xref: /freebsd/sys/kern/vfs_bio.c (revision 5bd73b51076b5cb5a2c9810f76c1d7ed20c4460e)
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/resourcevar.h>
65 #include <sys/rwlock.h>
66 #include <sys/sysctl.h>
67 #include <sys/vmem.h>
68 #include <sys/vmmeter.h>
69 #include <sys/vnode.h>
70 #include <geom/geom.h>
71 #include <vm/vm.h>
72 #include <vm/vm_param.h>
73 #include <vm/vm_kern.h>
74 #include <vm/vm_pageout.h>
75 #include <vm/vm_page.h>
76 #include <vm/vm_object.h>
77 #include <vm/vm_extern.h>
78 #include <vm/vm_map.h>
79 #include "opt_compat.h"
80 #include "opt_swap.h"
81 
82 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
83 
84 struct	bio_ops bioops;		/* I/O operation notification */
85 
86 struct	buf_ops buf_ops_bio = {
87 	.bop_name	=	"buf_ops_bio",
88 	.bop_write	=	bufwrite,
89 	.bop_strategy	=	bufstrategy,
90 	.bop_sync	=	bufsync,
91 	.bop_bdflush	=	bufbdflush,
92 };
93 
94 /*
95  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
96  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
97  */
98 struct buf *buf;		/* buffer header pool */
99 caddr_t unmapped_buf;
100 
101 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
102 struct proc *bufdaemonproc;
103 
104 static int inmem(struct vnode *vp, daddr_t blkno);
105 static void vm_hold_free_pages(struct buf *bp, int newbsize);
106 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
107 		vm_offset_t to);
108 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
109 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
110 		vm_page_t m);
111 static void vfs_clean_pages_dirty_buf(struct buf *bp);
112 static void vfs_setdirty_locked_object(struct buf *bp);
113 static void vfs_vmio_release(struct buf *bp);
114 static int vfs_bio_clcheck(struct vnode *vp, int size,
115 		daddr_t lblkno, daddr_t blkno);
116 static int buf_flush(int);
117 static int flushbufqueues(int, int);
118 static void buf_daemon(void);
119 static void bremfreel(struct buf *bp);
120 static __inline void bd_wakeup(void);
121 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
122 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
123     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
124 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
125 #endif
126 
127 int vmiodirenable = TRUE;
128 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
129     "Use the VM system for directory writes");
130 long runningbufspace;
131 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
132     "Amount of presently outstanding async buffer io");
133 static long bufspace;
134 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
135     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
136 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
137     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
138 #else
139 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
140     "Virtual memory used for buffers");
141 #endif
142 static long unmapped_bufspace;
143 SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD,
144     &unmapped_bufspace, 0,
145     "Amount of unmapped buffers, inclusive in the bufspace");
146 static long maxbufspace;
147 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
148     "Maximum allowed value of bufspace (including buf_daemon)");
149 static long bufmallocspace;
150 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
151     "Amount of malloced memory for buffers");
152 static long maxbufmallocspace;
153 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
154     "Maximum amount of malloced memory for buffers");
155 static long lobufspace;
156 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
157     "Minimum amount of buffers we want to have");
158 long hibufspace;
159 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
160     "Maximum allowed value of bufspace (excluding buf_daemon)");
161 static int bufreusecnt;
162 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
163     "Number of times we have reused a buffer");
164 static int buffreekvacnt;
165 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
166     "Number of times we have freed the KVA space from some buffer");
167 static int bufdefragcnt;
168 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
169     "Number of times we have had to repeat buffer allocation to defragment");
170 static long lorunningspace;
171 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
172     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
173     "Minimum preferred space used for in-progress I/O");
174 static long hirunningspace;
175 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
176     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
177     "Maximum amount of space to use for in-progress I/O");
178 int dirtybufferflushes;
179 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
180     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
181 int bdwriteskip;
182 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
183     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
184 int altbufferflushes;
185 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
186     0, "Number of fsync flushes to limit dirty buffers");
187 static int recursiveflushes;
188 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
189     0, "Number of flushes skipped due to being recursive");
190 static int numdirtybuffers;
191 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
192     "Number of buffers that are dirty (has unwritten changes) at the moment");
193 static int lodirtybuffers;
194 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
195     "How many buffers we want to have free before bufdaemon can sleep");
196 static int hidirtybuffers;
197 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
198     "When the number of dirty buffers is considered severe");
199 int dirtybufthresh;
200 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
201     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
202 static int numfreebuffers;
203 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
204     "Number of free buffers");
205 static int lofreebuffers;
206 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
207    "XXX Unused");
208 static int hifreebuffers;
209 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
210    "XXX Complicatedly unused");
211 static int getnewbufcalls;
212 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
213    "Number of calls to getnewbuf");
214 static int getnewbufrestarts;
215 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
216     "Number of times getnewbuf has had to restart a buffer aquisition");
217 static int mappingrestarts;
218 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
219     "Number of times getblk has had to restart a buffer mapping for "
220     "unmapped buffer");
221 static int flushbufqtarget = 100;
222 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
223     "Amount of work to do in flushbufqueues when helping bufdaemon");
224 static long notbufdflushes;
225 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes, 0,
226     "Number of dirty buffer flushes done by the bufdaemon helpers");
227 static long barrierwrites;
228 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
229     "Number of barrier writes");
230 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
231     &unmapped_buf_allowed, 0,
232     "Permit the use of the unmapped i/o");
233 
234 /*
235  * Lock for the non-dirty bufqueues
236  */
237 static struct mtx_padalign bqclean;
238 
239 /*
240  * Lock for the dirty queue.
241  */
242 static struct mtx_padalign bqdirty;
243 
244 /*
245  * This lock synchronizes access to bd_request.
246  */
247 static struct mtx_padalign bdlock;
248 
249 /*
250  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
251  * waitrunningbufspace().
252  */
253 static struct mtx_padalign rbreqlock;
254 
255 /*
256  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
257  */
258 static struct rwlock_padalign nblock;
259 
260 /*
261  * Lock that protects bdirtywait.
262  */
263 static struct mtx_padalign bdirtylock;
264 
265 /*
266  * Wakeup point for bufdaemon, as well as indicator of whether it is already
267  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
268  * is idling.
269  */
270 static int bd_request;
271 
272 /*
273  * Request for the buf daemon to write more buffers than is indicated by
274  * lodirtybuf.  This may be necessary to push out excess dependencies or
275  * defragment the address space where a simple count of the number of dirty
276  * buffers is insufficient to characterize the demand for flushing them.
277  */
278 static int bd_speedupreq;
279 
280 /*
281  * bogus page -- for I/O to/from partially complete buffers
282  * this is a temporary solution to the problem, but it is not
283  * really that bad.  it would be better to split the buffer
284  * for input in the case of buffers partially already in memory,
285  * but the code is intricate enough already.
286  */
287 vm_page_t bogus_page;
288 
289 /*
290  * Synchronization (sleep/wakeup) variable for active buffer space requests.
291  * Set when wait starts, cleared prior to wakeup().
292  * Used in runningbufwakeup() and waitrunningbufspace().
293  */
294 static int runningbufreq;
295 
296 /*
297  * Synchronization (sleep/wakeup) variable for buffer requests.
298  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
299  * by and/or.
300  * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(),
301  * getnewbuf(), and getblk().
302  */
303 static volatile int needsbuffer;
304 
305 /*
306  * Synchronization for bwillwrite() waiters.
307  */
308 static int bdirtywait;
309 
310 /*
311  * Definitions for the buffer free lists.
312  */
313 #define BUFFER_QUEUES	5	/* number of free buffer queues */
314 
315 #define QUEUE_NONE	0	/* on no queue */
316 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
317 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
318 #define QUEUE_EMPTYKVA	3	/* empty buffer headers w/KVA assignment */
319 #define QUEUE_EMPTY	4	/* empty buffer headers */
320 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
321 
322 /* Queues for free buffers with various properties */
323 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
324 #ifdef INVARIANTS
325 static int bq_len[BUFFER_QUEUES];
326 #endif
327 
328 /*
329  * Single global constant for BUF_WMESG, to avoid getting multiple references.
330  * buf_wmesg is referred from macros.
331  */
332 const char *buf_wmesg = BUF_WMESG;
333 
334 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
335 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
336 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
337 
338 static int
339 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
340 {
341 	long value;
342 	int error;
343 
344 	value = *(long *)arg1;
345 	error = sysctl_handle_long(oidp, &value, 0, req);
346 	if (error != 0 || req->newptr == NULL)
347 		return (error);
348 	mtx_lock(&rbreqlock);
349 	if (arg1 == &hirunningspace) {
350 		if (value < lorunningspace)
351 			error = EINVAL;
352 		else
353 			hirunningspace = value;
354 	} else {
355 		KASSERT(arg1 == &lorunningspace,
356 		    ("%s: unknown arg1", __func__));
357 		if (value > hirunningspace)
358 			error = EINVAL;
359 		else
360 			lorunningspace = value;
361 	}
362 	mtx_unlock(&rbreqlock);
363 	return (error);
364 }
365 
366 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
367     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
368 static int
369 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
370 {
371 	long lvalue;
372 	int ivalue;
373 
374 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
375 		return (sysctl_handle_long(oidp, arg1, arg2, req));
376 	lvalue = *(long *)arg1;
377 	if (lvalue > INT_MAX)
378 		/* On overflow, still write out a long to trigger ENOMEM. */
379 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
380 	ivalue = lvalue;
381 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
382 }
383 #endif
384 
385 /*
386  *	bqlock:
387  *
388  *	Return the appropriate queue lock based on the index.
389  */
390 static inline struct mtx *
391 bqlock(int qindex)
392 {
393 
394 	if (qindex == QUEUE_DIRTY)
395 		return (struct mtx *)(&bqdirty);
396 	return (struct mtx *)(&bqclean);
397 }
398 
399 /*
400  *	bdirtywakeup:
401  *
402  *	Wakeup any bwillwrite() waiters.
403  */
404 static void
405 bdirtywakeup(void)
406 {
407 	mtx_lock(&bdirtylock);
408 	if (bdirtywait) {
409 		bdirtywait = 0;
410 		wakeup(&bdirtywait);
411 	}
412 	mtx_unlock(&bdirtylock);
413 }
414 
415 /*
416  *	bdirtysub:
417  *
418  *	Decrement the numdirtybuffers count by one and wakeup any
419  *	threads blocked in bwillwrite().
420  */
421 static void
422 bdirtysub(void)
423 {
424 
425 	if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
426 	    (lodirtybuffers + hidirtybuffers) / 2)
427 		bdirtywakeup();
428 }
429 
430 /*
431  *	bdirtyadd:
432  *
433  *	Increment the numdirtybuffers count by one and wakeup the buf
434  *	daemon if needed.
435  */
436 static void
437 bdirtyadd(void)
438 {
439 
440 	/*
441 	 * Only do the wakeup once as we cross the boundary.  The
442 	 * buf daemon will keep running until the condition clears.
443 	 */
444 	if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
445 	    (lodirtybuffers + hidirtybuffers) / 2)
446 		bd_wakeup();
447 }
448 
449 /*
450  *	bufspacewakeup:
451  *
452  *	Called when buffer space is potentially available for recovery.
453  *	getnewbuf() will block on this flag when it is unable to free
454  *	sufficient buffer space.  Buffer space becomes recoverable when
455  *	bp's get placed back in the queues.
456  */
457 
458 static __inline void
459 bufspacewakeup(void)
460 {
461 	int need_wakeup, on;
462 
463 	/*
464 	 * If someone is waiting for BUF space, wake them up.  Even
465 	 * though we haven't freed the kva space yet, the waiting
466 	 * process will be able to now.
467 	 */
468 	rw_rlock(&nblock);
469 	for (;;) {
470 		need_wakeup = 0;
471 		on = needsbuffer;
472 		if ((on & VFS_BIO_NEED_BUFSPACE) == 0)
473 			break;
474 		need_wakeup = 1;
475 		if (atomic_cmpset_rel_int(&needsbuffer, on,
476 		    on & ~VFS_BIO_NEED_BUFSPACE))
477 			break;
478 	}
479 	if (need_wakeup)
480 		wakeup(__DEVOLATILE(void *, &needsbuffer));
481 	rw_runlock(&nblock);
482 }
483 
484 /*
485  *	runningwakeup:
486  *
487  *	Wake up processes that are waiting on asynchronous writes to fall
488  *	below lorunningspace.
489  */
490 static void
491 runningwakeup(void)
492 {
493 
494 	mtx_lock(&rbreqlock);
495 	if (runningbufreq) {
496 		runningbufreq = 0;
497 		wakeup(&runningbufreq);
498 	}
499 	mtx_unlock(&rbreqlock);
500 }
501 
502 /*
503  *	runningbufwakeup:
504  *
505  *	Decrement the outstanding write count according.
506  */
507 void
508 runningbufwakeup(struct buf *bp)
509 {
510 	long space, bspace;
511 
512 	bspace = bp->b_runningbufspace;
513 	if (bspace == 0)
514 		return;
515 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
516 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
517 	    space, bspace));
518 	bp->b_runningbufspace = 0;
519 	/*
520 	 * Only acquire the lock and wakeup on the transition from exceeding
521 	 * the threshold to falling below it.
522 	 */
523 	if (space < lorunningspace)
524 		return;
525 	if (space - bspace > lorunningspace)
526 		return;
527 	runningwakeup();
528 }
529 
530 /*
531  *	bufcountadd:
532  *
533  *	Called when a buffer has been added to one of the free queues to
534  *	account for the buffer and to wakeup anyone waiting for free buffers.
535  *	This typically occurs when large amounts of metadata are being handled
536  *	by the buffer cache ( else buffer space runs out first, usually ).
537  */
538 static __inline void
539 bufcountadd(struct buf *bp)
540 {
541 	int mask, need_wakeup, old, on;
542 
543 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
544 	    ("buf %p already counted as free", bp));
545 	bp->b_flags |= B_INFREECNT;
546 	old = atomic_fetchadd_int(&numfreebuffers, 1);
547 	KASSERT(old >= 0 && old < nbuf,
548 	    ("numfreebuffers climbed to %d", old + 1));
549 	mask = VFS_BIO_NEED_ANY;
550 	if (numfreebuffers >= hifreebuffers)
551 		mask |= VFS_BIO_NEED_FREE;
552 	rw_rlock(&nblock);
553 	for (;;) {
554 		need_wakeup = 0;
555 		on = needsbuffer;
556 		if (on == 0)
557 			break;
558 		need_wakeup = 1;
559 		if (atomic_cmpset_rel_int(&needsbuffer, on, on & ~mask))
560 			break;
561 	}
562 	if (need_wakeup)
563 		wakeup(__DEVOLATILE(void *, &needsbuffer));
564 	rw_runlock(&nblock);
565 }
566 
567 /*
568  *	bufcountsub:
569  *
570  *	Decrement the numfreebuffers count as needed.
571  */
572 static void
573 bufcountsub(struct buf *bp)
574 {
575 	int old;
576 
577 	/*
578 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
579 	 * delayed-write, the buffer was free and we must decrement
580 	 * numfreebuffers.
581 	 */
582 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
583 		KASSERT((bp->b_flags & B_INFREECNT) != 0,
584 		    ("buf %p not counted in numfreebuffers", bp));
585 		bp->b_flags &= ~B_INFREECNT;
586 		old = atomic_fetchadd_int(&numfreebuffers, -1);
587 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
588 	}
589 }
590 
591 /*
592  *	waitrunningbufspace()
593  *
594  *	runningbufspace is a measure of the amount of I/O currently
595  *	running.  This routine is used in async-write situations to
596  *	prevent creating huge backups of pending writes to a device.
597  *	Only asynchronous writes are governed by this function.
598  *
599  *	This does NOT turn an async write into a sync write.  It waits
600  *	for earlier writes to complete and generally returns before the
601  *	caller's write has reached the device.
602  */
603 void
604 waitrunningbufspace(void)
605 {
606 
607 	mtx_lock(&rbreqlock);
608 	while (runningbufspace > hirunningspace) {
609 		runningbufreq = 1;
610 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
611 	}
612 	mtx_unlock(&rbreqlock);
613 }
614 
615 
616 /*
617  *	vfs_buf_test_cache:
618  *
619  *	Called when a buffer is extended.  This function clears the B_CACHE
620  *	bit if the newly extended portion of the buffer does not contain
621  *	valid data.
622  */
623 static __inline
624 void
625 vfs_buf_test_cache(struct buf *bp,
626 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
627 		  vm_page_t m)
628 {
629 
630 	VM_OBJECT_ASSERT_LOCKED(m->object);
631 	if (bp->b_flags & B_CACHE) {
632 		int base = (foff + off) & PAGE_MASK;
633 		if (vm_page_is_valid(m, base, size) == 0)
634 			bp->b_flags &= ~B_CACHE;
635 	}
636 }
637 
638 /* Wake up the buffer daemon if necessary */
639 static __inline void
640 bd_wakeup(void)
641 {
642 
643 	mtx_lock(&bdlock);
644 	if (bd_request == 0) {
645 		bd_request = 1;
646 		wakeup(&bd_request);
647 	}
648 	mtx_unlock(&bdlock);
649 }
650 
651 /*
652  * bd_speedup - speedup the buffer cache flushing code
653  */
654 void
655 bd_speedup(void)
656 {
657 	int needwake;
658 
659 	mtx_lock(&bdlock);
660 	needwake = 0;
661 	if (bd_speedupreq == 0 || bd_request == 0)
662 		needwake = 1;
663 	bd_speedupreq = 1;
664 	bd_request = 1;
665 	if (needwake)
666 		wakeup(&bd_request);
667 	mtx_unlock(&bdlock);
668 }
669 
670 #ifndef NSWBUF_MIN
671 #define	NSWBUF_MIN	16
672 #endif
673 
674 #ifdef __i386__
675 #define	TRANSIENT_DENOM	5
676 #else
677 #define	TRANSIENT_DENOM 10
678 #endif
679 
680 /*
681  * Calculating buffer cache scaling values and reserve space for buffer
682  * headers.  This is called during low level kernel initialization and
683  * may be called more then once.  We CANNOT write to the memory area
684  * being reserved at this time.
685  */
686 caddr_t
687 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
688 {
689 	int tuned_nbuf;
690 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
691 
692 	/*
693 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
694 	 * PAGE_SIZE is >= 1K)
695 	 */
696 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
697 
698 	/*
699 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
700 	 * For the first 64MB of ram nominally allocate sufficient buffers to
701 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
702 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
703 	 * the buffer cache we limit the eventual kva reservation to
704 	 * maxbcache bytes.
