xref: /freebsd/sys/kern/vfs_bio.c (revision daceb336172a6b0572de864b97e70b28451ca636)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2004 Poul-Henning Kamp
5  * Copyright (c) 1994,1997 John S. Dyson
6  * Copyright (c) 2013 The FreeBSD Foundation
7  * All rights reserved.
8  *
9  * Portions of this software were developed by Konstantin Belousov
10  * under sponsorship from the FreeBSD Foundation.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * this file contains a new buffer I/O scheme implementing a coherent
36  * VM object and buffer cache scheme.  Pains have been taken to make
37  * sure that the performance degradation associated with schemes such
38  * as this is not realized.
39  *
40  * Author:  John S. Dyson
41  * Significant help during the development and debugging phases
42  * had been provided by David Greenman, also of the FreeBSD core team.
43  *
44  * see man buf(9) for more info.
45  */
46 
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/bio.h>
53 #include <sys/bitset.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/buf.h>
57 #include <sys/devicestat.h>
58 #include <sys/eventhandler.h>
59 #include <sys/fail.h>
60 #include <sys/limits.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/mount.h>
64 #include <sys/mutex.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/proc.h>
68 #include <sys/racct.h>
69 #include <sys/resourcevar.h>
70 #include <sys/rwlock.h>
71 #include <sys/smp.h>
72 #include <sys/sysctl.h>
73 #include <sys/sysproto.h>
74 #include <sys/vmem.h>
75 #include <sys/vmmeter.h>
76 #include <sys/vnode.h>
77 #include <sys/watchdog.h>
78 #include <geom/geom.h>
79 #include <vm/vm.h>
80 #include <vm/vm_param.h>
81 #include <vm/vm_kern.h>
82 #include <vm/vm_object.h>
83 #include <vm/vm_page.h>
84 #include <vm/vm_pageout.h>
85 #include <vm/vm_pager.h>
86 #include <vm/vm_extern.h>
87 #include <vm/vm_map.h>
88 #include <vm/swap_pager.h>
89 
90 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
91 
92 struct	bio_ops bioops;		/* I/O operation notification */
93 
94 struct	buf_ops buf_ops_bio = {
95 	.bop_name	=	"buf_ops_bio",
96 	.bop_write	=	bufwrite,
97 	.bop_strategy	=	bufstrategy,
98 	.bop_sync	=	bufsync,
99 	.bop_bdflush	=	bufbdflush,
100 };
101 
102 struct bufqueue {
103 	struct mtx_padalign	bq_lock;
104 	TAILQ_HEAD(, buf)	bq_queue;
105 	uint8_t			bq_index;
106 	uint16_t		bq_subqueue;
107 	int			bq_len;
108 } __aligned(CACHE_LINE_SIZE);
109 
110 #define	BQ_LOCKPTR(bq)		(&(bq)->bq_lock)
111 #define	BQ_LOCK(bq)		mtx_lock(BQ_LOCKPTR((bq)))
112 #define	BQ_UNLOCK(bq)		mtx_unlock(BQ_LOCKPTR((bq)))
113 #define	BQ_ASSERT_LOCKED(bq)	mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
114 
115 struct bufdomain {
116 	struct bufqueue	bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
117 	struct bufqueue bd_dirtyq;
118 	struct bufqueue	*bd_cleanq;
119 	struct mtx_padalign bd_run_lock;
120 	/* Constants */
121 	long		bd_maxbufspace;
122 	long		bd_hibufspace;
123 	long 		bd_lobufspace;
124 	long 		bd_bufspacethresh;
125 	int		bd_hifreebuffers;
126 	int		bd_lofreebuffers;
127 	int		bd_hidirtybuffers;
128 	int		bd_lodirtybuffers;
129 	int		bd_dirtybufthresh;
130 	int		bd_lim;
131 	/* atomics */
132 	int		bd_wanted;
133 	int __aligned(CACHE_LINE_SIZE)	bd_numdirtybuffers;
134 	int __aligned(CACHE_LINE_SIZE)	bd_running;
135 	long __aligned(CACHE_LINE_SIZE) bd_bufspace;
136 	int __aligned(CACHE_LINE_SIZE)	bd_freebuffers;
137 } __aligned(CACHE_LINE_SIZE);
138 
139 #define	BD_LOCKPTR(bd)		(&(bd)->bd_cleanq->bq_lock)
140 #define	BD_LOCK(bd)		mtx_lock(BD_LOCKPTR((bd)))
141 #define	BD_UNLOCK(bd)		mtx_unlock(BD_LOCKPTR((bd)))
142 #define	BD_ASSERT_LOCKED(bd)	mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
143 #define	BD_RUN_LOCKPTR(bd)	(&(bd)->bd_run_lock)
144 #define	BD_RUN_LOCK(bd)		mtx_lock(BD_RUN_LOCKPTR((bd)))
145 #define	BD_RUN_UNLOCK(bd)	mtx_unlock(BD_RUN_LOCKPTR((bd)))
146 #define	BD_DOMAIN(bd)		(bd - bdomain)
147 
148 static struct buf *buf;		/* buffer header pool */
149 extern struct buf *swbuf;	/* Swap buffer header pool. */
150 caddr_t unmapped_buf;
151 
152 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
153 struct proc *bufdaemonproc;
154 
155 static int inmem(struct vnode *vp, daddr_t blkno);
156 static void vm_hold_free_pages(struct buf *bp, int newbsize);
157 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
158 		vm_offset_t to);
159 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
160 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
161 		vm_page_t m);
162 static void vfs_clean_pages_dirty_buf(struct buf *bp);
163 static void vfs_setdirty_locked_object(struct buf *bp);
164 static void vfs_vmio_invalidate(struct buf *bp);
165 static void vfs_vmio_truncate(struct buf *bp, int npages);
166 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
167 static int vfs_bio_clcheck(struct vnode *vp, int size,
168 		daddr_t lblkno, daddr_t blkno);
169 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
170 		void (*)(struct buf *));
171 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
172 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
173 static void buf_daemon(void);
174 static __inline void bd_wakeup(void);
175 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
176 static void bufkva_reclaim(vmem_t *, int);
177 static void bufkva_free(struct buf *);
178 static int buf_import(void *, void **, int, int, int);
179 static void buf_release(void *, void **, int);
180 static void maxbcachebuf_adjust(void);
181 static inline struct bufdomain *bufdomain(struct buf *);
182 static void bq_remove(struct bufqueue *bq, struct buf *bp);
183 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
184 static int buf_recycle(struct bufdomain *, bool kva);
185 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
186 	    const char *lockname);
187 static void bd_init(struct bufdomain *bd);
188 static int bd_flushall(struct bufdomain *bd);
189 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
190 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
191 
192 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
193 int vmiodirenable = TRUE;
194 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
195     "Use the VM system for directory writes");
196 long runningbufspace;
197 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
198     "Amount of presently outstanding async buffer io");
199 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
200     NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
201 static counter_u64_t bufkvaspace;
202 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
203     "Kernel virtual memory used for buffers");
204 static long maxbufspace;
205 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
206     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
207     __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
208     "Maximum allowed value of bufspace (including metadata)");
209 static long bufmallocspace;
210 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
211     "Amount of malloced memory for buffers");
212 static long maxbufmallocspace;
213 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
214     0, "Maximum amount of malloced memory for buffers");
215 static long lobufspace;
216 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
217     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
218     __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
219     "Minimum amount of buffers we want to have");
220 long hibufspace;
221 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
222     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
223     __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
224     "Maximum allowed value of bufspace (excluding metadata)");
225 long bufspacethresh;
226 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
227     CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
228     __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
229     "Bufspace consumed before waking the daemon to free some");
230 static counter_u64_t buffreekvacnt;
231 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
232     "Number of times we have freed the KVA space from some buffer");
233 static counter_u64_t bufdefragcnt;
234 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
235     "Number of times we have had to repeat buffer allocation to defragment");
236 static long lorunningspace;
237 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
238     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
239     "Minimum preferred space used for in-progress I/O");
240 static long hirunningspace;
241 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
242     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
243     "Maximum amount of space to use for in-progress I/O");
244 int dirtybufferflushes;
245 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
246     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
247 int bdwriteskip;
248 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
249     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
250 int altbufferflushes;
251 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
252     0, "Number of fsync flushes to limit dirty buffers");
253 static int recursiveflushes;
254 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
255     0, "Number of flushes skipped due to being recursive");
256 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
257 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
258     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
259     "Number of buffers that are dirty (has unwritten changes) at the moment");
260 static int lodirtybuffers;
261 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
262     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
263     __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
264     "How many buffers we want to have free before bufdaemon can sleep");
265 static int hidirtybuffers;
266 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
267     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
268     __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
269     "When the number of dirty buffers is considered severe");
270 int dirtybufthresh;
271 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
272     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
273     __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
274     "Number of bdwrite to bawrite conversions to clear dirty buffers");
275 static int numfreebuffers;
276 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
277     "Number of free buffers");
278 static int lofreebuffers;
279 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
280     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
281     __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
282    "Target number of free buffers");
283 static int hifreebuffers;
284 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
285     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
286     __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
287    "Threshold for clean buffer recycling");
288 static counter_u64_t getnewbufcalls;
289 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
290    &getnewbufcalls, "Number of calls to getnewbuf");
291 static counter_u64_t getnewbufrestarts;
292 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
293     &getnewbufrestarts,
294     "Number of times getnewbuf has had to restart a buffer acquisition");
295 static counter_u64_t mappingrestarts;
296 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
297     &mappingrestarts,
298     "Number of times getblk has had to restart a buffer mapping for "
299     "unmapped buffer");
300 static counter_u64_t numbufallocfails;
301 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
302     &numbufallocfails, "Number of times buffer allocations failed");
303 static int flushbufqtarget = 100;
304 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
305     "Amount of work to do in flushbufqueues when helping bufdaemon");
306 static counter_u64_t notbufdflushes;
307 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes,
308     "Number of dirty buffer flushes done by the bufdaemon helpers");
309 static long barrierwrites;
310 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
311     "Number of barrier writes");
312 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
313     &unmapped_buf_allowed, 0,
314     "Permit the use of the unmapped i/o");
315 int maxbcachebuf = MAXBCACHEBUF;
316 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
317     "Maximum size of a buffer cache block");
318 
319 /*
320  * This lock synchronizes access to bd_request.
321  */
322 static struct mtx_padalign __exclusive_cache_line bdlock;
323 
324 /*
325  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
326  * waitrunningbufspace().
327  */
328 static struct mtx_padalign __exclusive_cache_line rbreqlock;
329 
330 /*
331  * Lock that protects bdirtywait.
332  */
333 static struct mtx_padalign __exclusive_cache_line bdirtylock;
334 
335 /*
336  * Wakeup point for bufdaemon, as well as indicator of whether it is already
337  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
338  * is idling.
339  */
340 static int bd_request;
341 
342 /*
343  * Request for the buf daemon to write more buffers than is indicated by
344  * lodirtybuf.  This may be necessary to push out excess dependencies or
345  * defragment the address space where a simple count of the number of dirty
346  * buffers is insufficient to characterize the demand for flushing them.
347  */
348 static int bd_speedupreq;
349 
350 /*
351  * Synchronization (sleep/wakeup) variable for active buffer space requests.
352  * Set when wait starts, cleared prior to wakeup().
353  * Used in runningbufwakeup() and waitrunningbufspace().
354  */
355 static int runningbufreq;
356 
357 /*
358  * Synchronization for bwillwrite() waiters.
359  */
360 static int bdirtywait;
361 
362 /*
363  * Definitions for the buffer free lists.
364  */
365 #define QUEUE_NONE	0	/* on no queue */
366 #define QUEUE_EMPTY	1	/* empty buffer headers */
367 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
368 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
369 #define QUEUE_SENTINEL	4	/* not an queue index, but mark for sentinel */
370 
371 /* Maximum number of buffer domains. */
372 #define	BUF_DOMAINS	8
373 
374 struct bufdomainset bdlodirty;		/* Domains > lodirty */
375 struct bufdomainset bdhidirty;		/* Domains > hidirty */
376 
377 /* Configured number of clean queues. */
378 static int __read_mostly buf_domains;
379 
380 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
381 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
382 struct bufqueue __exclusive_cache_line bqempty;
383 
384 /*
385  * per-cpu empty buffer cache.
386  */
387 uma_zone_t buf_zone;
388 
389 /*
390  * Single global constant for BUF_WMESG, to avoid getting multiple references.
391  * buf_wmesg is referred from macros.
392  */
393 const char *buf_wmesg = BUF_WMESG;
394 
395 static int
396 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
397 {
398 	long value;
399 	int error;
400 
401 	value = *(long *)arg1;
402 	error = sysctl_handle_long(oidp, &value, 0, req);
403 	if (error != 0 || req->newptr == NULL)
404 		return (error);
405 	mtx_lock(&rbreqlock);
406 	if (arg1 == &hirunningspace) {
407 		if (value < lorunningspace)
408 			error = EINVAL;
409 		else
410 			hirunningspace = value;
411 	} else {
412 		KASSERT(arg1 == &lorunningspace,
413 		    ("%s: unknown arg1", __func__));
414 		if (value > hirunningspace)
415 			error = EINVAL;
416 		else
417 			lorunningspace = value;
418 	}
419 	mtx_unlock(&rbreqlock);
420 	return (error);
421 }
422 
423 static int
424 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
425 {
426 	int error;
427 	int value;
428 	int i;
429 
430 	value = *(int *)arg1;
431 	error = sysctl_handle_int(oidp, &value, 0, req);
432 	if (error != 0 || req->newptr == NULL)
433 		return (error);
434 	*(int *)arg1 = value;
435 	for (i = 0; i < buf_domains; i++)
436 		*(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
437 		    value / buf_domains;
438 
439 	return (error);
440 }
441 
442 static int
443 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
444 {
445 	long value;
446 	int error;
447 	int i;
448 
449 	value = *(long *)arg1;
450 	error = sysctl_handle_long(oidp, &value, 0, req);
451 	if (error != 0 || req->newptr == NULL)
452 		return (error);
453 	*(long *)arg1 = value;
454 	for (i = 0; i < buf_domains; i++)
455 		*(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
456 		    value / buf_domains;
457 
458 	return (error);
459 }
460 
461 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
462     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
463 static int
464 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
465 {
466 	long lvalue;
467 	int ivalue;
468 	int i;
469 
470 	lvalue = 0;
471 	for (i = 0; i < buf_domains; i++)
472 		lvalue += bdomain[i].bd_bufspace;
473 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
474 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
475 	if (lvalue > INT_MAX)
476 		/* On overflow, still write out a long to trigger ENOMEM. */
477 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
478 	ivalue = lvalue;
479 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
480 }
481 #else
482 static int
483 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
484 {
485 	long lvalue;
486 	int i;
487 
488 	lvalue = 0;
489 	for (i = 0; i < buf_domains; i++)
490 		lvalue += bdomain[i].bd_bufspace;
491 	return (sysctl_handle_long(oidp, &lvalue, 0, req));
492 }
493 #endif
494 
495 static int
496 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
497 {
498 	int value;
499 	int i;
500 
501 	value = 0;
502 	for (i = 0; i < buf_domains; i++)
503 		value += bdomain[i].bd_numdirtybuffers;
504 	return (sysctl_handle_int(oidp, &value, 0, req));
505 }
506 
507 /*
508  *	bdirtywakeup:
509  *
510  *	Wakeup any bwillwrite() waiters.
511  */
512 static void
513 bdirtywakeup(void)
514 {
515 	mtx_lock(&bdirtylock);
516 	if (bdirtywait) {
517 		bdirtywait = 0;
518 		wakeup(&bdirtywait);
519 	}
520 	mtx_unlock(&bdirtylock);
521 }
522 
523 /*
524  *	bd_clear:
525  *
526  *	Clear a domain from the appropriate bitsets when dirtybuffers
527  *	is decremented.
528  */
529 static void
530 bd_clear(struct bufdomain *bd)
531 {
532 
533 	mtx_lock(&bdirtylock);
534 	if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
535 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
536 	if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
537 		BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
538 	mtx_unlock(&bdirtylock);
539 }
540 
541 /*
542  *	bd_set:
543  *
544  *	Set a domain in the appropriate bitsets when dirtybuffers
545  *	is incremented.
546  */
547 static void
548 bd_set(struct bufdomain *bd)
549 {
550 
551 	mtx_lock(&bdirtylock);
552 	if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
553 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
554 	if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
555 		BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
556 	mtx_unlock(&bdirtylock);
557 }
558 
559 /*
560  *	bdirtysub:
561  *
562  *	Decrement the numdirtybuffers count by one and wakeup any
563  *	threads blocked in bwillwrite().
564  */
565 static void
566 bdirtysub(struct buf *bp)
567 {
568 	struct bufdomain *bd;
569 	int num;
570 
571 	bd = bufdomain(bp);
572 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
573 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
574 		bdirtywakeup();
575 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
576 		bd_clear(bd);
577 }
578 
579 /*
580  *	bdirtyadd:
581  *
582  *	Increment the numdirtybuffers count by one and wakeup the buf
583  *	daemon if needed.
584  */
585 static void
586 bdirtyadd(struct buf *bp)
587 {
588 	struct bufdomain *bd;
589 	int num;
590 
591 	/*
592 	 * Only do the wakeup once as we cross the boundary.  The
593 	 * buf daemon will keep running until the condition clears.
594 	 */
595 	bd = bufdomain(bp);
596 	num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
597 	if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
598 		bd_wakeup();
599 	if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
600 		bd_set(bd);
601 }
602 
603 /*
604  *	bufspace_daemon_wakeup:
605  *
606  *	Wakeup the daemons responsible for freeing clean bufs.
607  */
608 static void
609 bufspace_daemon_wakeup(struct bufdomain *bd)
610 {
611 
612 	/*
613 	 * avoid the lock if the daemon is running.
614 	 */
615 	if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
616 		BD_RUN_LOCK(bd);
617 		atomic_store_int(&bd->bd_running, 1);
618 		wakeup(&bd->bd_running);
619 		BD_RUN_UNLOCK(bd);
620 	}
621 }
622 
623 /*
624  *	bufspace_daemon_wait:
625  *
626  *	Sleep until the domain falls below a limit or one second passes.
627  */
628 static void
629 bufspace_daemon_wait(struct bufdomain *bd)
630 {
631 	/*
632 	 * Re-check our limits and sleep.  bd_running must be
633 	 * cleared prior to checking the limits to avoid missed
634 	 * wakeups.  The waker will adjust one of bufspace or
635 	 * freebuffers prior to checking bd_running.
636 	 */
637 	BD_RUN_LOCK(bd);
638 	atomic_store_int(&bd->bd_running, 0);
639 	if (bd->bd_bufspace < bd->bd_bufspacethresh &&
640 	    bd->bd_freebuffers > bd->bd_lofreebuffers) {
641 		msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
642 		    "-", hz);
643 	} else {
644 		/* Avoid spurious wakeups while running. */
645 		atomic_store_int(&bd->bd_running, 1);
646 		BD_RUN_UNLOCK(bd);
647 	}
648 }
649 
650 /*
651  *	bufspace_adjust:
652  *
653  *	Adjust the reported bufspace for a KVA managed buffer, possibly
654  * 	waking any waiters.
655  */
656 static void
657 bufspace_adjust(struct buf *bp, int bufsize)
658 {
659 	struct bufdomain *bd;
660 	long space;
661 	int diff;
662 
663 	KASSERT((bp->b_flags & B_MALLOC) == 0,
664 	    ("bufspace_adjust: malloc buf %p", bp));
665 	bd = bufdomain(bp);
666 	diff = bufsize - bp->b_bufsize;
667 	if (diff < 0) {
668 		atomic_subtract_long(&bd->bd_bufspace, -diff);
669 	} else if (diff > 0) {
670 		space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
671 		/* Wake up the daemon on the transition. */
672 		if (space < bd->bd_bufspacethresh &&
673 		    space + diff >= bd->bd_bufspacethresh)
674 			bufspace_daemon_wakeup(bd);
675 	}
676 	bp->b_bufsize = bufsize;
677 }
678 
679 /*
680  *	bufspace_reserve:
681  *
682  *	Reserve bufspace before calling allocbuf().  metadata has a
683  *	different space limit than data.
684  */
685 static int
686 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
687 {
688 	long limit, new;
689 	long space;
690 
691 	if (metadata)
692 		limit = bd->bd_maxbufspace;
693 	else
694 		limit = bd->bd_hibufspace;
695 	space = atomic_fetchadd_long(&bd->bd_bufspace, size);
696 	new = space + size;
697 	if (new > limit) {
698 		atomic_subtract_long(&bd->bd_bufspace, size);
699 		return (ENOSPC);
700 	}
701 
702 	/* Wake up the daemon on the transition. */
703 	if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
704 		bufspace_daemon_wakeup(bd);
705 
706 	return (0);
707 }
708 
709 /*
710  *	bufspace_release:
711  *
712  *	Release reserved bufspace after bufspace_adjust() has consumed it.
713  */
714 static void
715 bufspace_release(struct bufdomain *bd, int size)
716 {
717 
718 	atomic_subtract_long(&bd->bd_bufspace, size);
719 }
720 
721 /*
722  *	bufspace_wait:
723  *
724  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
725  *	supplied.  bd_wanted must be set prior to polling for space.  The
726  *	operation must be re-tried on return.
727  */
728 static void
729 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
730     int slpflag, int slptimeo)
731 {
732 	struct thread *td;
733 	int error, fl, norunbuf;
734 
735 	if ((gbflags & GB_NOWAIT_BD) != 0)
736 		return;
737 
738 	td = curthread;
739 	BD_LOCK(bd);
740 	while (bd->bd_wanted) {
741 		if (vp != NULL && vp->v_type != VCHR &&
742 		    (td->td_pflags & TDP_BUFNEED) == 0) {
743 			BD_UNLOCK(bd);
744 			/*
745 			 * getblk() is called with a vnode locked, and
746 			 * some majority of the dirty buffers may as
747 			 * well belong to the vnode.  Flushing the
748 			 * buffers there would make a progress that
749 			 * cannot be achieved by the buf_daemon, that
750 			 * cannot lock the vnode.
