xref: /freebsd/sys/kern/vfs_bio.c (revision 49dae58b287906be26f56ba3e3dc693c3ba8cf37)
1 /*-
2  * Copyright (c) 2004 Poul-Henning Kamp
3  * Copyright (c) 1994,1997 John S. Dyson
4  * Copyright (c) 2013 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * Portions of this software were developed by Konstantin Belousov
8  * under sponsorship from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  * this file contains a new buffer I/O scheme implementing a coherent
34  * VM object and buffer cache scheme.  Pains have been taken to make
35  * sure that the performance degradation associated with schemes such
36  * as this is not realized.
37  *
38  * Author:  John S. Dyson
39  * Significant help during the development and debugging phases
40  * had been provided by David Greenman, also of the FreeBSD core team.
41  *
42  * see man buf(9) for more info.
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/conf.h>
52 #include <sys/buf.h>
53 #include <sys/devicestat.h>
54 #include <sys/eventhandler.h>
55 #include <sys/fail.h>
56 #include <sys/limits.h>
57 #include <sys/lock.h>
58 #include <sys/malloc.h>
59 #include <sys/mount.h>
60 #include <sys/mutex.h>
61 #include <sys/kernel.h>
62 #include <sys/kthread.h>
63 #include <sys/proc.h>
64 #include <sys/racct.h>
65 #include <sys/resourcevar.h>
66 #include <sys/rwlock.h>
67 #include <sys/smp.h>
68 #include <sys/sysctl.h>
69 #include <sys/sysproto.h>
70 #include <sys/vmem.h>
71 #include <sys/vmmeter.h>
72 #include <sys/vnode.h>
73 #include <sys/watchdog.h>
74 #include <geom/geom.h>
75 #include <vm/vm.h>
76 #include <vm/vm_param.h>
77 #include <vm/vm_kern.h>
78 #include <vm/vm_pageout.h>
79 #include <vm/vm_page.h>
80 #include <vm/vm_object.h>
81 #include <vm/vm_extern.h>
82 #include <vm/vm_map.h>
83 #include <vm/swap_pager.h>
84 #include "opt_compat.h"
85 #include "opt_swap.h"
86 
87 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
88 
89 struct	bio_ops bioops;		/* I/O operation notification */
90 
91 struct	buf_ops buf_ops_bio = {
92 	.bop_name	=	"buf_ops_bio",
93 	.bop_write	=	bufwrite,
94 	.bop_strategy	=	bufstrategy,
95 	.bop_sync	=	bufsync,
96 	.bop_bdflush	=	bufbdflush,
97 };
98 
99 static struct buf *buf;		/* buffer header pool */
100 extern struct buf *swbuf;	/* Swap buffer header pool. */
101 caddr_t unmapped_buf;
102 
103 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
104 struct proc *bufdaemonproc;
105 struct proc *bufspacedaemonproc;
106 
107 static int inmem(struct vnode *vp, daddr_t blkno);
108 static void vm_hold_free_pages(struct buf *bp, int newbsize);
109 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
110 		vm_offset_t to);
111 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
112 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
113 		vm_page_t m);
114 static void vfs_clean_pages_dirty_buf(struct buf *bp);
115 static void vfs_setdirty_locked_object(struct buf *bp);
116 static void vfs_vmio_invalidate(struct buf *bp);
117 static void vfs_vmio_truncate(struct buf *bp, int npages);
118 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
119 static int vfs_bio_clcheck(struct vnode *vp, int size,
120 		daddr_t lblkno, daddr_t blkno);
121 static int buf_flush(struct vnode *vp, int);
122 static int buf_recycle(bool);
123 static int buf_scan(bool);
124 static int flushbufqueues(struct vnode *, int, int);
125 static void buf_daemon(void);
126 static void bremfreel(struct buf *bp);
127 static __inline void bd_wakeup(void);
128 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
129 static void bufkva_reclaim(vmem_t *, int);
130 static void bufkva_free(struct buf *);
131 static int buf_import(void *, void **, int, int);
132 static void buf_release(void *, void **, int);
133 
134 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
135     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
136 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
137 #endif
138 
139 int vmiodirenable = TRUE;
140 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
141     "Use the VM system for directory writes");
142 long runningbufspace;
143 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
144     "Amount of presently outstanding async buffer io");
145 static long bufspace;
146 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
147     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
148 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
149     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
150 #else
151 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
152     "Physical memory used for buffers");
153 #endif
154 static long bufkvaspace;
155 SYSCTL_LONG(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 0,
156     "Kernel virtual memory used for buffers");
157 static long maxbufspace;
158 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, &maxbufspace, 0,
159     "Maximum allowed value of bufspace (including metadata)");
160 static long bufmallocspace;
161 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
162     "Amount of malloced memory for buffers");
163 static long maxbufmallocspace;
164 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
165     0, "Maximum amount of malloced memory for buffers");
166 static long lobufspace;
167 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RW, &lobufspace, 0,
168     "Minimum amount of buffers we want to have");
169 long hibufspace;
170 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RW, &hibufspace, 0,
171     "Maximum allowed value of bufspace (excluding metadata)");
172 long bufspacethresh;
173 SYSCTL_LONG(_vfs, OID_AUTO, bufspacethresh, CTLFLAG_RW, &bufspacethresh,
174     0, "Bufspace consumed before waking the daemon to free some");
175 static int buffreekvacnt;
176 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
177     "Number of times we have freed the KVA space from some buffer");
178 static int bufdefragcnt;
179 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
180     "Number of times we have had to repeat buffer allocation to defragment");
181 static long lorunningspace;
182 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
183     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
184     "Minimum preferred space used for in-progress I/O");
185 static long hirunningspace;
186 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
187     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
188     "Maximum amount of space to use for in-progress I/O");
189 int dirtybufferflushes;
190 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
191     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
192 int bdwriteskip;
193 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
194     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
195 int altbufferflushes;
196 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
197     0, "Number of fsync flushes to limit dirty buffers");
198 static int recursiveflushes;
199 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
200     0, "Number of flushes skipped due to being recursive");
201 static int numdirtybuffers;
202 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
203     "Number of buffers that are dirty (has unwritten changes) at the moment");
204 static int lodirtybuffers;
205 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
206     "How many buffers we want to have free before bufdaemon can sleep");
207 static int hidirtybuffers;
208 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
209     "When the number of dirty buffers is considered severe");
210 int dirtybufthresh;
211 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
212     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
213 static int numfreebuffers;
214 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
215     "Number of free buffers");
216 static int lofreebuffers;
217 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
218    "Target number of free buffers");
219 static int hifreebuffers;
220 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
221    "Threshold for clean buffer recycling");
222 static int getnewbufcalls;
223 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
224    "Number of calls to getnewbuf");
225 static int getnewbufrestarts;
226 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
227     "Number of times getnewbuf has had to restart a buffer aquisition");
228 static int mappingrestarts;
229 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
230     "Number of times getblk has had to restart a buffer mapping for "
231     "unmapped buffer");
232 static int numbufallocfails;
233 SYSCTL_INT(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, &numbufallocfails, 0,
234     "Number of times buffer allocations failed");
235 static int flushbufqtarget = 100;
236 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
237     "Amount of work to do in flushbufqueues when helping bufdaemon");
238 static long notbufdflushes;
239 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes, 0,
240     "Number of dirty buffer flushes done by the bufdaemon helpers");
241 static long barrierwrites;
242 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
243     "Number of barrier writes");
244 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
245     &unmapped_buf_allowed, 0,
246     "Permit the use of the unmapped i/o");
247 
248 /*
249  * This lock synchronizes access to bd_request.
250  */
251 static struct mtx_padalign bdlock;
252 
253 /*
254  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
255  * waitrunningbufspace().
256  */
257 static struct mtx_padalign rbreqlock;
258 
259 /*
260  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
261  */
262 static struct rwlock_padalign nblock;
263 
264 /*
265  * Lock that protects bdirtywait.
266  */
267 static struct mtx_padalign bdirtylock;
268 
269 /*
270  * Wakeup point for bufdaemon, as well as indicator of whether it is already
271  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
272  * is idling.
273  */
274 static int bd_request;
275 
276 /*
277  * Request/wakeup point for the bufspace daemon.
278  */
279 static int bufspace_request;
280 
281 /*
282  * Request for the buf daemon to write more buffers than is indicated by
283  * lodirtybuf.  This may be necessary to push out excess dependencies or
284  * defragment the address space where a simple count of the number of dirty
285  * buffers is insufficient to characterize the demand for flushing them.
286  */
287 static int bd_speedupreq;
288 
289 /*
290  * bogus page -- for I/O to/from partially complete buffers
291  * this is a temporary solution to the problem, but it is not
292  * really that bad.  it would be better to split the buffer
293  * for input in the case of buffers partially already in memory,
294  * but the code is intricate enough already.
295  */
296 vm_page_t bogus_page;
297 
298 /*
299  * Synchronization (sleep/wakeup) variable for active buffer space requests.
300  * Set when wait starts, cleared prior to wakeup().
301  * Used in runningbufwakeup() and waitrunningbufspace().
302  */
303 static int runningbufreq;
304 
305 /*
306  * Synchronization (sleep/wakeup) variable for buffer requests.
307  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
308  * by and/or.
309  * Used in numdirtywakeup(), bufspace_wakeup(), bwillwrite(),
310  * getnewbuf(), and getblk().
311  */
312 static volatile int needsbuffer;
313 
314 /*
315  * Synchronization for bwillwrite() waiters.
316  */
317 static int bdirtywait;
318 
319 /*
320  * Definitions for the buffer free lists.
321  */
322 #define QUEUE_NONE	0	/* on no queue */
323 #define QUEUE_EMPTY	1	/* empty buffer headers */
324 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
325 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
326 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
327 
328 /* Maximum number of clean buffer queues. */
329 #define	CLEAN_QUEUES	16
330 
331 /* Configured number of clean queues. */
332 static int clean_queues;
333 
334 /* Maximum number of buffer queues. */
335 #define BUFFER_QUEUES	(QUEUE_CLEAN + CLEAN_QUEUES)
336 
337 /* Queues for free buffers with various properties */
338 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
339 #ifdef INVARIANTS
340 static int bq_len[BUFFER_QUEUES];
341 #endif
342 
343 /*
344  * Lock for each bufqueue
345  */
346 static struct mtx_padalign bqlocks[BUFFER_QUEUES];
347 
348 /*
349  * per-cpu empty buffer cache.
350  */
351 uma_zone_t buf_zone;
352 
353 /*
354  * Single global constant for BUF_WMESG, to avoid getting multiple references.
355  * buf_wmesg is referred from macros.
356  */
357 const char *buf_wmesg = BUF_WMESG;
358 
359 static int
360 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
361 {
362 	long value;
363 	int error;
364 
365 	value = *(long *)arg1;
366 	error = sysctl_handle_long(oidp, &value, 0, req);
367 	if (error != 0 || req->newptr == NULL)
368 		return (error);
369 	mtx_lock(&rbreqlock);
370 	if (arg1 == &hirunningspace) {
371 		if (value < lorunningspace)
372 			error = EINVAL;
373 		else
374 			hirunningspace = value;
375 	} else {
376 		KASSERT(arg1 == &lorunningspace,
377 		    ("%s: unknown arg1", __func__));
378 		if (value > hirunningspace)
379 			error = EINVAL;
380 		else
381 			lorunningspace = value;
382 	}
383 	mtx_unlock(&rbreqlock);
384 	return (error);
385 }
386 
387 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
388     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
389 static int
390 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
391 {
392 	long lvalue;
393 	int ivalue;
394 
395 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
396 		return (sysctl_handle_long(oidp, arg1, arg2, req));
397 	lvalue = *(long *)arg1;
398 	if (lvalue > INT_MAX)
399 		/* On overflow, still write out a long to trigger ENOMEM. */
400 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
401 	ivalue = lvalue;
402 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
403 }
404 #endif
405 
406 static int
407 bqcleanq(void)
408 {
409 	static int nextq;
410 
411 	return ((atomic_fetchadd_int(&nextq, 1) % clean_queues) + QUEUE_CLEAN);
412 }
413 
414 static int
415 bqisclean(int qindex)
416 {
417 
418 	return (qindex >= QUEUE_CLEAN && qindex < QUEUE_CLEAN + CLEAN_QUEUES);
419 }
420 
421 /*
422  *	bqlock:
423  *
424  *	Return the appropriate queue lock based on the index.
425  */
426 static inline struct mtx *
427 bqlock(int qindex)
428 {
429 
430 	return (struct mtx *)&bqlocks[qindex];
431 }
432 
433 /*
434  *	bdirtywakeup:
435  *
436  *	Wakeup any bwillwrite() waiters.
437  */
438 static void
439 bdirtywakeup(void)
440 {
441 	mtx_lock(&bdirtylock);
442 	if (bdirtywait) {
443 		bdirtywait = 0;
444 		wakeup(&bdirtywait);
445 	}
446 	mtx_unlock(&bdirtylock);
447 }
448 
449 /*
450  *	bdirtysub:
451  *
452  *	Decrement the numdirtybuffers count by one and wakeup any
453  *	threads blocked in bwillwrite().
454  */
455 static void
456 bdirtysub(void)
457 {
458 
459 	if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
460 	    (lodirtybuffers + hidirtybuffers) / 2)
461 		bdirtywakeup();
462 }
463 
464 /*
465  *	bdirtyadd:
466  *
467  *	Increment the numdirtybuffers count by one and wakeup the buf
468  *	daemon if needed.
469  */
470 static void
471 bdirtyadd(void)
472 {
473 
474 	/*
475 	 * Only do the wakeup once as we cross the boundary.  The
476 	 * buf daemon will keep running until the condition clears.
477 	 */
478 	if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
479 	    (lodirtybuffers + hidirtybuffers) / 2)
480 		bd_wakeup();
481 }
482 
483 /*
484  *	bufspace_wakeup:
485  *
486  *	Called when buffer space is potentially available for recovery.
487  *	getnewbuf() will block on this flag when it is unable to free
488  *	sufficient buffer space.  Buffer space becomes recoverable when
489  *	bp's get placed back in the queues.
490  */
491 static void
492 bufspace_wakeup(void)
493 {
494 
495 	/*
496 	 * If someone is waiting for bufspace, wake them up.
497 	 *
498 	 * Since needsbuffer is set prior to doing an additional queue
499 	 * scan it is safe to check for the flag prior to acquiring the
500 	 * lock.  The thread that is preparing to scan again before
501 	 * blocking would discover the buf we released.
502 	 */
503 	if (needsbuffer) {
504 		rw_rlock(&nblock);
505 		if (atomic_cmpset_int(&needsbuffer, 1, 0) == 1)
506 			wakeup(__DEVOLATILE(void *, &needsbuffer));
507 		rw_runlock(&nblock);
508 	}
509 }
510 
511 /*
512  *	bufspace_daemonwakeup:
513  *
514  *	Wakeup the daemon responsible for freeing clean bufs.
515  */
516 static void
517 bufspace_daemonwakeup(void)
518 {
519 	rw_rlock(&nblock);
520 	if (bufspace_request == 0) {
521 		bufspace_request = 1;
522 		wakeup(&bufspace_request);
523 	}
524 	rw_runlock(&nblock);
525 }
526 
527 /*
528  *	bufspace_adjust:
529  *
530  *	Adjust the reported bufspace for a KVA managed buffer, possibly
531  * 	waking any waiters.
532  */
533 static void
534 bufspace_adjust(struct buf *bp, int bufsize)
535 {
536 	long space;
537 	int diff;
538 
539 	KASSERT((bp->b_flags & B_MALLOC) == 0,
540 	    ("bufspace_adjust: malloc buf %p", bp));
541 	diff = bufsize - bp->b_bufsize;
542 	if (diff < 0) {
543 		atomic_subtract_long(&bufspace, -diff);
544 		bufspace_wakeup();
545 	} else {
546 		space = atomic_fetchadd_long(&bufspace, diff);
547 		/* Wake up the daemon on the transition. */
548 		if (space < bufspacethresh && space + diff >= bufspacethresh)
549 			bufspace_daemonwakeup();
550 	}
551 	bp->b_bufsize = bufsize;
552 }
553 
554 /*
555  *	bufspace_reserve:
556  *
557  *	Reserve bufspace before calling allocbuf().  metadata has a
558  *	different space limit than data.
559  */
560 static int
561 bufspace_reserve(int size, bool metadata)
562 {
563 	long limit;
564 	long space;
565 
566 	if (metadata)
567 		limit = maxbufspace;
568 	else
569 		limit = hibufspace;
570 	do {
571 		space = bufspace;
572 		if (space + size > limit)
573 			return (ENOSPC);
574 	} while (atomic_cmpset_long(&bufspace, space, space + size) == 0);
575 
576 	/* Wake up the daemon on the transition. */
577 	if (space < bufspacethresh && space + size >= bufspacethresh)
578 		bufspace_daemonwakeup();
579 
580 	return (0);
581 }
582 
583 /*
584  *	bufspace_release:
585  *
586  *	Release reserved bufspace after bufspace_adjust() has consumed it.
587  */
588 static void
589 bufspace_release(int size)
590 {
591 	atomic_subtract_long(&bufspace, size);
592 	bufspace_wakeup();
593 }
594 
595 /*
596  *	bufspace_wait:
597  *
598  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
599  *	supplied.  needsbuffer must be set in a safe fashion prior to
600  *	polling for space.  The operation must be re-tried on return.
601  */
602 static void
603 bufspace_wait(struct vnode *vp, int gbflags, int slpflag, int slptimeo)
604 {
605 	struct thread *td;
606 	int error, fl, norunbuf;
607 
608 	if ((gbflags & GB_NOWAIT_BD) != 0)
609 		return;
610 
611 	td = curthread;
612 	rw_wlock(&nblock);
613 	while (needsbuffer != 0) {
614 		if (vp != NULL && vp->v_type != VCHR &&
615 		    (td->td_pflags & TDP_BUFNEED) == 0) {
616 			rw_wunlock(&nblock);
617 			/*
618 			 * getblk() is called with a vnode locked, and
619 			 * some majority of the dirty buffers may as
620 			 * well belong to the vnode.  Flushing the
621 			 * buffers there would make a progress that
622 			 * cannot be achieved by the buf_daemon, that
623 			 * cannot lock the vnode.
624 			 */
625 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
626 			    (td->td_pflags & TDP_NORUNNINGBUF);
627 
628 			/*
629 			 * Play bufdaemon.  The getnewbuf() function
630 			 * may be called while the thread owns lock
631 			 * for another dirty buffer for the same
632 			 * vnode, which makes it impossible to use
633 			 * VOP_FSYNC() there, due to the buffer lock
634 			 * recursion.
635 			 */
636 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
637 			fl = buf_flush(vp, flushbufqtarget);
638 			td->td_pflags &= norunbuf;
639 			rw_wlock(&nblock);
640 			if (fl != 0)
641 				continue;
642 			if (needsbuffer == 0)
643 				break;
644 		}
645 		error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock,
646 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
647 		if (error != 0)
648 			break;
649 	}
650 	rw_wunlock(&nblock);
651 }
652 
653 
654 /*
655  *	bufspace_daemon:
656  *
657  *	buffer space management daemon.  Tries to maintain some marginal
658  *	amount of free buffer space so that requesting processes neither
659  *	block nor work to reclaim buffers.
