838 	 */
2839 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2840 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2841 		bp->b_flags |= B_INVAL;
2842 	if (bp->b_flags & B_INVAL) {
2843 		if (bp->b_flags & B_DELWRI)
2844 			bundirty(bp);
2845 		if (bp->b_vp)
2846 			brelvp(bp);
2847 	}
2848 
2849 	buf_track(bp, __func__);
2850 
2851 	/* buffers with no memory */
2852 	if (bp->b_bufsize == 0) {
2853 		buf_free(bp);
2854 		return;
2855 	}
2856 	/* buffers with junk contents */
2857 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2858 	    (bp->b_ioflags & BIO_ERROR)) {
2859 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2860 		if (bp->b_vflags & BV_BKGRDINPROG)
2861 			panic("losing buffer 2");
2862 		qindex = QUEUE_CLEAN;
2863 		bp->b_flags |= B_AGE;
2864 	/* remaining buffers */
2865 	} else if (bp->b_flags & B_DELWRI)
2866 		qindex = QUEUE_DIRTY;
2867 	else
2868 		qindex = QUEUE_CLEAN;
2869 
2870 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2871 		panic("brelse: not dirty");
2872 
2873 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2874 	bp->b_xflags &= ~(BX_CVTENXIO);
2875 	/* binsfree unlocks bp. */
2876 	binsfree(bp, qindex);
2877 }
2878 
2879 /*
2880  * Release a buffer back to the appropriate queue but do not try to free
2881  * it.  The buffer is expected to be used again soon.
2882  *
2883  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2884  * biodone() to requeue an async I/O on completion.  It is also used when
2885  * known good buffers need to be requeued but we think we may need the data
2886  * again soon.
2887  *
2888  * XXX we should be able to leave the B_RELBUF hint set on completion.
2889  */
2890 void
2891 bqrelse(struct buf *bp)
2892 {
2893 	int qindex;
2894 
2895 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2896 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2897 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2898 
2899 	qindex = QUEUE_NONE;
2900 	if (BUF_LOCKRECURSED(bp)) {
2901 		/* do not release to free list */
2902 		BUF_UNLOCK(bp);
2903 		return;
2904 	}
2905 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2906 	bp->b_xflags &= ~(BX_CVTENXIO);
2907 
2908 	if (LIST_EMPTY(&bp->b_dep)) {
2909 		bp->b_flags &= ~B_IOSTARTED;
2910 	} else {
2911 		KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2912 		    ("bqrelse: SU io not finished bp %p", bp));
2913 	}
2914 
2915 	if (bp->b_flags & B_MANAGED) {
2916 		if (bp->b_flags & B_REMFREE)
2917 			bremfreef(bp);
2918 		goto out;
2919 	}
2920 
2921 	/* buffers with stale but valid contents */
2922 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2923 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2924 		BO_LOCK(bp->b_bufobj);
2925 		bp->b_vflags &= ~BV_BKGRDERR;
2926 		BO_UNLOCK(bp->b_bufobj);
2927 		qindex = QUEUE_DIRTY;
2928 	} else {
2929 		if ((bp->b_flags & B_DELWRI) == 0 &&
2930 		    (bp->b_xflags & BX_VNDIRTY))
2931 			panic("bqrelse: not dirty");
2932 		if ((bp->b_flags & B_NOREUSE) != 0) {
2933 			brelse(bp);
2934 			return;
2935 		}
2936 		qindex = QUEUE_CLEAN;
2937 	}
2938 	buf_track(bp, __func__);
2939 	/* binsfree unlocks bp. */
2940 	binsfree(bp, qindex);
2941 	return;
2942 
2943 out:
2944 	buf_track(bp, __func__);
2945 	/* unlock */
2946 	BUF_UNLOCK(bp);
2947 }
2948 
2949 /*
2950  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2951  * restore bogus pages.
2952  */
2953 static void
2954 vfs_vmio_iodone(struct buf *bp)
2955 {
2956 	vm_ooffset_t foff;
2957 	vm_page_t m;
2958 	vm_object_t obj;
2959 	struct vnode *vp __unused;
2960 	int i, iosize, resid;
2961 	bool bogus;
2962 
2963 	obj = bp->b_bufobj->bo_object;
2964 	KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2965 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2966 	    blockcount_read(&obj->paging_in_progress), bp->b_npages));
2967 
2968 	vp = bp->b_vp;
2969 	VNPASS(vp->v_holdcnt > 0, vp);
2970 	VNPASS(vp->v_object != NULL, vp);
2971 
2972 	foff = bp->b_offset;
2973 	KASSERT(bp->b_offset != NOOFFSET,
2974 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2975 
2976 	bogus = false;
2977 	iosize = bp->b_bcount - bp->b_resid;
2978 	for (i = 0; i < bp->b_npages; i++) {
2979 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2980 		if (resid > iosize)
2981 			resid = iosize;
2982 
2983 		/*
2984 		 * cleanup bogus pages, restoring the originals
2985 		 */
2986 		m = bp->b_pages[i];
2987 		if (m == bogus_page) {
2988 			bogus = true;
2989 			m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2990 			if (m == NULL)
2991 				panic("biodone: page disappeared!");
2992 			bp->b_pages[i] = m;
2993 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2994 			/*
2995 			 * In the write case, the valid and clean bits are
2996 			 * already changed correctly ( see bdwrite() ), so we
2997 			 * only need to do this here in the read case.
2998 			 */
2999 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
3000 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
3001 			    "has unexpected dirty bits", m));
3002 			vfs_page_set_valid(bp, foff, m);
3003 		}
3004 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
3005 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
3006 		    (intmax_t)foff, (uintmax_t)m->pindex));
3007 
3008 		vm_page_sunbusy(m);
3009 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3010 		iosize -= resid;
3011 	}
3012 	vm_object_pip_wakeupn(obj, bp->b_npages);
3013 	if (bogus && buf_mapped(bp)) {
3014 		BUF_CHECK_MAPPED(bp);
3015 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3016 		    bp->b_pages, bp->b_npages);
3017 	}
3018 }
3019 
3020 /*
3021  * Perform page invalidation when a buffer is released.  The fully invalid
3022  * pages will be reclaimed later in vfs_vmio_truncate().
3023  */
3024 static void
3025 vfs_vmio_invalidate(struct buf *bp)
3026 {
3027 	vm_object_t obj;
3028 	vm_page_t m;
3029 	int flags, i, resid, poffset, presid;
3030 
3031 	if (buf_mapped(bp)) {
3032 		BUF_CHECK_MAPPED(bp);
3033 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
3034 	} else
3035 		BUF_CHECK_UNMAPPED(bp);
3036 	/*
3037 	 * Get the base offset and length of the buffer.  Note that
3038 	 * in the VMIO case if the buffer block size is not
3039 	 * page-aligned then b_data pointer may not be page-aligned.
3040 	 * But our b_pages[] array *IS* page aligned.
3041 	 *
3042 	 * block sizes less then DEV_BSIZE (usually 512) are not
3043 	 * supported due to the page granularity bits (m->valid,
3044 	 * m->dirty, etc...).
3045 	 *
3046 	 * See man buf(9) for more information
3047 	 */
3048 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3049 	obj = bp->b_bufobj->bo_object;
3050 	resid = bp->b_bufsize;
3051 	poffset = bp->b_offset & PAGE_MASK;
3052 	VM_OBJECT_WLOCK(obj);
3053 	for (i = 0; i < bp->b_npages; i++) {
3054 		m = bp->b_pages[i];
3055 		if (m == bogus_page)
3056 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
3057 		bp->b_pages[i] = NULL;
3058 
3059 		presid = resid > (PAGE_SIZE - poffset) ?
3060 		    (PAGE_SIZE - poffset) : resid;
3061 		KASSERT(presid >= 0, ("brelse: extra page"));
3062 		vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3063 		if (pmap_page_wired_mappings(m) == 0)
3064 			vm_page_set_invalid(m, poffset, presid);
3065 		vm_page_sunbusy(m);
3066 		vm_page_release_locked(m, flags);
3067 		resid -= presid;
3068 		poffset = 0;
3069 	}
3070 	VM_OBJECT_WUNLOCK(obj);
3071 	bp->b_npages = 0;
3072 }
3073 
3074 /*
3075  * Page-granular truncation of an existing VMIO buffer.
3076  */
3077 static void
3078 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3079 {
3080 	vm_object_t obj;
3081 	vm_page_t m;
3082 	int flags, i;
3083 
3084 	if (bp->b_npages == desiredpages)
3085 		return;
3086 
3087 	if (buf_mapped(bp)) {
3088 		BUF_CHECK_MAPPED(bp);
3089 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3090 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3091 	} else
3092 		BUF_CHECK_UNMAPPED(bp);
3093 
3094 	/*
3095 	 * The object lock is needed only if we will attempt to free pages.
3096 	 */
3097 	flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3098 	if ((bp->b_flags & B_DIRECT) != 0) {
3099 		flags |= VPR_TRYFREE;
3100 		obj = bp->b_bufobj->bo_object;
3101 		VM_OBJECT_WLOCK(obj);
3102 	} else {
3103 		obj = NULL;
3104 	}
3105 	for (i = desiredpages; i < bp->b_npages; i++) {
3106 		m = bp->b_pages[i];
3107 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3108 		bp->b_pages[i] = NULL;
3109 		if (obj != NULL)
3110 			vm_page_release_locked(m, flags);
3111 		else
3112 			vm_page_release(m, flags);
3113 	}
3114 	if (obj != NULL)
3115 		VM_OBJECT_WUNLOCK(obj);
3116 	bp->b_npages = desiredpages;
3117 }
3118 
3119 /*
3120  * Byte granular extension of VMIO buffers.
3121  */
3122 static void
3123 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3124 {
3125 	/*
3126 	 * We are growing the buffer, possibly in a
3127 	 * byte-granular fashion.
3128 	 */
3129 	vm_object_t obj;
3130 	vm_offset_t toff;
3131 	vm_offset_t tinc;
3132 	vm_page_t m;
3133 
3134 	/*
3135 	 * Step 1, bring in the VM pages from the object, allocating
3136 	 * them if necessary.  We must clear B_CACHE if these pages
3137 	 * are not valid for the range covered by the buffer.
3138 	 */
3139 	obj = bp->b_bufobj->bo_object;
3140 	if (bp->b_npages < desiredpages) {
3141 		KASSERT(desiredpages <= atop(maxbcachebuf),
3142 		    ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3143 		    bp, desiredpages, maxbcachebuf));
3144 
3145 		/*
3146 		 * We must allocate system pages since blocking
3147 		 * here could interfere with paging I/O, no
3148 		 * matter which process we are.
3149 		 *
3150 		 * Only exclusive busy can be tested here.
3151 		 * Blocking on shared busy might lead to
3152 		 * deadlocks once allocbuf() is called after
3153 		 * pages are vfs_busy_pages().
3154 		 */
3155 		(void)vm_page_grab_pages_unlocked(obj,
3156 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3157 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3158 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3159 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3160 		bp->b_npages = desiredpages;
3161 	}
3162 
3163 	/*
3164 	 * Step 2.  We've loaded the pages into the buffer,
3165 	 * we have to figure out if we can still have B_CACHE
3166 	 * set.  Note that B_CACHE is set according to the
3167 	 * byte-granular range ( bcount and size ), not the
3168 	 * aligned range ( newbsize ).