751 			 */
752 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
753 			    (td->td_pflags & TDP_NORUNNINGBUF);
754 
755 			/*
756 			 * Play bufdaemon.  The getnewbuf() function
757 			 * may be called while the thread owns lock
758 			 * for another dirty buffer for the same
759 			 * vnode, which makes it impossible to use
760 			 * VOP_FSYNC() there, due to the buffer lock
761 			 * recursion.
762 			 */
763 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
764 			fl = buf_flush(vp, bd, flushbufqtarget);
765 			td->td_pflags &= norunbuf;
766 			BD_LOCK(bd);
767 			if (fl != 0)
768 				continue;
769 			if (bd->bd_wanted == 0)
770 				break;
771 		}
772 		error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
773 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
774 		if (error != 0)
775 			break;
776 	}
777 	BD_UNLOCK(bd);
778 }
779 
780 
781 /*
782  *	bufspace_daemon:
783  *
784  *	buffer space management daemon.  Tries to maintain some marginal
785  *	amount of free buffer space so that requesting processes neither
786  *	block nor work to reclaim buffers.
787  */
788 static void
789 bufspace_daemon(void *arg)
790 {
791 	struct bufdomain *bd;
792 
793 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
794 	    SHUTDOWN_PRI_LAST + 100);
795 
796 	bd = arg;
797 	for (;;) {
798 		kthread_suspend_check();
799 
800 		/*
801 		 * Free buffers from the clean queue until we meet our
802 		 * targets.
803 		 *
804 		 * Theory of operation:  The buffer cache is most efficient
805 		 * when some free buffer headers and space are always
806 		 * available to getnewbuf().  This daemon attempts to prevent
807 		 * the excessive blocking and synchronization associated
808 		 * with shortfall.  It goes through three phases according
809 		 * demand:
810 		 *
811 		 * 1)	The daemon wakes up voluntarily once per-second
812 		 *	during idle periods when the counters are below
813 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
814 		 *
815 		 * 2)	The daemon wakes up as we cross the thresholds
816 		 *	ahead of any potential blocking.  This may bounce
817 		 *	slightly according to the rate of consumption and
818 		 *	release.
819 		 *
820 		 * 3)	The daemon and consumers are starved for working
821 		 *	clean buffers.  This is the 'bufspace' sleep below
822 		 *	which will inefficiently trade bufs with bqrelse
823 		 *	until we return to condition 2.
824 		 */
825 		while (bd->bd_bufspace > bd->bd_lobufspace ||
826 		    bd->bd_freebuffers < bd->bd_hifreebuffers) {
827 			if (buf_recycle(bd, false) != 0) {
828 				if (bd_flushall(bd))
829 					continue;
830 				/*
831 				 * Speedup dirty if we've run out of clean
832 				 * buffers.  This is possible in particular
833 				 * because softdep may held many bufs locked
834 				 * pending writes to other bufs which are
835 				 * marked for delayed write, exhausting
836 				 * clean space until they are written.
837 				 */
838 				bd_speedup();
839 				BD_LOCK(bd);
840 				if (bd->bd_wanted) {
841 					msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
842 					    PRIBIO|PDROP, "bufspace", hz/10);
843 				} else
844 					BD_UNLOCK(bd);
845 			}
846 			maybe_yield();
847 		}
848 		bufspace_daemon_wait(bd);
849 	}
850 }
851 
852 /*
853  *	bufmallocadjust:
854  *
855  *	Adjust the reported bufspace for a malloc managed buffer, possibly
856  *	waking any waiters.
857  */
858 static void
859 bufmallocadjust(struct buf *bp, int bufsize)
860 {
861 	int diff;
862 
863 	KASSERT((bp->b_flags & B_MALLOC) != 0,
864 	    ("bufmallocadjust: non-malloc buf %p", bp));
865 	diff = bufsize - bp->b_bufsize;
866 	if (diff < 0)
867 		atomic_subtract_long(&bufmallocspace, -diff);
868 	else
869 		atomic_add_long(&bufmallocspace, diff);
870 	bp->b_bufsize = bufsize;
871 }
872 
873 /*
874  *	runningwakeup:
875  *
876  *	Wake up processes that are waiting on asynchronous writes to fall
877  *	below lorunningspace.
878  */
879 static void
880 runningwakeup(void)
881 {
882 
883 	mtx_lock(&rbreqlock);
884 	if (runningbufreq) {
885 		runningbufreq = 0;
886 		wakeup(&runningbufreq);
887 	}
888 	mtx_unlock(&rbreqlock);
889 }
890 
891 /*
892  *	runningbufwakeup:
893  *
894  *	Decrement the outstanding write count according.
895  */
896 void
897 runningbufwakeup(struct buf *bp)
898 {
899 	long space, bspace;
900 
901 	bspace = bp->b_runningbufspace;
902 	if (bspace == 0)
903 		return;
904 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
905 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
906 	    space, bspace));
907 	bp->b_runningbufspace = 0;
908 	/*
909 	 * Only acquire the lock and wakeup on the transition from exceeding
910 	 * the threshold to falling below it.
911 	 */
912 	if (space < lorunningspace)
913 		return;
914 	if (space - bspace > lorunningspace)
915 		return;
916 	runningwakeup();
917 }
918 
919 /*
920  *	waitrunningbufspace()
921  *
922  *	runningbufspace is a measure of the amount of I/O currently
923  *	running.  This routine is used in async-write situations to
924  *	prevent creating huge backups of pending writes to a device.
925  *	Only asynchronous writes are governed by this function.
926  *
927  *	This does NOT turn an async write into a sync write.  It waits
928  *	for earlier writes to complete and generally returns before the
929  *	caller's write has reached the device.
930  */
931 void
932 waitrunningbufspace(void)
933 {
934 
935 	mtx_lock(&rbreqlock);
936 	while (runningbufspace > hirunningspace) {
937 		runningbufreq = 1;
938 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
939 	}
940 	mtx_unlock(&rbreqlock);
941 }
942 
943 
944 /*
945  *	vfs_buf_test_cache:
946  *
947  *	Called when a buffer is extended.  This function clears the B_CACHE
948  *	bit if the newly extended portion of the buffer does not contain
949  *	valid data.
950  */
951 static __inline void
952 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
953     vm_offset_t size, vm_page_t m)
954 {
955 
956 	VM_OBJECT_ASSERT_LOCKED(m->object);
957 	if (bp->b_flags & B_CACHE) {
958 		int base = (foff + off) & PAGE_MASK;
959 		if (vm_page_is_valid(m, base, size) == 0)
960 			bp->b_flags &= ~B_CACHE;
961 	}
962 }
963 
964 /* Wake up the buffer daemon if necessary */
965 static void
966 bd_wakeup(void)
967 {
968 
969 	mtx_lock(&bdlock);
970 	if (bd_request == 0) {
971 		bd_request = 1;
972 		wakeup(&bd_request);
973 	}
974 	mtx_unlock(&bdlock);
975 }
976 
977 /*
978  * Adjust the maxbcachbuf tunable.
979  */
980 static void
981 maxbcachebuf_adjust(void)
982 {
983 	int i;
984 
985 	/*
986 	 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
987 	 */
988 	i = 2;
989 	while (i * 2 <= maxbcachebuf)
990 		i *= 2;
991 	maxbcachebuf = i;
992 	if (maxbcachebuf < MAXBSIZE)
993 		maxbcachebuf = MAXBSIZE;
994 	if (maxbcachebuf > MAXPHYS)
995 		maxbcachebuf = MAXPHYS;
996 	if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
997 		printf("maxbcachebuf=%d\n", maxbcachebuf);
998 }
999 
1000 /*
1001  * bd_speedup - speedup the buffer cache flushing code
1002  */
1003 void
1004 bd_speedup(void)
1005 {
1006 	int needwake;
1007 
1008 	mtx_lock(&bdlock);
1009 	needwake = 0;
1010 	if (bd_speedupreq == 0 || bd_request == 0)
1011 		needwake = 1;
1012 	bd_speedupreq = 1;
1013 	bd_request = 1;
1014 	if (needwake)
1015 		wakeup(&bd_request);
1016 	mtx_unlock(&bdlock);
1017 }
1018 
1019 #ifdef __i386__
1020 #define	TRANSIENT_DENOM	5
1021 #else
1022 #define	TRANSIENT_DENOM 10
1023 #endif
1024 
1025 /*
1026  * Calculating buffer cache scaling values and reserve space for buffer
1027  * headers.  This is called during low level kernel initialization and
1028  * may be called more then once.  We CANNOT write to the memory area
1029  * being reserved at this time.
1030  */
1031 caddr_t
1032 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1033 {
1034 	int tuned_nbuf;
1035 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
1036 
1037 	/*
1038 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
1039 	 * PAGE_SIZE is >= 1K)
1040 	 */
1041 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
1042 
1043 	maxbcachebuf_adjust();
1044 	/*
1045 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1046 	 * For the first 64MB of ram nominally allocate sufficient buffers to
1047 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
1048 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
1049 	 * the buffer cache we limit the eventual kva reservation to
1050 	 * maxbcache bytes.
1051 	 *
1052 	 * factor represents the 1/4 x ram conversion.
1053 	 */
1054 	if (nbuf == 0) {
1055 		int factor = 4 * BKVASIZE / 1024;
1056 
1057 		nbuf = 50;
1058 		if (physmem_est > 4096)
1059 			nbuf += min((physmem_est - 4096) / factor,
1060 			    65536 / factor);
1061 		if (physmem_est > 65536)
1062 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1063 			    32 * 1024 * 1024 / (factor * 5));
1064 
1065 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
1066 			nbuf = maxbcache / BKVASIZE;
1067 		tuned_nbuf = 1;
1068 	} else
1069 		tuned_nbuf = 0;
1070 
1071 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
1072 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
1073 	if (nbuf > maxbuf) {
1074 		if (!tuned_nbuf)
1075 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1076 			    maxbuf);
1077 		nbuf = maxbuf;
1078 	}
1079 
1080 	/*
1081 	 * Ideal allocation size for the transient bio submap is 10%
1082 	 * of the maximal space buffer map.  This roughly corresponds
1083 	 * to the amount of the buffer mapped for typical UFS load.
1084 	 *
1085 	 * Clip the buffer map to reserve space for the transient
1086 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1087 	 * maximum buffer map extent on the platform.
1088 	 *
1089 	 * The fall-back to the maxbuf in case of maxbcache unset,
1090 	 * allows to not trim the buffer KVA for the architectures
1091 	 * with ample KVA space.
1092 	 */
1093 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1094 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1095 		buf_sz = (long)nbuf * BKVASIZE;
1096 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1097 		    (TRANSIENT_DENOM - 1)) {
1098 			/*
1099 			 * There is more KVA than memory.  Do not
1100 			 * adjust buffer map size, and assign the rest
1101 			 * of maxbuf to transient map.
1102 			 */
1103 			biotmap_sz = maxbuf_sz - buf_sz;
1104 		} else {
1105 			/*
1106 			 * Buffer map spans all KVA we could afford on
1107 			 * this platform.  Give 10% (20% on i386) of
1108 			 * the buffer map to the transient bio map.
1109 			 */
1110 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
1111 			buf_sz -= biotmap_sz;
1112 		}
1113 		if (biotmap_sz / INT_MAX > MAXPHYS)
1114 			bio_transient_maxcnt = INT_MAX;
1115 		else
1116 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1117 		/*
1118 		 * Artificially limit to 1024 simultaneous in-flight I/Os
1119 		 * using the transient mapping.
1120 		 */
1121 		if (bio_transient_maxcnt > 1024)
1122 			bio_transient_maxcnt = 1024;
1123 		if (tuned_nbuf)
1124 			nbuf = buf_sz / BKVASIZE;
1125 	}
1126 
1127 	if (nswbuf == 0) {
1128 		nswbuf = min(nbuf / 4, 256);
1129 		if (nswbuf < NSWBUF_MIN)
1130 			nswbuf = NSWBUF_MIN;
1131 	}
1132 
1133 	/*
1134 	 * Reserve space for the buffer cache buffers
1135 	 */
1136 	buf = (void *)v;
1137 	v = (caddr_t)(buf + nbuf);
1138 
1139 	return(v);
1140 }
1141 
1142 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1143 void
1144 bufinit(void)
1145 {
1146 	struct buf *bp;
1147 	int i;
1148 
1149 	KASSERT(maxbcachebuf >= MAXBSIZE,
1150 	    ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1151 	    MAXBSIZE));
1152 	bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1153 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1154 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1155 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1156 
1157 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1158 
1159 	/* finally, initialize each buffer header and stick on empty q */
1160 	for (i = 0; i < nbuf; i++) {
1161 		bp = &buf[i];
1162 		bzero(bp, sizeof *bp);
1163 		bp->b_flags = B_INVAL;
1164 		bp->b_rcred = NOCRED;
1165 		bp->b_wcred = NOCRED;
1166 		bp->b_qindex = QUEUE_NONE;
1167 		bp->b_domain = -1;
1168 		bp->b_subqueue = mp_maxid + 1;
1169 		bp->b_xflags = 0;
1170 		bp->b_data = bp->b_kvabase = unmapped_buf;
1171 		LIST_INIT(&bp->b_dep);
1172 		BUF_LOCKINIT(bp);
1173 		bq_insert(&bqempty, bp, false);
1174 	}
1175 
1176 	/*
1177 	 * maxbufspace is the absolute maximum amount of buffer space we are
1178 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1179 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1180 	 * used by most other requests.  The differential is required to
1181 	 * ensure that metadata deadlocks don't occur.
1182 	 *
1183 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1184 	 * this may result in KVM fragmentation which is not handled optimally
1185 	 * by the system. XXX This is less true with vmem.  We could use
1186 	 * PAGE_SIZE.
1187 	 */
1188 	maxbufspace = (long)nbuf * BKVASIZE;
1189 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1190 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1191 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1192 
1193 	/*
1194 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1195 	 * arbitrarily and may need further tuning. It corresponds to
1196 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1197 	 * which fits with many RAID controllers' tagged queuing limits.
1198 	 * The lower 1 MiB limit is the historical upper limit for
1199 	 * hirunningspace.
1200 	 */
1201 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1202 	    16 * 1024 * 1024), 1024 * 1024);
1203 	lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1204 
1205 	/*
1206 	 * Limit the amount of malloc memory since it is wired permanently into
1207 	 * the kernel space.  Even though this is accounted for in the buffer
1208 	 * allocation, we don't want the malloced region to grow uncontrolled.
1209 	 * The malloc scheme improves memory utilization significantly on
1210 	 * average (small) directories.
1211 	 */
1212 	maxbufmallocspace = hibufspace / 20;
1213 
1214 	/*
1215 	 * Reduce the chance of a deadlock occurring by limiting the number
1216 	 * of delayed-write dirty buffers we allow to stack up.
1217 	 */
1218 	hidirtybuffers = nbuf / 4 + 20;
1219 	dirtybufthresh = hidirtybuffers * 9 / 10;
1220 	/*
1221 	 * To support extreme low-memory systems, make sure hidirtybuffers
1222 	 * cannot eat up all available buffer space.  This occurs when our
1223 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1224 	 * buffer space assuming BKVASIZE'd buffers.
1225 	 */
1226 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1227 		hidirtybuffers >>= 1;
1228 	}
1229 	lodirtybuffers = hidirtybuffers / 2;
1230 
1231 	/*
1232 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1233 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1234 	 * are counted as free but will be unavailable to threads executing
1235 	 * on other cpus.
1236 	 *
1237 	 * hifreebuffers is the free target for the bufspace daemon.  This
1238 	 * should be set appropriately to limit work per-iteration.
1239 	 */
1240 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1241 	hifreebuffers = (3 * lofreebuffers) / 2;
1242 	numfreebuffers = nbuf;
1243 
1244 	/* Setup the kva and free list allocators. */
1245 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1246 	buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1247 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1248 
1249 	/*
1250 	 * Size the clean queue according to the amount of buffer space.
1251 	 * One queue per-256mb up to the max.  More queues gives better
1252 	 * concurrency but less accurate LRU.
1253 	 */
1254 	buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1255 	for (i = 0 ; i < buf_domains; i++) {
1256 		struct bufdomain *bd;
1257 
1258 		bd = &bdomain[i];
1259 		bd_init(bd);
1260 		bd->bd_freebuffers = nbuf / buf_domains;
1261 		bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1262 		bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1263 		bd->bd_bufspace = 0;
1264 		bd->bd_maxbufspace = maxbufspace / buf_domains;
1265 		bd->bd_hibufspace = hibufspace / buf_domains;
1266 		bd->bd_lobufspace = lobufspace / buf_domains;
1267 		bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1268 		bd->bd_numdirtybuffers = 0;
1269 		bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1270 		bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1271 		bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1272 		/* Don't allow more than 2% of bufs in the per-cpu caches. */
1273 		bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1274 	}
1275 	getnewbufcalls = counter_u64_alloc(M_WAITOK);
1276 	getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1277 	mappingrestarts = counter_u64_alloc(M_WAITOK);
1278 	numbufallocfails = counter_u64_alloc(M_WAITOK);
1279 	notbufdflushes = counter_u64_alloc(M_WAITOK);
1280 	buffreekvacnt = counter_u64_alloc(M_WAITOK);
1281 	bufdefragcnt = counter_u64_alloc(M_WAITOK);
1282 	bufkvaspace = counter_u64_alloc(M_WAITOK);
1283 }
1284 
1285 #ifdef INVARIANTS
1286 static inline void
1287 vfs_buf_check_mapped(struct buf *bp)
1288 {
1289 
1290 	KASSERT(bp->b_kvabase != unmapped_buf,
1291 	    ("mapped buf: b_kvabase was not updated %p", bp));
1292 	KASSERT(bp->b_data != unmapped_buf,
1293 	    ("mapped buf: b_data was not updated %p", bp));
1294 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1295 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1296 }
1297 
1298 static inline void
1299 vfs_buf_check_unmapped(struct buf *bp)
1300 {
1301 
1302 	KASSERT(bp->b_data == unmapped_buf,
1303 	    ("unmapped buf: corrupted b_data %p", bp));
1304 }
1305 
1306 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1307 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1308 #else
1309 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1310 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1311 #endif
1312 
1313 static int
1314 isbufbusy(struct buf *bp)
1315 {
1316 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1317 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1318 		return (1);
1319 	return (0);
1320 }
1321 
1322 /*
1323  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1324  */
1325 void
1326 bufshutdown(int show_busybufs)
1327 {
1328 	static int first_buf_printf = 1;
1329 	struct buf *bp;
1330 	int iter, nbusy, pbusy;
1331 #ifndef PREEMPTION
1332 	int subiter;
1333 #endif
1334 
1335 	/*
1336 	 * Sync filesystems for shutdown
1337 	 */
1338 	wdog_kern_pat(WD_LASTVAL);
1339 	sys_sync(curthread, NULL);
1340 
1341 	/*
1342 	 * With soft updates, some buffers that are
1343 	 * written will be remarked as dirty until other
1344 	 * buffers are written.
1345 	 */
1346 	for (iter = pbusy = 0; iter < 20; iter++) {
1347 		nbusy = 0;
1348 		for (bp = &buf[nbuf]; --bp >= buf; )
1349 			if (isbufbusy(bp))
1350 				nbusy++;
1351 		if (nbusy == 0) {
1352 			if (first_buf_printf)
1353 				printf("All buffers synced.");
1354 			break;
1355 		}
1356 		if (first_buf_printf) {
1357 			printf("Syncing disks, buffers remaining... ");
1358 			first_buf_printf = 0;
1359 		}
1360 		printf("%d ", nbusy);
1361 		if (nbusy < pbusy)
1362 			iter = 0;
1363 		pbusy = nbusy;
1364 
1365 		wdog_kern_pat(WD_LASTVAL);
1366 		sys_sync(curthread, NULL);
1367 
1368 #ifdef PREEMPTION
1369 		/*
1370 		 * Spin for a while to allow interrupt threads to run.
1371 		 */
1372 		DELAY(50000 * iter);
1373 #else
1374 		/*
1375 		 * Context switch several times to allow interrupt
1376 		 * threads to run.
1377 		 */
1378 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1379 			thread_lock(curthread);
1380 			mi_switch(SW_VOL, NULL);
1381 			thread_unlock(curthread);
1382 			DELAY(1000);
1383 		}
1384 #endif
1385 	}
1386 	printf("\n");
1387 	/*
1388 	 * Count only busy local buffers to prevent forcing
1389 	 * a fsck if we're just a client of a wedged NFS server
1390 	 */
1391 	nbusy = 0;
1392 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1393 		if (isbufbusy(bp)) {
1394 #if 0
1395 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1396 			if (bp->b_dev == NULL) {
1397 				TAILQ_REMOVE(&mountlist,
1398 				    bp->b_vp->v_mount, mnt_list);
1399 				continue;
1400 			}
1401 #endif
1402 			nbusy++;
1403 			if (show_busybufs > 0) {
1404 				printf(
1405 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1406 				    nbusy, bp, bp->b_vp, bp->b_flags,
1407 				    (intmax_t)bp->b_blkno,
1408 				    (intmax_t)bp->b_lblkno);
1409 				BUF_LOCKPRINTINFO(bp);
1410 				if (show_busybufs > 1)
1411 					vn_printf(bp->b_vp,
1412 					    "vnode content: ");
1413 			}
1414 		}
1415 	}
1416 	if (nbusy) {
1417 		/*
1418 		 * Failed to sync all blocks. Indicate this and don't
1419 		 * unmount filesystems (thus forcing an fsck on reboot).