660  */
661 static void
662 bufspace_daemon(void)
663 {
664 	for (;;) {
665 		kproc_suspend_check(bufspacedaemonproc);
666 
667 		/*
668 		 * Free buffers from the clean queue until we meet our
669 		 * targets.
670 		 *
671 		 * Theory of operation:  The buffer cache is most efficient
672 		 * when some free buffer headers and space are always
673 		 * available to getnewbuf().  This daemon attempts to prevent
674 		 * the excessive blocking and synchronization associated
675 		 * with shortfall.  It goes through three phases according
676 		 * demand:
677 		 *
678 		 * 1)	The daemon wakes up voluntarily once per-second
679 		 *	during idle periods when the counters are below
680 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
681 		 *
682 		 * 2)	The daemon wakes up as we cross the thresholds
683 		 *	ahead of any potential blocking.  This may bounce
684 		 *	slightly according to the rate of consumption and
685 		 *	release.
686 		 *
687 		 * 3)	The daemon and consumers are starved for working
688 		 *	clean buffers.  This is the 'bufspace' sleep below
689 		 *	which will inefficiently trade bufs with bqrelse
690 		 *	until we return to condition 2.
691 		 */
692 		while (bufspace > lobufspace ||
693 		    numfreebuffers < hifreebuffers) {
694 			if (buf_recycle(false) != 0) {
695 				atomic_set_int(&needsbuffer, 1);
696 				if (buf_recycle(false) != 0) {
697 					rw_wlock(&nblock);
698 					if (needsbuffer)
699 						rw_sleep(__DEVOLATILE(void *,
700 						    &needsbuffer), &nblock,
701 						    PRIBIO|PDROP, "bufspace",
702 						    hz/10);
703 					else
704 						rw_wunlock(&nblock);
705 				}
706 			}
707 			maybe_yield();
708 		}
709 
710 		/*
711 		 * Re-check our limits under the exclusive nblock.
712 		 */
713 		rw_wlock(&nblock);
714 		if (bufspace < bufspacethresh &&
715 		    numfreebuffers > lofreebuffers) {
716 			bufspace_request = 0;
717 			rw_sleep(&bufspace_request, &nblock, PRIBIO|PDROP,
718 			    "-", hz);
719 		} else
720 			rw_wunlock(&nblock);
721 	}
722 }
723 
724 static struct kproc_desc bufspace_kp = {
725 	"bufspacedaemon",
726 	bufspace_daemon,
727 	&bufspacedaemonproc
728 };
729 SYSINIT(bufspacedaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start,
730     &bufspace_kp);
731 
732 /*
733  *	bufmallocadjust:
734  *
735  *	Adjust the reported bufspace for a malloc managed buffer, possibly
736  *	waking any waiters.
737  */
738 static void
739 bufmallocadjust(struct buf *bp, int bufsize)
740 {
741 	int diff;
742 
743 	KASSERT((bp->b_flags & B_MALLOC) != 0,
744 	    ("bufmallocadjust: non-malloc buf %p", bp));
745 	diff = bufsize - bp->b_bufsize;
746 	if (diff < 0)
747 		atomic_subtract_long(&bufmallocspace, -diff);
748 	else
749 		atomic_add_long(&bufmallocspace, diff);
750 	bp->b_bufsize = bufsize;
751 }
752 
753 /*
754  *	runningwakeup:
755  *
756  *	Wake up processes that are waiting on asynchronous writes to fall
757  *	below lorunningspace.
758  */
759 static void
760 runningwakeup(void)
761 {
762 
763 	mtx_lock(&rbreqlock);
764 	if (runningbufreq) {
765 		runningbufreq = 0;
766 		wakeup(&runningbufreq);
767 	}
768 	mtx_unlock(&rbreqlock);
769 }
770 
771 /*
772  *	runningbufwakeup:
773  *
774  *	Decrement the outstanding write count according.
775  */
776 void
777 runningbufwakeup(struct buf *bp)
778 {
779 	long space, bspace;
780 
781 	bspace = bp->b_runningbufspace;
782 	if (bspace == 0)
783 		return;
784 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
785 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
786 	    space, bspace));
787 	bp->b_runningbufspace = 0;
788 	/*
789 	 * Only acquire the lock and wakeup on the transition from exceeding
790 	 * the threshold to falling below it.
791 	 */
792 	if (space < lorunningspace)
793 		return;
794 	if (space - bspace > lorunningspace)
795 		return;
796 	runningwakeup();
797 }
798 
799 /*
800  *	waitrunningbufspace()
801  *
802  *	runningbufspace is a measure of the amount of I/O currently
803  *	running.  This routine is used in async-write situations to
804  *	prevent creating huge backups of pending writes to a device.
805  *	Only asynchronous writes are governed by this function.
806  *
807  *	This does NOT turn an async write into a sync write.  It waits
808  *	for earlier writes to complete and generally returns before the
809  *	caller's write has reached the device.
810  */
811 void
812 waitrunningbufspace(void)
813 {
814 
815 	mtx_lock(&rbreqlock);
816 	while (runningbufspace > hirunningspace) {
817 		runningbufreq = 1;
818 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
819 	}
820 	mtx_unlock(&rbreqlock);
821 }
822 
823 
824 /*
825  *	vfs_buf_test_cache:
826  *
827  *	Called when a buffer is extended.  This function clears the B_CACHE
828  *	bit if the newly extended portion of the buffer does not contain
829  *	valid data.
830  */
831 static __inline void
832 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
833     vm_offset_t size, vm_page_t m)
834 {
835 
836 	VM_OBJECT_ASSERT_LOCKED(m->object);
837 	if (bp->b_flags & B_CACHE) {
838 		int base = (foff + off) & PAGE_MASK;
839 		if (vm_page_is_valid(m, base, size) == 0)
840 			bp->b_flags &= ~B_CACHE;
841 	}
842 }
843 
844 /* Wake up the buffer daemon if necessary */
845 static __inline void
846 bd_wakeup(void)
847 {
848 
849 	mtx_lock(&bdlock);
850 	if (bd_request == 0) {
851 		bd_request = 1;
852 		wakeup(&bd_request);
853 	}
854 	mtx_unlock(&bdlock);
855 }
856 
857 /*
858  * bd_speedup - speedup the buffer cache flushing code
859  */
860 void
861 bd_speedup(void)
862 {
863 	int needwake;
864 
865 	mtx_lock(&bdlock);
866 	needwake = 0;
867 	if (bd_speedupreq == 0 || bd_request == 0)
868 		needwake = 1;
869 	bd_speedupreq = 1;
870 	bd_request = 1;
871 	if (needwake)
872 		wakeup(&bd_request);
873 	mtx_unlock(&bdlock);
874 }
875 
876 #ifndef NSWBUF_MIN
877 #define	NSWBUF_MIN	16
878 #endif
879 
880 #ifdef __i386__
881 #define	TRANSIENT_DENOM	5
882 #else
883 #define	TRANSIENT_DENOM 10
884 #endif
885 
886 /*
887  * Calculating buffer cache scaling values and reserve space for buffer
888  * headers.  This is called during low level kernel initialization and
889  * may be called more then once.  We CANNOT write to the memory area
890  * being reserved at this time.
891  */
892 caddr_t
893 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
894 {
895 	int tuned_nbuf;
896 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
897 
898 	/*
899 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
900 	 * PAGE_SIZE is >= 1K)
901 	 */
902 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
903 
904 	/*
905 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
906 	 * For the first 64MB of ram nominally allocate sufficient buffers to
907 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
908 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
909 	 * the buffer cache we limit the eventual kva reservation to
910 	 * maxbcache bytes.
911 	 *
912 	 * factor represents the 1/4 x ram conversion.
913 	 */
914 	if (nbuf == 0) {
915 		int factor = 4 * BKVASIZE / 1024;
916 
917 		nbuf = 50;
918 		if (physmem_est > 4096)
919 			nbuf += min((physmem_est - 4096) / factor,
920 			    65536 / factor);
921 		if (physmem_est > 65536)
922 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
923 			    32 * 1024 * 1024 / (factor * 5));
924 
925 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
926 			nbuf = maxbcache / BKVASIZE;
927 		tuned_nbuf = 1;
928 	} else
929 		tuned_nbuf = 0;
930 
931 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
932 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
933 	if (nbuf > maxbuf) {
934 		if (!tuned_nbuf)
935 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
936 			    maxbuf);
937 		nbuf = maxbuf;
938 	}
939 
940 	/*
941 	 * Ideal allocation size for the transient bio submap is 10%
942 	 * of the maximal space buffer map.  This roughly corresponds
943 	 * to the amount of the buffer mapped for typical UFS load.
944 	 *
945 	 * Clip the buffer map to reserve space for the transient
946 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
947 	 * maximum buffer map extent on the platform.
948 	 *
949 	 * The fall-back to the maxbuf in case of maxbcache unset,
950 	 * allows to not trim the buffer KVA for the architectures
951 	 * with ample KVA space.
952 	 */
953 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
954 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
955 		buf_sz = (long)nbuf * BKVASIZE;
956 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
957 		    (TRANSIENT_DENOM - 1)) {
958 			/*
959 			 * There is more KVA than memory.  Do not
960 			 * adjust buffer map size, and assign the rest
961 			 * of maxbuf to transient map.
962 			 */
963 			biotmap_sz = maxbuf_sz - buf_sz;
964 		} else {
965 			/*
966 			 * Buffer map spans all KVA we could afford on
967 			 * this platform.  Give 10% (20% on i386) of
968 			 * the buffer map to the transient bio map.
969 			 */
970 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
971 			buf_sz -= biotmap_sz;
972 		}
973 		if (biotmap_sz / INT_MAX > MAXPHYS)
974 			bio_transient_maxcnt = INT_MAX;
975 		else
976 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
977 		/*
978 		 * Artifically limit to 1024 simultaneous in-flight I/Os
979 		 * using the transient mapping.
980 		 */
981 		if (bio_transient_maxcnt > 1024)
982 			bio_transient_maxcnt = 1024;
983 		if (tuned_nbuf)
984 			nbuf = buf_sz / BKVASIZE;
985 	}
986 
987 	/*
988 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
989 	 * We have no less then 16 and no more then 256.
990 	 */
991 	nswbuf = min(nbuf / 4, 256);
992 	TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf);
993 	if (nswbuf < NSWBUF_MIN)
994 		nswbuf = NSWBUF_MIN;
995 
996 	/*
997 	 * Reserve space for the buffer cache buffers
998 	 */
999 	swbuf = (void *)v;
1000 	v = (caddr_t)(swbuf + nswbuf);
1001 	buf = (void *)v;
1002 	v = (caddr_t)(buf + nbuf);
1003 
1004 	return(v);
1005 }
1006 
1007 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1008 void
1009 bufinit(void)
1010 {
1011 	struct buf *bp;
1012 	int i;
1013 
1014 	CTASSERT(MAXBCACHEBUF >= MAXBSIZE);
1015 	mtx_init(&bqlocks[QUEUE_DIRTY], "bufq dirty lock", NULL, MTX_DEF);
1016 	mtx_init(&bqlocks[QUEUE_EMPTY], "bufq empty lock", NULL, MTX_DEF);
1017 	for (i = QUEUE_CLEAN; i < QUEUE_CLEAN + CLEAN_QUEUES; i++)
1018 		mtx_init(&bqlocks[i], "bufq clean lock", NULL, MTX_DEF);
1019 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1020 	rw_init(&nblock, "needsbuffer lock");
1021 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1022 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1023 
1024 	/* next, make a null set of free lists */
1025 	for (i = 0; i < BUFFER_QUEUES; i++)
1026 		TAILQ_INIT(&bufqueues[i]);
1027 
1028 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1029 
1030 	/* finally, initialize each buffer header and stick on empty q */
1031 	for (i = 0; i < nbuf; i++) {
1032 		bp = &buf[i];
1033 		bzero(bp, sizeof *bp);
1034 		bp->b_flags = B_INVAL;
1035 		bp->b_rcred = NOCRED;
1036 		bp->b_wcred = NOCRED;
1037 		bp->b_qindex = QUEUE_EMPTY;
1038 		bp->b_xflags = 0;
1039 		bp->b_data = bp->b_kvabase = unmapped_buf;
1040 		LIST_INIT(&bp->b_dep);
1041 		BUF_LOCKINIT(bp);
1042 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
1043 #ifdef INVARIANTS
1044 		bq_len[QUEUE_EMPTY]++;
1045 #endif
1046 	}
1047 
1048 	/*
1049 	 * maxbufspace is the absolute maximum amount of buffer space we are
1050 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1051 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1052 	 * used by most other requests.  The differential is required to
1053 	 * ensure that metadata deadlocks don't occur.
1054 	 *
1055 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1056 	 * this may result in KVM fragmentation which is not handled optimally
1057 	 * by the system. XXX This is less true with vmem.  We could use
1058 	 * PAGE_SIZE.
1059 	 */
1060 	maxbufspace = (long)nbuf * BKVASIZE;
1061 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBCACHEBUF * 10);
1062 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1063 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1064 
1065 	/*
1066 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1067 	 * arbitrarily and may need further tuning. It corresponds to
1068 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1069 	 * which fits with many RAID controllers' tagged queuing limits.
1070 	 * The lower 1 MiB limit is the historical upper limit for
1071 	 * hirunningspace.
1072 	 */
1073 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBCACHEBUF),
1074 	    16 * 1024 * 1024), 1024 * 1024);
1075 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBCACHEBUF);
1076 
1077 	/*
1078 	 * Limit the amount of malloc memory since it is wired permanently into
1079 	 * the kernel space.  Even though this is accounted for in the buffer
1080 	 * allocation, we don't want the malloced region to grow uncontrolled.
1081 	 * The malloc scheme improves memory utilization significantly on
1082 	 * average (small) directories.
1083 	 */
1084 	maxbufmallocspace = hibufspace / 20;
1085 
1086 	/*
1087 	 * Reduce the chance of a deadlock occuring by limiting the number
1088 	 * of delayed-write dirty buffers we allow to stack up.
1089 	 */
1090 	hidirtybuffers = nbuf / 4 + 20;
1091 	dirtybufthresh = hidirtybuffers * 9 / 10;
1092 	numdirtybuffers = 0;
1093 	/*
1094 	 * To support extreme low-memory systems, make sure hidirtybuffers
1095 	 * cannot eat up all available buffer space.  This occurs when our
1096 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1097 	 * buffer space assuming BKVASIZE'd buffers.
1098 	 */
1099 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1100 		hidirtybuffers >>= 1;
1101 	}
1102 	lodirtybuffers = hidirtybuffers / 2;
1103 
1104 	/*
1105 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1106 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1107 	 * are counted as free but will be unavailable to threads executing
1108 	 * on other cpus.
1109 	 *
1110 	 * hifreebuffers is the free target for the bufspace daemon.  This
1111 	 * should be set appropriately to limit work per-iteration.
1112 	 */
1113 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1114 	hifreebuffers = (3 * lofreebuffers) / 2;
1115 	numfreebuffers = nbuf;
1116 
1117 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
1118 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
1119 
1120 	/* Setup the kva and free list allocators. */
1121 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1122 	buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1123 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1124 
1125 	/*
1126 	 * Size the clean queue according to the amount of buffer space.
1127 	 * One queue per-256mb up to the max.  More queues gives better
1128 	 * concurrency but less accurate LRU.
1129 	 */
1130 	clean_queues = MIN(howmany(maxbufspace, 256*1024*1024), CLEAN_QUEUES);
1131 
1132 }
1133 
1134 #ifdef INVARIANTS
1135 static inline void
1136 vfs_buf_check_mapped(struct buf *bp)
1137 {
1138 
1139 	KASSERT(bp->b_kvabase != unmapped_buf,
1140 	    ("mapped buf: b_kvabase was not updated %p", bp));
1141 	KASSERT(bp->b_data != unmapped_buf,
1142 	    ("mapped buf: b_data was not updated %p", bp));
1143 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1144 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1145 }
1146 
1147 static inline void
1148 vfs_buf_check_unmapped(struct buf *bp)
1149 {
1150 
1151 	KASSERT(bp->b_data == unmapped_buf,
1152 	    ("unmapped buf: corrupted b_data %p", bp));
1153 }
1154 
1155 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1156 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1157 #else
1158 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1159 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1160 #endif
1161 
1162 static int
1163 isbufbusy(struct buf *bp)
1164 {
1165 	if (((bp->b_flags & (B_INVAL | B_PERSISTENT)) == 0 &&
1166 	    BUF_ISLOCKED(bp)) ||
1167 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1168 		return (1);
1169 	return (0);
1170 }
1171 
1172 /*
1173  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1174  */
1175 void
1176 bufshutdown(int show_busybufs)
1177 {
1178 	static int first_buf_printf = 1;
1179 	struct buf *bp;
1180 	int iter, nbusy, pbusy;
1181 #ifndef PREEMPTION
1182 	int subiter;
1183 #endif
1184 
1185 	/*
1186 	 * Sync filesystems for shutdown
1187 	 */
1188 	wdog_kern_pat(WD_LASTVAL);
1189 	sys_sync(curthread, NULL);
1190 
1191 	/*
1192 	 * With soft updates, some buffers that are
1193 	 * written will be remarked as dirty until other
1194 	 * buffers are written.
1195 	 */
1196 	for (iter = pbusy = 0; iter < 20; iter++) {
1197 		nbusy = 0;
1198 		for (bp = &buf[nbuf]; --bp >= buf; )
1199 			if (isbufbusy(bp))
1200 				nbusy++;
1201 		if (nbusy == 0) {
1202 			if (first_buf_printf)
1203 				printf("All buffers synced.");
1204 			break;
1205 		}
1206 		if (first_buf_printf) {
1207 			printf("Syncing disks, buffers remaining... ");
1208 			first_buf_printf = 0;
1209 		}
1210 		printf("%d ", nbusy);
1211 		if (nbusy < pbusy)
1212 			iter = 0;
1213 		pbusy = nbusy;
1214 
1215 		wdog_kern_pat(WD_LASTVAL);
1216 		sys_sync(curthread, NULL);
1217 
1218 #ifdef PREEMPTION
1219 		/*
1220 		 * Drop Giant and spin for a while to allow
1221 		 * interrupt threads to run.
1222 		 */
1223 		DROP_GIANT();
1224 		DELAY(50000 * iter);
1225 		PICKUP_GIANT();
1226 #else
1227 		/*
1228 		 * Drop Giant and context switch several times to
1229 		 * allow interrupt threads to run.