3169 	 *
3170 	 * The VM test is against m->valid, which is DEV_BSIZE
3171 	 * aligned.  Needless to say, the validity of the data
3172 	 * needs to also be DEV_BSIZE aligned.  Note that this
3173 	 * fails with NFS if the server or some other client
3174 	 * extends the file's EOF.  If our buffer is resized,
3175 	 * B_CACHE may remain set! XXX
3176 	 */
3177 	toff = bp->b_bcount;
3178 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3179 	while ((bp->b_flags & B_CACHE) && toff < size) {
3180 		vm_pindex_t pi;
3181 
3182 		if (tinc > (size - toff))
3183 			tinc = size - toff;
3184 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3185 		m = bp->b_pages[pi];
3186 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3187 		toff += tinc;
3188 		tinc = PAGE_SIZE;
3189 	}
3190 
3191 	/*
3192 	 * Step 3, fixup the KVA pmap.
3193 	 */
3194 	if (buf_mapped(bp))
3195 		bpmap_qenter(bp);
3196 	else
3197 		BUF_CHECK_UNMAPPED(bp);
3198 }
3199 
3200 /*
3201  * Check to see if a block at a particular lbn is available for a clustered
3202  * write.
3203  */
3204 static int
3205 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3206 {
3207 	struct buf *bpa;
3208 	int match;
3209 
3210 	match = 0;
3211 
3212 	/* If the buf isn't in core skip it */
3213 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3214 		return (0);
3215 
3216 	/* If the buf is busy we don't want to wait for it */
3217 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3218 		return (0);
3219 
3220 	/* Only cluster with valid clusterable delayed write buffers */
3221 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3222 	    (B_DELWRI | B_CLUSTEROK))
3223 		goto done;
3224 
3225 	if (bpa->b_bufsize != size)
3226 		goto done;
3227 
3228 	/*
3229 	 * Check to see if it is in the expected place on disk and that the
3230 	 * block has been mapped.
3231 	 */
3232 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3233 		match = 1;
3234 done:
3235 	BUF_UNLOCK(bpa);
3236 	return (match);
3237 }
3238 
3239 /*
3240  *	vfs_bio_awrite:
3241  *
3242  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3243  *	This is much better then the old way of writing only one buffer at
3244  *	a time.  Note that we may not be presented with the buffers in the
3245  *	correct order, so we search for the cluster in both directions.
3246  */
3247 int
3248 vfs_bio_awrite(struct buf *bp)
3249 {
3250 	struct bufobj *bo;
3251 	int i;
3252 	int j;
3253 	daddr_t lblkno = bp->b_lblkno;
3254 	struct vnode *vp = bp->b_vp;
3255 	int ncl;
3256 	int nwritten;
3257 	int size;
3258 	int maxcl;
3259 	int gbflags;
3260 
3261 	bo = &vp->v_bufobj;
3262 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3263 	/*
3264 	 * right now we support clustered writing only to regular files.  If
3265 	 * we find a clusterable block we could be in the middle of a cluster
3266 	 * rather then at the beginning.
3267 	 */
3268 	if ((vp->v_type == VREG) &&
3269 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3270 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3271 		size = vp->v_mount->mnt_stat.f_iosize;
3272 		maxcl = maxphys / size;
3273 
3274 		BO_RLOCK(bo);
3275 		for (i = 1; i < maxcl; i++)
3276 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3277 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3278 				break;
3279 
3280 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3281 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3282 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3283 				break;
3284 		BO_RUNLOCK(bo);
3285 		--j;
3286 		ncl = i + j;
3287 		/*
3288 		 * this is a possible cluster write
3289 		 */
3290 		if (ncl != 1) {
3291 			BUF_UNLOCK(bp);
3292 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3293 			    gbflags);
3294 			return (nwritten);
3295 		}
3296 	}
3297 	bremfree(bp);
3298 	bp->b_flags |= B_ASYNC;
3299 	/*
3300 	 * default (old) behavior, writing out only one block
3301 	 *
3302 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3303 	 */
3304 	nwritten = bp->b_bufsize;
3305 	(void) bwrite(bp);
3306 
3307 	return (nwritten);
3308 }
3309 
3310 /*
3311  *	getnewbuf_kva:
3312  *
3313  *	Allocate KVA for an empty buf header according to gbflags.
3314  */
3315 static int
3316 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3317 {
3318 
3319 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3320 		/*
3321 		 * In order to keep fragmentation sane we only allocate kva
3322 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3323 		 */
3324 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3325 
3326 		if (maxsize != bp->b_kvasize &&
3327 		    bufkva_alloc(bp, maxsize, gbflags))
3328 			return (ENOSPC);
3329 	}
3330 	return (0);
3331 }
3332 
3333 /*
3334  *	getnewbuf:
3335  *
3336  *	Find and initialize a new buffer header, freeing up existing buffers
3337  *	in the bufqueues as necessary.  The new buffer is returned locked.
3338  *
3339  *	We block if:
3340  *		We have insufficient buffer headers
3341  *		We have insufficient buffer space
3342  *		buffer_arena is too fragmented ( space reservation fails )
3343  *		If we have to flush dirty buffers ( but we try to avoid this )
3344  *
3345  *	The caller is responsible for releasing the reserved bufspace after
3346  *	allocbuf() is called.
3347  */
3348 static struct buf *
3349 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3350 {
3351 	struct bufdomain *bd;
3352 	struct buf *bp;
3353 	bool metadata, reserved;
3354 
3355 	bp = NULL;
3356 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3357 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3358 	if (!unmapped_buf_allowed)
3359 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3360 
3361 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3362 	    vp->v_type == VCHR)
3363 		metadata = true;
3364 	else
3365 		metadata = false;
3366 	if (vp == NULL)
3367 		bd = &bdomain[0];
3368 	else
3369 		bd = &bdomain[vp->v_bufobj.bo_domain];
3370 
3371 	counter_u64_add(getnewbufcalls, 1);
3372 	reserved = false;
3373 	do {
3374 		if (reserved == false &&
3375 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3376 			counter_u64_add(getnewbufrestarts, 1);
3377 			continue;
3378 		}
3379 		reserved = true;
3380 		if ((bp = buf_alloc(bd)) == NULL) {
3381 			counter_u64_add(getnewbufrestarts, 1);
3382 			continue;
3383 		}
3384 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3385 			return (bp);
3386 		break;
3387 	} while (buf_recycle(bd, false) == 0);
3388 
3389 	if (reserved)
3390 		bufspace_release(bd, maxsize);
3391 	if (bp != NULL) {
3392 		bp->b_flags |= B_INVAL;
3393 		brelse(bp);
3394 	}
3395 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3396 
3397 	return (NULL);
3398 }
3399 
3400 /*
3401  *	buf_daemon:
3402  *
3403  *	buffer flushing daemon.  Buffers are normally flushed by the
3404  *	update daemon but if it cannot keep up this process starts to
3405  *	take the load in an attempt to prevent getnewbuf() from blocking.
3406  */
3407 static struct kproc_desc buf_kp = {
3408 	"bufdaemon",
3409 	buf_daemon,
3410 	&bufdaemonproc
3411 };
3412 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3413 
3414 static int
3415 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3416 {
3417 	int flushed;
3418 
3419 	flushed = flushbufqueues(vp, bd, target, 0);
3420 	if (flushed == 0) {
3421 		/*
3422 		 * Could not find any buffers without rollback
3423 		 * dependencies, so just write the first one
3424 		 * in the hopes of eventually making progress.
3425 		 */
3426 		if (vp != NULL && target > 2)
3427 			target /= 2;
3428 		flushbufqueues(vp, bd, target, 1);
3429 	}
3430 	return (flushed);
3431 }
3432 
3433 static void
3434 buf_daemon_shutdown(void *arg __unused, int howto __unused)
3435 {
3436 	int error;
3437 
3438 	if (KERNEL_PANICKED())
3439 		return;
3440 
3441 	mtx_lock(&bdlock);
3442 	bd_shutdown = true;
3443 	wakeup(&bd_request);
3444 	error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown",
3445 	    60 * hz);
3446 	mtx_unlock(&bdlock);
3447 	if (error != 0)
3448 		printf("bufdaemon wait error: %d\n", error);
3449 }
3450 
3451 static void
3452 buf_daemon(void)
3453 {
3454 	struct bufdomain *bd;
3455 	int speedupreq;
3456 	int lodirty;
3457 	int i;
3458 
3459 	/*
3460 	 * This process needs to be suspended prior to shutdown sync.
3461 	 */
3462 	EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL,
3463 	    SHUTDOWN_PRI_LAST + 100);
3464 
3465 	/*
3466 	 * Start the buf clean daemons as children threads.
3467 	 */
3468 	for (i = 0 ; i < buf_domains; i++) {
3469 		int error;
3470 
3471 		error = kthread_add((void (*)(void *))bufspace_daemon,
3472 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3473 		if (error)
3474 			panic("error %d spawning bufspace daemon", error);
3475 	}
3476 
3477 	/*
3478 	 * This process is allowed to take the buffer cache to the limit
3479 	 */
3480 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3481 	mtx_lock(&bdlock);
3482 	while (!bd_shutdown) {
3483 		bd_request = 0;
3484 		mtx_unlock(&bdlock);
3485 
3486 		/*
3487 		 * Save speedupreq for this pass and reset to capture new
3488 		 * requests.
3489 		 */
3490 		speedupreq = bd_speedupreq;
3491 		bd_speedupreq = 0;
3492 
3493 		/*
3494 		 * Flush each domain sequentially according to its level and
3495 		 * the speedup request.
3496 		 */
3497 		for (i = 0; i < buf_domains; i++) {
3498 			bd = &bdomain[i];
3499 			if (speedupreq)
3500 				lodirty = bd->bd_numdirtybuffers / 2;
3501 			else
3502 				lodirty = bd->bd_lodirtybuffers;
3503 			while (bd->bd_numdirtybuffers > lodirty) {
3504 				if (buf_flush(NULL, bd,
3505 				    bd->bd_numdirtybuffers - lodirty) == 0)
3506 					break;
3507 				kern_yield(PRI_USER);
3508 			}
3509 		}
3510 
3511 		/*
3512 		 * Only clear bd_request if we have reached our low water
3513 		 * mark.  The buf_daemon normally waits 1 second and
3514 		 * then incrementally flushes any dirty buffers that have
3515 		 * built up, within reason.
3516 		 *
3517 		 * If we were unable to hit our low water mark and couldn't
3518 		 * find any flushable buffers, we sleep for a short period
3519 		 * to avoid endless loops on unlockable buffers.
3520 		 */
3521 		mtx_lock(&bdlock);
3522 		if (bd_shutdown)
3523 			break;
3524 		if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3525 			/*
3526 			 * We reached our low water mark, reset the
3527 			 * request and sleep until we are needed again.
3528 			 * The sleep is just so the suspend code works.
3529 			 */
3530 			bd_request = 0;
3531 			/*
3532 			 * Do an extra wakeup in case dirty threshold
3533 			 * changed via sysctl and the explicit transition
3534 			 * out of shortfall was missed.
3535 			 */
3536 			bdirtywakeup();
3537 			if (runningbufspace <= lorunningspace)
3538 				runningwakeup();
3539 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3540 		} else {
3541 			/*
3542 			 * We couldn't find any flushable dirty buffers but
3543 			 * still have too many dirty buffers, we
3544 			 * have to sleep and try again.  (rare)
3545 			 */
3546 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3547 		}
3548 	}
3549 	wakeup(&bd_shutdown);
3550 	mtx_unlock(&bdlock);
3551 	kthread_exit();
3552 }
3553 
3554 /*
3555  *	flushbufqueues:
3556  *
3557  *	Try to flush a buffer in the dirty queue.  We must be careful to
3558  *	free up B_INVAL buffers instead of write them, which NFS is
3559  *	particularly sensitive to.