1420 		 */
1421 		printf("Giving up on %d buffers\n", nbusy);
1422 		DELAY(5000000);	/* 5 seconds */
1423 	} else {
1424 		if (!first_buf_printf)
1425 			printf("Final sync complete\n");
1426 		/*
1427 		 * Unmount filesystems
1428 		 */
1429 		if (panicstr == NULL)
1430 			vfs_unmountall();
1431 	}
1432 	swapoff_all();
1433 	DELAY(100000);		/* wait for console output to finish */
1434 }
1435 
1436 static void
1437 bpmap_qenter(struct buf *bp)
1438 {
1439 
1440 	BUF_CHECK_MAPPED(bp);
1441 
1442 	/*
1443 	 * bp->b_data is relative to bp->b_offset, but
1444 	 * bp->b_offset may be offset into the first page.
1445 	 */
1446 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1447 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1448 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1449 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1450 }
1451 
1452 static inline struct bufdomain *
1453 bufdomain(struct buf *bp)
1454 {
1455 
1456 	return (&bdomain[bp->b_domain]);
1457 }
1458 
1459 static struct bufqueue *
1460 bufqueue(struct buf *bp)
1461 {
1462 
1463 	switch (bp->b_qindex) {
1464 	case QUEUE_NONE:
1465 		/* FALLTHROUGH */
1466 	case QUEUE_SENTINEL:
1467 		return (NULL);
1468 	case QUEUE_EMPTY:
1469 		return (&bqempty);
1470 	case QUEUE_DIRTY:
1471 		return (&bufdomain(bp)->bd_dirtyq);
1472 	case QUEUE_CLEAN:
1473 		return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1474 	default:
1475 		break;
1476 	}
1477 	panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1478 }
1479 
1480 /*
1481  * Return the locked bufqueue that bp is a member of.
1482  */
1483 static struct bufqueue *
1484 bufqueue_acquire(struct buf *bp)
1485 {
1486 	struct bufqueue *bq, *nbq;
1487 
1488 	/*
1489 	 * bp can be pushed from a per-cpu queue to the
1490 	 * cleanq while we're waiting on the lock.  Retry
1491 	 * if the queues don't match.
1492 	 */
1493 	bq = bufqueue(bp);
1494 	BQ_LOCK(bq);
1495 	for (;;) {
1496 		nbq = bufqueue(bp);
1497 		if (bq == nbq)
1498 			break;
1499 		BQ_UNLOCK(bq);
1500 		BQ_LOCK(nbq);
1501 		bq = nbq;
1502 	}
1503 	return (bq);
1504 }
1505 
1506 /*
1507  *	binsfree:
1508  *
1509  *	Insert the buffer into the appropriate free list.  Requires a
1510  *	locked buffer on entry and buffer is unlocked before return.
1511  */
1512 static void
1513 binsfree(struct buf *bp, int qindex)
1514 {
1515 	struct bufdomain *bd;
1516 	struct bufqueue *bq;
1517 
1518 	KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1519 	    ("binsfree: Invalid qindex %d", qindex));
1520 	BUF_ASSERT_XLOCKED(bp);
1521 
1522 	/*
1523 	 * Handle delayed bremfree() processing.
1524 	 */
1525 	if (bp->b_flags & B_REMFREE) {
1526 		if (bp->b_qindex == qindex) {
1527 			bp->b_flags |= B_REUSE;
1528 			bp->b_flags &= ~B_REMFREE;
1529 			BUF_UNLOCK(bp);
1530 			return;
1531 		}
1532 		bq = bufqueue_acquire(bp);
1533 		bq_remove(bq, bp);
1534 		BQ_UNLOCK(bq);
1535 	}
1536 	bd = bufdomain(bp);
1537 	if (qindex == QUEUE_CLEAN) {
1538 		if (bd->bd_lim != 0)
1539 			bq = &bd->bd_subq[PCPU_GET(cpuid)];
1540 		else
1541 			bq = bd->bd_cleanq;
1542 	} else
1543 		bq = &bd->bd_dirtyq;
1544 	bq_insert(bq, bp, true);
1545 }
1546 
1547 /*
1548  * buf_free:
1549  *
1550  *	Free a buffer to the buf zone once it no longer has valid contents.
1551  */
1552 static void
1553 buf_free(struct buf *bp)
1554 {
1555 
1556 	if (bp->b_flags & B_REMFREE)
1557 		bremfreef(bp);
1558 	if (bp->b_vflags & BV_BKGRDINPROG)
1559 		panic("losing buffer 1");
1560 	if (bp->b_rcred != NOCRED) {
1561 		crfree(bp->b_rcred);
1562 		bp->b_rcred = NOCRED;
1563 	}
1564 	if (bp->b_wcred != NOCRED) {
1565 		crfree(bp->b_wcred);
1566 		bp->b_wcred = NOCRED;
1567 	}
1568 	if (!LIST_EMPTY(&bp->b_dep))
1569 		buf_deallocate(bp);
1570 	bufkva_free(bp);
1571 	atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1572 	BUF_UNLOCK(bp);
1573 	uma_zfree(buf_zone, bp);
1574 }
1575 
1576 /*
1577  * buf_import:
1578  *
1579  *	Import bufs into the uma cache from the buf list.  The system still
1580  *	expects a static array of bufs and much of the synchronization
1581  *	around bufs assumes type stable storage.  As a result, UMA is used
1582  *	only as a per-cpu cache of bufs still maintained on a global list.
1583  */
1584 static int
1585 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1586 {
1587 	struct buf *bp;
1588 	int i;
1589 
1590 	BQ_LOCK(&bqempty);
1591 	for (i = 0; i < cnt; i++) {
1592 		bp = TAILQ_FIRST(&bqempty.bq_queue);
1593 		if (bp == NULL)
1594 			break;
1595 		bq_remove(&bqempty, bp);
1596 		store[i] = bp;
1597 	}
1598 	BQ_UNLOCK(&bqempty);
1599 
1600 	return (i);
1601 }
1602 
1603 /*
1604  * buf_release:
1605  *
1606  *	Release bufs from the uma cache back to the buffer queues.
1607  */
1608 static void
1609 buf_release(void *arg, void **store, int cnt)
1610 {
1611 	struct bufqueue *bq;
1612 	struct buf *bp;
1613         int i;
1614 
1615 	bq = &bqempty;
1616 	BQ_LOCK(bq);
1617         for (i = 0; i < cnt; i++) {
1618 		bp = store[i];
1619 		/* Inline bq_insert() to batch locking. */
1620 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1621 		bp->b_flags &= ~(B_AGE | B_REUSE);
1622 		bq->bq_len++;
1623 		bp->b_qindex = bq->bq_index;
1624 	}
1625 	BQ_UNLOCK(bq);
1626 }
1627 
1628 /*
1629  * buf_alloc:
1630  *
1631  *	Allocate an empty buffer header.
1632  */
1633 static struct buf *
1634 buf_alloc(struct bufdomain *bd)
1635 {
1636 	struct buf *bp;
1637 	int freebufs;
1638 
1639 	/*
1640 	 * We can only run out of bufs in the buf zone if the average buf
1641 	 * is less than BKVASIZE.  In this case the actual wait/block will
1642 	 * come from buf_reycle() failing to flush one of these small bufs.
1643 	 */
1644 	bp = NULL;
1645 	freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1646 	if (freebufs > 0)
1647 		bp = uma_zalloc(buf_zone, M_NOWAIT);
1648 	if (bp == NULL) {
1649 		atomic_add_int(&bd->bd_freebuffers, 1);
1650 		bufspace_daemon_wakeup(bd);
1651 		counter_u64_add(numbufallocfails, 1);
1652 		return (NULL);
1653 	}
1654 	/*
1655 	 * Wake-up the bufspace daemon on transition below threshold.
1656 	 */
1657 	if (freebufs == bd->bd_lofreebuffers)
1658 		bufspace_daemon_wakeup(bd);
1659 
1660 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1661 		panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1662 
1663 	KASSERT(bp->b_vp == NULL,
1664 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1665 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1666 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1667 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1668 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1669 	KASSERT(bp->b_npages == 0,
1670 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1671 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1672 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1673 
1674 	bp->b_domain = BD_DOMAIN(bd);
1675 	bp->b_flags = 0;
1676 	bp->b_ioflags = 0;
1677 	bp->b_xflags = 0;
1678 	bp->b_vflags = 0;
1679 	bp->b_vp = NULL;
1680 	bp->b_blkno = bp->b_lblkno = 0;
1681 	bp->b_offset = NOOFFSET;
1682 	bp->b_iodone = 0;
1683 	bp->b_error = 0;
1684 	bp->b_resid = 0;
1685 	bp->b_bcount = 0;
1686 	bp->b_npages = 0;
1687 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1688 	bp->b_bufobj = NULL;
1689 	bp->b_data = bp->b_kvabase = unmapped_buf;
1690 	bp->b_fsprivate1 = NULL;
1691 	bp->b_fsprivate2 = NULL;
1692 	bp->b_fsprivate3 = NULL;
1693 	LIST_INIT(&bp->b_dep);
1694 
1695 	return (bp);
1696 }
1697 
1698 /*
1699  *	buf_recycle:
1700  *
1701  *	Free a buffer from the given bufqueue.  kva controls whether the
1702  *	freed buf must own some kva resources.  This is used for
1703  *	defragmenting.
1704  */
1705 static int
1706 buf_recycle(struct bufdomain *bd, bool kva)
1707 {
1708 	struct bufqueue *bq;
1709 	struct buf *bp, *nbp;
1710 
1711 	if (kva)
1712 		counter_u64_add(bufdefragcnt, 1);
1713 	nbp = NULL;
1714 	bq = bd->bd_cleanq;
1715 	BQ_LOCK(bq);
1716 	KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1717 	    ("buf_recycle: Locks don't match"));
1718 	nbp = TAILQ_FIRST(&bq->bq_queue);
1719 
1720 	/*
1721 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1722 	 * depending.
1723 	 */
1724 	while ((bp = nbp) != NULL) {
1725 		/*
1726 		 * Calculate next bp (we can only use it if we do not
1727 		 * release the bqlock).
1728 		 */
1729 		nbp = TAILQ_NEXT(bp, b_freelist);
1730 
1731 		/*
1732 		 * If we are defragging then we need a buffer with
1733 		 * some kva to reclaim.
1734 		 */
1735 		if (kva && bp->b_kvasize == 0)
1736 			continue;
1737 
1738 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1739 			continue;
1740 
1741 		/*
1742 		 * Implement a second chance algorithm for frequently
1743 		 * accessed buffers.
1744 		 */
1745 		if ((bp->b_flags & B_REUSE) != 0) {
1746 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1747 			TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1748 			bp->b_flags &= ~B_REUSE;
1749 			BUF_UNLOCK(bp);
1750 			continue;
1751 		}
1752 
1753 		/*
1754 		 * Skip buffers with background writes in progress.
1755 		 */
1756 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1757 			BUF_UNLOCK(bp);
1758 			continue;
1759 		}
1760 
1761 		KASSERT(bp->b_qindex == QUEUE_CLEAN,
1762 		    ("buf_recycle: inconsistent queue %d bp %p",
1763 		    bp->b_qindex, bp));
1764 		KASSERT(bp->b_domain == BD_DOMAIN(bd),
1765 		    ("getnewbuf: queue domain %d doesn't match request %d",
1766 		    bp->b_domain, (int)BD_DOMAIN(bd)));
1767 		/*
1768 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1769 		 * the scan from this point on.
1770 		 */
1771 		bq_remove(bq, bp);
1772 		BQ_UNLOCK(bq);
1773 
1774 		/*
1775 		 * Requeue the background write buffer with error and
1776 		 * restart the scan.
1777 		 */
1778 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1779 			bqrelse(bp);
1780 			BQ_LOCK(bq);
1781 			nbp = TAILQ_FIRST(&bq->bq_queue);
1782 			continue;
1783 		}
1784 		bp->b_flags |= B_INVAL;
1785 		brelse(bp);
1786 		return (0);
1787 	}
1788 	bd->bd_wanted = 1;
1789 	BQ_UNLOCK(bq);
1790 
1791 	return (ENOBUFS);
1792 }
1793 
1794 /*
1795  *	bremfree:
1796  *
1797  *	Mark the buffer for removal from the appropriate free list.
1798  *
1799  */
1800 void
1801 bremfree(struct buf *bp)
1802 {
1803 
1804 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1805 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1806 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1807 	KASSERT(bp->b_qindex != QUEUE_NONE,
1808 	    ("bremfree: buffer %p not on a queue.", bp));
1809 	BUF_ASSERT_XLOCKED(bp);
1810 
1811 	bp->b_flags |= B_REMFREE;
1812 }
1813 
1814 /*
1815  *	bremfreef:
1816  *
1817  *	Force an immediate removal from a free list.  Used only in nfs when
1818  *	it abuses the b_freelist pointer.
1819  */
1820 void
1821 bremfreef(struct buf *bp)
1822 {
1823 	struct bufqueue *bq;
1824 
1825 	bq = bufqueue_acquire(bp);
1826 	bq_remove(bq, bp);
1827 	BQ_UNLOCK(bq);
1828 }
1829 
1830 static void
1831 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1832 {
1833 
1834 	mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1835 	TAILQ_INIT(&bq->bq_queue);
1836 	bq->bq_len = 0;
1837 	bq->bq_index = qindex;
1838 	bq->bq_subqueue = subqueue;
1839 }
1840 
1841 static void
1842 bd_init(struct bufdomain *bd)
1843 {
1844 	int i;
1845 
1846 	bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1847 	bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1848 	bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1849 	for (i = 0; i <= mp_maxid; i++)
1850 		bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1851 		    "bufq clean subqueue lock");
1852 	mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1853 }
1854 
1855 /*
1856  *	bq_remove:
1857  *
1858  *	Removes a buffer from the free list, must be called with the
1859  *	correct qlock held.
1860  */
1861 static void
1862 bq_remove(struct bufqueue *bq, struct buf *bp)
1863 {
1864 
1865 	CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1866 	    bp, bp->b_vp, bp->b_flags);
1867 	KASSERT(bp->b_qindex != QUEUE_NONE,
1868 	    ("bq_remove: buffer %p not on a queue.", bp));
1869 	KASSERT(bufqueue(bp) == bq,
1870 	    ("bq_remove: Remove buffer %p from wrong queue.", bp));
1871 
1872 	BQ_ASSERT_LOCKED(bq);
1873 	if (bp->b_qindex != QUEUE_EMPTY) {
1874 		BUF_ASSERT_XLOCKED(bp);
1875 	}
1876 	KASSERT(bq->bq_len >= 1,
1877 	    ("queue %d underflow", bp->b_qindex));
1878 	TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1879 	bq->bq_len--;
1880 	bp->b_qindex = QUEUE_NONE;
1881 	bp->b_flags &= ~(B_REMFREE | B_REUSE);
1882 }
1883 
1884 static void
1885 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1886 {
1887 	struct buf *bp;
1888 
1889 	BQ_ASSERT_LOCKED(bq);
1890 	if (bq != bd->bd_cleanq) {
1891 		BD_LOCK(bd);
1892 		while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1893 			TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1894 			TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1895 			    b_freelist);
1896 			bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1897 		}
1898 		bd->bd_cleanq->bq_len += bq->bq_len;
1899 		bq->bq_len = 0;
1900 	}
1901 	if (bd->bd_wanted) {
1902 		bd->bd_wanted = 0;
1903 		wakeup(&bd->bd_wanted);
1904 	}
1905 	if (bq != bd->bd_cleanq)
1906 		BD_UNLOCK(bd);
1907 }
1908 
1909 static int
1910 bd_flushall(struct bufdomain *bd)
1911 {
1912 	struct bufqueue *bq;
1913 	int flushed;
1914 	int i;
1915 
1916 	if (bd->bd_lim == 0)
1917 		return (0);
1918 	flushed = 0;
1919 	for (i = 0; i <= mp_maxid; i++) {
1920 		bq = &bd->bd_subq[i];
1921 		if (bq->bq_len == 0)
1922 			continue;
1923 		BQ_LOCK(bq);
1924 		bd_flush(bd, bq);
1925 		BQ_UNLOCK(bq);
1926 		flushed++;
1927 	}
1928 
1929 	return (flushed);
1930 }
1931 
1932 static void
1933 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1934 {
1935 	struct bufdomain *bd;
1936 
1937 	if (bp->b_qindex != QUEUE_NONE)
1938 		panic("bq_insert: free buffer %p onto another queue?", bp);
1939 
1940 	bd = bufdomain(bp);
1941 	if (bp->b_flags & B_AGE) {
1942 		/* Place this buf directly on the real queue. */
1943 		if (bq->bq_index == QUEUE_CLEAN)
1944 			bq = bd->bd_cleanq;
1945 		BQ_LOCK(bq);
1946 		TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1947 	} else {
1948 		BQ_LOCK(bq);
1949 		TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1950 	}
1951 	bp->b_flags &= ~(B_AGE | B_REUSE);
1952 	bq->bq_len++;
1953 	bp->b_qindex = bq->bq_index;
1954 	bp->b_subqueue = bq->bq_subqueue;
1955 
1956 	/*
1957 	 * Unlock before we notify so that we don't wakeup a waiter that
1958 	 * fails a trylock on the buf and sleeps again.
1959 	 */
1960 	if (unlock)
1961 		BUF_UNLOCK(bp);
1962 
1963 	if (bp->b_qindex == QUEUE_CLEAN) {
1964 		/*
1965 		 * Flush the per-cpu queue and notify any waiters.
1966 		 */
1967 		if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1968 		    bq->bq_len >= bd->bd_lim))
1969 			bd_flush(bd, bq);
1970 	}
1971 	BQ_UNLOCK(bq);
1972 }
1973 
1974 /*
1975  *	bufkva_free:
1976  *
1977  *	Free the kva allocation for a buffer.
1978  *
1979  */
1980 static void
1981 bufkva_free(struct buf *bp)
1982 {
1983 
1984 #ifdef INVARIANTS
1985 	if (bp->b_kvasize == 0) {
1986 		KASSERT(bp->b_kvabase == unmapped_buf &&
1987 		    bp->b_data == unmapped_buf,
1988 		    ("Leaked KVA space on %p", bp));
1989 	} else if (buf_mapped(bp))
1990 		BUF_CHECK_MAPPED(bp);
1991 	else
1992 		BUF_CHECK_UNMAPPED(bp);
1993 #endif
1994 	if (bp->b_kvasize == 0)
1995 		return;
1996 
1997 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
1998 	counter_u64_add(bufkvaspace, -bp->b_kvasize);
1999 	counter_u64_add(buffreekvacnt, 1);
2000 	bp->b_data = bp->b_kvabase = unmapped_buf;
2001 	bp->b_kvasize = 0;
2002 }
2003 
2004 /*
2005  *	bufkva_alloc:
2006  *
2007  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
2008  */
2009 static int
2010 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2011 {
2012 	vm_offset_t addr;
2013 	int error;
2014 
2015 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2016 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2017 
2018 	bufkva_free(bp);
2019 
2020 	addr = 0;
2021 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2022 	if (error != 0) {
2023 		/*
2024 		 * Buffer map is too fragmented.  Request the caller
2025 		 * to defragment the map.
2026 		 */
2027 		return (error);
2028 	}
2029 	bp->b_kvabase = (caddr_t)addr;
2030 	bp->b_kvasize = maxsize;
2031 	counter_u64_add(bufkvaspace, bp->b_kvasize);
2032 	if ((gbflags & GB_UNMAPPED) != 0) {
2033 		bp->b_data = unmapped_buf;
2034 		BUF_CHECK_UNMAPPED(bp);
2035 	} else {
2036 		bp->b_data = bp->b_kvabase;
2037 		BUF_CHECK_MAPPED(bp);
2038 	}
2039 	return (0);
2040 }
2041 
2042 /*
2043  *	bufkva_reclaim:
2044  *
2045  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
2046  *	callback that fires to avoid returning failure.
2047  */
2048 static void
2049 bufkva_reclaim(vmem_t *vmem, int flags)
2050 {
2051 	bool done;
2052 	int q;
2053 	int i;
2054 
2055 	done = false;
2056 	for (i = 0; i < 5; i++) {
2057 		for (q = 0; q < buf_domains; q++)
2058 			if (buf_recycle(&bdomain[q], true) != 0)
2059 				done = true;
2060 		if (done)
2061 			break;
2062 	}
2063 	return;
2064 }
2065 
2066 /*
2067  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
2068  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2069  * the buffer is valid and we do not have to do anything.
2070  */
2071 static void
2072 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2073     struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2074 {
2075 	struct buf *rabp;
2076 	struct thread *td;
2077 	int i;
2078 
2079 	td = curthread;
2080 
2081 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2082 		if (inmem(vp, *rablkno))
2083 			continue;
2084 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2085 		if ((rabp->b_flags & B_CACHE) != 0) {
2086 			brelse(rabp);
2087 			continue;
2088 		}
2089 #ifdef RACCT
2090 		if (racct_enable) {
2091 			PROC_LOCK(curproc);
2092 			racct_add_buf(curproc, rabp, 0);
2093 			PROC_UNLOCK(curproc);
2094 		}
2095 #endif /* RACCT */
2096 		td->td_ru.ru_inblock++;
2097 		rabp->b_flags |= B_ASYNC;
2098 		rabp->b_flags &= ~B_INVAL;
2099 		if ((flags & GB_CKHASH) != 0) {
2100 			rabp->b_flags |= B_CKHASH;
2101 			rabp->b_ckhashcalc = ckhashfunc;
2102 		}
2103 		rabp->b_ioflags &= ~BIO_ERROR;
2104 		rabp->b_iocmd = BIO_READ;
2105 		if (rabp->b_rcred == NOCRED && cred != NOCRED)
2106 			rabp->b_rcred = crhold(cred);
2107 		vfs_busy_pages(rabp, 0);
2108 		BUF_KERNPROC(rabp);
2109 		rabp->b_iooffset = dbtob(rabp->b_blkno);
2110 		bstrategy(rabp);
2111 	}
2112 }
2113 
2114 /*
2115  * Entry point for bread() and breadn() via #defines in sys/buf.h.