1230 		 */
1231 		DROP_GIANT();
1232 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1233 			thread_lock(curthread);
1234 			mi_switch(SW_VOL, NULL);
1235 			thread_unlock(curthread);
1236 			DELAY(1000);
1237 		}
1238 		PICKUP_GIANT();
1239 #endif
1240 	}
1241 	printf("\n");
1242 	/*
1243 	 * Count only busy local buffers to prevent forcing
1244 	 * a fsck if we're just a client of a wedged NFS server
1245 	 */
1246 	nbusy = 0;
1247 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1248 		if (isbufbusy(bp)) {
1249 #if 0
1250 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1251 			if (bp->b_dev == NULL) {
1252 				TAILQ_REMOVE(&mountlist,
1253 				    bp->b_vp->v_mount, mnt_list);
1254 				continue;
1255 			}
1256 #endif
1257 			nbusy++;
1258 			if (show_busybufs > 0) {
1259 				printf(
1260 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1261 				    nbusy, bp, bp->b_vp, bp->b_flags,
1262 				    (intmax_t)bp->b_blkno,
1263 				    (intmax_t)bp->b_lblkno);
1264 				BUF_LOCKPRINTINFO(bp);
1265 				if (show_busybufs > 1)
1266 					vn_printf(bp->b_vp,
1267 					    "vnode content: ");
1268 			}
1269 		}
1270 	}
1271 	if (nbusy) {
1272 		/*
1273 		 * Failed to sync all blocks. Indicate this and don't
1274 		 * unmount filesystems (thus forcing an fsck on reboot).
1275 		 */
1276 		printf("Giving up on %d buffers\n", nbusy);
1277 		DELAY(5000000);	/* 5 seconds */
1278 	} else {
1279 		if (!first_buf_printf)
1280 			printf("Final sync complete\n");
1281 		/*
1282 		 * Unmount filesystems
1283 		 */
1284 		if (panicstr == NULL)
1285 			vfs_unmountall();
1286 	}
1287 	swapoff_all();
1288 	DELAY(100000);		/* wait for console output to finish */
1289 }
1290 
1291 static void
1292 bpmap_qenter(struct buf *bp)
1293 {
1294 
1295 	BUF_CHECK_MAPPED(bp);
1296 
1297 	/*
1298 	 * bp->b_data is relative to bp->b_offset, but
1299 	 * bp->b_offset may be offset into the first page.
1300 	 */
1301 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1302 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1303 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1304 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1305 }
1306 
1307 /*
1308  *	binsfree:
1309  *
1310  *	Insert the buffer into the appropriate free list.
1311  */
1312 static void
1313 binsfree(struct buf *bp, int qindex)
1314 {
1315 	struct mtx *olock, *nlock;
1316 
1317 	if (qindex != QUEUE_EMPTY) {
1318 		BUF_ASSERT_XLOCKED(bp);
1319 	}
1320 
1321 	/*
1322 	 * Stick to the same clean queue for the lifetime of the buf to
1323 	 * limit locking below.  Otherwise pick ont sequentially.
1324 	 */
1325 	if (qindex == QUEUE_CLEAN) {
1326 		if (bqisclean(bp->b_qindex))
1327 			qindex = bp->b_qindex;
1328 		else
1329 			qindex = bqcleanq();
1330 	}
1331 
1332 	/*
1333 	 * Handle delayed bremfree() processing.
1334 	 */
1335 	nlock = bqlock(qindex);
1336 	if (bp->b_flags & B_REMFREE) {
1337 		olock = bqlock(bp->b_qindex);
1338 		mtx_lock(olock);
1339 		bremfreel(bp);
1340 		if (olock != nlock) {
1341 			mtx_unlock(olock);
1342 			mtx_lock(nlock);
1343 		}
1344 	} else
1345 		mtx_lock(nlock);
1346 
1347 	if (bp->b_qindex != QUEUE_NONE)
1348 		panic("binsfree: free buffer onto another queue???");
1349 
1350 	bp->b_qindex = qindex;
1351 	if (bp->b_flags & B_AGE)
1352 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1353 	else
1354 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1355 #ifdef INVARIANTS
1356 	bq_len[bp->b_qindex]++;
1357 #endif
1358 	mtx_unlock(nlock);
1359 }
1360 
1361 /*
1362  * buf_free:
1363  *
1364  *	Free a buffer to the buf zone once it no longer has valid contents.
1365  */
1366 static void
1367 buf_free(struct buf *bp)
1368 {
1369 
1370 	if (bp->b_flags & B_REMFREE)
1371 		bremfreef(bp);
1372 	if (bp->b_vflags & BV_BKGRDINPROG)
1373 		panic("losing buffer 1");
1374 	if (bp->b_rcred != NOCRED) {
1375 		crfree(bp->b_rcred);
1376 		bp->b_rcred = NOCRED;
1377 	}
1378 	if (bp->b_wcred != NOCRED) {
1379 		crfree(bp->b_wcred);
1380 		bp->b_wcred = NOCRED;
1381 	}
1382 	if (!LIST_EMPTY(&bp->b_dep))
1383 		buf_deallocate(bp);
1384 	bufkva_free(bp);
1385 	BUF_UNLOCK(bp);
1386 	uma_zfree(buf_zone, bp);
1387 	atomic_add_int(&numfreebuffers, 1);
1388 	bufspace_wakeup();
1389 }
1390 
1391 /*
1392  * buf_import:
1393  *
1394  *	Import bufs into the uma cache from the buf list.  The system still
1395  *	expects a static array of bufs and much of the synchronization
1396  *	around bufs assumes type stable storage.  As a result, UMA is used
1397  *	only as a per-cpu cache of bufs still maintained on a global list.
1398  */
1399 static int
1400 buf_import(void *arg, void **store, int cnt, int flags)
1401 {
1402 	struct buf *bp;
1403 	int i;
1404 
1405 	mtx_lock(&bqlocks[QUEUE_EMPTY]);
1406 	for (i = 0; i < cnt; i++) {
1407 		bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1408 		if (bp == NULL)
1409 			break;
1410 		bremfreel(bp);
1411 		store[i] = bp;
1412 	}
1413 	mtx_unlock(&bqlocks[QUEUE_EMPTY]);
1414 
1415 	return (i);
1416 }
1417 
1418 /*
1419  * buf_release:
1420  *
1421  *	Release bufs from the uma cache back to the buffer queues.
1422  */
1423 static void
1424 buf_release(void *arg, void **store, int cnt)
1425 {
1426         int i;
1427 
1428         for (i = 0; i < cnt; i++)
1429 		binsfree(store[i], QUEUE_EMPTY);
1430 }
1431 
1432 /*
1433  * buf_alloc:
1434  *
1435  *	Allocate an empty buffer header.
1436  */
1437 static struct buf *
1438 buf_alloc(void)
1439 {
1440 	struct buf *bp;
1441 
1442 	bp = uma_zalloc(buf_zone, M_NOWAIT);
1443 	if (bp == NULL) {
1444 		bufspace_daemonwakeup();
1445 		atomic_add_int(&numbufallocfails, 1);
1446 		return (NULL);
1447 	}
1448 
1449 	/*
1450 	 * Wake-up the bufspace daemon on transition.
1451 	 */
1452 	if (atomic_fetchadd_int(&numfreebuffers, -1) == lofreebuffers)
1453 		bufspace_daemonwakeup();
1454 
1455 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1456 		panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1457 
1458 	KASSERT(bp->b_vp == NULL,
1459 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1460 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1461 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1462 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1463 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1464 	KASSERT(bp->b_npages == 0,
1465 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1466 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1467 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1468 
1469 	bp->b_flags = 0;
1470 	bp->b_ioflags = 0;
1471 	bp->b_xflags = 0;
1472 	bp->b_vflags = 0;
1473 	bp->b_vp = NULL;
1474 	bp->b_blkno = bp->b_lblkno = 0;
1475 	bp->b_offset = NOOFFSET;
1476 	bp->b_iodone = 0;
1477 	bp->b_error = 0;
1478 	bp->b_resid = 0;
1479 	bp->b_bcount = 0;
1480 	bp->b_npages = 0;
1481 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1482 	bp->b_bufobj = NULL;
1483 	bp->b_pin_count = 0;
1484 	bp->b_data = bp->b_kvabase = unmapped_buf;
1485 	bp->b_fsprivate1 = NULL;
1486 	bp->b_fsprivate2 = NULL;
1487 	bp->b_fsprivate3 = NULL;
1488 	LIST_INIT(&bp->b_dep);
1489 
1490 	return (bp);
1491 }
1492 
1493 /*
1494  *	buf_qrecycle:
1495  *
1496  *	Free a buffer from the given bufqueue.  kva controls whether the
1497  *	freed buf must own some kva resources.  This is used for
1498  *	defragmenting.
1499  */
1500 static int
1501 buf_qrecycle(int qindex, bool kva)
1502 {
1503 	struct buf *bp, *nbp;
1504 
1505 	if (kva)
1506 		atomic_add_int(&bufdefragcnt, 1);
1507 	nbp = NULL;
1508 	mtx_lock(&bqlocks[qindex]);
1509 	nbp = TAILQ_FIRST(&bufqueues[qindex]);
1510 
1511 	/*
1512 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1513 	 * depending.
1514 	 */
1515 	while ((bp = nbp) != NULL) {
1516 		/*
1517 		 * Calculate next bp (we can only use it if we do not
1518 		 * release the bqlock).
1519 		 */
1520 		nbp = TAILQ_NEXT(bp, b_freelist);
1521 
1522 		/*
1523 		 * If we are defragging then we need a buffer with
1524 		 * some kva to reclaim.
1525 		 */
1526 		if (kva && bp->b_kvasize == 0)
1527 			continue;
1528 
1529 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1530 			continue;
1531 
1532 		/*
1533 		 * Skip buffers with background writes in progress.
1534 		 */
1535 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1536 			BUF_UNLOCK(bp);
1537 			continue;
1538 		}
1539 
1540 		KASSERT(bp->b_qindex == qindex,
1541 		    ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1542 		/*
1543 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1544 		 * the scan from this point on.
1545 		 */
1546 		bremfreel(bp);
1547 		mtx_unlock(&bqlocks[qindex]);
1548 
1549 		/*
1550 		 * Requeue the background write buffer with error and
1551 		 * restart the scan.
1552 		 */
1553 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1554 			bqrelse(bp);
1555 			mtx_lock(&bqlocks[qindex]);
1556 			nbp = TAILQ_FIRST(&bufqueues[qindex]);
1557 			continue;
1558 		}
1559 		bp->b_flags |= B_INVAL;
1560 		brelse(bp);
1561 		return (0);
1562 	}
1563 	mtx_unlock(&bqlocks[qindex]);
1564 
1565 	return (ENOBUFS);
1566 }
1567 
1568 /*
1569  *	buf_recycle:
1570  *
1571  *	Iterate through all clean queues until we find a buf to recycle or
1572  *	exhaust the search.
1573  */
1574 static int
1575 buf_recycle(bool kva)
1576 {
1577 	int qindex, first_qindex;
1578 
1579 	qindex = first_qindex = bqcleanq();
1580 	do {
1581 		if (buf_qrecycle(qindex, kva) == 0)
1582 			return (0);
1583 		if (++qindex == QUEUE_CLEAN + clean_queues)
1584 			qindex = QUEUE_CLEAN;
1585 	} while (qindex != first_qindex);
1586 
1587 	return (ENOBUFS);
1588 }
1589 
1590 /*
1591  *	buf_scan:
1592  *
1593  *	Scan the clean queues looking for a buffer to recycle.  needsbuffer
1594  *	is set on failure so that the caller may optionally bufspace_wait()
1595  *	in a race-free fashion.
1596  */
1597 static int
1598 buf_scan(bool defrag)
1599 {
1600 	int error;
1601 
1602 	/*
1603 	 * To avoid heavy synchronization and wakeup races we set
1604 	 * needsbuffer and re-poll before failing.  This ensures that
1605 	 * no frees can be missed between an unsuccessful poll and
1606 	 * going to sleep in a synchronized fashion.
1607 	 */
1608 	if ((error = buf_recycle(defrag)) != 0) {
1609 		atomic_set_int(&needsbuffer, 1);
1610 		bufspace_daemonwakeup();
1611 		error = buf_recycle(defrag);
1612 	}
1613 	if (error == 0)
1614 		atomic_add_int(&getnewbufrestarts, 1);
1615 	return (error);
1616 }
1617 
1618 /*
1619  *	bremfree:
1620  *
1621  *	Mark the buffer for removal from the appropriate free list.
1622  *
1623  */
1624 void
1625 bremfree(struct buf *bp)
1626 {
1627 
1628 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1629 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1630 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1631 	KASSERT(bp->b_qindex != QUEUE_NONE,
1632 	    ("bremfree: buffer %p not on a queue.", bp));
1633 	BUF_ASSERT_XLOCKED(bp);
1634 
1635 	bp->b_flags |= B_REMFREE;
1636 }
1637 
1638 /*
1639  *	bremfreef:
1640  *
1641  *	Force an immediate removal from a free list.  Used only in nfs when
1642  *	it abuses the b_freelist pointer.
1643  */
1644 void
1645 bremfreef(struct buf *bp)
1646 {
1647 	struct mtx *qlock;
1648 
1649 	qlock = bqlock(bp->b_qindex);
1650 	mtx_lock(qlock);
1651 	bremfreel(bp);
1652 	mtx_unlock(qlock);
1653 }
1654 
1655 /*
1656  *	bremfreel:
1657  *
1658  *	Removes a buffer from the free list, must be called with the
1659  *	correct qlock held.
1660  */
1661 static void
1662 bremfreel(struct buf *bp)
1663 {
1664 
1665 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1666 	    bp, bp->b_vp, bp->b_flags);
1667 	KASSERT(bp->b_qindex != QUEUE_NONE,
1668 	    ("bremfreel: buffer %p not on a queue.", bp));
1669 	if (bp->b_qindex != QUEUE_EMPTY) {
1670 		BUF_ASSERT_XLOCKED(bp);
1671 	}
1672 	mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1673 
1674 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1675 #ifdef INVARIANTS
1676 	KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1677 	    bp->b_qindex));
1678 	bq_len[bp->b_qindex]--;
1679 #endif
1680 	bp->b_qindex = QUEUE_NONE;
1681 	bp->b_flags &= ~B_REMFREE;
1682 }
1683 
1684 /*
1685  *	bufkva_free:
1686  *
1687  *	Free the kva allocation for a buffer.
1688  *
1689  */
1690 static void
1691 bufkva_free(struct buf *bp)
1692 {
1693 
1694 #ifdef INVARIANTS
1695 	if (bp->b_kvasize == 0) {
1696 		KASSERT(bp->b_kvabase == unmapped_buf &&
1697 		    bp->b_data == unmapped_buf,
1698 		    ("Leaked KVA space on %p", bp));
1699 	} else if (buf_mapped(bp))
1700 		BUF_CHECK_MAPPED(bp);
1701 	else
1702 		BUF_CHECK_UNMAPPED(bp);
1703 #endif
1704 	if (bp->b_kvasize == 0)
1705 		return;
1706 
1707 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
1708 	atomic_subtract_long(&bufkvaspace, bp->b_kvasize);
1709 	atomic_add_int(&buffreekvacnt, 1);
1710 	bp->b_data = bp->b_kvabase = unmapped_buf;
1711 	bp->b_kvasize = 0;
1712 }
1713 
1714 /*
1715  *	bufkva_alloc:
1716  *
1717  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
1718  */
1719 static int
1720 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
1721 {
1722 	vm_offset_t addr;
1723 	int error;
1724 
1725 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
1726 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
1727 
1728 	bufkva_free(bp);
1729 
1730 	addr = 0;
1731 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
1732 	if (error != 0) {
1733 		/*
1734 		 * Buffer map is too fragmented.  Request the caller
1735 		 * to defragment the map.
1736 		 */
1737 		return (error);
1738 	}
1739 	bp->b_kvabase = (caddr_t)addr;
1740 	bp->b_kvasize = maxsize;
1741 	atomic_add_long(&bufkvaspace, bp->b_kvasize);
1742 	if ((gbflags & GB_UNMAPPED) != 0) {
1743 		bp->b_data = unmapped_buf;
1744 		BUF_CHECK_UNMAPPED(bp);
1745 	} else {
1746 		bp->b_data = bp->b_kvabase;
1747 		BUF_CHECK_MAPPED(bp);
1748 	}
1749 	return (0);
1750 }
1751 
1752 /*
1753  *	bufkva_reclaim:
1754  *
1755  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
1756  *	callback that fires to avoid returning failure.
1757  */
1758 static void
1759 bufkva_reclaim(vmem_t *vmem, int flags)
1760 {
1761 	int i;
1762 
1763 	for (i = 0; i < 5; i++)
1764 		if (buf_scan(true) != 0)
1765 			break;
1766 	return;
1767 }
1768 
1769 
1770 /*
1771  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
1772  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1773  * the buffer is valid and we do not have to do anything.
1774  */
1775 void
1776 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1777     int cnt, struct ucred * cred)
1778 {
1779 	struct buf *rabp;
1780 	int i;
1781 
1782 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1783 		if (inmem(vp, *rablkno))
1784 			continue;
1785 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1786 
1787 		if ((rabp->b_flags & B_CACHE) == 0) {
1788 			if (!TD_IS_IDLETHREAD(curthread)) {
1789 #ifdef RACCT
1790 				if (racct_enable) {
1791 					PROC_LOCK(curproc);
1792 					racct_add_buf(curproc, rabp, 0);
1793 					PROC_UNLOCK(curproc);
1794 				}
1795 #endif /* RACCT */
1796 				curthread->td_ru.ru_inblock++;
1797 			}
1798 			rabp->b_flags |= B_ASYNC;
1799 			rabp->b_flags &= ~B_INVAL;
1800 			rabp->b_ioflags &= ~BIO_ERROR;
1801 			rabp->b_iocmd = BIO_READ;
1802 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
1803 				rabp->b_rcred = crhold(cred);
1804 			vfs_busy_pages(rabp, 0);
1805 			BUF_KERNPROC(rabp);
1806 			rabp->b_iooffset = dbtob(rabp->b_blkno);
1807 			bstrategy(rabp);
1808 		} else {
1809 			brelse(rabp);
1810 		}
1811 	}
1812 }
1813 
1814 /*
1815  * Entry point for bread() and breadn() via #defines in sys/buf.h.
1816  *
1817  * Get a buffer with the specified data.  Look in the cache first.  We
1818  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
1819  * is set, the buffer is valid and we do not have to do anything, see
1820  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1821  *
1822  * Always return a NULL buffer pointer (in bpp) when returning an error.