3560  */
3561 static int flushwithdeps = 0;
3562 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3563     &flushwithdeps, 0,
3564     "Number of buffers flushed with dependencies that require rollbacks");
3565 
3566 static int
3567 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3568     int flushdeps)
3569 {
3570 	struct bufqueue *bq;
3571 	struct buf *sentinel;
3572 	struct vnode *vp;
3573 	struct mount *mp;
3574 	struct buf *bp;
3575 	int hasdeps;
3576 	int flushed;
3577 	int error;
3578 	bool unlock;
3579 
3580 	flushed = 0;
3581 	bq = &bd->bd_dirtyq;
3582 	bp = NULL;
3583 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3584 	sentinel->b_qindex = QUEUE_SENTINEL;
3585 	BQ_LOCK(bq);
3586 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3587 	BQ_UNLOCK(bq);
3588 	while (flushed != target) {
3589 		maybe_yield();
3590 		BQ_LOCK(bq);
3591 		bp = TAILQ_NEXT(sentinel, b_freelist);
3592 		if (bp != NULL) {
3593 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3594 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3595 			    b_freelist);
3596 		} else {
3597 			BQ_UNLOCK(bq);
3598 			break;
3599 		}
3600 		/*
3601 		 * Skip sentinels inserted by other invocations of the
3602 		 * flushbufqueues(), taking care to not reorder them.
3603 		 *
3604 		 * Only flush the buffers that belong to the
3605 		 * vnode locked by the curthread.
3606 		 */
3607 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3608 		    bp->b_vp != lvp)) {
3609 			BQ_UNLOCK(bq);
3610 			continue;
3611 		}
3612 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3613 		BQ_UNLOCK(bq);
3614 		if (error != 0)
3615 			continue;
3616 
3617 		/*
3618 		 * BKGRDINPROG can only be set with the buf and bufobj
3619 		 * locks both held.  We tolerate a race to clear it here.
3620 		 */
3621 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3622 		    (bp->b_flags & B_DELWRI) == 0) {
3623 			BUF_UNLOCK(bp);
3624 			continue;
3625 		}
3626 		if (bp->b_flags & B_INVAL) {
3627 			bremfreef(bp);
3628 			brelse(bp);
3629 			flushed++;
3630 			continue;
3631 		}
3632 
3633 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3634 			if (flushdeps == 0) {
3635 				BUF_UNLOCK(bp);
3636 				continue;
3637 			}
3638 			hasdeps = 1;
3639 		} else
3640 			hasdeps = 0;
3641 		/*
3642 		 * We must hold the lock on a vnode before writing
3643 		 * one of its buffers. Otherwise we may confuse, or
3644 		 * in the case of a snapshot vnode, deadlock the
3645 		 * system.
3646 		 *
3647 		 * The lock order here is the reverse of the normal
3648 		 * of vnode followed by buf lock.  This is ok because
3649 		 * the NOWAIT will prevent deadlock.
3650 		 */
3651 		vp = bp->b_vp;
3652 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3653 			BUF_UNLOCK(bp);
3654 			continue;
3655 		}
3656 		if (lvp == NULL) {
3657 			unlock = true;
3658 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3659 		} else {
3660 			ASSERT_VOP_LOCKED(vp, "getbuf");
3661 			unlock = false;
3662 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3663 			    vn_lock(vp, LK_TRYUPGRADE);
3664 		}
3665 		if (error == 0) {
3666 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3667 			    bp, bp->b_vp, bp->b_flags);
3668 			if (curproc == bufdaemonproc) {
3669 				vfs_bio_awrite(bp);
3670 			} else {
3671 				bremfree(bp);
3672 				bwrite(bp);
3673 				counter_u64_add(notbufdflushes, 1);
3674 			}
3675 			vn_finished_write(mp);
3676 			if (unlock)
3677 				VOP_UNLOCK(vp);
3678 			flushwithdeps += hasdeps;
3679 			flushed++;
3680 
3681 			/*
3682 			 * Sleeping on runningbufspace while holding
3683 			 * vnode lock leads to deadlock.
3684 			 */
3685 			if (curproc == bufdaemonproc &&
3686 			    runningbufspace > hirunningspace)
3687 				waitrunningbufspace();
3688 			continue;
3689 		}
3690 		vn_finished_write(mp);
3691 		BUF_UNLOCK(bp);
3692 	}
3693 	BQ_LOCK(bq);
3694 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3695 	BQ_UNLOCK(bq);
3696 	free(sentinel, M_TEMP);
3697 	return (flushed);
3698 }
3699 
3700 /*
3701  * Check to see if a block is currently memory resident.
3702  */
3703 struct buf *
3704 incore(struct bufobj *bo, daddr_t blkno)
3705 {
3706 	return (gbincore_unlocked(bo, blkno));
3707 }
3708 
3709 /*
3710  * Returns true if no I/O is needed to access the
3711  * associated VM object.  This is like incore except
3712  * it also hunts around in the VM system for the data.
3713  */
3714 bool
3715 inmem(struct vnode * vp, daddr_t blkno)
3716 {
3717 	vm_object_t obj;
3718 	vm_offset_t toff, tinc, size;
3719 	vm_page_t m, n;
3720 	vm_ooffset_t off;
3721 	int valid;
3722 
3723 	ASSERT_VOP_LOCKED(vp, "inmem");
3724 
3725 	if (incore(&vp->v_bufobj, blkno))
3726 		return (true);
3727 	if (vp->v_mount == NULL)
3728 		return (false);
3729 	obj = vp->v_object;
3730 	if (obj == NULL)
3731 		return (false);
3732 
3733 	size = PAGE_SIZE;
3734 	if (size > vp->v_mount->mnt_stat.f_iosize)
3735 		size = vp->v_mount->mnt_stat.f_iosize;
3736 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3737 
3738 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3739 		m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3740 recheck:
3741 		if (m == NULL)
3742 			return (false);
3743 
3744 		tinc = size;
3745 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3746 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3747 		/*
3748 		 * Consider page validity only if page mapping didn't change
3749 		 * during the check.
3750 		 */
3751 		valid = vm_page_is_valid(m,
3752 		    (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3753 		n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3754 		if (m != n) {
3755 			m = n;
3756 			goto recheck;
3757 		}
3758 		if (!valid)
3759 			return (false);
3760 	}
3761 	return (true);
3762 }
3763 
3764 /*
3765  * Set the dirty range for a buffer based on the status of the dirty
3766  * bits in the pages comprising the buffer.  The range is limited
3767  * to the size of the buffer.
3768  *
3769  * Tell the VM system that the pages associated with this buffer
3770  * are clean.  This is used for delayed writes where the data is
3771  * going to go to disk eventually without additional VM intevention.
3772  *
3773  * Note that while we only really need to clean through to b_bcount, we
3774  * just go ahead and clean through to b_bufsize.
3775  */
3776 static void
3777 vfs_clean_pages_dirty_buf(struct buf *bp)
3778 {
3779 	vm_ooffset_t foff, noff, eoff;
3780 	vm_page_t m;
3781 	int i;
3782 
3783 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3784 		return;
3785 
3786 	foff = bp->b_offset;
3787 	KASSERT(bp->b_offset != NOOFFSET,
3788 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3789 
3790 	vfs_busy_pages_acquire(bp);
3791 	vfs_setdirty_range(bp);
3792 	for (i = 0; i < bp->b_npages; i++) {
3793 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3794 		eoff = noff;
3795 		if (eoff > bp->b_offset + bp->b_bufsize)
3796 			eoff = bp->b_offset + bp->b_bufsize;
3797 		m = bp->b_pages[i];
3798 		vfs_page_set_validclean(bp, foff, m);
3799 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3800 		foff = noff;
3801 	}
3802 	vfs_busy_pages_release(bp);
3803 }
3804 
3805 static void
3806 vfs_setdirty_range(struct buf *bp)
3807 {
3808 	vm_offset_t boffset;
3809 	vm_offset_t eoffset;
3810 	int i;
3811 
3812 	/*
3813 	 * test the pages to see if they have been modified directly
3814 	 * by users through the VM system.
3815 	 */
3816 	for (i = 0; i < bp->b_npages; i++)
3817 		vm_page_test_dirty(bp->b_pages[i]);
3818 
3819 	/*
3820 	 * Calculate the encompassing dirty range, boffset and eoffset,
3821 	 * (eoffset - boffset) bytes.
3822 	 */
3823 
3824 	for (i = 0; i < bp->b_npages; i++) {
3825 		if (bp->b_pages[i]->dirty)
3826 			break;
3827 	}
3828 	boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3829 
3830 	for (i = bp->b_npages - 1; i >= 0; --i) {
3831 		if (bp->b_pages[i]->dirty) {
3832 			break;
3833 		}
3834 	}
3835 	eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3836 
3837 	/*
3838 	 * Fit it to the buffer.
3839 	 */
3840 
3841 	if (eoffset > bp->b_bcount)
3842 		eoffset = bp->b_bcount;
3843 
3844 	/*
3845 	 * If we have a good dirty range, merge with the existing
3846 	 * dirty range.
3847 	 */
3848 
3849 	if (boffset < eoffset) {
3850 		if (bp->b_dirtyoff > boffset)
3851 			bp->b_dirtyoff = boffset;
3852 		if (bp->b_dirtyend < eoffset)
3853 			bp->b_dirtyend = eoffset;
3854 	}
3855 }
3856 
3857 /*
3858  * Allocate the KVA mapping for an existing buffer.
3859  * If an unmapped buffer is provided but a mapped buffer is requested, take
3860  * also care to properly setup mappings between pages and KVA.
3861  */
3862 static void
3863 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3864 {
3865 	int bsize, maxsize, need_mapping, need_kva;
3866 	off_t offset;
3867 
3868 	need_mapping = bp->b_data == unmapped_buf &&
3869 	    (gbflags & GB_UNMAPPED) == 0;
3870 	need_kva = bp->b_kvabase == unmapped_buf &&
3871 	    bp->b_data == unmapped_buf &&
3872 	    (gbflags & GB_KVAALLOC) != 0;
3873 	if (!need_mapping && !need_kva)
3874 		return;
3875 
3876 	BUF_CHECK_UNMAPPED(bp);
3877 
3878 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3879 		/*
3880 		 * Buffer is not mapped, but the KVA was already
3881 		 * reserved at the time of the instantiation.  Use the
3882 		 * allocated space.
3883 		 */
3884 		goto has_addr;
3885 	}
3886 
3887 	/*
3888 	 * Calculate the amount of the address space we would reserve
3889 	 * if the buffer was mapped.
3890 	 */
3891 	bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3892 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3893 	offset = blkno * bsize;
3894 	maxsize = size + (offset & PAGE_MASK);
3895 	maxsize = imax(maxsize, bsize);
3896 
3897 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3898 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3899 			/*
3900 			 * XXXKIB: defragmentation cannot
3901 			 * succeed, not sure what else to do.