2116  *
2117  * Get a buffer with the specified data.  Look in the cache first.  We
2118  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
2119  * is set, the buffer is valid and we do not have to do anything, see
2120  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2121  *
2122  * Always return a NULL buffer pointer (in bpp) when returning an error.
2123  */
2124 int
2125 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
2126     int *rabsize, int cnt, struct ucred *cred, int flags,
2127     void (*ckhashfunc)(struct buf *), struct buf **bpp)
2128 {
2129 	struct buf *bp;
2130 	struct thread *td;
2131 	int error, readwait, rv;
2132 
2133 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2134 	td = curthread;
2135 	/*
2136 	 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2137 	 * are specified.
2138 	 */
2139 	error = getblkx(vp, blkno, size, 0, 0, flags, &bp);
2140 	if (error != 0) {
2141 		*bpp = NULL;
2142 		return (error);
2143 	}
2144 	flags &= ~GB_NOSPARSE;
2145 	*bpp = bp;
2146 
2147 	/*
2148 	 * If not found in cache, do some I/O
2149 	 */
2150 	readwait = 0;
2151 	if ((bp->b_flags & B_CACHE) == 0) {
2152 #ifdef RACCT
2153 		if (racct_enable) {
2154 			PROC_LOCK(td->td_proc);
2155 			racct_add_buf(td->td_proc, bp, 0);
2156 			PROC_UNLOCK(td->td_proc);
2157 		}
2158 #endif /* RACCT */
2159 		td->td_ru.ru_inblock++;
2160 		bp->b_iocmd = BIO_READ;
2161 		bp->b_flags &= ~B_INVAL;
2162 		if ((flags & GB_CKHASH) != 0) {
2163 			bp->b_flags |= B_CKHASH;
2164 			bp->b_ckhashcalc = ckhashfunc;
2165 		}
2166 		bp->b_ioflags &= ~BIO_ERROR;
2167 		if (bp->b_rcred == NOCRED && cred != NOCRED)
2168 			bp->b_rcred = crhold(cred);
2169 		vfs_busy_pages(bp, 0);
2170 		bp->b_iooffset = dbtob(bp->b_blkno);
2171 		bstrategy(bp);
2172 		++readwait;
2173 	}
2174 
2175 	/*
2176 	 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2177 	 */
2178 	breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2179 
2180 	rv = 0;
2181 	if (readwait) {
2182 		rv = bufwait(bp);
2183 		if (rv != 0) {
2184 			brelse(bp);
2185 			*bpp = NULL;
2186 		}
2187 	}
2188 	return (rv);
2189 }
2190 
2191 /*
2192  * Write, release buffer on completion.  (Done by iodone
2193  * if async).  Do not bother writing anything if the buffer
2194  * is invalid.
2195  *
2196  * Note that we set B_CACHE here, indicating that buffer is
2197  * fully valid and thus cacheable.  This is true even of NFS
2198  * now so we set it generally.  This could be set either here
2199  * or in biodone() since the I/O is synchronous.  We put it
2200  * here.
2201  */
2202 int
2203 bufwrite(struct buf *bp)
2204 {
2205 	int oldflags;
2206 	struct vnode *vp;
2207 	long space;
2208 	int vp_md;
2209 
2210 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2211 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2212 		bp->b_flags |= B_INVAL | B_RELBUF;
2213 		bp->b_flags &= ~B_CACHE;
2214 		brelse(bp);
2215 		return (ENXIO);
2216 	}
2217 	if (bp->b_flags & B_INVAL) {
2218 		brelse(bp);
2219 		return (0);
2220 	}
2221 
2222 	if (bp->b_flags & B_BARRIER)
2223 		atomic_add_long(&barrierwrites, 1);
2224 
2225 	oldflags = bp->b_flags;
2226 
2227 	BUF_ASSERT_HELD(bp);
2228 
2229 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2230 	    ("FFS background buffer should not get here %p", bp));
2231 
2232 	vp = bp->b_vp;
2233 	if (vp)
2234 		vp_md = vp->v_vflag & VV_MD;
2235 	else
2236 		vp_md = 0;
2237 
2238 	/*
2239 	 * Mark the buffer clean.  Increment the bufobj write count
2240 	 * before bundirty() call, to prevent other thread from seeing
2241 	 * empty dirty list and zero counter for writes in progress,
2242 	 * falsely indicating that the bufobj is clean.
2243 	 */
2244 	bufobj_wref(bp->b_bufobj);
2245 	bundirty(bp);
2246 
2247 	bp->b_flags &= ~B_DONE;
2248 	bp->b_ioflags &= ~BIO_ERROR;
2249 	bp->b_flags |= B_CACHE;
2250 	bp->b_iocmd = BIO_WRITE;
2251 
2252 	vfs_busy_pages(bp, 1);
2253 
2254 	/*
2255 	 * Normal bwrites pipeline writes
2256 	 */
2257 	bp->b_runningbufspace = bp->b_bufsize;
2258 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2259 
2260 #ifdef RACCT
2261 	if (racct_enable) {
2262 		PROC_LOCK(curproc);
2263 		racct_add_buf(curproc, bp, 1);
2264 		PROC_UNLOCK(curproc);
2265 	}
2266 #endif /* RACCT */
2267 	curthread->td_ru.ru_oublock++;
2268 	if (oldflags & B_ASYNC)
2269 		BUF_KERNPROC(bp);
2270 	bp->b_iooffset = dbtob(bp->b_blkno);
2271 	buf_track(bp, __func__);
2272 	bstrategy(bp);
2273 
2274 	if ((oldflags & B_ASYNC) == 0) {
2275 		int rtval = bufwait(bp);
2276 		brelse(bp);
2277 		return (rtval);
2278 	} else if (space > hirunningspace) {
2279 		/*
2280 		 * don't allow the async write to saturate the I/O
2281 		 * system.  We will not deadlock here because
2282 		 * we are blocking waiting for I/O that is already in-progress
2283 		 * to complete. We do not block here if it is the update
2284 		 * or syncer daemon trying to clean up as that can lead
2285 		 * to deadlock.
2286 		 */
2287 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2288 			waitrunningbufspace();
2289 	}
2290 
2291 	return (0);
2292 }
2293 
2294 void
2295 bufbdflush(struct bufobj *bo, struct buf *bp)
2296 {
2297 	struct buf *nbp;
2298 
2299 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2300 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2301 		altbufferflushes++;
2302 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2303 		BO_LOCK(bo);
2304 		/*
2305 		 * Try to find a buffer to flush.
2306 		 */
2307 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2308 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2309 			    BUF_LOCK(nbp,
2310 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2311 				continue;
2312 			if (bp == nbp)
2313 				panic("bdwrite: found ourselves");
2314 			BO_UNLOCK(bo);
2315 			/* Don't countdeps with the bo lock held. */
2316 			if (buf_countdeps(nbp, 0)) {
2317 				BO_LOCK(bo);
2318 				BUF_UNLOCK(nbp);
2319 				continue;
2320 			}
2321 			if (nbp->b_flags & B_CLUSTEROK) {
2322 				vfs_bio_awrite(nbp);
2323 			} else {
2324 				bremfree(nbp);
2325 				bawrite(nbp);
2326 			}
2327 			dirtybufferflushes++;
2328 			break;
2329 		}
2330 		if (nbp == NULL)
2331 			BO_UNLOCK(bo);
2332 	}
2333 }
2334 
2335 /*
2336  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2337  * anything if the buffer is marked invalid.
2338  *
2339  * Note that since the buffer must be completely valid, we can safely
2340  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2341  * biodone() in order to prevent getblk from writing the buffer
2342  * out synchronously.
2343  */
2344 void
2345 bdwrite(struct buf *bp)
2346 {
2347 	struct thread *td = curthread;
2348 	struct vnode *vp;
2349 	struct bufobj *bo;
2350 
2351 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2352 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2353 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2354 	    ("Barrier request in delayed write %p", bp));
2355 	BUF_ASSERT_HELD(bp);
2356 
2357 	if (bp->b_flags & B_INVAL) {
2358 		brelse(bp);
2359 		return;
2360 	}
2361 
2362 	/*
2363 	 * If we have too many dirty buffers, don't create any more.
2364 	 * If we are wildly over our limit, then force a complete
2365 	 * cleanup. Otherwise, just keep the situation from getting
2366 	 * out of control. Note that we have to avoid a recursive
2367 	 * disaster and not try to clean up after our own cleanup!
2368 	 */
2369 	vp = bp->b_vp;
2370 	bo = bp->b_bufobj;
2371 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2372 		td->td_pflags |= TDP_INBDFLUSH;
2373 		BO_BDFLUSH(bo, bp);
2374 		td->td_pflags &= ~TDP_INBDFLUSH;
2375 	} else
2376 		recursiveflushes++;
2377 
2378 	bdirty(bp);
2379 	/*
2380 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2381 	 * true even of NFS now.
2382 	 */
2383 	bp->b_flags |= B_CACHE;
2384 
2385 	/*
2386 	 * This bmap keeps the system from needing to do the bmap later,
2387 	 * perhaps when the system is attempting to do a sync.  Since it
2388 	 * is likely that the indirect block -- or whatever other datastructure
2389 	 * that the filesystem needs is still in memory now, it is a good
2390 	 * thing to do this.  Note also, that if the pageout daemon is
2391 	 * requesting a sync -- there might not be enough memory to do
2392 	 * the bmap then...  So, this is important to do.
2393 	 */
2394 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2395 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2396 	}
2397 
2398 	buf_track(bp, __func__);
2399 
2400 	/*
2401 	 * Set the *dirty* buffer range based upon the VM system dirty
2402 	 * pages.
2403 	 *
2404 	 * Mark the buffer pages as clean.  We need to do this here to
2405 	 * satisfy the vnode_pager and the pageout daemon, so that it
2406 	 * thinks that the pages have been "cleaned".  Note that since
2407 	 * the pages are in a delayed write buffer -- the VFS layer
2408 	 * "will" see that the pages get written out on the next sync,
2409 	 * or perhaps the cluster will be completed.
2410 	 */
2411 	vfs_clean_pages_dirty_buf(bp);
2412 	bqrelse(bp);
2413 
2414 	/*
2415 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2416 	 * due to the softdep code.
2417 	 */
2418 }
2419 
2420 /*
2421  *	bdirty:
2422  *
2423  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2424  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2425  *	itself to properly update it in the dirty/clean lists.  We mark it
2426  *	B_DONE to ensure that any asynchronization of the buffer properly
2427  *	clears B_DONE ( else a panic will occur later ).
2428  *
2429  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2430  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2431  *	should only be called if the buffer is known-good.
2432  *
2433  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2434  *	count.
2435  *
2436  *	The buffer must be on QUEUE_NONE.
2437  */
2438 void
2439 bdirty(struct buf *bp)
2440 {
2441 
2442 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2443 	    bp, bp->b_vp, bp->b_flags);
2444 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2445 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2446 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2447 	BUF_ASSERT_HELD(bp);
2448 	bp->b_flags &= ~(B_RELBUF);
2449 	bp->b_iocmd = BIO_WRITE;
2450 
2451 	if ((bp->b_flags & B_DELWRI) == 0) {
2452 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2453 		reassignbuf(bp);
2454 		bdirtyadd(bp);
2455 	}
2456 }
2457 
2458 /*
2459  *	bundirty:
2460  *
2461  *	Clear B_DELWRI for buffer.
2462  *
2463  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2464  *	count.
2465  *
2466  *	The buffer must be on QUEUE_NONE.
2467  */
2468 
2469 void
2470 bundirty(struct buf *bp)
2471 {
2472 
2473 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2474 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2475 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2476 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2477 	BUF_ASSERT_HELD(bp);
2478 
2479 	if (bp->b_flags & B_DELWRI) {
2480 		bp->b_flags &= ~B_DELWRI;
2481 		reassignbuf(bp);
2482 		bdirtysub(bp);
2483 	}
2484 	/*
2485 	 * Since it is now being written, we can clear its deferred write flag.
2486 	 */
2487 	bp->b_flags &= ~B_DEFERRED;
2488 }
2489 
2490 /*
2491  *	bawrite:
2492  *
2493  *	Asynchronous write.  Start output on a buffer, but do not wait for
2494  *	it to complete.  The buffer is released when the output completes.
2495  *
2496  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2497  *	B_INVAL buffers.  Not us.
2498  */
2499 void
2500 bawrite(struct buf *bp)
2501 {
2502 
2503 	bp->b_flags |= B_ASYNC;
2504 	(void) bwrite(bp);
2505 }
2506 
2507 /*
2508  *	babarrierwrite:
2509  *
2510  *	Asynchronous barrier write.  Start output on a buffer, but do not
2511  *	wait for it to complete.  Place a write barrier after this write so
2512  *	that this buffer and all buffers written before it are committed to
2513  *	the disk before any buffers written after this write are committed
2514  *	to the disk.  The buffer is released when the output completes.
2515  */
2516 void
2517 babarrierwrite(struct buf *bp)
2518 {
2519 
2520 	bp->b_flags |= B_ASYNC | B_BARRIER;
2521 	(void) bwrite(bp);
2522 }
2523 
2524 /*
2525  *	bbarrierwrite:
2526  *
2527  *	Synchronous barrier write.  Start output on a buffer and wait for
2528  *	it to complete.  Place a write barrier after this write so that
2529  *	this buffer and all buffers written before it are committed to
2530  *	the disk before any buffers written after this write are committed
2531  *	to the disk.  The buffer is released when the output completes.
2532  */
2533 int
2534 bbarrierwrite(struct buf *bp)
2535 {
2536 
2537 	bp->b_flags |= B_BARRIER;
2538 	return (bwrite(bp));
2539 }
2540 
2541 /*
2542  *	bwillwrite:
2543  *
2544  *	Called prior to the locking of any vnodes when we are expecting to
2545  *	write.  We do not want to starve the buffer cache with too many
2546  *	dirty buffers so we block here.  By blocking prior to the locking
2547  *	of any vnodes we attempt to avoid the situation where a locked vnode
2548  *	prevents the various system daemons from flushing related buffers.
2549  */
2550 void
2551 bwillwrite(void)
2552 {
2553 
2554 	if (buf_dirty_count_severe()) {
2555 		mtx_lock(&bdirtylock);
2556 		while (buf_dirty_count_severe()) {
2557 			bdirtywait = 1;
2558 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2559 			    "flswai", 0);
2560 		}
2561 		mtx_unlock(&bdirtylock);
2562 	}
2563 }
2564 
2565 /*
2566  * Return true if we have too many dirty buffers.
2567  */
2568 int
2569 buf_dirty_count_severe(void)
2570 {
2571 
2572 	return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2573 }
2574 
2575 /*
2576  *	brelse:
2577  *
2578  *	Release a busy buffer and, if requested, free its resources.  The
2579  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2580  *	to be accessed later as a cache entity or reused for other purposes.
2581  */
2582 void
2583 brelse(struct buf *bp)
2584 {
2585 	struct mount *v_mnt;
2586 	int qindex;
2587 
2588 	/*
2589 	 * Many functions erroneously call brelse with a NULL bp under rare
2590 	 * error conditions. Simply return when called with a NULL bp.
2591 	 */
2592 	if (bp == NULL)
2593 		return;
2594 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2595 	    bp, bp->b_vp, bp->b_flags);
2596 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2597 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2598 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2599 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2600 
2601 	if (BUF_LOCKRECURSED(bp)) {
2602 		/*
2603 		 * Do not process, in particular, do not handle the
2604 		 * B_INVAL/B_RELBUF and do not release to free list.
2605 		 */
2606 		BUF_UNLOCK(bp);
2607 		return;
2608 	}
2609 
2610 	if (bp->b_flags & B_MANAGED) {
2611 		bqrelse(bp);
2612 		return;
2613 	}
2614 
2615 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2616 		BO_LOCK(bp->b_bufobj);
2617 		bp->b_vflags &= ~BV_BKGRDERR;
2618 		BO_UNLOCK(bp->b_bufobj);
2619 		bdirty(bp);
2620 	}
2621 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2622 	    (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2623 	    !(bp->b_flags & B_INVAL)) {
2624 		/*
2625 		 * Failed write, redirty.  All errors except ENXIO (which
2626 		 * means the device is gone) are treated as being
2627 		 * transient.
2628 		 *
2629 		 * XXX Treating EIO as transient is not correct; the
2630 		 * contract with the local storage device drivers is that
2631 		 * they will only return EIO once the I/O is no longer
2632 		 * retriable.  Network I/O also respects this through the
2633 		 * guarantees of TCP and/or the internal retries of NFS.
2634 		 * ENOMEM might be transient, but we also have no way of
2635 		 * knowing when its ok to retry/reschedule.  In general,
2636 		 * this entire case should be made obsolete through better
2637 		 * error handling/recovery and resource scheduling.
2638 		 *
2639 		 * Do this also for buffers that failed with ENXIO, but have
2640 		 * non-empty dependencies - the soft updates code might need
2641 		 * to access the buffer to untangle them.
2642 		 *
2643 		 * Must clear BIO_ERROR to prevent pages from being scrapped.
2644 		 */
2645 		bp->b_ioflags &= ~BIO_ERROR;
2646 		bdirty(bp);
2647 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2648 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2649 		/*
2650 		 * Either a failed read I/O, or we were asked to free or not
2651 		 * cache the buffer, or we failed to write to a device that's
2652 		 * no longer present.
2653 		 */
2654 		bp->b_flags |= B_INVAL;
2655 		if (!LIST_EMPTY(&bp->b_dep))
2656 			buf_deallocate(bp);
2657 		if (bp->b_flags & B_DELWRI)
2658 			bdirtysub(bp);
2659 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2660 		if ((bp->b_flags & B_VMIO) == 0) {
2661 			allocbuf(bp, 0);
2662 			if (bp->b_vp)
2663 				brelvp(bp);
2664 		}
2665 	}
2666 
2667 	/*
2668 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2669 	 * is called with B_DELWRI set, the underlying pages may wind up
2670 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2671 	 * because pages associated with a B_DELWRI bp are marked clean.
2672 	 *
2673 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2674 	 * if B_DELWRI is set.
2675 	 */
2676 	if (bp->b_flags & B_DELWRI)
2677 		bp->b_flags &= ~B_RELBUF;
2678 
2679 	/*
2680 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2681 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2682 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2683 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2684 	 *
2685 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2686 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2687 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2688 	 *
2689 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2690 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2691 	 * the commit state and we cannot afford to lose the buffer. If the
2692 	 * buffer has a background write in progress, we need to keep it
2693 	 * around to prevent it from being reconstituted and starting a second
2694 	 * background write.
2695 	 */
2696 
2697 	v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2698 
2699 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2700 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2701 	    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2702 	    vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2703 		vfs_vmio_invalidate(bp);
2704 		allocbuf(bp, 0);
2705 	}
2706 
2707 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2708 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2709 		allocbuf(bp, 0);
2710 		bp->b_flags &= ~B_NOREUSE;
2711 		if (bp->b_vp != NULL)
2712 			brelvp(bp);
2713 	}
2714 
2715 	/*
2716 	 * If the buffer has junk contents signal it and eventually
2717 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2718 	 * doesn't find it.
2719 	 */
2720 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2721 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2722 		bp->b_flags |= B_INVAL;
2723 	if (bp->b_flags & B_INVAL) {
2724 		if (bp->b_flags & B_DELWRI)
2725 			bundirty(bp);
2726 		if (bp->b_vp)
2727 			brelvp(bp);
2728 	}
2729 
2730 	buf_track(bp, __func__);
2731 
2732 	/* buffers with no memory */
2733 	if (bp->b_bufsize == 0) {
2734 		buf_free(bp);
2735 		return;
2736 	}
2737 	/* buffers with junk contents */
2738 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2739 	    (bp->b_ioflags & BIO_ERROR)) {
2740 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2741 		if (bp->b_vflags & BV_BKGRDINPROG)
2742 			panic("losing buffer 2");
2743 		qindex = QUEUE_CLEAN;
2744 		bp->b_flags |= B_AGE;
2745 	/* remaining buffers */
2746 	} else if (bp->b_flags & B_DELWRI)
2747 		qindex = QUEUE_DIRTY;
2748 	else
2749 		qindex = QUEUE_CLEAN;
2750 
2751 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2752 		panic("brelse: not dirty");
2753 
2754 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2755 	/* binsfree unlocks bp. */
2756 	binsfree(bp, qindex);
2757 }
2758 
2759 /*
2760  * Release a buffer back to the appropriate queue but do not try to free
2761  * it.  The buffer is expected to be used again soon.
2762  *
2763  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2764  * biodone() to requeue an async I/O on completion.  It is also used when
2765  * known good buffers need to be requeued but we think we may need the data
2766  * again soon.
2767  *
2768  * XXX we should be able to leave the B_RELBUF hint set on completion.
2769  */
2770 void
2771 bqrelse(struct buf *bp)
2772 {
2773 	int qindex;
2774 
2775 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2776 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2777 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2778 
2779 	qindex = QUEUE_NONE;
2780 	if (BUF_LOCKRECURSED(bp)) {
2781 		/* do not release to free list */
2782 		BUF_UNLOCK(bp);
2783 		return;
2784 	}
2785 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2786 
2787 	if (bp->b_flags & B_MANAGED) {
2788 		if (bp->b_flags & B_REMFREE)
2789 			bremfreef(bp);
2790 		goto out;
2791 	}
2792 
2793 	/* buffers with stale but valid contents */
2794 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2795 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2796 		BO_LOCK(bp->b_bufobj);
2797 		bp->b_vflags &= ~BV_BKGRDERR;
2798 		BO_UNLOCK(bp->b_bufobj);
2799 		qindex = QUEUE_DIRTY;
2800 	} else {
2801 		if ((bp->b_flags & B_DELWRI) == 0 &&
2802 		    (bp->b_xflags & BX_VNDIRTY))
2803 			panic("bqrelse: not dirty");
2804 		if ((bp->b_flags & B_NOREUSE) != 0) {
2805 			brelse(bp);
2806 			return;
2807 		}
2808 		qindex = QUEUE_CLEAN;
2809 	}
2810 	buf_track(bp, __func__);
2811 	/* binsfree unlocks bp. */
2812 	binsfree(bp, qindex);
2813 	return;
2814 
2815 out:
2816 	buf_track(bp, __func__);
2817 	/* unlock */
2818 	BUF_UNLOCK(bp);
2819 }
2820 
2821 /*
2822  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2823  * restore bogus pages.