1823  */
1824 int
1825 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1826     int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1827 {
1828 	struct buf *bp;
1829 	int rv = 0, readwait = 0;
1830 
1831 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1832 	/*
1833 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1834 	 */
1835 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1836 	if (bp == NULL)
1837 		return (EBUSY);
1838 
1839 	/* if not found in cache, do some I/O */
1840 	if ((bp->b_flags & B_CACHE) == 0) {
1841 		if (!TD_IS_IDLETHREAD(curthread)) {
1842 #ifdef RACCT
1843 			if (racct_enable) {
1844 				PROC_LOCK(curproc);
1845 				racct_add_buf(curproc, bp, 0);
1846 				PROC_UNLOCK(curproc);
1847 			}
1848 #endif /* RACCT */
1849 			curthread->td_ru.ru_inblock++;
1850 		}
1851 		bp->b_iocmd = BIO_READ;
1852 		bp->b_flags &= ~B_INVAL;
1853 		bp->b_ioflags &= ~BIO_ERROR;
1854 		if (bp->b_rcred == NOCRED && cred != NOCRED)
1855 			bp->b_rcred = crhold(cred);
1856 		vfs_busy_pages(bp, 0);
1857 		bp->b_iooffset = dbtob(bp->b_blkno);
1858 		bstrategy(bp);
1859 		++readwait;
1860 	}
1861 
1862 	breada(vp, rablkno, rabsize, cnt, cred);
1863 
1864 	if (readwait) {
1865 		rv = bufwait(bp);
1866 		if (rv != 0) {
1867 			brelse(bp);
1868 			*bpp = NULL;
1869 		}
1870 	}
1871 	return (rv);
1872 }
1873 
1874 /*
1875  * Write, release buffer on completion.  (Done by iodone
1876  * if async).  Do not bother writing anything if the buffer
1877  * is invalid.
1878  *
1879  * Note that we set B_CACHE here, indicating that buffer is
1880  * fully valid and thus cacheable.  This is true even of NFS
1881  * now so we set it generally.  This could be set either here
1882  * or in biodone() since the I/O is synchronous.  We put it
1883  * here.
1884  */
1885 int
1886 bufwrite(struct buf *bp)
1887 {
1888 	int oldflags;
1889 	struct vnode *vp;
1890 	long space;
1891 	int vp_md;
1892 
1893 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1894 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
1895 		bp->b_flags |= B_INVAL | B_RELBUF;
1896 		bp->b_flags &= ~B_CACHE;
1897 		brelse(bp);
1898 		return (ENXIO);
1899 	}
1900 	if (bp->b_flags & B_INVAL) {
1901 		brelse(bp);
1902 		return (0);
1903 	}
1904 
1905 	if (bp->b_flags & B_BARRIER)
1906 		barrierwrites++;
1907 
1908 	oldflags = bp->b_flags;
1909 
1910 	BUF_ASSERT_HELD(bp);
1911 
1912 	if (bp->b_pin_count > 0)
1913 		bunpin_wait(bp);
1914 
1915 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1916 	    ("FFS background buffer should not get here %p", bp));
1917 
1918 	vp = bp->b_vp;
1919 	if (vp)
1920 		vp_md = vp->v_vflag & VV_MD;
1921 	else
1922 		vp_md = 0;
1923 
1924 	/*
1925 	 * Mark the buffer clean.  Increment the bufobj write count
1926 	 * before bundirty() call, to prevent other thread from seeing
1927 	 * empty dirty list and zero counter for writes in progress,
1928 	 * falsely indicating that the bufobj is clean.
1929 	 */
1930 	bufobj_wref(bp->b_bufobj);
1931 	bundirty(bp);
1932 
1933 	bp->b_flags &= ~B_DONE;
1934 	bp->b_ioflags &= ~BIO_ERROR;
1935 	bp->b_flags |= B_CACHE;
1936 	bp->b_iocmd = BIO_WRITE;
1937 
1938 	vfs_busy_pages(bp, 1);
1939 
1940 	/*
1941 	 * Normal bwrites pipeline writes
1942 	 */
1943 	bp->b_runningbufspace = bp->b_bufsize;
1944 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1945 
1946 	if (!TD_IS_IDLETHREAD(curthread)) {
1947 #ifdef RACCT
1948 		if (racct_enable) {
1949 			PROC_LOCK(curproc);
1950 			racct_add_buf(curproc, bp, 1);
1951 			PROC_UNLOCK(curproc);
1952 		}
1953 #endif /* RACCT */
1954 		curthread->td_ru.ru_oublock++;
1955 	}
1956 	if (oldflags & B_ASYNC)
1957 		BUF_KERNPROC(bp);
1958 	bp->b_iooffset = dbtob(bp->b_blkno);
1959 	bstrategy(bp);
1960 
1961 	if ((oldflags & B_ASYNC) == 0) {
1962 		int rtval = bufwait(bp);
1963 		brelse(bp);
1964 		return (rtval);
1965 	} else if (space > hirunningspace) {
1966 		/*
1967 		 * don't allow the async write to saturate the I/O
1968 		 * system.  We will not deadlock here because
1969 		 * we are blocking waiting for I/O that is already in-progress
1970 		 * to complete. We do not block here if it is the update
1971 		 * or syncer daemon trying to clean up as that can lead
1972 		 * to deadlock.
1973 		 */
1974 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1975 			waitrunningbufspace();
1976 	}
1977 
1978 	return (0);
1979 }
1980 
1981 void
1982 bufbdflush(struct bufobj *bo, struct buf *bp)
1983 {
1984 	struct buf *nbp;
1985 
1986 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1987 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1988 		altbufferflushes++;
1989 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1990 		BO_LOCK(bo);
1991 		/*
1992 		 * Try to find a buffer to flush.
1993 		 */
1994 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1995 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1996 			    BUF_LOCK(nbp,
1997 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
1998 				continue;
1999 			if (bp == nbp)
2000 				panic("bdwrite: found ourselves");
2001 			BO_UNLOCK(bo);
2002 			/* Don't countdeps with the bo lock held. */
2003 			if (buf_countdeps(nbp, 0)) {
2004 				BO_LOCK(bo);
2005 				BUF_UNLOCK(nbp);
2006 				continue;
2007 			}
2008 			if (nbp->b_flags & B_CLUSTEROK) {
2009 				vfs_bio_awrite(nbp);
2010 			} else {
2011 				bremfree(nbp);
2012 				bawrite(nbp);
2013 			}
2014 			dirtybufferflushes++;
2015 			break;
2016 		}
2017 		if (nbp == NULL)
2018 			BO_UNLOCK(bo);
2019 	}
2020 }
2021 
2022 /*
2023  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2024  * anything if the buffer is marked invalid.
2025  *
2026  * Note that since the buffer must be completely valid, we can safely
2027  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2028  * biodone() in order to prevent getblk from writing the buffer
2029  * out synchronously.
2030  */
2031 void
2032 bdwrite(struct buf *bp)
2033 {
2034 	struct thread *td = curthread;
2035 	struct vnode *vp;
2036 	struct bufobj *bo;
2037 
2038 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2039 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2040 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2041 	    ("Barrier request in delayed write %p", bp));
2042 	BUF_ASSERT_HELD(bp);
2043 
2044 	if (bp->b_flags & B_INVAL) {
2045 		brelse(bp);
2046 		return;
2047 	}
2048 
2049 	/*
2050 	 * If we have too many dirty buffers, don't create any more.
2051 	 * If we are wildly over our limit, then force a complete
2052 	 * cleanup. Otherwise, just keep the situation from getting
2053 	 * out of control. Note that we have to avoid a recursive
2054 	 * disaster and not try to clean up after our own cleanup!
2055 	 */
2056 	vp = bp->b_vp;
2057 	bo = bp->b_bufobj;
2058 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2059 		td->td_pflags |= TDP_INBDFLUSH;
2060 		BO_BDFLUSH(bo, bp);
2061 		td->td_pflags &= ~TDP_INBDFLUSH;
2062 	} else
2063 		recursiveflushes++;
2064 
2065 	bdirty(bp);
2066 	/*
2067 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2068 	 * true even of NFS now.
2069 	 */
2070 	bp->b_flags |= B_CACHE;
2071 
2072 	/*
2073 	 * This bmap keeps the system from needing to do the bmap later,
2074 	 * perhaps when the system is attempting to do a sync.  Since it
2075 	 * is likely that the indirect block -- or whatever other datastructure
2076 	 * that the filesystem needs is still in memory now, it is a good
2077 	 * thing to do this.  Note also, that if the pageout daemon is
2078 	 * requesting a sync -- there might not be enough memory to do
2079 	 * the bmap then...  So, this is important to do.
2080 	 */
2081 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2082 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2083 	}
2084 
2085 	/*
2086 	 * Set the *dirty* buffer range based upon the VM system dirty
2087 	 * pages.
2088 	 *
2089 	 * Mark the buffer pages as clean.  We need to do this here to
2090 	 * satisfy the vnode_pager and the pageout daemon, so that it
2091 	 * thinks that the pages have been "cleaned".  Note that since
2092 	 * the pages are in a delayed write buffer -- the VFS layer
2093 	 * "will" see that the pages get written out on the next sync,
2094 	 * or perhaps the cluster will be completed.
2095 	 */
2096 	vfs_clean_pages_dirty_buf(bp);
2097 	bqrelse(bp);
2098 
2099 	/*
2100 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2101 	 * due to the softdep code.
2102 	 */
2103 }
2104 
2105 /*
2106  *	bdirty:
2107  *
2108  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2109  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2110  *	itself to properly update it in the dirty/clean lists.  We mark it
2111  *	B_DONE to ensure that any asynchronization of the buffer properly
2112  *	clears B_DONE ( else a panic will occur later ).
2113  *
2114  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2115  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2116  *	should only be called if the buffer is known-good.
2117  *
2118  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2119  *	count.
2120  *
2121  *	The buffer must be on QUEUE_NONE.
2122  */
2123 void
2124 bdirty(struct buf *bp)
2125 {
2126 
2127 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2128 	    bp, bp->b_vp, bp->b_flags);
2129 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2130 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2131 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2132 	BUF_ASSERT_HELD(bp);
2133 	bp->b_flags &= ~(B_RELBUF);
2134 	bp->b_iocmd = BIO_WRITE;
2135 
2136 	if ((bp->b_flags & B_DELWRI) == 0) {
2137 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2138 		reassignbuf(bp);
2139 		bdirtyadd();
2140 	}
2141 }
2142 
2143 /*
2144  *	bundirty:
2145  *
2146  *	Clear B_DELWRI for buffer.
2147  *
2148  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2149  *	count.
2150  *
2151  *	The buffer must be on QUEUE_NONE.
2152  */
2153 
2154 void
2155 bundirty(struct buf *bp)
2156 {
2157 
2158 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2159 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2160 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2161 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2162 	BUF_ASSERT_HELD(bp);
2163 
2164 	if (bp->b_flags & B_DELWRI) {
2165 		bp->b_flags &= ~B_DELWRI;
2166 		reassignbuf(bp);
2167 		bdirtysub();
2168 	}
2169 	/*
2170 	 * Since it is now being written, we can clear its deferred write flag.
2171 	 */
2172 	bp->b_flags &= ~B_DEFERRED;
2173 }
2174 
2175 /*
2176  *	bawrite:
2177  *
2178  *	Asynchronous write.  Start output on a buffer, but do not wait for
2179  *	it to complete.  The buffer is released when the output completes.
2180  *
2181  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2182  *	B_INVAL buffers.  Not us.
2183  */
2184 void
2185 bawrite(struct buf *bp)
2186 {
2187 
2188 	bp->b_flags |= B_ASYNC;
2189 	(void) bwrite(bp);
2190 }
2191 
2192 /*
2193  *	babarrierwrite:
2194  *
2195  *	Asynchronous barrier write.  Start output on a buffer, but do not
2196  *	wait for it to complete.  Place a write barrier after this write so
2197  *	that this buffer and all buffers written before it are committed to
2198  *	the disk before any buffers written after this write are committed
2199  *	to the disk.  The buffer is released when the output completes.
2200  */
2201 void
2202 babarrierwrite(struct buf *bp)
2203 {
2204 
2205 	bp->b_flags |= B_ASYNC | B_BARRIER;
2206 	(void) bwrite(bp);
2207 }
2208 
2209 /*
2210  *	bbarrierwrite:
2211  *
2212  *	Synchronous barrier write.  Start output on a buffer and wait for
2213  *	it to complete.  Place a write barrier after this write so that
2214  *	this buffer and all buffers written before it are committed to
2215  *	the disk before any buffers written after this write are committed
2216  *	to the disk.  The buffer is released when the output completes.
2217  */
2218 int
2219 bbarrierwrite(struct buf *bp)
2220 {
2221 
2222 	bp->b_flags |= B_BARRIER;
2223 	return (bwrite(bp));
2224 }
2225 
2226 /*
2227  *	bwillwrite:
2228  *
2229  *	Called prior to the locking of any vnodes when we are expecting to
2230  *	write.  We do not want to starve the buffer cache with too many
2231  *	dirty buffers so we block here.  By blocking prior to the locking
2232  *	of any vnodes we attempt to avoid the situation where a locked vnode
2233  *	prevents the various system daemons from flushing related buffers.
2234  */
2235 void
2236 bwillwrite(void)
2237 {
2238 
2239 	if (numdirtybuffers >= hidirtybuffers) {
2240 		mtx_lock(&bdirtylock);
2241 		while (numdirtybuffers >= hidirtybuffers) {
2242 			bdirtywait = 1;
2243 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2244 			    "flswai", 0);
2245 		}
2246 		mtx_unlock(&bdirtylock);
2247 	}
2248 }
2249 
2250 /*
2251  * Return true if we have too many dirty buffers.
2252  */
2253 int
2254 buf_dirty_count_severe(void)
2255 {
2256 
2257 	return(numdirtybuffers >= hidirtybuffers);
2258 }
2259 
2260 /*
2261  *	brelse:
2262  *
2263  *	Release a busy buffer and, if requested, free its resources.  The
2264  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2265  *	to be accessed later as a cache entity or reused for other purposes.
2266  */
2267 void
2268 brelse(struct buf *bp)
2269 {
2270 	int qindex;
2271 
2272 	/*
2273 	 * Many functions erroneously call brelse with a NULL bp under rare
2274 	 * error conditions. Simply return when called with a NULL bp.
2275 	 */
2276 	if (bp == NULL)
2277 		return;
2278 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2279 	    bp, bp->b_vp, bp->b_flags);
2280 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2281 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2282 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2283 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2284 
2285 	if (BUF_LOCKRECURSED(bp)) {
2286 		/*
2287 		 * Do not process, in particular, do not handle the
2288 		 * B_INVAL/B_RELBUF and do not release to free list.
2289 		 */
2290 		BUF_UNLOCK(bp);
2291 		return;
2292 	}
2293 
2294 	if (bp->b_flags & B_MANAGED) {
2295 		bqrelse(bp);
2296 		return;
2297 	}
2298 
2299 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2300 		BO_LOCK(bp->b_bufobj);
2301 		bp->b_vflags &= ~BV_BKGRDERR;
2302 		BO_UNLOCK(bp->b_bufobj);
2303 		bdirty(bp);
2304 	}
2305 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2306 	    !(bp->b_flags & B_INVAL)) {
2307 		/*
2308 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
2309 		 * pages from being scrapped.
2310 		 */
2311 		bp->b_ioflags &= ~BIO_ERROR;
2312 		bdirty(bp);
2313 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2314 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2315 		/*
2316 		 * Either a failed read I/O or we were asked to free or not
2317 		 * cache the buffer.
2318 		 */
2319 		bp->b_flags |= B_INVAL;
2320 		if (!LIST_EMPTY(&bp->b_dep))
2321 			buf_deallocate(bp);
2322 		if (bp->b_flags & B_DELWRI)
2323 			bdirtysub();
2324 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2325 		if ((bp->b_flags & B_VMIO) == 0) {
2326 			allocbuf(bp, 0);
2327 			if (bp->b_vp)
2328 				brelvp(bp);
2329 		}
2330 	}
2331 
2332 	/*
2333 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2334 	 * is called with B_DELWRI set, the underlying pages may wind up
2335 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2336 	 * because pages associated with a B_DELWRI bp are marked clean.
2337 	 *
2338 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2339 	 * if B_DELWRI is set.
2340 	 */
2341 	if (bp->b_flags & B_DELWRI)
2342 		bp->b_flags &= ~B_RELBUF;
2343 
2344 	/*
2345 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2346 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2347 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2348 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2349 	 *
2350 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2351 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2352 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2353 	 *
2354 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2355 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2356 	 * the commit state and we cannot afford to lose the buffer. If the
2357 	 * buffer has a background write in progress, we need to keep it
2358 	 * around to prevent it from being reconstituted and starting a second
2359 	 * background write.
2360 	 */
2361 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2362 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2363 	    !(bp->b_vp->v_mount != NULL &&
2364 	    (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2365 	    !vn_isdisk(bp->b_vp, NULL) && (bp->b_flags & B_DELWRI))) {
2366 		vfs_vmio_invalidate(bp);
2367 		allocbuf(bp, 0);
2368 	}
2369 
2370 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2371 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2372 		allocbuf(bp, 0);
2373 		bp->b_flags &= ~B_NOREUSE;
2374 		if (bp->b_vp != NULL)
2375 			brelvp(bp);
2376 	}
2377 
2378 	/*
2379 	 * If the buffer has junk contents signal it and eventually
2380 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2381 	 * doesn't find it.
2382 	 */
2383 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2384 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2385 		bp->b_flags |= B_INVAL;
2386 	if (bp->b_flags & B_INVAL) {
2387 		if (bp->b_flags & B_DELWRI)
2388 			bundirty(bp);
2389 		if (bp->b_vp)
2390 			brelvp(bp);
2391 	}
2392 
2393 	/* buffers with no memory */
2394 	if (bp->b_bufsize == 0) {
2395 		buf_free(bp);
2396 		return;
2397 	}
2398 	/* buffers with junk contents */
2399 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2400 	    (bp->b_ioflags & BIO_ERROR)) {
2401 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2402 		if (bp->b_vflags & BV_BKGRDINPROG)
2403 			panic("losing buffer 2");
2404 		qindex = QUEUE_CLEAN;
2405 		bp->b_flags |= B_AGE;
2406 	/* remaining buffers */
2407 	} else if (bp->b_flags & B_DELWRI)
2408 		qindex = QUEUE_DIRTY;
2409 	else
2410 		qindex = QUEUE_CLEAN;
2411 
2412 	binsfree(bp, qindex);
2413 
2414 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
2415 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2416 		panic("brelse: not dirty");
2417 	/* unlock */
2418 	BUF_UNLOCK(bp);
2419 	if (qindex == QUEUE_CLEAN)
2420 		bufspace_wakeup();
2421 }
2422 
2423 /*
2424  * Release a buffer back to the appropriate queue but do not try to free
2425  * it.  The buffer is expected to be used again soon.
2426  *
2427  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2428  * biodone() to requeue an async I/O on completion.  It is also used when
2429  * known good buffers need to be requeued but we think we may need the data
2430  * again soon.
2431  *
2432  * XXX we should be able to leave the B_RELBUF hint set on completion.