705 	 *
706 	 * factor represents the 1/4 x ram conversion.
707 	 */
708 	if (nbuf == 0) {
709 		int factor = 4 * BKVASIZE / 1024;
710 
711 		nbuf = 50;
712 		if (physmem_est > 4096)
713 			nbuf += min((physmem_est - 4096) / factor,
714 			    65536 / factor);
715 		if (physmem_est > 65536)
716 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
717 			    32 * 1024 * 1024 / (factor * 5));
718 
719 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
720 			nbuf = maxbcache / BKVASIZE;
721 		tuned_nbuf = 1;
722 	} else
723 		tuned_nbuf = 0;
724 
725 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
726 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
727 	if (nbuf > maxbuf) {
728 		if (!tuned_nbuf)
729 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
730 			    maxbuf);
731 		nbuf = maxbuf;
732 	}
733 
734 	/*
735 	 * Ideal allocation size for the transient bio submap is 10%
736 	 * of the maximal space buffer map.  This roughly corresponds
737 	 * to the amount of the buffer mapped for typical UFS load.
738 	 *
739 	 * Clip the buffer map to reserve space for the transient
740 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
741 	 * maximum buffer map extent on the platform.
742 	 *
743 	 * The fall-back to the maxbuf in case of maxbcache unset,
744 	 * allows to not trim the buffer KVA for the architectures
745 	 * with ample KVA space.
746 	 */
747 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
748 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
749 		buf_sz = (long)nbuf * BKVASIZE;
750 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
751 		    (TRANSIENT_DENOM - 1)) {
752 			/*
753 			 * There is more KVA than memory.  Do not
754 			 * adjust buffer map size, and assign the rest
755 			 * of maxbuf to transient map.
756 			 */
757 			biotmap_sz = maxbuf_sz - buf_sz;
758 		} else {
759 			/*
760 			 * Buffer map spans all KVA we could afford on
761 			 * this platform.  Give 10% (20% on i386) of
762 			 * the buffer map to the transient bio map.
763 			 */
764 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
765 			buf_sz -= biotmap_sz;
766 		}
767 		if (biotmap_sz / INT_MAX > MAXPHYS)
768 			bio_transient_maxcnt = INT_MAX;
769 		else
770 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
771 		/*
772 		 * Artifically limit to 1024 simultaneous in-flight I/Os
773 		 * using the transient mapping.
774 		 */
775 		if (bio_transient_maxcnt > 1024)
776 			bio_transient_maxcnt = 1024;
777 		if (tuned_nbuf)
778 			nbuf = buf_sz / BKVASIZE;
779 	}
780 
781 	/*
782 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
783 	 * We have no less then 16 and no more then 256.
784 	 */
785 	nswbuf = min(nbuf / 4, 256);
786 	TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf);
787 	if (nswbuf < NSWBUF_MIN)
788 		nswbuf = NSWBUF_MIN;
789 
790 	/*
791 	 * Reserve space for the buffer cache buffers
792 	 */
793 	swbuf = (void *)v;
794 	v = (caddr_t)(swbuf + nswbuf);
795 	buf = (void *)v;
796 	v = (caddr_t)(buf + nbuf);
797 
798 	return(v);
799 }
800 
801 /* Initialize the buffer subsystem.  Called before use of any buffers. */
802 void
803 bufinit(void)
804 {
805 	struct buf *bp;
806 	int i;
807 
808 	mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF);
809 	mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF);
810 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
811 	rw_init(&nblock, "needsbuffer lock");
812 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
813 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
814 
815 	/* next, make a null set of free lists */
816 	for (i = 0; i < BUFFER_QUEUES; i++)
817 		TAILQ_INIT(&bufqueues[i]);
818 
819 	/* finally, initialize each buffer header and stick on empty q */
820 	for (i = 0; i < nbuf; i++) {
821 		bp = &buf[i];
822 		bzero(bp, sizeof *bp);
823 		bp->b_flags = B_INVAL | B_INFREECNT;
824 		bp->b_rcred = NOCRED;
825 		bp->b_wcred = NOCRED;
826 		bp->b_qindex = QUEUE_EMPTY;
827 		bp->b_xflags = 0;
828 		LIST_INIT(&bp->b_dep);
829 		BUF_LOCKINIT(bp);
830 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
831 #ifdef INVARIANTS
832 		bq_len[QUEUE_EMPTY]++;
833 #endif
834 	}
835 
836 	/*
837 	 * maxbufspace is the absolute maximum amount of buffer space we are
838 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
839 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
840 	 * used by most other processes.  The differential is required to
841 	 * ensure that buf_daemon is able to run when other processes might
842 	 * be blocked waiting for buffer space.
843 	 *
844 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
845 	 * this may result in KVM fragmentation which is not handled optimally
846 	 * by the system.
847 	 */
848 	maxbufspace = (long)nbuf * BKVASIZE;
849 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
850 	lobufspace = hibufspace - MAXBSIZE;
851 
852 	/*
853 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
854 	 * arbitrarily and may need further tuning. It corresponds to
855 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
856 	 * which fits with many RAID controllers' tagged queuing limits.
857 	 * The lower 1 MiB limit is the historical upper limit for
858 	 * hirunningspace.
859 	 */
860 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
861 	    16 * 1024 * 1024), 1024 * 1024);
862 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
863 
864 /*
865  * Limit the amount of malloc memory since it is wired permanently into
866  * the kernel space.  Even though this is accounted for in the buffer
867  * allocation, we don't want the malloced region to grow uncontrolled.
868  * The malloc scheme improves memory utilization significantly on average
869  * (small) directories.
870  */
871 	maxbufmallocspace = hibufspace / 20;
872 
873 /*
874  * Reduce the chance of a deadlock occuring by limiting the number
875  * of delayed-write dirty buffers we allow to stack up.
876  */
877 	hidirtybuffers = nbuf / 4 + 20;
878 	dirtybufthresh = hidirtybuffers * 9 / 10;
879 	numdirtybuffers = 0;
880 /*
881  * To support extreme low-memory systems, make sure hidirtybuffers cannot
882  * eat up all available buffer space.  This occurs when our minimum cannot
883  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
884  * BKVASIZE'd buffers.
885  */
886 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
887 		hidirtybuffers >>= 1;
888 	}
889 	lodirtybuffers = hidirtybuffers / 2;
890 
891 /*
892  * Try to keep the number of free buffers in the specified range,
893  * and give special processes (e.g. like buf_daemon) access to an
894  * emergency reserve.
895  */
896 	lofreebuffers = nbuf / 18 + 5;
897 	hifreebuffers = 2 * lofreebuffers;
898 	numfreebuffers = nbuf;
899 
900 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
901 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
902 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
903 }
904 
905 #ifdef INVARIANTS
906 static inline void
907 vfs_buf_check_mapped(struct buf *bp)
908 {
909 
910 	KASSERT((bp->b_flags & B_UNMAPPED) == 0,
911 	    ("mapped buf %p %x", bp, bp->b_flags));
912 	KASSERT(bp->b_kvabase != unmapped_buf,
913 	    ("mapped buf: b_kvabase was not updated %p", bp));
914 	KASSERT(bp->b_data != unmapped_buf,
915 	    ("mapped buf: b_data was not updated %p", bp));
916 }
917 
918 static inline void
919 vfs_buf_check_unmapped(struct buf *bp)
920 {
921 
922 	KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED,
923 	    ("unmapped buf %p %x", bp, bp->b_flags));
924 	KASSERT(bp->b_kvabase == unmapped_buf,
925 	    ("unmapped buf: corrupted b_kvabase %p", bp));
926 	KASSERT(bp->b_data == unmapped_buf,
927 	    ("unmapped buf: corrupted b_data %p", bp));
928 }
929 
930 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
931 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
932 #else
933 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
934 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
935 #endif
936 
937 static void
938 bpmap_qenter(struct buf *bp)
939 {
940 
941 	BUF_CHECK_MAPPED(bp);
942 
943 	/*
944 	 * bp->b_data is relative to bp->b_offset, but
945 	 * bp->b_offset may be offset into the first page.
946 	 */
947 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
948 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
949 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
950 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
951 }
952 
953 /*
954  * bfreekva() - free the kva allocation for a buffer.
955  *
956  *	Since this call frees up buffer space, we call bufspacewakeup().
957  */
958 static void
959 bfreekva(struct buf *bp)
960 {
961 
962 	if (bp->b_kvasize == 0)
963 		return;
964 
965 	atomic_add_int(&buffreekvacnt, 1);
966 	atomic_subtract_long(&bufspace, bp->b_kvasize);
967 	if ((bp->b_flags & B_UNMAPPED) == 0) {
968 		BUF_CHECK_MAPPED(bp);
969 		vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase,
970 		    bp->b_kvasize);
971 	} else {
972 		BUF_CHECK_UNMAPPED(bp);
973 		if ((bp->b_flags & B_KVAALLOC) != 0) {
974 			vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc,
975 			    bp->b_kvasize);
976 		}
977 		atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
978 		bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
979 	}
980 	bp->b_kvasize = 0;
981 	bufspacewakeup();
982 }
983 
984 /*
985  *	binsfree:
986  *
987  *	Insert the buffer into the appropriate free list.
988  */
989 static void
990 binsfree(struct buf *bp, int qindex)
991 {
992 	struct mtx *olock, *nlock;
993 
994 	BUF_ASSERT_XLOCKED(bp);
995 
996 	olock = bqlock(bp->b_qindex);
997 	nlock = bqlock(qindex);
998 	mtx_lock(olock);
999 	/* Handle delayed bremfree() processing. */
1000 	if (bp->b_flags & B_REMFREE)
1001 		bremfreel(bp);
1002 
1003 	if (bp->b_qindex != QUEUE_NONE)
1004 		panic("binsfree: free buffer onto another queue???");
1005 
1006 	bp->b_qindex = qindex;
1007 	if (olock != nlock) {
1008 		mtx_unlock(olock);
1009 		mtx_lock(nlock);
1010 	}
1011 	if (bp->b_flags & B_AGE)
1012 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1013 	else
1014 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1015 #ifdef INVARIANTS
1016 	bq_len[bp->b_qindex]++;
1017 #endif
1018 	mtx_unlock(nlock);
1019 
1020 	/*
1021 	 * Something we can maybe free or reuse.
1022 	 */
1023 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1024 		bufspacewakeup();
1025 
1026 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1027 		bufcountadd(bp);
1028 }
1029 
1030 /*
1031  *	bremfree:
1032  *
1033  *	Mark the buffer for removal from the appropriate free list.
1034  *
1035  */
1036 void
1037 bremfree(struct buf *bp)
1038 {
1039 
1040 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1041 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1042 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1043 	KASSERT(bp->b_qindex != QUEUE_NONE,
1044 	    ("bremfree: buffer %p not on a queue.", bp));
1045 	BUF_ASSERT_XLOCKED(bp);
1046 
1047 	bp->b_flags |= B_REMFREE;
1048 	bufcountsub(bp);
1049 }
1050 
1051 /*
1052  *	bremfreef:
1053  *
1054  *	Force an immediate removal from a free list.  Used only in nfs when
1055  *	it abuses the b_freelist pointer.
1056  */
1057 void
1058 bremfreef(struct buf *bp)
1059 {
1060 	struct mtx *qlock;
1061 
1062 	qlock = bqlock(bp->b_qindex);
1063 	mtx_lock(qlock);
1064 	bremfreel(bp);
1065 	mtx_unlock(qlock);
1066 }
1067 
1068 /*
1069  *	bremfreel:
1070  *
1071  *	Removes a buffer from the free list, must be called with the
1072  *	correct qlock held.
1073  */
1074 static void
1075 bremfreel(struct buf *bp)
1076 {
1077 
1078 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1079 	    bp, bp->b_vp, bp->b_flags);
1080 	KASSERT(bp->b_qindex != QUEUE_NONE,
1081 	    ("bremfreel: buffer %p not on a queue.", bp));
1082 	BUF_ASSERT_XLOCKED(bp);
1083 	mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1084 
1085 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1086 #ifdef INVARIANTS
1087 	KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1088 	    bp->b_qindex));
1089 	bq_len[bp->b_qindex]--;
1090 #endif
1091 	bp->b_qindex = QUEUE_NONE;
1092 	/*
1093 	 * If this was a delayed bremfree() we only need to remove the buffer
1094 	 * from the queue and return the stats are already done.
1095 	 */
1096 	if (bp->b_flags & B_REMFREE) {
1097 		bp->b_flags &= ~B_REMFREE;
1098 		return;
1099 	}
1100 	bufcountsub(bp);
1101 }
1102 
1103 /*
1104  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
1105  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1106  * the buffer is valid and we do not have to do anything.
1107  */
1108 void
1109 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1110     int cnt, struct ucred * cred)
1111 {
1112 	struct buf *rabp;
1113 	int i;
1114 
1115 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1116 		if (inmem(vp, *rablkno))
1117 			continue;
1118 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1119 
1120 		if ((rabp->b_flags & B_CACHE) == 0) {
1121 			if (!TD_IS_IDLETHREAD(curthread))
1122 				curthread->td_ru.ru_inblock++;
1123 			rabp->b_flags |= B_ASYNC;
1124 			rabp->b_flags &= ~B_INVAL;
1125 			rabp->b_ioflags &= ~BIO_ERROR;
1126 			rabp->b_iocmd = BIO_READ;
1127 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
1128 				rabp->b_rcred = crhold(cred);
1129 			vfs_busy_pages(rabp, 0);
1130 			BUF_KERNPROC(rabp);
1131 			rabp->b_iooffset = dbtob(rabp->b_blkno);
1132 			bstrategy(rabp);
1133 		} else {
1134 			brelse(rabp);
1135 		}
1136 	}
1137 }
1138 
1139 /*
1140  * Entry point for bread() and breadn() via #defines in sys/buf.h.
1141  *
1142  * Get a buffer with the specified data.  Look in the cache first.  We
1143  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
1144  * is set, the buffer is valid and we do not have to do anything, see
1145  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1146  */
1147 int
1148 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1149     int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1150 {
1151 	struct buf *bp;
1152 	int rv = 0, readwait = 0;
1153 
1154 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1155 	/*
1156 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1157 	 */
1158 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1159 	if (bp == NULL)
1160 		return (EBUSY);
1161 
1162 	/* if not found in cache, do some I/O */
1163 	if ((bp->b_flags & B_CACHE) == 0) {
1164 		if (!TD_IS_IDLETHREAD(curthread))
1165 			curthread->td_ru.ru_inblock++;
1166 		bp->b_iocmd = BIO_READ;
1167 		bp->b_flags &= ~B_INVAL;
1168 		bp->b_ioflags &= ~BIO_ERROR;
1169 		if (bp->b_rcred == NOCRED && cred != NOCRED)
1170 			bp->b_rcred = crhold(cred);
1171 		vfs_busy_pages(bp, 0);
1172 		bp->b_iooffset = dbtob(bp->b_blkno);
1173 		bstrategy(bp);
1174 		++readwait;
1175 	}
1176 
1177 	breada(vp, rablkno, rabsize, cnt, cred);
1178 
1179 	if (readwait) {
1180 		rv = bufwait(bp);
1181 	}
1182 	return (rv);
1183 }
1184 
1185 /*
1186  * Write, release buffer on completion.  (Done by iodone
1187  * if async).  Do not bother writing anything if the buffer
1188  * is invalid.
1189  *
1190  * Note that we set B_CACHE here, indicating that buffer is
1191  * fully valid and thus cacheable.  This is true even of NFS
1192  * now so we set it generally.  This could be set either here
1193  * or in biodone() since the I/O is synchronous.  We put it
1194  * here.
1195  */
1196 int
1197 bufwrite(struct buf *bp)
1198 {
1199 	int oldflags;
1200 	struct vnode *vp;
1201 	long space;
1202 	int vp_md;
1203 
1204 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1205 	if (bp->b_flags & B_INVAL) {
1206 		brelse(bp);
1207 		return (0);
1208 	}
1209 
1210 	if (bp->b_flags & B_BARRIER)
1211 		barrierwrites++;
1212 
1213 	oldflags = bp->b_flags;
1214 
1215 	BUF_ASSERT_HELD(bp);
1216 
1217 	if (bp->b_pin_count > 0)
1218 		bunpin_wait(bp);
1219 
1220 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1221 	    ("FFS background buffer should not get here %p", bp));
1222 
1223 	vp = bp->b_vp;
1224 	if (vp)
1225 		vp_md = vp->v_vflag & VV_MD;
1226 	else
1227 		vp_md = 0;
1228 
1229 	/*
1230 	 * Mark the buffer clean.  Increment the bufobj write count
1231 	 * before bundirty() call, to prevent other thread from seeing
1232 	 * empty dirty list and zero counter for writes in progress,
1233 	 * falsely indicating that the bufobj is clean.
1234 	 */
1235 	bufobj_wref(bp->b_bufobj);
1236 	bundirty(bp);
1237 
1238 	bp->b_flags &= ~B_DONE;
1239 	bp->b_ioflags &= ~BIO_ERROR;
1240 	bp->b_flags |= B_CACHE;
1241 	bp->b_iocmd = BIO_WRITE;
1242 
1243 	vfs_busy_pages(bp, 1);
1244 
1245 	/*
1246 	 * Normal bwrites pipeline writes
1247 	 */
1248 	bp->b_runningbufspace = bp->b_bufsize;
1249 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1250 
1251 	if (!TD_IS_IDLETHREAD(curthread))
1252 		curthread->td_ru.ru_oublock++;
1253 	if (oldflags & B_ASYNC)
1254 		BUF_KERNPROC(bp);
1255 	bp->b_iooffset = dbtob(bp->b_blkno);
1256 	bstrategy(bp);
1257 
1258 	if ((oldflags & B_ASYNC) == 0) {
1259 		int rtval = bufwait(bp);
1260 		brelse(bp);
1261 		return (rtval);
1262 	} else if (space > hirunningspace) {
1263 		/*
1264 		 * don't allow the async write to saturate the I/O
1265 		 * system.  We will not deadlock here because
1266 		 * we are blocking waiting for I/O that is already in-progress
1267 		 * to complete. We do not block here if it is the update
1268 		 * or syncer daemon trying to clean up as that can lead
1269 		 * to deadlock.