2824  */
2825 static void
2826 vfs_vmio_iodone(struct buf *bp)
2827 {
2828 	vm_ooffset_t foff;
2829 	vm_page_t m;
2830 	vm_object_t obj;
2831 	struct vnode *vp __unused;
2832 	int i, iosize, resid;
2833 	bool bogus;
2834 
2835 	obj = bp->b_bufobj->bo_object;
2836 	KASSERT(obj->paging_in_progress >= bp->b_npages,
2837 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2838 	    obj->paging_in_progress, bp->b_npages));
2839 
2840 	vp = bp->b_vp;
2841 	KASSERT(vp->v_holdcnt > 0,
2842 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2843 	KASSERT(vp->v_object != NULL,
2844 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2845 
2846 	foff = bp->b_offset;
2847 	KASSERT(bp->b_offset != NOOFFSET,
2848 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2849 
2850 	bogus = false;
2851 	iosize = bp->b_bcount - bp->b_resid;
2852 	VM_OBJECT_WLOCK(obj);
2853 	for (i = 0; i < bp->b_npages; i++) {
2854 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2855 		if (resid > iosize)
2856 			resid = iosize;
2857 
2858 		/*
2859 		 * cleanup bogus pages, restoring the originals
2860 		 */
2861 		m = bp->b_pages[i];
2862 		if (m == bogus_page) {
2863 			bogus = true;
2864 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2865 			if (m == NULL)
2866 				panic("biodone: page disappeared!");
2867 			bp->b_pages[i] = m;
2868 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2869 			/*
2870 			 * In the write case, the valid and clean bits are
2871 			 * already changed correctly ( see bdwrite() ), so we
2872 			 * only need to do this here in the read case.
2873 			 */
2874 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2875 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2876 			    "has unexpected dirty bits", m));
2877 			vfs_page_set_valid(bp, foff, m);
2878 		}
2879 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2880 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2881 		    (intmax_t)foff, (uintmax_t)m->pindex));
2882 
2883 		vm_page_sunbusy(m);
2884 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2885 		iosize -= resid;
2886 	}
2887 	vm_object_pip_wakeupn(obj, bp->b_npages);
2888 	VM_OBJECT_WUNLOCK(obj);
2889 	if (bogus && buf_mapped(bp)) {
2890 		BUF_CHECK_MAPPED(bp);
2891 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2892 		    bp->b_pages, bp->b_npages);
2893 	}
2894 }
2895 
2896 /*
2897  * Unwire a page held by a buf and either free it or update the page queues to
2898  * reflect its recent use.
2899  */
2900 static void
2901 vfs_vmio_unwire(struct buf *bp, vm_page_t m)
2902 {
2903 	bool freed;
2904 
2905 	vm_page_lock(m);
2906 	if (vm_page_unwire_noq(m)) {
2907 		if ((bp->b_flags & B_DIRECT) != 0)
2908 			freed = vm_page_try_to_free(m);
2909 		else
2910 			freed = false;
2911 		if (!freed) {
2912 			/*
2913 			 * Use a racy check of the valid bits to determine
2914 			 * whether we can accelerate reclamation of the page.
2915 			 * The valid bits will be stable unless the page is
2916 			 * being mapped or is referenced by multiple buffers,
2917 			 * and in those cases we expect races to be rare.  At
2918 			 * worst we will either accelerate reclamation of a
2919 			 * valid page and violate LRU, or unnecessarily defer
2920 			 * reclamation of an invalid page.
2921 			 *
2922 			 * The B_NOREUSE flag marks data that is not expected to
2923 			 * be reused, so accelerate reclamation in that case
2924 			 * too.  Otherwise, maintain LRU.
2925 			 */
2926 			if (m->valid == 0 || (bp->b_flags & B_NOREUSE) != 0)
2927 				vm_page_deactivate_noreuse(m);
2928 			else if (vm_page_active(m))
2929 				vm_page_reference(m);
2930 			else
2931 				vm_page_deactivate(m);
2932 		}
2933 	}
2934 	vm_page_unlock(m);
2935 }
2936 
2937 /*
2938  * Perform page invalidation when a buffer is released.  The fully invalid
2939  * pages will be reclaimed later in vfs_vmio_truncate().
2940  */
2941 static void
2942 vfs_vmio_invalidate(struct buf *bp)
2943 {
2944 	vm_object_t obj;
2945 	vm_page_t m;
2946 	int i, resid, poffset, presid;
2947 
2948 	if (buf_mapped(bp)) {
2949 		BUF_CHECK_MAPPED(bp);
2950 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2951 	} else
2952 		BUF_CHECK_UNMAPPED(bp);
2953 	/*
2954 	 * Get the base offset and length of the buffer.  Note that
2955 	 * in the VMIO case if the buffer block size is not
2956 	 * page-aligned then b_data pointer may not be page-aligned.
2957 	 * But our b_pages[] array *IS* page aligned.
2958 	 *
2959 	 * block sizes less then DEV_BSIZE (usually 512) are not
2960 	 * supported due to the page granularity bits (m->valid,
2961 	 * m->dirty, etc...).
2962 	 *
2963 	 * See man buf(9) for more information
2964 	 */
2965 	obj = bp->b_bufobj->bo_object;
2966 	resid = bp->b_bufsize;
2967 	poffset = bp->b_offset & PAGE_MASK;
2968 	VM_OBJECT_WLOCK(obj);
2969 	for (i = 0; i < bp->b_npages; i++) {
2970 		m = bp->b_pages[i];
2971 		if (m == bogus_page)
2972 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2973 		bp->b_pages[i] = NULL;
2974 
2975 		presid = resid > (PAGE_SIZE - poffset) ?
2976 		    (PAGE_SIZE - poffset) : resid;
2977 		KASSERT(presid >= 0, ("brelse: extra page"));
2978 		while (vm_page_xbusied(m)) {
2979 			vm_page_lock(m);
2980 			VM_OBJECT_WUNLOCK(obj);
2981 			vm_page_busy_sleep(m, "mbncsh", true);
2982 			VM_OBJECT_WLOCK(obj);
2983 		}
2984 		if (pmap_page_wired_mappings(m) == 0)
2985 			vm_page_set_invalid(m, poffset, presid);
2986 		vfs_vmio_unwire(bp, m);
2987 		resid -= presid;
2988 		poffset = 0;
2989 	}
2990 	VM_OBJECT_WUNLOCK(obj);
2991 	bp->b_npages = 0;
2992 }
2993 
2994 /*
2995  * Page-granular truncation of an existing VMIO buffer.
2996  */
2997 static void
2998 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2999 {
3000 	vm_object_t obj;
3001 	vm_page_t m;
3002 	int i;
3003 
3004 	if (bp->b_npages == desiredpages)
3005 		return;
3006 
3007 	if (buf_mapped(bp)) {
3008 		BUF_CHECK_MAPPED(bp);
3009 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3010 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3011 	} else
3012 		BUF_CHECK_UNMAPPED(bp);
3013 
3014 	/*
3015 	 * The object lock is needed only if we will attempt to free pages.
3016 	 */
3017 	obj = (bp->b_flags & B_DIRECT) != 0 ? bp->b_bufobj->bo_object : NULL;
3018 	if (obj != NULL)
3019 		VM_OBJECT_WLOCK(obj);
3020 	for (i = desiredpages; i < bp->b_npages; i++) {
3021 		m = bp->b_pages[i];
3022 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3023 		bp->b_pages[i] = NULL;
3024 		vfs_vmio_unwire(bp, m);
3025 	}
3026 	if (obj != NULL)
3027 		VM_OBJECT_WUNLOCK(obj);
3028 	bp->b_npages = desiredpages;
3029 }
3030 
3031 /*
3032  * Byte granular extension of VMIO buffers.
3033  */
3034 static void
3035 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3036 {
3037 	/*
3038 	 * We are growing the buffer, possibly in a
3039 	 * byte-granular fashion.
3040 	 */
3041 	vm_object_t obj;
3042 	vm_offset_t toff;
3043 	vm_offset_t tinc;
3044 	vm_page_t m;
3045 
3046 	/*
3047 	 * Step 1, bring in the VM pages from the object, allocating
3048 	 * them if necessary.  We must clear B_CACHE if these pages
3049 	 * are not valid for the range covered by the buffer.
3050 	 */
3051 	obj = bp->b_bufobj->bo_object;
3052 	VM_OBJECT_WLOCK(obj);
3053 	if (bp->b_npages < desiredpages) {
3054 		/*
3055 		 * We must allocate system pages since blocking
3056 		 * here could interfere with paging I/O, no
3057 		 * matter which process we are.
3058 		 *
3059 		 * Only exclusive busy can be tested here.
3060 		 * Blocking on shared busy might lead to
3061 		 * deadlocks once allocbuf() is called after
3062 		 * pages are vfs_busy_pages().
3063 		 */
3064 		(void)vm_page_grab_pages(obj,
3065 		    OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3066 		    VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3067 		    VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3068 		    &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3069 		bp->b_npages = desiredpages;
3070 	}
3071 
3072 	/*
3073 	 * Step 2.  We've loaded the pages into the buffer,
3074 	 * we have to figure out if we can still have B_CACHE
3075 	 * set.  Note that B_CACHE is set according to the
3076 	 * byte-granular range ( bcount and size ), not the
3077 	 * aligned range ( newbsize ).
3078 	 *
3079 	 * The VM test is against m->valid, which is DEV_BSIZE
3080 	 * aligned.  Needless to say, the validity of the data
3081 	 * needs to also be DEV_BSIZE aligned.  Note that this
3082 	 * fails with NFS if the server or some other client
3083 	 * extends the file's EOF.  If our buffer is resized,
3084 	 * B_CACHE may remain set! XXX
3085 	 */
3086 	toff = bp->b_bcount;
3087 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3088 	while ((bp->b_flags & B_CACHE) && toff < size) {
3089 		vm_pindex_t pi;
3090 
3091 		if (tinc > (size - toff))
3092 			tinc = size - toff;
3093 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3094 		m = bp->b_pages[pi];
3095 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3096 		toff += tinc;
3097 		tinc = PAGE_SIZE;
3098 	}
3099 	VM_OBJECT_WUNLOCK(obj);
3100 
3101 	/*
3102 	 * Step 3, fixup the KVA pmap.
3103 	 */
3104 	if (buf_mapped(bp))
3105 		bpmap_qenter(bp);
3106 	else
3107 		BUF_CHECK_UNMAPPED(bp);
3108 }
3109 
3110 /*
3111  * Check to see if a block at a particular lbn is available for a clustered
3112  * write.
3113  */
3114 static int
3115 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3116 {
3117 	struct buf *bpa;
3118 	int match;
3119 
3120 	match = 0;
3121 
3122 	/* If the buf isn't in core skip it */
3123 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3124 		return (0);
3125 
3126 	/* If the buf is busy we don't want to wait for it */
3127 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3128 		return (0);
3129 
3130 	/* Only cluster with valid clusterable delayed write buffers */
3131 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3132 	    (B_DELWRI | B_CLUSTEROK))
3133 		goto done;
3134 
3135 	if (bpa->b_bufsize != size)
3136 		goto done;
3137 
3138 	/*
3139 	 * Check to see if it is in the expected place on disk and that the
3140 	 * block has been mapped.
3141 	 */
3142 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3143 		match = 1;
3144 done:
3145 	BUF_UNLOCK(bpa);
3146 	return (match);
3147 }
3148 
3149 /*
3150  *	vfs_bio_awrite:
3151  *
3152  *	Implement clustered async writes for clearing out B_DELWRI buffers.
3153  *	This is much better then the old way of writing only one buffer at
3154  *	a time.  Note that we may not be presented with the buffers in the
3155  *	correct order, so we search for the cluster in both directions.
3156  */
3157 int
3158 vfs_bio_awrite(struct buf *bp)
3159 {
3160 	struct bufobj *bo;
3161 	int i;
3162 	int j;
3163 	daddr_t lblkno = bp->b_lblkno;
3164 	struct vnode *vp = bp->b_vp;
3165 	int ncl;
3166 	int nwritten;
3167 	int size;
3168 	int maxcl;
3169 	int gbflags;
3170 
3171 	bo = &vp->v_bufobj;
3172 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3173 	/*
3174 	 * right now we support clustered writing only to regular files.  If
3175 	 * we find a clusterable block we could be in the middle of a cluster
3176 	 * rather then at the beginning.
3177 	 */
3178 	if ((vp->v_type == VREG) &&
3179 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
3180 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3181 
3182 		size = vp->v_mount->mnt_stat.f_iosize;
3183 		maxcl = MAXPHYS / size;
3184 
3185 		BO_RLOCK(bo);
3186 		for (i = 1; i < maxcl; i++)
3187 			if (vfs_bio_clcheck(vp, size, lblkno + i,
3188 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3189 				break;
3190 
3191 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3192 			if (vfs_bio_clcheck(vp, size, lblkno - j,
3193 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3194 				break;
3195 		BO_RUNLOCK(bo);
3196 		--j;
3197 		ncl = i + j;
3198 		/*
3199 		 * this is a possible cluster write
3200 		 */
3201 		if (ncl != 1) {
3202 			BUF_UNLOCK(bp);
3203 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3204 			    gbflags);
3205 			return (nwritten);
3206 		}
3207 	}
3208 	bremfree(bp);
3209 	bp->b_flags |= B_ASYNC;
3210 	/*
3211 	 * default (old) behavior, writing out only one block
3212 	 *
3213 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
3214 	 */
3215 	nwritten = bp->b_bufsize;
3216 	(void) bwrite(bp);
3217 
3218 	return (nwritten);
3219 }
3220 
3221 /*
3222  *	getnewbuf_kva:
3223  *
3224  *	Allocate KVA for an empty buf header according to gbflags.
3225  */
3226 static int
3227 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3228 {
3229 
3230 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3231 		/*
3232 		 * In order to keep fragmentation sane we only allocate kva
3233 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
3234 		 */
3235 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3236 
3237 		if (maxsize != bp->b_kvasize &&
3238 		    bufkva_alloc(bp, maxsize, gbflags))
3239 			return (ENOSPC);
3240 	}
3241 	return (0);
3242 }
3243 
3244 /*
3245  *	getnewbuf:
3246  *
3247  *	Find and initialize a new buffer header, freeing up existing buffers
3248  *	in the bufqueues as necessary.  The new buffer is returned locked.
3249  *
3250  *	We block if:
3251  *		We have insufficient buffer headers
3252  *		We have insufficient buffer space
3253  *		buffer_arena is too fragmented ( space reservation fails )
3254  *		If we have to flush dirty buffers ( but we try to avoid this )
3255  *
3256  *	The caller is responsible for releasing the reserved bufspace after
3257  *	allocbuf() is called.
3258  */
3259 static struct buf *
3260 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3261 {
3262 	struct bufdomain *bd;
3263 	struct buf *bp;
3264 	bool metadata, reserved;
3265 
3266 	bp = NULL;
3267 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3268 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3269 	if (!unmapped_buf_allowed)
3270 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3271 
3272 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3273 	    vp->v_type == VCHR)
3274 		metadata = true;
3275 	else
3276 		metadata = false;
3277 	if (vp == NULL)
3278 		bd = &bdomain[0];
3279 	else
3280 		bd = &bdomain[vp->v_bufobj.bo_domain];
3281 
3282 	counter_u64_add(getnewbufcalls, 1);
3283 	reserved = false;
3284 	do {
3285 		if (reserved == false &&
3286 		    bufspace_reserve(bd, maxsize, metadata) != 0) {
3287 			counter_u64_add(getnewbufrestarts, 1);
3288 			continue;
3289 		}
3290 		reserved = true;
3291 		if ((bp = buf_alloc(bd)) == NULL) {
3292 			counter_u64_add(getnewbufrestarts, 1);
3293 			continue;
3294 		}
3295 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3296 			return (bp);
3297 		break;
3298 	} while (buf_recycle(bd, false) == 0);
3299 
3300 	if (reserved)
3301 		bufspace_release(bd, maxsize);
3302 	if (bp != NULL) {
3303 		bp->b_flags |= B_INVAL;
3304 		brelse(bp);
3305 	}
3306 	bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3307 
3308 	return (NULL);
3309 }
3310 
3311 /*
3312  *	buf_daemon:
3313  *
3314  *	buffer flushing daemon.  Buffers are normally flushed by the
3315  *	update daemon but if it cannot keep up this process starts to
3316  *	take the load in an attempt to prevent getnewbuf() from blocking.
3317  */
3318 static struct kproc_desc buf_kp = {
3319 	"bufdaemon",
3320 	buf_daemon,
3321 	&bufdaemonproc
3322 };
3323 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3324 
3325 static int
3326 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3327 {
3328 	int flushed;
3329 
3330 	flushed = flushbufqueues(vp, bd, target, 0);
3331 	if (flushed == 0) {
3332 		/*
3333 		 * Could not find any buffers without rollback
3334 		 * dependencies, so just write the first one
3335 		 * in the hopes of eventually making progress.
3336 		 */
3337 		if (vp != NULL && target > 2)
3338 			target /= 2;
3339 		flushbufqueues(vp, bd, target, 1);
3340 	}
3341 	return (flushed);
3342 }
3343 
3344 static void
3345 buf_daemon()
3346 {
3347 	struct bufdomain *bd;
3348 	int speedupreq;
3349 	int lodirty;
3350 	int i;
3351 
3352 	/*
3353 	 * This process needs to be suspended prior to shutdown sync.
3354 	 */
3355 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3356 	    SHUTDOWN_PRI_LAST + 100);
3357 
3358 	/*
3359 	 * Start the buf clean daemons as children threads.
3360 	 */
3361 	for (i = 0 ; i < buf_domains; i++) {
3362 		int error;
3363 
3364 		error = kthread_add((void (*)(void *))bufspace_daemon,
3365 		    &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3366 		if (error)
3367 			panic("error %d spawning bufspace daemon", error);
3368 	}
3369 
3370 	/*
3371 	 * This process is allowed to take the buffer cache to the limit
3372 	 */
3373 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3374 	mtx_lock(&bdlock);
3375 	for (;;) {
3376 		bd_request = 0;
3377 		mtx_unlock(&bdlock);
3378 
3379 		kthread_suspend_check();
3380 
3381 		/*
3382 		 * Save speedupreq for this pass and reset to capture new
3383 		 * requests.
3384 		 */
3385 		speedupreq = bd_speedupreq;
3386 		bd_speedupreq = 0;
3387 
3388 		/*
3389 		 * Flush each domain sequentially according to its level and
3390 		 * the speedup request.
3391 		 */
3392 		for (i = 0; i < buf_domains; i++) {
3393 			bd = &bdomain[i];
3394 			if (speedupreq)
3395 				lodirty = bd->bd_numdirtybuffers / 2;
3396 			else
3397 				lodirty = bd->bd_lodirtybuffers;
3398 			while (bd->bd_numdirtybuffers > lodirty) {
3399 				if (buf_flush(NULL, bd,
3400 				    bd->bd_numdirtybuffers - lodirty) == 0)
3401 					break;
3402 				kern_yield(PRI_USER);
3403 			}
3404 		}
3405 
3406 		/*
3407 		 * Only clear bd_request if we have reached our low water
3408 		 * mark.  The buf_daemon normally waits 1 second and
3409 		 * then incrementally flushes any dirty buffers that have
3410 		 * built up, within reason.
3411 		 *
3412 		 * If we were unable to hit our low water mark and couldn't
3413 		 * find any flushable buffers, we sleep for a short period
3414 		 * to avoid endless loops on unlockable buffers.
3415 		 */
3416 		mtx_lock(&bdlock);
3417 		if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3418 			/*
3419 			 * We reached our low water mark, reset the
3420 			 * request and sleep until we are needed again.
3421 			 * The sleep is just so the suspend code works.
3422 			 */
3423 			bd_request = 0;
3424 			/*
3425 			 * Do an extra wakeup in case dirty threshold
3426 			 * changed via sysctl and the explicit transition
3427 			 * out of shortfall was missed.
3428 			 */
3429 			bdirtywakeup();
3430 			if (runningbufspace <= lorunningspace)
3431 				runningwakeup();
3432 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3433 		} else {
3434 			/*
3435 			 * We couldn't find any flushable dirty buffers but
3436 			 * still have too many dirty buffers, we
3437 			 * have to sleep and try again.  (rare)
3438 			 */
3439 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3440 		}
3441 	}
3442 }
3443 
3444 /*
3445  *	flushbufqueues:
3446  *
3447  *	Try to flush a buffer in the dirty queue.  We must be careful to
3448  *	free up B_INVAL buffers instead of write them, which NFS is
3449  *	particularly sensitive to.
3450  */
3451 static int flushwithdeps = 0;
3452 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
3453     0, "Number of buffers flushed with dependecies that require rollbacks");
3454 
3455 static int
3456 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3457     int flushdeps)
3458 {
3459 	struct bufqueue *bq;
3460 	struct buf *sentinel;
3461 	struct vnode *vp;
3462 	struct mount *mp;
3463 	struct buf *bp;
3464 	int hasdeps;
3465 	int flushed;
3466 	int error;
3467 	bool unlock;
3468 
3469 	flushed = 0;
3470 	bq = &bd->bd_dirtyq;
3471 	bp = NULL;
3472 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3473 	sentinel->b_qindex = QUEUE_SENTINEL;
3474 	BQ_LOCK(bq);
3475 	TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3476 	BQ_UNLOCK(bq);
3477 	while (flushed != target) {
3478 		maybe_yield();
3479 		BQ_LOCK(bq);
3480 		bp = TAILQ_NEXT(sentinel, b_freelist);
3481 		if (bp != NULL) {
3482 			TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3483 			TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3484 			    b_freelist);
3485 		} else {
3486 			BQ_UNLOCK(bq);
3487 			break;
3488 		}
3489 		/*
3490 		 * Skip sentinels inserted by other invocations of the
3491 		 * flushbufqueues(), taking care to not reorder them.