2433  */
2434 void
2435 bqrelse(struct buf *bp)
2436 {
2437 	int qindex;
2438 
2439 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2440 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2441 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2442 
2443 	qindex = QUEUE_NONE;
2444 	if (BUF_LOCKRECURSED(bp)) {
2445 		/* do not release to free list */
2446 		BUF_UNLOCK(bp);
2447 		return;
2448 	}
2449 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2450 
2451 	if (bp->b_flags & B_MANAGED) {
2452 		if (bp->b_flags & B_REMFREE)
2453 			bremfreef(bp);
2454 		goto out;
2455 	}
2456 
2457 	/* buffers with stale but valid contents */
2458 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2459 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2460 		BO_LOCK(bp->b_bufobj);
2461 		bp->b_vflags &= ~BV_BKGRDERR;
2462 		BO_UNLOCK(bp->b_bufobj);
2463 		qindex = QUEUE_DIRTY;
2464 	} else {
2465 		if ((bp->b_flags & B_DELWRI) == 0 &&
2466 		    (bp->b_xflags & BX_VNDIRTY))
2467 			panic("bqrelse: not dirty");
2468 		if ((bp->b_flags & B_NOREUSE) != 0) {
2469 			brelse(bp);
2470 			return;
2471 		}
2472 		qindex = QUEUE_CLEAN;
2473 	}
2474 	binsfree(bp, qindex);
2475 
2476 out:
2477 	/* unlock */
2478 	BUF_UNLOCK(bp);
2479 	if (qindex == QUEUE_CLEAN)
2480 		bufspace_wakeup();
2481 }
2482 
2483 /*
2484  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2485  * restore bogus pages.
2486  */
2487 static void
2488 vfs_vmio_iodone(struct buf *bp)
2489 {
2490 	vm_ooffset_t foff;
2491 	vm_page_t m;
2492 	vm_object_t obj;
2493 	struct vnode *vp;
2494 	int bogus, i, iosize;
2495 
2496 	obj = bp->b_bufobj->bo_object;
2497 	KASSERT(obj->paging_in_progress >= bp->b_npages,
2498 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2499 	    obj->paging_in_progress, bp->b_npages));
2500 
2501 	vp = bp->b_vp;
2502 	KASSERT(vp->v_holdcnt > 0,
2503 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2504 	KASSERT(vp->v_object != NULL,
2505 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2506 
2507 	foff = bp->b_offset;
2508 	KASSERT(bp->b_offset != NOOFFSET,
2509 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2510 
2511 	bogus = 0;
2512 	iosize = bp->b_bcount - bp->b_resid;
2513 	VM_OBJECT_WLOCK(obj);
2514 	for (i = 0; i < bp->b_npages; i++) {
2515 		int resid;
2516 
2517 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2518 		if (resid > iosize)
2519 			resid = iosize;
2520 
2521 		/*
2522 		 * cleanup bogus pages, restoring the originals
2523 		 */
2524 		m = bp->b_pages[i];
2525 		if (m == bogus_page) {
2526 			bogus = 1;
2527 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2528 			if (m == NULL)
2529 				panic("biodone: page disappeared!");
2530 			bp->b_pages[i] = m;
2531 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2532 			/*
2533 			 * In the write case, the valid and clean bits are
2534 			 * already changed correctly ( see bdwrite() ), so we
2535 			 * only need to do this here in the read case.
2536 			 */
2537 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2538 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2539 			    "has unexpected dirty bits", m));
2540 			vfs_page_set_valid(bp, foff, m);
2541 		}
2542 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2543 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2544 		    (intmax_t)foff, (uintmax_t)m->pindex));
2545 
2546 		vm_page_sunbusy(m);
2547 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2548 		iosize -= resid;
2549 	}
2550 	vm_object_pip_wakeupn(obj, bp->b_npages);
2551 	VM_OBJECT_WUNLOCK(obj);
2552 	if (bogus && buf_mapped(bp)) {
2553 		BUF_CHECK_MAPPED(bp);
2554 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2555 		    bp->b_pages, bp->b_npages);
2556 	}
2557 }
2558 
2559 /*
2560  * Unwire a page held by a buf and place it on the appropriate vm queue.
2561  */
2562 static void
2563 vfs_vmio_unwire(struct buf *bp, vm_page_t m)
2564 {
2565 	bool freed;
2566 
2567 	vm_page_lock(m);
2568 	if (vm_page_unwire(m, PQ_NONE)) {
2569 		/*
2570 		 * Determine if the page should be freed before adding
2571 		 * it to the inactive queue.
2572 		 */
2573 		if (m->valid == 0) {
2574 			freed = !vm_page_busied(m);
2575 			if (freed)
2576 				vm_page_free(m);
2577 		} else if ((bp->b_flags & B_DIRECT) != 0)
2578 			freed = vm_page_try_to_free(m);
2579 		else
2580 			freed = false;
2581 		if (!freed) {
2582 			/*
2583 			 * If the page is unlikely to be reused, let the
2584 			 * VM know.  Otherwise, maintain LRU page
2585 			 * ordering and put the page at the tail of the
2586 			 * inactive queue.
2587 			 */
2588 			if ((bp->b_flags & B_NOREUSE) != 0)
2589 				vm_page_deactivate_noreuse(m);
2590 			else
2591 				vm_page_deactivate(m);
2592 		}
2593 	}
2594 	vm_page_unlock(m);
2595 }
2596 
2597 /*
2598  * Perform page invalidation when a buffer is released.  The fully invalid
2599  * pages will be reclaimed later in vfs_vmio_truncate().
2600  */
2601 static void
2602 vfs_vmio_invalidate(struct buf *bp)
2603 {
2604 	vm_object_t obj;
2605 	vm_page_t m;
2606 	int i, resid, poffset, presid;
2607 
2608 	if (buf_mapped(bp)) {
2609 		BUF_CHECK_MAPPED(bp);
2610 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2611 	} else
2612 		BUF_CHECK_UNMAPPED(bp);
2613 	/*
2614 	 * Get the base offset and length of the buffer.  Note that
2615 	 * in the VMIO case if the buffer block size is not
2616 	 * page-aligned then b_data pointer may not be page-aligned.
2617 	 * But our b_pages[] array *IS* page aligned.
2618 	 *
2619 	 * block sizes less then DEV_BSIZE (usually 512) are not
2620 	 * supported due to the page granularity bits (m->valid,
2621 	 * m->dirty, etc...).
2622 	 *
2623 	 * See man buf(9) for more information
2624 	 */
2625 	obj = bp->b_bufobj->bo_object;
2626 	resid = bp->b_bufsize;
2627 	poffset = bp->b_offset & PAGE_MASK;
2628 	VM_OBJECT_WLOCK(obj);
2629 	for (i = 0; i < bp->b_npages; i++) {
2630 		m = bp->b_pages[i];
2631 		if (m == bogus_page)
2632 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2633 		bp->b_pages[i] = NULL;
2634 
2635 		presid = resid > (PAGE_SIZE - poffset) ?
2636 		    (PAGE_SIZE - poffset) : resid;
2637 		KASSERT(presid >= 0, ("brelse: extra page"));
2638 		while (vm_page_xbusied(m)) {
2639 			vm_page_lock(m);
2640 			VM_OBJECT_WUNLOCK(obj);
2641 			vm_page_busy_sleep(m, "mbncsh");
2642 			VM_OBJECT_WLOCK(obj);
2643 		}
2644 		if (pmap_page_wired_mappings(m) == 0)
2645 			vm_page_set_invalid(m, poffset, presid);
2646 		vfs_vmio_unwire(bp, m);
2647 		resid -= presid;
2648 		poffset = 0;
2649 	}
2650 	VM_OBJECT_WUNLOCK(obj);
2651 	bp->b_npages = 0;
2652 }
2653 
2654 /*
2655  * Page-granular truncation of an existing VMIO buffer.
2656  */
2657 static void
2658 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2659 {
2660 	vm_object_t obj;
2661 	vm_page_t m;
2662 	int i;
2663 
2664 	if (bp->b_npages == desiredpages)
2665 		return;
2666 
2667 	if (buf_mapped(bp)) {
2668 		BUF_CHECK_MAPPED(bp);
2669 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2670 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2671 	} else
2672 		BUF_CHECK_UNMAPPED(bp);
2673 	obj = bp->b_bufobj->bo_object;
2674 	if (obj != NULL)
2675 		VM_OBJECT_WLOCK(obj);
2676 	for (i = desiredpages; i < bp->b_npages; i++) {
2677 		m = bp->b_pages[i];
2678 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
2679 		bp->b_pages[i] = NULL;
2680 		vfs_vmio_unwire(bp, m);
2681 	}
2682 	if (obj != NULL)
2683 		VM_OBJECT_WUNLOCK(obj);
2684 	bp->b_npages = desiredpages;
2685 }
2686 
2687 /*
2688  * Byte granular extension of VMIO buffers.
2689  */
2690 static void
2691 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
2692 {
2693 	/*
2694 	 * We are growing the buffer, possibly in a
2695 	 * byte-granular fashion.
2696 	 */
2697 	vm_object_t obj;
2698 	vm_offset_t toff;
2699 	vm_offset_t tinc;
2700 	vm_page_t m;
2701 
2702 	/*
2703 	 * Step 1, bring in the VM pages from the object, allocating
2704 	 * them if necessary.  We must clear B_CACHE if these pages
2705 	 * are not valid for the range covered by the buffer.
2706 	 */
2707 	obj = bp->b_bufobj->bo_object;
2708 	VM_OBJECT_WLOCK(obj);
2709 	while (bp->b_npages < desiredpages) {
2710 		/*
2711 		 * We must allocate system pages since blocking
2712 		 * here could interfere with paging I/O, no
2713 		 * matter which process we are.
2714 		 *
2715 		 * Only exclusive busy can be tested here.
2716 		 * Blocking on shared busy might lead to
2717 		 * deadlocks once allocbuf() is called after
2718 		 * pages are vfs_busy_pages().
2719 		 */
2720 		m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + bp->b_npages,
2721 		    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2722 		    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY |
2723 		    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
2724 		if (m->valid == 0)
2725 			bp->b_flags &= ~B_CACHE;
2726 		bp->b_pages[bp->b_npages] = m;
2727 		++bp->b_npages;
2728 	}
2729 
2730 	/*
2731 	 * Step 2.  We've loaded the pages into the buffer,
2732 	 * we have to figure out if we can still have B_CACHE
2733 	 * set.  Note that B_CACHE is set according to the
2734 	 * byte-granular range ( bcount and size ), not the
2735 	 * aligned range ( newbsize ).
2736 	 *
2737 	 * The VM test is against m->valid, which is DEV_BSIZE
2738 	 * aligned.  Needless to say, the validity of the data
2739 	 * needs to also be DEV_BSIZE aligned.  Note that this
2740 	 * fails with NFS if the server or some other client
2741 	 * extends the file's EOF.  If our buffer is resized,
2742 	 * B_CACHE may remain set! XXX
2743 	 */
2744 	toff = bp->b_bcount;
2745 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2746 	while ((bp->b_flags & B_CACHE) && toff < size) {
2747 		vm_pindex_t pi;
2748 
2749 		if (tinc > (size - toff))
2750 			tinc = size - toff;
2751 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
2752 		m = bp->b_pages[pi];
2753 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
2754 		toff += tinc;
2755 		tinc = PAGE_SIZE;
2756 	}
2757 	VM_OBJECT_WUNLOCK(obj);
2758 
2759 	/*
2760 	 * Step 3, fixup the KVA pmap.
2761 	 */
2762 	if (buf_mapped(bp))
2763 		bpmap_qenter(bp);
2764 	else
2765 		BUF_CHECK_UNMAPPED(bp);
2766 }
2767 
2768 /*
2769  * Check to see if a block at a particular lbn is available for a clustered
2770  * write.
2771  */
2772 static int
2773 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
2774 {
2775 	struct buf *bpa;
2776 	int match;
2777 
2778 	match = 0;
2779 
2780 	/* If the buf isn't in core skip it */
2781 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
2782 		return (0);
2783 
2784 	/* If the buf is busy we don't want to wait for it */
2785 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2786 		return (0);
2787 
2788 	/* Only cluster with valid clusterable delayed write buffers */
2789 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
2790 	    (B_DELWRI | B_CLUSTEROK))
2791 		goto done;
2792 
2793 	if (bpa->b_bufsize != size)
2794 		goto done;
2795 
2796 	/*
2797 	 * Check to see if it is in the expected place on disk and that the
2798 	 * block has been mapped.
2799 	 */
2800 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
2801 		match = 1;
2802 done:
2803 	BUF_UNLOCK(bpa);
2804 	return (match);
2805 }
2806 
2807 /*
2808  *	vfs_bio_awrite:
2809  *
2810  *	Implement clustered async writes for clearing out B_DELWRI buffers.
2811  *	This is much better then the old way of writing only one buffer at
2812  *	a time.  Note that we may not be presented with the buffers in the
2813  *	correct order, so we search for the cluster in both directions.
2814  */
2815 int
2816 vfs_bio_awrite(struct buf *bp)
2817 {
2818 	struct bufobj *bo;
2819 	int i;
2820 	int j;
2821 	daddr_t lblkno = bp->b_lblkno;
2822 	struct vnode *vp = bp->b_vp;
2823 	int ncl;
2824 	int nwritten;
2825 	int size;
2826 	int maxcl;
2827 	int gbflags;
2828 
2829 	bo = &vp->v_bufobj;
2830 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
2831 	/*
2832 	 * right now we support clustered writing only to regular files.  If
2833 	 * we find a clusterable block we could be in the middle of a cluster
2834 	 * rather then at the beginning.
2835 	 */
2836 	if ((vp->v_type == VREG) &&
2837 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
2838 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
2839 
2840 		size = vp->v_mount->mnt_stat.f_iosize;
2841 		maxcl = MAXPHYS / size;
2842 
2843 		BO_RLOCK(bo);
2844 		for (i = 1; i < maxcl; i++)
2845 			if (vfs_bio_clcheck(vp, size, lblkno + i,
2846 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
2847 				break;
2848 
2849 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
2850 			if (vfs_bio_clcheck(vp, size, lblkno - j,
2851 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
2852 				break;
2853 		BO_RUNLOCK(bo);
2854 		--j;
2855 		ncl = i + j;
2856 		/*
2857 		 * this is a possible cluster write
2858 		 */
2859 		if (ncl != 1) {
2860 			BUF_UNLOCK(bp);
2861 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
2862 			    gbflags);
2863 			return (nwritten);
2864 		}
2865 	}
2866 	bremfree(bp);
2867 	bp->b_flags |= B_ASYNC;
2868 	/*
2869 	 * default (old) behavior, writing out only one block
2870 	 *
2871 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
2872 	 */
2873 	nwritten = bp->b_bufsize;
2874 	(void) bwrite(bp);
2875 
2876 	return (nwritten);
2877 }
2878 
2879 /*
2880  *	getnewbuf_kva:
2881  *
2882  *	Allocate KVA for an empty buf header according to gbflags.
2883  */
2884 static int
2885 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
2886 {
2887 
2888 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
2889 		/*
2890 		 * In order to keep fragmentation sane we only allocate kva
2891 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
2892 		 */
2893 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2894 
2895 		if (maxsize != bp->b_kvasize &&
2896 		    bufkva_alloc(bp, maxsize, gbflags))
2897 			return (ENOSPC);
2898 	}
2899 	return (0);
2900 }
2901 
2902 /*
2903  *	getnewbuf:
2904  *
2905  *	Find and initialize a new buffer header, freeing up existing buffers
2906  *	in the bufqueues as necessary.  The new buffer is returned locked.
2907  *
2908  *	We block if:
2909  *		We have insufficient buffer headers
2910  *		We have insufficient buffer space
2911  *		buffer_arena is too fragmented ( space reservation fails )
2912  *		If we have to flush dirty buffers ( but we try to avoid this )
2913  *
2914  *	The caller is responsible for releasing the reserved bufspace after
2915  *	allocbuf() is called.
2916  */
2917 static struct buf *
2918 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
2919 {
2920 	struct buf *bp;
2921 	bool metadata, reserved;
2922 
2923 	bp = NULL;
2924 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2925 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2926 	if (!unmapped_buf_allowed)
2927 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2928 
2929 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2930 	    vp->v_type == VCHR)
2931 		metadata = true;
2932 	else
2933 		metadata = false;
2934 	atomic_add_int(&getnewbufcalls, 1);
2935 	reserved = false;
2936 	do {
2937 		if (reserved == false &&
2938 		    bufspace_reserve(maxsize, metadata) != 0)
2939 			continue;
2940 		reserved = true;
2941 		if ((bp = buf_alloc()) == NULL)
2942 			continue;
2943 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
2944 			return (bp);
2945 		break;
2946 	} while(buf_scan(false) == 0);
2947 
2948 	if (reserved)
2949 		atomic_subtract_long(&bufspace, maxsize);
2950 	if (bp != NULL) {
2951 		bp->b_flags |= B_INVAL;
2952 		brelse(bp);
2953 	}
2954 	bufspace_wait(vp, gbflags, slpflag, slptimeo);
2955 
2956 	return (NULL);
2957 }
2958 
2959 /*
2960  *	buf_daemon:
2961  *
2962  *	buffer flushing daemon.  Buffers are normally flushed by the
2963  *	update daemon but if it cannot keep up this process starts to
2964  *	take the load in an attempt to prevent getnewbuf() from blocking.
2965  */
2966 static struct kproc_desc buf_kp = {
2967 	"bufdaemon",
2968 	buf_daemon,
2969 	&bufdaemonproc
2970 };
2971 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2972 
2973 static int
2974 buf_flush(struct vnode *vp, int target)
2975 {
2976 	int flushed;
2977 
2978 	flushed = flushbufqueues(vp, target, 0);
2979 	if (flushed == 0) {
2980 		/*
2981 		 * Could not find any buffers without rollback
2982 		 * dependencies, so just write the first one
2983 		 * in the hopes of eventually making progress.
2984 		 */
2985 		if (vp != NULL && target > 2)
2986 			target /= 2;
2987 		flushbufqueues(vp, target, 1);
2988 	}
2989 	return (flushed);
2990 }
2991 
2992 static void
2993 buf_daemon()
2994 {
2995 	int lodirty;
2996 
2997 	/*
2998 	 * This process needs to be suspended prior to shutdown sync.
2999 	 */
3000 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
3001 	    SHUTDOWN_PRI_LAST);
3002 
3003 	/*
3004 	 * This process is allowed to take the buffer cache to the limit
3005 	 */
3006 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3007 	mtx_lock(&bdlock);
3008 	for (;;) {
3009 		bd_request = 0;
3010 		mtx_unlock(&bdlock);
3011 
3012 		kproc_suspend_check(bufdaemonproc);
3013 		lodirty = lodirtybuffers;
3014 		if (bd_speedupreq) {
3015 			lodirty = numdirtybuffers / 2;
3016 			bd_speedupreq = 0;
3017 		}
3018 		/*
3019 		 * Do the flush.  Limit the amount of in-transit I/O we
3020 		 * allow to build up, otherwise we would completely saturate
3021 		 * the I/O system.