1270 		 */
1271 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1272 			waitrunningbufspace();
1273 	}
1274 
1275 	return (0);
1276 }
1277 
1278 void
1279 bufbdflush(struct bufobj *bo, struct buf *bp)
1280 {
1281 	struct buf *nbp;
1282 
1283 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1284 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1285 		altbufferflushes++;
1286 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1287 		BO_LOCK(bo);
1288 		/*
1289 		 * Try to find a buffer to flush.
1290 		 */
1291 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1292 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1293 			    BUF_LOCK(nbp,
1294 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
1295 				continue;
1296 			if (bp == nbp)
1297 				panic("bdwrite: found ourselves");
1298 			BO_UNLOCK(bo);
1299 			/* Don't countdeps with the bo lock held. */
1300 			if (buf_countdeps(nbp, 0)) {
1301 				BO_LOCK(bo);
1302 				BUF_UNLOCK(nbp);
1303 				continue;
1304 			}
1305 			if (nbp->b_flags & B_CLUSTEROK) {
1306 				vfs_bio_awrite(nbp);
1307 			} else {
1308 				bremfree(nbp);
1309 				bawrite(nbp);
1310 			}
1311 			dirtybufferflushes++;
1312 			break;
1313 		}
1314 		if (nbp == NULL)
1315 			BO_UNLOCK(bo);
1316 	}
1317 }
1318 
1319 /*
1320  * Delayed write. (Buffer is marked dirty).  Do not bother writing
1321  * anything if the buffer is marked invalid.
1322  *
1323  * Note that since the buffer must be completely valid, we can safely
1324  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
1325  * biodone() in order to prevent getblk from writing the buffer
1326  * out synchronously.
1327  */
1328 void
1329 bdwrite(struct buf *bp)
1330 {
1331 	struct thread *td = curthread;
1332 	struct vnode *vp;
1333 	struct bufobj *bo;
1334 
1335 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1336 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1337 	KASSERT((bp->b_flags & B_BARRIER) == 0,
1338 	    ("Barrier request in delayed write %p", bp));
1339 	BUF_ASSERT_HELD(bp);
1340 
1341 	if (bp->b_flags & B_INVAL) {
1342 		brelse(bp);
1343 		return;
1344 	}
1345 
1346 	/*
1347 	 * If we have too many dirty buffers, don't create any more.
1348 	 * If we are wildly over our limit, then force a complete
1349 	 * cleanup. Otherwise, just keep the situation from getting
1350 	 * out of control. Note that we have to avoid a recursive
1351 	 * disaster and not try to clean up after our own cleanup!
1352 	 */
1353 	vp = bp->b_vp;
1354 	bo = bp->b_bufobj;
1355 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1356 		td->td_pflags |= TDP_INBDFLUSH;
1357 		BO_BDFLUSH(bo, bp);
1358 		td->td_pflags &= ~TDP_INBDFLUSH;
1359 	} else
1360 		recursiveflushes++;
1361 
1362 	bdirty(bp);
1363 	/*
1364 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1365 	 * true even of NFS now.
1366 	 */
1367 	bp->b_flags |= B_CACHE;
1368 
1369 	/*
1370 	 * This bmap keeps the system from needing to do the bmap later,
1371 	 * perhaps when the system is attempting to do a sync.  Since it
1372 	 * is likely that the indirect block -- or whatever other datastructure
1373 	 * that the filesystem needs is still in memory now, it is a good
1374 	 * thing to do this.  Note also, that if the pageout daemon is
1375 	 * requesting a sync -- there might not be enough memory to do
1376 	 * the bmap then...  So, this is important to do.
1377 	 */
1378 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1379 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1380 	}
1381 
1382 	/*
1383 	 * Set the *dirty* buffer range based upon the VM system dirty
1384 	 * pages.
1385 	 *
1386 	 * Mark the buffer pages as clean.  We need to do this here to
1387 	 * satisfy the vnode_pager and the pageout daemon, so that it
1388 	 * thinks that the pages have been "cleaned".  Note that since
1389 	 * the pages are in a delayed write buffer -- the VFS layer
1390 	 * "will" see that the pages get written out on the next sync,
1391 	 * or perhaps the cluster will be completed.
1392 	 */
1393 	vfs_clean_pages_dirty_buf(bp);
1394 	bqrelse(bp);
1395 
1396 	/*
1397 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1398 	 * due to the softdep code.
1399 	 */
1400 }
1401 
1402 /*
1403  *	bdirty:
1404  *
1405  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1406  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1407  *	itself to properly update it in the dirty/clean lists.  We mark it
1408  *	B_DONE to ensure that any asynchronization of the buffer properly
1409  *	clears B_DONE ( else a panic will occur later ).
1410  *
1411  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1412  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1413  *	should only be called if the buffer is known-good.
1414  *
1415  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1416  *	count.
1417  *
1418  *	The buffer must be on QUEUE_NONE.
1419  */
1420 void
1421 bdirty(struct buf *bp)
1422 {
1423 
1424 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1425 	    bp, bp->b_vp, bp->b_flags);
1426 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1427 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1428 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1429 	BUF_ASSERT_HELD(bp);
1430 	bp->b_flags &= ~(B_RELBUF);
1431 	bp->b_iocmd = BIO_WRITE;
1432 
1433 	if ((bp->b_flags & B_DELWRI) == 0) {
1434 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1435 		reassignbuf(bp);
1436 		bdirtyadd();
1437 	}
1438 }
1439 
1440 /*
1441  *	bundirty:
1442  *
1443  *	Clear B_DELWRI for buffer.
1444  *
1445  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1446  *	count.
1447  *
1448  *	The buffer must be on QUEUE_NONE.
1449  */
1450 
1451 void
1452 bundirty(struct buf *bp)
1453 {
1454 
1455 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1456 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1457 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1458 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1459 	BUF_ASSERT_HELD(bp);
1460 
1461 	if (bp->b_flags & B_DELWRI) {
1462 		bp->b_flags &= ~B_DELWRI;
1463 		reassignbuf(bp);
1464 		bdirtysub();
1465 	}
1466 	/*
1467 	 * Since it is now being written, we can clear its deferred write flag.
1468 	 */
1469 	bp->b_flags &= ~B_DEFERRED;
1470 }
1471 
1472 /*
1473  *	bawrite:
1474  *
1475  *	Asynchronous write.  Start output on a buffer, but do not wait for
1476  *	it to complete.  The buffer is released when the output completes.
1477  *
1478  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1479  *	B_INVAL buffers.  Not us.
1480  */
1481 void
1482 bawrite(struct buf *bp)
1483 {
1484 
1485 	bp->b_flags |= B_ASYNC;
1486 	(void) bwrite(bp);
1487 }
1488 
1489 /*
1490  *	babarrierwrite:
1491  *
1492  *	Asynchronous barrier write.  Start output on a buffer, but do not
1493  *	wait for it to complete.  Place a write barrier after this write so
1494  *	that this buffer and all buffers written before it are committed to
1495  *	the disk before any buffers written after this write are committed
1496  *	to the disk.  The buffer is released when the output completes.
1497  */
1498 void
1499 babarrierwrite(struct buf *bp)
1500 {
1501 
1502 	bp->b_flags |= B_ASYNC | B_BARRIER;
1503 	(void) bwrite(bp);
1504 }
1505 
1506 /*
1507  *	bbarrierwrite:
1508  *
1509  *	Synchronous barrier write.  Start output on a buffer and wait for
1510  *	it to complete.  Place a write barrier after this write so that
1511  *	this buffer and all buffers written before it are committed to
1512  *	the disk before any buffers written after this write are committed
1513  *	to the disk.  The buffer is released when the output completes.
1514  */
1515 int
1516 bbarrierwrite(struct buf *bp)
1517 {
1518 
1519 	bp->b_flags |= B_BARRIER;
1520 	return (bwrite(bp));
1521 }
1522 
1523 /*
1524  *	bwillwrite:
1525  *
1526  *	Called prior to the locking of any vnodes when we are expecting to
1527  *	write.  We do not want to starve the buffer cache with too many
1528  *	dirty buffers so we block here.  By blocking prior to the locking
1529  *	of any vnodes we attempt to avoid the situation where a locked vnode
1530  *	prevents the various system daemons from flushing related buffers.
1531  */
1532 void
1533 bwillwrite(void)
1534 {
1535 
1536 	if (numdirtybuffers >= hidirtybuffers) {
1537 		mtx_lock(&bdirtylock);
1538 		while (numdirtybuffers >= hidirtybuffers) {
1539 			bdirtywait = 1;
1540 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
1541 			    "flswai", 0);
1542 		}
1543 		mtx_unlock(&bdirtylock);
1544 	}
1545 }
1546 
1547 /*
1548  * Return true if we have too many dirty buffers.
1549  */
1550 int
1551 buf_dirty_count_severe(void)
1552 {
1553 
1554 	return(numdirtybuffers >= hidirtybuffers);
1555 }
1556 
1557 static __noinline int
1558 buf_vm_page_count_severe(void)
1559 {
1560 
1561 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1562 
1563 	return vm_page_count_severe();
1564 }
1565 
1566 /*
1567  *	brelse:
1568  *
1569  *	Release a busy buffer and, if requested, free its resources.  The
1570  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1571  *	to be accessed later as a cache entity or reused for other purposes.
1572  */
1573 void
1574 brelse(struct buf *bp)
1575 {
1576 	int qindex;
1577 
1578 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1579 	    bp, bp->b_vp, bp->b_flags);
1580 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1581 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1582 
1583 	if (BUF_LOCKRECURSED(bp)) {
1584 		/*
1585 		 * Do not process, in particular, do not handle the
1586 		 * B_INVAL/B_RELBUF and do not release to free list.
1587 		 */
1588 		BUF_UNLOCK(bp);
1589 		return;
1590 	}
1591 
1592 	if (bp->b_flags & B_MANAGED) {
1593 		bqrelse(bp);
1594 		return;
1595 	}
1596 
1597 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1598 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1599 		/*
1600 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1601 		 * pages from being scrapped.  If the error is anything
1602 		 * other than an I/O error (EIO), assume that retrying
1603 		 * is futile.
1604 		 */
1605 		bp->b_ioflags &= ~BIO_ERROR;
1606 		bdirty(bp);
1607 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1608 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1609 		/*
1610 		 * Either a failed I/O or we were asked to free or not
1611 		 * cache the buffer.
1612 		 */
1613 		bp->b_flags |= B_INVAL;
1614 		if (!LIST_EMPTY(&bp->b_dep))
1615 			buf_deallocate(bp);
1616 		if (bp->b_flags & B_DELWRI)
1617 			bdirtysub();
1618 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1619 		if ((bp->b_flags & B_VMIO) == 0) {
1620 			if (bp->b_bufsize)
1621 				allocbuf(bp, 0);
1622 			if (bp->b_vp)
1623 				brelvp(bp);
1624 		}
1625 	}
1626 
1627 	/*
1628 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1629 	 * is called with B_DELWRI set, the underlying pages may wind up
1630 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1631 	 * because pages associated with a B_DELWRI bp are marked clean.
1632 	 *
1633 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1634 	 * if B_DELWRI is set.
1635 	 *
1636 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1637 	 * on pages to return pages to the VM page queues.
1638 	 */
1639 	if (bp->b_flags & B_DELWRI)
1640 		bp->b_flags &= ~B_RELBUF;
1641 	else if (buf_vm_page_count_severe()) {
1642 		/*
1643 		 * BKGRDINPROG can only be set with the buf and bufobj
1644 		 * locks both held.  We tolerate a race to clear it here.
1645 		 */
1646 		if (!(bp->b_vflags & BV_BKGRDINPROG))
1647 			bp->b_flags |= B_RELBUF;
1648 	}
1649 
1650 	/*
1651 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1652 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1653 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1654 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1655 	 *
1656 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1657 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1658 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1659 	 *
1660 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1661 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1662 	 * the commit state and we cannot afford to lose the buffer. If the
1663 	 * buffer has a background write in progress, we need to keep it
1664 	 * around to prevent it from being reconstituted and starting a second
1665 	 * background write.
1666 	 */
1667 	if ((bp->b_flags & B_VMIO)
1668 	    && !(bp->b_vp->v_mount != NULL &&
1669 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1670 		 !vn_isdisk(bp->b_vp, NULL) &&
1671 		 (bp->b_flags & B_DELWRI))
1672 	    ) {
1673 
1674 		int i, j, resid;
1675 		vm_page_t m;
1676 		off_t foff;
1677 		vm_pindex_t poff;
1678 		vm_object_t obj;
1679 
1680 		obj = bp->b_bufobj->bo_object;
1681 
1682 		/*
1683 		 * Get the base offset and length of the buffer.  Note that
1684 		 * in the VMIO case if the buffer block size is not
1685 		 * page-aligned then b_data pointer may not be page-aligned.
1686 		 * But our b_pages[] array *IS* page aligned.
1687 		 *
1688 		 * block sizes less then DEV_BSIZE (usually 512) are not
1689 		 * supported due to the page granularity bits (m->valid,
1690 		 * m->dirty, etc...).
1691 		 *
1692 		 * See man buf(9) for more information
1693 		 */
1694 		resid = bp->b_bufsize;
1695 		foff = bp->b_offset;
1696 		for (i = 0; i < bp->b_npages; i++) {
1697 			int had_bogus = 0;
1698 
1699 			m = bp->b_pages[i];
1700 
1701 			/*
1702 			 * If we hit a bogus page, fixup *all* the bogus pages
1703 			 * now.
1704 			 */
1705 			if (m == bogus_page) {
1706 				poff = OFF_TO_IDX(bp->b_offset);
1707 				had_bogus = 1;
1708 
1709 				VM_OBJECT_RLOCK(obj);
1710 				for (j = i; j < bp->b_npages; j++) {
1711 					vm_page_t mtmp;
1712 					mtmp = bp->b_pages[j];
1713 					if (mtmp == bogus_page) {
1714 						mtmp = vm_page_lookup(obj, poff + j);
1715 						if (!mtmp) {
1716 							panic("brelse: page missing\n");
1717 						}
1718 						bp->b_pages[j] = mtmp;
1719 					}
1720 				}
1721 				VM_OBJECT_RUNLOCK(obj);
1722 
1723 				if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) {
1724 					BUF_CHECK_MAPPED(bp);
1725 					pmap_qenter(
1726 					    trunc_page((vm_offset_t)bp->b_data),
1727 					    bp->b_pages, bp->b_npages);
1728 				}
1729 				m = bp->b_pages[i];
1730 			}
1731 			if ((bp->b_flags & B_NOCACHE) ||
1732 			    (bp->b_ioflags & BIO_ERROR &&
1733 			     bp->b_iocmd == BIO_READ)) {
1734 				int poffset = foff & PAGE_MASK;
1735 				int presid = resid > (PAGE_SIZE - poffset) ?
1736 					(PAGE_SIZE - poffset) : resid;
1737 
1738 				KASSERT(presid >= 0, ("brelse: extra page"));
1739 				VM_OBJECT_WLOCK(obj);
1740 				while (vm_page_xbusied(m)) {
1741 					vm_page_lock(m);
1742 					VM_OBJECT_WUNLOCK(obj);
1743 					vm_page_busy_sleep(m, "mbncsh");
1744 					VM_OBJECT_WLOCK(obj);
1745 				}
1746 				if (pmap_page_wired_mappings(m) == 0)
1747 					vm_page_set_invalid(m, poffset, presid);
1748 				VM_OBJECT_WUNLOCK(obj);
1749 				if (had_bogus)
1750 					printf("avoided corruption bug in bogus_page/brelse code\n");
1751 			}
1752 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1753 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1754 		}
1755 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1756 			vfs_vmio_release(bp);
1757 
1758 	} else if (bp->b_flags & B_VMIO) {
1759 
1760 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1761 			vfs_vmio_release(bp);
1762 		}
1763 
1764 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1765 		if (bp->b_bufsize != 0)
1766 			allocbuf(bp, 0);
1767 		if (bp->b_vp != NULL)
1768 			brelvp(bp);
1769 	}
1770 
1771 	/*
1772 	 * If the buffer has junk contents signal it and eventually
1773 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1774 	 * doesn't find it.
1775 	 */
1776 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1777 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1778 		bp->b_flags |= B_INVAL;
1779 	if (bp->b_flags & B_INVAL) {
1780 		if (bp->b_flags & B_DELWRI)
1781 			bundirty(bp);
1782 		if (bp->b_vp)
1783 			brelvp(bp);
1784 	}
1785 
1786 	/* buffers with no memory */
1787 	if (bp->b_bufsize == 0) {
1788 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1789 		if (bp->b_vflags & BV_BKGRDINPROG)
1790 			panic("losing buffer 1");
1791 		if (bp->b_kvasize)
1792 			qindex = QUEUE_EMPTYKVA;
1793 		else
1794 			qindex = QUEUE_EMPTY;
1795 		bp->b_flags |= B_AGE;
1796 	/* buffers with junk contents */
1797 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1798 	    (bp->b_ioflags & BIO_ERROR)) {
1799 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1800 		if (bp->b_vflags & BV_BKGRDINPROG)
1801 			panic("losing buffer 2");
1802 		qindex = QUEUE_CLEAN;
1803 		bp->b_flags |= B_AGE;
1804 	/* remaining buffers */
1805 	} else if (bp->b_flags & B_DELWRI)
1806 		qindex = QUEUE_DIRTY;
1807 	else
1808 		qindex = QUEUE_CLEAN;
1809 
1810 	binsfree(bp, qindex);
1811 
1812 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1813 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1814 		panic("brelse: not dirty");
1815 	/* unlock */
1816 	BUF_UNLOCK(bp);
1817 }
1818 
1819 /*
1820  * Release a buffer back to the appropriate queue but do not try to free
1821  * it.  The buffer is expected to be used again soon.
1822  *
1823  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1824  * biodone() to requeue an async I/O on completion.  It is also used when
1825  * known good buffers need to be requeued but we think we may need the data
1826  * again soon.
1827  *
1828  * XXX we should be able to leave the B_RELBUF hint set on completion.
1829  */
1830 void
1831 bqrelse(struct buf *bp)
1832 {
1833 	int qindex;
1834 
1835 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1836 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1837 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1838 
1839 	if (BUF_LOCKRECURSED(bp)) {
1840 		/* do not release to free list */
1841 		BUF_UNLOCK(bp);
1842 		return;
1843 	}
1844 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1845 
1846 	if (bp->b_flags & B_MANAGED) {
1847 		if (bp->b_flags & B_REMFREE)
1848 			bremfreef(bp);
1849 		goto out;
1850 	}
1851 
1852 	/* buffers with stale but valid contents */
1853 	if (bp->b_flags & B_DELWRI) {
1854 		qindex = QUEUE_DIRTY;
1855 	} else {
1856 		if ((bp->b_flags & B_DELWRI) == 0 &&
1857 		    (bp->b_xflags & BX_VNDIRTY))
1858 			panic("bqrelse: not dirty");
1859 		/*
1860 		 * BKGRDINPROG can only be set with the buf and bufobj
1861 		 * locks both held.  We tolerate a race to clear it here.