3492 		 *
3493 		 * Only flush the buffers that belong to the
3494 		 * vnode locked by the curthread.
3495 		 */
3496 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3497 		    bp->b_vp != lvp)) {
3498 			BQ_UNLOCK(bq);
3499 			continue;
3500 		}
3501 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3502 		BQ_UNLOCK(bq);
3503 		if (error != 0)
3504 			continue;
3505 
3506 		/*
3507 		 * BKGRDINPROG can only be set with the buf and bufobj
3508 		 * locks both held.  We tolerate a race to clear it here.
3509 		 */
3510 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3511 		    (bp->b_flags & B_DELWRI) == 0) {
3512 			BUF_UNLOCK(bp);
3513 			continue;
3514 		}
3515 		if (bp->b_flags & B_INVAL) {
3516 			bremfreef(bp);
3517 			brelse(bp);
3518 			flushed++;
3519 			continue;
3520 		}
3521 
3522 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3523 			if (flushdeps == 0) {
3524 				BUF_UNLOCK(bp);
3525 				continue;
3526 			}
3527 			hasdeps = 1;
3528 		} else
3529 			hasdeps = 0;
3530 		/*
3531 		 * We must hold the lock on a vnode before writing
3532 		 * one of its buffers. Otherwise we may confuse, or
3533 		 * in the case of a snapshot vnode, deadlock the
3534 		 * system.
3535 		 *
3536 		 * The lock order here is the reverse of the normal
3537 		 * of vnode followed by buf lock.  This is ok because
3538 		 * the NOWAIT will prevent deadlock.
3539 		 */
3540 		vp = bp->b_vp;
3541 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3542 			BUF_UNLOCK(bp);
3543 			continue;
3544 		}
3545 		if (lvp == NULL) {
3546 			unlock = true;
3547 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3548 		} else {
3549 			ASSERT_VOP_LOCKED(vp, "getbuf");
3550 			unlock = false;
3551 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3552 			    vn_lock(vp, LK_TRYUPGRADE);
3553 		}
3554 		if (error == 0) {
3555 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3556 			    bp, bp->b_vp, bp->b_flags);
3557 			if (curproc == bufdaemonproc) {
3558 				vfs_bio_awrite(bp);
3559 			} else {
3560 				bremfree(bp);
3561 				bwrite(bp);
3562 				counter_u64_add(notbufdflushes, 1);
3563 			}
3564 			vn_finished_write(mp);
3565 			if (unlock)
3566 				VOP_UNLOCK(vp, 0);
3567 			flushwithdeps += hasdeps;
3568 			flushed++;
3569 
3570 			/*
3571 			 * Sleeping on runningbufspace while holding
3572 			 * vnode lock leads to deadlock.
3573 			 */
3574 			if (curproc == bufdaemonproc &&
3575 			    runningbufspace > hirunningspace)
3576 				waitrunningbufspace();
3577 			continue;
3578 		}
3579 		vn_finished_write(mp);
3580 		BUF_UNLOCK(bp);
3581 	}
3582 	BQ_LOCK(bq);
3583 	TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3584 	BQ_UNLOCK(bq);
3585 	free(sentinel, M_TEMP);
3586 	return (flushed);
3587 }
3588 
3589 /*
3590  * Check to see if a block is currently memory resident.
3591  */
3592 struct buf *
3593 incore(struct bufobj *bo, daddr_t blkno)
3594 {
3595 	struct buf *bp;
3596 
3597 	BO_RLOCK(bo);
3598 	bp = gbincore(bo, blkno);
3599 	BO_RUNLOCK(bo);
3600 	return (bp);
3601 }
3602 
3603 /*
3604  * Returns true if no I/O is needed to access the
3605  * associated VM object.  This is like incore except
3606  * it also hunts around in the VM system for the data.
3607  */
3608 
3609 static int
3610 inmem(struct vnode * vp, daddr_t blkno)
3611 {
3612 	vm_object_t obj;
3613 	vm_offset_t toff, tinc, size;
3614 	vm_page_t m;
3615 	vm_ooffset_t off;
3616 
3617 	ASSERT_VOP_LOCKED(vp, "inmem");
3618 
3619 	if (incore(&vp->v_bufobj, blkno))
3620 		return 1;
3621 	if (vp->v_mount == NULL)
3622 		return 0;
3623 	obj = vp->v_object;
3624 	if (obj == NULL)
3625 		return (0);
3626 
3627 	size = PAGE_SIZE;
3628 	if (size > vp->v_mount->mnt_stat.f_iosize)
3629 		size = vp->v_mount->mnt_stat.f_iosize;
3630 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3631 
3632 	VM_OBJECT_RLOCK(obj);
3633 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3634 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3635 		if (!m)
3636 			goto notinmem;
3637 		tinc = size;
3638 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3639 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3640 		if (vm_page_is_valid(m,
3641 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3642 			goto notinmem;
3643 	}
3644 	VM_OBJECT_RUNLOCK(obj);
3645 	return 1;
3646 
3647 notinmem:
3648 	VM_OBJECT_RUNLOCK(obj);
3649 	return (0);
3650 }
3651 
3652 /*
3653  * Set the dirty range for a buffer based on the status of the dirty
3654  * bits in the pages comprising the buffer.  The range is limited
3655  * to the size of the buffer.
3656  *
3657  * Tell the VM system that the pages associated with this buffer
3658  * are clean.  This is used for delayed writes where the data is
3659  * going to go to disk eventually without additional VM intevention.
3660  *
3661  * Note that while we only really need to clean through to b_bcount, we
3662  * just go ahead and clean through to b_bufsize.
3663  */
3664 static void
3665 vfs_clean_pages_dirty_buf(struct buf *bp)
3666 {
3667 	vm_ooffset_t foff, noff, eoff;
3668 	vm_page_t m;
3669 	int i;
3670 
3671 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3672 		return;
3673 
3674 	foff = bp->b_offset;
3675 	KASSERT(bp->b_offset != NOOFFSET,
3676 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3677 
3678 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3679 	vfs_drain_busy_pages(bp);
3680 	vfs_setdirty_locked_object(bp);
3681 	for (i = 0; i < bp->b_npages; i++) {
3682 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3683 		eoff = noff;
3684 		if (eoff > bp->b_offset + bp->b_bufsize)
3685 			eoff = bp->b_offset + bp->b_bufsize;
3686 		m = bp->b_pages[i];
3687 		vfs_page_set_validclean(bp, foff, m);
3688 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3689 		foff = noff;
3690 	}
3691 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3692 }
3693 
3694 static void
3695 vfs_setdirty_locked_object(struct buf *bp)
3696 {
3697 	vm_object_t object;
3698 	int i;
3699 
3700 	object = bp->b_bufobj->bo_object;
3701 	VM_OBJECT_ASSERT_WLOCKED(object);
3702 
3703 	/*
3704 	 * We qualify the scan for modified pages on whether the
3705 	 * object has been flushed yet.
3706 	 */
3707 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3708 		vm_offset_t boffset;
3709 		vm_offset_t eoffset;
3710 
3711 		/*
3712 		 * test the pages to see if they have been modified directly
3713 		 * by users through the VM system.
3714 		 */
3715 		for (i = 0; i < bp->b_npages; i++)
3716 			vm_page_test_dirty(bp->b_pages[i]);
3717 
3718 		/*
3719 		 * Calculate the encompassing dirty range, boffset and eoffset,
3720 		 * (eoffset - boffset) bytes.
3721 		 */
3722 
3723 		for (i = 0; i < bp->b_npages; i++) {
3724 			if (bp->b_pages[i]->dirty)
3725 				break;
3726 		}
3727 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3728 
3729 		for (i = bp->b_npages - 1; i >= 0; --i) {
3730 			if (bp->b_pages[i]->dirty) {
3731 				break;
3732 			}
3733 		}
3734 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3735 
3736 		/*
3737 		 * Fit it to the buffer.
3738 		 */
3739 
3740 		if (eoffset > bp->b_bcount)
3741 			eoffset = bp->b_bcount;
3742 
3743 		/*
3744 		 * If we have a good dirty range, merge with the existing
3745 		 * dirty range.
3746 		 */
3747 
3748 		if (boffset < eoffset) {
3749 			if (bp->b_dirtyoff > boffset)
3750 				bp->b_dirtyoff = boffset;
3751 			if (bp->b_dirtyend < eoffset)
3752 				bp->b_dirtyend = eoffset;
3753 		}
3754 	}
3755 }
3756 
3757 /*
3758  * Allocate the KVA mapping for an existing buffer.
3759  * If an unmapped buffer is provided but a mapped buffer is requested, take
3760  * also care to properly setup mappings between pages and KVA.
3761  */
3762 static void
3763 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3764 {
3765 	int bsize, maxsize, need_mapping, need_kva;
3766 	off_t offset;
3767 
3768 	need_mapping = bp->b_data == unmapped_buf &&
3769 	    (gbflags & GB_UNMAPPED) == 0;
3770 	need_kva = bp->b_kvabase == unmapped_buf &&
3771 	    bp->b_data == unmapped_buf &&
3772 	    (gbflags & GB_KVAALLOC) != 0;
3773 	if (!need_mapping && !need_kva)
3774 		return;
3775 
3776 	BUF_CHECK_UNMAPPED(bp);
3777 
3778 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3779 		/*
3780 		 * Buffer is not mapped, but the KVA was already
3781 		 * reserved at the time of the instantiation.  Use the
3782 		 * allocated space.
3783 		 */
3784 		goto has_addr;
3785 	}
3786 
3787 	/*
3788 	 * Calculate the amount of the address space we would reserve
3789 	 * if the buffer was mapped.
3790 	 */
3791 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3792 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3793 	offset = blkno * bsize;
3794 	maxsize = size + (offset & PAGE_MASK);
3795 	maxsize = imax(maxsize, bsize);
3796 
3797 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3798 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3799 			/*
3800 			 * XXXKIB: defragmentation cannot
3801 			 * succeed, not sure what else to do.
3802 			 */
3803 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3804 		}
3805 		counter_u64_add(mappingrestarts, 1);
3806 		bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3807 	}
3808 has_addr:
3809 	if (need_mapping) {
3810 		/* b_offset is handled by bpmap_qenter. */
3811 		bp->b_data = bp->b_kvabase;
3812 		BUF_CHECK_MAPPED(bp);
3813 		bpmap_qenter(bp);
3814 	}
3815 }
3816 
3817 struct buf *
3818 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3819     int flags)
3820 {
3821 	struct buf *bp;
3822 	int error;
3823 
3824 	error = getblkx(vp, blkno, size, slpflag, slptimeo, flags, &bp);
3825 	if (error != 0)
3826 		return (NULL);
3827 	return (bp);
3828 }
3829 
3830 /*
3831  *	getblkx:
3832  *
3833  *	Get a block given a specified block and offset into a file/device.
3834  *	The buffers B_DONE bit will be cleared on return, making it almost
3835  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3836  *	return.  The caller should clear B_INVAL prior to initiating a
3837  *	READ.
3838  *
3839  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3840  *	an existing buffer.
3841  *
3842  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3843  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3844  *	and then cleared based on the backing VM.  If the previous buffer is
3845  *	non-0-sized but invalid, B_CACHE will be cleared.
3846  *
3847  *	If getblk() must create a new buffer, the new buffer is returned with
3848  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3849  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3850  *	backing VM.
3851  *
3852  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3853  *	B_CACHE bit is clear.
3854  *
3855  *	What this means, basically, is that the caller should use B_CACHE to
3856  *	determine whether the buffer is fully valid or not and should clear
3857  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3858  *	the buffer by loading its data area with something, the caller needs
3859  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3860  *	the caller should set B_CACHE ( as an optimization ), else the caller
3861  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3862  *	a write attempt or if it was a successful read.  If the caller
3863  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3864  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3865  */
3866 int
3867 getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3868     int flags, struct buf **bpp)
3869 {
3870 	struct buf *bp;
3871 	struct bufobj *bo;
3872 	daddr_t d_blkno;
3873 	int bsize, error, maxsize, vmio;
3874 	off_t offset;
3875 
3876 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3877 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3878 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3879 	ASSERT_VOP_LOCKED(vp, "getblk");
3880 	if (size > maxbcachebuf)
3881 		panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3882 		    maxbcachebuf);
3883 	if (!unmapped_buf_allowed)
3884 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3885 
3886 	bo = &vp->v_bufobj;
3887 	d_blkno = blkno;
3888 loop:
3889 	BO_RLOCK(bo);
3890 	bp = gbincore(bo, blkno);
3891 	if (bp != NULL) {
3892 		int lockflags;
3893 		/*
3894 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3895 		 * it must be on a queue.
3896 		 */
3897 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3898 
3899 		if ((flags & GB_LOCK_NOWAIT) != 0)
3900 			lockflags |= LK_NOWAIT;
3901 
3902 		error = BUF_TIMELOCK(bp, lockflags,
3903 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3904 
3905 		/*
3906 		 * If we slept and got the lock we have to restart in case
3907 		 * the buffer changed identities.
3908 		 */
3909 		if (error == ENOLCK)
3910 			goto loop;
3911 		/* We timed out or were interrupted. */
3912 		else if (error != 0)
3913 			return (error);
3914 		/* If recursed, assume caller knows the rules. */
3915 		else if (BUF_LOCKRECURSED(bp))
3916 			goto end;
3917 
3918 		/*
3919 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3920 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3921 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3922 		 * backing VM cache.
3923 		 */
3924 		if (bp->b_flags & B_INVAL)
3925 			bp->b_flags &= ~B_CACHE;
3926 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3927 			bp->b_flags |= B_CACHE;
3928 		if (bp->b_flags & B_MANAGED)
3929 			MPASS(bp->b_qindex == QUEUE_NONE);
3930 		else
3931 			bremfree(bp);
3932 
3933 		/*
3934 		 * check for size inconsistencies for non-VMIO case.
3935 		 */
3936 		if (bp->b_bcount != size) {
3937 			if ((bp->b_flags & B_VMIO) == 0 ||
3938 			    (size > bp->b_kvasize)) {
3939 				if (bp->b_flags & B_DELWRI) {
3940 					bp->b_flags |= B_NOCACHE;
3941 					bwrite(bp);
3942 				} else {
3943 					if (LIST_EMPTY(&bp->b_dep)) {
3944 						bp->b_flags |= B_RELBUF;
3945 						brelse(bp);
3946 					} else {
3947 						bp->b_flags |= B_NOCACHE;
3948 						bwrite(bp);
3949 					}
3950 				}
3951 				goto loop;
3952 			}
3953 		}
3954 
3955 		/*
3956 		 * Handle the case of unmapped buffer which should
3957 		 * become mapped, or the buffer for which KVA
3958 		 * reservation is requested.
3959 		 */
3960 		bp_unmapped_get_kva(bp, blkno, size, flags);
3961 
3962 		/*
3963 		 * If the size is inconsistent in the VMIO case, we can resize
3964 		 * the buffer.  This might lead to B_CACHE getting set or
3965 		 * cleared.  If the size has not changed, B_CACHE remains
3966 		 * unchanged from its previous state.
3967 		 */
3968 		allocbuf(bp, size);
3969 
3970 		KASSERT(bp->b_offset != NOOFFSET,
3971 		    ("getblk: no buffer offset"));
3972 
3973 		/*
3974 		 * A buffer with B_DELWRI set and B_CACHE clear must
3975 		 * be committed before we can return the buffer in
3976 		 * order to prevent the caller from issuing a read
3977 		 * ( due to B_CACHE not being set ) and overwriting
3978 		 * it.
3979 		 *
3980 		 * Most callers, including NFS and FFS, need this to
3981 		 * operate properly either because they assume they
3982 		 * can issue a read if B_CACHE is not set, or because
3983 		 * ( for example ) an uncached B_DELWRI might loop due
3984 		 * to softupdates re-dirtying the buffer.  In the latter
3985 		 * case, B_CACHE is set after the first write completes,
3986 		 * preventing further loops.
3987 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3988 		 * above while extending the buffer, we cannot allow the
3989 		 * buffer to remain with B_CACHE set after the write
3990 		 * completes or it will represent a corrupt state.  To
3991 		 * deal with this we set B_NOCACHE to scrap the buffer
3992 		 * after the write.
3993 		 *
3994 		 * We might be able to do something fancy, like setting
3995 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3996 		 * so the below call doesn't set B_CACHE, but that gets real
3997 		 * confusing.  This is much easier.
3998 		 */
3999 
4000 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4001 			bp->b_flags |= B_NOCACHE;
4002 			bwrite(bp);
4003 			goto loop;
4004 		}
4005 		bp->b_flags &= ~B_DONE;
4006 	} else {
4007 		/*
4008 		 * Buffer is not in-core, create new buffer.  The buffer
4009 		 * returned by getnewbuf() is locked.  Note that the returned
4010 		 * buffer is also considered valid (not marked B_INVAL).
4011 		 */
4012 		BO_RUNLOCK(bo);
4013 		/*
4014 		 * If the user does not want us to create the buffer, bail out
4015 		 * here.
4016 		 */
4017 		if (flags & GB_NOCREAT)
4018 			return (EEXIST);
4019 
4020 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
4021 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4022 		offset = blkno * bsize;
4023 		vmio = vp->v_object != NULL;
4024 		if (vmio) {
4025 			maxsize = size + (offset & PAGE_MASK);
4026 		} else {
4027 			maxsize = size;
4028 			/* Do not allow non-VMIO notmapped buffers. */
4029 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4030 		}
4031 		maxsize = imax(maxsize, bsize);
4032 		if ((flags & GB_NOSPARSE) != 0 && vmio &&
4033 		    !vn_isdisk(vp, NULL)) {
4034 			error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4035 			KASSERT(error != EOPNOTSUPP,
4036 			    ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4037 			    vp));
4038 			if (error != 0)
4039 				return (error);
4040 			if (d_blkno == -1)
4041 				return (EJUSTRETURN);
4042 		}
4043 
4044 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4045 		if (bp == NULL) {
4046 			if (slpflag || slptimeo)
4047 				return (ETIMEDOUT);
4048 			/*
4049 			 * XXX This is here until the sleep path is diagnosed
4050 			 * enough to work under very low memory conditions.
4051 			 *
4052 			 * There's an issue on low memory, 4BSD+non-preempt
4053 			 * systems (eg MIPS routers with 32MB RAM) where buffer
4054 			 * exhaustion occurs without sleeping for buffer
4055 			 * reclaimation.  This just sticks in a loop and
4056 			 * constantly attempts to allocate a buffer, which
4057 			 * hits exhaustion and tries to wakeup bufdaemon.
4058 			 * This never happens because we never yield.
4059 			 *
4060 			 * The real solution is to identify and fix these cases
4061 			 * so we aren't effectively busy-waiting in a loop
4062 			 * until the reclaimation path has cycles to run.
4063 			 */
4064 			kern_yield(PRI_USER);
4065 			goto loop;
4066 		}
4067 
4068 		/*
4069 		 * This code is used to make sure that a buffer is not
4070 		 * created while the getnewbuf routine is blocked.
4071 		 * This can be a problem whether the vnode is locked or not.
4072 		 * If the buffer is created out from under us, we have to
4073 		 * throw away the one we just created.
4074 		 *
4075 		 * Note: this must occur before we associate the buffer
4076 		 * with the vp especially considering limitations in
4077 		 * the splay tree implementation when dealing with duplicate
4078 		 * lblkno's.
4079 		 */
4080 		BO_LOCK(bo);
4081 		if (gbincore(bo, blkno)) {
4082 			BO_UNLOCK(bo);
4083 			bp->b_flags |= B_INVAL;
4084 			bufspace_release(bufdomain(bp), maxsize);
4085 			brelse(bp);
4086 			goto loop;
4087 		}
4088 
4089 		/*
4090 		 * Insert the buffer into the hash, so that it can
4091 		 * be found by incore.
4092 		 */
4093 		bp->b_lblkno = blkno;
4094 		bp->b_blkno = d_blkno;
4095 		bp->b_offset = offset;
4096 		bgetvp(vp, bp);
4097 		BO_UNLOCK(bo);
4098 
4099 		/*
4100 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
4101 		 * buffer size starts out as 0, B_CACHE will be set by
4102 		 * allocbuf() for the VMIO case prior to it testing the
4103 		 * backing store for validity.
4104 		 */
4105 
4106 		if (vmio) {
4107 			bp->b_flags |= B_VMIO;
4108 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4109 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4110 			    bp, vp->v_object, bp->b_bufobj->bo_object));
4111 		} else {
4112 			bp->b_flags &= ~B_VMIO;
4113 			KASSERT(bp->b_bufobj->bo_object == NULL,
4114 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
4115 			    bp, bp->b_bufobj->bo_object));
4116 			BUF_CHECK_MAPPED(bp);
4117 		}
4118 
4119 		allocbuf(bp, size);
4120 		bufspace_release(bufdomain(bp), maxsize);
4121 		bp->b_flags &= ~B_DONE;
4122 	}
4123 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4124 	BUF_ASSERT_HELD(bp);
4125 end:
4126 	buf_track(bp, __func__);
4127 	KASSERT(bp->b_bufobj == bo,
4128 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4129 	*bpp = bp;
4130 	return (0);
4131 }
4132 
4133 /*
4134  * Get an empty, disassociated buffer of given size.  The buffer is initially
4135  * set to B_INVAL.