3022 		 */
3023 		while (numdirtybuffers > lodirty) {
3024 			if (buf_flush(NULL, numdirtybuffers - lodirty) == 0)
3025 				break;
3026 			kern_yield(PRI_USER);
3027 		}
3028 
3029 		/*
3030 		 * Only clear bd_request if we have reached our low water
3031 		 * mark.  The buf_daemon normally waits 1 second and
3032 		 * then incrementally flushes any dirty buffers that have
3033 		 * built up, within reason.
3034 		 *
3035 		 * If we were unable to hit our low water mark and couldn't
3036 		 * find any flushable buffers, we sleep for a short period
3037 		 * to avoid endless loops on unlockable buffers.
3038 		 */
3039 		mtx_lock(&bdlock);
3040 		if (numdirtybuffers <= lodirtybuffers) {
3041 			/*
3042 			 * We reached our low water mark, reset the
3043 			 * request and sleep until we are needed again.
3044 			 * The sleep is just so the suspend code works.
3045 			 */
3046 			bd_request = 0;
3047 			/*
3048 			 * Do an extra wakeup in case dirty threshold
3049 			 * changed via sysctl and the explicit transition
3050 			 * out of shortfall was missed.
3051 			 */
3052 			bdirtywakeup();
3053 			if (runningbufspace <= lorunningspace)
3054 				runningwakeup();
3055 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3056 		} else {
3057 			/*
3058 			 * We couldn't find any flushable dirty buffers but
3059 			 * still have too many dirty buffers, we
3060 			 * have to sleep and try again.  (rare)
3061 			 */
3062 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3063 		}
3064 	}
3065 }
3066 
3067 /*
3068  *	flushbufqueues:
3069  *
3070  *	Try to flush a buffer in the dirty queue.  We must be careful to
3071  *	free up B_INVAL buffers instead of write them, which NFS is
3072  *	particularly sensitive to.
3073  */
3074 static int flushwithdeps = 0;
3075 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
3076     0, "Number of buffers flushed with dependecies that require rollbacks");
3077 
3078 static int
3079 flushbufqueues(struct vnode *lvp, int target, int flushdeps)
3080 {
3081 	struct buf *sentinel;
3082 	struct vnode *vp;
3083 	struct mount *mp;
3084 	struct buf *bp;
3085 	int hasdeps;
3086 	int flushed;
3087 	int queue;
3088 	int error;
3089 	bool unlock;
3090 
3091 	flushed = 0;
3092 	queue = QUEUE_DIRTY;
3093 	bp = NULL;
3094 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3095 	sentinel->b_qindex = QUEUE_SENTINEL;
3096 	mtx_lock(&bqlocks[queue]);
3097 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
3098 	mtx_unlock(&bqlocks[queue]);
3099 	while (flushed != target) {
3100 		maybe_yield();
3101 		mtx_lock(&bqlocks[queue]);
3102 		bp = TAILQ_NEXT(sentinel, b_freelist);
3103 		if (bp != NULL) {
3104 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
3105 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
3106 			    b_freelist);
3107 		} else {
3108 			mtx_unlock(&bqlocks[queue]);
3109 			break;
3110 		}
3111 		/*
3112 		 * Skip sentinels inserted by other invocations of the
3113 		 * flushbufqueues(), taking care to not reorder them.
3114 		 *
3115 		 * Only flush the buffers that belong to the
3116 		 * vnode locked by the curthread.
3117 		 */
3118 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3119 		    bp->b_vp != lvp)) {
3120 			mtx_unlock(&bqlocks[queue]);
3121  			continue;
3122 		}
3123 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3124 		mtx_unlock(&bqlocks[queue]);
3125 		if (error != 0)
3126 			continue;
3127 		if (bp->b_pin_count > 0) {
3128 			BUF_UNLOCK(bp);
3129 			continue;
3130 		}
3131 		/*
3132 		 * BKGRDINPROG can only be set with the buf and bufobj
3133 		 * locks both held.  We tolerate a race to clear it here.
3134 		 */
3135 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3136 		    (bp->b_flags & B_DELWRI) == 0) {
3137 			BUF_UNLOCK(bp);
3138 			continue;
3139 		}
3140 		if (bp->b_flags & B_INVAL) {
3141 			bremfreef(bp);
3142 			brelse(bp);
3143 			flushed++;
3144 			continue;
3145 		}
3146 
3147 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3148 			if (flushdeps == 0) {
3149 				BUF_UNLOCK(bp);
3150 				continue;
3151 			}
3152 			hasdeps = 1;
3153 		} else
3154 			hasdeps = 0;
3155 		/*
3156 		 * We must hold the lock on a vnode before writing
3157 		 * one of its buffers. Otherwise we may confuse, or
3158 		 * in the case of a snapshot vnode, deadlock the
3159 		 * system.
3160 		 *
3161 		 * The lock order here is the reverse of the normal
3162 		 * of vnode followed by buf lock.  This is ok because
3163 		 * the NOWAIT will prevent deadlock.
3164 		 */
3165 		vp = bp->b_vp;
3166 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3167 			BUF_UNLOCK(bp);
3168 			continue;
3169 		}
3170 		if (lvp == NULL) {
3171 			unlock = true;
3172 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3173 		} else {
3174 			ASSERT_VOP_LOCKED(vp, "getbuf");
3175 			unlock = false;
3176 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3177 			    vn_lock(vp, LK_TRYUPGRADE);
3178 		}
3179 		if (error == 0) {
3180 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3181 			    bp, bp->b_vp, bp->b_flags);
3182 			if (curproc == bufdaemonproc) {
3183 				vfs_bio_awrite(bp);
3184 			} else {
3185 				bremfree(bp);
3186 				bwrite(bp);
3187 				notbufdflushes++;
3188 			}
3189 			vn_finished_write(mp);
3190 			if (unlock)
3191 				VOP_UNLOCK(vp, 0);
3192 			flushwithdeps += hasdeps;
3193 			flushed++;
3194 
3195 			/*
3196 			 * Sleeping on runningbufspace while holding
3197 			 * vnode lock leads to deadlock.
3198 			 */
3199 			if (curproc == bufdaemonproc &&
3200 			    runningbufspace > hirunningspace)
3201 				waitrunningbufspace();
3202 			continue;
3203 		}
3204 		vn_finished_write(mp);
3205 		BUF_UNLOCK(bp);
3206 	}
3207 	mtx_lock(&bqlocks[queue]);
3208 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
3209 	mtx_unlock(&bqlocks[queue]);
3210 	free(sentinel, M_TEMP);
3211 	return (flushed);
3212 }
3213 
3214 /*
3215  * Check to see if a block is currently memory resident.
3216  */
3217 struct buf *
3218 incore(struct bufobj *bo, daddr_t blkno)
3219 {
3220 	struct buf *bp;
3221 
3222 	BO_RLOCK(bo);
3223 	bp = gbincore(bo, blkno);
3224 	BO_RUNLOCK(bo);
3225 	return (bp);
3226 }
3227 
3228 /*
3229  * Returns true if no I/O is needed to access the
3230  * associated VM object.  This is like incore except
3231  * it also hunts around in the VM system for the data.
3232  */
3233 
3234 static int
3235 inmem(struct vnode * vp, daddr_t blkno)
3236 {
3237 	vm_object_t obj;
3238 	vm_offset_t toff, tinc, size;
3239 	vm_page_t m;
3240 	vm_ooffset_t off;
3241 
3242 	ASSERT_VOP_LOCKED(vp, "inmem");
3243 
3244 	if (incore(&vp->v_bufobj, blkno))
3245 		return 1;
3246 	if (vp->v_mount == NULL)
3247 		return 0;
3248 	obj = vp->v_object;
3249 	if (obj == NULL)
3250 		return (0);
3251 
3252 	size = PAGE_SIZE;
3253 	if (size > vp->v_mount->mnt_stat.f_iosize)
3254 		size = vp->v_mount->mnt_stat.f_iosize;
3255 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3256 
3257 	VM_OBJECT_RLOCK(obj);
3258 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3259 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3260 		if (!m)
3261 			goto notinmem;
3262 		tinc = size;
3263 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3264 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3265 		if (vm_page_is_valid(m,
3266 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3267 			goto notinmem;
3268 	}
3269 	VM_OBJECT_RUNLOCK(obj);
3270 	return 1;
3271 
3272 notinmem:
3273 	VM_OBJECT_RUNLOCK(obj);
3274 	return (0);
3275 }
3276 
3277 /*
3278  * Set the dirty range for a buffer based on the status of the dirty
3279  * bits in the pages comprising the buffer.  The range is limited
3280  * to the size of the buffer.
3281  *
3282  * Tell the VM system that the pages associated with this buffer
3283  * are clean.  This is used for delayed writes where the data is
3284  * going to go to disk eventually without additional VM intevention.
3285  *
3286  * Note that while we only really need to clean through to b_bcount, we
3287  * just go ahead and clean through to b_bufsize.
3288  */
3289 static void
3290 vfs_clean_pages_dirty_buf(struct buf *bp)
3291 {
3292 	vm_ooffset_t foff, noff, eoff;
3293 	vm_page_t m;
3294 	int i;
3295 
3296 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3297 		return;
3298 
3299 	foff = bp->b_offset;
3300 	KASSERT(bp->b_offset != NOOFFSET,
3301 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3302 
3303 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3304 	vfs_drain_busy_pages(bp);
3305 	vfs_setdirty_locked_object(bp);
3306 	for (i = 0; i < bp->b_npages; i++) {
3307 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3308 		eoff = noff;
3309 		if (eoff > bp->b_offset + bp->b_bufsize)
3310 			eoff = bp->b_offset + bp->b_bufsize;
3311 		m = bp->b_pages[i];
3312 		vfs_page_set_validclean(bp, foff, m);
3313 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3314 		foff = noff;
3315 	}
3316 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3317 }
3318 
3319 static void
3320 vfs_setdirty_locked_object(struct buf *bp)
3321 {
3322 	vm_object_t object;
3323 	int i;
3324 
3325 	object = bp->b_bufobj->bo_object;
3326 	VM_OBJECT_ASSERT_WLOCKED(object);
3327 
3328 	/*
3329 	 * We qualify the scan for modified pages on whether the
3330 	 * object has been flushed yet.
3331 	 */
3332 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3333 		vm_offset_t boffset;
3334 		vm_offset_t eoffset;
3335 
3336 		/*
3337 		 * test the pages to see if they have been modified directly
3338 		 * by users through the VM system.
3339 		 */
3340 		for (i = 0; i < bp->b_npages; i++)
3341 			vm_page_test_dirty(bp->b_pages[i]);
3342 
3343 		/*
3344 		 * Calculate the encompassing dirty range, boffset and eoffset,
3345 		 * (eoffset - boffset) bytes.
3346 		 */
3347 
3348 		for (i = 0; i < bp->b_npages; i++) {
3349 			if (bp->b_pages[i]->dirty)
3350 				break;
3351 		}
3352 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3353 
3354 		for (i = bp->b_npages - 1; i >= 0; --i) {
3355 			if (bp->b_pages[i]->dirty) {
3356 				break;
3357 			}
3358 		}
3359 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3360 
3361 		/*
3362 		 * Fit it to the buffer.
3363 		 */
3364 
3365 		if (eoffset > bp->b_bcount)
3366 			eoffset = bp->b_bcount;
3367 
3368 		/*
3369 		 * If we have a good dirty range, merge with the existing
3370 		 * dirty range.
3371 		 */
3372 
3373 		if (boffset < eoffset) {
3374 			if (bp->b_dirtyoff > boffset)
3375 				bp->b_dirtyoff = boffset;
3376 			if (bp->b_dirtyend < eoffset)
3377 				bp->b_dirtyend = eoffset;
3378 		}
3379 	}
3380 }
3381 
3382 /*
3383  * Allocate the KVA mapping for an existing buffer.
3384  * If an unmapped buffer is provided but a mapped buffer is requested, take
3385  * also care to properly setup mappings between pages and KVA.
3386  */
3387 static void
3388 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3389 {
3390 	int bsize, maxsize, need_mapping, need_kva;
3391 	off_t offset;
3392 
3393 	need_mapping = bp->b_data == unmapped_buf &&
3394 	    (gbflags & GB_UNMAPPED) == 0;
3395 	need_kva = bp->b_kvabase == unmapped_buf &&
3396 	    bp->b_data == unmapped_buf &&
3397 	    (gbflags & GB_KVAALLOC) != 0;
3398 	if (!need_mapping && !need_kva)
3399 		return;
3400 
3401 	BUF_CHECK_UNMAPPED(bp);
3402 
3403 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3404 		/*
3405 		 * Buffer is not mapped, but the KVA was already
3406 		 * reserved at the time of the instantiation.  Use the
3407 		 * allocated space.
3408 		 */
3409 		goto has_addr;
3410 	}
3411 
3412 	/*
3413 	 * Calculate the amount of the address space we would reserve
3414 	 * if the buffer was mapped.
3415 	 */
3416 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3417 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3418 	offset = blkno * bsize;
3419 	maxsize = size + (offset & PAGE_MASK);
3420 	maxsize = imax(maxsize, bsize);
3421 
3422 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3423 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3424 			/*
3425 			 * XXXKIB: defragmentation cannot
3426 			 * succeed, not sure what else to do.
3427 			 */
3428 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3429 		}
3430 		atomic_add_int(&mappingrestarts, 1);
3431 		bufspace_wait(bp->b_vp, gbflags, 0, 0);
3432 	}
3433 has_addr:
3434 	if (need_mapping) {
3435 		/* b_offset is handled by bpmap_qenter. */
3436 		bp->b_data = bp->b_kvabase;
3437 		BUF_CHECK_MAPPED(bp);
3438 		bpmap_qenter(bp);
3439 	}
3440 }
3441 
3442 /*
3443  *	getblk:
3444  *
3445  *	Get a block given a specified block and offset into a file/device.
3446  *	The buffers B_DONE bit will be cleared on return, making it almost
3447  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3448  *	return.  The caller should clear B_INVAL prior to initiating a
3449  *	READ.
3450  *
3451  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3452  *	an existing buffer.
3453  *
3454  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3455  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3456  *	and then cleared based on the backing VM.  If the previous buffer is
3457  *	non-0-sized but invalid, B_CACHE will be cleared.
3458  *
3459  *	If getblk() must create a new buffer, the new buffer is returned with
3460  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3461  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3462  *	backing VM.
3463  *
3464  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3465  *	B_CACHE bit is clear.
3466  *
3467  *	What this means, basically, is that the caller should use B_CACHE to
3468  *	determine whether the buffer is fully valid or not and should clear
3469  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3470  *	the buffer by loading its data area with something, the caller needs
3471  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3472  *	the caller should set B_CACHE ( as an optimization ), else the caller
3473  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3474  *	a write attempt or if it was a successfull read.  If the caller
3475  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3476  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3477  */
3478 struct buf *
3479 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3480     int flags)
3481 {
3482 	struct buf *bp;
3483 	struct bufobj *bo;
3484 	int bsize, error, maxsize, vmio;
3485 	off_t offset;
3486 
3487 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3488 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3489 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3490 	ASSERT_VOP_LOCKED(vp, "getblk");
3491 	if (size > MAXBCACHEBUF)
3492 		panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size,
3493 		    MAXBCACHEBUF);
3494 	if (!unmapped_buf_allowed)
3495 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3496 
3497 	bo = &vp->v_bufobj;
3498 loop:
3499 	BO_RLOCK(bo);
3500 	bp = gbincore(bo, blkno);
3501 	if (bp != NULL) {
3502 		int lockflags;
3503 		/*
3504 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3505 		 * it must be on a queue.
3506 		 */
3507 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3508 
3509 		if (flags & GB_LOCK_NOWAIT)
3510 			lockflags |= LK_NOWAIT;
3511 
3512 		error = BUF_TIMELOCK(bp, lockflags,
3513 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3514 
3515 		/*
3516 		 * If we slept and got the lock we have to restart in case
3517 		 * the buffer changed identities.
3518 		 */
3519 		if (error == ENOLCK)
3520 			goto loop;
3521 		/* We timed out or were interrupted. */
3522 		else if (error)
3523 			return (NULL);
3524 		/* If recursed, assume caller knows the rules. */
3525 		else if (BUF_LOCKRECURSED(bp))
3526 			goto end;
3527 
3528 		/*
3529 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3530 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3531 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3532 		 * backing VM cache.
3533 		 */
3534 		if (bp->b_flags & B_INVAL)
3535 			bp->b_flags &= ~B_CACHE;
3536 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3537 			bp->b_flags |= B_CACHE;
3538 		if (bp->b_flags & B_MANAGED)
3539 			MPASS(bp->b_qindex == QUEUE_NONE);
3540 		else
3541 			bremfree(bp);
3542 
3543 		/*
3544 		 * check for size inconsistencies for non-VMIO case.
3545 		 */
3546 		if (bp->b_bcount != size) {
3547 			if ((bp->b_flags & B_VMIO) == 0 ||
3548 			    (size > bp->b_kvasize)) {
3549 				if (bp->b_flags & B_DELWRI) {
3550 					/*
3551 					 * If buffer is pinned and caller does
3552 					 * not want sleep  waiting for it to be
3553 					 * unpinned, bail out
3554 					 * */
3555 					if (bp->b_pin_count > 0) {
3556 						if (flags & GB_LOCK_NOWAIT) {
3557 							bqrelse(bp);
3558 							return (NULL);
3559 						} else {
3560 							bunpin_wait(bp);
3561 						}
3562 					}
3563 					bp->b_flags |= B_NOCACHE;
3564 					bwrite(bp);
3565 				} else {
3566 					if (LIST_EMPTY(&bp->b_dep)) {
3567 						bp->b_flags |= B_RELBUF;
3568 						brelse(bp);
3569 					} else {
3570 						bp->b_flags |= B_NOCACHE;
3571 						bwrite(bp);
3572 					}
3573 				}
3574 				goto loop;
3575 			}
3576 		}
3577 
3578 		/*
3579 		 * Handle the case of unmapped buffer which should
3580 		 * become mapped, or the buffer for which KVA
3581 		 * reservation is requested.
3582 		 */
3583 		bp_unmapped_get_kva(bp, blkno, size, flags);
3584 
3585 		/*
3586 		 * If the size is inconsistant in the VMIO case, we can resize
3587 		 * the buffer.  This might lead to B_CACHE getting set or
3588 		 * cleared.  If the size has not changed, B_CACHE remains
3589 		 * unchanged from its previous state.
3590 		 */
3591 		allocbuf(bp, size);
3592 
3593 		KASSERT(bp->b_offset != NOOFFSET,
3594 		    ("getblk: no buffer offset"));
3595 
3596 		/*
3597 		 * A buffer with B_DELWRI set and B_CACHE clear must
3598 		 * be committed before we can return the buffer in
3599 		 * order to prevent the caller from issuing a read
3600 		 * ( due to B_CACHE not being set ) and overwriting
3601 		 * it.
3602 		 *
3603 		 * Most callers, including NFS and FFS, need this to
3604 		 * operate properly either because they assume they
3605 		 * can issue a read if B_CACHE is not set, or because
3606 		 * ( for example ) an uncached B_DELWRI might loop due
3607 		 * to softupdates re-dirtying the buffer.  In the latter
3608 		 * case, B_CACHE is set after the first write completes,
3609 		 * preventing further loops.