1862 		 */
1863 		if (buf_vm_page_count_severe() &&
1864 		    (bp->b_vflags & BV_BKGRDINPROG) == 0) {
1865 			/*
1866 			 * We are too low on memory, we have to try to free
1867 			 * the buffer (most importantly: the wired pages
1868 			 * making up its backing store) *now*.
1869 			 */
1870 			brelse(bp);
1871 			return;
1872 		}
1873 		qindex = QUEUE_CLEAN;
1874 	}
1875 	binsfree(bp, qindex);
1876 
1877 out:
1878 	/* unlock */
1879 	BUF_UNLOCK(bp);
1880 }
1881 
1882 /* Give pages used by the bp back to the VM system (where possible) */
1883 static void
1884 vfs_vmio_release(struct buf *bp)
1885 {
1886 	int i;
1887 	vm_page_t m;
1888 
1889 	if ((bp->b_flags & B_UNMAPPED) == 0) {
1890 		BUF_CHECK_MAPPED(bp);
1891 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1892 	} else
1893 		BUF_CHECK_UNMAPPED(bp);
1894 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
1895 	for (i = 0; i < bp->b_npages; i++) {
1896 		m = bp->b_pages[i];
1897 		bp->b_pages[i] = NULL;
1898 		/*
1899 		 * In order to keep page LRU ordering consistent, put
1900 		 * everything on the inactive queue.
1901 		 */
1902 		vm_page_lock(m);
1903 		vm_page_unwire(m, PQ_INACTIVE);
1904 
1905 		/*
1906 		 * Might as well free the page if we can and it has
1907 		 * no valid data.  We also free the page if the
1908 		 * buffer was used for direct I/O
1909 		 */
1910 		if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1911 			if (m->wire_count == 0 && !vm_page_busied(m))
1912 				vm_page_free(m);
1913 		} else if (bp->b_flags & B_DIRECT)
1914 			vm_page_try_to_free(m);
1915 		else if (buf_vm_page_count_severe())
1916 			vm_page_try_to_cache(m);
1917 		vm_page_unlock(m);
1918 	}
1919 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
1920 
1921 	if (bp->b_bufsize) {
1922 		bufspacewakeup();
1923 		bp->b_bufsize = 0;
1924 	}
1925 	bp->b_npages = 0;
1926 	bp->b_flags &= ~B_VMIO;
1927 	if (bp->b_vp)
1928 		brelvp(bp);
1929 }
1930 
1931 /*
1932  * Check to see if a block at a particular lbn is available for a clustered
1933  * write.
1934  */
1935 static int
1936 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1937 {
1938 	struct buf *bpa;
1939 	int match;
1940 
1941 	match = 0;
1942 
1943 	/* If the buf isn't in core skip it */
1944 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1945 		return (0);
1946 
1947 	/* If the buf is busy we don't want to wait for it */
1948 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1949 		return (0);
1950 
1951 	/* Only cluster with valid clusterable delayed write buffers */
1952 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1953 	    (B_DELWRI | B_CLUSTEROK))
1954 		goto done;
1955 
1956 	if (bpa->b_bufsize != size)
1957 		goto done;
1958 
1959 	/*
1960 	 * Check to see if it is in the expected place on disk and that the
1961 	 * block has been mapped.
1962 	 */
1963 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1964 		match = 1;
1965 done:
1966 	BUF_UNLOCK(bpa);
1967 	return (match);
1968 }
1969 
1970 /*
1971  *	vfs_bio_awrite:
1972  *
1973  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1974  *	This is much better then the old way of writing only one buffer at
1975  *	a time.  Note that we may not be presented with the buffers in the
1976  *	correct order, so we search for the cluster in both directions.
1977  */
1978 int
1979 vfs_bio_awrite(struct buf *bp)
1980 {
1981 	struct bufobj *bo;
1982 	int i;
1983 	int j;
1984 	daddr_t lblkno = bp->b_lblkno;
1985 	struct vnode *vp = bp->b_vp;
1986 	int ncl;
1987 	int nwritten;
1988 	int size;
1989 	int maxcl;
1990 	int gbflags;
1991 
1992 	bo = &vp->v_bufobj;
1993 	gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
1994 	/*
1995 	 * right now we support clustered writing only to regular files.  If
1996 	 * we find a clusterable block we could be in the middle of a cluster
1997 	 * rather then at the beginning.
1998 	 */
1999 	if ((vp->v_type == VREG) &&
2000 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
2001 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
2002 
2003 		size = vp->v_mount->mnt_stat.f_iosize;
2004 		maxcl = MAXPHYS / size;
2005 
2006 		BO_RLOCK(bo);
2007 		for (i = 1; i < maxcl; i++)
2008 			if (vfs_bio_clcheck(vp, size, lblkno + i,
2009 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
2010 				break;
2011 
2012 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
2013 			if (vfs_bio_clcheck(vp, size, lblkno - j,
2014 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
2015 				break;
2016 		BO_RUNLOCK(bo);
2017 		--j;
2018 		ncl = i + j;
2019 		/*
2020 		 * this is a possible cluster write
2021 		 */
2022 		if (ncl != 1) {
2023 			BUF_UNLOCK(bp);
2024 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
2025 			    gbflags);
2026 			return (nwritten);
2027 		}
2028 	}
2029 	bremfree(bp);
2030 	bp->b_flags |= B_ASYNC;
2031 	/*
2032 	 * default (old) behavior, writing out only one block
2033 	 *
2034 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
2035 	 */
2036 	nwritten = bp->b_bufsize;
2037 	(void) bwrite(bp);
2038 
2039 	return (nwritten);
2040 }
2041 
2042 static void
2043 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags)
2044 {
2045 
2046 	KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2047 	    bp->b_kvasize == 0, ("call bfreekva(%p)", bp));
2048 	if ((gbflags & GB_UNMAPPED) == 0) {
2049 		bp->b_kvabase = (caddr_t)addr;
2050 	} else if ((gbflags & GB_KVAALLOC) != 0) {
2051 		KASSERT((gbflags & GB_UNMAPPED) != 0,
2052 		    ("GB_KVAALLOC without GB_UNMAPPED"));
2053 		bp->b_kvaalloc = (caddr_t)addr;
2054 		bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2055 		atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2056 	}
2057 	bp->b_kvasize = maxsize;
2058 }
2059 
2060 /*
2061  * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if
2062  * needed.
2063  */
2064 static int
2065 allocbufkva(struct buf *bp, int maxsize, int gbflags)
2066 {
2067 	vm_offset_t addr;
2068 
2069 	bfreekva(bp);
2070 	addr = 0;
2071 
2072 	if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) {
2073 		/*
2074 		 * Buffer map is too fragmented.  Request the caller
2075 		 * to defragment the map.
2076 		 */
2077 		atomic_add_int(&bufdefragcnt, 1);
2078 		return (1);
2079 	}
2080 	setbufkva(bp, addr, maxsize, gbflags);
2081 	atomic_add_long(&bufspace, bp->b_kvasize);
2082 	return (0);
2083 }
2084 
2085 /*
2086  * Ask the bufdaemon for help, or act as bufdaemon itself, when a
2087  * locked vnode is supplied.
2088  */
2089 static void
2090 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo,
2091     int defrag)
2092 {
2093 	struct thread *td;
2094 	char *waitmsg;
2095 	int cnt, error, flags, norunbuf, wait;
2096 
2097 	mtx_assert(&bqclean, MA_OWNED);
2098 
2099 	if (defrag) {
2100 		flags = VFS_BIO_NEED_BUFSPACE;
2101 		waitmsg = "nbufkv";
2102 	} else if (bufspace >= hibufspace) {
2103 		waitmsg = "nbufbs";
2104 		flags = VFS_BIO_NEED_BUFSPACE;
2105 	} else {
2106 		waitmsg = "newbuf";
2107 		flags = VFS_BIO_NEED_ANY;
2108 	}
2109 	atomic_set_int(&needsbuffer, flags);
2110 	mtx_unlock(&bqclean);
2111 
2112 	bd_speedup();	/* heeeelp */
2113 	if ((gbflags & GB_NOWAIT_BD) != 0)
2114 		return;
2115 
2116 	td = curthread;
2117 	cnt = 0;
2118 	wait = MNT_NOWAIT;
2119 	rw_wlock(&nblock);
2120 	while ((needsbuffer & flags) != 0) {
2121 		if (vp != NULL && vp->v_type != VCHR &&
2122 		    (td->td_pflags & TDP_BUFNEED) == 0) {
2123 			rw_wunlock(&nblock);
2124 			/*
2125 			 * getblk() is called with a vnode locked, and
2126 			 * some majority of the dirty buffers may as
2127 			 * well belong to the vnode.  Flushing the
2128 			 * buffers there would make a progress that
2129 			 * cannot be achieved by the buf_daemon, that
2130 			 * cannot lock the vnode.
2131 			 */
2132 			if (cnt++ > 2)
2133 				wait = MNT_WAIT;
2134 			ASSERT_VOP_LOCKED(vp, "bufd_helper");
2135 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
2136 			    vn_lock(vp, LK_TRYUPGRADE);
2137 			if (error == 0) {
2138 				/* play bufdaemon */
2139 				norunbuf = curthread_pflags_set(TDP_BUFNEED |
2140 				    TDP_NORUNNINGBUF);
2141 				VOP_FSYNC(vp, wait, td);
2142 				atomic_add_long(&notbufdflushes, 1);
2143 				curthread_pflags_restore(norunbuf);
2144 			}
2145 			rw_wlock(&nblock);
2146 			if ((needsbuffer & flags) == 0)
2147 				break;
2148 		}
2149 		error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock,
2150 		    (PRIBIO + 4) | slpflag, waitmsg, slptimeo);
2151 		if (error != 0)
2152 			break;
2153 	}
2154 	rw_wunlock(&nblock);
2155 }
2156 
2157 static void
2158 getnewbuf_reuse_bp(struct buf *bp, int qindex)
2159 {
2160 
2161 	CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
2162 	    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
2163 	     bp->b_kvasize, bp->b_bufsize, qindex);
2164 	mtx_assert(&bqclean, MA_NOTOWNED);
2165 
2166 	/*
2167 	 * Note: we no longer distinguish between VMIO and non-VMIO
2168 	 * buffers.
2169 	 */
2170 	KASSERT((bp->b_flags & B_DELWRI) == 0,
2171 	    ("delwri buffer %p found in queue %d", bp, qindex));
2172 
2173 	if (qindex == QUEUE_CLEAN) {
2174 		if (bp->b_flags & B_VMIO) {
2175 			bp->b_flags &= ~B_ASYNC;
2176 			vfs_vmio_release(bp);
2177 		}
2178 		if (bp->b_vp != NULL)
2179 			brelvp(bp);
2180 	}
2181 
2182 	/*
2183 	 * Get the rest of the buffer freed up.  b_kva* is still valid
2184 	 * after this operation.
2185 	 */
2186 
2187 	if (bp->b_rcred != NOCRED) {
2188 		crfree(bp->b_rcred);
2189 		bp->b_rcred = NOCRED;
2190 	}
2191 	if (bp->b_wcred != NOCRED) {
2192 		crfree(bp->b_wcred);
2193 		bp->b_wcred = NOCRED;
2194 	}
2195 	if (!LIST_EMPTY(&bp->b_dep))
2196 		buf_deallocate(bp);
2197 	if (bp->b_vflags & BV_BKGRDINPROG)
2198 		panic("losing buffer 3");
2199 	KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p.  qindex: %d",
2200 	    bp, bp->b_vp, qindex));
2201 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2202 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
2203 
2204 	if (bp->b_bufsize)
2205 		allocbuf(bp, 0);
2206 
2207 	bp->b_flags &= B_UNMAPPED | B_KVAALLOC;
2208 	bp->b_ioflags = 0;
2209 	bp->b_xflags = 0;
2210 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
2211 	    ("buf %p still counted as free?", bp));
2212 	bp->b_vflags = 0;
2213 	bp->b_vp = NULL;
2214 	bp->b_blkno = bp->b_lblkno = 0;
2215 	bp->b_offset = NOOFFSET;
2216 	bp->b_iodone = 0;
2217 	bp->b_error = 0;
2218 	bp->b_resid = 0;
2219 	bp->b_bcount = 0;
2220 	bp->b_npages = 0;
2221 	bp->b_dirtyoff = bp->b_dirtyend = 0;
2222 	bp->b_bufobj = NULL;
2223 	bp->b_pin_count = 0;
2224 	bp->b_fsprivate1 = NULL;
2225 	bp->b_fsprivate2 = NULL;
2226 	bp->b_fsprivate3 = NULL;
2227 
2228 	LIST_INIT(&bp->b_dep);
2229 }
2230 
2231 static int flushingbufs;
2232 
2233 static struct buf *
2234 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata)
2235 {
2236 	struct buf *bp, *nbp;
2237 	int nqindex, qindex, pass;
2238 
2239 	KASSERT(!unmapped || !defrag, ("both unmapped and defrag"));
2240 
2241 	pass = 1;
2242 restart:
2243 	atomic_add_int(&getnewbufrestarts, 1);
2244 
2245 	/*
2246 	 * Setup for scan.  If we do not have enough free buffers,
2247 	 * we setup a degenerate case that immediately fails.  Note
2248 	 * that if we are specially marked process, we are allowed to
2249 	 * dip into our reserves.
2250 	 *
2251 	 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2252 	 * for the allocation of the mapped buffer.  For unmapped, the
2253 	 * easiest is to start with EMPTY outright.
2254 	 *
2255 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
2256 	 * However, there are a number of cases (defragging, reusing, ...)
2257 	 * where we cannot backup.
2258 	 */
2259 	nbp = NULL;
2260 	mtx_lock(&bqclean);
2261 	if (!defrag && unmapped) {
2262 		nqindex = QUEUE_EMPTY;
2263 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2264 	}
2265 	if (nbp == NULL) {
2266 		nqindex = QUEUE_EMPTYKVA;
2267 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2268 	}
2269 
2270 	/*
2271 	 * If no EMPTYKVA buffers and we are either defragging or
2272 	 * reusing, locate a CLEAN buffer to free or reuse.  If
2273 	 * bufspace useage is low skip this step so we can allocate a
2274 	 * new buffer.
2275 	 */
2276 	if (nbp == NULL && (defrag || bufspace >= lobufspace)) {
2277 		nqindex = QUEUE_CLEAN;
2278 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2279 	}
2280 
2281 	/*
2282 	 * If we could not find or were not allowed to reuse a CLEAN
2283 	 * buffer, check to see if it is ok to use an EMPTY buffer.
2284 	 * We can only use an EMPTY buffer if allocating its KVA would
2285 	 * not otherwise run us out of buffer space.  No KVA is needed
2286 	 * for the unmapped allocation.
2287 	 */
2288 	if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace ||
2289 	    metadata)) {
2290 		nqindex = QUEUE_EMPTY;
2291 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2292 	}
2293 
2294 	/*
2295 	 * All available buffers might be clean, retry ignoring the
2296 	 * lobufspace as the last resort.
2297 	 */
2298 	if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) {
2299 		nqindex = QUEUE_CLEAN;
2300 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2301 	}
2302 
2303 	/*
2304 	 * Run scan, possibly freeing data and/or kva mappings on the fly
2305 	 * depending.
2306 	 */
2307 	while ((bp = nbp) != NULL) {
2308 		qindex = nqindex;
2309 
2310 		/*
2311 		 * Calculate next bp (we can only use it if we do not
2312 		 * block or do other fancy things).
2313 		 */
2314 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
2315 			switch (qindex) {
2316 			case QUEUE_EMPTY:
2317 				nqindex = QUEUE_EMPTYKVA;
2318 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2319 				if (nbp != NULL)
2320 					break;
2321 				/* FALLTHROUGH */
2322 			case QUEUE_EMPTYKVA:
2323 				nqindex = QUEUE_CLEAN;
2324 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2325 				if (nbp != NULL)
2326 					break;
2327 				/* FALLTHROUGH */
2328 			case QUEUE_CLEAN:
2329 				if (metadata && pass == 1) {
2330 					pass = 2;
2331 					nqindex = QUEUE_EMPTY;
2332 					nbp = TAILQ_FIRST(
2333 					    &bufqueues[QUEUE_EMPTY]);
2334 				}
2335 				/*
2336 				 * nbp is NULL.
2337 				 */
2338 				break;
2339 			}
2340 		}
2341 		/*
2342 		 * If we are defragging then we need a buffer with
2343 		 * b_kvasize != 0.  XXX this situation should no longer
2344 		 * occur, if defrag is non-zero the buffer's b_kvasize
2345 		 * should also be non-zero at this point.  XXX
2346 		 */
2347 		if (defrag && bp->b_kvasize == 0) {
2348 			printf("Warning: defrag empty buffer %p\n", bp);
2349 			continue;
2350 		}
2351 
2352 		/*
2353 		 * Start freeing the bp.  This is somewhat involved.  nbp
2354 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
2355 		 */
2356 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2357 			continue;
2358 		/*
2359 		 * BKGRDINPROG can only be set with the buf and bufobj
2360 		 * locks both held.  We tolerate a race to clear it here.
2361 		 */
2362 		if (bp->b_vflags & BV_BKGRDINPROG) {
2363 			BUF_UNLOCK(bp);
2364 			continue;
2365 		}
2366 
2367 		KASSERT(bp->b_qindex == qindex,
2368 		    ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2369 
2370 		bremfreel(bp);
2371 		mtx_unlock(&bqclean);
2372 		/*
2373 		 * NOTE:  nbp is now entirely invalid.  We can only restart
2374 		 * the scan from this point on.
2375 		 */
2376 
2377 		getnewbuf_reuse_bp(bp, qindex);
2378 		mtx_assert(&bqclean, MA_NOTOWNED);
2379 
2380 		/*
2381 		 * If we are defragging then free the buffer.