3902 			 */
3903 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3904 		}
3905 		counter_u64_add(mappingrestarts, 1);
3906 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3907 	}
3908 has_addr:
3909 	if (need_mapping) {
3910 		/* b_offset is handled by bpmap_qenter. */
3911 		bp->b_data = bp->b_kvabase;
3912 		BUF_CHECK_MAPPED(bp);
3913 		bpmap_qenter(bp);
3914 	}
3915 }
3916 
3917 struct buf *
3918 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3919     int flags)
3920 {
3921 	struct buf *bp;
3922 	int error;
3923 
3924 	error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3925 	if (error != 0)
3926 		return (NULL);
3927 	return (bp);
3928 }
3929 
3930 /*
3931  *	getblkx:
3932  *
3933  *	Get a block given a specified block and offset into a file/device.
3934  *	The buffers B_DONE bit will be cleared on return, making it almost
3935  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3936  *	return.  The caller should clear B_INVAL prior to initiating a
3937  *	READ.
3938  *
3939  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3940  *	an existing buffer.
3941  *
3942  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3943  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3944  *	and then cleared based on the backing VM.  If the previous buffer is
3945  *	non-0-sized but invalid, B_CACHE will be cleared.
3946  *
3947  *	If getblk() must create a new buffer, the new buffer is returned with
3948  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3949  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3950  *	backing VM.
3951  *
3952  *	getblk() also forces a bwrite() for any B_DELWRI buffer whose
3953  *	B_CACHE bit is clear.
3954  *
3955  *	What this means, basically, is that the caller should use B_CACHE to
3956  *	determine whether the buffer is fully valid or not and should clear
3957  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3958  *	the buffer by loading its data area with something, the caller needs
3959  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3960  *	the caller should set B_CACHE ( as an optimization ), else the caller
3961  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3962  *	a write attempt or if it was a successful read.  If the caller
3963  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3964  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3965  *
3966  *	The blkno parameter is the logical block being requested. Normally
3967  *	the mapping of logical block number to disk block address is done
3968  *	by calling VOP_BMAP(). However, if the mapping is already known, the
3969  *	disk block address can be passed using the dblkno parameter. If the
3970  *	disk block address is not known, then the same value should be passed
3971  *	for blkno and dblkno.
3972  */
3973 int
3974 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3975     int slptimeo, int flags, struct buf **bpp)
3976 {
3977 	struct buf *bp;
3978 	struct bufobj *bo;
3979 	daddr_t d_blkno;
3980 	int bsize, error, maxsize, vmio;
3981 	off_t offset;
3982 
3983 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3984 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3985 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3986 	if (vp->v_type != VCHR)
3987 		ASSERT_VOP_LOCKED(vp, "getblk");
3988 	if (size > maxbcachebuf) {
3989 		printf("getblkx: size(%d) > maxbcachebuf(%d)\n", size,
3990 		    maxbcachebuf);
3991 		return (EIO);
3992 	}
3993 	if (!unmapped_buf_allowed)
3994 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3995 
3996 	bo = &vp->v_bufobj;
3997 	d_blkno = dblkno;
3998 
3999 	/* Attempt lockless lookup first. */
4000 	bp = gbincore_unlocked(bo, blkno);
4001 	if (bp == NULL) {
4002 		/*
4003 		 * With GB_NOCREAT we must be sure about not finding the buffer
4004 		 * as it may have been reassigned during unlocked lookup.
4005 		 */
4006 		if ((flags & GB_NOCREAT) != 0)
4007 			goto loop;
4008 		goto newbuf_unlocked;
4009 	}
4010 
4011 	error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
4012 	    0);
4013 	if (error != 0)
4014 		goto loop;
4015 
4016 	/* Verify buf identify has not changed since lookup. */
4017 	if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
4018 		goto foundbuf_fastpath;
4019 
4020 	/* It changed, fallback to locked lookup. */
4021 	BUF_UNLOCK_RAW(bp);
4022 
4023 loop:
4024 	BO_RLOCK(bo);
4025 	bp = gbincore(bo, blkno);
4026 	if (bp != NULL) {
4027 		int lockflags;
4028 
4029 		/*
4030 		 * Buffer is in-core.  If the buffer is not busy nor managed,
4031 		 * it must be on a queue.
4032 		 */
4033 		lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
4034 		    ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL);
4035 #ifdef WITNESS
4036 		lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0;
4037 #endif
4038 
4039 		error = BUF_TIMELOCK(bp, lockflags,
4040 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
4041 
4042 		/*
4043 		 * If we slept and got the lock we have to restart in case
4044 		 * the buffer changed identities.
4045 		 */
4046 		if (error == ENOLCK)
4047 			goto loop;
4048 		/* We timed out or were interrupted. */
4049 		else if (error != 0)
4050 			return (error);
4051 
4052 foundbuf_fastpath:
4053 		/* If recursed, assume caller knows the rules. */
4054 		if (BUF_LOCKRECURSED(bp))
4055 			goto end;
4056 
4057 		/*
4058 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
4059 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
4060 		 * and for a VMIO buffer B_CACHE is adjusted according to the
4061 		 * backing VM cache.
4062 		 */
4063 		if (bp->b_flags & B_INVAL)
4064 			bp->b_flags &= ~B_CACHE;
4065 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4066 			bp->b_flags |= B_CACHE;
4067 		if (bp->b_flags & B_MANAGED)
4068 			MPASS(bp->b_qindex == QUEUE_NONE);
4069 		else
4070 			bremfree(bp);
4071 
4072 		/*
4073 		 * check for size inconsistencies for non-VMIO case.
4074 		 */
4075 		if (bp->b_bcount != size) {
4076 			if ((bp->b_flags & B_VMIO) == 0 ||
4077 			    (size > bp->b_kvasize)) {
4078 				if (bp->b_flags & B_DELWRI) {
4079 					bp->b_flags |= B_NOCACHE;
4080 					bwrite(bp);
4081 				} else {
4082 					if (LIST_EMPTY(&bp->b_dep)) {
4083 						bp->b_flags |= B_RELBUF;
4084 						brelse(bp);
4085 					} else {
4086 						bp->b_flags |= B_NOCACHE;
4087 						bwrite(bp);
4088 					}
4089 				}
4090 				goto loop;
4091 			}
4092 		}
4093 
4094 		/*
4095 		 * Handle the case of unmapped buffer which should
4096 		 * become mapped, or the buffer for which KVA
4097 		 * reservation is requested.
4098 		 */
4099 		bp_unmapped_get_kva(bp, blkno, size, flags);
4100 
4101 		/*
4102 		 * If the size is inconsistent in the VMIO case, we can resize
4103 		 * the buffer.  This might lead to B_CACHE getting set or
4104 		 * cleared.  If the size has not changed, B_CACHE remains
4105 		 * unchanged from its previous state.
4106 		 */
4107 		allocbuf(bp, size);
4108 
4109 		KASSERT(bp->b_offset != NOOFFSET,
4110 		    ("getblk: no buffer offset"));
4111 
4112 		/*
4113 		 * A buffer with B_DELWRI set and B_CACHE clear must
4114 		 * be committed before we can return the buffer in
4115 		 * order to prevent the caller from issuing a read
4116 		 * ( due to B_CACHE not being set ) and overwriting
4117 		 * it.
4118 		 *
4119 		 * Most callers, including NFS and FFS, need this to
4120 		 * operate properly either because they assume they
4121 		 * can issue a read if B_CACHE is not set, or because
4122 		 * ( for example ) an uncached B_DELWRI might loop due
4123 		 * to softupdates re-dirtying the buffer.  In the latter
4124 		 * case, B_CACHE is set after the first write completes,
4125 		 * preventing further loops.
4126 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
4127 		 * above while extending the buffer, we cannot allow the
4128 		 * buffer to remain with B_CACHE set after the write
4129 		 * completes or it will represent a corrupt state.  To
4130 		 * deal with this we set B_NOCACHE to scrap the buffer
4131 		 * after the write.
4132 		 *
4133 		 * We might be able to do something fancy, like setting
4134 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
4135 		 * so the below call doesn't set B_CACHE, but that gets real
4136 		 * confusing.  This is much easier.
4137 		 */
4138 
4139 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4140 			bp->b_flags |= B_NOCACHE;
4141 			bwrite(bp);
4142 			goto loop;
4143 		}
4144 		bp->b_flags &= ~B_DONE;
4145 	} else {
4146 		/*
4147 		 * Buffer is not in-core, create new buffer.  The buffer
4148 		 * returned by getnewbuf() is locked.  Note that the returned
4149 		 * buffer is also considered valid (not marked B_INVAL).
4150 		 */
4151 		BO_RUNLOCK(bo);
4152 newbuf_unlocked:
4153 		/*
4154 		 * If the user does not want us to create the buffer, bail out
4155 		 * here.
4156 		 */
4157 		if (flags & GB_NOCREAT)
4158 			return (EEXIST);
4159 
4160 		bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4161 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4162 		offset = blkno * bsize;
4163 		vmio = vp->v_object != NULL;
4164 		if (vmio) {
4165 			maxsize = size + (offset & PAGE_MASK);
4166 			if (maxsize > maxbcachebuf) {
4167 				printf(
4168 			    "getblkx: maxsize(%d) > maxbcachebuf(%d)\n",
4169 				    maxsize, maxbcachebuf);
4170 				return (EIO);
4171 			}
4172 		} else {
4173 			maxsize = size;
4174 			/* Do not allow non-VMIO notmapped buffers. */
4175 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4176 		}
4177 		maxsize = imax(maxsize, bsize);
4178 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4179 		    !vn_isdisk(vp)) {
4180 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4181 			KASSERT(error != EOPNOTSUPP,
4182 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4183 			    vp));
4184 			if (error != 0)
4185 				return (error);
4186 			if (d_blkno == -1)
4187 				return (EJUSTRETURN);
4188 		}
4189 
4190 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4191 		if (bp == NULL) {
4192 			if (slpflag || slptimeo)
4193 				return (ETIMEDOUT);
4194 			/*
4195 			 * XXX This is here until the sleep path is diagnosed
4196 			 * enough to work under very low memory conditions.
4197 			 *
4198 			 * There's an issue on low memory, 4BSD+non-preempt
4199 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4200 			 * exhaustion occurs without sleeping for buffer
4201 			 * reclaimation.  This just sticks in a loop and
4202 			 * constantly attempts to allocate a buffer, which
4203 			 * hits exhaustion and tries to wakeup bufdaemon.
4204 			 * This never happens because we never yield.
4205 			 *
4206 			 * The real solution is to identify and fix these cases
4207 			 * so we aren't effectively busy-waiting in a loop
4208 			 * until the reclaimation path has cycles to run.
4209 			 */
4210 			kern_yield(PRI_USER);
4211 			goto loop;
4212 		}
4213 
4214 		/*
4215 		 * This code is used to make sure that a buffer is not
4216 		 * created while the getnewbuf routine is blocked.
4217 		 * This can be a problem whether the vnode is locked or not.
4218 		 * If the buffer is created out from under us, we have to
4219 		 * throw away the one we just created.
4220 		 *
4221 		 * Note: this must occur before we associate the buffer
4222 		 * with the vp especially considering limitations in
4223 		 * the splay tree implementation when dealing with duplicate
4224 		 * lblkno's.
4225 		 */
4226 		BO_LOCK(bo);
4227 		if (gbincore(bo, blkno)) {
4228 			BO_UNLOCK(bo);
4229 			bp->b_flags |= B_INVAL;
4230 			bufspace_release(bufdomain(bp), maxsize);
4231 			brelse(bp);
4232 			goto loop;
4233 		}
4234 
4235 		/*
4236 		 * Insert the buffer into the hash, so that it can
4237 		 * be found by incore.