3610 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3611 		 * above while extending the buffer, we cannot allow the
3612 		 * buffer to remain with B_CACHE set after the write
3613 		 * completes or it will represent a corrupt state.  To
3614 		 * deal with this we set B_NOCACHE to scrap the buffer
3615 		 * after the write.
3616 		 *
3617 		 * We might be able to do something fancy, like setting
3618 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3619 		 * so the below call doesn't set B_CACHE, but that gets real
3620 		 * confusing.  This is much easier.
3621 		 */
3622 
3623 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3624 			bp->b_flags |= B_NOCACHE;
3625 			bwrite(bp);
3626 			goto loop;
3627 		}
3628 		bp->b_flags &= ~B_DONE;
3629 	} else {
3630 		/*
3631 		 * Buffer is not in-core, create new buffer.  The buffer
3632 		 * returned by getnewbuf() is locked.  Note that the returned
3633 		 * buffer is also considered valid (not marked B_INVAL).
3634 		 */
3635 		BO_RUNLOCK(bo);
3636 		/*
3637 		 * If the user does not want us to create the buffer, bail out
3638 		 * here.
3639 		 */
3640 		if (flags & GB_NOCREAT)
3641 			return NULL;
3642 		if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3643 			return NULL;
3644 
3645 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3646 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3647 		offset = blkno * bsize;
3648 		vmio = vp->v_object != NULL;
3649 		if (vmio) {
3650 			maxsize = size + (offset & PAGE_MASK);
3651 		} else {
3652 			maxsize = size;
3653 			/* Do not allow non-VMIO notmapped buffers. */
3654 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3655 		}
3656 		maxsize = imax(maxsize, bsize);
3657 
3658 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
3659 		if (bp == NULL) {
3660 			if (slpflag || slptimeo)
3661 				return NULL;
3662 			/*
3663 			 * XXX This is here until the sleep path is diagnosed
3664 			 * enough to work under very low memory conditions.
3665 			 *
3666 			 * There's an issue on low memory, 4BSD+non-preempt
3667 			 * systems (eg MIPS routers with 32MB RAM) where buffer
3668 			 * exhaustion occurs without sleeping for buffer
3669 			 * reclaimation.  This just sticks in a loop and
3670 			 * constantly attempts to allocate a buffer, which
3671 			 * hits exhaustion and tries to wakeup bufdaemon.
3672 			 * This never happens because we never yield.
3673 			 *
3674 			 * The real solution is to identify and fix these cases
3675 			 * so we aren't effectively busy-waiting in a loop
3676 			 * until the reclaimation path has cycles to run.
3677 			 */
3678 			kern_yield(PRI_USER);
3679 			goto loop;
3680 		}
3681 
3682 		/*
3683 		 * This code is used to make sure that a buffer is not
3684 		 * created while the getnewbuf routine is blocked.
3685 		 * This can be a problem whether the vnode is locked or not.
3686 		 * If the buffer is created out from under us, we have to
3687 		 * throw away the one we just created.
3688 		 *
3689 		 * Note: this must occur before we associate the buffer
3690 		 * with the vp especially considering limitations in
3691 		 * the splay tree implementation when dealing with duplicate
3692 		 * lblkno's.
3693 		 */
3694 		BO_LOCK(bo);
3695 		if (gbincore(bo, blkno)) {
3696 			BO_UNLOCK(bo);
3697 			bp->b_flags |= B_INVAL;
3698 			brelse(bp);
3699 			bufspace_release(maxsize);
3700 			goto loop;
3701 		}
3702 
3703 		/*
3704 		 * Insert the buffer into the hash, so that it can
3705 		 * be found by incore.
3706 		 */
3707 		bp->b_blkno = bp->b_lblkno = blkno;
3708 		bp->b_offset = offset;
3709 		bgetvp(vp, bp);
3710 		BO_UNLOCK(bo);
3711 
3712 		/*
3713 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
3714 		 * buffer size starts out as 0, B_CACHE will be set by
3715 		 * allocbuf() for the VMIO case prior to it testing the
3716 		 * backing store for validity.
3717 		 */
3718 
3719 		if (vmio) {
3720 			bp->b_flags |= B_VMIO;
3721 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3722 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3723 			    bp, vp->v_object, bp->b_bufobj->bo_object));
3724 		} else {
3725 			bp->b_flags &= ~B_VMIO;
3726 			KASSERT(bp->b_bufobj->bo_object == NULL,
3727 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
3728 			    bp, bp->b_bufobj->bo_object));
3729 			BUF_CHECK_MAPPED(bp);
3730 		}
3731 
3732 		allocbuf(bp, size);
3733 		bufspace_release(maxsize);
3734 		bp->b_flags &= ~B_DONE;
3735 	}
3736 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3737 	BUF_ASSERT_HELD(bp);
3738 end:
3739 	KASSERT(bp->b_bufobj == bo,
3740 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3741 	return (bp);
3742 }
3743 
3744 /*
3745  * Get an empty, disassociated buffer of given size.  The buffer is initially
3746  * set to B_INVAL.
3747  */
3748 struct buf *
3749 geteblk(int size, int flags)
3750 {
3751 	struct buf *bp;
3752 	int maxsize;
3753 
3754 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
3755 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
3756 		if ((flags & GB_NOWAIT_BD) &&
3757 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
3758 			return (NULL);
3759 	}
3760 	allocbuf(bp, size);
3761 	bufspace_release(maxsize);
3762 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
3763 	BUF_ASSERT_HELD(bp);
3764 	return (bp);
3765 }
3766 
3767 /*
3768  * Truncate the backing store for a non-vmio buffer.
3769  */
3770 static void
3771 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
3772 {
3773 
3774 	if (bp->b_flags & B_MALLOC) {
3775 		/*
3776 		 * malloced buffers are not shrunk
3777 		 */
3778 		if (newbsize == 0) {
3779 			bufmallocadjust(bp, 0);
3780 			free(bp->b_data, M_BIOBUF);
3781 			bp->b_data = bp->b_kvabase;
3782 			bp->b_flags &= ~B_MALLOC;
3783 		}
3784 		return;
3785 	}
3786 	vm_hold_free_pages(bp, newbsize);
3787 	bufspace_adjust(bp, newbsize);
3788 }
3789 
3790 /*
3791  * Extend the backing for a non-VMIO buffer.
3792  */
3793 static void
3794 vfs_nonvmio_extend(struct buf *bp, int newbsize)
3795 {
3796 	caddr_t origbuf;
3797 	int origbufsize;
3798 
3799 	/*
3800 	 * We only use malloced memory on the first allocation.
3801 	 * and revert to page-allocated memory when the buffer
3802 	 * grows.
3803 	 *
3804 	 * There is a potential smp race here that could lead
3805 	 * to bufmallocspace slightly passing the max.  It
3806 	 * is probably extremely rare and not worth worrying
3807 	 * over.
3808 	 */
3809 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
3810 	    bufmallocspace < maxbufmallocspace) {
3811 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
3812 		bp->b_flags |= B_MALLOC;
3813 		bufmallocadjust(bp, newbsize);
3814 		return;
3815 	}
3816 
3817 	/*
3818 	 * If the buffer is growing on its other-than-first
3819 	 * allocation then we revert to the page-allocation
3820 	 * scheme.
3821 	 */
3822 	origbuf = NULL;
3823 	origbufsize = 0;
3824 	if (bp->b_flags & B_MALLOC) {
3825 		origbuf = bp->b_data;
3826 		origbufsize = bp->b_bufsize;
3827 		bp->b_data = bp->b_kvabase;
3828 		bufmallocadjust(bp, 0);
3829 		bp->b_flags &= ~B_MALLOC;
3830 		newbsize = round_page(newbsize);
3831 	}
3832 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
3833 	    (vm_offset_t) bp->b_data + newbsize);
3834 	if (origbuf != NULL) {
3835 		bcopy(origbuf, bp->b_data, origbufsize);
3836 		free(origbuf, M_BIOBUF);
3837 	}
3838 	bufspace_adjust(bp, newbsize);
3839 }
3840 
3841 /*
3842  * This code constitutes the buffer memory from either anonymous system
3843  * memory (in the case of non-VMIO operations) or from an associated
3844  * VM object (in the case of VMIO operations).  This code is able to
3845  * resize a buffer up or down.
3846  *
3847  * Note that this code is tricky, and has many complications to resolve
3848  * deadlock or inconsistant data situations.  Tread lightly!!!
3849  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3850  * the caller.  Calling this code willy nilly can result in the loss of data.
3851  *
3852  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
3853  * B_CACHE for the non-VMIO case.
3854  */
3855 int
3856 allocbuf(struct buf *bp, int size)
3857 {
3858 	int newbsize;
3859 
3860 	BUF_ASSERT_HELD(bp);
3861 
3862 	if (bp->b_bcount == size)
3863 		return (1);
3864 
3865 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
3866 		panic("allocbuf: buffer too small");
3867 
3868 	newbsize = roundup2(size, DEV_BSIZE);
3869 	if ((bp->b_flags & B_VMIO) == 0) {
3870 		if ((bp->b_flags & B_MALLOC) == 0)
3871 			newbsize = round_page(newbsize);
3872 		/*
3873 		 * Just get anonymous memory from the kernel.  Don't
3874 		 * mess with B_CACHE.
3875 		 */
3876 		if (newbsize < bp->b_bufsize)
3877 			vfs_nonvmio_truncate(bp, newbsize);
3878 		else if (newbsize > bp->b_bufsize)
3879 			vfs_nonvmio_extend(bp, newbsize);
3880 	} else {
3881 		int desiredpages;
3882 
3883 		desiredpages = (size == 0) ? 0 :
3884 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3885 
3886 		if (bp->b_flags & B_MALLOC)
3887 			panic("allocbuf: VMIO buffer can't be malloced");
3888 		/*
3889 		 * Set B_CACHE initially if buffer is 0 length or will become
3890 		 * 0-length.
3891 		 */
3892 		if (size == 0 || bp->b_bufsize == 0)
3893 			bp->b_flags |= B_CACHE;
3894 
3895 		if (newbsize < bp->b_bufsize)
3896 			vfs_vmio_truncate(bp, desiredpages);
3897 		/* XXX This looks as if it should be newbsize > b_bufsize */
3898 		else if (size > bp->b_bcount)
3899 			vfs_vmio_extend(bp, desiredpages, size);
3900 		bufspace_adjust(bp, newbsize);
3901 	}
3902 	bp->b_bcount = size;		/* requested buffer size. */
3903 	return (1);
3904 }
3905 
3906 extern int inflight_transient_maps;
3907 
3908 void
3909 biodone(struct bio *bp)
3910 {
3911 	struct mtx *mtxp;
3912 	void (*done)(struct bio *);
3913 	vm_offset_t start, end;
3914 
3915 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3916 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3917 		bp->bio_flags |= BIO_UNMAPPED;
3918 		start = trunc_page((vm_offset_t)bp->bio_data);
3919 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3920 		bp->bio_data = unmapped_buf;
3921 		pmap_qremove(start, OFF_TO_IDX(end - start));
3922 		vmem_free(transient_arena, start, end - start);
3923 		atomic_add_int(&inflight_transient_maps, -1);
3924 	}
3925 	done = bp->bio_done;
3926 	if (done == NULL) {
3927 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
3928 		mtx_lock(mtxp);
3929 		bp->bio_flags |= BIO_DONE;
3930 		wakeup(bp);
3931 		mtx_unlock(mtxp);
3932 	} else {
3933 		bp->bio_flags |= BIO_DONE;
3934 		done(bp);
3935 	}
3936 }
3937 
3938 /*
3939  * Wait for a BIO to finish.
3940  */
3941 int
3942 biowait(struct bio *bp, const char *wchan)
3943 {
3944 	struct mtx *mtxp;
3945 
3946 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3947 	mtx_lock(mtxp);
3948 	while ((bp->bio_flags & BIO_DONE) == 0)
3949 		msleep(bp, mtxp, PRIBIO, wchan, 0);
3950 	mtx_unlock(mtxp);
3951 	if (bp->bio_error != 0)
3952 		return (bp->bio_error);
3953 	if (!(bp->bio_flags & BIO_ERROR))
3954 		return (0);
3955 	return (EIO);
3956 }
3957 
3958 void
3959 biofinish(struct bio *bp, struct devstat *stat, int error)
3960 {
3961 
3962 	if (error) {
3963 		bp->bio_error = error;
3964 		bp->bio_flags |= BIO_ERROR;
3965 	}
3966 	if (stat != NULL)
3967 		devstat_end_transaction_bio(stat, bp);
3968 	biodone(bp);
3969 }
3970 
3971 /*
3972  *	bufwait:
3973  *
3974  *	Wait for buffer I/O completion, returning error status.  The buffer
3975  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3976  *	error and cleared.
3977  */
3978 int
3979 bufwait(struct buf *bp)
3980 {
3981 	if (bp->b_iocmd == BIO_READ)
3982 		bwait(bp, PRIBIO, "biord");
3983 	else
3984 		bwait(bp, PRIBIO, "biowr");
3985 	if (bp->b_flags & B_EINTR) {
3986 		bp->b_flags &= ~B_EINTR;
3987 		return (EINTR);
3988 	}
3989 	if (bp->b_ioflags & BIO_ERROR) {
3990 		return (bp->b_error ? bp->b_error : EIO);
3991 	} else {
3992 		return (0);
3993 	}
3994 }
3995 
3996 /*
3997  *	bufdone:
3998  *
3999  *	Finish I/O on a buffer, optionally calling a completion function.
4000  *	This is usually called from an interrupt so process blocking is
4001  *	not allowed.
4002  *
4003  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4004  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4005  *	assuming B_INVAL is clear.
4006  *
4007  *	For the VMIO case, we set B_CACHE if the op was a read and no
4008  *	read error occured, or if the op was a write.  B_CACHE is never
4009  *	set if the buffer is invalid or otherwise uncacheable.
4010  *
4011  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
4012  *	initiator to leave B_INVAL set to brelse the buffer out of existance
4013  *	in the biodone routine.
4014  */
4015 void
4016 bufdone(struct buf *bp)
4017 {
4018 	struct bufobj *dropobj;
4019 	void    (*biodone)(struct buf *);
4020 
4021 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4022 	dropobj = NULL;
4023 
4024 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4025 	BUF_ASSERT_HELD(bp);
4026 
4027 	runningbufwakeup(bp);
4028 	if (bp->b_iocmd == BIO_WRITE)
4029 		dropobj = bp->b_bufobj;
4030 	/* call optional completion function if requested */
4031 	if (bp->b_iodone != NULL) {
4032 		biodone = bp->b_iodone;
4033 		bp->b_iodone = NULL;
4034 		(*biodone) (bp);
4035 		if (dropobj)
4036 			bufobj_wdrop(dropobj);
4037 		return;
4038 	}
4039 
4040 	bufdone_finish(bp);
4041 
4042 	if (dropobj)
4043 		bufobj_wdrop(dropobj);
4044 }
4045 
4046 void
4047 bufdone_finish(struct buf *bp)
4048 {
4049 	BUF_ASSERT_HELD(bp);
4050 
4051 	if (!LIST_EMPTY(&bp->b_dep))
4052 		buf_complete(bp);
4053 
4054 	if (bp->b_flags & B_VMIO) {
4055 		/*
4056 		 * Set B_CACHE if the op was a normal read and no error
4057 		 * occured.  B_CACHE is set for writes in the b*write()
4058 		 * routines.
4059 		 */
4060 		if (bp->b_iocmd == BIO_READ &&
4061 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4062 		    !(bp->b_ioflags & BIO_ERROR))
4063 			bp->b_flags |= B_CACHE;
4064 		vfs_vmio_iodone(bp);
4065 	}
4066 
4067 	/*
4068 	 * For asynchronous completions, release the buffer now. The brelse
4069 	 * will do a wakeup there if necessary - so no need to do a wakeup
4070 	 * here in the async case. The sync case always needs to do a wakeup.
4071 	 */
4072 	if (bp->b_flags & B_ASYNC) {
4073 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4074 		    (bp->b_ioflags & BIO_ERROR))
4075 			brelse(bp);
4076 		else
4077 			bqrelse(bp);
4078 	} else
4079 		bdone(bp);
4080 }
4081 
4082 /*
4083  * This routine is called in lieu of iodone in the case of
4084  * incomplete I/O.  This keeps the busy status for pages
4085  * consistant.
4086  */
4087 void
4088 vfs_unbusy_pages(struct buf *bp)
4089 {
4090 	int i;
4091 	vm_object_t obj;
4092 	vm_page_t m;
4093 
4094 	runningbufwakeup(bp);
4095 	if (!(bp->b_flags & B_VMIO))
4096 		return;
4097 
4098 	obj = bp->b_bufobj->bo_object;
4099 	VM_OBJECT_WLOCK(obj);
4100 	for (i = 0; i < bp->b_npages; i++) {
4101 		m = bp->b_pages[i];
4102 		if (m == bogus_page) {
4103 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4104 			if (!m)
4105 				panic("vfs_unbusy_pages: page missing\n");
4106 			bp->b_pages[i] = m;
4107 			if (buf_mapped(bp)) {
4108 				BUF_CHECK_MAPPED(bp);
4109 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4110 				    bp->b_pages, bp->b_npages);
4111 			} else
4112 				BUF_CHECK_UNMAPPED(bp);
4113 		}
4114 		vm_page_sunbusy(m);
4115 	}
4116 	vm_object_pip_wakeupn(obj, bp->b_npages);
4117 	VM_OBJECT_WUNLOCK(obj);
4118 }
4119 
4120 /*
4121  * vfs_page_set_valid:
4122  *
4123  *	Set the valid bits in a page based on the supplied offset.   The
4124  *	range is restricted to the buffer's size.
4125  *
4126  *	This routine is typically called after a read completes.
4127  */
4128 static void
4129 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4130 {
4131 	vm_ooffset_t eoff;
4132 
4133 	/*
4134 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4135 	 * page boundary and eoff is not greater than the end of the buffer.
4136 	 * The end of the buffer, in this case, is our file EOF, not the
4137 	 * allocation size of the buffer.
4138 	 */
4139 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4140 	if (eoff > bp->b_offset + bp->b_bcount)
4141 		eoff = bp->b_offset + bp->b_bcount;
4142 
4143 	/*
4144 	 * Set valid range.  This is typically the entire buffer and thus the
4145 	 * entire page.
4146 	 */
4147 	if (eoff > off)
4148 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4149 }
4150 
4151 /*
4152  * vfs_page_set_validclean:
4153  *
4154  *	Set the valid bits and clear the dirty bits in a page based on the
4155  *	supplied offset.   The range is restricted to the buffer's size.
4156  */
4157 static void
4158 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4159 {
4160 	vm_ooffset_t soff, eoff;
4161 
4162 	/*
4163 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4164 	 * page boundry or cross the end of the buffer.  The end of the
4165 	 * buffer, in this case, is our file EOF, not the allocation size
4166 	 * of the buffer.