2382 		 */
2383 		if (defrag) {
2384 			bp->b_flags |= B_INVAL;
2385 			bfreekva(bp);
2386 			brelse(bp);
2387 			defrag = 0;
2388 			goto restart;
2389 		}
2390 
2391 		/*
2392 		 * Notify any waiters for the buffer lock about
2393 		 * identity change by freeing the buffer.
2394 		 */
2395 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2396 			bp->b_flags |= B_INVAL;
2397 			bfreekva(bp);
2398 			brelse(bp);
2399 			goto restart;
2400 		}
2401 
2402 		if (metadata)
2403 			break;
2404 
2405 		/*
2406 		 * If we are overcomitted then recover the buffer and its
2407 		 * KVM space.  This occurs in rare situations when multiple
2408 		 * processes are blocked in getnewbuf() or allocbuf().
2409 		 */
2410 		if (bufspace >= hibufspace)
2411 			flushingbufs = 1;
2412 		if (flushingbufs && bp->b_kvasize != 0) {
2413 			bp->b_flags |= B_INVAL;
2414 			bfreekva(bp);
2415 			brelse(bp);
2416 			goto restart;
2417 		}
2418 		if (bufspace < lobufspace)
2419 			flushingbufs = 0;
2420 		break;
2421 	}
2422 	return (bp);
2423 }
2424 
2425 /*
2426  *	getnewbuf:
2427  *
2428  *	Find and initialize a new buffer header, freeing up existing buffers
2429  *	in the bufqueues as necessary.  The new buffer is returned locked.
2430  *
2431  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
2432  *	buffer away, the caller must set B_INVAL prior to calling brelse().
2433  *
2434  *	We block if:
2435  *		We have insufficient buffer headers
2436  *		We have insufficient buffer space
2437  *		buffer_arena is too fragmented ( space reservation fails )
2438  *		If we have to flush dirty buffers ( but we try to avoid this )
2439  */
2440 static struct buf *
2441 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
2442     int gbflags)
2443 {
2444 	struct buf *bp;
2445 	int defrag, metadata;
2446 
2447 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2448 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2449 	if (!unmapped_buf_allowed)
2450 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2451 
2452 	defrag = 0;
2453 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2454 	    vp->v_type == VCHR)
2455 		metadata = 1;
2456 	else
2457 		metadata = 0;
2458 	/*
2459 	 * We can't afford to block since we might be holding a vnode lock,
2460 	 * which may prevent system daemons from running.  We deal with
2461 	 * low-memory situations by proactively returning memory and running
2462 	 * async I/O rather then sync I/O.
2463 	 */
2464 	atomic_add_int(&getnewbufcalls, 1);
2465 	atomic_subtract_int(&getnewbufrestarts, 1);
2466 restart:
2467 	bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED |
2468 	    GB_KVAALLOC)) == GB_UNMAPPED, metadata);
2469 	if (bp != NULL)
2470 		defrag = 0;
2471 
2472 	/*
2473 	 * If we exhausted our list, sleep as appropriate.  We may have to
2474 	 * wakeup various daemons and write out some dirty buffers.
2475 	 *
2476 	 * Generally we are sleeping due to insufficient buffer space.
2477 	 */
2478 	if (bp == NULL) {
2479 		mtx_assert(&bqclean, MA_OWNED);
2480 		getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag);
2481 		mtx_assert(&bqclean, MA_NOTOWNED);
2482 	} else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) {
2483 		mtx_assert(&bqclean, MA_NOTOWNED);
2484 
2485 		bfreekva(bp);
2486 		bp->b_flags |= B_UNMAPPED;
2487 		bp->b_kvabase = bp->b_data = unmapped_buf;
2488 		bp->b_kvasize = maxsize;
2489 		atomic_add_long(&bufspace, bp->b_kvasize);
2490 		atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2491 		atomic_add_int(&bufreusecnt, 1);
2492 	} else {
2493 		mtx_assert(&bqclean, MA_NOTOWNED);
2494 
2495 		/*
2496 		 * We finally have a valid bp.  We aren't quite out of the
2497 		 * woods, we still have to reserve kva space.  In order
2498 		 * to keep fragmentation sane we only allocate kva in
2499 		 * BKVASIZE chunks.
2500 		 */
2501 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2502 
2503 		if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED |
2504 		    B_KVAALLOC)) == B_UNMAPPED) {
2505 			if (allocbufkva(bp, maxsize, gbflags)) {
2506 				defrag = 1;
2507 				bp->b_flags |= B_INVAL;
2508 				brelse(bp);
2509 				goto restart;
2510 			}
2511 			atomic_add_int(&bufreusecnt, 1);
2512 		} else if ((bp->b_flags & B_KVAALLOC) != 0 &&
2513 		    (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) {
2514 			/*
2515 			 * If the reused buffer has KVA allocated,
2516 			 * reassign b_kvaalloc to b_kvabase.
2517 			 */
2518 			bp->b_kvabase = bp->b_kvaalloc;
2519 			bp->b_flags &= ~B_KVAALLOC;
2520 			atomic_subtract_long(&unmapped_bufspace,
2521 			    bp->b_kvasize);
2522 			atomic_add_int(&bufreusecnt, 1);
2523 		} else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2524 		    (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED |
2525 		    GB_KVAALLOC)) {
2526 			/*
2527 			 * The case of reused buffer already have KVA
2528 			 * mapped, but the request is for unmapped
2529 			 * buffer with KVA allocated.
2530 			 */
2531 			bp->b_kvaalloc = bp->b_kvabase;
2532 			bp->b_data = bp->b_kvabase = unmapped_buf;
2533 			bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2534 			atomic_add_long(&unmapped_bufspace,
2535 			    bp->b_kvasize);
2536 			atomic_add_int(&bufreusecnt, 1);
2537 		}
2538 		if ((gbflags & GB_UNMAPPED) == 0) {
2539 			bp->b_saveaddr = bp->b_kvabase;
2540 			bp->b_data = bp->b_saveaddr;
2541 			bp->b_flags &= ~B_UNMAPPED;
2542 			BUF_CHECK_MAPPED(bp);
2543 		}
2544 	}
2545 	return (bp);
2546 }
2547 
2548 /*
2549  *	buf_daemon:
2550  *
2551  *	buffer flushing daemon.  Buffers are normally flushed by the
2552  *	update daemon but if it cannot keep up this process starts to
2553  *	take the load in an attempt to prevent getnewbuf() from blocking.
2554  */
2555 
2556 static struct kproc_desc buf_kp = {
2557 	"bufdaemon",
2558 	buf_daemon,
2559 	&bufdaemonproc
2560 };
2561 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2562 
2563 static int
2564 buf_flush(int target)
2565 {
2566 	int flushed;
2567 
2568 	flushed = flushbufqueues(target, 0);
2569 	if (flushed == 0) {
2570 		/*
2571 		 * Could not find any buffers without rollback
2572 		 * dependencies, so just write the first one
2573 		 * in the hopes of eventually making progress.
2574 		 */
2575 		flushed = flushbufqueues(target, 1);
2576 	}
2577 	return (flushed);
2578 }
2579 
2580 static void
2581 buf_daemon()
2582 {
2583 	int lodirty;
2584 
2585 	/*
2586 	 * This process needs to be suspended prior to shutdown sync.
2587 	 */
2588 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2589 	    SHUTDOWN_PRI_LAST);
2590 
2591 	/*
2592 	 * This process is allowed to take the buffer cache to the limit
2593 	 */
2594 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2595 	mtx_lock(&bdlock);
2596 	for (;;) {
2597 		bd_request = 0;
2598 		mtx_unlock(&bdlock);
2599 
2600 		kproc_suspend_check(bufdaemonproc);
2601 		lodirty = lodirtybuffers;
2602 		if (bd_speedupreq) {
2603 			lodirty = numdirtybuffers / 2;
2604 			bd_speedupreq = 0;
2605 		}
2606 		/*
2607 		 * Do the flush.  Limit the amount of in-transit I/O we
2608 		 * allow to build up, otherwise we would completely saturate
2609 		 * the I/O system.
2610 		 */
2611 		while (numdirtybuffers > lodirty) {
2612 			if (buf_flush(numdirtybuffers - lodirty) == 0)
2613 				break;
2614 			kern_yield(PRI_USER);
2615 		}
2616 
2617 		/*
2618 		 * Only clear bd_request if we have reached our low water
2619 		 * mark.  The buf_daemon normally waits 1 second and
2620 		 * then incrementally flushes any dirty buffers that have
2621 		 * built up, within reason.
2622 		 *
2623 		 * If we were unable to hit our low water mark and couldn't
2624 		 * find any flushable buffers, we sleep for a short period
2625 		 * to avoid endless loops on unlockable buffers.
2626 		 */
2627 		mtx_lock(&bdlock);
2628 		if (numdirtybuffers <= lodirtybuffers) {
2629 			/*
2630 			 * We reached our low water mark, reset the
2631 			 * request and sleep until we are needed again.
2632 			 * The sleep is just so the suspend code works.
2633 			 */
2634 			bd_request = 0;
2635 			/*
2636 			 * Do an extra wakeup in case dirty threshold
2637 			 * changed via sysctl and the explicit transition
2638 			 * out of shortfall was missed.
2639 			 */
2640 			bdirtywakeup();
2641 			if (runningbufspace <= lorunningspace)
2642 				runningwakeup();
2643 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2644 		} else {
2645 			/*
2646 			 * We couldn't find any flushable dirty buffers but
2647 			 * still have too many dirty buffers, we
2648 			 * have to sleep and try again.  (rare)
2649 			 */
2650 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2651 		}
2652 	}
2653 }
2654 
2655 /*
2656  *	flushbufqueues:
2657  *
2658  *	Try to flush a buffer in the dirty queue.  We must be careful to
2659  *	free up B_INVAL buffers instead of write them, which NFS is
2660  *	particularly sensitive to.
2661  */
2662 static int flushwithdeps = 0;
2663 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2664     0, "Number of buffers flushed with dependecies that require rollbacks");
2665 
2666 static int
2667 flushbufqueues(int target, int flushdeps)
2668 {
2669 	struct buf *sentinel;
2670 	struct vnode *vp;
2671 	struct mount *mp;
2672 	struct buf *bp;
2673 	int hasdeps;
2674 	int flushed;
2675 	int queue;
2676 	int error;
2677 
2678 	flushed = 0;
2679 	queue = QUEUE_DIRTY;
2680 	bp = NULL;
2681 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2682 	sentinel->b_qindex = QUEUE_SENTINEL;
2683 	mtx_lock(&bqdirty);
2684 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2685 	mtx_unlock(&bqdirty);
2686 	while (flushed != target) {
2687 		maybe_yield();
2688 		mtx_lock(&bqdirty);
2689 		bp = TAILQ_NEXT(sentinel, b_freelist);
2690 		if (bp != NULL) {
2691 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2692 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2693 			    b_freelist);
2694 		} else {
2695 			mtx_unlock(&bqdirty);
2696 			break;
2697 		}
2698 		KASSERT(bp->b_qindex != QUEUE_SENTINEL,
2699 		    ("parallel calls to flushbufqueues() bp %p", bp));
2700 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
2701 		mtx_unlock(&bqdirty);
2702 		if (error != 0)
2703 			continue;
2704 		if (bp->b_pin_count > 0) {
2705 			BUF_UNLOCK(bp);
2706 			continue;
2707 		}
2708 		/*
2709 		 * BKGRDINPROG can only be set with the buf and bufobj
2710 		 * locks both held.  We tolerate a race to clear it here.
2711 		 */
2712 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2713 		    (bp->b_flags & B_DELWRI) == 0) {
2714 			BUF_UNLOCK(bp);
2715 			continue;
2716 		}
2717 		if (bp->b_flags & B_INVAL) {
2718 			bremfreef(bp);
2719 			brelse(bp);
2720 			flushed++;
2721 			continue;
2722 		}
2723 
2724 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2725 			if (flushdeps == 0) {
2726 				BUF_UNLOCK(bp);
2727 				continue;
2728 			}
2729 			hasdeps = 1;
2730 		} else
2731 			hasdeps = 0;
2732 		/*
2733 		 * We must hold the lock on a vnode before writing
2734 		 * one of its buffers. Otherwise we may confuse, or
2735 		 * in the case of a snapshot vnode, deadlock the
2736 		 * system.
2737 		 *
2738 		 * The lock order here is the reverse of the normal
2739 		 * of vnode followed by buf lock.  This is ok because
2740 		 * the NOWAIT will prevent deadlock.
2741 		 */
2742 		vp = bp->b_vp;
2743 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2744 			BUF_UNLOCK(bp);
2745 			continue;
2746 		}
2747 		error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
2748 		if (error == 0) {
2749 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2750 			    bp, bp->b_vp, bp->b_flags);
2751 			vfs_bio_awrite(bp);
2752 			vn_finished_write(mp);
2753 			VOP_UNLOCK(vp, 0);
2754 			flushwithdeps += hasdeps;
2755 			flushed++;
2756 			if (runningbufspace > hirunningspace)
2757 				waitrunningbufspace();
2758 			continue;
2759 		}
2760 		vn_finished_write(mp);
2761 		BUF_UNLOCK(bp);
2762 	}
2763 	mtx_lock(&bqdirty);
2764 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2765 	mtx_unlock(&bqdirty);
2766 	free(sentinel, M_TEMP);
2767 	return (flushed);
2768 }
2769 
2770 /*
2771  * Check to see if a block is currently memory resident.
2772  */
2773 struct buf *
2774 incore(struct bufobj *bo, daddr_t blkno)
2775 {
2776 	struct buf *bp;
2777 
2778 	BO_RLOCK(bo);
2779 	bp = gbincore(bo, blkno);
2780 	BO_RUNLOCK(bo);
2781 	return (bp);
2782 }
2783 
2784 /*
2785  * Returns true if no I/O is needed to access the
2786  * associated VM object.  This is like incore except
2787  * it also hunts around in the VM system for the data.
2788  */
2789 
2790 static int
2791 inmem(struct vnode * vp, daddr_t blkno)
2792 {
2793 	vm_object_t obj;
2794 	vm_offset_t toff, tinc, size;
2795 	vm_page_t m;
2796 	vm_ooffset_t off;
2797 
2798 	ASSERT_VOP_LOCKED(vp, "inmem");
2799 
2800 	if (incore(&vp->v_bufobj, blkno))
2801 		return 1;
2802 	if (vp->v_mount == NULL)
2803 		return 0;
2804 	obj = vp->v_object;
2805 	if (obj == NULL)
2806 		return (0);
2807 
2808 	size = PAGE_SIZE;
2809 	if (size > vp->v_mount->mnt_stat.f_iosize)
2810 		size = vp->v_mount->mnt_stat.f_iosize;
2811 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2812 
2813 	VM_OBJECT_RLOCK(obj);
2814 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2815 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2816 		if (!m)
2817 			goto notinmem;
2818 		tinc = size;
2819 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2820 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2821 		if (vm_page_is_valid(m,
2822 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2823 			goto notinmem;
2824 	}
2825 	VM_OBJECT_RUNLOCK(obj);
2826 	return 1;
2827 
2828 notinmem:
2829 	VM_OBJECT_RUNLOCK(obj);
2830 	return (0);
2831 }
2832 
2833 /*
2834  * Set the dirty range for a buffer based on the status of the dirty
2835  * bits in the pages comprising the buffer.  The range is limited
2836  * to the size of the buffer.
2837  *
2838  * Tell the VM system that the pages associated with this buffer
2839  * are clean.  This is used for delayed writes where the data is
2840  * going to go to disk eventually without additional VM intevention.
2841  *
2842  * Note that while we only really need to clean through to b_bcount, we
2843  * just go ahead and clean through to b_bufsize.
2844  */
2845 static void
2846 vfs_clean_pages_dirty_buf(struct buf *bp)
2847 {
2848 	vm_ooffset_t foff, noff, eoff;
2849 	vm_page_t m;
2850 	int i;
2851 
2852 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2853 		return;
2854 
2855 	foff = bp->b_offset;
2856 	KASSERT(bp->b_offset != NOOFFSET,
2857 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
2858 
2859 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
2860 	vfs_drain_busy_pages(bp);
2861 	vfs_setdirty_locked_object(bp);
2862 	for (i = 0; i < bp->b_npages; i++) {
2863 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2864 		eoff = noff;
2865 		if (eoff > bp->b_offset + bp->b_bufsize)
2866 			eoff = bp->b_offset + bp->b_bufsize;
2867 		m = bp->b_pages[i];
2868 		vfs_page_set_validclean(bp, foff, m);
2869 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2870 		foff = noff;
2871 	}
2872 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
2873 }
2874 
2875 static void
2876 vfs_setdirty_locked_object(struct buf *bp)
2877 {
2878 	vm_object_t object;
2879 	int i;
2880 
2881 	object = bp->b_bufobj->bo_object;
2882 	VM_OBJECT_ASSERT_WLOCKED(object);
2883 
2884 	/*
2885 	 * We qualify the scan for modified pages on whether the
2886 	 * object has been flushed yet.
2887 	 */
2888 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2889 		vm_offset_t boffset;
2890 		vm_offset_t eoffset;
2891 
2892 		/*
2893 		 * test the pages to see if they have been modified directly
2894 		 * by users through the VM system.
2895 		 */
2896 		for (i = 0; i < bp->b_npages; i++)
2897 			vm_page_test_dirty(bp->b_pages[i]);
2898 
2899 		/*
2900 		 * Calculate the encompassing dirty range, boffset and eoffset,
2901 		 * (eoffset - boffset) bytes.
2902 		 */
2903 
2904 		for (i = 0; i < bp->b_npages; i++) {
2905 			if (bp->b_pages[i]->dirty)
2906 				break;
2907 		}
2908 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2909 
2910 		for (i = bp->b_npages - 1; i >= 0; --i) {
2911 			if (bp->b_pages[i]->dirty) {
2912 				break;
2913 			}
2914 		}
2915 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2916 
2917 		/*
2918 		 * Fit it to the buffer.
2919 		 */
2920 
2921 		if (eoffset > bp->b_bcount)
2922 			eoffset = bp->b_bcount;
2923 
2924 		/*
2925 		 * If we have a good dirty range, merge with the existing
2926 		 * dirty range.
2927 		 */
2928 
2929 		if (boffset < eoffset) {
2930 			if (bp->b_dirtyoff > boffset)
2931 				bp->b_dirtyoff = boffset;
2932 			if (bp->b_dirtyend < eoffset)
2933 				bp->b_dirtyend = eoffset;
2934 		}
2935 	}
2936 }
2937 
2938 /*
2939  * Allocate the KVA mapping for an existing buffer. It handles the
2940  * cases of both B_UNMAPPED buffer, and buffer with the preallocated
2941  * KVA which is not mapped (B_KVAALLOC).