4238 		 */
4239 		bp->b_lblkno = blkno;
4240 		bp->b_blkno = d_blkno;
4241 		bp->b_offset = offset;
4242 		bgetvp(vp, bp);
4243 		BO_UNLOCK(bo);
4244 
4245 		/*
4246 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4247 		 * buffer size starts out as 0, B_CACHE will be set by
4248 		 * allocbuf() for the VMIO case prior to it testing the
4249 		 * backing store for validity.
4250 		 */
4251 
4252 		if (vmio) {
4253 			bp->b_flags |= B_VMIO;
4254 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4255 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4256 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4257 		} else {
4258 			bp->b_flags &= ~B_VMIO;
4259 			KASSERT(bp->b_bufobj->bo_object == NULL,
4260 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4261 			    bp, bp->b_bufobj->bo_object));
4262 			BUF_CHECK_MAPPED(bp);
4263 		}
4264 
4265 		allocbuf(bp, size);
4266 		bufspace_release(bufdomain(bp), maxsize);
4267 		bp->b_flags &= ~B_DONE;
4268 	}
4269 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4270 end:
4271 	buf_track(bp, __func__);
4272 	KASSERT(bp->b_bufobj == bo,
4273 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4274 	*bpp = bp;
4275 	return (0);
4276 }
4277 
4278 /*
4279  * Get an empty, disassociated buffer of given size.  The buffer is initially
4280  * set to B_INVAL.
4281  */
4282 struct buf *
4283 geteblk(int size, int flags)
4284 {
4285 	struct buf *bp;
4286 	int maxsize;
4287 
4288 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4289 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4290 		if ((flags & GB_NOWAIT_BD) &&
4291 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4292 			return (NULL);
4293 	}
4294 	allocbuf(bp, size);
4295 	bufspace_release(bufdomain(bp), maxsize);
4296 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4297 	return (bp);
4298 }
4299 
4300 /*
4301  * Truncate the backing store for a non-vmio buffer.
4302  */
4303 static void
4304 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4305 {
4306 
4307 	if (bp->b_flags & B_MALLOC) {
4308 		/*
4309 		 * malloced buffers are not shrunk
4310 		 */
4311 		if (newbsize == 0) {
4312 			bufmallocadjust(bp, 0);
4313 			free(bp->b_data, M_BIOBUF);
4314 			bp->b_data = bp->b_kvabase;
4315 			bp->b_flags &= ~B_MALLOC;
4316 		}
4317 		return;
4318 	}
4319 	vm_hold_free_pages(bp, newbsize);
4320 	bufspace_adjust(bp, newbsize);
4321 }
4322 
4323 /*
4324  * Extend the backing for a non-VMIO buffer.
4325  */
4326 static void
4327 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4328 {
4329 	caddr_t origbuf;
4330 	int origbufsize;
4331 
4332 	/*
4333 	 * We only use malloced memory on the first allocation.
4334 	 * and revert to page-allocated memory when the buffer
4335 	 * grows.
4336 	 *
4337 	 * There is a potential smp race here that could lead
4338 	 * to bufmallocspace slightly passing the max.  It
4339 	 * is probably extremely rare and not worth worrying
4340 	 * over.
4341 	 */
4342 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4343 	    bufmallocspace < maxbufmallocspace) {
4344 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4345 		bp->b_flags |= B_MALLOC;
4346 		bufmallocadjust(bp, newbsize);
4347 		return;
4348 	}
4349 
4350 	/*
4351 	 * If the buffer is growing on its other-than-first
4352 	 * allocation then we revert to the page-allocation
4353 	 * scheme.
4354 	 */
4355 	origbuf = NULL;
4356 	origbufsize = 0;
4357 	if (bp->b_flags & B_MALLOC) {
4358 		origbuf = bp->b_data;
4359 		origbufsize = bp->b_bufsize;
4360 		bp->b_data = bp->b_kvabase;
4361 		bufmallocadjust(bp, 0);
4362 		bp->b_flags &= ~B_MALLOC;
4363 		newbsize = round_page(newbsize);
4364 	}
4365 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4366 	    (vm_offset_t) bp->b_data + newbsize);
4367 	if (origbuf != NULL) {
4368 		bcopy(origbuf, bp->b_data, origbufsize);
4369 		free(origbuf, M_BIOBUF);
4370 	}
4371 	bufspace_adjust(bp, newbsize);
4372 }
4373 
4374 /*
4375  * This code constitutes the buffer memory from either anonymous system
4376  * memory (in the case of non-VMIO operations) or from an associated
4377  * VM object (in the case of VMIO operations).  This code is able to
4378  * resize a buffer up or down.
4379  *
4380  * Note that this code is tricky, and has many complications to resolve
4381  * deadlock or inconsistent data situations.  Tread lightly!!!
4382  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4383  * the caller.  Calling this code willy nilly can result in the loss of data.
4384  *
4385  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4386  * B_CACHE for the non-VMIO case.
4387  */
4388 int
4389 allocbuf(struct buf *bp, int size)
4390 {
4391 	int newbsize;
4392 
4393 	if (bp->b_bcount == size)
4394 		return (1);
4395 
4396 	KASSERT(bp->b_kvasize == 0 || bp->b_kvasize >= size,
4397 	    ("allocbuf: buffer too small %p %#x %#x",
4398 	    bp, bp->b_kvasize, size));
4399 
4400 	newbsize = roundup2(size, DEV_BSIZE);
4401 	if ((bp->b_flags & B_VMIO) == 0) {
4402 		if ((bp->b_flags & B_MALLOC) == 0)
4403 			newbsize = round_page(newbsize);
4404 		/*
4405 		 * Just get anonymous memory from the kernel.  Don't
4406 		 * mess with B_CACHE.
4407 		 */
4408 		if (newbsize < bp->b_bufsize)
4409 			vfs_nonvmio_truncate(bp, newbsize);
4410 		else if (newbsize > bp->b_bufsize)
4411 			vfs_nonvmio_extend(bp, newbsize);
4412 	} else {
4413 		int desiredpages;
4414 
4415 		desiredpages = size == 0 ? 0 :
4416 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4417 
4418 		KASSERT((bp->b_flags & B_MALLOC) == 0,
4419 		    ("allocbuf: VMIO buffer can't be malloced %p", bp));
4420 
4421 		/*
4422 		 * Set B_CACHE initially if buffer is 0 length or will become
4423 		 * 0-length.
4424 		 */
4425 		if (size == 0 || bp->b_bufsize == 0)
4426 			bp->b_flags |= B_CACHE;
4427 
4428 		if (newbsize < bp->b_bufsize)
4429 			vfs_vmio_truncate(bp, desiredpages);
4430 		/* XXX This looks as if it should be newbsize > b_bufsize */
4431 		else if (size > bp->b_bcount)
4432 			vfs_vmio_extend(bp, desiredpages, size);
4433 		bufspace_adjust(bp, newbsize);
4434 	}
4435 	bp->b_bcount = size;		/* requested buffer size. */
4436 	return (1);
4437 }
4438 
4439 extern int inflight_transient_maps;
4440 
4441 static struct bio_queue nondump_bios;
4442 
4443 void
4444 biodone(struct bio *bp)
4445 {
4446 	struct mtx *mtxp;
4447 	void (*done)(struct bio *);
4448 	vm_offset_t start, end;
4449 
4450 	biotrack(bp, __func__);
4451 
4452 	/*
4453 	 * Avoid completing I/O when dumping after a panic since that may
4454 	 * result in a deadlock in the filesystem or pager code.  Note that
4455 	 * this doesn't affect dumps that were started manually since we aim
4456 	 * to keep the system usable after it has been resumed.
4457 	 */
4458 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4459 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4460 		return;
4461 	}
4462 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4463 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4464 		bp->bio_flags |= BIO_UNMAPPED;
4465 		start = trunc_page((vm_offset_t)bp->bio_data);
4466 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4467 		bp->bio_data = unmapped_buf;
4468 		pmap_qremove(start, atop(end - start));
4469 		vmem_free(transient_arena, start, end - start);
4470 		atomic_add_int(&inflight_transient_maps, -1);
4471 	}
4472 	done = bp->bio_done;
4473 	/*
4474 	 * The check for done == biodone is to allow biodone to be
4475 	 * used as a bio_done routine.
4476 	 */
4477 	if (done == NULL || done == biodone) {
4478 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4479 		mtx_lock(mtxp);
4480 		bp->bio_flags |= BIO_DONE;
4481 		wakeup(bp);
4482 		mtx_unlock(mtxp);
4483 	} else
4484 		done(bp);
4485 }
4486 
4487 /*
4488  * Wait for a BIO to finish.
4489  */
4490 int
4491 biowait(struct bio *bp, const char *wmesg)
4492 {
4493 	struct mtx *mtxp;
4494 
4495 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4496 	mtx_lock(mtxp);
4497 	while ((bp->bio_flags & BIO_DONE) == 0)
4498 		msleep(bp, mtxp, PRIBIO, wmesg, 0);
4499 	mtx_unlock(mtxp);
4500 	if (bp->bio_error != 0)
4501 		return (bp->bio_error);
4502 	if (!(bp->bio_flags & BIO_ERROR))
4503 		return (0);
4504 	return (EIO);
4505 }
4506 
4507 void
4508 biofinish(struct bio *bp, struct devstat *stat, int error)
4509 {
4510 
4511 	if (error) {
4512 		bp->bio_error = error;
4513 		bp->bio_flags |= BIO_ERROR;
4514 	}
4515 	if (stat != NULL)
4516 		devstat_end_transaction_bio(stat, bp);
4517 	biodone(bp);
4518 }
4519 
4520 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4521 void
4522 biotrack_buf(struct bio *bp, const char *location)
4523 {
4524 
4525 	buf_track(bp->bio_track_bp, location);
4526 }
4527 #endif
4528 
4529 /*
4530  *	bufwait:
4531  *
4532  *	Wait for buffer I/O completion, returning error status.  The buffer
4533  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4534  *	error and cleared.
4535  */
4536 int
4537 bufwait(struct buf *bp)
4538 {
4539 	if (bp->b_iocmd == BIO_READ)
4540 		bwait(bp, PRIBIO, "biord");
4541 	else
4542 		bwait(bp, PRIBIO, "biowr");
4543 	if (bp->b_flags & B_EINTR) {
4544 		bp->b_flags &= ~B_EINTR;
4545 		return (EINTR);
4546 	}
4547 	if (bp->b_ioflags & BIO_ERROR) {
4548 		return (bp->b_error ? bp->b_error : EIO);
4549 	} else {
4550 		return (0);
4551 	}
4552 }
4553 
4554 /*
4555  *	bufdone:
4556  *
4557  *	Finish I/O on a buffer, optionally calling a completion function.
4558  *	This is usually called from an interrupt so process blocking is
4559  *	not allowed.
4560  *
4561  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4562  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4563  *	assuming B_INVAL is clear.
4564  *
4565  *	For the VMIO case, we set B_CACHE if the op was a read and no
4566  *	read error occurred, or if the op was a write.  B_CACHE is never
4567  *	set if the buffer is invalid or otherwise uncacheable.
4568  *
4569  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4570  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4571  *	in the biodone routine.