4167 	 */
4168 	soff = off;
4169 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4170 	if (eoff > bp->b_offset + bp->b_bcount)
4171 		eoff = bp->b_offset + bp->b_bcount;
4172 
4173 	/*
4174 	 * Set valid range.  This is typically the entire buffer and thus the
4175 	 * entire page.
4176 	 */
4177 	if (eoff > soff) {
4178 		vm_page_set_validclean(
4179 		    m,
4180 		   (vm_offset_t) (soff & PAGE_MASK),
4181 		   (vm_offset_t) (eoff - soff)
4182 		);
4183 	}
4184 }
4185 
4186 /*
4187  * Ensure that all buffer pages are not exclusive busied.  If any page is
4188  * exclusive busy, drain it.
4189  */
4190 void
4191 vfs_drain_busy_pages(struct buf *bp)
4192 {
4193 	vm_page_t m;
4194 	int i, last_busied;
4195 
4196 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4197 	last_busied = 0;
4198 	for (i = 0; i < bp->b_npages; i++) {
4199 		m = bp->b_pages[i];
4200 		if (vm_page_xbusied(m)) {
4201 			for (; last_busied < i; last_busied++)
4202 				vm_page_sbusy(bp->b_pages[last_busied]);
4203 			while (vm_page_xbusied(m)) {
4204 				vm_page_lock(m);
4205 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4206 				vm_page_busy_sleep(m, "vbpage");
4207 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4208 			}
4209 		}
4210 	}
4211 	for (i = 0; i < last_busied; i++)
4212 		vm_page_sunbusy(bp->b_pages[i]);
4213 }
4214 
4215 /*
4216  * This routine is called before a device strategy routine.
4217  * It is used to tell the VM system that paging I/O is in
4218  * progress, and treat the pages associated with the buffer
4219  * almost as being exclusive busy.  Also the object paging_in_progress
4220  * flag is handled to make sure that the object doesn't become
4221  * inconsistant.
4222  *
4223  * Since I/O has not been initiated yet, certain buffer flags
4224  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
4225  * and should be ignored.
4226  */
4227 void
4228 vfs_busy_pages(struct buf *bp, int clear_modify)
4229 {
4230 	int i, bogus;
4231 	vm_object_t obj;
4232 	vm_ooffset_t foff;
4233 	vm_page_t m;
4234 
4235 	if (!(bp->b_flags & B_VMIO))
4236 		return;
4237 
4238 	obj = bp->b_bufobj->bo_object;
4239 	foff = bp->b_offset;
4240 	KASSERT(bp->b_offset != NOOFFSET,
4241 	    ("vfs_busy_pages: no buffer offset"));
4242 	VM_OBJECT_WLOCK(obj);
4243 	vfs_drain_busy_pages(bp);
4244 	if (bp->b_bufsize != 0)
4245 		vfs_setdirty_locked_object(bp);
4246 	bogus = 0;
4247 	for (i = 0; i < bp->b_npages; i++) {
4248 		m = bp->b_pages[i];
4249 
4250 		if ((bp->b_flags & B_CLUSTER) == 0) {
4251 			vm_object_pip_add(obj, 1);
4252 			vm_page_sbusy(m);
4253 		}
4254 		/*
4255 		 * When readying a buffer for a read ( i.e
4256 		 * clear_modify == 0 ), it is important to do
4257 		 * bogus_page replacement for valid pages in
4258 		 * partially instantiated buffers.  Partially
4259 		 * instantiated buffers can, in turn, occur when
4260 		 * reconstituting a buffer from its VM backing store
4261 		 * base.  We only have to do this if B_CACHE is
4262 		 * clear ( which causes the I/O to occur in the
4263 		 * first place ).  The replacement prevents the read
4264 		 * I/O from overwriting potentially dirty VM-backed
4265 		 * pages.  XXX bogus page replacement is, uh, bogus.
4266 		 * It may not work properly with small-block devices.
4267 		 * We need to find a better way.
4268 		 */
4269 		if (clear_modify) {
4270 			pmap_remove_write(m);
4271 			vfs_page_set_validclean(bp, foff, m);
4272 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4273 		    (bp->b_flags & B_CACHE) == 0) {
4274 			bp->b_pages[i] = bogus_page;
4275 			bogus++;
4276 		}
4277 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4278 	}
4279 	VM_OBJECT_WUNLOCK(obj);
4280 	if (bogus && buf_mapped(bp)) {
4281 		BUF_CHECK_MAPPED(bp);
4282 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4283 		    bp->b_pages, bp->b_npages);
4284 	}
4285 }
4286 
4287 /*
4288  *	vfs_bio_set_valid:
4289  *
4290  *	Set the range within the buffer to valid.  The range is
4291  *	relative to the beginning of the buffer, b_offset.  Note that
4292  *	b_offset itself may be offset from the beginning of the first
4293  *	page.
4294  */
4295 void
4296 vfs_bio_set_valid(struct buf *bp, int base, int size)
4297 {
4298 	int i, n;
4299 	vm_page_t m;
4300 
4301 	if (!(bp->b_flags & B_VMIO))
4302 		return;
4303 
4304 	/*
4305 	 * Fixup base to be relative to beginning of first page.
4306 	 * Set initial n to be the maximum number of bytes in the
4307 	 * first page that can be validated.
4308 	 */
4309 	base += (bp->b_offset & PAGE_MASK);
4310 	n = PAGE_SIZE - (base & PAGE_MASK);
4311 
4312 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4313 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4314 		m = bp->b_pages[i];
4315 		if (n > size)
4316 			n = size;
4317 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4318 		base += n;
4319 		size -= n;
4320 		n = PAGE_SIZE;
4321 	}
4322 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4323 }
4324 
4325 /*
4326  *	vfs_bio_clrbuf:
4327  *
4328  *	If the specified buffer is a non-VMIO buffer, clear the entire
4329  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4330  *	validate only the previously invalid portions of the buffer.
4331  *	This routine essentially fakes an I/O, so we need to clear
4332  *	BIO_ERROR and B_INVAL.
4333  *
4334  *	Note that while we only theoretically need to clear through b_bcount,
4335  *	we go ahead and clear through b_bufsize.
4336  */
4337 void
4338 vfs_bio_clrbuf(struct buf *bp)
4339 {
4340 	int i, j, mask, sa, ea, slide;
4341 
4342 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4343 		clrbuf(bp);
4344 		return;
4345 	}
4346 	bp->b_flags &= ~B_INVAL;
4347 	bp->b_ioflags &= ~BIO_ERROR;
4348 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4349 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4350 	    (bp->b_offset & PAGE_MASK) == 0) {
4351 		if (bp->b_pages[0] == bogus_page)
4352 			goto unlock;
4353 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4354 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4355 		if ((bp->b_pages[0]->valid & mask) == mask)
4356 			goto unlock;
4357 		if ((bp->b_pages[0]->valid & mask) == 0) {
4358 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4359 			bp->b_pages[0]->valid |= mask;
4360 			goto unlock;
4361 		}
4362 	}
4363 	sa = bp->b_offset & PAGE_MASK;
4364 	slide = 0;
4365 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4366 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4367 		ea = slide & PAGE_MASK;
4368 		if (ea == 0)
4369 			ea = PAGE_SIZE;
4370 		if (bp->b_pages[i] == bogus_page)
4371 			continue;
4372 		j = sa / DEV_BSIZE;
4373 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4374 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4375 		if ((bp->b_pages[i]->valid & mask) == mask)
4376 			continue;
4377 		if ((bp->b_pages[i]->valid & mask) == 0)
4378 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4379 		else {
4380 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4381 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4382 					pmap_zero_page_area(bp->b_pages[i],
4383 					    sa, DEV_BSIZE);
4384 				}
4385 			}
4386 		}
4387 		bp->b_pages[i]->valid |= mask;
4388 	}
4389 unlock:
4390 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4391 	bp->b_resid = 0;
4392 }
4393 
4394 void
4395 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4396 {
4397 	vm_page_t m;
4398 	int i, n;
4399 
4400 	if (buf_mapped(bp)) {
4401 		BUF_CHECK_MAPPED(bp);
4402 		bzero(bp->b_data + base, size);
4403 	} else {
4404 		BUF_CHECK_UNMAPPED(bp);
4405 		n = PAGE_SIZE - (base & PAGE_MASK);
4406 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4407 			m = bp->b_pages[i];
4408 			if (n > size)
4409 				n = size;
4410 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4411 			base += n;
4412 			size -= n;
4413 			n = PAGE_SIZE;
4414 		}
4415 	}
4416 }
4417 
4418 /*
4419  * vm_hold_load_pages and vm_hold_free_pages get pages into
4420  * a buffers address space.  The pages are anonymous and are
4421  * not associated with a file object.
4422  */
4423 static void
4424 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4425 {
4426 	vm_offset_t pg;
4427 	vm_page_t p;
4428 	int index;
4429 
4430 	BUF_CHECK_MAPPED(bp);
4431 
4432 	to = round_page(to);
4433 	from = round_page(from);
4434 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4435 
4436 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4437 tryagain:
4438 		/*
4439 		 * note: must allocate system pages since blocking here
4440 		 * could interfere with paging I/O, no matter which
4441 		 * process we are.
4442 		 */
4443 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4444 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4445 		if (p == NULL) {
4446 			VM_WAIT;
4447 			goto tryagain;
4448 		}
4449 		pmap_qenter(pg, &p, 1);
4450 		bp->b_pages[index] = p;
4451 	}
4452 	bp->b_npages = index;
4453 }
4454 
4455 /* Return pages associated with this buf to the vm system */
4456 static void
4457 vm_hold_free_pages(struct buf *bp, int newbsize)
4458 {
4459 	vm_offset_t from;
4460 	vm_page_t p;
4461 	int index, newnpages;
4462 
4463 	BUF_CHECK_MAPPED(bp);
4464 
4465 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4466 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4467 	if (bp->b_npages > newnpages)
4468 		pmap_qremove(from, bp->b_npages - newnpages);
4469 	for (index = newnpages; index < bp->b_npages; index++) {
4470 		p = bp->b_pages[index];
4471 		bp->b_pages[index] = NULL;
4472 		if (vm_page_sbusied(p))
4473 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4474 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4475 		p->wire_count--;
4476 		vm_page_free(p);
4477 		atomic_subtract_int(&vm_cnt.v_wire_count, 1);
4478 	}
4479 	bp->b_npages = newnpages;
4480 }
4481 
4482 /*
4483  * Map an IO request into kernel virtual address space.
4484  *
4485  * All requests are (re)mapped into kernel VA space.
4486  * Notice that we use b_bufsize for the size of the buffer
4487  * to be mapped.  b_bcount might be modified by the driver.
4488  *
4489  * Note that even if the caller determines that the address space should
4490  * be valid, a race or a smaller-file mapped into a larger space may
4491  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4492  * check the return value.
4493  *
4494  * This function only works with pager buffers.
4495  */
4496 int
4497 vmapbuf(struct buf *bp, int mapbuf)
4498 {
4499 	vm_prot_t prot;
4500 	int pidx;
4501 
4502 	if (bp->b_bufsize < 0)
4503 		return (-1);
4504 	prot = VM_PROT_READ;
4505 	if (bp->b_iocmd == BIO_READ)
4506 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4507 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4508 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4509 	    btoc(MAXPHYS))) < 0)
4510 		return (-1);
4511 	bp->b_npages = pidx;
4512 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4513 	if (mapbuf || !unmapped_buf_allowed) {
4514 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4515 		bp->b_data = bp->b_kvabase + bp->b_offset;
4516 	} else
4517 		bp->b_data = unmapped_buf;
4518 	return(0);
4519 }
4520 
4521 /*
4522  * Free the io map PTEs associated with this IO operation.
4523  * We also invalidate the TLB entries and restore the original b_addr.
4524  *
4525  * This function only works with pager buffers.
4526  */
4527 void
4528 vunmapbuf(struct buf *bp)
4529 {
4530 	int npages;
4531 
4532 	npages = bp->b_npages;
4533 	if (buf_mapped(bp))
4534 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4535 	vm_page_unhold_pages(bp->b_pages, npages);
4536 
4537 	bp->b_data = unmapped_buf;
4538 }
4539 
4540 void
4541 bdone(struct buf *bp)
4542 {
4543 	struct mtx *mtxp;
4544 
4545 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4546 	mtx_lock(mtxp);
4547 	bp->b_flags |= B_DONE;
4548 	wakeup(bp);
4549 	mtx_unlock(mtxp);
4550 }
4551 
4552 void
4553 bwait(struct buf *bp, u_char pri, const char *wchan)
4554 {
4555 	struct mtx *mtxp;
4556 
4557 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4558 	mtx_lock(mtxp);
4559 	while ((bp->b_flags & B_DONE) == 0)
4560 		msleep(bp, mtxp, pri, wchan, 0);
4561 	mtx_unlock(mtxp);
4562 }
4563 
4564 int
4565 bufsync(struct bufobj *bo, int waitfor)
4566 {
4567 
4568 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4569 }
4570 
4571 void
4572 bufstrategy(struct bufobj *bo, struct buf *bp)
4573 {
4574 	int i = 0;
4575 	struct vnode *vp;
4576 
4577 	vp = bp->b_vp;
4578 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4579 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4580 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4581 	i = VOP_STRATEGY(vp, bp);
4582 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4583 }
4584 
4585 void
4586 bufobj_wrefl(struct bufobj *bo)
4587 {
4588 
4589 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4590 	ASSERT_BO_WLOCKED(bo);
4591 	bo->bo_numoutput++;
4592 }
4593 
4594 void
4595 bufobj_wref(struct bufobj *bo)
4596 {
4597 
4598 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4599 	BO_LOCK(bo);
4600 	bo->bo_numoutput++;
4601 	BO_UNLOCK(bo);
4602 }
4603 
4604 void
4605 bufobj_wdrop(struct bufobj *bo)
4606 {
4607 
4608 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4609 	BO_LOCK(bo);
4610 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4611 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4612 		bo->bo_flag &= ~BO_WWAIT;
4613 		wakeup(&bo->bo_numoutput);
4614 	}
4615 	BO_UNLOCK(bo);
4616 }
4617 
4618 int
4619 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4620 {
4621 	int error;
4622 
4623 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4624 	ASSERT_BO_WLOCKED(bo);
4625 	error = 0;
4626 	while (bo->bo_numoutput) {
4627 		bo->bo_flag |= BO_WWAIT;
4628 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4629 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4630 		if (error)
4631 			break;
4632 	}
4633 	return (error);
4634 }
4635 
4636 void
4637 bpin(struct buf *bp)
4638 {
4639 	struct mtx *mtxp;
4640 
4641 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4642 	mtx_lock(mtxp);
4643 	bp->b_pin_count++;
4644 	mtx_unlock(mtxp);
4645 }
4646 
4647 void
4648 bunpin(struct buf *bp)
4649 {
4650 	struct mtx *mtxp;
4651 
4652 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4653 	mtx_lock(mtxp);
4654 	if (--bp->b_pin_count == 0)
4655 		wakeup(bp);
4656 	mtx_unlock(mtxp);
4657 }
4658 
4659 void
4660 bunpin_wait(struct buf *bp)
4661 {
4662 	struct mtx *mtxp;
4663 
4664 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4665 	mtx_lock(mtxp);
4666 	while (bp->b_pin_count > 0)
4667 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4668 	mtx_unlock(mtxp);
4669 }
4670 
4671 /*
4672  * Set bio_data or bio_ma for struct bio from the struct buf.
4673  */
4674 void
4675 bdata2bio(struct buf *bp, struct bio *bip)
4676 {
4677 
4678 	if (!buf_mapped(bp)) {
4679 		KASSERT(unmapped_buf_allowed, ("unmapped"));
4680 		bip->bio_ma = bp->b_pages;
4681 		bip->bio_ma_n = bp->b_npages;
4682 		bip->bio_data = unmapped_buf;
4683 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4684 		bip->bio_flags |= BIO_UNMAPPED;
4685 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4686 		    PAGE_SIZE == bp->b_npages,
4687 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4688 		    (long long)bip->bio_length, bip->bio_ma_n));
4689 	} else {
4690 		bip->bio_data = bp->b_data;
4691 		bip->bio_ma = NULL;
4692 	}
4693 }
4694 
4695 #include "opt_ddb.h"
4696 #ifdef DDB
4697 #include <ddb/ddb.h>
4698 
4699 /* DDB command to show buffer data */
4700 DB_SHOW_COMMAND(buffer, db_show_buffer)
4701 {
4702 	/* get args */
4703 	struct buf *bp = (struct buf *)addr;
4704 
4705 	if (!have_addr) {
4706 		db_printf("usage: show buffer <addr>\n");
4707 		return;
4708 	}
4709 
4710 	db_printf("buf at %p\n", bp);
4711 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4712 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4713 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4714 	db_printf(
4715 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4716 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4717 	    "b_dep = %p\n",
4718 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4719 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4720 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4721 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
4722 	    bp->b_kvabase, bp->b_kvasize);
4723 	if (bp->b_npages) {
4724 		int i;
4725 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4726 		for (i = 0; i < bp->b_npages; i++) {
4727 			vm_page_t m;
4728 			m = bp->b_pages[i];
4729 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4730 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4731 			if ((i + 1) < bp->b_npages)
4732 				db_printf(",");
4733 		}
4734 		db_printf("\n");
4735 	}
4736 	db_printf(" ");
4737 	BUF_LOCKPRINTINFO(bp);
4738 }
4739 
4740 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4741 {
4742 	struct buf *bp;
4743 	int i;
4744 
4745 	for (i = 0; i < nbuf; i++) {
4746 		bp = &buf[i];
4747 		if (BUF_ISLOCKED(bp)) {
4748 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4749 			db_printf("\n");
4750 		}
4751 	}
4752 }
4753 
4754 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4755 {
4756 	struct vnode *vp;
4757 	struct buf *bp;
4758 
4759 	if (!have_addr) {
4760 		db_printf("usage: show vnodebufs <addr>\n");
4761 		return;
4762 	}
4763 	vp = (struct vnode *)addr;
4764 	db_printf("Clean buffers:\n");
4765 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4766 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4767 		db_printf("\n");
4768 	}
4769 	db_printf("Dirty buffers:\n");
4770 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4771 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4772 		db_printf("\n");
4773 	}
4774 }
4775 
4776 DB_COMMAND(countfreebufs, db_coundfreebufs)
4777 {
4778 	struct buf *bp;
4779 	int i, used = 0, nfree = 0;
4780 
4781 	if (have_addr) {
4782 		db_printf("usage: countfreebufs\n");
4783 		return;
4784 	}
4785 
4786 	for (i = 0; i < nbuf; i++) {
4787 		bp = &buf[i];
4788 		if (bp->b_qindex == QUEUE_EMPTY)
4789 			nfree++;
4790 		else
4791 			used++;
4792 	}
4793 
4794 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4795 	    nfree + used);
4796 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4797 }
4798 #endif /* DDB */
4799