2942  */
2943 static void
2944 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
2945 {
2946 	struct buf *scratch_bp;
2947 	int bsize, maxsize, need_mapping, need_kva;
2948 	off_t offset;
2949 
2950 	need_mapping = (bp->b_flags & B_UNMAPPED) != 0 &&
2951 	    (gbflags & GB_UNMAPPED) == 0;
2952 	need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED &&
2953 	    (gbflags & GB_KVAALLOC) != 0;
2954 	if (!need_mapping && !need_kva)
2955 		return;
2956 
2957 	BUF_CHECK_UNMAPPED(bp);
2958 
2959 	if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) {
2960 		/*
2961 		 * Buffer is not mapped, but the KVA was already
2962 		 * reserved at the time of the instantiation.  Use the
2963 		 * allocated space.
2964 		 */
2965 		bp->b_flags &= ~B_KVAALLOC;
2966 		KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0"));
2967 		bp->b_kvabase = bp->b_kvaalloc;
2968 		atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
2969 		goto has_addr;
2970 	}
2971 
2972 	/*
2973 	 * Calculate the amount of the address space we would reserve
2974 	 * if the buffer was mapped.
2975 	 */
2976 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
2977 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
2978 	offset = blkno * bsize;
2979 	maxsize = size + (offset & PAGE_MASK);
2980 	maxsize = imax(maxsize, bsize);
2981 
2982 mapping_loop:
2983 	if (allocbufkva(bp, maxsize, gbflags)) {
2984 		/*
2985 		 * Request defragmentation. getnewbuf() returns us the
2986 		 * allocated space by the scratch buffer KVA.
2987 		 */
2988 		scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags |
2989 		    (GB_UNMAPPED | GB_KVAALLOC));
2990 		if (scratch_bp == NULL) {
2991 			if ((gbflags & GB_NOWAIT_BD) != 0) {
2992 				/*
2993 				 * XXXKIB: defragmentation cannot
2994 				 * succeed, not sure what else to do.
2995 				 */
2996 				panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp);
2997 			}
2998 			atomic_add_int(&mappingrestarts, 1);
2999 			goto mapping_loop;
3000 		}
3001 		KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0,
3002 		    ("scratch bp !B_KVAALLOC %p", scratch_bp));
3003 		setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc,
3004 		    scratch_bp->b_kvasize, gbflags);
3005 
3006 		/* Get rid of the scratch buffer. */
3007 		scratch_bp->b_kvasize = 0;
3008 		scratch_bp->b_flags |= B_INVAL;
3009 		scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
3010 		brelse(scratch_bp);
3011 	}
3012 	if (!need_mapping)
3013 		return;
3014 
3015 has_addr:
3016 	bp->b_saveaddr = bp->b_kvabase;
3017 	bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */
3018 	bp->b_flags &= ~B_UNMAPPED;
3019 	BUF_CHECK_MAPPED(bp);
3020 	bpmap_qenter(bp);
3021 }
3022 
3023 /*
3024  *	getblk:
3025  *
3026  *	Get a block given a specified block and offset into a file/device.
3027  *	The buffers B_DONE bit will be cleared on return, making it almost
3028  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3029  *	return.  The caller should clear B_INVAL prior to initiating a
3030  *	READ.
3031  *
3032  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3033  *	an existing buffer.
3034  *
3035  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3036  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3037  *	and then cleared based on the backing VM.  If the previous buffer is
3038  *	non-0-sized but invalid, B_CACHE will be cleared.
3039  *
3040  *	If getblk() must create a new buffer, the new buffer is returned with
3041  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3042  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3043  *	backing VM.
3044  *
3045  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3046  *	B_CACHE bit is clear.
3047  *
3048  *	What this means, basically, is that the caller should use B_CACHE to
3049  *	determine whether the buffer is fully valid or not and should clear
3050  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3051  *	the buffer by loading its data area with something, the caller needs
3052  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3053  *	the caller should set B_CACHE ( as an optimization ), else the caller
3054  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3055  *	a write attempt or if it was a successfull read.  If the caller
3056  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3057  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3058  */
3059 struct buf *
3060 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3061     int flags)
3062 {
3063 	struct buf *bp;
3064 	struct bufobj *bo;
3065 	int bsize, error, maxsize, vmio;
3066 	off_t offset;
3067 
3068 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3069 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3070 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3071 	ASSERT_VOP_LOCKED(vp, "getblk");
3072 	if (size > MAXBSIZE)
3073 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
3074 	if (!unmapped_buf_allowed)
3075 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3076 
3077 	bo = &vp->v_bufobj;
3078 loop:
3079 	BO_RLOCK(bo);
3080 	bp = gbincore(bo, blkno);
3081 	if (bp != NULL) {
3082 		int lockflags;
3083 		/*
3084 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3085 		 * it must be on a queue.
3086 		 */
3087 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3088 
3089 		if (flags & GB_LOCK_NOWAIT)
3090 			lockflags |= LK_NOWAIT;
3091 
3092 		error = BUF_TIMELOCK(bp, lockflags,
3093 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3094 
3095 		/*
3096 		 * If we slept and got the lock we have to restart in case
3097 		 * the buffer changed identities.
3098 		 */
3099 		if (error == ENOLCK)
3100 			goto loop;
3101 		/* We timed out or were interrupted. */
3102 		else if (error)
3103 			return (NULL);
3104 		/* If recursed, assume caller knows the rules. */
3105 		else if (BUF_LOCKRECURSED(bp))
3106 			goto end;
3107 
3108 		/*
3109 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3110 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3111 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3112 		 * backing VM cache.
3113 		 */
3114 		if (bp->b_flags & B_INVAL)
3115 			bp->b_flags &= ~B_CACHE;
3116 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3117 			bp->b_flags |= B_CACHE;
3118 		if (bp->b_flags & B_MANAGED)
3119 			MPASS(bp->b_qindex == QUEUE_NONE);
3120 		else
3121 			bremfree(bp);
3122 
3123 		/*
3124 		 * check for size inconsistencies for non-VMIO case.
3125 		 */
3126 		if (bp->b_bcount != size) {
3127 			if ((bp->b_flags & B_VMIO) == 0 ||
3128 			    (size > bp->b_kvasize)) {
3129 				if (bp->b_flags & B_DELWRI) {
3130 					/*
3131 					 * If buffer is pinned and caller does
3132 					 * not want sleep  waiting for it to be
3133 					 * unpinned, bail out
3134 					 * */
3135 					if (bp->b_pin_count > 0) {
3136 						if (flags & GB_LOCK_NOWAIT) {
3137 							bqrelse(bp);
3138 							return (NULL);
3139 						} else {
3140 							bunpin_wait(bp);
3141 						}
3142 					}
3143 					bp->b_flags |= B_NOCACHE;
3144 					bwrite(bp);
3145 				} else {
3146 					if (LIST_EMPTY(&bp->b_dep)) {
3147 						bp->b_flags |= B_RELBUF;
3148 						brelse(bp);
3149 					} else {
3150 						bp->b_flags |= B_NOCACHE;
3151 						bwrite(bp);
3152 					}
3153 				}
3154 				goto loop;
3155 			}
3156 		}
3157 
3158 		/*
3159 		 * Handle the case of unmapped buffer which should
3160 		 * become mapped, or the buffer for which KVA
3161 		 * reservation is requested.
3162 		 */
3163 		bp_unmapped_get_kva(bp, blkno, size, flags);
3164 
3165 		/*
3166 		 * If the size is inconsistant in the VMIO case, we can resize
3167 		 * the buffer.  This might lead to B_CACHE getting set or
3168 		 * cleared.  If the size has not changed, B_CACHE remains
3169 		 * unchanged from its previous state.
3170 		 */
3171 		if (bp->b_bcount != size)
3172 			allocbuf(bp, size);
3173 
3174 		KASSERT(bp->b_offset != NOOFFSET,
3175 		    ("getblk: no buffer offset"));
3176 
3177 		/*
3178 		 * A buffer with B_DELWRI set and B_CACHE clear must
3179 		 * be committed before we can return the buffer in
3180 		 * order to prevent the caller from issuing a read
3181 		 * ( due to B_CACHE not being set ) and overwriting
3182 		 * it.
3183 		 *
3184 		 * Most callers, including NFS and FFS, need this to
3185 		 * operate properly either because they assume they
3186 		 * can issue a read if B_CACHE is not set, or because
3187 		 * ( for example ) an uncached B_DELWRI might loop due
3188 		 * to softupdates re-dirtying the buffer.  In the latter
3189 		 * case, B_CACHE is set after the first write completes,
3190 		 * preventing further loops.
3191 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3192 		 * above while extending the buffer, we cannot allow the
3193 		 * buffer to remain with B_CACHE set after the write
3194 		 * completes or it will represent a corrupt state.  To
3195 		 * deal with this we set B_NOCACHE to scrap the buffer
3196 		 * after the write.
3197 		 *
3198 		 * We might be able to do something fancy, like setting
3199 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3200 		 * so the below call doesn't set B_CACHE, but that gets real
3201 		 * confusing.  This is much easier.
3202 		 */
3203 
3204 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3205 			bp->b_flags |= B_NOCACHE;
3206 			bwrite(bp);
3207 			goto loop;
3208 		}
3209 		bp->b_flags &= ~B_DONE;
3210 	} else {
3211 		/*
3212 		 * Buffer is not in-core, create new buffer.  The buffer
3213 		 * returned by getnewbuf() is locked.  Note that the returned
3214 		 * buffer is also considered valid (not marked B_INVAL).
3215 		 */
3216 		BO_RUNLOCK(bo);
3217 		/*
3218 		 * If the user does not want us to create the buffer, bail out
3219 		 * here.
3220 		 */
3221 		if (flags & GB_NOCREAT)
3222 			return NULL;
3223 		if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3224 			return NULL;
3225 
3226 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3227 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3228 		offset = blkno * bsize;
3229 		vmio = vp->v_object != NULL;
3230 		if (vmio) {
3231 			maxsize = size + (offset & PAGE_MASK);
3232 		} else {
3233 			maxsize = size;
3234 			/* Do not allow non-VMIO notmapped buffers. */
3235 			flags &= ~GB_UNMAPPED;
3236 		}
3237 		maxsize = imax(maxsize, bsize);
3238 
3239 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
3240 		if (bp == NULL) {
3241 			if (slpflag || slptimeo)
3242 				return NULL;
3243 			goto loop;
3244 		}
3245 
3246 		/*
3247 		 * This code is used to make sure that a buffer is not
3248 		 * created while the getnewbuf routine is blocked.
3249 		 * This can be a problem whether the vnode is locked or not.
3250 		 * If the buffer is created out from under us, we have to
3251 		 * throw away the one we just created.
3252 		 *
3253 		 * Note: this must occur before we associate the buffer
3254 		 * with the vp especially considering limitations in
3255 		 * the splay tree implementation when dealing with duplicate
3256 		 * lblkno's.
3257 		 */
3258 		BO_LOCK(bo);
3259 		if (gbincore(bo, blkno)) {
3260 			BO_UNLOCK(bo);
3261 			bp->b_flags |= B_INVAL;
3262 			brelse(bp);
3263 			goto loop;
3264 		}
3265 
3266 		/*
3267 		 * Insert the buffer into the hash, so that it can
3268 		 * be found by incore.
3269 		 */
3270 		bp->b_blkno = bp->b_lblkno = blkno;
3271 		bp->b_offset = offset;
3272 		bgetvp(vp, bp);
3273 		BO_UNLOCK(bo);
3274 
3275 		/*
3276 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
3277 		 * buffer size starts out as 0, B_CACHE will be set by
3278 		 * allocbuf() for the VMIO case prior to it testing the
3279 		 * backing store for validity.
3280 		 */
3281 
3282 		if (vmio) {
3283 			bp->b_flags |= B_VMIO;
3284 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3285 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3286 			    bp, vp->v_object, bp->b_bufobj->bo_object));
3287 		} else {
3288 			bp->b_flags &= ~B_VMIO;
3289 			KASSERT(bp->b_bufobj->bo_object == NULL,
3290 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
3291 			    bp, bp->b_bufobj->bo_object));
3292 			BUF_CHECK_MAPPED(bp);
3293 		}
3294 
3295 		allocbuf(bp, size);
3296 		bp->b_flags &= ~B_DONE;
3297 	}
3298 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3299 	BUF_ASSERT_HELD(bp);
3300 end:
3301 	KASSERT(bp->b_bufobj == bo,
3302 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3303 	return (bp);
3304 }
3305 
3306 /*
3307  * Get an empty, disassociated buffer of given size.  The buffer is initially
3308  * set to B_INVAL.
3309  */
3310 struct buf *
3311 geteblk(int size, int flags)
3312 {
3313 	struct buf *bp;
3314 	int maxsize;
3315 
3316 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
3317 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
3318 		if ((flags & GB_NOWAIT_BD) &&
3319 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
3320 			return (NULL);
3321 	}
3322 	allocbuf(bp, size);
3323 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
3324 	BUF_ASSERT_HELD(bp);
3325 	return (bp);
3326 }
3327 
3328 
3329 /*
3330  * This code constitutes the buffer memory from either anonymous system
3331  * memory (in the case of non-VMIO operations) or from an associated
3332  * VM object (in the case of VMIO operations).  This code is able to
3333  * resize a buffer up or down.
3334  *
3335  * Note that this code is tricky, and has many complications to resolve
3336  * deadlock or inconsistant data situations.  Tread lightly!!!
3337  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3338  * the caller.  Calling this code willy nilly can result in the loss of data.
3339  *
3340  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
3341  * B_CACHE for the non-VMIO case.
3342  */
3343 
3344 int
3345 allocbuf(struct buf *bp, int size)
3346 {
3347 	int newbsize, mbsize;
3348 	int i;
3349 
3350 	BUF_ASSERT_HELD(bp);
3351 
3352 	if (bp->b_kvasize < size)
3353 		panic("allocbuf: buffer too small");
3354 
3355 	if ((bp->b_flags & B_VMIO) == 0) {
3356 		caddr_t origbuf;
3357 		int origbufsize;
3358 		/*
3359 		 * Just get anonymous memory from the kernel.  Don't
3360 		 * mess with B_CACHE.
3361 		 */
3362 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3363 		if (bp->b_flags & B_MALLOC)
3364 			newbsize = mbsize;
3365 		else
3366 			newbsize = round_page(size);
3367 
3368 		if (newbsize < bp->b_bufsize) {
3369 			/*
3370 			 * malloced buffers are not shrunk
3371 			 */
3372 			if (bp->b_flags & B_MALLOC) {
3373 				if (newbsize) {
3374 					bp->b_bcount = size;
3375 				} else {
3376 					free(bp->b_data, M_BIOBUF);
3377 					if (bp->b_bufsize) {
3378 						atomic_subtract_long(
3379 						    &bufmallocspace,
3380 						    bp->b_bufsize);
3381 						bufspacewakeup();
3382 						bp->b_bufsize = 0;
3383 					}
3384 					bp->b_saveaddr = bp->b_kvabase;
3385 					bp->b_data = bp->b_saveaddr;
3386 					bp->b_bcount = 0;
3387 					bp->b_flags &= ~B_MALLOC;
3388 				}
3389 				return 1;
3390 			}
3391 			vm_hold_free_pages(bp, newbsize);
3392 		} else if (newbsize > bp->b_bufsize) {
3393 			/*
3394 			 * We only use malloced memory on the first allocation.
3395 			 * and revert to page-allocated memory when the buffer
3396 			 * grows.
3397 			 */
3398 			/*
3399 			 * There is a potential smp race here that could lead
3400 			 * to bufmallocspace slightly passing the max.  It
3401 			 * is probably extremely rare and not worth worrying
3402 			 * over.
3403 			 */
3404 			if ( (bufmallocspace < maxbufmallocspace) &&
3405 				(bp->b_bufsize == 0) &&
3406 				(mbsize <= PAGE_SIZE/2)) {
3407 
3408 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
3409 				bp->b_bufsize = mbsize;
3410 				bp->b_bcount = size;
3411 				bp->b_flags |= B_MALLOC;
3412 				atomic_add_long(&bufmallocspace, mbsize);
3413 				return 1;
3414 			}
3415 			origbuf = NULL;
3416 			origbufsize = 0;
3417 			/*
3418 			 * If the buffer is growing on its other-than-first allocation,
3419 			 * then we revert to the page-allocation scheme.
3420 			 */
3421 			if (bp->b_flags & B_MALLOC) {
3422 				origbuf = bp->b_data;
3423 				origbufsize = bp->b_bufsize;
3424 				bp->b_data = bp->b_kvabase;
3425 				if (bp->b_bufsize) {
3426 					atomic_subtract_long(&bufmallocspace,
3427 					    bp->b_bufsize);
3428 					bufspacewakeup();
3429 					bp->b_bufsize = 0;
3430 				}
3431 				bp->b_flags &= ~B_MALLOC;
3432 				newbsize = round_page(newbsize);
3433 			}
3434 			vm_hold_load_pages(
3435 			    bp,
3436 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
3437 			    (vm_offset_t) bp->b_data + newbsize);
3438 			if (origbuf) {
3439 				bcopy(origbuf, bp->b_data, origbufsize);
3440 				free(origbuf, M_BIOBUF);
3441 			}
3442 		}
3443 	} else {
3444 		int desiredpages;
3445 
3446 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3447 		desiredpages = (size == 0) ? 0 :
3448 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3449 
3450 		if (bp->b_flags & B_MALLOC)
3451 			panic("allocbuf: VMIO buffer can't be malloced");
3452 		/*
3453 		 * Set B_CACHE initially if buffer is 0 length or will become
3454 		 * 0-length.
3455 		 */
3456 		if (size == 0 || bp->b_bufsize == 0)
3457 			bp->b_flags |= B_CACHE;
3458 
3459 		if (newbsize < bp->b_bufsize) {
3460 			/*
3461 			 * DEV_BSIZE aligned new buffer size is less then the
3462 			 * DEV_BSIZE aligned existing buffer size.  Figure out
3463 			 * if we have to remove any pages.