4572  */
4573 void
4574 bufdone(struct buf *bp)
4575 {
4576 	struct bufobj *dropobj;
4577 	void    (*biodone)(struct buf *);
4578 
4579 	buf_track(bp, __func__);
4580 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4581 	dropobj = NULL;
4582 
4583 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4584 
4585 	runningbufwakeup(bp);
4586 	if (bp->b_iocmd == BIO_WRITE)
4587 		dropobj = bp->b_bufobj;
4588 	/* call optional completion function if requested */
4589 	if (bp->b_iodone != NULL) {
4590 		biodone = bp->b_iodone;
4591 		bp->b_iodone = NULL;
4592 		(*biodone) (bp);
4593 		if (dropobj)
4594 			bufobj_wdrop(dropobj);
4595 		return;
4596 	}
4597 	if (bp->b_flags & B_VMIO) {
4598 		/*
4599 		 * Set B_CACHE if the op was a normal read and no error
4600 		 * occurred.  B_CACHE is set for writes in the b*write()
4601 		 * routines.
4602 		 */
4603 		if (bp->b_iocmd == BIO_READ &&
4604 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4605 		    !(bp->b_ioflags & BIO_ERROR))
4606 			bp->b_flags |= B_CACHE;
4607 		vfs_vmio_iodone(bp);
4608 	}
4609 	if (!LIST_EMPTY(&bp->b_dep))
4610 		buf_complete(bp);
4611 	if ((bp->b_flags & B_CKHASH) != 0) {
4612 		KASSERT(bp->b_iocmd == BIO_READ,
4613 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4614 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4615 		(*bp->b_ckhashcalc)(bp);
4616 	}
4617 	/*
4618 	 * For asynchronous completions, release the buffer now. The brelse
4619 	 * will do a wakeup there if necessary - so no need to do a wakeup
4620 	 * here in the async case. The sync case always needs to do a wakeup.
4621 	 */
4622 	if (bp->b_flags & B_ASYNC) {
4623 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4624 		    (bp->b_ioflags & BIO_ERROR))
4625 			brelse(bp);
4626 		else
4627 			bqrelse(bp);
4628 	} else
4629 		bdone(bp);
4630 	if (dropobj)
4631 		bufobj_wdrop(dropobj);
4632 }
4633 
4634 /*
4635  * This routine is called in lieu of iodone in the case of
4636  * incomplete I/O.  This keeps the busy status for pages
4637  * consistent.
4638  */
4639 void
4640 vfs_unbusy_pages(struct buf *bp)
4641 {
4642 	int i;
4643 	vm_object_t obj;
4644 	vm_page_t m;
4645 
4646 	runningbufwakeup(bp);
4647 	if (!(bp->b_flags & B_VMIO))
4648 		return;
4649 
4650 	obj = bp->b_bufobj->bo_object;
4651 	for (i = 0; i < bp->b_npages; i++) {
4652 		m = bp->b_pages[i];
4653 		if (m == bogus_page) {
4654 			m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4655 			if (!m)
4656 				panic("vfs_unbusy_pages: page missing\n");
4657 			bp->b_pages[i] = m;
4658 			if (buf_mapped(bp)) {
4659 				BUF_CHECK_MAPPED(bp);
4660 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4661 				    bp->b_pages, bp->b_npages);
4662 			} else
4663 				BUF_CHECK_UNMAPPED(bp);
4664 		}
4665 		vm_page_sunbusy(m);
4666 	}
4667 	vm_object_pip_wakeupn(obj, bp->b_npages);
4668 }
4669 
4670 /*
4671  * vfs_page_set_valid:
4672  *
4673  *	Set the valid bits in a page based on the supplied offset.   The
4674  *	range is restricted to the buffer's size.
4675  *
4676  *	This routine is typically called after a read completes.
4677  */
4678 static void
4679 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4680 {
4681 	vm_ooffset_t eoff;
4682 
4683 	/*
4684 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4685 	 * page boundary and eoff is not greater than the end of the buffer.
4686 	 * The end of the buffer, in this case, is our file EOF, not the
4687 	 * allocation size of the buffer.
4688 	 */
4689 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4690 	if (eoff > bp->b_offset + bp->b_bcount)
4691 		eoff = bp->b_offset + bp->b_bcount;
4692 
4693 	/*
4694 	 * Set valid range.  This is typically the entire buffer and thus the
4695 	 * entire page.
4696 	 */
4697 	if (eoff > off)
4698 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4699 }
4700 
4701 /*
4702  * vfs_page_set_validclean:
4703  *
4704  *	Set the valid bits and clear the dirty bits in a page based on the
4705  *	supplied offset.   The range is restricted to the buffer's size.
4706  */
4707 static void
4708 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4709 {
4710 	vm_ooffset_t soff, eoff;
4711 
4712 	/*
4713 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4714 	 * page boundary or cross the end of the buffer.  The end of the
4715 	 * buffer, in this case, is our file EOF, not the allocation size
4716 	 * of the buffer.
4717 	 */
4718 	soff = off;
4719 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4720 	if (eoff > bp->b_offset + bp->b_bcount)
4721 		eoff = bp->b_offset + bp->b_bcount;
4722 
4723 	/*
4724 	 * Set valid range.  This is typically the entire buffer and thus the
4725 	 * entire page.
4726 	 */
4727 	if (eoff > soff) {
4728 		vm_page_set_validclean(
4729 		    m,
4730 		   (vm_offset_t) (soff & PAGE_MASK),
4731 		   (vm_offset_t) (eoff - soff)
4732 		);
4733 	}
4734 }
4735 
4736 /*
4737  * Acquire a shared busy on all pages in the buf.
4738  */
4739 void
4740 vfs_busy_pages_acquire(struct buf *bp)
4741 {
4742 	int i;
4743 
4744 	for (i = 0; i < bp->b_npages; i++)
4745 		vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4746 }
4747 
4748 void
4749 vfs_busy_pages_release(struct buf *bp)
4750 {
4751 	int i;
4752 
4753 	for (i = 0; i < bp->b_npages; i++)
4754 		vm_page_sunbusy(bp->b_pages[i]);
4755 }
4756 
4757 /*
4758  * This routine is called before a device strategy routine.
4759  * It is used to tell the VM system that paging I/O is in
4760  * progress, and treat the pages associated with the buffer
4761  * almost as being exclusive busy.  Also the object paging_in_progress
4762  * flag is handled to make sure that the object doesn't become
4763  * inconsistent.
4764  *
4765  * Since I/O has not been initiated yet, certain buffer flags
4766  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4767  * and should be ignored.
4768  */
4769 void
4770 vfs_busy_pages(struct buf *bp, int clear_modify)
4771 {
4772 	vm_object_t obj;
4773 	vm_ooffset_t foff;
4774 	vm_page_t m;
4775 	int i;
4776 	bool bogus;
4777 
4778 	if (!(bp->b_flags & B_VMIO))
4779 		return;
4780 
4781 	obj = bp->b_bufobj->bo_object;
4782 	foff = bp->b_offset;
4783 	KASSERT(bp->b_offset != NOOFFSET,
4784 	    ("vfs_busy_pages: no buffer offset"));
4785 	if ((bp->b_flags & B_CLUSTER) == 0) {
4786 		vm_object_pip_add(obj, bp->b_npages);
4787 		vfs_busy_pages_acquire(bp);
4788 	}
4789 	if (bp->b_bufsize != 0)
4790 		vfs_setdirty_range(bp);
4791 	bogus = false;
4792 	for (i = 0; i < bp->b_npages; i++) {
4793 		m = bp->b_pages[i];
4794 		vm_page_assert_sbusied(m);
4795 
4796 		/*
4797 		 * When readying a buffer for a read ( i.e
4798 		 * clear_modify == 0 ), it is important to do
4799 		 * bogus_page replacement for valid pages in
4800 		 * partially instantiated buffers.  Partially
4801 		 * instantiated buffers can, in turn, occur when
4802 		 * reconstituting a buffer from its VM backing store
4803 		 * base.  We only have to do this if B_CACHE is
4804 		 * clear ( which causes the I/O to occur in the
4805 		 * first place ).  The replacement prevents the read
4806 		 * I/O from overwriting potentially dirty VM-backed
4807 		 * pages.  XXX bogus page replacement is, uh, bogus.
4808 		 * It may not work properly with small-block devices.
4809 		 * We need to find a better way.
4810 		 */
4811 		if (clear_modify) {
4812 			pmap_remove_write(m);
4813 			vfs_page_set_validclean(bp, foff, m);
4814 		} else if (vm_page_all_valid(m) &&
4815 		    (bp->b_flags & B_CACHE) == 0) {
4816 			bp->b_pages[i] = bogus_page;
4817 			bogus = true;
4818 		}
4819 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4820 	}
4821 	if (bogus && buf_mapped(bp)) {
4822 		BUF_CHECK_MAPPED(bp);
4823 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4824 		    bp->b_pages, bp->b_npages);
4825 	}
4826 }
4827 
4828 /*
4829  *	vfs_bio_set_valid:
4830  *
4831  *	Set the range within the buffer to valid.  The range is
4832  *	relative to the beginning of the buffer, b_offset.  Note that
4833  *	b_offset itself may be offset from the beginning of the first
4834  *	page.
4835  */
4836 void
4837 vfs_bio_set_valid(struct buf *bp, int base, int size)
4838 {
4839 	int i, n;
4840 	vm_page_t m;
4841 
4842 	if (!(bp->b_flags & B_VMIO))
4843 		return;
4844 
4845 	/*
4846 	 * Fixup base to be relative to beginning of first page.
4847 	 * Set initial n to be the maximum number of bytes in the
4848 	 * first page that can be validated.
4849 	 */
4850 	base += (bp->b_offset & PAGE_MASK);
4851 	n = PAGE_SIZE - (base & PAGE_MASK);
4852 
4853 	/*
4854 	 * Busy may not be strictly necessary here because the pages are
4855 	 * unlikely to be fully valid and the vnode lock will synchronize
4856 	 * their access via getpages.  It is grabbed for consistency with
4857 	 * other page validation.
4858 	 */
4859 	vfs_busy_pages_acquire(bp);
4860 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4861 		m = bp->b_pages[i];
4862 		if (n > size)
4863 			n = size;
4864 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4865 		base += n;
4866 		size -= n;
4867 		n = PAGE_SIZE;
4868 	}
4869 	vfs_busy_pages_release(bp);
4870 }
4871 
4872 /*
4873  *	vfs_bio_clrbuf:
4874  *
4875  *	If the specified buffer is a non-VMIO buffer, clear the entire
4876  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4877  *	validate only the previously invalid portions of the buffer.
4878  *	This routine essentially fakes an I/O, so we need to clear
4879  *	BIO_ERROR and B_INVAL.
4880  *
4881  *	Note that while we only theoretically need to clear through b_bcount,
4882  *	we go ahead and clear through b_bufsize.