4136  */
4137 struct buf *
4138 geteblk(int size, int flags)
4139 {
4140 	struct buf *bp;
4141 	int maxsize;
4142 
4143 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
4144 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4145 		if ((flags & GB_NOWAIT_BD) &&
4146 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
4147 			return (NULL);
4148 	}
4149 	allocbuf(bp, size);
4150 	bufspace_release(bufdomain(bp), maxsize);
4151 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
4152 	BUF_ASSERT_HELD(bp);
4153 	return (bp);
4154 }
4155 
4156 /*
4157  * Truncate the backing store for a non-vmio buffer.
4158  */
4159 static void
4160 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4161 {
4162 
4163 	if (bp->b_flags & B_MALLOC) {
4164 		/*
4165 		 * malloced buffers are not shrunk
4166 		 */
4167 		if (newbsize == 0) {
4168 			bufmallocadjust(bp, 0);
4169 			free(bp->b_data, M_BIOBUF);
4170 			bp->b_data = bp->b_kvabase;
4171 			bp->b_flags &= ~B_MALLOC;
4172 		}
4173 		return;
4174 	}
4175 	vm_hold_free_pages(bp, newbsize);
4176 	bufspace_adjust(bp, newbsize);
4177 }
4178 
4179 /*
4180  * Extend the backing for a non-VMIO buffer.
4181  */
4182 static void
4183 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4184 {
4185 	caddr_t origbuf;
4186 	int origbufsize;
4187 
4188 	/*
4189 	 * We only use malloced memory on the first allocation.
4190 	 * and revert to page-allocated memory when the buffer
4191 	 * grows.
4192 	 *
4193 	 * There is a potential smp race here that could lead
4194 	 * to bufmallocspace slightly passing the max.  It
4195 	 * is probably extremely rare and not worth worrying
4196 	 * over.
4197 	 */
4198 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4199 	    bufmallocspace < maxbufmallocspace) {
4200 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4201 		bp->b_flags |= B_MALLOC;
4202 		bufmallocadjust(bp, newbsize);
4203 		return;
4204 	}
4205 
4206 	/*
4207 	 * If the buffer is growing on its other-than-first
4208 	 * allocation then we revert to the page-allocation
4209 	 * scheme.
4210 	 */
4211 	origbuf = NULL;
4212 	origbufsize = 0;
4213 	if (bp->b_flags & B_MALLOC) {
4214 		origbuf = bp->b_data;
4215 		origbufsize = bp->b_bufsize;
4216 		bp->b_data = bp->b_kvabase;
4217 		bufmallocadjust(bp, 0);
4218 		bp->b_flags &= ~B_MALLOC;
4219 		newbsize = round_page(newbsize);
4220 	}
4221 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4222 	    (vm_offset_t) bp->b_data + newbsize);
4223 	if (origbuf != NULL) {
4224 		bcopy(origbuf, bp->b_data, origbufsize);
4225 		free(origbuf, M_BIOBUF);
4226 	}
4227 	bufspace_adjust(bp, newbsize);
4228 }
4229 
4230 /*
4231  * This code constitutes the buffer memory from either anonymous system
4232  * memory (in the case of non-VMIO operations) or from an associated
4233  * VM object (in the case of VMIO operations).  This code is able to
4234  * resize a buffer up or down.
4235  *
4236  * Note that this code is tricky, and has many complications to resolve
4237  * deadlock or inconsistent data situations.  Tread lightly!!!
4238  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4239  * the caller.  Calling this code willy nilly can result in the loss of data.
4240  *
4241  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
4242  * B_CACHE for the non-VMIO case.
4243  */
4244 int
4245 allocbuf(struct buf *bp, int size)
4246 {
4247 	int newbsize;
4248 
4249 	BUF_ASSERT_HELD(bp);
4250 
4251 	if (bp->b_bcount == size)
4252 		return (1);
4253 
4254 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4255 		panic("allocbuf: buffer too small");
4256 
4257 	newbsize = roundup2(size, DEV_BSIZE);
4258 	if ((bp->b_flags & B_VMIO) == 0) {
4259 		if ((bp->b_flags & B_MALLOC) == 0)
4260 			newbsize = round_page(newbsize);
4261 		/*
4262 		 * Just get anonymous memory from the kernel.  Don't
4263 		 * mess with B_CACHE.
4264 		 */
4265 		if (newbsize < bp->b_bufsize)
4266 			vfs_nonvmio_truncate(bp, newbsize);
4267 		else if (newbsize > bp->b_bufsize)
4268 			vfs_nonvmio_extend(bp, newbsize);
4269 	} else {
4270 		int desiredpages;
4271 
4272 		desiredpages = (size == 0) ? 0 :
4273 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4274 
4275 		if (bp->b_flags & B_MALLOC)
4276 			panic("allocbuf: VMIO buffer can't be malloced");
4277 		/*
4278 		 * Set B_CACHE initially if buffer is 0 length or will become
4279 		 * 0-length.
4280 		 */
4281 		if (size == 0 || bp->b_bufsize == 0)
4282 			bp->b_flags |= B_CACHE;
4283 
4284 		if (newbsize < bp->b_bufsize)
4285 			vfs_vmio_truncate(bp, desiredpages);
4286 		/* XXX This looks as if it should be newbsize > b_bufsize */
4287 		else if (size > bp->b_bcount)
4288 			vfs_vmio_extend(bp, desiredpages, size);
4289 		bufspace_adjust(bp, newbsize);
4290 	}
4291 	bp->b_bcount = size;		/* requested buffer size. */
4292 	return (1);
4293 }
4294 
4295 extern int inflight_transient_maps;
4296 
4297 static struct bio_queue nondump_bios;
4298 
4299 void
4300 biodone(struct bio *bp)
4301 {
4302 	struct mtx *mtxp;
4303 	void (*done)(struct bio *);
4304 	vm_offset_t start, end;
4305 
4306 	biotrack(bp, __func__);
4307 
4308 	/*
4309 	 * Avoid completing I/O when dumping after a panic since that may
4310 	 * result in a deadlock in the filesystem or pager code.  Note that
4311 	 * this doesn't affect dumps that were started manually since we aim
4312 	 * to keep the system usable after it has been resumed.
4313 	 */
4314 	if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4315 		TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4316 		return;
4317 	}
4318 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4319 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4320 		bp->bio_flags |= BIO_UNMAPPED;
4321 		start = trunc_page((vm_offset_t)bp->bio_data);
4322 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4323 		bp->bio_data = unmapped_buf;
4324 		pmap_qremove(start, atop(end - start));
4325 		vmem_free(transient_arena, start, end - start);
4326 		atomic_add_int(&inflight_transient_maps, -1);
4327 	}
4328 	done = bp->bio_done;
4329 	if (done == NULL) {
4330 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
4331 		mtx_lock(mtxp);
4332 		bp->bio_flags |= BIO_DONE;
4333 		wakeup(bp);
4334 		mtx_unlock(mtxp);
4335 	} else
4336 		done(bp);
4337 }
4338 
4339 /*
4340  * Wait for a BIO to finish.
4341  */
4342 int
4343 biowait(struct bio *bp, const char *wchan)
4344 {
4345 	struct mtx *mtxp;
4346 
4347 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4348 	mtx_lock(mtxp);
4349 	while ((bp->bio_flags & BIO_DONE) == 0)
4350 		msleep(bp, mtxp, PRIBIO, wchan, 0);
4351 	mtx_unlock(mtxp);
4352 	if (bp->bio_error != 0)
4353 		return (bp->bio_error);
4354 	if (!(bp->bio_flags & BIO_ERROR))
4355 		return (0);
4356 	return (EIO);
4357 }
4358 
4359 void
4360 biofinish(struct bio *bp, struct devstat *stat, int error)
4361 {
4362 
4363 	if (error) {
4364 		bp->bio_error = error;
4365 		bp->bio_flags |= BIO_ERROR;
4366 	}
4367 	if (stat != NULL)
4368 		devstat_end_transaction_bio(stat, bp);
4369 	biodone(bp);
4370 }
4371 
4372 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4373 void
4374 biotrack_buf(struct bio *bp, const char *location)
4375 {
4376 
4377 	buf_track(bp->bio_track_bp, location);
4378 }
4379 #endif
4380 
4381 /*
4382  *	bufwait:
4383  *
4384  *	Wait for buffer I/O completion, returning error status.  The buffer
4385  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
4386  *	error and cleared.
4387  */
4388 int
4389 bufwait(struct buf *bp)
4390 {
4391 	if (bp->b_iocmd == BIO_READ)
4392 		bwait(bp, PRIBIO, "biord");
4393 	else
4394 		bwait(bp, PRIBIO, "biowr");
4395 	if (bp->b_flags & B_EINTR) {
4396 		bp->b_flags &= ~B_EINTR;
4397 		return (EINTR);
4398 	}
4399 	if (bp->b_ioflags & BIO_ERROR) {
4400 		return (bp->b_error ? bp->b_error : EIO);
4401 	} else {
4402 		return (0);
4403 	}
4404 }
4405 
4406 /*
4407  *	bufdone:
4408  *
4409  *	Finish I/O on a buffer, optionally calling a completion function.
4410  *	This is usually called from an interrupt so process blocking is
4411  *	not allowed.
4412  *
4413  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4414  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4415  *	assuming B_INVAL is clear.
4416  *
4417  *	For the VMIO case, we set B_CACHE if the op was a read and no
4418  *	read error occurred, or if the op was a write.  B_CACHE is never
4419  *	set if the buffer is invalid or otherwise uncacheable.
4420  *
4421  *	bufdone does not mess with B_INVAL, allowing the I/O routine or the
4422  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4423  *	in the biodone routine.
4424  */
4425 void
4426 bufdone(struct buf *bp)
4427 {
4428 	struct bufobj *dropobj;
4429 	void    (*biodone)(struct buf *);
4430 
4431 	buf_track(bp, __func__);
4432 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4433 	dropobj = NULL;
4434 
4435 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4436 	BUF_ASSERT_HELD(bp);
4437 
4438 	runningbufwakeup(bp);
4439 	if (bp->b_iocmd == BIO_WRITE)
4440 		dropobj = bp->b_bufobj;
4441 	/* call optional completion function if requested */
4442 	if (bp->b_iodone != NULL) {
4443 		biodone = bp->b_iodone;
4444 		bp->b_iodone = NULL;
4445 		(*biodone) (bp);
4446 		if (dropobj)
4447 			bufobj_wdrop(dropobj);
4448 		return;
4449 	}
4450 	if (bp->b_flags & B_VMIO) {
4451 		/*
4452 		 * Set B_CACHE if the op was a normal read and no error
4453 		 * occurred.  B_CACHE is set for writes in the b*write()
4454 		 * routines.
4455 		 */
4456 		if (bp->b_iocmd == BIO_READ &&
4457 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4458 		    !(bp->b_ioflags & BIO_ERROR))
4459 			bp->b_flags |= B_CACHE;
4460 		vfs_vmio_iodone(bp);
4461 	}
4462 	if (!LIST_EMPTY(&bp->b_dep))
4463 		buf_complete(bp);
4464 	if ((bp->b_flags & B_CKHASH) != 0) {
4465 		KASSERT(bp->b_iocmd == BIO_READ,
4466 		    ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4467 		KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4468 		(*bp->b_ckhashcalc)(bp);
4469 	}
4470 	/*
4471 	 * For asynchronous completions, release the buffer now. The brelse
4472 	 * will do a wakeup there if necessary - so no need to do a wakeup
4473 	 * here in the async case. The sync case always needs to do a wakeup.
4474 	 */
4475 	if (bp->b_flags & B_ASYNC) {
4476 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4477 		    (bp->b_ioflags & BIO_ERROR))
4478 			brelse(bp);
4479 		else
4480 			bqrelse(bp);
4481 	} else
4482 		bdone(bp);
4483 	if (dropobj)
4484 		bufobj_wdrop(dropobj);
4485 }
4486 
4487 /*
4488  * This routine is called in lieu of iodone in the case of
4489  * incomplete I/O.  This keeps the busy status for pages
4490  * consistent.
4491  */
4492 void
4493 vfs_unbusy_pages(struct buf *bp)
4494 {
4495 	int i;
4496 	vm_object_t obj;
4497 	vm_page_t m;
4498 
4499 	runningbufwakeup(bp);
4500 	if (!(bp->b_flags & B_VMIO))
4501 		return;
4502 
4503 	obj = bp->b_bufobj->bo_object;
4504 	VM_OBJECT_WLOCK(obj);
4505 	for (i = 0; i < bp->b_npages; i++) {
4506 		m = bp->b_pages[i];
4507 		if (m == bogus_page) {
4508 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4509 			if (!m)
4510 				panic("vfs_unbusy_pages: page missing\n");
4511 			bp->b_pages[i] = m;
4512 			if (buf_mapped(bp)) {
4513 				BUF_CHECK_MAPPED(bp);
4514 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4515 				    bp->b_pages, bp->b_npages);
4516 			} else
4517 				BUF_CHECK_UNMAPPED(bp);
4518 		}
4519 		vm_page_sunbusy(m);
4520 	}
4521 	vm_object_pip_wakeupn(obj, bp->b_npages);
4522 	VM_OBJECT_WUNLOCK(obj);
4523 }
4524 
4525 /*
4526  * vfs_page_set_valid:
4527  *
4528  *	Set the valid bits in a page based on the supplied offset.   The
4529  *	range is restricted to the buffer's size.
4530  *
4531  *	This routine is typically called after a read completes.
4532  */
4533 static void
4534 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4535 {
4536 	vm_ooffset_t eoff;
4537 
4538 	/*
4539 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4540 	 * page boundary and eoff is not greater than the end of the buffer.
4541 	 * The end of the buffer, in this case, is our file EOF, not the
4542 	 * allocation size of the buffer.
4543 	 */
4544 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4545 	if (eoff > bp->b_offset + bp->b_bcount)
4546 		eoff = bp->b_offset + bp->b_bcount;
4547 
4548 	/*
4549 	 * Set valid range.  This is typically the entire buffer and thus the
4550 	 * entire page.
4551 	 */
4552 	if (eoff > off)
4553 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4554 }
4555 
4556 /*
4557  * vfs_page_set_validclean:
4558  *
4559  *	Set the valid bits and clear the dirty bits in a page based on the
4560  *	supplied offset.   The range is restricted to the buffer's size.
4561  */
4562 static void
4563 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4564 {
4565 	vm_ooffset_t soff, eoff;
4566 
4567 	/*
4568 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4569 	 * page boundary or cross the end of the buffer.  The end of the
4570 	 * buffer, in this case, is our file EOF, not the allocation size
4571 	 * of the buffer.
4572 	 */
4573 	soff = off;
4574 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4575 	if (eoff > bp->b_offset + bp->b_bcount)
4576 		eoff = bp->b_offset + bp->b_bcount;
4577 
4578 	/*
4579 	 * Set valid range.  This is typically the entire buffer and thus the
4580 	 * entire page.
4581 	 */
4582 	if (eoff > soff) {
4583 		vm_page_set_validclean(
4584 		    m,
4585 		   (vm_offset_t) (soff & PAGE_MASK),
4586 		   (vm_offset_t) (eoff - soff)
4587 		);
4588 	}
4589 }
4590 
4591 /*
4592  * Ensure that all buffer pages are not exclusive busied.  If any page is
4593  * exclusive busy, drain it.
4594  */
4595 void
4596 vfs_drain_busy_pages(struct buf *bp)
4597 {
4598 	vm_page_t m;
4599 	int i, last_busied;
4600 
4601 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4602 	last_busied = 0;
4603 	for (i = 0; i < bp->b_npages; i++) {
4604 		m = bp->b_pages[i];
4605 		if (vm_page_xbusied(m)) {
4606 			for (; last_busied < i; last_busied++)
4607 				vm_page_sbusy(bp->b_pages[last_busied]);
4608 			while (vm_page_xbusied(m)) {
4609 				vm_page_lock(m);
4610 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4611 				vm_page_busy_sleep(m, "vbpage", true);
4612 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4613 			}
4614 		}
4615 	}
4616 	for (i = 0; i < last_busied; i++)
4617 		vm_page_sunbusy(bp->b_pages[i]);
4618 }
4619 
4620 /*
4621  * This routine is called before a device strategy routine.
4622  * It is used to tell the VM system that paging I/O is in
4623  * progress, and treat the pages associated with the buffer
4624  * almost as being exclusive busy.  Also the object paging_in_progress
4625  * flag is handled to make sure that the object doesn't become
4626  * inconsistent.
4627  *
4628  * Since I/O has not been initiated yet, certain buffer flags
4629  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4630  * and should be ignored.
4631  */
4632 void
4633 vfs_busy_pages(struct buf *bp, int clear_modify)
4634 {
4635 	vm_object_t obj;
4636 	vm_ooffset_t foff;
4637 	vm_page_t m;
4638 	int i;
4639 	bool bogus;
4640 
4641 	if (!(bp->b_flags & B_VMIO))
4642 		return;
4643 
4644 	obj = bp->b_bufobj->bo_object;
4645 	foff = bp->b_offset;
4646 	KASSERT(bp->b_offset != NOOFFSET,
4647 	    ("vfs_busy_pages: no buffer offset"));
4648 	VM_OBJECT_WLOCK(obj);
4649 	vfs_drain_busy_pages(bp);
4650 	if (bp->b_bufsize != 0)
4651 		vfs_setdirty_locked_object(bp);
4652 	bogus = false;
4653 	for (i = 0; i < bp->b_npages; i++) {
4654 		m = bp->b_pages[i];
4655 
4656 		if ((bp->b_flags & B_CLUSTER) == 0) {
4657 			vm_object_pip_add(obj, 1);
4658 			vm_page_sbusy(m);
4659 		}
4660 		/*
4661 		 * When readying a buffer for a read ( i.e
4662 		 * clear_modify == 0 ), it is important to do
4663 		 * bogus_page replacement for valid pages in
4664 		 * partially instantiated buffers.  Partially
4665 		 * instantiated buffers can, in turn, occur when
4666 		 * reconstituting a buffer from its VM backing store
4667 		 * base.  We only have to do this if B_CACHE is
4668 		 * clear ( which causes the I/O to occur in the
4669 		 * first place ).  The replacement prevents the read
4670 		 * I/O from overwriting potentially dirty VM-backed
4671 		 * pages.  XXX bogus page replacement is, uh, bogus.
4672 		 * It may not work properly with small-block devices.
4673 		 * We need to find a better way.
4674 		 */
4675 		if (clear_modify) {
4676 			pmap_remove_write(m);
4677 			vfs_page_set_validclean(bp, foff, m);
4678 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4679 		    (bp->b_flags & B_CACHE) == 0) {
4680 			bp->b_pages[i] = bogus_page;
4681 			bogus = true;
4682 		}
4683 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4684 	}
4685 	VM_OBJECT_WUNLOCK(obj);
4686 	if (bogus && buf_mapped(bp)) {
4687 		BUF_CHECK_MAPPED(bp);
4688 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4689 		    bp->b_pages, bp->b_npages);
4690 	}
4691 }
4692 
4693 /*
4694  *	vfs_bio_set_valid:
4695  *
4696  *	Set the range within the buffer to valid.  The range is
4697  *	relative to the beginning of the buffer, b_offset.  Note that
4698  *	b_offset itself may be offset from the beginning of the first
4699  *	page.
4700  */
4701 void
4702 vfs_bio_set_valid(struct buf *bp, int base, int size)
4703 {
4704 	int i, n;
4705 	vm_page_t m;
4706 
4707 	if (!(bp->b_flags & B_VMIO))
4708 		return;
4709 
4710 	/*
4711 	 * Fixup base to be relative to beginning of first page.
4712 	 * Set initial n to be the maximum number of bytes in the
4713 	 * first page that can be validated.
4714 	 */
4715 	base += (bp->b_offset & PAGE_MASK);
4716 	n = PAGE_SIZE - (base & PAGE_MASK);
4717 
4718 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4719 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4720 		m = bp->b_pages[i];
4721 		if (n > size)
4722 			n = size;
4723 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4724 		base += n;
4725 		size -= n;
4726 		n = PAGE_SIZE;
4727 	}
4728 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4729 }
4730 
4731 /*
4732  *	vfs_bio_clrbuf:
4733  *
4734  *	If the specified buffer is a non-VMIO buffer, clear the entire
4735  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4736  *	validate only the previously invalid portions of the buffer.
4737  *	This routine essentially fakes an I/O, so we need to clear
4738  *	BIO_ERROR and B_INVAL.
4739  *
4740  *	Note that while we only theoretically need to clear through b_bcount,
4741  *	we go ahead and clear through b_bufsize.
4742  */
4743 void
4744 vfs_bio_clrbuf(struct buf *bp)
4745 {
4746 	int i, j, mask, sa, ea, slide;
4747 
4748 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4749 		clrbuf(bp);
4750 		return;
4751 	}
4752 	bp->b_flags &= ~B_INVAL;
4753 	bp->b_ioflags &= ~BIO_ERROR;
4754 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4755 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4756 	    (bp->b_offset & PAGE_MASK) == 0) {
4757 		if (bp->b_pages[0] == bogus_page)
4758 			goto unlock;
4759 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4760 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4761 		if ((bp->b_pages[0]->valid & mask) == mask)
4762 			goto unlock;
4763 		if ((bp->b_pages[0]->valid & mask) == 0) {
4764 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4765 			bp->b_pages[0]->valid |= mask;
4766 			goto unlock;
4767 		}
4768 	}
4769 	sa = bp->b_offset & PAGE_MASK;
4770 	slide = 0;
4771 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4772 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4773 		ea = slide & PAGE_MASK;
4774 		if (ea == 0)
4775 			ea = PAGE_SIZE;
4776 		if (bp->b_pages[i] == bogus_page)
4777 			continue;
4778 		j = sa / DEV_BSIZE;
4779 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4780 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4781 		if ((bp->b_pages[i]->valid & mask) == mask)
4782 			continue;
4783 		if ((bp->b_pages[i]->valid & mask) == 0)
4784 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4785 		else {
4786 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4787 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4788 					pmap_zero_page_area(bp->b_pages[i],
4789 					    sa, DEV_BSIZE);
4790 				}
4791 			}
4792 		}
4793 		bp->b_pages[i]->valid |= mask;
4794 	}
4795 unlock:
4796 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4797 	bp->b_resid = 0;
4798 }
4799 
4800 void
4801 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4802 {
4803 	vm_page_t m;
4804 	int i, n;
4805 
4806 	if (buf_mapped(bp)) {
4807 		BUF_CHECK_MAPPED(bp);
4808 		bzero(bp->b_data + base, size);
4809 	} else {
4810 		BUF_CHECK_UNMAPPED(bp);
4811 		n = PAGE_SIZE - (base & PAGE_MASK);
4812 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4813 			m = bp->b_pages[i];
4814 			if (n > size)
4815 				n = size;
4816 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4817 			base += n;
4818 			size -= n;
4819 			n = PAGE_SIZE;
4820 		}
4821 	}
4822 }
4823 
4824 /*
4825  * Update buffer flags based on I/O request parameters, optionally releasing the
4826  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4827  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4828  * I/O).  Otherwise the buffer is released to the cache.