3464 			 */
3465 			if (desiredpages < bp->b_npages) {
3466 				vm_page_t m;
3467 
3468 				if ((bp->b_flags & B_UNMAPPED) == 0) {
3469 					BUF_CHECK_MAPPED(bp);
3470 					pmap_qremove((vm_offset_t)trunc_page(
3471 					    (vm_offset_t)bp->b_data) +
3472 					    (desiredpages << PAGE_SHIFT),
3473 					    (bp->b_npages - desiredpages));
3474 				} else
3475 					BUF_CHECK_UNMAPPED(bp);
3476 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3477 				for (i = desiredpages; i < bp->b_npages; i++) {
3478 					/*
3479 					 * the page is not freed here -- it
3480 					 * is the responsibility of
3481 					 * vnode_pager_setsize
3482 					 */
3483 					m = bp->b_pages[i];
3484 					KASSERT(m != bogus_page,
3485 					    ("allocbuf: bogus page found"));
3486 					while (vm_page_sleep_if_busy(m,
3487 					    "biodep"))
3488 						continue;
3489 
3490 					bp->b_pages[i] = NULL;
3491 					vm_page_lock(m);
3492 					vm_page_unwire(m, PQ_INACTIVE);
3493 					vm_page_unlock(m);
3494 				}
3495 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3496 				bp->b_npages = desiredpages;
3497 			}
3498 		} else if (size > bp->b_bcount) {
3499 			/*
3500 			 * We are growing the buffer, possibly in a
3501 			 * byte-granular fashion.
3502 			 */
3503 			vm_object_t obj;
3504 			vm_offset_t toff;
3505 			vm_offset_t tinc;
3506 
3507 			/*
3508 			 * Step 1, bring in the VM pages from the object,
3509 			 * allocating them if necessary.  We must clear
3510 			 * B_CACHE if these pages are not valid for the
3511 			 * range covered by the buffer.
3512 			 */
3513 
3514 			obj = bp->b_bufobj->bo_object;
3515 
3516 			VM_OBJECT_WLOCK(obj);
3517 			while (bp->b_npages < desiredpages) {
3518 				vm_page_t m;
3519 
3520 				/*
3521 				 * We must allocate system pages since blocking
3522 				 * here could interfere with paging I/O, no
3523 				 * matter which process we are.
3524 				 *
3525 				 * Only exclusive busy can be tested here.
3526 				 * Blocking on shared busy might lead to
3527 				 * deadlocks once allocbuf() is called after
3528 				 * pages are vfs_busy_pages().
3529 				 */
3530 				m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3531 				    bp->b_npages, VM_ALLOC_NOBUSY |
3532 				    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3533 				    VM_ALLOC_IGN_SBUSY |
3534 				    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3535 				if (m->valid == 0)
3536 					bp->b_flags &= ~B_CACHE;
3537 				bp->b_pages[bp->b_npages] = m;
3538 				++bp->b_npages;
3539 			}
3540 
3541 			/*
3542 			 * Step 2.  We've loaded the pages into the buffer,
3543 			 * we have to figure out if we can still have B_CACHE
3544 			 * set.  Note that B_CACHE is set according to the
3545 			 * byte-granular range ( bcount and size ), new the
3546 			 * aligned range ( newbsize ).
3547 			 *
3548 			 * The VM test is against m->valid, which is DEV_BSIZE
3549 			 * aligned.  Needless to say, the validity of the data
3550 			 * needs to also be DEV_BSIZE aligned.  Note that this
3551 			 * fails with NFS if the server or some other client
3552 			 * extends the file's EOF.  If our buffer is resized,
3553 			 * B_CACHE may remain set! XXX
3554 			 */
3555 
3556 			toff = bp->b_bcount;
3557 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3558 
3559 			while ((bp->b_flags & B_CACHE) && toff < size) {
3560 				vm_pindex_t pi;
3561 
3562 				if (tinc > (size - toff))
3563 					tinc = size - toff;
3564 
3565 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3566 				    PAGE_SHIFT;
3567 
3568 				vfs_buf_test_cache(
3569 				    bp,
3570 				    bp->b_offset,
3571 				    toff,
3572 				    tinc,
3573 				    bp->b_pages[pi]
3574 				);
3575 				toff += tinc;
3576 				tinc = PAGE_SIZE;
3577 			}
3578 			VM_OBJECT_WUNLOCK(obj);
3579 
3580 			/*
3581 			 * Step 3, fixup the KVM pmap.
3582 			 */
3583 			if ((bp->b_flags & B_UNMAPPED) == 0)
3584 				bpmap_qenter(bp);
3585 			else
3586 				BUF_CHECK_UNMAPPED(bp);
3587 		}
3588 	}
3589 	if (newbsize < bp->b_bufsize)
3590 		bufspacewakeup();
3591 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3592 	bp->b_bcount = size;		/* requested buffer size	*/
3593 	return 1;
3594 }
3595 
3596 extern int inflight_transient_maps;
3597 
3598 void
3599 biodone(struct bio *bp)
3600 {
3601 	struct mtx *mtxp;
3602 	void (*done)(struct bio *);
3603 	vm_offset_t start, end;
3604 
3605 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3606 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3607 		bp->bio_flags |= BIO_UNMAPPED;
3608 		start = trunc_page((vm_offset_t)bp->bio_data);
3609 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3610 		pmap_qremove(start, OFF_TO_IDX(end - start));
3611 		vmem_free(transient_arena, start, end - start);
3612 		atomic_add_int(&inflight_transient_maps, -1);
3613 	}
3614 	done = bp->bio_done;
3615 	if (done == NULL) {
3616 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
3617 		mtx_lock(mtxp);
3618 		bp->bio_flags |= BIO_DONE;
3619 		wakeup(bp);
3620 		mtx_unlock(mtxp);
3621 	} else {
3622 		bp->bio_flags |= BIO_DONE;
3623 		done(bp);
3624 	}
3625 }
3626 
3627 /*
3628  * Wait for a BIO to finish.
3629  */
3630 int
3631 biowait(struct bio *bp, const char *wchan)
3632 {
3633 	struct mtx *mtxp;
3634 
3635 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3636 	mtx_lock(mtxp);
3637 	while ((bp->bio_flags & BIO_DONE) == 0)
3638 		msleep(bp, mtxp, PRIBIO, wchan, 0);
3639 	mtx_unlock(mtxp);
3640 	if (bp->bio_error != 0)
3641 		return (bp->bio_error);
3642 	if (!(bp->bio_flags & BIO_ERROR))
3643 		return (0);
3644 	return (EIO);
3645 }
3646 
3647 void
3648 biofinish(struct bio *bp, struct devstat *stat, int error)
3649 {
3650 
3651 	if (error) {
3652 		bp->bio_error = error;
3653 		bp->bio_flags |= BIO_ERROR;
3654 	}
3655 	if (stat != NULL)
3656 		devstat_end_transaction_bio(stat, bp);
3657 	biodone(bp);
3658 }
3659 
3660 /*
3661  *	bufwait:
3662  *
3663  *	Wait for buffer I/O completion, returning error status.  The buffer
3664  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3665  *	error and cleared.
3666  */
3667 int
3668 bufwait(struct buf *bp)
3669 {
3670 	if (bp->b_iocmd == BIO_READ)
3671 		bwait(bp, PRIBIO, "biord");
3672 	else
3673 		bwait(bp, PRIBIO, "biowr");
3674 	if (bp->b_flags & B_EINTR) {
3675 		bp->b_flags &= ~B_EINTR;
3676 		return (EINTR);
3677 	}
3678 	if (bp->b_ioflags & BIO_ERROR) {
3679 		return (bp->b_error ? bp->b_error : EIO);
3680 	} else {
3681 		return (0);
3682 	}
3683 }
3684 
3685  /*
3686   * Call back function from struct bio back up to struct buf.
3687   */
3688 static void
3689 bufdonebio(struct bio *bip)
3690 {
3691 	struct buf *bp;
3692 
3693 	bp = bip->bio_caller2;
3694 	bp->b_resid = bip->bio_resid;
3695 	bp->b_ioflags = bip->bio_flags;
3696 	bp->b_error = bip->bio_error;
3697 	if (bp->b_error)
3698 		bp->b_ioflags |= BIO_ERROR;
3699 	bufdone(bp);
3700 	g_destroy_bio(bip);
3701 }
3702 
3703 void
3704 dev_strategy(struct cdev *dev, struct buf *bp)
3705 {
3706 	struct cdevsw *csw;
3707 	int ref;
3708 
3709 	KASSERT(dev->si_refcount > 0,
3710 	    ("dev_strategy on un-referenced struct cdev *(%s) %p",
3711 	    devtoname(dev), dev));
3712 
3713 	csw = dev_refthread(dev, &ref);
3714 	dev_strategy_csw(dev, csw, bp);
3715 	dev_relthread(dev, ref);
3716 }
3717 
3718 void
3719 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp)
3720 {
3721 	struct bio *bip;
3722 
3723 	KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE,
3724 	    ("b_iocmd botch"));
3725 	KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) ||
3726 	    dev->si_threadcount > 0,
3727 	    ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev),
3728 	    dev));
3729 	if (csw == NULL) {
3730 		bp->b_error = ENXIO;
3731 		bp->b_ioflags = BIO_ERROR;
3732 		bufdone(bp);
3733 		return;
3734 	}
3735 	for (;;) {
3736 		bip = g_new_bio();
3737 		if (bip != NULL)
3738 			break;
3739 		/* Try again later */
3740 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3741 	}
3742 	bip->bio_cmd = bp->b_iocmd;
3743 	bip->bio_offset = bp->b_iooffset;
3744 	bip->bio_length = bp->b_bcount;
3745 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3746 	bdata2bio(bp, bip);
3747 	bip->bio_done = bufdonebio;
3748 	bip->bio_caller2 = bp;
3749 	bip->bio_dev = dev;
3750 	(*csw->d_strategy)(bip);
3751 }
3752 
3753 /*
3754  *	bufdone:
3755  *
3756  *	Finish I/O on a buffer, optionally calling a completion function.
3757  *	This is usually called from an interrupt so process blocking is
3758  *	not allowed.
3759  *
3760  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3761  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3762  *	assuming B_INVAL is clear.
3763  *
3764  *	For the VMIO case, we set B_CACHE if the op was a read and no
3765  *	read error occured, or if the op was a write.  B_CACHE is never
3766  *	set if the buffer is invalid or otherwise uncacheable.
3767  *
3768  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3769  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3770  *	in the biodone routine.
3771  */
3772 void
3773 bufdone(struct buf *bp)
3774 {
3775 	struct bufobj *dropobj;
3776 	void    (*biodone)(struct buf *);
3777 
3778 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3779 	dropobj = NULL;
3780 
3781 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3782 	BUF_ASSERT_HELD(bp);
3783 
3784 	runningbufwakeup(bp);
3785 	if (bp->b_iocmd == BIO_WRITE)
3786 		dropobj = bp->b_bufobj;
3787 	/* call optional completion function if requested */
3788 	if (bp->b_iodone != NULL) {
3789 		biodone = bp->b_iodone;
3790 		bp->b_iodone = NULL;
3791 		(*biodone) (bp);
3792 		if (dropobj)
3793 			bufobj_wdrop(dropobj);
3794 		return;
3795 	}
3796 
3797 	bufdone_finish(bp);
3798 
3799 	if (dropobj)
3800 		bufobj_wdrop(dropobj);
3801 }
3802 
3803 void
3804 bufdone_finish(struct buf *bp)
3805 {
3806 	BUF_ASSERT_HELD(bp);
3807 
3808 	if (!LIST_EMPTY(&bp->b_dep))
3809 		buf_complete(bp);
3810 
3811 	if (bp->b_flags & B_VMIO) {
3812 		vm_ooffset_t foff;
3813 		vm_page_t m;
3814 		vm_object_t obj;
3815 		struct vnode *vp;
3816 		int bogus, i, iosize;
3817 
3818 		obj = bp->b_bufobj->bo_object;
3819 		KASSERT(obj->paging_in_progress >= bp->b_npages,
3820 		    ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3821 		    obj->paging_in_progress, bp->b_npages));
3822 
3823 		vp = bp->b_vp;
3824 		KASSERT(vp->v_holdcnt > 0,
3825 		    ("biodone_finish: vnode %p has zero hold count", vp));
3826 		KASSERT(vp->v_object != NULL,
3827 		    ("biodone_finish: vnode %p has no vm_object", vp));
3828 
3829 		foff = bp->b_offset;
3830 		KASSERT(bp->b_offset != NOOFFSET,
3831 		    ("biodone_finish: bp %p has no buffer offset", bp));
3832 
3833 		/*
3834 		 * Set B_CACHE if the op was a normal read and no error
3835 		 * occured.  B_CACHE is set for writes in the b*write()
3836 		 * routines.
3837 		 */
3838 		iosize = bp->b_bcount - bp->b_resid;
3839 		if (bp->b_iocmd == BIO_READ &&
3840 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3841 		    !(bp->b_ioflags & BIO_ERROR)) {
3842 			bp->b_flags |= B_CACHE;
3843 		}
3844 		bogus = 0;
3845 		VM_OBJECT_WLOCK(obj);
3846 		for (i = 0; i < bp->b_npages; i++) {
3847 			int bogusflag = 0;
3848 			int resid;
3849 
3850 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3851 			if (resid > iosize)
3852 				resid = iosize;
3853 
3854 			/*
3855 			 * cleanup bogus pages, restoring the originals
3856 			 */
3857 			m = bp->b_pages[i];
3858 			if (m == bogus_page) {
3859 				bogus = bogusflag = 1;
3860 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3861 				if (m == NULL)
3862 					panic("biodone: page disappeared!");
3863 				bp->b_pages[i] = m;
3864 			}
3865 			KASSERT(OFF_TO_IDX(foff) == m->pindex,
3866 			    ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3867 			    (intmax_t)foff, (uintmax_t)m->pindex));
3868 
3869 			/*
3870 			 * In the write case, the valid and clean bits are
3871 			 * already changed correctly ( see bdwrite() ), so we
3872 			 * only need to do this here in the read case.
3873 			 */
3874 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3875 				KASSERT((m->dirty & vm_page_bits(foff &
3876 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3877 				    " page %p has unexpected dirty bits", m));
3878 				vfs_page_set_valid(bp, foff, m);
3879 			}
3880 
3881 			vm_page_sunbusy(m);
3882 			vm_object_pip_subtract(obj, 1);
3883 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3884 			iosize -= resid;
3885 		}
3886 		vm_object_pip_wakeupn(obj, 0);
3887 		VM_OBJECT_WUNLOCK(obj);
3888 		if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
3889 			BUF_CHECK_MAPPED(bp);
3890 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3891 			    bp->b_pages, bp->b_npages);
3892 		}
3893 	}
3894 
3895 	/*
3896 	 * For asynchronous completions, release the buffer now. The brelse
3897 	 * will do a wakeup there if necessary - so no need to do a wakeup
3898 	 * here in the async case. The sync case always needs to do a wakeup.
3899 	 */
3900 
3901 	if (bp->b_flags & B_ASYNC) {
3902 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3903 			brelse(bp);
3904 		else
3905 			bqrelse(bp);
3906 	} else
3907 		bdone(bp);
3908 }
3909 
3910 /*
3911  * This routine is called in lieu of iodone in the case of
3912  * incomplete I/O.  This keeps the busy status for pages
3913  * consistant.
3914  */
3915 void
3916 vfs_unbusy_pages(struct buf *bp)
3917 {
3918 	int i;
3919 	vm_object_t obj;
3920 	vm_page_t m;
3921 
3922 	runningbufwakeup(bp);
3923 	if (!(bp->b_flags & B_VMIO))
3924 		return;
3925 
3926 	obj = bp->b_bufobj->bo_object;
3927 	VM_OBJECT_WLOCK(obj);
3928 	for (i = 0; i < bp->b_npages; i++) {
3929 		m = bp->b_pages[i];
3930 		if (m == bogus_page) {
3931 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3932 			if (!m)
3933 				panic("vfs_unbusy_pages: page missing\n");
3934 			bp->b_pages[i] = m;
3935 			if ((bp->b_flags & B_UNMAPPED) == 0) {
3936 				BUF_CHECK_MAPPED(bp);
3937 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3938 				    bp->b_pages, bp->b_npages);
3939 			} else
3940 				BUF_CHECK_UNMAPPED(bp);
3941 		}
3942 		vm_object_pip_subtract(obj, 1);
3943 		vm_page_sunbusy(m);
3944 	}
3945 	vm_object_pip_wakeupn(obj, 0);
3946 	VM_OBJECT_WUNLOCK(obj);
3947 }
3948 
3949 /*
3950  * vfs_page_set_valid:
3951  *
3952  *	Set the valid bits in a page based on the supplied offset.   The
3953  *	range is restricted to the buffer's size.
3954  *
3955  *	This routine is typically called after a read completes.
3956  */
3957 static void
3958 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3959 {
3960 	vm_ooffset_t eoff;
3961 
3962 	/*
3963 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3964 	 * page boundary and eoff is not greater than the end of the buffer.
3965 	 * The end of the buffer, in this case, is our file EOF, not the
3966 	 * allocation size of the buffer.
3967 	 */
3968 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3969 	if (eoff > bp->b_offset + bp->b_bcount)
3970 		eoff = bp->b_offset + bp->b_bcount;
3971 
3972 	/*
3973 	 * Set valid range.  This is typically the entire buffer and thus the
3974 	 * entire page.
3975 	 */
3976 	if (eoff > off)
3977 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3978 }
3979 
3980 /*
3981  * vfs_page_set_validclean:
3982  *
3983  *	Set the valid bits and clear the dirty bits in a page based on the
3984  *	supplied offset.   The range is restricted to the buffer's size.
3985  */
3986 static void
3987 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3988 {
3989 	vm_ooffset_t soff, eoff;
3990 
3991 	/*
3992 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3993 	 * page boundry or cross the end of the buffer.  The end of the
3994 	 * buffer, in this case, is our file EOF, not the allocation size
3995 	 * of the buffer.
3996 	 */
3997 	soff = off;
3998 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3999 	if (eoff > bp->b_offset + bp->b_bcount)
4000 		eoff = bp->b_offset + bp->b_bcount;
4001 
4002 	/*
4003 	 * Set valid range.  This is typically the entire buffer and thus the
4004 	 * entire page.
4005 	 */
4006 	if (eoff > soff) {
4007 		vm_page_set_validclean(
4008 		    m,
4009 		   (vm_offset_t) (soff & PAGE_MASK),
4010 		   (vm_offset_t) (eoff - soff)
4011 		);
4012 	}
4013 }
4014 
4015 /*
4016  * Ensure that all buffer pages are not exclusive busied.  If any page is
4017  * exclusive busy, drain it.