4883  */
4884 void
4885 vfs_bio_clrbuf(struct buf *bp)
4886 {
4887 	int i, j, sa, ea, slide, zbits;
4888 	vm_page_bits_t mask;
4889 
4890 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4891 		clrbuf(bp);
4892 		return;
4893 	}
4894 	bp->b_flags &= ~B_INVAL;
4895 	bp->b_ioflags &= ~BIO_ERROR;
4896 	vfs_busy_pages_acquire(bp);
4897 	sa = bp->b_offset & PAGE_MASK;
4898 	slide = 0;
4899 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4900 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4901 		ea = slide & PAGE_MASK;
4902 		if (ea == 0)
4903 			ea = PAGE_SIZE;
4904 		if (bp->b_pages[i] == bogus_page)
4905 			continue;
4906 		j = sa / DEV_BSIZE;
4907 		zbits = (sizeof(vm_page_bits_t) * NBBY) -
4908 		    (ea - sa) / DEV_BSIZE;
4909 		mask = (VM_PAGE_BITS_ALL >> zbits) << j;
4910 		if ((bp->b_pages[i]->valid & mask) == mask)
4911 			continue;
4912 		if ((bp->b_pages[i]->valid & mask) == 0)
4913 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4914 		else {
4915 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4916 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4917 					pmap_zero_page_area(bp->b_pages[i],
4918 					    sa, DEV_BSIZE);
4919 				}
4920 			}
4921 		}
4922 		vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4923 		    roundup2(ea - sa, DEV_BSIZE));
4924 	}
4925 	vfs_busy_pages_release(bp);
4926 	bp->b_resid = 0;
4927 }
4928 
4929 void
4930 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4931 {
4932 	vm_page_t m;
4933 	int i, n;
4934 
4935 	if (buf_mapped(bp)) {
4936 		BUF_CHECK_MAPPED(bp);
4937 		bzero(bp->b_data + base, size);
4938 	} else {
4939 		BUF_CHECK_UNMAPPED(bp);
4940 		n = PAGE_SIZE - (base & PAGE_MASK);
4941 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4942 			m = bp->b_pages[i];
4943 			if (n > size)
4944 				n = size;
4945 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4946 			base += n;
4947 			size -= n;
4948 			n = PAGE_SIZE;
4949 		}
4950 	}
4951 }
4952 
4953 /*
4954  * Update buffer flags based on I/O request parameters, optionally releasing the
4955  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4956  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4957  * I/O).  Otherwise the buffer is released to the cache.
4958  */
4959 static void
4960 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4961 {
4962 
4963 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4964 	    ("buf %p non-VMIO noreuse", bp));
4965 
4966 	if ((ioflag & IO_DIRECT) != 0)
4967 		bp->b_flags |= B_DIRECT;
4968 	if ((ioflag & IO_EXT) != 0)
4969 		bp->b_xflags |= BX_ALTDATA;
4970 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4971 		bp->b_flags |= B_RELBUF;
4972 		if ((ioflag & IO_NOREUSE) != 0)
4973 			bp->b_flags |= B_NOREUSE;
4974 		if (release)
4975 			brelse(bp);
4976 	} else if (release)
4977 		bqrelse(bp);
4978 }
4979 
4980 void
4981 vfs_bio_brelse(struct buf *bp, int ioflag)
4982 {
4983 
4984 	b_io_dismiss(bp, ioflag, true);
4985 }
4986 
4987 void
4988 vfs_bio_set_flags(struct buf *bp, int ioflag)
4989 {
4990 
4991 	b_io_dismiss(bp, ioflag, false);
4992 }
4993 
4994 /*
4995  * vm_hold_load_pages and vm_hold_free_pages get pages into
4996  * a buffers address space.  The pages are anonymous and are
4997  * not associated with a file object.
4998  */
4999 static void
5000 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
5001 {
5002 	vm_offset_t pg;
5003 	vm_page_t p;
5004 	int index;
5005 
5006 	BUF_CHECK_MAPPED(bp);
5007 
5008 	to = round_page(to);
5009 	from = round_page(from);
5010 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
5011 	MPASS((bp->b_flags & B_MAXPHYS) == 0);
5012 	KASSERT(to - from <= maxbcachebuf,
5013 	    ("vm_hold_load_pages too large %p %#jx %#jx %u",
5014 	    bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
5015 
5016 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
5017 		/*
5018 		 * note: must allocate system pages since blocking here
5019 		 * could interfere with paging I/O, no matter which
5020 		 * process we are.
5021 		 */
5022 		p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
5023 		    VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
5024 		pmap_qenter(pg, &p, 1);
5025 		bp->b_pages[index] = p;
5026 	}
5027 	bp->b_npages = index;
5028 }
5029 
5030 /* Return pages associated with this buf to the vm system */
5031 static void
5032 vm_hold_free_pages(struct buf *bp, int newbsize)
5033 {
5034 	vm_offset_t from;
5035 	vm_page_t p;
5036 	int index, newnpages;
5037 
5038 	BUF_CHECK_MAPPED(bp);
5039 
5040 	from = round_page((vm_offset_t)bp->b_data + newbsize);
5041 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
5042 	if (bp->b_npages > newnpages)
5043 		pmap_qremove(from, bp->b_npages - newnpages);
5044 	for (index = newnpages; index < bp->b_npages; index++) {
5045 		p = bp->b_pages[index];
5046 		bp->b_pages[index] = NULL;
5047 		vm_page_unwire_noq(p);
5048 		vm_page_free(p);
5049 	}
5050 	bp->b_npages = newnpages;
5051 }
5052 
5053 /*
5054  * Map an IO request into kernel virtual address space.
5055  *
5056  * All requests are (re)mapped into kernel VA space.
5057  * Notice that we use b_bufsize for the size of the buffer
5058  * to be mapped.  b_bcount might be modified by the driver.
5059  *
5060  * Note that even if the caller determines that the address space should
5061  * be valid, a race or a smaller-file mapped into a larger space may
5062  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
5063  * check the return value.
5064  *
5065  * This function only works with pager buffers.
5066  */
5067 int
5068 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
5069 {
5070 	vm_prot_t prot;
5071 	int pidx;
5072 
5073 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
5074 	prot = VM_PROT_READ;
5075 	if (bp->b_iocmd == BIO_READ)
5076 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
5077 	pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5078 	    (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5079 	if (pidx < 0)
5080 		return (-1);
5081 	bp->b_bufsize = len;
5082 	bp->b_npages = pidx;
5083 	bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5084 	if (mapbuf || !unmapped_buf_allowed) {
5085 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5086 		bp->b_data = bp->b_kvabase + bp->b_offset;
5087 	} else
5088 		bp->b_data = unmapped_buf;
5089 	return (0);
5090 }
5091 
5092 /*
5093  * Free the io map PTEs associated with this IO operation.
5094  * We also invalidate the TLB entries and restore the original b_addr.
5095  *
5096  * This function only works with pager buffers.
5097  */
5098 void
5099 vunmapbuf(struct buf *bp)
5100 {
5101 	int npages;
5102 
5103 	npages = bp->b_npages;
5104 	if (buf_mapped(bp))
5105 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5106 	vm_page_unhold_pages(bp->b_pages, npages);
5107 
5108 	bp->b_data = unmapped_buf;
5109 }
5110 
5111 void
5112 bdone(struct buf *bp)
5113 {
5114 	struct mtx *mtxp;
5115 
5116 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5117 	mtx_lock(mtxp);
5118 	bp->b_flags |= B_DONE;
5119 	wakeup(bp);
5120 	mtx_unlock(mtxp);
5121 }
5122 
5123 void
5124 bwait(struct buf *bp, u_char pri, const char *wchan)
5125 {
5126 	struct mtx *mtxp;
5127 
5128 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
5129 	mtx_lock(mtxp);
5130 	while ((bp->b_flags & B_DONE) == 0)
5131 		msleep(bp, mtxp, pri, wchan, 0);
5132 	mtx_unlock(mtxp);
5133 }
5134 
5135 int
5136 bufsync(struct bufobj *bo, int waitfor)
5137 {
5138 
5139 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5140 }
5141 
5142 void
5143 bufstrategy(struct bufobj *bo, struct buf *bp)
5144 {
5145 	int i __unused;
5146 	struct vnode *vp;
5147 
5148 	vp = bp->b_vp;
5149 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5150 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5151 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5152 	i = VOP_STRATEGY(vp, bp);
5153 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5154 }
5155 
5156 /*
5157  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
5158  */
5159 void
5160 bufobj_init(struct bufobj *bo, void *private)
5161 {
5162 	static volatile int bufobj_cleanq;
5163 
5164         bo->bo_domain =
5165             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5166         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5167         bo->bo_private = private;
5168         TAILQ_INIT(&bo->bo_clean.bv_hd);
5169 	pctrie_init(&bo->bo_clean.bv_root);
5170         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5171 	pctrie_init(&bo->bo_dirty.bv_root);
5172 }
5173 
5174 void
5175 bufobj_wrefl(struct bufobj *bo)
5176 {
5177 
5178 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5179 	ASSERT_BO_WLOCKED(bo);
5180 	bo->bo_numoutput++;
5181 }
5182 
5183 void
5184 bufobj_wref(struct bufobj *bo)
5185 {
5186 
5187 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5188 	BO_LOCK(bo);
5189 	bo->bo_numoutput++;
5190 	BO_UNLOCK(bo);
5191 }
5192 
5193 void
5194 bufobj_wdrop(struct bufobj *bo)
5195 {
5196 
5197 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5198 	BO_LOCK(bo);
5199 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5200 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5201 		bo->bo_flag &= ~BO_WWAIT;
5202 		wakeup(&bo->bo_numoutput);
5203 	}
5204 	BO_UNLOCK(bo);
5205 }
5206 
5207 int
5208 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5209 {
5210 	int error;
5211 
5212 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5213 	ASSERT_BO_WLOCKED(bo);
5214 	error = 0;
5215 	while (bo->bo_numoutput) {
5216 		bo->bo_flag |= BO_WWAIT;
5217 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5218 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5219 		if (error)
5220 			break;
5221 	}
5222 	return (error);
5223 }
5224 
5225 /*
5226  * Set bio_data or bio_ma for struct bio from the struct buf.
5227  */
5228 void
5229 bdata2bio(struct buf *bp, struct bio *bip)
5230 {
5231 
5232 	if (!buf_mapped(bp)) {
5233 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5234 		bip->bio_ma = bp->b_pages;
5235 		bip->bio_ma_n = bp->b_npages;
5236 		bip->bio_data = unmapped_buf;
5237 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5238 		bip->bio_flags |= BIO_UNMAPPED;
5239 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5240 		    PAGE_SIZE == bp->b_npages,
5241 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5242 		    (long long)bip->bio_length, bip->bio_ma_n));
5243 	} else {
5244 		bip->bio_data = bp->b_data;
5245 		bip->bio_ma = NULL;
5246 	}
5247 }
5248 
5249 struct memdesc
5250 memdesc_bio(struct bio *bio)
5251 {
5252 	if ((bio->bio_flags & BIO_VLIST) != 0)
5253 		return (memdesc_vlist((struct bus_dma_segment *)bio->bio_data,
5254 		    bio->bio_ma_n));
5255 
5256 	if ((bio->bio_flags & BIO_UNMAPPED) != 0)
5257 		return (memdesc_vmpages(bio->bio_ma, bio->bio_bcount,
5258 		    bio->bio_ma_offset));
5259 
5260 	return (memdesc_vaddr(bio->bio_data, bio->bio_bcount));
5261 }
5262 
5263 static int buf_pager_relbuf;
5264 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5265     &buf_pager_relbuf, 0,
5266     "Make buffer pager release buffers after reading");
5267 
5268 /*
5269  * The buffer pager.  It uses buffer reads to validate pages.
5270  *
5271  * In contrast to the generic local pager from vm/vnode_pager.c, this
5272  * pager correctly and easily handles volumes where the underlying
5273  * device block size is greater than the machine page size.  The
5274  * buffer cache transparently extends the requested page run to be
5275  * aligned at the block boundary, and does the necessary bogus page
5276  * replacements in the addends to avoid obliterating already valid
5277  * pages.
5278  *
5279  * The only non-trivial issue is that the exclusive busy state for
5280  * pages, which is assumed by the vm_pager_getpages() interface, is
5281  * incompatible with the VMIO buffer cache's desire to share-busy the
5282  * pages.  This function performs a trivial downgrade of the pages'
5283  * state before reading buffers, and a less trivial upgrade from the
5284  * shared-busy to excl-busy state after the read.