4829  */
4830 static void
4831 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4832 {
4833 
4834 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4835 	    ("buf %p non-VMIO noreuse", bp));
4836 
4837 	if ((ioflag & IO_DIRECT) != 0)
4838 		bp->b_flags |= B_DIRECT;
4839 	if ((ioflag & IO_EXT) != 0)
4840 		bp->b_xflags |= BX_ALTDATA;
4841 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4842 		bp->b_flags |= B_RELBUF;
4843 		if ((ioflag & IO_NOREUSE) != 0)
4844 			bp->b_flags |= B_NOREUSE;
4845 		if (release)
4846 			brelse(bp);
4847 	} else if (release)
4848 		bqrelse(bp);
4849 }
4850 
4851 void
4852 vfs_bio_brelse(struct buf *bp, int ioflag)
4853 {
4854 
4855 	b_io_dismiss(bp, ioflag, true);
4856 }
4857 
4858 void
4859 vfs_bio_set_flags(struct buf *bp, int ioflag)
4860 {
4861 
4862 	b_io_dismiss(bp, ioflag, false);
4863 }
4864 
4865 /*
4866  * vm_hold_load_pages and vm_hold_free_pages get pages into
4867  * a buffers address space.  The pages are anonymous and are
4868  * not associated with a file object.
4869  */
4870 static void
4871 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4872 {
4873 	vm_offset_t pg;
4874 	vm_page_t p;
4875 	int index;
4876 
4877 	BUF_CHECK_MAPPED(bp);
4878 
4879 	to = round_page(to);
4880 	from = round_page(from);
4881 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4882 
4883 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4884 		/*
4885 		 * note: must allocate system pages since blocking here
4886 		 * could interfere with paging I/O, no matter which
4887 		 * process we are.
4888 		 */
4889 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4890 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4891 		    VM_ALLOC_WAITOK);
4892 		pmap_qenter(pg, &p, 1);
4893 		bp->b_pages[index] = p;
4894 	}
4895 	bp->b_npages = index;
4896 }
4897 
4898 /* Return pages associated with this buf to the vm system */
4899 static void
4900 vm_hold_free_pages(struct buf *bp, int newbsize)
4901 {
4902 	vm_offset_t from;
4903 	vm_page_t p;
4904 	int index, newnpages;
4905 
4906 	BUF_CHECK_MAPPED(bp);
4907 
4908 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4909 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4910 	if (bp->b_npages > newnpages)
4911 		pmap_qremove(from, bp->b_npages - newnpages);
4912 	for (index = newnpages; index < bp->b_npages; index++) {
4913 		p = bp->b_pages[index];
4914 		bp->b_pages[index] = NULL;
4915 		p->wire_count--;
4916 		vm_page_free(p);
4917 	}
4918 	vm_wire_sub(bp->b_npages - newnpages);
4919 	bp->b_npages = newnpages;
4920 }
4921 
4922 /*
4923  * Map an IO request into kernel virtual address space.
4924  *
4925  * All requests are (re)mapped into kernel VA space.
4926  * Notice that we use b_bufsize for the size of the buffer
4927  * to be mapped.  b_bcount might be modified by the driver.
4928  *
4929  * Note that even if the caller determines that the address space should
4930  * be valid, a race or a smaller-file mapped into a larger space may
4931  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4932  * check the return value.
4933  *
4934  * This function only works with pager buffers.
4935  */
4936 int
4937 vmapbuf(struct buf *bp, int mapbuf)
4938 {
4939 	vm_prot_t prot;
4940 	int pidx;
4941 
4942 	if (bp->b_bufsize < 0)
4943 		return (-1);
4944 	prot = VM_PROT_READ;
4945 	if (bp->b_iocmd == BIO_READ)
4946 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4947 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4948 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4949 	    btoc(MAXPHYS))) < 0)
4950 		return (-1);
4951 	bp->b_npages = pidx;
4952 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4953 	if (mapbuf || !unmapped_buf_allowed) {
4954 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4955 		bp->b_data = bp->b_kvabase + bp->b_offset;
4956 	} else
4957 		bp->b_data = unmapped_buf;
4958 	return(0);
4959 }
4960 
4961 /*
4962  * Free the io map PTEs associated with this IO operation.
4963  * We also invalidate the TLB entries and restore the original b_addr.
4964  *
4965  * This function only works with pager buffers.
4966  */
4967 void
4968 vunmapbuf(struct buf *bp)
4969 {
4970 	int npages;
4971 
4972 	npages = bp->b_npages;
4973 	if (buf_mapped(bp))
4974 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4975 	vm_page_unhold_pages(bp->b_pages, npages);
4976 
4977 	bp->b_data = unmapped_buf;
4978 }
4979 
4980 void
4981 bdone(struct buf *bp)
4982 {
4983 	struct mtx *mtxp;
4984 
4985 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4986 	mtx_lock(mtxp);
4987 	bp->b_flags |= B_DONE;
4988 	wakeup(bp);
4989 	mtx_unlock(mtxp);
4990 }
4991 
4992 void
4993 bwait(struct buf *bp, u_char pri, const char *wchan)
4994 {
4995 	struct mtx *mtxp;
4996 
4997 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4998 	mtx_lock(mtxp);
4999 	while ((bp->b_flags & B_DONE) == 0)
5000 		msleep(bp, mtxp, pri, wchan, 0);
5001 	mtx_unlock(mtxp);
5002 }
5003 
5004 int
5005 bufsync(struct bufobj *bo, int waitfor)
5006 {
5007 
5008 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5009 }
5010 
5011 void
5012 bufstrategy(struct bufobj *bo, struct buf *bp)
5013 {
5014 	int i __unused;
5015 	struct vnode *vp;
5016 
5017 	vp = bp->b_vp;
5018 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5019 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5020 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5021 	i = VOP_STRATEGY(vp, bp);
5022 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5023 }
5024 
5025 /*
5026  * Initialize a struct bufobj before use.  Memory is assumed zero filled.
5027  */
5028 void
5029 bufobj_init(struct bufobj *bo, void *private)
5030 {
5031 	static volatile int bufobj_cleanq;
5032 
5033         bo->bo_domain =
5034             atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5035         rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5036         bo->bo_private = private;
5037         TAILQ_INIT(&bo->bo_clean.bv_hd);
5038         TAILQ_INIT(&bo->bo_dirty.bv_hd);
5039 }
5040 
5041 void
5042 bufobj_wrefl(struct bufobj *bo)
5043 {
5044 
5045 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5046 	ASSERT_BO_WLOCKED(bo);
5047 	bo->bo_numoutput++;
5048 }
5049 
5050 void
5051 bufobj_wref(struct bufobj *bo)
5052 {
5053 
5054 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5055 	BO_LOCK(bo);
5056 	bo->bo_numoutput++;
5057 	BO_UNLOCK(bo);
5058 }
5059 
5060 void
5061 bufobj_wdrop(struct bufobj *bo)
5062 {
5063 
5064 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5065 	BO_LOCK(bo);
5066 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5067 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5068 		bo->bo_flag &= ~BO_WWAIT;
5069 		wakeup(&bo->bo_numoutput);
5070 	}
5071 	BO_UNLOCK(bo);
5072 }
5073 
5074 int
5075 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5076 {
5077 	int error;
5078 
5079 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5080 	ASSERT_BO_WLOCKED(bo);
5081 	error = 0;
5082 	while (bo->bo_numoutput) {
5083 		bo->bo_flag |= BO_WWAIT;
5084 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5085 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5086 		if (error)
5087 			break;
5088 	}
5089 	return (error);
5090 }
5091 
5092 /*
5093  * Set bio_data or bio_ma for struct bio from the struct buf.
5094  */
5095 void
5096 bdata2bio(struct buf *bp, struct bio *bip)
5097 {
5098 
5099 	if (!buf_mapped(bp)) {
5100 		KASSERT(unmapped_buf_allowed, ("unmapped"));
5101 		bip->bio_ma = bp->b_pages;
5102 		bip->bio_ma_n = bp->b_npages;
5103 		bip->bio_data = unmapped_buf;
5104 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5105 		bip->bio_flags |= BIO_UNMAPPED;
5106 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5107 		    PAGE_SIZE == bp->b_npages,
5108 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5109 		    (long long)bip->bio_length, bip->bio_ma_n));
5110 	} else {
5111 		bip->bio_data = bp->b_data;
5112 		bip->bio_ma = NULL;
5113 	}
5114 }
5115 
5116 /*
5117  * The MIPS pmap code currently doesn't handle aliased pages.
5118  * The VIPT caches may not handle page aliasing themselves, leading
5119  * to data corruption.
5120  *
5121  * As such, this code makes a system extremely unhappy if said
5122  * system doesn't support unaliasing the above situation in hardware.
5123  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5124  * this feature at build time, so it has to be handled in software.
5125  *
5126  * Once the MIPS pmap/cache code grows to support this function on
5127  * earlier chips, it should be flipped back off.
5128  */
5129 #ifdef	__mips__
5130 static int buf_pager_relbuf = 1;
5131 #else
5132 static int buf_pager_relbuf = 0;
5133 #endif
5134 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5135     &buf_pager_relbuf, 0,
5136     "Make buffer pager release buffers after reading");
5137 
5138 /*
5139  * The buffer pager.  It uses buffer reads to validate pages.
5140  *
5141  * In contrast to the generic local pager from vm/vnode_pager.c, this
5142  * pager correctly and easily handles volumes where the underlying
5143  * device block size is greater than the machine page size.  The
5144  * buffer cache transparently extends the requested page run to be
5145  * aligned at the block boundary, and does the necessary bogus page
5146  * replacements in the addends to avoid obliterating already valid
5147  * pages.
5148  *
5149  * The only non-trivial issue is that the exclusive busy state for
5150  * pages, which is assumed by the vm_pager_getpages() interface, is
5151  * incompatible with the VMIO buffer cache's desire to share-busy the
5152  * pages.  This function performs a trivial downgrade of the pages'
5153  * state before reading buffers, and a less trivial upgrade from the
5154  * shared-busy to excl-busy state after the read.
5155  */
5156 int
5157 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5158     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5159     vbg_get_blksize_t get_blksize)
5160 {
5161 	vm_page_t m;
5162 	vm_object_t object;
5163 	struct buf *bp;
5164 	struct mount *mp;
5165 	daddr_t lbn, lbnp;
5166 	vm_ooffset_t la, lb, poff, poffe;
5167 	long bsize;
5168 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5169 	bool redo, lpart;
5170 
5171 	object = vp->v_object;
5172 	mp = vp->v_mount;
5173 	error = 0;
5174 	la = IDX_TO_OFF(ma[count - 1]->pindex);
5175 	if (la >= object->un_pager.vnp.vnp_size)
5176 		return (VM_PAGER_BAD);
5177 
5178 	/*
5179 	 * Change the meaning of la from where the last requested page starts
5180 	 * to where it ends, because that's the end of the requested region
5181 	 * and the start of the potential read-ahead region.
5182 	 */
5183 	la += PAGE_SIZE;
5184 	lpart = la > object->un_pager.vnp.vnp_size;
5185 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5186 
5187 	/*
5188 	 * Calculate read-ahead, behind and total pages.
5189 	 */
5190 	pgsin = count;
5191 	lb = IDX_TO_OFF(ma[0]->pindex);
5192 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5193 	pgsin += pgsin_b;
5194 	if (rbehind != NULL)
5195 		*rbehind = pgsin_b;
5196 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5197 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5198 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5199 		    PAGE_SIZE) - la);
5200 	pgsin += pgsin_a;
5201 	if (rahead != NULL)
5202 		*rahead = pgsin_a;
5203 	VM_CNT_INC(v_vnodein);
5204 	VM_CNT_ADD(v_vnodepgsin, pgsin);
5205 
5206 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5207 	    != 0) ? GB_UNMAPPED : 0;
5208 	VM_OBJECT_WLOCK(object);
5209 again:
5210 	for (i = 0; i < count; i++)
5211 		vm_page_busy_downgrade(ma[i]);
5212 	VM_OBJECT_WUNLOCK(object);
5213 
5214 	lbnp = -1;
5215 	for (i = 0; i < count; i++) {
5216 		m = ma[i];
5217 
5218 		/*
5219 		 * Pages are shared busy and the object lock is not
5220 		 * owned, which together allow for the pages'
5221 		 * invalidation.  The racy test for validity avoids
5222 		 * useless creation of the buffer for the most typical
5223 		 * case when invalidation is not used in redo or for
5224 		 * parallel read.  The shared->excl upgrade loop at
5225 		 * the end of the function catches the race in a
5226 		 * reliable way (protected by the object lock).
5227 		 */
5228 		if (m->valid == VM_PAGE_BITS_ALL)
5229 			continue;
5230 
5231 		poff = IDX_TO_OFF(m->pindex);
5232 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5233 		for (; poff < poffe; poff += bsize) {
5234 			lbn = get_lblkno(vp, poff);
5235 			if (lbn == lbnp)
5236 				goto next_page;
5237 			lbnp = lbn;
5238 
5239 			bsize = get_blksize(vp, lbn);
5240 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5241 			    br_flags, &bp);
5242 			if (error != 0)
5243 				goto end_pages;
5244 			if (LIST_EMPTY(&bp->b_dep)) {
5245 				/*
5246 				 * Invalidation clears m->valid, but
5247 				 * may leave B_CACHE flag if the
5248 				 * buffer existed at the invalidation
5249 				 * time.  In this case, recycle the
5250 				 * buffer to do real read on next
5251 				 * bread() after redo.
5252 				 *
5253 				 * Otherwise B_RELBUF is not strictly
5254 				 * necessary, enable to reduce buf
5255 				 * cache pressure.
5256 				 */
5257 				if (buf_pager_relbuf ||
5258 				    m->valid != VM_PAGE_BITS_ALL)
5259 					bp->b_flags |= B_RELBUF;
5260 
5261 				bp->b_flags &= ~B_NOCACHE;
5262 				brelse(bp);
5263 			} else {
5264 				bqrelse(bp);
5265 			}
5266 		}
5267 		KASSERT(1 /* racy, enable for debugging */ ||
5268 		    m->valid == VM_PAGE_BITS_ALL || i == count - 1,
5269 		    ("buf %d %p invalid", i, m));
5270 		if (i == count - 1 && lpart) {
5271 			VM_OBJECT_WLOCK(object);
5272 			if (m->valid != 0 &&
5273 			    m->valid != VM_PAGE_BITS_ALL)
5274 				vm_page_zero_invalid(m, TRUE);
5275 			VM_OBJECT_WUNLOCK(object);
5276 		}
5277 next_page:;
5278 	}
5279 end_pages:
5280 
5281 	VM_OBJECT_WLOCK(object);
5282 	redo = false;
5283 	for (i = 0; i < count; i++) {
5284 		vm_page_sunbusy(ma[i]);
5285 		ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
5286 
5287 		/*
5288 		 * Since the pages were only sbusy while neither the
5289 		 * buffer nor the object lock was held by us, or
5290 		 * reallocated while vm_page_grab() slept for busy
5291 		 * relinguish, they could have been invalidated.
5292 		 * Recheck the valid bits and re-read as needed.
5293 		 *
5294 		 * Note that the last page is made fully valid in the
5295 		 * read loop, and partial validity for the page at
5296 		 * index count - 1 could mean that the page was
5297 		 * invalidated or removed, so we must restart for
5298 		 * safety as well.
5299 		 */
5300 		if (ma[i]->valid != VM_PAGE_BITS_ALL)
5301 			redo = true;
5302 	}
5303 	if (redo && error == 0)
5304 		goto again;
5305 	VM_OBJECT_WUNLOCK(object);
5306 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5307 }
5308 
5309 #include "opt_ddb.h"
5310 #ifdef DDB
5311 #include <ddb/ddb.h>
5312 
5313 /* DDB command to show buffer data */
5314 DB_SHOW_COMMAND(buffer, db_show_buffer)
5315 {
5316 	/* get args */
5317 	struct buf *bp = (struct buf *)addr;
5318 #ifdef FULL_BUF_TRACKING
5319 	uint32_t i, j;
5320 #endif
5321 
5322 	if (!have_addr) {
5323 		db_printf("usage: show buffer <addr>\n");
5324 		return;
5325 	}
5326 
5327 	db_printf("buf at %p\n", bp);
5328 	db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5329 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5330 	    (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5331 	db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5332 	    (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5333 	    (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5334 	db_printf(
5335 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5336 	    "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5337 	    "b_vp = %p, b_dep = %p\n",
5338 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5339 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5340 	    (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5341 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
5342 	    bp->b_kvabase, bp->b_kvasize);
5343 	if (bp->b_npages) {
5344 		int i;
5345 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5346 		for (i = 0; i < bp->b_npages; i++) {
5347 			vm_page_t m;
5348 			m = bp->b_pages[i];
5349 			if (m != NULL)
5350 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5351 				    (u_long)m->pindex,
5352 				    (u_long)VM_PAGE_TO_PHYS(m));
5353 			else
5354 				db_printf("( ??? )");
5355 			if ((i + 1) < bp->b_npages)
5356 				db_printf(",");
5357 		}
5358 		db_printf("\n");
5359 	}
5360 	BUF_LOCKPRINTINFO(bp);
5361 #if defined(FULL_BUF_TRACKING)
5362 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5363 
5364 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5365 	for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5366 		if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5367 			continue;
5368 		db_printf(" %2u: %s\n", j,
5369 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5370 	}
5371 #elif defined(BUF_TRACKING)
5372 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5373 #endif
5374 	db_printf(" ");
5375 }
5376 
5377 DB_SHOW_COMMAND(bufqueues, bufqueues)
5378 {
5379 	struct bufdomain *bd;
5380 	struct buf *bp;
5381 	long total;
5382 	int i, j, cnt;
5383 
5384 	db_printf("bqempty: %d\n", bqempty.bq_len);
5385 
5386 	for (i = 0; i < buf_domains; i++) {
5387 		bd = &bdomain[i];
5388 		db_printf("Buf domain %d\n", i);
5389 		db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5390 		db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5391 		db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5392 		db_printf("\n");
5393 		db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5394 		db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5395 		db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5396 		db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5397 		db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5398 		db_printf("\n");
5399 		db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5400 		db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5401 		db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5402 		db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5403 		db_printf("\n");
5404 		total = 0;
5405 		TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5406 			total += bp->b_bufsize;
5407 		db_printf("\tcleanq count\t%d (%ld)\n",
5408 		    bd->bd_cleanq->bq_len, total);
5409 		total = 0;
5410 		TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5411 			total += bp->b_bufsize;
5412 		db_printf("\tdirtyq count\t%d (%ld)\n",
5413 		    bd->bd_dirtyq.bq_len, total);
5414 		db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5415 		db_printf("\tlim\t\t%d\n", bd->bd_lim);
5416 		db_printf("\tCPU ");
5417 		for (j = 0; j <= mp_maxid; j++)
5418 			db_printf("%d, ", bd->bd_subq[j].bq_len);
5419 		db_printf("\n");
5420 		cnt = 0;
5421 		total = 0;
5422 		for (j = 0; j < nbuf; j++)
5423 			if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5424 				cnt++;
5425 				total += buf[j].b_bufsize;
5426 			}
5427 		db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5428 		cnt = 0;
5429 		total = 0;
5430 		for (j = 0; j < nbuf; j++)
5431 			if (buf[j].b_domain == i) {
5432 				cnt++;
5433 				total += buf[j].b_bufsize;
5434 			}
5435 		db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5436 	}
5437 }
5438 
5439 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5440 {
5441 	struct buf *bp;
5442 	int i;
5443 
5444 	for (i = 0; i < nbuf; i++) {
5445 		bp = &buf[i];
5446 		if (BUF_ISLOCKED(bp)) {
5447 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5448 			db_printf("\n");
5449 			if (db_pager_quit)
5450 				break;
5451 		}
5452 	}
5453 }
5454 
5455 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5456 {
5457 	struct vnode *vp;
5458 	struct buf *bp;
5459 
5460 	if (!have_addr) {
5461 		db_printf("usage: show vnodebufs <addr>\n");
5462 		return;
5463 	}
5464 	vp = (struct vnode *)addr;
5465 	db_printf("Clean buffers:\n");
5466 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5467 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5468 		db_printf("\n");
5469 	}
5470 	db_printf("Dirty buffers:\n");
5471 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5472 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5473 		db_printf("\n");
5474 	}
5475 }
5476 
5477 DB_COMMAND(countfreebufs, db_coundfreebufs)
5478 {
5479 	struct buf *bp;
5480 	int i, used = 0, nfree = 0;
5481 
5482 	if (have_addr) {
5483 		db_printf("usage: countfreebufs\n");
5484 		return;
5485 	}
5486 
5487 	for (i = 0; i < nbuf; i++) {
5488 		bp = &buf[i];
5489 		if (bp->b_qindex == QUEUE_EMPTY)
5490 			nfree++;
5491 		else
5492 			used++;
5493 	}
5494 
5495 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5496 	    nfree + used);
5497 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5498 }
5499 #endif /* DDB */
5500