4018  */
4019 void
4020 vfs_drain_busy_pages(struct buf *bp)
4021 {
4022 	vm_page_t m;
4023 	int i, last_busied;
4024 
4025 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4026 	last_busied = 0;
4027 	for (i = 0; i < bp->b_npages; i++) {
4028 		m = bp->b_pages[i];
4029 		if (vm_page_xbusied(m)) {
4030 			for (; last_busied < i; last_busied++)
4031 				vm_page_sbusy(bp->b_pages[last_busied]);
4032 			while (vm_page_xbusied(m)) {
4033 				vm_page_lock(m);
4034 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4035 				vm_page_busy_sleep(m, "vbpage");
4036 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4037 			}
4038 		}
4039 	}
4040 	for (i = 0; i < last_busied; i++)
4041 		vm_page_sunbusy(bp->b_pages[i]);
4042 }
4043 
4044 /*
4045  * This routine is called before a device strategy routine.
4046  * It is used to tell the VM system that paging I/O is in
4047  * progress, and treat the pages associated with the buffer
4048  * almost as being exclusive busy.  Also the object paging_in_progress
4049  * flag is handled to make sure that the object doesn't become
4050  * inconsistant.
4051  *
4052  * Since I/O has not been initiated yet, certain buffer flags
4053  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
4054  * and should be ignored.
4055  */
4056 void
4057 vfs_busy_pages(struct buf *bp, int clear_modify)
4058 {
4059 	int i, bogus;
4060 	vm_object_t obj;
4061 	vm_ooffset_t foff;
4062 	vm_page_t m;
4063 
4064 	if (!(bp->b_flags & B_VMIO))
4065 		return;
4066 
4067 	obj = bp->b_bufobj->bo_object;
4068 	foff = bp->b_offset;
4069 	KASSERT(bp->b_offset != NOOFFSET,
4070 	    ("vfs_busy_pages: no buffer offset"));
4071 	VM_OBJECT_WLOCK(obj);
4072 	vfs_drain_busy_pages(bp);
4073 	if (bp->b_bufsize != 0)
4074 		vfs_setdirty_locked_object(bp);
4075 	bogus = 0;
4076 	for (i = 0; i < bp->b_npages; i++) {
4077 		m = bp->b_pages[i];
4078 
4079 		if ((bp->b_flags & B_CLUSTER) == 0) {
4080 			vm_object_pip_add(obj, 1);
4081 			vm_page_sbusy(m);
4082 		}
4083 		/*
4084 		 * When readying a buffer for a read ( i.e
4085 		 * clear_modify == 0 ), it is important to do
4086 		 * bogus_page replacement for valid pages in
4087 		 * partially instantiated buffers.  Partially
4088 		 * instantiated buffers can, in turn, occur when
4089 		 * reconstituting a buffer from its VM backing store
4090 		 * base.  We only have to do this if B_CACHE is
4091 		 * clear ( which causes the I/O to occur in the
4092 		 * first place ).  The replacement prevents the read
4093 		 * I/O from overwriting potentially dirty VM-backed
4094 		 * pages.  XXX bogus page replacement is, uh, bogus.
4095 		 * It may not work properly with small-block devices.
4096 		 * We need to find a better way.
4097 		 */
4098 		if (clear_modify) {
4099 			pmap_remove_write(m);
4100 			vfs_page_set_validclean(bp, foff, m);
4101 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4102 		    (bp->b_flags & B_CACHE) == 0) {
4103 			bp->b_pages[i] = bogus_page;
4104 			bogus++;
4105 		}
4106 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4107 	}
4108 	VM_OBJECT_WUNLOCK(obj);
4109 	if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
4110 		BUF_CHECK_MAPPED(bp);
4111 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4112 		    bp->b_pages, bp->b_npages);
4113 	}
4114 }
4115 
4116 /*
4117  *	vfs_bio_set_valid:
4118  *
4119  *	Set the range within the buffer to valid.  The range is
4120  *	relative to the beginning of the buffer, b_offset.  Note that
4121  *	b_offset itself may be offset from the beginning of the first
4122  *	page.
4123  */
4124 void
4125 vfs_bio_set_valid(struct buf *bp, int base, int size)
4126 {
4127 	int i, n;
4128 	vm_page_t m;
4129 
4130 	if (!(bp->b_flags & B_VMIO))
4131 		return;
4132 
4133 	/*
4134 	 * Fixup base to be relative to beginning of first page.
4135 	 * Set initial n to be the maximum number of bytes in the
4136 	 * first page that can be validated.
4137 	 */
4138 	base += (bp->b_offset & PAGE_MASK);
4139 	n = PAGE_SIZE - (base & PAGE_MASK);
4140 
4141 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4142 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4143 		m = bp->b_pages[i];
4144 		if (n > size)
4145 			n = size;
4146 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4147 		base += n;
4148 		size -= n;
4149 		n = PAGE_SIZE;
4150 	}
4151 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4152 }
4153 
4154 /*
4155  *	vfs_bio_clrbuf:
4156  *
4157  *	If the specified buffer is a non-VMIO buffer, clear the entire
4158  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4159  *	validate only the previously invalid portions of the buffer.
4160  *	This routine essentially fakes an I/O, so we need to clear
4161  *	BIO_ERROR and B_INVAL.
4162  *
4163  *	Note that while we only theoretically need to clear through b_bcount,
4164  *	we go ahead and clear through b_bufsize.
4165  */
4166 void
4167 vfs_bio_clrbuf(struct buf *bp)
4168 {
4169 	int i, j, mask, sa, ea, slide;
4170 
4171 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4172 		clrbuf(bp);
4173 		return;
4174 	}
4175 	bp->b_flags &= ~B_INVAL;
4176 	bp->b_ioflags &= ~BIO_ERROR;
4177 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4178 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4179 	    (bp->b_offset & PAGE_MASK) == 0) {
4180 		if (bp->b_pages[0] == bogus_page)
4181 			goto unlock;
4182 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4183 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4184 		if ((bp->b_pages[0]->valid & mask) == mask)
4185 			goto unlock;
4186 		if ((bp->b_pages[0]->valid & mask) == 0) {
4187 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4188 			bp->b_pages[0]->valid |= mask;
4189 			goto unlock;
4190 		}
4191 	}
4192 	sa = bp->b_offset & PAGE_MASK;
4193 	slide = 0;
4194 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4195 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4196 		ea = slide & PAGE_MASK;
4197 		if (ea == 0)
4198 			ea = PAGE_SIZE;
4199 		if (bp->b_pages[i] == bogus_page)
4200 			continue;
4201 		j = sa / DEV_BSIZE;
4202 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4203 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4204 		if ((bp->b_pages[i]->valid & mask) == mask)
4205 			continue;
4206 		if ((bp->b_pages[i]->valid & mask) == 0)
4207 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4208 		else {
4209 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4210 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4211 					pmap_zero_page_area(bp->b_pages[i],
4212 					    sa, DEV_BSIZE);
4213 				}
4214 			}
4215 		}
4216 		bp->b_pages[i]->valid |= mask;
4217 	}
4218 unlock:
4219 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4220 	bp->b_resid = 0;
4221 }
4222 
4223 void
4224 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4225 {
4226 	vm_page_t m;
4227 	int i, n;
4228 
4229 	if ((bp->b_flags & B_UNMAPPED) == 0) {
4230 		BUF_CHECK_MAPPED(bp);
4231 		bzero(bp->b_data + base, size);
4232 	} else {
4233 		BUF_CHECK_UNMAPPED(bp);
4234 		n = PAGE_SIZE - (base & PAGE_MASK);
4235 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4236 			m = bp->b_pages[i];
4237 			if (n > size)
4238 				n = size;
4239 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4240 			base += n;
4241 			size -= n;
4242 			n = PAGE_SIZE;
4243 		}
4244 	}
4245 }
4246 
4247 /*
4248  * vm_hold_load_pages and vm_hold_free_pages get pages into
4249  * a buffers address space.  The pages are anonymous and are
4250  * not associated with a file object.
4251  */
4252 static void
4253 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4254 {
4255 	vm_offset_t pg;
4256 	vm_page_t p;
4257 	int index;
4258 
4259 	BUF_CHECK_MAPPED(bp);
4260 
4261 	to = round_page(to);
4262 	from = round_page(from);
4263 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4264 
4265 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4266 tryagain:
4267 		/*
4268 		 * note: must allocate system pages since blocking here
4269 		 * could interfere with paging I/O, no matter which
4270 		 * process we are.
4271 		 */
4272 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4273 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4274 		if (p == NULL) {
4275 			VM_WAIT;
4276 			goto tryagain;
4277 		}
4278 		pmap_qenter(pg, &p, 1);
4279 		bp->b_pages[index] = p;
4280 	}
4281 	bp->b_npages = index;
4282 }
4283 
4284 /* Return pages associated with this buf to the vm system */
4285 static void
4286 vm_hold_free_pages(struct buf *bp, int newbsize)
4287 {
4288 	vm_offset_t from;
4289 	vm_page_t p;
4290 	int index, newnpages;
4291 
4292 	BUF_CHECK_MAPPED(bp);
4293 
4294 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4295 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4296 	if (bp->b_npages > newnpages)
4297 		pmap_qremove(from, bp->b_npages - newnpages);
4298 	for (index = newnpages; index < bp->b_npages; index++) {
4299 		p = bp->b_pages[index];
4300 		bp->b_pages[index] = NULL;
4301 		if (vm_page_sbusied(p))
4302 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4303 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4304 		p->wire_count--;
4305 		vm_page_free(p);
4306 		atomic_subtract_int(&vm_cnt.v_wire_count, 1);
4307 	}
4308 	bp->b_npages = newnpages;
4309 }
4310 
4311 /*
4312  * Map an IO request into kernel virtual address space.
4313  *
4314  * All requests are (re)mapped into kernel VA space.
4315  * Notice that we use b_bufsize for the size of the buffer
4316  * to be mapped.  b_bcount might be modified by the driver.
4317  *
4318  * Note that even if the caller determines that the address space should
4319  * be valid, a race or a smaller-file mapped into a larger space may
4320  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4321  * check the return value.
4322  */
4323 int
4324 vmapbuf(struct buf *bp, int mapbuf)
4325 {
4326 	caddr_t kva;
4327 	vm_prot_t prot;
4328 	int pidx;
4329 
4330 	if (bp->b_bufsize < 0)
4331 		return (-1);
4332 	prot = VM_PROT_READ;
4333 	if (bp->b_iocmd == BIO_READ)
4334 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4335 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4336 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4337 	    btoc(MAXPHYS))) < 0)
4338 		return (-1);
4339 	bp->b_npages = pidx;
4340 	if (mapbuf || !unmapped_buf_allowed) {
4341 		pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
4342 		kva = bp->b_saveaddr;
4343 		bp->b_saveaddr = bp->b_data;
4344 		bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK);
4345 		bp->b_flags &= ~B_UNMAPPED;
4346 	} else {
4347 		bp->b_flags |= B_UNMAPPED;
4348 		bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4349 		bp->b_saveaddr = bp->b_data;
4350 		bp->b_data = unmapped_buf;
4351 	}
4352 	return(0);
4353 }
4354 
4355 /*
4356  * Free the io map PTEs associated with this IO operation.
4357  * We also invalidate the TLB entries and restore the original b_addr.
4358  */
4359 void
4360 vunmapbuf(struct buf *bp)
4361 {
4362 	int npages;
4363 
4364 	npages = bp->b_npages;
4365 	if (bp->b_flags & B_UNMAPPED)
4366 		bp->b_flags &= ~B_UNMAPPED;
4367 	else
4368 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4369 	vm_page_unhold_pages(bp->b_pages, npages);
4370 
4371 	bp->b_data = bp->b_saveaddr;
4372 }
4373 
4374 void
4375 bdone(struct buf *bp)
4376 {
4377 	struct mtx *mtxp;
4378 
4379 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4380 	mtx_lock(mtxp);
4381 	bp->b_flags |= B_DONE;
4382 	wakeup(bp);
4383 	mtx_unlock(mtxp);
4384 }
4385 
4386 void
4387 bwait(struct buf *bp, u_char pri, const char *wchan)
4388 {
4389 	struct mtx *mtxp;
4390 
4391 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4392 	mtx_lock(mtxp);
4393 	while ((bp->b_flags & B_DONE) == 0)
4394 		msleep(bp, mtxp, pri, wchan, 0);
4395 	mtx_unlock(mtxp);
4396 }
4397 
4398 int
4399 bufsync(struct bufobj *bo, int waitfor)
4400 {
4401 
4402 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4403 }
4404 
4405 void
4406 bufstrategy(struct bufobj *bo, struct buf *bp)
4407 {
4408 	int i = 0;
4409 	struct vnode *vp;
4410 
4411 	vp = bp->b_vp;
4412 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4413 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4414 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4415 	i = VOP_STRATEGY(vp, bp);
4416 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4417 }
4418 
4419 void
4420 bufobj_wrefl(struct bufobj *bo)
4421 {
4422 
4423 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4424 	ASSERT_BO_WLOCKED(bo);
4425 	bo->bo_numoutput++;
4426 }
4427 
4428 void
4429 bufobj_wref(struct bufobj *bo)
4430 {
4431 
4432 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4433 	BO_LOCK(bo);
4434 	bo->bo_numoutput++;
4435 	BO_UNLOCK(bo);
4436 }
4437 
4438 void
4439 bufobj_wdrop(struct bufobj *bo)
4440 {
4441 
4442 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4443 	BO_LOCK(bo);
4444 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4445 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4446 		bo->bo_flag &= ~BO_WWAIT;
4447 		wakeup(&bo->bo_numoutput);
4448 	}
4449 	BO_UNLOCK(bo);
4450 }
4451 
4452 int
4453 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4454 {
4455 	int error;
4456 
4457 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4458 	ASSERT_BO_WLOCKED(bo);
4459 	error = 0;
4460 	while (bo->bo_numoutput) {
4461 		bo->bo_flag |= BO_WWAIT;
4462 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4463 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4464 		if (error)
4465 			break;
4466 	}
4467 	return (error);
4468 }
4469 
4470 void
4471 bpin(struct buf *bp)
4472 {
4473 	struct mtx *mtxp;
4474 
4475 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4476 	mtx_lock(mtxp);
4477 	bp->b_pin_count++;
4478 	mtx_unlock(mtxp);
4479 }
4480 
4481 void
4482 bunpin(struct buf *bp)
4483 {
4484 	struct mtx *mtxp;
4485 
4486 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4487 	mtx_lock(mtxp);
4488 	if (--bp->b_pin_count == 0)
4489 		wakeup(bp);
4490 	mtx_unlock(mtxp);
4491 }
4492 
4493 void
4494 bunpin_wait(struct buf *bp)
4495 {
4496 	struct mtx *mtxp;
4497 
4498 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4499 	mtx_lock(mtxp);
4500 	while (bp->b_pin_count > 0)
4501 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4502 	mtx_unlock(mtxp);
4503 }
4504 
4505 /*
4506  * Set bio_data or bio_ma for struct bio from the struct buf.
4507  */
4508 void
4509 bdata2bio(struct buf *bp, struct bio *bip)
4510 {
4511 
4512 	if ((bp->b_flags & B_UNMAPPED) != 0) {
4513 		KASSERT(unmapped_buf_allowed, ("unmapped"));
4514 		bip->bio_ma = bp->b_pages;
4515 		bip->bio_ma_n = bp->b_npages;
4516 		bip->bio_data = unmapped_buf;
4517 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4518 		bip->bio_flags |= BIO_UNMAPPED;
4519 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4520 		    PAGE_SIZE == bp->b_npages,
4521 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4522 		    (long long)bip->bio_length, bip->bio_ma_n));
4523 	} else {
4524 		bip->bio_data = bp->b_data;
4525 		bip->bio_ma = NULL;
4526 	}
4527 }
4528 
4529 #include "opt_ddb.h"
4530 #ifdef DDB
4531 #include <ddb/ddb.h>
4532 
4533 /* DDB command to show buffer data */
4534 DB_SHOW_COMMAND(buffer, db_show_buffer)
4535 {
4536 	/* get args */
4537 	struct buf *bp = (struct buf *)addr;
4538 
4539 	if (!have_addr) {
4540 		db_printf("usage: show buffer <addr>\n");
4541 		return;
4542 	}
4543 
4544 	db_printf("buf at %p\n", bp);
4545 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4546 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4547 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4548 	db_printf(
4549 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4550 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4551 	    "b_dep = %p\n",
4552 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4553 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4554 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4555 	if (bp->b_npages) {
4556 		int i;
4557 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4558 		for (i = 0; i < bp->b_npages; i++) {
4559 			vm_page_t m;
4560 			m = bp->b_pages[i];
4561 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4562 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4563 			if ((i + 1) < bp->b_npages)
4564 				db_printf(",");
4565 		}
4566 		db_printf("\n");
4567 	}
4568 	db_printf(" ");
4569 	BUF_LOCKPRINTINFO(bp);
4570 }
4571 
4572 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4573 {
4574 	struct buf *bp;
4575 	int i;
4576 
4577 	for (i = 0; i < nbuf; i++) {
4578 		bp = &buf[i];
4579 		if (BUF_ISLOCKED(bp)) {
4580 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4581 			db_printf("\n");
4582 		}
4583 	}
4584 }
4585 
4586 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4587 {
4588 	struct vnode *vp;
4589 	struct buf *bp;
4590 
4591 	if (!have_addr) {
4592 		db_printf("usage: show vnodebufs <addr>\n");
4593 		return;
4594 	}
4595 	vp = (struct vnode *)addr;
4596 	db_printf("Clean buffers:\n");
4597 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4598 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4599 		db_printf("\n");
4600 	}
4601 	db_printf("Dirty buffers:\n");
4602 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4603 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4604 		db_printf("\n");
4605 	}
4606 }
4607 
4608 DB_COMMAND(countfreebufs, db_coundfreebufs)
4609 {
4610 	struct buf *bp;
4611 	int i, used = 0, nfree = 0;
4612 
4613 	if (have_addr) {
4614 		db_printf("usage: countfreebufs\n");
4615 		return;
4616 	}
4617 
4618 	for (i = 0; i < nbuf; i++) {
4619 		bp = &buf[i];
4620 		if ((bp->b_flags & B_INFREECNT) != 0)
4621 			nfree++;
4622 		else
4623 			used++;
4624 	}
4625 
4626 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4627 	    nfree + used);
4628 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4629 }
4630 #endif /* DDB */
4631