5285  */
5286 int
5287 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5288     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5289     vbg_get_blksize_t get_blksize)
5290 {
5291 	vm_page_t m;
5292 	vm_object_t object;
5293 	struct buf *bp;
5294 	struct mount *mp;
5295 	daddr_t lbn, lbnp;
5296 	vm_ooffset_t la, lb, poff, poffe;
5297 	long bo_bs, bsize;
5298 	int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5299 	bool redo, lpart;
5300 
5301 	object = vp->v_object;
5302 	mp = vp->v_mount;
5303 	error = 0;
5304 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5305 	if (la >= object->un_pager.vnp.vnp_size)
5306 		return (VM_PAGER_BAD);
5307 
5308 	/*
5309 	 * Change the meaning of la from where the last requested page starts
5310 	 * to where it ends, because that's the end of the requested region
5311 	 * and the start of the potential read-ahead region.
5312 	 */
5313 	la += PAGE_SIZE;
5314 	lpart = la > object->un_pager.vnp.vnp_size;
5315 	error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5316 	    &bo_bs);
5317 	if (error != 0)
5318 		return (VM_PAGER_ERROR);
5319 
5320 	/*
5321 	 * Calculate read-ahead, behind and total pages.
5322 	 */
5323 	pgsin = count;
5324 	lb = IDX_TO_OFF(ma[0]->pindex);
5325 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5326 	pgsin += pgsin_b;
5327 	if (rbehind != NULL)
5328 		*rbehind = pgsin_b;
5329 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5330 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5331 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5332 		    PAGE_SIZE) - la);
5333 	pgsin += pgsin_a;
5334 	if (rahead != NULL)
5335 		*rahead = pgsin_a;
5336 	VM_CNT_INC(v_vnodein);
5337 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5338 
5339 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5340 	    != 0) ? GB_UNMAPPED : 0;
5341 again:
5342 	for (i = 0; i < count; i++) {
5343 		if (ma[i] != bogus_page)
5344 			vm_page_busy_downgrade(ma[i]);
5345 	}
5346 
5347 	lbnp = -1;
5348 	for (i = 0; i < count; i++) {
5349 		m = ma[i];
5350 		if (m == bogus_page)
5351 			continue;
5352 
5353 		/*
5354 		 * Pages are shared busy and the object lock is not
5355 		 * owned, which together allow for the pages'
5356 		 * invalidation.  The racy test for validity avoids
5357 		 * useless creation of the buffer for the most typical
5358 		 * case when invalidation is not used in redo or for
5359 		 * parallel read.  The shared->excl upgrade loop at
5360 		 * the end of the function catches the race in a
5361 		 * reliable way (protected by the object lock).
5362 		 */
5363 		if (vm_page_all_valid(m))
5364 			continue;
5365 
5366 		poff = IDX_TO_OFF(m->pindex);
5367 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5368 		for (; poff < poffe; poff += bsize) {
5369 			lbn = get_lblkno(vp, poff);
5370 			if (lbn == lbnp)
5371 				goto next_page;
5372 			lbnp = lbn;
5373 
5374 			error = get_blksize(vp, lbn, &bsize);
5375 			if (error == 0)
5376 				error = bread_gb(vp, lbn, bsize,
5377 				    curthread->td_ucred, br_flags, &bp);
5378 			if (error != 0)
5379 				goto end_pages;
5380 			if (bp->b_rcred == curthread->td_ucred) {
5381 				crfree(bp->b_rcred);
5382 				bp->b_rcred = NOCRED;
5383 			}
5384 			if (LIST_EMPTY(&bp->b_dep)) {
5385 				/*
5386 				 * Invalidation clears m->valid, but
5387 				 * may leave B_CACHE flag if the
5388 				 * buffer existed at the invalidation
5389 				 * time.  In this case, recycle the
5390 				 * buffer to do real read on next
5391 				 * bread() after redo.
5392 				 *
5393 				 * Otherwise B_RELBUF is not strictly
5394 				 * necessary, enable to reduce buf
5395 				 * cache pressure.
5396 				 */
5397 				if (buf_pager_relbuf ||
5398 				    !vm_page_all_valid(m))
5399 					bp->b_flags |= B_RELBUF;
5400 
5401 				bp->b_flags &= ~B_NOCACHE;
5402 				brelse(bp);
5403 			} else {
5404 				bqrelse(bp);
5405 			}
5406 		}
5407 		KASSERT(1 /* racy, enable for debugging */ ||
5408 		    vm_page_all_valid(m) || i == count - 1,
5409 		    ("buf %d %p invalid", i, m));
5410 		if (i == count - 1 && lpart) {
5411 			if (!vm_page_none_valid(m) &&
5412 			    !vm_page_all_valid(m))
5413 				vm_page_zero_invalid(m, TRUE);
5414 		}
5415 next_page:;
5416 	}
5417 end_pages:
5418 
5419 	redo = false;
5420 	for (i = 0; i < count; i++) {
5421 		if (ma[i] == bogus_page)
5422 			continue;
5423 		if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5424 			vm_page_sunbusy(ma[i]);
5425 			ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5426 			    VM_ALLOC_NORMAL);
5427 		}
5428 
5429 		/*
5430 		 * Since the pages were only sbusy while neither the
5431 		 * buffer nor the object lock was held by us, or
5432 		 * reallocated while vm_page_grab() slept for busy
5433 		 * relinguish, they could have been invalidated.
5434 		 * Recheck the valid bits and re-read as needed.
5435 		 *
5436 		 * Note that the last page is made fully valid in the
5437 		 * read loop, and partial validity for the page at
5438 		 * index count - 1 could mean that the page was
5439 		 * invalidated or removed, so we must restart for
5440 		 * safety as well.
5441 		 */
5442 		if (!vm_page_all_valid(ma[i]))
5443 			redo = true;
5444 	}
5445 	if (redo && error == 0)
5446 		goto again;
5447 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5448 }
5449 
5450 #include "opt_ddb.h"
5451 #ifdef DDB
5452 #include <ddb/ddb.h>
5453 
5454 /* DDB command to show buffer data */
5455 DB_SHOW_COMMAND(buffer, db_show_buffer)
5456 {
5457 	/* get args */
5458 	struct buf *bp = (struct buf *)addr;
5459 #ifdef FULL_BUF_TRACKING
5460 	uint32_t i, j;
5461 #endif
5462 
5463 	if (!have_addr) {
5464 		db_printf("usage: show buffer <addr>\n");
5465 		return;
5466 	}
5467 
5468 	db_printf("buf at %p\n", bp);
5469 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5470 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5471 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5472 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5473 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5474 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5475 	db_printf(
5476 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5477 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5478 	    "b_vp = %p, b_dep = %p\n",
5479 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5480 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5481 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5482 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5483 	    bp->b_kvabase, bp->b_kvasize);
5484 	if (bp->b_npages) {
5485 		int i;
5486 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5487 		for (i = 0; i < bp->b_npages; i++) {
5488 			vm_page_t m;
5489 			m = bp->b_pages[i];
5490 			if (m != NULL)
5491 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5492 				    (u_long)m->pindex,
5493 				    (u_long)VM_PAGE_TO_PHYS(m));
5494 			else
5495 				db_printf("( ??? )");
5496 			if ((i + 1) < bp->b_npages)
5497 				db_printf(",");
5498 		}
5499 		db_printf("\n");
5500 	}
5501 	BUF_LOCKPRINTINFO(bp);
5502 #if defined(FULL_BUF_TRACKING)
5503 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5504 
5505 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5506 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5507 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5508 			continue;
5509 		db_printf(" %2u: %s\n", j,
5510 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5511 	}
5512 #elif defined(BUF_TRACKING)
5513 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5514 #endif
5515 	db_printf(" ");
5516 }
5517 
5518 DB_SHOW_COMMAND_FLAGS(bufqueues, bufqueues, DB_CMD_MEMSAFE)
5519 {
5520 	struct bufdomain *bd;
5521 	struct buf *bp;
5522 	long total;
5523 	int i, j, cnt;
5524 
5525 	db_printf("bqempty: %d\n", bqempty.bq_len);
5526 
5527 	for (i = 0; i < buf_domains; i++) {
5528 		bd = &bdomain[i];
5529 		db_printf("Buf domain %d\n", i);
5530 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5531 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5532 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5533 		db_printf("\n");
5534 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5535 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5536 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5537 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5538 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5539 		db_printf("\n");
5540 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5541 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5542 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5543 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5544 		db_printf("\n");
5545 		total = 0;
5546 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5547 			total += bp->b_bufsize;
5548 		db_printf("\tcleanq count\t%d (%ld)\n",
5549 		    bd->bd_cleanq->bq_len, total);
5550 		total = 0;
5551 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5552 			total += bp->b_bufsize;
5553 		db_printf("\tdirtyq count\t%d (%ld)\n",
5554 		    bd->bd_dirtyq.bq_len, total);
5555 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5556 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5557 		db_printf("\tCPU ");
5558 		for (j = 0; j <= mp_maxid; j++)
5559 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5560 		db_printf("\n");
5561 		cnt = 0;
5562 		total = 0;
5563 		for (j = 0; j < nbuf; j++) {
5564 			bp = nbufp(j);
5565 			if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5566 				cnt++;
5567 				total += bp->b_bufsize;
5568 			}
5569 		}
5570 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5571 		cnt = 0;
5572 		total = 0;
5573 		for (j = 0; j < nbuf; j++) {
5574 			bp = nbufp(j);
5575 			if (bp->b_domain == i) {
5576 				cnt++;
5577 				total += bp->b_bufsize;
5578 			}
5579 		}
5580 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5581 	}
5582 }
5583 
5584 DB_SHOW_COMMAND_FLAGS(lockedbufs, lockedbufs, DB_CMD_MEMSAFE)
5585 {
5586 	struct buf *bp;
5587 	int i;
5588 
5589 	for (i = 0; i < nbuf; i++) {
5590 		bp = nbufp(i);
5591 		if (BUF_ISLOCKED(bp)) {
5592 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5593 			db_printf("\n");
5594 			if (db_pager_quit)
5595 				break;
5596 		}
5597 	}
5598 }
5599 
5600 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5601 {
5602 	struct vnode *vp;
5603 	struct buf *bp;
5604 
5605 	if (!have_addr) {
5606 		db_printf("usage: show vnodebufs <addr>\n");
5607 		return;
5608 	}
5609 	vp = (struct vnode *)addr;
5610 	db_printf("Clean buffers:\n");
5611 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5612 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5613 		db_printf("\n");
5614 	}
5615 	db_printf("Dirty buffers:\n");
5616 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5617 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5618 		db_printf("\n");
5619 	}
5620 }
5621 
5622 DB_COMMAND_FLAGS(countfreebufs, db_coundfreebufs, DB_CMD_MEMSAFE)
5623 {
5624 	struct buf *bp;
5625 	int i, used = 0, nfree = 0;
5626 
5627 	if (have_addr) {
5628 		db_printf("usage: countfreebufs\n");
5629 		return;
5630 	}
5631 
5632 	for (i = 0; i < nbuf; i++) {
5633 		bp = nbufp(i);
5634 		if (bp->b_qindex == QUEUE_EMPTY)
5635 			nfree++;
5636 		else
5637 			used++;
5638 	}
5639 
5640 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5641 	    nfree + used);
5642 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5643 }
5644 #endif /* DDB */
5645